refactor(webui): reuse sidebar selection highlight

When an LLM response arrives with finish_reason='length' and has_tool_calls
but blank text content (e.g. the model spent its whole output budget on a
tool call whose closing tag was truncated), the runner dropped the tool
calls and then misrouted the blank response into the empty-response retry
branch. Retrying the same prompt cannot recover from output-budget
exhaustion, so every retry hit the same length ceiling and the turn ended
in the generic apology.

The length-recovery branch was gated on 'finish_reason == length and not
is_blank_text(clean)', so a blank-but-truncated turn could never reach it.

- The empty-response retry branch now excludes finish_reason == 'length'
  (in addition to 'error').
- The length-recovery branch no longer requires non-blank content, so a
  blank-but-truncated turn enters recovery and appends
  build_length_recovery_message (which handles a blank tail safely).

Adds a regression test asserting the length-recovery path is taken; it
fails on the unfixed code and passes with the fix.

Fixes #5133

docs/configuration.md

@@ -348,6 +348,20 @@ Valid `apiType` values are exactly `auto`, `chat_completions`, and `responses`.
 
 </details>
 
+<a id="responses-state-and-compaction"></a>
+
+### Responses conversation state and compaction
+
+Providers that use the Responses API can keep reasoning context across a
+conversation, which helps with multi-step tasks. Supported providers can also
+compact long conversations automatically.
+
+nanobot preserves Responses conversation state automatically for OpenAI
+Responses, OpenAI Codex, Azure OpenAI, and compatible GitHub Copilot models.
+Native compaction is also automatic when the provider supports it. The
+threshold is derived from the active model's context window and reserved output
+headroom; no provider configuration is required.
+
 <details>
 <summary><b>Azure OpenAI</b></summary>
 

docs/providers.md

@@ -229,7 +229,7 @@ Arbitrary custom provider names are OpenAI-compatible only; they do not use the
 }
 ```
 
-`providers.openai.apiType` may be set when you need to force a specific OpenAI API surface. Other providers reject `apiType`; leave it unset outside `providers.openai`. Replace the model with a model ID available to your OpenAI account.
+`providers.openai.apiType` may be set when you need to force a specific OpenAI API surface. Other providers reject `apiType`; leave it unset outside `providers.openai`. Replace the model with a model ID available to your OpenAI account. Direct OpenAI Responses, OpenAI Codex, Azure OpenAI Responses, and eligible GitHub Copilot models share [opaque Responses state retention](./configuration.md#responses-state-and-compaction); native compaction is enabled only where the backend supports it.
 
 ### Custom OpenAI-Compatible Endpoint
 
@@ -458,7 +458,7 @@ For GitHub Copilot:
 nanobot provider login github-copilot --set-main
 ```
 
-Each command authenticates the selected provider and makes its current default model active. OAuth providers are not valid automatic fallbacks. See [`troubleshooting.md`](./troubleshooting.md#provider-and-model-problems) for proxy, headless-login, model-name, and config-key errors.
+Each command authenticates the selected provider and makes its current default model active. OpenAI Codex and eligible GitHub Copilot models participate in [Responses state retention](./configuration.md#responses-state-and-compaction), while native compaction remains provider-capability-specific. OAuth providers are not valid automatic fallbacks. See [`troubleshooting.md`](./troubleshooting.md#provider-and-model-problems) for proxy, headless-login, model-name, and config-key errors.
 
 ## Provider Resolution
 

nanobot/agent/autocompact.py

@@ -31,9 +31,19 @@ class AutoCompact:
                     now: datetime | None = None) -> bool:
         if self._ttl <= 0 or not ts:
             return False
-        if isinstance(ts, str):
+        try:
-            ts = datetime.fromisoformat(ts)
+            if isinstance(ts, str):
-        return ((now or datetime.now()) - ts).total_seconds() >= self._ttl * 60
+                ts = datetime.fromisoformat(ts)
+            current = now or datetime.now()
+            if getattr(ts, "tzinfo", None) is not None or current.tzinfo is not None:
+                idle_seconds = current.timestamp() - ts.timestamp()
+            else:
+                idle_seconds = (current - ts).total_seconds()
+        except (OSError, OverflowError, TypeError, ValueError):
+            # list_sessions() forwards raw persisted metadata; an unusable value
+            # must not escape the idle scan and stop the agent loop.
+            return False
+        return idle_seconds >= self._ttl * 60
 
     def _has_compactable_idle_tail(self, key: str) -> bool:
         session = self.sessions.get_or_create(key)

nanobot/agent/context.py

@@ -225,9 +225,6 @@ class ContextBuilder:
             if current_role == "user"
             else []
         )
-        user_content = self.build_user_content(current_message, image_paths=media)
-        blocks = list(runtime_context_blocks or ()) if current_role == "user" else []
-        merged, runtime_context_meta = append_runtime_context(user_content, blocks)
         messages: list[dict[str, Any]] = [
             {
                 "role": "system",
@@ -243,21 +240,47 @@ class ContextBuilder:
             },
             *history,
         ]
+        current = self.build_current_message(
+            current_message,
+            media=media,
+            current_role=current_role,
+            runtime_context_blocks=runtime_context_blocks,
+        )
         if messages[-1].get("role") == current_role:
             last = dict(messages[-1])
-            last["content"] = self._merge_message_content(last.get("content"), merged)
+            last["content"] = self._merge_message_content(
-            if current_role == "user" and runtime_context_meta is not None:
+                last.get("content"),
+                current.get("content"),
+            )
+            current_meta = current.get("_meta")
+            if current_role == "user" and isinstance(current_meta, dict):
                 internal_meta = dict(last.get("_meta") or {})
-                internal_meta[RUNTIME_CONTEXT_MESSAGE_META] = runtime_context_meta
+                internal_meta.update(cast(dict[str, Any], current_meta))
                 last["_meta"] = internal_meta
             messages[-1] = last
             return messages
-        current: dict[str, Any] = {"role": current_role, "content": merged}
-        if current_role == "user" and runtime_context_meta is not None:
-            current["_meta"] = {RUNTIME_CONTEXT_MESSAGE_META: runtime_context_meta}
         messages.append(current)
         return messages
 
+    def build_current_message(
+        self,
+        current_message: str,
+        *,
+        media: list[str] | None = None,
+        current_role: str = "user",
+        runtime_context_blocks: Sequence[RuntimeContextBlock] | None = None,
+    ) -> dict[str, Any]:
+        """Build only the fresh turn message without merging it into history."""
+        content = self.build_user_content(current_message, image_paths=media)
+        blocks = list(runtime_context_blocks or ()) if current_role == "user" else []
+        merged, runtime_context_meta = append_runtime_context(content, blocks)
+        current: dict[str, Any] = {"role": current_role, "content": merged}
+        if current_role == "user" and runtime_context_meta is not None:
+            current["_meta"] = {
+                RUNTIME_CONTEXT_MESSAGE_META: runtime_context_meta,
+            }
+        return current
+
     def build_user_content(
         self,
         text: str,

nanobot/agent/loop.py

@@ -9,6 +9,7 @@ import dataclasses
 import inspect
 import os
 import time
+import weakref
 from collections.abc import Coroutine, Iterable, Mapping
 from contextlib import AbstractContextManager, ExitStack, nullcontext, suppress
 from dataclasses import dataclass, field
@@ -48,7 +49,7 @@ from nanobot.bus.queue import MessageBus
 from nanobot.bus.runtime_events import RuntimeEventBus
 from nanobot.command import CommandContext, CommandRouter, register_builtin_commands
 from nanobot.config.schema import AgentDefaults, ModelPresetConfig
-from nanobot.providers.base import LLMProvider
+from nanobot.providers.base import LLMProvider, ProviderConversationState
 from nanobot.providers.factory import ProviderSnapshot
 from nanobot.runtime_context import (
     RUNTIME_CONTEXT_HISTORY_META,
@@ -105,6 +106,7 @@ if TYPE_CHECKING:
     from nanobot.triggers.local_store import LocalTriggerStore
 
 _T = TypeVar("_T")
+_SUBAGENT_PROVIDER_TASK_META = "subagent_provider_task_id"
 
 
 class TurnKind(Enum):
@@ -125,6 +127,7 @@ class TurnContext:
 
     history: list[dict[str, Any]] = field(default_factory=list)
     initial_messages: list[dict[str, Any]] = field(default_factory=list)
+    provider_state: ProviderConversationState | None = field(default=None, repr=False)
     request_context: RequestContext | None = None
     runtime_context_blocks: list[RuntimeContextBlock] = field(default_factory=list)
     attributes: dict[str, Any] = field(default_factory=dict)
@@ -242,6 +245,8 @@ class AgentLoop:
 
     _RUNTIME_CHECKPOINT_KEY = "runtime_checkpoint"
     _PENDING_USER_TURN_KEY = "pending_user_turn"
+    _PROVIDER_STATE_CHECKPOINT_VERSION_KEY = "provider_state_checkpoint_version"
+    _PROVIDER_STATE_CHECKPOINT_VERSION = "v1"
 
     def __init__(
         self,
@@ -394,7 +399,9 @@ class AgentLoop:
         self._runtime_context_providers: list[RuntimeContextProvider] = []
         self._active_tasks: dict[str, set[asyncio.Task[Any]]] = {}
         self._background_tasks: set[asyncio.Task[Any]] = set()
-        self._session_locks: dict[str, asyncio.Lock] = {}
+        self._session_locks: weakref.WeakValueDictionary[str, asyncio.Lock] = (
+            weakref.WeakValueDictionary()
+        )
         # Per-session pending queues for mid-turn message injection.
         # When a session has an active task, new messages for that session
         # are routed here instead of creating a new task.
@@ -854,6 +861,7 @@ class AgentLoop:
         turn_scopes: list[AbstractContextManager[Any]] | None = None,
         tools: ToolRegistry | None = None,
         request_context: RequestContext | None = None,
+        provider_state: ProviderConversationState | None = None,
     ) -> tuple[str | None, list[str], list[dict[str, Any]], str, bool]:
         """Run the agent iteration loop.
 
@@ -869,7 +877,18 @@ class AgentLoop:
         async def _checkpoint(payload: dict[str, Any]) -> None:
             if session is None:
                 return
-            self._set_runtime_checkpoint(session, payload)
+            public_payload = dict(payload)
+            private_state = public_payload.pop("provider_state", None)
+            public_payload.pop(self._PROVIDER_STATE_CHECKPOINT_VERSION_KEY, None)
+            if "provider_state" in payload and (
+                private_state is None
+                or isinstance(private_state, ProviderConversationState)
+            ):
+                session.provider_state = private_state
+                public_payload[self._PROVIDER_STATE_CHECKPOINT_VERSION_KEY] = (
+                    self._PROVIDER_STATE_CHECKPOINT_VERSION
+                )
+            self._set_runtime_checkpoint(session, public_payload)
 
         async def _drain_pending(*, limit: int = _MAX_INJECTIONS_PER_TURN) -> list[dict[str, Any]]:
             """Drain follow-up messages from the pending queue.
@@ -1067,6 +1086,7 @@ class AgentLoop:
                     session_metadata=session_metadata,
                     message_metadata=metadata,
                 ),
+                provider_state=provider_state,
             ))
         finally:
             turn_scope_stack.close()
@@ -1074,6 +1094,8 @@ class AgentLoop:
             reset_request_context(request_token)
             reset_file_states(file_state_token)
         self._last_usage = result.usage
+        if session is not None and not ephemeral:
+            session.provider_state = result.provider_state
         if result.stop_reason == "max_iterations":
             logger.warning("Max iterations ({}) reached", self.max_iterations)
             should_stream = turn_continuation.should_stream_budget_response(
@@ -1206,7 +1228,7 @@ class AgentLoop:
         session_key = self._effective_session_key(msg)
         if session_key != msg.session_key:
             msg = dataclasses.replace(msg, session_key_override=session_key)
-        lock = self._session_locks.setdefault(session_key, asyncio.Lock())
+        lock = self._get_session_lock(session_key)
         gate = self._concurrency_gate or nullcontext()
 
         delivery = self.turn_delivery_factory.unrouted(msg, session_key)
@@ -1657,14 +1679,24 @@ class AgentLoop:
             "extend_to_user": is_subagent,
         }
         ctx.history = session.get_history(**_hist_kwargs)
+        stored_state = session.provider_state
+        subagent_followup_persisted = False
         if is_subagent:
             # Keep the durable internal delivery as an assistant record, but
             # present this completion to the model as fresh follow-up input.
             # Providers without assistant-prefill support drop trailing
             # assistant messages, so using the persisted record as the current
             # prompt would hide an independently dispatched subagent result.
-            if self._persist_subagent_followup(session, ctx.msg):
+            subagent_followup_persisted = self._persist_subagent_followup(
+                session,
+                ctx.msg,
+            )
+            if subagent_followup_persisted:
                 logger.debug("Subagent result persisted for session {}", ctx.session_key)
+                # Establish a durable, replay-safe baseline before any fallible
+                # provider compatibility or prompt assembly work. A compatible
+                # staged state replaces this in a second atomic save below.
+                session.provider_state = None
                 self.sessions.save(session)
             ctx.input_persisted_early = True
         ctx.delivery.record_runtime(runtime)
@@ -1672,13 +1704,65 @@ class AgentLoop:
         ctx.request_context = self._request_context_for_turn(ctx)
         if ctx.kind is TurnKind.USER:
             ctx.runtime_context_blocks = await self._resolve_runtime_context_for_turn(ctx)
-        ctx.initial_messages = self._build_initial_messages(ctx)
+        staged_provider_state = False
+        if stored_state is not None and runtime.provider.can_resume_conversation_state(
+            stored_state,
+            runtime.model,
+        ):
+            current_provider_message = self.context.build_current_message(
+                ctx.msg.content,
+                media=ctx.msg.media if ctx.kind is TurnKind.USER and ctx.msg.media else None,
+                runtime_context_blocks=ctx.runtime_context_blocks,
+            )
+            task_id = ctx.msg.metadata.get("subagent_task_id") if is_subagent else None
+            already_staged = False
+            if isinstance(task_id, str) and task_id:
+                internal_meta = current_provider_message.get("_meta")
+                current_provider_message["_meta"] = {
+                    **(
+                        cast(dict[str, Any], internal_meta)
+                        if isinstance(internal_meta, dict)
+                        else {}
+                    ),
+                    _SUBAGENT_PROVIDER_TASK_META: task_id,
+                }
+                already_staged = any(
+                    isinstance(message.get("_meta"), dict)
+                    and cast(dict[str, Any], message["_meta"]).get(
+                        _SUBAGENT_PROVIDER_TASK_META
+                    )
+                    == task_id
+                    for message in stored_state.pending_messages
+                )
+            ctx.provider_state = (
+                stored_state
+                if already_staged
+                else stored_state.with_pending_messages([
+                    *stored_state.pending_messages,
+                    current_provider_message,
+                ])
+            )
+            if (
+                not ctx.ephemeral
+                and (ctx.kind is TurnKind.USER or subagent_followup_persisted)
+            ):
+                session.provider_state = ctx.provider_state
+                staged_provider_state = True
+        elif stored_state is not None:
+            session.provider_state = None
         if ctx.kind is TurnKind.USER:
             ctx.input_persisted_early = self._persist_user_message_early(
                 ctx.msg,
                 session,
                 runtime_context_blocks=ctx.runtime_context_blocks,
             )
+            if staged_provider_state and not ctx.input_persisted_early:
+                session.provider_state = stored_state
+        elif subagent_followup_persisted and staged_provider_state:
+            # Upgrade the replay-safe baseline to the resumable state before
+            # prompt assembly and the first model checkpoint.
+            self.sessions.save(session)
+        ctx.initial_messages = self._build_initial_messages(ctx)
 
         if ctx.on_progress is None:
             ctx.on_progress = ctx.delivery.progress_callback()
@@ -1712,6 +1796,7 @@ class AgentLoop:
             turn_scopes=ctx.turn_scopes,
             tools=ctx.tools,
             request_context=ctx.request_context,
+            provider_state=ctx.provider_state,
         )
         final_content, _, all_msgs, stop_reason, had_injections = result
         ctx.final_content = final_content
@@ -2049,7 +2134,36 @@ class AgentLoop:
             ):
                 overlap = size
                 break
-        session.messages.extend(restored_messages[overlap:])
+        appended_messages = restored_messages[overlap:]
+        session.messages.extend(appended_messages)
+        assistant_message_data = (
+            cast(dict[str, Any], assistant_message)
+            if isinstance(assistant_message, dict)
+            else None
+        )
+        provider_state_is_synchronized = (
+            checkpoint_data.get(self._PROVIDER_STATE_CHECKPOINT_VERSION_KEY)
+            == self._PROVIDER_STATE_CHECKPOINT_VERSION
+        )
+        phase = checkpoint_data.get("phase")
+        exact_final_response = (
+            phase == "final_response"
+            and assistant_message_data is not None
+            and assistant_message_data.get("role") == "assistant"
+            and not bool(checkpoint_data.get("completed_tool_results"))
+            and not bool(checkpoint_data.get("pending_tool_calls"))
+        )
+        exact_completed_tools = (
+            phase == "tools_completed"
+            and assistant_message_data is not None
+            and assistant_message_data.get("role") == "assistant"
+            and not bool(checkpoint_data.get("pending_tool_calls"))
+        )
+        if not (
+            provider_state_is_synchronized
+            and (exact_final_response or exact_completed_tools)
+        ):
+            session.provider_state = None
 
         self._clear_pending_user_turn(session)
         self._clear_runtime_checkpoint(session)
@@ -2070,6 +2184,7 @@ class AgentLoop:
                     "timestamp": datetime.now().isoformat(),
                 }
             )
+            session.provider_state = None
             session.updated_at = datetime.now()
 
         self._clear_pending_user_turn(session)
@@ -2108,7 +2223,7 @@ class AgentLoop:
             content=content, media=media or [], metadata=metadata,
         )
         # Share the dispatch lock so direct calls serialize with bus turns.
-        lock = self._session_locks.setdefault(session_key, asyncio.Lock())
+        lock = self._get_session_lock(session_key)
         try:
             async with lock:
                 kwargs: dict[str, Any] = {
@@ -2139,3 +2254,11 @@ class AgentLoop:
         finally:
             await self.runtime_event_publisher.run_status_changed(msg, session_key, "idle")
             self.runtime_event_publisher.clear_turn(session_key)
+
+    def _get_session_lock(self, session_key: str) -> asyncio.Lock:
+        """Return the shared lock while allowing idle session entries to expire."""
+        lock = self._session_locks.get(session_key)
+        if lock is None:
+            lock = asyncio.Lock()
+            self._session_locks[session_key] = lock
+        return lock

nanobot/agent/memory.py

@@ -931,6 +931,7 @@ class Consolidator:
             session_key=session.key,
         )
         session.last_consolidated = end_idx
+        session.provider_state = None
         self.sessions.save(session)
         return summary
 
@@ -1136,6 +1137,7 @@ class Consolidator:
                 if summary:
                     last_summary = summary
                 session.last_consolidated = end_idx
+                session.provider_state = None
                 self.sessions.save(session)
                 if not summary:
                     # LLM is degraded — stop hammering it this call;
@@ -1205,6 +1207,7 @@ class Consolidator:
 
             # Preserve history and advance only the replay boundary.
             session.last_consolidated = len(session.messages) - len(visible_suffix)
+            session.provider_state = None
             self.sessions.save(session)
 
             logger.info(

nanobot/agent/runner.py

@@ -19,7 +19,17 @@ from nanobot.agent.context_governance import (
 )
 from nanobot.agent.hook import AgentHook, AgentHookContext, AgentRunHookContext
 from nanobot.agent.tools.registry import ToolRegistry, is_tool_error_result
-from nanobot.providers.base import LLMProvider, LLMResponse, ToolCallRequest
+from nanobot.providers.base import (
+    LLMProvider,
+    LLMResponse,
+    ProviderCallContext,
+    ProviderConversationState,
+    ToolCallRequest,
+)
+from nanobot.providers.conversation_state import (
+    ProviderConversationStateController,
+    allows_conversation_message_merge,
+)
 from nanobot.runtime_context import (
     RUNTIME_CONTEXT_MESSAGE_META,
     detach_runtime_context,
@@ -104,6 +114,7 @@ class AgentRunSpec:
     goal_active_predicate: Callable[[], bool] | None = None
     goal_continue_message: GoalContinueMessage | None = None
     finalize_on_max_iterations: bool = True
+    provider_state: ProviderConversationState | None = None
 
 
 @dataclass(slots=True)
@@ -120,6 +131,7 @@ class AgentRunResult:
     had_injections: bool = False
     # Terminal tail to emit when the preceding final-content prefix was already streamed.
     pending_stream_content: str | None = None
+    provider_state: ProviderConversationState | None = field(default=None, repr=False)
 
 
 class AgentRunner:
@@ -161,6 +173,7 @@ class AgentRunner:
                 and messages[-1].get("role") == "user"
                 and not is_hidden_history_message(injection)
                 and not is_hidden_history_message(messages[-1])
+                and allows_conversation_message_merge(messages[-1])
             ):
                 merged = dict(messages[-1])
                 left_meta = merged.get("_meta")
@@ -231,6 +244,7 @@ class AgentRunner:
         assistant_message: dict[str, Any] | None,
         injection_cycles: int,
         *,
+        conversation_state: ProviderConversationStateController | None = None,
         phase: str = "after error",
         iteration: int | None = None,
         allow_goal_continue: bool = False,
@@ -258,16 +272,21 @@ class AgentRunner:
         if assistant_message is not None:
             messages.append(assistant_message)
             if iteration is not None:
+                checkpoint: dict[str, Any] = {
+                    "phase": "final_response",
+                    "iteration": iteration,
+                    "model": spec.runtime.model,
+                    "assistant_message": assistant_message,
+                    "completed_tool_results": [],
+                    "pending_tool_calls": [],
+                }
+                if conversation_state is not None:
+                    checkpoint["provider_state"] = conversation_state.checkpoint(
+                        messages
+                    )
                 await self._emit_checkpoint(
                     spec,
-                    {
+                    checkpoint,
-                        "phase": "final_response",
-                        "iteration": iteration,
-                        "model": spec.runtime.model,
-                        "assistant_message": assistant_message,
-                        "completed_tool_results": [],
-                        "pending_tool_calls": [],
-                    },
                 )
         self._append_injected_messages(messages, injections)
         if real_injection:
@@ -420,6 +439,12 @@ class AgentRunner:
         injection_cycles = 0
         compacted_tool_call_ids: set[str] = set()
         pending_stream_content: str | None = None
+        conversation_state = ProviderConversationStateController(
+            provider=spec.runtime.provider,
+            model=spec.runtime.model,
+            messages=messages,
+            state=spec.provider_state,
+        )
         governance_config = ContextGovernanceConfig(
             provider=spec.runtime.provider,
             model=spec.runtime.model,
@@ -450,7 +475,20 @@ class AgentRunner:
                 session_key=spec.session_key,
             )
             await hook.before_iteration(context)
-            response = await self._request_model(spec, messages_for_model, hook, context)
+            provider_context = conversation_state.prepare_request(
+                messages,
+                context_window_tokens=spec.runtime.context_window_tokens,
+                model_messages=messages_for_model,
+            )
+            response = await self._request_model(
+                spec,
+                messages_for_model,
+                hook,
+                context,
+                conversation_state=conversation_state,
+                provider_context=provider_context,
+            )
+            conversation_state.observe_response(response, messages)
             context.response = response
             context.tool_calls = list(response.tool_calls)
 
@@ -480,6 +518,10 @@ class AgentRunner:
                     reasoning_content=response.reasoning_content,
                     thinking_blocks=response.thinking_blocks,
                 )
+                assistant_message = conversation_state.project_response_message(
+                    assistant_message,
+                    response,
+                )
                 messages.append(assistant_message)
                 await self._emit_checkpoint(
                     spec,
@@ -544,6 +586,15 @@ class AgentRunner:
                         length_recovery_parts.clear()
                         continue
                     break
+                checkpoint_model_messages = (
+                    self.context_governor.prepare_for_model(
+                        governance_config,
+                        messages,
+                        compacted_tool_call_ids,
+                    )
+                    if response.provider_state is not None
+                    else None
+                )
                 await self._emit_checkpoint(
                     spec,
                     {
@@ -553,6 +604,10 @@ class AgentRunner:
                         "assistant_message": assistant_message,
                         "completed_tool_results": completed_tool_results,
                         "pending_tool_calls": [],
+                        "provider_state": conversation_state.checkpoint(
+                            messages,
+                            model_messages=checkpoint_model_messages,
+                        ),
                     },
                 )
                 empty_content_retries = 0
@@ -575,7 +630,11 @@ class AgentRunner:
                 )
 
             clean = hook.finalize_content(context, response.content)
-            if response.finish_reason != "error" and is_blank_text(clean):
+            if (
+                response.finish_reason
+                not in {"error", "length", "refusal", "content_filter"}
+                and is_blank_text(clean)
+            ):
                 empty_content_retries += 1
                 if empty_content_retries < _MAX_EMPTY_RETRIES:
                     logger.warning(
@@ -598,7 +657,12 @@ class AgentRunner:
                 if hook.wants_streaming():
                     await hook.on_stream_end(context, resuming=False)
                 retry_messages = self._finalization_retry_messages(messages_for_model)
-                response = await self._request_finalization_retry(spec, messages_for_model)
+                response = await self._request_finalization_retry(
+                    spec,
+                    messages_for_model,
+                    transcript=messages,
+                    conversation_state=conversation_state,
+                )
                 retry_usage = self._usage_or_estimate(spec, retry_messages, response)
                 self._accumulate_usage(usage, retry_usage)
                 raw_usage = self._merge_usage(raw_usage, retry_usage)
@@ -608,7 +672,7 @@ class AgentRunner:
                 original_content = response.content
                 clean = hook.finalize_content(context, response.content)
 
-            if response.finish_reason == "length" and not is_blank_text(clean):
+            if response.finish_reason == "length":
                 if len(length_recovery_parts) < _MAX_LENGTH_RECOVERIES:
                     length_recovery_parts.append(
                         _restore_outer_whitespace(clean or "", original_content)
@@ -623,10 +687,13 @@ class AgentRunner:
                     if hook.wants_streaming():
                         context.stream_continues_current_message = True
                         await hook.on_stream_end(context, resuming=True)
-                    messages.append(build_assistant_message(
+                    messages.append(conversation_state.project_response_message(
-                        clean,
+                        build_assistant_message(
-                        reasoning_content=response.reasoning_content,
+                            clean,
-                        thinking_blocks=response.thinking_blocks,
+                            reasoning_content=response.reasoning_content,
+                            thinking_blocks=response.thinking_blocks,
+                        ),
+                        response,
                     ))
                     messages.append(build_length_recovery_message(clean or ""))
                     await hook.after_iteration(context)
@@ -656,15 +723,22 @@ class AgentRunner:
                     reasoning_content=response.reasoning_content,
                     thinking_blocks=response.thinking_blocks,
                 )
+                assistant_message = conversation_state.project_response_message(
+                    assistant_message,
+                    response,
+                )
 
             # Check for mid-turn injections BEFORE signaling stream end.
             # If injections are found we keep the stream alive (resuming=True)
             # so streaming channels don't prematurely finalize the card.
             should_continue, injection_cycles = await self._try_drain_injections(
                 spec, messages, assistant_message, injection_cycles,
+                conversation_state=conversation_state,
                 phase="after final response",
                 iteration=iteration,
-                allow_goal_continue=True,
+                allow_goal_continue=(
+                    response.finish_reason not in {"refusal", "content_filter"}
+                ),
             )
             if should_continue:
                 had_injections = True
@@ -717,11 +791,17 @@ class AgentRunner:
                     continue
                 break
 
-            messages.append(assistant_message or build_assistant_message(
+            messages.append(
-                clean,
+                assistant_message
-                reasoning_content=response.reasoning_content,
+                or conversation_state.project_response_message(
-                thinking_blocks=response.thinking_blocks,
+                    build_assistant_message(
-            ))
+                        clean,
+                        reasoning_content=response.reasoning_content,
+                        thinking_blocks=response.thinking_blocks,
+                    ),
+                    response,
+                )
+            )
             await self._emit_checkpoint(
                 spec,
                 {
@@ -731,6 +811,7 @@ class AgentRunner:
                     "assistant_message": messages[-1],
                     "completed_tool_results": [],
                     "pending_tool_calls": [],
+                    "provider_state": conversation_state.checkpoint(messages),
                 },
             )
             if length_recovery_parts:
@@ -764,6 +845,7 @@ class AgentRunner:
                     hook,
                     messages,
                     usage,
+                    conversation_state,
                 )
             if terminal_content is None:
                 terminal_content = self._max_iterations_fallback(spec)
@@ -787,6 +869,7 @@ class AgentRunner:
             tool_events=tool_events,
             had_injections=had_injections,
             pending_stream_content=pending_stream_content,
+            provider_state=conversation_state.finish(messages),
         )
 
     def _build_request_kwargs(
@@ -817,6 +900,8 @@ class AgentRunner:
         context: AgentHookContext,
         *,
         malformed_retry: bool = False,
+        conversation_state: ProviderConversationStateController,
+        provider_context: ProviderCallContext | None = None,
     ) -> LLMResponse:
         timeout_s: float | None = spec.llm_timeout_s
         if timeout_s is None:
@@ -886,6 +971,7 @@ class AgentRunner:
 
             coro = spec.runtime.provider.chat_stream_with_retry(
                 **kwargs,
+                provider_context=provider_context,
                 on_content_delta=_stream,
                 on_thinking_delta=_thinking,
                 on_tool_call_delta=_provider_tool_event,
@@ -920,11 +1006,15 @@ class AgentRunner:
 
             coro = spec.runtime.provider.chat_stream_with_retry(
                 **kwargs,
+                provider_context=provider_context,
                 on_content_delta=_stream_progress,
                 on_tool_call_delta=_provider_tool_event,
             )
         else:
-            coro = spec.runtime.provider.chat_with_retry(**kwargs)
+            coro = spec.runtime.provider.chat_with_retry(
+                **kwargs,
+                provider_context=provider_context,
+            )
 
         # Streaming requests also have provider-level idle timeouts
         # (NANOBOT_STREAM_IDLE_TIMEOUT_S), but a stream that keeps producing
@@ -986,6 +1076,10 @@ class AgentRunner:
             return await self._request_model(
                 spec, retry_messages, hook, context,
                 malformed_retry=True,
+                conversation_state=conversation_state,
+                provider_context=conversation_state.independent_request_context(
+                    context_window_tokens=spec.runtime.context_window_tokens,
+                ),
             )
         if (
             all_dropped
@@ -998,7 +1092,13 @@ class AgentRunner:
             fallback_messages = self._malformed_tool_call_retry_messages(
                 messages, response.content,
             )
-            return await self._request_no_tools(spec, fallback_messages)
+            return await self._request_no_tools(
+                spec,
+                fallback_messages,
+                provider_context=conversation_state.independent_request_context(
+                    context_window_tokens=spec.runtime.context_window_tokens,
+                ),
+            )
         return response
 
     @staticmethod
@@ -1031,6 +1131,10 @@ class AgentRunner:
             original_finish_reason,
         )
         response.tool_calls = valid
+        # The opaque candidate still contains every raw function_call item.
+        # Advancing it after dropping even one call would replay an unmatched
+        # call without a corresponding tool output on the next request.
+        response.provider_state = None
         if not valid:
             response.finish_reason = "stop"
         return (dropped, not valid, original_finish_reason)
@@ -1060,9 +1164,27 @@ class AgentRunner:
         self,
         spec: AgentRunSpec,
         messages: list[dict[str, Any]],
+        *,
+        transcript: list[dict[str, Any]],
+        conversation_state: ProviderConversationStateController,
     ) -> LLMResponse:
         retry_messages = self._finalization_retry_messages(messages)
-        return await self._request_no_tools(spec, retry_messages)
+        provider_context = conversation_state.prepare_request(
+            transcript,
+            context_window_tokens=spec.runtime.context_window_tokens,
+            supplemental_messages=[retry_messages[-1]],
+        )
+        response = await self._request_no_tools(
+            spec,
+            retry_messages,
+            provider_context=provider_context,
+        )
+        conversation_state.observe_response(
+            response,
+            transcript,
+            adopt_candidate_state=False,
+        )
+        return response
 
     @staticmethod
     def _finalization_retry_messages(messages: list[dict[str, Any]]) -> list[dict[str, Any]]:
@@ -1076,10 +1198,17 @@ class AgentRunner:
         hook: AgentHook,
         messages: list[dict[str, Any]],
         usage: dict[str, int],
+        conversation_state: ProviderConversationStateController,
     ) -> str | None:
         retry_messages = self._budget_exhausted_finalization_messages(messages)
         try:
-            response = await self._request_no_tools(spec, retry_messages)
+            response = await self._request_no_tools(
+                spec,
+                retry_messages,
+                provider_context=conversation_state.independent_request_context(
+                    context_window_tokens=spec.runtime.context_window_tokens,
+                ),
+            )
         except Exception:
             logger.exception(
                 "Budget-exhausted finalization failed for {}; using fallback",
@@ -1115,9 +1244,18 @@ class AgentRunner:
         self,
         spec: AgentRunSpec,
         messages: list[dict[str, Any]],
+        *,
+        provider_context: ProviderCallContext | None = None,
     ) -> LLMResponse:
-        kwargs = self._build_request_kwargs(spec, messages, tools=None)
+        kwargs = self._build_request_kwargs(
-        return await spec.runtime.provider.chat_with_retry(**kwargs)
+            spec,
+            messages,
+            tools=None,
+        )
+        return await spec.runtime.provider.chat_with_retry(
+            **kwargs,
+            provider_context=provider_context,
+        )
 
     @staticmethod
     def _budget_exhausted_finalization_messages(

nanobot/agent/tools/exec_session.py

@@ -5,6 +5,7 @@ from __future__ import annotations
 import asyncio
 import time
 import uuid
+from collections import deque
 from contextlib import suppress
 from dataclasses import dataclass
 from typing import Any
@@ -51,6 +52,66 @@ class ExecSessionInfo:
     owner_session_key: str | None = None
 
 
+class _BoundedOutputBuffer:
+    """Keep the first and most recent characters within a fixed budget."""
+
+    def __init__(self, max_chars: int) -> None:
+        self.max_chars = max_chars
+        self._content = ""
+        self._tail: deque[str] = deque()
+        self._tail_chars = 0
+        self._total_chars = 0
+        self._truncated = False
+
+    @property
+    def has_output(self) -> bool:
+        return self._total_chars > 0
+
+    @property
+    def retained_chars(self) -> int:
+        return len(self._content) + self._tail_chars
+
+    def append(self, text: str) -> None:
+        if not text:
+            return
+        self._total_chars += len(text)
+        if not self._truncated:
+            combined = self._content + text
+            if len(combined) <= self.max_chars:
+                self._content = combined
+                return
+            head_chars = self.max_chars // 2
+            tail_chars = self.max_chars - head_chars
+            self._content = combined[:head_chars]
+            self._tail.append(combined[-tail_chars:])
+            self._tail_chars = tail_chars
+            self._truncated = True
+            return
+
+        tail_chars = self.max_chars - len(self._content)
+        self._tail.append(text)
+        self._tail_chars += len(text)
+        while self._tail_chars > tail_chars:
+            excess = self._tail_chars - tail_chars
+            first = self._tail[0]
+            if len(first) <= excess:
+                self._tail.popleft()
+                self._tail_chars -= len(first)
+            else:
+                self._tail[0] = first[excess:]
+                self._tail_chars -= excess
+
+    def drain(self) -> tuple[str, int]:
+        output = self._content + "".join(self._tail)
+        truncated_chars = self._total_chars - len(output)
+        self._content = ""
+        self._tail.clear()
+        self._tail_chars = 0
+        self._total_chars = 0
+        self._truncated = False
+        return output, truncated_chars
+
+
 class _ExecSession:
     def __init__(
         self,
@@ -73,30 +134,27 @@ class _ExecSession:
         # timeout None/0 means no limit; an infinite deadline is never reached.
         self.deadline = time.monotonic() + timeout if timeout else float("inf")
         self.last_access = time.monotonic()
-        self._chunks: list[str] = []
+        self._stdout = _BoundedOutputBuffer(MAX_OUTPUT_CHARS)
+        self._stderr = _BoundedOutputBuffer(MAX_OUTPUT_CHARS)
         self._lock = asyncio.Lock()
         self._timed_out = False
-        self._stdout_task = asyncio.create_task(self._read_stream(process.stdout, ""))
+        self._stdout_task = asyncio.create_task(self._read_stream(process.stdout, self._stdout))
-        self._stderr_task = asyncio.create_task(self._read_stream(process.stderr, "STDERR:\n"))
+        self._stderr_task = asyncio.create_task(self._read_stream(process.stderr, self._stderr))
 
     async def _read_stream(
         self,
         stream: asyncio.StreamReader | None,
-        prefix: str,
+        buffer: _BoundedOutputBuffer,
     ) -> None:
         if stream is None:
             return
-        first = True
         while True:
             chunk = await stream.read(4096)
             if not chunk:
                 break
             text = chunk.decode("utf-8", errors="replace")
-            if prefix and first:
-                text = prefix + text
-                first = False
             async with self._lock:
-                self._chunks.append(text)
+                buffer.append(text)
 
     async def write(self, chars: str) -> str | None:
         if self.process.returncode is not None:
@@ -157,10 +215,14 @@ class _ExecSession:
             await self._wait_for_buffered_output()
 
         async with self._lock:
-            output = "".join(self._chunks)
+            stdout, stdout_truncated = self._stdout.drain()
-            self._chunks.clear()
+            stderr, stderr_truncated = self._stderr.drain()
 
-        output, truncated = _truncate_output(output, max_output_chars)
+        output_parts = [stdout] if stdout else []
+        if stderr:
+            output_parts.append(f"STDERR:\n{stderr}")
+        output = "\n".join(output_parts)
+        output, response_truncated = _truncate_output(output, max_output_chars)
         return _SessionPoll(
             output=output,
             done=self.process.returncode is not None,
@@ -169,7 +231,7 @@ class _ExecSession:
             timed_out=self._timed_out,
             terminated=terminated,
             stdin_closed=stdin_closed,
-            truncated_chars=truncated,
+            truncated_chars=stdout_truncated + stderr_truncated + response_truncated,
         )
 
     async def kill(self) -> None:
@@ -195,7 +257,7 @@ class _ExecSession:
         deadline = time.monotonic() + OUTPUT_DRAIN_GRACE_S
         while time.monotonic() < deadline:
             async with self._lock:
-                if self._chunks:
+                if self._stdout.has_output or self._stderr.has_output:
                     return
             await asyncio.sleep(0.01)
 
@@ -403,20 +465,16 @@ def clamp_session_int(value: int | None, default: int, minimum: int, maximum: in
 def _truncate_output(output: str, max_output_chars: int) -> tuple[str, int]:
     if len(output) <= max_output_chars:
         return output, 0
-    half = max_output_chars // 2
+    head_chars = max_output_chars // 2
+    tail_chars = max_output_chars - head_chars
     omitted = len(output) - max_output_chars
-    return (
+    return output[:head_chars] + output[-tail_chars:], omitted
-        output[:half]
-        + f"\n\n... ({omitted:,} chars truncated) ...\n\n"
-        + output[-half:],
-        omitted,
-    )
 
 
 def format_session_poll(session_id: str, poll: _SessionPoll) -> str:
     parts = [poll.output] if poll.output else []
     if poll.truncated_chars:
-        parts.append(f"(output truncated by {poll.truncated_chars:,} chars)")
+        parts.append(f"({poll.truncated_chars:,} chars truncated from output)")
     if poll.timed_out:
         parts.append("Error: Command timed out; session was terminated.")
     if poll.terminated and not poll.timed_out:
@@ -587,7 +645,9 @@ class WriteStdinTool(Tool):
         max_output_chars: int,
     ) -> str:
         deadline = time.monotonic() + (wait_timeout_ms / 1000)
-        aggregate: list[str] = []
+        aggregate = _BoundedOutputBuffer(max_output_chars)
+        upstream_truncated = 0
+        search_overlap = ""
         first = True
         poll: _SessionPoll | None = None
 
@@ -604,15 +664,20 @@ class WriteStdinTool(Tool):
                 owner_session_key=current_request_session_key(),
             )
             first = False
+            upstream_truncated += poll.truncated_chars
             if poll.output:
                 aggregate.append(poll.output)
-                joined = "".join(aggregate)
+                searchable = search_overlap + poll.output
-                if wait_for in joined:
+                if wait_for in searchable:
-                    poll.output = joined
+                    poll.output, aggregate_truncated = aggregate.drain()
+                    poll.truncated_chars = upstream_truncated + aggregate_truncated
                     result = format_session_poll(session_id, poll)
                     return ToolResult.error(result) if poll.timed_out else result
+                overlap_chars = max(0, len(wait_for) - 1)
+                search_overlap = searchable[-overlap_chars:] if overlap_chars else ""
             if poll.done or remaining_ms <= 0:
-                poll.output = "".join(aggregate)
+                poll.output, aggregate_truncated = aggregate.drain()
+                poll.truncated_chars = upstream_truncated + aggregate_truncated
                 result = format_session_poll(session_id, poll)
                 if wait_for not in poll.output:
                     result += f"\nWait target not observed: {wait_for!r}"

nanobot/channels/base.py

@@ -248,7 +248,15 @@ class BaseChannel(ABC):
         permission_id = authorization_id if authorization_id is not None else sender_id
         if not self.is_allowed(permission_id):
             if is_dm:
-                code = generate_code(self.name, str(sender_id))
+                try:
+                    code = generate_code(self.name, str(sender_id))
+                except OSError:
+                    # Transient pairing-store I/O failure: skip the pairing
+                    # reply for this message rather than crash the handler.
+                    self.logger.warning(
+                        "Pairing store unavailable; dropping DM from {}", sender_id
+                    )
+                    return
                 await self.send(
                     OutboundMessage(
                         channel=self.name,

nanobot/channels/websocket/runtime.py

@@ -1216,6 +1216,7 @@ class WebSocketChannel(BaseChannel):
             body,
             metadata=meta,
             phase="answer",
+            include_source=True,
         )
         raw = json.dumps(body, ensure_ascii=False)
         if not conns:

nanobot/channels/websocket/tests/test_websocket_channel.py

@@ -51,6 +51,7 @@ from nanobot.webui.http_utils import (
 )
 from nanobot.webui.metadata import (
     WEBSOCKET_TURN_OWNER_METADATA_KEY,
+    WEBUI_MESSAGE_SOURCE_METADATA_KEY,
     WEBUI_SYSTEM_COMMAND_TURN_PREFIX,
     WEBUI_TURN_METADATA_KEY,
 )
@@ -1350,6 +1351,35 @@ async def test_send_delta_emits_delta_and_stream_end() -> None:
     assert "text" not in second
 
 
+@pytest.mark.asyncio
+async def test_send_delta_preserves_webui_source_metadata() -> None:
+    bus = MagicMock()
+    channel = WebSocketChannel({"enabled": True, "allowFrom": ["*"], "streaming": True}, bus, gateway=_basic_handler(bus))
+    mock_ws = AsyncMock()
+    channel._attach(mock_ws, "chat-source-stream")
+    source = {"kind": "cron", "label": "Repo check"}
+    metadata = {WEBUI_MESSAGE_SOURCE_METADATA_KEY: source}
+
+    await channel.send_delta("chat-source-stream", "done", metadata=metadata, stream_id="sid")
+    await channel.send_delta(
+        "chat-source-stream",
+        "",
+        metadata=metadata,
+        stream_id="sid",
+        stream_end=True,
+    )
+
+    first = json.loads(mock_ws.send.call_args_list[0][0][0])
+    second = json.loads(mock_ws.send.call_args_list[1][0][0])
+    assert first["event"] == "delta"
+    assert first["source"] == source
+    assert second["event"] == "stream_end"
+    assert second["source"] == source
+    lines = read_transcript_lines("websocket:chat-source-stream")
+    assert lines[-2]["source"] == source
+    assert lines[-1]["source"] == source
+
+
 @pytest.mark.asyncio
 async def test_send_delta_marks_resuming_stream_end() -> None:
     bus = MagicMock()

nanobot/pairing/store.py

@@ -40,9 +40,15 @@ def _load() -> dict[str, Any]:
             data = json.load(f)
     except FileNotFoundError:
         return {"approved": {}, "pending": {}}
-    except (json.JSONDecodeError, OSError):
+    except json.JSONDecodeError:
         logger.warning("Corrupted pairing store, resetting")
         return {"approved": {}, "pending": {}}
+    except OSError:
+        # A transiently locked or busy file is not corruption. Propagate so
+        # mutating callers fail loudly instead of persisting an empty view
+        # that would erase every approved sender.
+        logger.warning("Pairing store temporarily unreadable: {}", path)
+        raise
     if not isinstance(data, dict):
         logger.warning("Corrupted pairing store, resetting")
         return {"approved": {}, "pending": {}}
@@ -171,7 +177,11 @@ def deny_code(code: str) -> bool:
 def is_approved(channel: str, sender_id: str) -> bool:
     """Check whether *sender_id* has been approved on *channel*."""
     with _LOCK:
-        data = _load()
+        try:
+            data = _load()
+        except OSError:
+            # Fail closed for this check; the store itself stays untouched.
+            return False
         approved: dict[str, set[str]] = data.get("approved", {})
         return str(sender_id) in approved.get(channel, set())
 
@@ -179,7 +189,10 @@ def is_approved(channel: str, sender_id: str) -> bool:
 def list_pending() -> list[dict[str, Any]]:
     """Return all non-expired pending pairing requests."""
     with _LOCK:
-        data = _load()
+        try:
+            data = _load()
+        except OSError:
+            return []
         _gc_pending(data)
         return [
             {"code": code, **info}
@@ -257,7 +270,10 @@ def clear_channel(channel: str) -> dict[str, int]:
 def get_approved(channel: str) -> list[str]:
     """Return all approved sender IDs for *channel*."""
     with _LOCK:
-        data = _load()
+        try:
+            data = _load()
+        except OSError:
+            return []
         return sorted(data.get("approved", {}).get(channel, set()))
 
 
@@ -283,6 +299,15 @@ def handle_pairing_command(channel: str, subcommand_text: str) -> str:
     This is a pure function (no side effects other than store mutations)
     so it can be used from both the CLI and the agent CommandRouter.
     """
+    try:
+        return _handle_pairing_subcommand(channel, subcommand_text)
+    except OSError:
+        # Mutations fail loudly on a transient I/O error instead of lying
+        # ("invalid code") or silently rewriting the store from an empty view.
+        return "The pairing store is temporarily unavailable. Please try again."
+
+
+def _handle_pairing_subcommand(channel: str, subcommand_text: str) -> str:
     parts = subcommand_text.split()
     sub = parts[0] if parts else "list"
     arg = parts[1] if len(parts) > 1 else None

nanobot/providers/azure_openai_provider.py

@@ -23,14 +23,26 @@ import uuid
 from collections.abc import Awaitable, Callable
 from typing import Any, cast
 
+from loguru import logger
 from openai import AsyncOpenAI
 
-from nanobot.providers.base import LLMProvider, LLMResponse
+from nanobot.providers.base import (
+    LLMProvider,
+    LLMResponse,
+    ProviderCallContext,
+    ProviderConversationState,
+)
 from nanobot.providers.openai_responses import (
+    ResponsesStreamCapture,
+    build_responses_state,
     consume_sdk_stream,
-    convert_messages,
     convert_tools,
+    is_compaction_compatibility_error,
+    is_replayable_finish_reason,
     parse_response_output,
+    prepare_responses_input,
+    resolve_compact_threshold,
+    responses_state_matches,
 )
 
 _AZURE_OPENAI_SCOPE = "https://cognitiveservices.azure.com/.default"
@@ -97,6 +109,7 @@ class AzureOpenAIProvider(LLMProvider):
     ):
         super().__init__(api_key, api_base)
         self.default_model = default_model
+        self._native_compaction_available = True
 
         if not api_base:
             raise ValueError("Azure OpenAI api_base is required")
@@ -142,6 +155,25 @@ class AzureOpenAIProvider(LLMProvider):
         name = deployment_name.lower()
         return not any(token in name for token in ("gpt-5", "o1", "o3", "o4"))
 
+    def _responses_state_provider(self) -> str:
+        return f"azure_openai:{str(self.api_base).rstrip('/')}"
+
+    def can_resume_conversation_state(
+        self,
+        state: ProviderConversationState,
+        model: str | None = None,
+    ) -> bool:
+        return responses_state_matches(
+            state,
+            provider=self._responses_state_provider(),
+            model=model or self.default_model,
+        )
+
+    def supports_native_compaction(self, model: str | None = None) -> bool:
+        """Azure's native Responses endpoint accepts context management."""
+        _ = model
+        return self._native_compaction_available
+
     def _build_body(
         self,
         messages: list[dict[str, Any]],
@@ -151,10 +183,26 @@ class AzureOpenAIProvider(LLMProvider):
         temperature: float,
         reasoning_effort: str | None,
         tool_choice: str | dict[str, Any] | None,
+        provider_context: ProviderCallContext | None = None,
     ) -> dict[str, Any]:
         """Build the Responses API request body from Chat-Completions-style args."""
         deployment = model or self.default_model
-        instructions, input_items = convert_messages(self._sanitize_empty_content(messages))
+        sanitized_messages = self._sanitize_empty_content(messages)
+        sanitized_state = (
+            provider_context.conversation_state
+            if provider_context is not None
+            else None
+        )
+        if sanitized_state is not None:
+            sanitized_state = sanitized_state.with_pending_messages(
+                self._sanitize_empty_content(sanitized_state.pending_messages)
+            )
+        instructions, input_items, replayed = prepare_responses_input(
+            sanitized_messages,
+            state=sanitized_state,
+            provider=self._responses_state_provider(),
+            model=deployment,
+        )
 
         body: dict[str, Any] = {
             "model": deployment,
@@ -164,13 +212,29 @@ class AzureOpenAIProvider(LLMProvider):
             "store": False,
             "stream": False,
         }
+        compact_threshold = resolve_compact_threshold(
+            (
+                provider_context.context_window_tokens
+                if provider_context is not None
+                else None
+            ),
+            max_tokens,
+        )
+        if self.supports_native_compaction(deployment) and compact_threshold is not None:
+            body["context_management"] = [{
+                "type": "compaction",
+                "compact_threshold": compact_threshold,
+            }]
 
         if self._supports_temperature(deployment, reasoning_effort):
             body["temperature"] = temperature
 
+        if not self._supports_temperature(deployment, reasoning_effort):
+            body["include"] = ["reasoning.encrypted_content"]
         if reasoning_effort and reasoning_effort.lower() != "none":
             body["reasoning"] = {"effort": reasoning_effort}
-            body["include"] = ["reasoning.encrypted_content"]
+        if replayed and "gpt-5.6" in deployment.lower():
+            body.setdefault("reasoning", {})["context"] = "all_turns"
 
         if tools:
             body["tools"] = convert_tools(tools)
@@ -178,21 +242,97 @@ class AzureOpenAIProvider(LLMProvider):
 
         return body
 
+    async def _create_response_with_compaction_fallback(
+        self,
+        body: dict[str, Any],
+    ) -> Any:
+        """Retry once without server compaction when Azure rejects the option."""
+        try:
+            return cast(Any, await self._client.responses.create(**body))
+        except Exception as exc:
+            if (
+                "context_management" not in body
+                or not is_compaction_compatibility_error(exc)
+            ):
+                raise
+            self._native_compaction_available = False
+            body.pop("context_management", None)
+            logger.warning(
+                "Azure Responses server compaction unsupported; disabled for this provider "
+                "instance (status={})",
+                getattr(exc, "status_code", None),
+            )
+            return cast(Any, await self._client.responses.create(**body))
+
     @staticmethod
     def _handle_error(e: Exception) -> LLMResponse:
         response = getattr(e, "response", None)
         body = getattr(e, "body", None) or getattr(response, "text", None)
         body_text = str(body).strip() if body is not None else ""
         msg = f"Error: {body_text[:500]}" if body_text else f"Error calling Azure OpenAI: {e}"
-        retry_after = LLMProvider._extract_retry_after_from_headers(getattr(response, "headers", None))
+        headers = getattr(response, "headers", None)
+        retry_after = LLMProvider._extract_retry_after_from_headers(headers)
         if retry_after is None:
             retry_after = LLMProvider._extract_retry_after(msg)
-        return LLMResponse(content=msg, finish_reason="error", retry_after=retry_after)
+        status_code = getattr(e, "status_code", None)
+        if status_code is None and response is not None:
+            status_code = getattr(response, "status_code", None)
+        error_type, error_code = LLMProvider._extract_error_type_code(body)
+        should_retry: bool | None = None
+        if headers is not None:
+            raw_should_retry = headers.get("x-should-retry")
+            if isinstance(raw_should_retry, str):
+                lowered = raw_should_retry.strip().lower()
+                if lowered == "true":
+                    should_retry = True
+                elif lowered == "false":
+                    should_retry = False
+        error_name = type(e).__name__.lower()
+        error_kind = (
+            "timeout"
+            if "timeout" in error_name
+            else "connection"
+            if "connection" in error_name
+            else None
+        )
+        return LLMResponse(
+            content=msg,
+            finish_reason="error",
+            retry_after=retry_after,
+            error_status_code=int(status_code) if status_code is not None else None,
+            error_kind=error_kind,
+            error_type=error_type,
+            error_code=error_code,
+            error_retry_after_s=retry_after,
+            error_should_retry=should_retry,
+        )
 
     # ------------------------------------------------------------------
     # Public API
     # ------------------------------------------------------------------
 
+    async def chat_with_context(
+        self,
+        *,
+        provider_context: ProviderCallContext,
+        **kwargs: Any,
+    ) -> LLMResponse:
+        return await self.chat(
+            **kwargs,
+            provider_context=provider_context,
+        )
+
+    async def chat_stream_with_context(
+        self,
+        *,
+        provider_context: ProviderCallContext,
+        **kwargs: Any,
+    ) -> LLMResponse:
+        return await self.chat_stream(
+            **kwargs,
+            provider_context=provider_context,
+        )
+
     async def chat(
         self,
         messages: list[dict[str, Any]],
@@ -202,14 +342,21 @@ class AzureOpenAIProvider(LLMProvider):
         temperature: float = 0.7,
         reasoning_effort: str | None = None,
         tool_choice: str | dict[str, Any] | None = None,
+        provider_context: ProviderCallContext | None = None,
     ) -> LLMResponse:
         body = self._build_body(
             messages, tools, model, max_tokens, temperature,
             reasoning_effort, tool_choice,
+            provider_context,
         )
         try:
-            response = cast(Any, await self._client.responses.create(**body))
+            response = await self._create_response_with_compaction_fallback(body)
-            return parse_response_output(response)
+            return parse_response_output(
+                response,
+                state_provider=self._responses_state_provider(),
+                state_model=str(body["model"]),
+                state_input_items=cast(list[dict[str, Any]], body["input"]),
+            )
         except Exception as e:
             return self._handle_error(e)
 
@@ -225,26 +372,43 @@ class AzureOpenAIProvider(LLMProvider):
         on_content_delta: Callable[[str], Awaitable[None]] | None = None,
         on_thinking_delta: Callable[[str], Awaitable[None]] | None = None,
         on_tool_call_delta: Callable[[dict[str, Any]], Awaitable[None]] | None = None,
+        provider_context: ProviderCallContext | None = None,
     ) -> LLMResponse:
         _ = on_thinking_delta
         body = self._build_body(
             messages, tools, model, max_tokens, temperature,
             reasoning_effort, tool_choice,
+            provider_context,
         )
         body["stream"] = True
 
         try:
-            stream = cast(Any, await self._client.responses.create(**body))
+            stream = await self._create_response_with_compaction_fallback(body)
+            capture = ResponsesStreamCapture()
             content, tool_calls, finish_reason, usage, reasoning_content = (
-                await consume_sdk_stream(stream, on_content_delta, on_tool_call_delta)
+                await consume_sdk_stream(
+                    stream,
+                    on_content_delta,
+                    on_tool_call_delta,
+                    capture=capture,
+                )
             )
-            return LLMResponse(
+            result = LLMResponse(
                 content=content or None,
                 tool_calls=tool_calls,
                 finish_reason=finish_reason,
                 usage=usage,
                 reasoning_content=reasoning_content,
             )
+            if capture.completed and is_replayable_finish_reason(finish_reason):
+                result.provider_state = build_responses_state(
+                    provider=self._responses_state_provider(),
+                    model=str(body["model"]),
+                    input_items=cast(list[dict[str, Any]], body["input"]),
+                    output_items=capture.output_items,
+                    usage=usage,
+                )
+            return result
         except Exception as e:
             return self._handle_error(e)
 

nanobot/providers/base.py

@@ -1,5 +1,7 @@
 """Base LLM provider interface."""
 
+from __future__ import annotations
+
 import asyncio
 import json
 import os
@@ -7,6 +9,7 @@ import re
 from abc import ABC, abstractmethod
 from collections.abc import Awaitable, Callable
 from contextlib import suppress
+from copy import deepcopy
 from dataclasses import dataclass, field
 from datetime import datetime, timezone
 from email.utils import parsedate_to_datetime
@@ -150,6 +153,104 @@ def tool_arguments_json_for_replay(arguments: Any) -> str:
     return json.dumps(tool_arguments_object_for_replay(arguments), ensure_ascii=False)
 
 
+@dataclass
+class ProviderConversationState:
+    """Opaque provider-owned continuation state.
+
+    ``payload`` may contain encrypted reasoning or other provider-private
+    protocol items. Keep it out of normal logs and public chat history.
+    ``pending_messages`` are Chat-style messages produced after the most
+    recent provider response and are materialized by the owning provider on
+    the next request.
+    """
+
+    kind: str
+    provider: str
+    model: str
+    version: int
+    payload: dict[str, Any] = field(default_factory=dict, repr=False)
+    pending_messages: list[dict[str, Any]] = field(default_factory=list, repr=False)
+
+    def with_pending_messages(
+        self,
+        messages: list[dict[str, Any]],
+    ) -> ProviderConversationState:
+        """Return a state copy with an isolated pending-message list."""
+        return ProviderConversationState(
+            kind=self.kind,
+            provider=self.provider,
+            model=self.model,
+            version=self.version,
+            payload=self.payload,
+            pending_messages=deepcopy(messages),
+        )
+
+    def to_private_record(self) -> dict[str, Any]:
+        """Serialize for the private session sidecar, never for public history."""
+        return {
+            "kind": self.kind,
+            "provider": self.provider,
+            "model": self.model,
+            "version": self.version,
+            "payload": deepcopy(self.payload),
+            "pending_messages": deepcopy(self.pending_messages),
+        }
+
+    @classmethod
+    def from_private_record(
+        cls,
+        value: object,
+    ) -> ProviderConversationState | None:
+        """Validate and deserialize a private session-sidecar value."""
+        if not isinstance(value, dict):
+            return None
+        data = cast(dict[str, Any], value)
+        kind = data.get("kind")
+        provider = data.get("provider")
+        model = data.get("model")
+        version = data.get("version")
+        payload = data.get("payload")
+        pending = data.get("pending_messages", [])
+        if (
+            not isinstance(kind, str)
+            or not kind
+            or not isinstance(provider, str)
+            or not provider
+            or not isinstance(model, str)
+            or not model
+            or isinstance(version, bool)
+            or not isinstance(version, int)
+            or not isinstance(payload, dict)
+            or not isinstance(pending, list)
+            or any(
+                not isinstance(message, dict)
+                for message in cast(list[object], pending)
+            )
+        ):
+            return None
+        return cls(
+            kind=kind,
+            provider=provider,
+            model=model,
+            version=version,
+            payload=deepcopy(cast(dict[str, Any], payload)),
+            pending_messages=deepcopy(cast(list[dict[str, Any]], pending)),
+        )
+
+
+@dataclass(frozen=True)
+class ProviderCallContext:
+    """Optional provider-owned continuation data for one model request.
+
+    The regular ``chat`` contract stays provider-agnostic. Responses-capable
+    providers consume this context through the opt-in ``chat_with_context``
+    hooks, while every other provider inherits the context-free delegation.
+    """
+
+    conversation_state: ProviderConversationState | None = field(default=None, repr=False)
+    context_window_tokens: int | None = None
+
+
 @dataclass
 class LLMResponse:
     """Response from an LLM provider."""
@@ -160,6 +261,10 @@ class LLMResponse:
     retry_after: float | None = None  # Provider supplied retry wait in seconds.
     reasoning_content: str | None = None  # Kimi, DeepSeek-R1, MiMo etc.
     thinking_blocks: list[dict[str, Any]] | None = None  # Anthropic extended thinking
+    provider_state: ProviderConversationState | None = field(default=None, repr=False)
+    # Routing wrappers may preserve or discard an incoming provider-owned
+    # continuation independently of the final fallback error's retry policy.
+    preserve_provider_state_on_error: bool | None = field(default=None, repr=False)
     # Structured error metadata used by retry policy when finish_reason == "error".
     error_status_code: int | None = None
     error_kind: str | None = None  # e.g. "timeout", "connection"
@@ -274,6 +379,18 @@ class LLMProvider(ABC):
         self.api_base = api_base
         self.generation: GenerationSettings = GenerationSettings()
 
+    def can_resume_conversation_state(
+        self,
+        state: ProviderConversationState,
+        model: str | None = None,
+    ) -> bool:
+        """Whether this provider can safely consume an opaque saved state."""
+        return False
+
+    def supports_native_compaction(self, model: str | None = None) -> bool:
+        """Whether requests may include provider-native context compaction."""
+        return False
+
     @staticmethod
     def _sanitize_empty_content(messages: list[dict[str, Any]]) -> list[dict[str, Any]]:
         """Sanitize message content: fix empty blocks, strip internal _meta fields.
@@ -416,7 +533,7 @@ class LLMProvider(ABC):
         return any(marker in err for marker in cls._TRANSIENT_ERROR_MARKERS)
 
     @classmethod
-    def _is_transient_response(cls, response: LLMResponse) -> bool:
+    def is_transient_response(cls, response: LLMResponse) -> bool:
         """Prefer structured error metadata, fallback to text markers for legacy providers."""
         if response.error_should_retry is not None:
             return bool(response.error_should_retry)
@@ -607,6 +724,21 @@ class LLMProvider(ABC):
                 result.append(msg)
         return result if found else None
 
+    @staticmethod
+    def _contains_image_content(value: object) -> bool:
+        """Return whether a JSON-like provider payload contains an input image."""
+        if isinstance(value, dict):
+            mapping = cast(dict[str, object], value)
+            if mapping.get("type") in {"image_url", "input_image"}:
+                return True
+            return any(LLMProvider._contains_image_content(item) for item in mapping.values())
+        if isinstance(value, list):
+            return any(
+                LLMProvider._contains_image_content(item)
+                for item in cast(list[object], value)
+            )
+        return False
+
     @staticmethod
     def _strip_image_content_inplace(messages: list[dict[str, Any]]) -> bool:
         """Replace image_url blocks with text placeholder *in-place*.
@@ -633,6 +765,12 @@ class LLMProvider(ABC):
     async def _safe_chat(self, **kwargs: Any) -> LLMResponse:
         """Call chat() and convert unexpected exceptions to error responses."""
         try:
+            provider_context = kwargs.pop("provider_context", None)
+            if isinstance(provider_context, ProviderCallContext):
+                return await self.chat_with_context(
+                    provider_context=provider_context,
+                    **kwargs,
+                )
             return await self.chat(**kwargs)
         except asyncio.CancelledError:
             raise
@@ -666,17 +804,47 @@ class LLMProvider(ABC):
         """
         _ = on_thinking_delta, on_tool_call_delta
         response = await self.chat(
-            messages=messages, tools=tools, model=model,
+            messages=messages,
-            max_tokens=max_tokens, temperature=temperature,
+            tools=tools,
-            reasoning_effort=reasoning_effort, tool_choice=tool_choice,
+            model=model,
+            max_tokens=max_tokens,
+            temperature=temperature,
+            reasoning_effort=reasoning_effort,
+            tool_choice=tool_choice,
         )
         if on_content_delta and response.content:
             await on_content_delta(response.content)
         return response
 
+    async def chat_with_context(
+        self,
+        *,
+        provider_context: ProviderCallContext,
+        **kwargs: Any,
+    ) -> LLMResponse:
+        """Opt-in continuation hook; ordinary providers delegate to ``chat``."""
+        _ = provider_context
+        return await self.chat(**kwargs)
+
+    async def chat_stream_with_context(
+        self,
+        *,
+        provider_context: ProviderCallContext,
+        **kwargs: Any,
+    ) -> LLMResponse:
+        """Streaming continuation hook with a context-free default."""
+        _ = provider_context
+        return await self.chat_stream(**kwargs)
+
     async def _safe_chat_stream(self, **kwargs: Any) -> LLMResponse:
         """Call chat_stream() and convert unexpected exceptions to error responses."""
         try:
+            provider_context = kwargs.pop("provider_context", None)
+            if isinstance(provider_context, ProviderCallContext):
+                return await self.chat_stream_with_context(
+                    provider_context=provider_context,
+                    **kwargs,
+                )
             return await self.chat_stream(**kwargs)
         except asyncio.CancelledError:
             raise
@@ -698,6 +866,7 @@ class LLMProvider(ABC):
         on_stream_recover: Callable[[], Awaitable[None]] | None = None,
         retry_mode: str = "standard",
         on_retry_wait: Callable[[str], Awaitable[None]] | None = None,
+        provider_context: ProviderCallContext | None = None,
     ) -> LLMResponse:
         """Call chat_stream() with retry on transient provider failures."""
         if max_tokens is self._SENTINEL or max_tokens is None:
@@ -730,6 +899,8 @@ class LLMProvider(ABC):
             on_thinking_delta=on_thinking_delta,
             on_tool_call_delta=on_tool_call_delta,
         )
+        if provider_context is not None:
+            kw["provider_context"] = provider_context
         if on_stream_recover and getattr(self, "supports_stream_recover_callback", False):
             kw["on_stream_recover"] = _recover_stream
         return await self._run_with_retry(
@@ -753,6 +924,7 @@ class LLMProvider(ABC):
         tool_choice: str | dict[str, Any] | None = None,
         retry_mode: str = "standard",
         on_retry_wait: Callable[[str], Awaitable[None]] | None = None,
+        provider_context: ProviderCallContext | None = None,
     ) -> LLMResponse:
         """Call chat() with retry on transient provider failures.
 
@@ -775,6 +947,8 @@ class LLMProvider(ABC):
             max_tokens=max_tokens, temperature=temperature,
             reasoning_effort=reasoning_effort, tool_choice=tool_choice,
         )
+        if provider_context is not None:
+            kw["provider_context"] = provider_context
         return await self._run_with_retry(
             self._safe_chat,
             kw,
@@ -932,14 +1106,33 @@ class LLMProvider(ABC):
                 last_error_key = error_key
                 identical_error_count = 1 if error_key else 0
 
-            if not self._is_transient_response(response):
+            if not self.is_transient_response(response):
-                stripped = self._strip_image_content(original_messages)
+                stripped = self._strip_image_content(kw["messages"])
-                if stripped is not None and stripped != kw["messages"]:
+                provider_context = kw.get("provider_context")
+                stripped_context: ProviderCallContext | None = None
+                if isinstance(provider_context, ProviderCallContext):
+                    state = provider_context.conversation_state
+                    if state is not None and (
+                        stripped is not None
+                        or self._strip_image_content(state.pending_messages) is not None
+                        or self._contains_image_content(state.payload)
+                    ):
+                        # Provider-owned payloads may retain earlier input_image items.
+                        # Rebuild from the stripped public transcript for this retry.
+                        stripped_context = ProviderCallContext(
+                            context_window_tokens=(
+                                provider_context.context_window_tokens
+                            ),
+                        )
+                if stripped is not None or stripped_context is not None:
                     logger.warning(
                         "Non-transient LLM error with image content, retrying without images"
                     )
                     retry_kw = dict(kw)
-                    retry_kw["messages"] = stripped
+                    if stripped is not None:
+                        retry_kw["messages"] = stripped
+                    if stripped_context is not None:
+                        retry_kw["provider_context"] = stripped_context
                     result = await call(**retry_kw)
                     # Permanently strip images from the original messages so
                     # subsequent iterations do not repeat the error-retry cycle.

nanobot/providers/conversation_state.py

@@ -0,0 +1,262 @@
+"""Provider-owned conversation-state lifecycle coordination."""
+
+from __future__ import annotations
+
+from copy import deepcopy
+from typing import Any, cast
+
+from nanobot.providers.base import (
+    LLMProvider,
+    LLMResponse,
+    ProviderCallContext,
+    ProviderConversationState,
+)
+
+_PROVIDER_STATE_OUTPUT_META = "provider_state_output"
+_PROVIDER_STATE_BOUNDARY_META = "provider_state_boundary"
+
+
+def allows_conversation_message_merge(message: dict[str, Any]) -> bool:
+    """Return whether new same-role input may merge into *message*."""
+    internal_meta = cast(object, message.get("_meta"))
+    return not (
+        isinstance(internal_meta, dict)
+        and cast(dict[str, Any], internal_meta).get(
+            _PROVIDER_STATE_BOUNDARY_META
+        ) is True
+    )
+
+
+class ProviderConversationStateController:
+    """Keep provider conversation-state semantics outside the agent runner.
+
+    The runner owns the tool loop and reports lifecycle events here. This
+    controller owns capability checks, transcript deltas, response projections,
+    retry transitions, and durable snapshots for provider-private state.
+    """
+
+    def __init__(
+        self,
+        *,
+        provider: LLMProvider,
+        model: str | None,
+        messages: list[dict[str, Any]],
+        state: ProviderConversationState | None = None,
+    ) -> None:
+        self._provider = provider
+        self._model = model
+        self._state = (
+            state
+            if state is not None
+            and provider.can_resume_conversation_state(state, model)
+            else None
+        )
+        self._boundary = len(messages)
+        self._request_messages: list[dict[str, Any]] = []
+
+    def independent_request_context(
+        self,
+        *,
+        context_window_tokens: int | None,
+    ) -> ProviderCallContext | None:
+        """Return typed provider context for a request that does not resume state."""
+        if context_window_tokens is None:
+            return None
+        return ProviderCallContext(context_window_tokens=context_window_tokens)
+
+    def prepare_request(
+        self,
+        messages: list[dict[str, Any]],
+        *,
+        context_window_tokens: int | None,
+        model_messages: list[dict[str, Any]] | None = None,
+        supplemental_messages: list[dict[str, Any]] | None = None,
+    ) -> ProviderCallContext | None:
+        """Build typed context for the next request and remember its durable delta."""
+        independent_context = self.independent_request_context(
+            context_window_tokens=context_window_tokens,
+        )
+        if self._state is None:
+            self._request_messages = []
+            return independent_context
+        if not self._provider.can_resume_conversation_state(
+            self._state,
+            self._model,
+        ):
+            self._state = None
+            self._request_messages = []
+            return independent_context
+
+        durable_messages = self._messages_after_boundary(messages)
+        governed_messages = (
+            self._model_messages_after_boundary(model_messages)
+            if model_messages is not None and durable_messages
+            else None
+        )
+        request_messages = (
+            governed_messages
+            if governed_messages is not None
+            else durable_messages
+        )
+        supplemental = deepcopy(supplemental_messages or [])
+        self._request_messages = deepcopy(request_messages)
+        request_state = self._state.with_pending_messages([
+            *self._state.pending_messages,
+            *request_messages,
+            *supplemental,
+        ])
+        return ProviderCallContext(
+            conversation_state=request_state,
+            context_window_tokens=(
+                independent_context.context_window_tokens
+                if independent_context is not None
+                else None
+            ),
+        )
+
+    def observe_response(
+        self,
+        response: LLMResponse,
+        messages: list[dict[str, Any]],
+        *,
+        adopt_candidate_state: bool = True,
+    ) -> None:
+        """Advance, preserve, or discard state after one provider response."""
+        candidate = response.provider_state if adopt_candidate_state else None
+        candidate_is_replayable = response.finish_reason in {
+            "stop",
+            "tool_calls",
+            "function_call",
+        }
+        if (
+            candidate is not None
+            and candidate_is_replayable
+            and self._provider.can_resume_conversation_state(
+                candidate,
+                self._model,
+            )
+        ):
+            self._state = candidate
+            self._boundary = len(messages)
+            self._seal_boundary(messages)
+        elif response.finish_reason == "error" and (
+            response.preserve_provider_state_on_error is True
+            or (
+                response.preserve_provider_state_on_error is None
+                and LLMProvider.is_transient_response(response)
+            )
+        ):
+            if self._state is not None and self._request_messages:
+                self._state = self._state.with_pending_messages([
+                    *self._state.pending_messages,
+                    *self._request_messages,
+                ])
+            self._boundary = len(messages)
+        else:
+            self._state = None
+            self._boundary = len(messages)
+        self._request_messages = []
+
+    @staticmethod
+    def project_response_message(
+        message: dict[str, Any],
+        response: LLMResponse,
+    ) -> dict[str, Any]:
+        """Mark a Chat projection already represented by provider output."""
+        if response.provider_state is None:
+            return message
+        internal_meta = dict(message.get("_meta") or {})
+        internal_meta[_PROVIDER_STATE_OUTPUT_META] = True
+        message["_meta"] = internal_meta
+        return message
+
+    def checkpoint(
+        self,
+        messages: list[dict[str, Any]],
+        *,
+        model_messages: list[dict[str, Any]] | None = None,
+    ) -> ProviderConversationState | None:
+        """Return a durable state snapshot without changing live state."""
+        if self._state is None:
+            return None
+        durable_messages = self._messages_after_boundary(messages)
+        governed_messages = (
+            self._model_messages_after_boundary(model_messages)
+            if model_messages is not None and durable_messages
+            else None
+        )
+        pending_messages = (
+            governed_messages
+            if governed_messages is not None
+            else durable_messages
+        )
+        return self._state.with_pending_messages([
+            *self._state.pending_messages,
+            *pending_messages,
+        ])
+
+    def finish(
+        self,
+        messages: list[dict[str, Any]],
+    ) -> ProviderConversationState | None:
+        """Return the final durable state after all runner messages are known."""
+        self._state = self.checkpoint(messages)
+        return self._state
+
+    def _messages_after_boundary(
+        self,
+        messages: list[dict[str, Any]],
+    ) -> list[dict[str, Any]]:
+        pending: list[dict[str, Any]] = []
+        for message in messages[self._boundary:]:
+            internal_meta = cast(object, message.get("_meta"))
+            if (
+                isinstance(internal_meta, dict)
+                and cast(dict[str, Any], internal_meta).get(
+                    _PROVIDER_STATE_OUTPUT_META
+                ) is True
+            ):
+                continue
+            pending.append(deepcopy(message))
+        return pending
+
+    @staticmethod
+    def _model_messages_after_boundary(
+        messages: list[dict[str, Any]],
+    ) -> list[dict[str, Any]] | None:
+        """Return the governed delta after the latest provider-owned boundary."""
+        boundary = None
+        for idx in range(len(messages) - 1, -1, -1):
+            internal_meta = cast(object, messages[idx].get("_meta"))
+            if (
+                isinstance(internal_meta, dict)
+                and cast(dict[str, Any], internal_meta).get(
+                    _PROVIDER_STATE_BOUNDARY_META
+                ) is True
+            ):
+                boundary = idx
+                break
+        if boundary is None:
+            return None
+
+        pending: list[dict[str, Any]] = []
+        for message in messages[boundary + 1:]:
+            internal_meta = cast(object, message.get("_meta"))
+            if (
+                isinstance(internal_meta, dict)
+                and cast(dict[str, Any], internal_meta).get(
+                    _PROVIDER_STATE_OUTPUT_META
+                ) is True
+            ):
+                continue
+            pending.append(deepcopy(message))
+        return pending
+
+    @staticmethod
+    def _seal_boundary(messages: list[dict[str, Any]]) -> None:
+        """Prevent later same-role injection merging across a state boundary."""
+        if not messages:
+            return
+        internal_meta = dict(messages[-1].get("_meta") or {})
+        internal_meta[_PROVIDER_STATE_BOUNDARY_META] = True
+        messages[-1]["_meta"] = internal_meta

nanobot/providers/factory.py

@@ -261,6 +261,7 @@ def make_provider(
             primary=provider,
             fallback_presets=fallback_presets,
             provider_factory=lambda fb: _make_provider_core(config, preset=fb),
+            primary_context_window_tokens=resolved.context_window_tokens,
         )
 
     return provider

nanobot/providers/fallback_provider.py

@@ -6,11 +6,18 @@ from __future__ import annotations
 
 import time
 from collections.abc import Awaitable, Callable
+from dataclasses import replace
 from typing import Any
 
 from loguru import logger
 
-from nanobot.providers.base import GenerationSettings, LLMProvider, LLMResponse
+from nanobot.providers.base import (
+    GenerationSettings,
+    LLMProvider,
+    LLMResponse,
+    ProviderCallContext,
+    ProviderConversationState,
+)
 
 # Circuit breaker tuned to match OpenAICompatProvider's Responses API breaker.
 _PRIMARY_FAILURE_THRESHOLD = 3
@@ -113,11 +120,13 @@ class FallbackProvider(LLMProvider):
         fallback_presets: list[Any],
         provider_factory: Callable[[Any], LLMProvider],
         fallback_model_observer: FallbackModelObserver | None = None,
+        primary_context_window_tokens: int | None = None,
     ):
         self._primary = primary
         self._fallback_presets = list(fallback_presets)
         self._provider_factory = provider_factory
         self._fallback_model_observer = fallback_model_observer
+        self._primary_context_window_tokens = primary_context_window_tokens
         self._has_fallbacks = bool(fallback_presets)
         self._primary_failures = 0
         self._primary_tripped_at: float | None = None
@@ -141,6 +150,33 @@ class FallbackProvider(LLMProvider):
     def supports_progress_deltas(self) -> bool:
         return bool(getattr(self._primary, "supports_progress_deltas", False))
 
+    def can_resume_conversation_state(
+        self,
+        state: ProviderConversationState,
+        model: str | None = None,
+    ) -> bool:
+        return self._primary.can_resume_conversation_state(state, model)
+
+    def supports_native_compaction(self, model: str | None = None) -> bool:
+        return self._primary.supports_native_compaction(model)
+
+    def _primary_call_context(
+        self,
+        provider_context: ProviderCallContext,
+        model: str | None,
+    ) -> ProviderCallContext:
+        context_window_tokens = (
+            self._primary_context_window_tokens
+            if self._primary_context_window_tokens is not None
+            else provider_context.context_window_tokens
+        )
+        if not self._primary.supports_native_compaction(model):
+            context_window_tokens = None
+        return ProviderCallContext(
+            conversation_state=provider_context.conversation_state,
+            context_window_tokens=context_window_tokens,
+        )
+
     def _primary_available(self) -> bool:
         """Return True if the primary provider is not currently tripped."""
         if self._primary_tripped_at is None:
@@ -157,6 +193,25 @@ class FallbackProvider(LLMProvider):
             lambda p, kw: p.chat(**kw), kwargs, has_streamed=None
         )
 
+    async def chat_with_context(
+        self,
+        *,
+        provider_context: ProviderCallContext,
+        **kwargs: Any,
+    ) -> LLMResponse:
+        call_kwargs: dict[str, Any] = dict(kwargs)
+        call_kwargs["provider_context"] = self._primary_call_context(
+            provider_context,
+            kwargs.get("model"),
+        )
+        if not self._has_fallbacks:
+            return await self._primary.chat_with_context(**call_kwargs)
+        return await self._try_with_fallback(
+            lambda p, kw: p.chat_with_context(**kw),
+            call_kwargs,
+            has_streamed=None,
+        )
+
     async def chat_stream(self, **kwargs: Any) -> LLMResponse:
         on_stream_recover = kwargs.pop("on_stream_recover", None)
         if not self._has_fallbacks:
@@ -179,6 +234,38 @@ class FallbackProvider(LLMProvider):
             on_stream_recover=on_stream_recover,
         )
 
+    async def chat_stream_with_context(
+        self,
+        *,
+        provider_context: ProviderCallContext,
+        **kwargs: Any,
+    ) -> LLMResponse:
+        on_stream_recover = kwargs.pop("on_stream_recover", None)
+        call_kwargs: dict[str, Any] = dict(kwargs)
+        call_kwargs["provider_context"] = self._primary_call_context(
+            provider_context,
+            kwargs.get("model"),
+        )
+        if not self._has_fallbacks:
+            return await self._primary.chat_stream_with_context(**call_kwargs)
+
+        has_streamed: list[bool] = [False]
+        original_delta = call_kwargs.get("on_content_delta")
+
+        async def _tracking_delta(text: str) -> None:
+            if text:
+                has_streamed[0] = True
+            if original_delta:
+                await original_delta(text)
+
+        call_kwargs["on_content_delta"] = _tracking_delta
+        return await self._try_with_fallback(
+            lambda p, kw: p.chat_stream_with_context(**kw),
+            call_kwargs,
+            has_streamed=has_streamed,
+            on_stream_recover=on_stream_recover,
+        )
+
     async def _try_with_fallback(
         self,
         call: Callable[[LLMProvider, dict[str, Any]], Awaitable[LLMResponse]],
@@ -189,6 +276,9 @@ class FallbackProvider(LLMProvider):
         primary_model = kwargs.get("model") or self._primary.get_default_model()
         primary_was_attempted = False
         primary_error = "unknown error"
+        # A primary error eligible for failover did not return a replacement
+        # continuation, so the incoming primary state remains reusable.
+        preserve_primary_state = True
 
         if self._primary_available():
             primary_was_attempted = True
@@ -286,6 +376,23 @@ class FallbackProvider(LLMProvider):
                 "max_tokens": fallback.max_tokens,
                 "temperature": fallback.temperature,
             }
+            provider_context = fallback_kwargs.get("provider_context")
+            if isinstance(provider_context, ProviderCallContext):
+                state = provider_context.conversation_state
+                if state is not None and not fallback_provider.can_resume_conversation_state(
+                    state,
+                    fallback_model,
+                ):
+                    state = None
+                context_window_tokens = (
+                    fallback.context_window_tokens
+                    if fallback_provider.supports_native_compaction(fallback_model)
+                    else None
+                )
+                fallback_kwargs["provider_context"] = ProviderCallContext(
+                    conversation_state=state,
+                    context_window_tokens=context_window_tokens,
+                )
             if fallback.reasoning_effort is None:
                 fallback_kwargs.pop("reasoning_effort", None)
             else:
@@ -312,11 +419,15 @@ class FallbackProvider(LLMProvider):
         )
         # Return the last error response we saw (primary or last fallback).
         if last_response is not None:
-            return last_response
+            return replace(
+                last_response,
+                preserve_provider_state_on_error=preserve_primary_state,
+            )
         # Primary was tripped and we have no fallbacks — synthesize an error.
         return LLMResponse(
             content=f"Primary model '{primary_model}' circuit open and no fallbacks available",
             finish_reason="error",
+            preserve_provider_state_on_error=preserve_primary_state,
         )
 
     async def _notify_fallback_model(self, model: str) -> None:

nanobot/providers/github_copilot_provider.py

@@ -16,7 +16,7 @@ import httpx
 from oauth_cli_kit.models import OAuthToken
 from oauth_cli_kit.storage import FileTokenStorage
 
-from nanobot.providers.base import LLMResponse
+from nanobot.providers.base import LLMResponse, ProviderCallContext
 from nanobot.providers.openai_compat_provider import OpenAICompatProvider
 
 DEFAULT_GITHUB_DEVICE_CODE_URL = "https://github.com/login/device/code"
@@ -248,6 +248,7 @@ class GitHubCopilotProvider(OpenAICompatProvider):
         temperature: float = 0.7,
         reasoning_effort: str | None = None,
         tool_choice: str | dict[str, Any] | None = None,
+        provider_context: ProviderCallContext | None = None,
     ) -> LLMResponse:
         await self._refresh_client_api_key()
         return await super().chat(
@@ -258,6 +259,7 @@ class GitHubCopilotProvider(OpenAICompatProvider):
             temperature=temperature,
             reasoning_effort=reasoning_effort,
             tool_choice=tool_choice,
+            provider_context=provider_context,
         )
 
     async def chat_stream(
@@ -272,6 +274,7 @@ class GitHubCopilotProvider(OpenAICompatProvider):
         on_content_delta: Callable[[str], Awaitable[None]] | None = None,
         on_thinking_delta: Callable[[str], Awaitable[None]] | None = None,
         on_tool_call_delta: Callable[[dict[str, Any]], Awaitable[None]] | None = None,
+        provider_context: ProviderCallContext | None = None,
     ) -> LLMResponse:
         await self._refresh_client_api_key()
         return await super().chat_stream(
@@ -285,4 +288,5 @@ class GitHubCopilotProvider(OpenAICompatProvider):
             on_content_delta=on_content_delta,
             on_thinking_delta=on_thinking_delta,
             on_tool_call_delta=on_tool_call_delta,
+            provider_context=provider_context,
         )

nanobot/providers/openai_codex_provider.py

@@ -17,17 +17,27 @@ from oauth_cli_kit import get_token as get_codex_token
 from nanobot.providers.base import (
     LLMProvider,
     LLMResponse,
-    ToolCallRequest,
+    ProviderCallContext,
+    ProviderConversationState,
     resolve_stream_idle_timeout_s,
 )
 from nanobot.providers.openai_responses import (
+    ResponsesStreamCapture,
+    build_responses_state,
     consume_sse_with_reasoning,
-    convert_messages,
     convert_tools,
+    is_compaction_compatibility_error,
+    is_replayable_finish_reason,
+    prepare_responses_input,
+    resolve_compact_threshold,
+    responses_state_context_tokens,
+    responses_state_items,
+    responses_state_matches,
 )
 
 DEFAULT_CODEX_URL = "https://chatgpt.com/backend-api/codex/responses"
 DEFAULT_ORIGINATOR = "nanobot"
+_COMPACTION_RETAINED_CHAR_BUDGET = 256_000
 
 
 class OpenAICodexProvider(LLMProvider):
@@ -45,21 +55,39 @@ class OpenAICodexProvider(LLMProvider):
         self.default_model = default_model
         self.proxy = proxy or None
         self._extra_body = dict(extra_body or {})
+        self._native_compaction_available = True
 
     async def _call_codex(
         self,
         messages: list[dict[str, Any]],
         tools: list[dict[str, Any]] | None,
         model: str | None,
+        max_tokens: int,
         reasoning_effort: str | None,
         tool_choice: str | dict[str, Any] | None,
         on_content_delta: Callable[[str], Awaitable[None]] | None = None,
         on_thinking_delta: Callable[[str], Awaitable[None]] | None = None,
         on_tool_call_delta: Callable[[dict[str, Any]], Awaitable[None]] | None = None,
+        provider_context: ProviderCallContext | None = None,
     ) -> LLMResponse:
         """Shared request logic for both chat() and chat_stream()."""
         model = model or self.default_model
-        system_prompt, input_items = convert_messages(messages)
+        sanitized_messages = self._sanitize_empty_content(messages)
+        sanitized_state = (
+            provider_context.conversation_state
+            if provider_context is not None
+            else None
+        )
+        if sanitized_state is not None:
+            sanitized_state = sanitized_state.with_pending_messages(
+                self._sanitize_empty_content(sanitized_state.pending_messages)
+            )
+        system_prompt, input_items, replayed = prepare_responses_input(
+            sanitized_messages,
+            state=sanitized_state,
+            provider=self._responses_state_provider(),
+            model=_strip_model_prefix(model),
+        )
 
         body: dict[str, Any] = {
             "model": _strip_model_prefix(model),
@@ -68,12 +96,15 @@ class OpenAICodexProvider(LLMProvider):
             "instructions": system_prompt,
             "input": input_items,
             "text": {"verbosity": "medium"},
-            "include": ["reasoning.encrypted_content"],
             "prompt_cache_key": _prompt_cache_key(messages[:2]),
             "tool_choice": tool_choice or "auto",
             "parallel_tool_calls": True,
         }
+        body["include"] = ["reasoning.encrypted_content"]
         reasoning_options = _build_reasoning_options(reasoning_effort)
+        if replayed and "gpt-5.6" in _strip_model_prefix(model).lower():
+            reasoning_options = dict(reasoning_options or {})
+            reasoning_options["context"] = "all_turns"
         if reasoning_options:
             body["reasoning"] = reasoning_options
         if tools:
@@ -87,33 +118,90 @@ class OpenAICodexProvider(LLMProvider):
             token = await asyncio.to_thread(get_codex_token, proxy=self.proxy)
             headers = _build_headers(cast(str, token.account_id), token.access)
 
-            stage = "codex_request"
+            async def _send(
-            try:
+                request_body: dict[str, Any],
-                content, tool_calls, finish_reason, usage, reasoning_content = await _request_codex(
+                *,
-                    DEFAULT_CODEX_URL, headers, body, verify=True,
+                emit_deltas: bool,
-                    proxy=self.proxy,
+            ) -> LLMResponse:
-                    on_content_delta=on_content_delta,
+                wire_body = _without_response_item_ids(request_body)
-                    on_thinking_delta=on_thinking_delta,
+                try:
-                    on_tool_call_delta=on_tool_call_delta,
+                    return await _request_codex(
-                )
+                        DEFAULT_CODEX_URL,
-            except Exception as e:
+                        headers,
-                if "CERTIFICATE_VERIFY_FAILED" not in str(e):
+                        wire_body,
-                    raise
+                        verify=True,
-                logger.warning("SSL verification failed for Codex API; retrying with verify=False")
+                        proxy=self.proxy,
-                content, tool_calls, finish_reason, usage, reasoning_content = await _request_codex(
+                        on_content_delta=on_content_delta if emit_deltas else None,
-                    DEFAULT_CODEX_URL, headers, body, verify=False,
+                        on_thinking_delta=on_thinking_delta if emit_deltas else None,
-                    proxy=self.proxy,
+                        on_tool_call_delta=on_tool_call_delta if emit_deltas else None,
-                    on_content_delta=on_content_delta,
+                    )
-                    on_thinking_delta=on_thinking_delta,
+                except Exception as exc:
-                    on_tool_call_delta=on_tool_call_delta,
+                    if "CERTIFICATE_VERIFY_FAILED" not in str(exc):
-                )
+                        raise
-            return LLMResponse(
+                    logger.warning(
-                content=content,
+                        "SSL verification failed for Codex API; retrying with verify=False"
-                tool_calls=tool_calls,
+                    )
-                finish_reason=finish_reason,
+                    return await _request_codex(
-                usage=usage,
+                        DEFAULT_CODEX_URL,
-                reasoning_content=reasoning_content,
+                        headers,
+                        wire_body,
+                        verify=False,
+                        proxy=self.proxy,
+                        on_content_delta=on_content_delta if emit_deltas else None,
+                        on_thinking_delta=on_thinking_delta if emit_deltas else None,
+                        on_tool_call_delta=on_tool_call_delta if emit_deltas else None,
+                    )
+
+            compact_threshold = resolve_compact_threshold(
+                (
+                    provider_context.context_window_tokens
+                    if provider_context is not None
+                    else None
+                ),
+                max_tokens,
             )
+            if (
+                self.supports_native_compaction(model)
+                and replayed
+                and sanitized_state is not None
+                and compact_threshold is not None
+                and responses_state_context_tokens(sanitized_state) >= compact_threshold
+            ):
+                stage = "codex_compaction"
+                compact_body = {
+                    **body,
+                    "input": [*input_items, {"type": "compaction_trigger"}],
+                }
+                try:
+                    compact_result = await _send(compact_body, emit_deltas=False)
+                    compact_items = (
+                        responses_state_items(compact_result.provider_state)
+                        if compact_result.provider_state is not None
+                        else None
+                    )
+                    if not compact_items or compact_items[-1].get("type") not in {
+                        "compaction",
+                        "compaction_summary",
+                        "context_compaction",
+                    }:
+                        raise RuntimeError("Codex compaction returned no compaction item")
+                    body["input"] = [
+                        *_retained_compaction_messages(input_items),
+                        *compact_items,
+                    ]
+                except Exception as compact_error:
+                    if is_compaction_compatibility_error(compact_error):
+                        self._native_compaction_available = False
+                    logger.warning(
+                        "Codex native compaction unavailable; continuing without it "
+                        "(type={} status={} disabled={})",
+                        type(compact_error).__name__,
+                        getattr(compact_error, "status_code", None),
+                        not self._native_compaction_available,
+                    )
+
+            stage = "codex_request"
+            return await _send(body, emit_deltas=True)
         except Exception as e:
             response = _codex_error_response(e)
             exc_type = "CodexHTTPError" if isinstance(e, _CodexHTTPError) else type(e).__name__
@@ -137,8 +225,28 @@ class OpenAICodexProvider(LLMProvider):
         model: str | None = None, max_tokens: int = 4096, temperature: float = 0.7,
         reasoning_effort: str | None = None,
         tool_choice: str | dict[str, Any] | None = None,
+        provider_context: ProviderCallContext | None = None,
     ) -> LLMResponse:
-        return await self._call_codex(messages, tools, model, reasoning_effort, tool_choice)
+        return await self._call_codex(
+            messages,
+            tools,
+            model,
+            max_tokens,
+            reasoning_effort,
+            tool_choice,
+            provider_context=provider_context,
+        )
+
+    async def chat_with_context(
+        self,
+        *,
+        provider_context: ProviderCallContext,
+        **kwargs: Any,
+    ) -> LLMResponse:
+        return await self.chat(
+            **kwargs,
+            provider_context=provider_context,
+        )
 
     async def chat_stream(
         self, messages: list[dict[str, Any]], tools: list[dict[str, Any]] | None = None,
@@ -148,21 +256,55 @@ class OpenAICodexProvider(LLMProvider):
         on_content_delta: Callable[[str], Awaitable[None]] | None = None,
         on_thinking_delta: Callable[[str], Awaitable[None]] | None = None,
         on_tool_call_delta: Callable[[dict[str, Any]], Awaitable[None]] | None = None,
+        provider_context: ProviderCallContext | None = None,
     ) -> LLMResponse:
         return await self._call_codex(
-            messages,
+            messages=messages,
-            tools,
+            tools=tools,
-            model,
+            model=model,
-            reasoning_effort,
+            max_tokens=max_tokens,
-            tool_choice,
+            reasoning_effort=reasoning_effort,
-            on_content_delta,
+            tool_choice=tool_choice,
-            on_thinking_delta,
+            on_content_delta=on_content_delta,
-            on_tool_call_delta,
+            on_thinking_delta=on_thinking_delta,
+            on_tool_call_delta=on_tool_call_delta,
+            provider_context=provider_context,
+        )
+
+    async def chat_stream_with_context(
+        self,
+        *,
+        provider_context: ProviderCallContext,
+        **kwargs: Any,
+    ) -> LLMResponse:
+        return await self.chat_stream(
+            **kwargs,
+            provider_context=provider_context,
         )
 
     def get_default_model(self) -> str:
         return self.default_model
 
+    @staticmethod
+    def _responses_state_provider() -> str:
+        return f"openai_codex:{DEFAULT_CODEX_URL.rstrip('/')}"
+
+    def can_resume_conversation_state(
+        self,
+        state: ProviderConversationState,
+        model: str | None = None,
+    ) -> bool:
+        return responses_state_matches(
+            state,
+            provider=self._responses_state_provider(),
+            model=_strip_model_prefix(model or self.default_model),
+        )
+
+    def supports_native_compaction(self, model: str | None = None) -> bool:
+        """Use the Codex backend's inline compaction trigger when needed."""
+        _ = model
+        return self._native_compaction_available
+
 
 def _strip_model_prefix(model: str) -> str:
     if model.startswith("openai-codex/") or model.startswith("openai_codex/"):
@@ -170,6 +312,58 @@ def _strip_model_prefix(model: str) -> str:
     return model
 
 
+def _without_response_item_ids(
+    request_body: dict[str, Any],
+) -> dict[str, Any]:
+    """Match Codex's default ``store=false`` request-item contract."""
+    if request_body.get("store") is True:
+        return request_body
+    raw_input = request_body.get("input")
+    if not isinstance(raw_input, list):
+        return request_body
+
+    input_items: list[object] = cast(list[object], raw_input)
+    sanitized_input: list[object] = []
+    for raw_item in input_items:
+        if not isinstance(raw_item, dict):
+            sanitized_input.append(raw_item)
+            continue
+        item = cast(dict[str, Any], raw_item)
+        sanitized_input.append({
+            key: value
+            for key, value in item.items()
+            if key != "id"
+        })
+
+    body = dict(request_body)
+    body["input"] = sanitized_input
+    return body
+
+
+def _retained_compaction_messages(
+    input_items: list[dict[str, Any]],
+) -> list[dict[str, Any]]:
+    """Mirror Codex's bounded retention of user/developer/system messages."""
+    retained_reversed: list[dict[str, Any]] = []
+    remaining = _COMPACTION_RETAINED_CHAR_BUDGET
+    for item in reversed(input_items):
+        if item.get("type") not in {None, "message"} or item.get("role") not in {
+            "user",
+            "developer",
+            "system",
+        }:
+            continue
+        size = len(json.dumps(item, ensure_ascii=False))
+        if size > remaining and retained_reversed:
+            continue
+        retained_reversed.append(item)
+        remaining = max(0, remaining - size)
+        if remaining == 0:
+            break
+    retained_reversed.reverse()
+    return retained_reversed
+
+
 def _build_reasoning_options(reasoning_effort: str | None) -> dict[str, str] | None:
     """Opt in to visible summaries without changing provider-default effort."""
     if reasoning_effort and reasoning_effort.lower() == "none":
@@ -202,6 +396,7 @@ class _CodexHTTPError(RuntimeError):
         error_type: str | None = None,
         error_code: str | None = None,
         should_retry: bool | None = None,
+        compaction_unsupported: bool = False,
     ):
         super().__init__(message)
         self.status_code = status_code
@@ -209,6 +404,7 @@ class _CodexHTTPError(RuntimeError):
         self.error_type = error_type
         self.error_code = error_code
         self.should_retry = should_retry
+        self.compaction_unsupported = compaction_unsupported
 
 
 async def _request_codex(
@@ -220,7 +416,7 @@ async def _request_codex(
     on_content_delta: Callable[[str], Awaitable[None]] | None = None,
     on_thinking_delta: Callable[[str], Awaitable[None]] | None = None,
     on_tool_call_delta: Callable[[dict[str, Any]], Awaitable[None]] | None = None,
-) -> tuple[str, list[ToolCallRequest], str, dict[str, int], str | None]:
+) -> LLMResponse:
     idle_timeout_s = resolve_stream_idle_timeout_s()
     client_kwargs: dict[str, Any] = {"timeout": idle_timeout_s, "verify": verify}
     if proxy:
@@ -233,6 +429,17 @@ async def _request_codex(
                 raw = text.decode("utf-8", "ignore")
                 retry_after = LLMProvider._extract_retry_after_from_headers(response.headers)
                 error_type, error_code = LLMProvider._extract_error_type_code(raw)
+                compaction_unsupported = (
+                    response.status_code in {400, 404, 422}
+                    and any(
+                        marker in raw.lower()
+                        for marker in (
+                            "context_management",
+                            "compact_threshold",
+                            "compaction_trigger",
+                        )
+                    )
+                )
                 raise _CodexHTTPError(
                     _friendly_error(response.status_code, raw),
                     status_code=response.status_code,
@@ -240,13 +447,38 @@ async def _request_codex(
                     error_type=error_type,
                     error_code=error_code,
                     should_retry=_should_retry_status(response.status_code, error_type, error_code, raw),
+                    compaction_unsupported=compaction_unsupported,
                 )
-            return await consume_sse_with_reasoning(
+            capture = ResponsesStreamCapture()
+            (
+                content,
+                tool_calls,
+                finish_reason,
+                usage,
+                reasoning_content,
+            ) = await consume_sse_with_reasoning(
                 response,
                 on_content_delta=on_content_delta,
                 on_tool_call_delta=on_tool_call_delta,
                 on_reasoning_delta=on_thinking_delta,
+                capture=capture,
             )
+            result = LLMResponse(
+                content=content,
+                tool_calls=tool_calls,
+                finish_reason=finish_reason,
+                usage=usage,
+                reasoning_content=reasoning_content,
+            )
+            if capture.completed and is_replayable_finish_reason(finish_reason):
+                result.provider_state = build_responses_state(
+                    provider=f"openai_codex:{url.rstrip('/')}",
+                    model=str(body.get("model") or ""),
+                    input_items=cast(list[dict[str, Any]], body.get("input") or []),
+                    output_items=capture.output_items,
+                    usage=usage,
+                )
+            return result
 
 
 def _prompt_cache_key(messages: list[dict[str, Any]]) -> str:

nanobot/providers/openai_compat_provider.py

@@ -26,16 +26,24 @@ from pydantic.alias_generators import to_snake
 from nanobot.providers.base import (
     LLMProvider,
     LLMResponse,
+    ProviderCallContext,
+    ProviderConversationState,
     ToolCallRequest,
     parse_tool_arguments,
     resolve_stream_idle_timeout_s,
     tool_arguments_json_for_replay,
 )
 from nanobot.providers.openai_responses import (
+    ResponsesStreamCapture,
+    build_responses_state,
     consume_sdk_stream,
-    convert_messages,
     convert_tools,
+    is_compaction_compatibility_error,
+    is_replayable_finish_reason,
     parse_response_output,
+    prepare_responses_input,
+    resolve_compact_threshold,
+    responses_state_matches,
 )
 
 if TYPE_CHECKING:
@@ -443,6 +451,8 @@ class OpenAICompatProvider(LLMProvider):
     registry lookups needed.
     """
 
+    _native_compaction_available = True
+
     def __init__(
         self,
         api_key: str | None = None,
@@ -463,6 +473,7 @@ class OpenAICompatProvider(LLMProvider):
         self._api_type = api_type if spec and spec.name == "openai" else "auto"
         self._extra_query = extra_query or {}
         self._proxy = proxy or None
+        self._native_compaction_available = True
 
         if api_key and spec and spec.env_key:
             self._setup_env(api_key, api_base)
@@ -971,6 +982,37 @@ class OpenAICompatProvider(LLMProvider):
 
         return self._responses_circuit_allows_probe(model, reasoning_effort)
 
+    def _responses_state_provider(self) -> str:
+        spec_name = self._spec.name if self._spec is not None else "custom"
+        effective_base = self._effective_base or "https://api.openai.com/v1"
+        return f"openai_compat:{spec_name}:{effective_base.rstrip('/')}"
+
+    def _responses_state_model(self, model: str | None) -> str:
+        return self._request_model_name(model or self.default_model)
+
+    def can_resume_conversation_state(
+        self,
+        state: ProviderConversationState,
+        model: str | None = None,
+    ) -> bool:
+        return responses_state_matches(
+            state,
+            provider=self._responses_state_provider(),
+            model=self._responses_state_model(model),
+        )
+
+    def supports_native_compaction(self, model: str | None = None) -> bool:
+        """Enable server compaction only on direct OpenAI Responses endpoints."""
+        _ = model
+        if (
+            not self._native_compaction_available
+            or self._api_type == "chat_completions"
+        ):
+            return False
+        if self._spec is not None and self._spec.name != "openai":
+            return False
+        return _is_direct_openai_base(self._effective_base)
+
     def _responses_circuit_allows_probe(
         self,
         model: str | None,
@@ -1040,12 +1082,29 @@ class OpenAICompatProvider(LLMProvider):
         temperature: float,
         reasoning_effort: str | None,
         tool_choice: str | dict[str, Any] | None,
+        provider_context: ProviderCallContext | None = None,
     ) -> dict[str, Any]:
         """Build a Responses API body for direct OpenAI requests."""
         model_name = model or self.default_model
         model_name = self._request_model_name(model_name)
         sanitized_messages = self._sanitize_messages(self._sanitize_empty_content(messages))
-        instructions, input_items = convert_messages(sanitized_messages)
+        sanitized_state = (
+            provider_context.conversation_state
+            if provider_context is not None
+            else None
+        )
+        if sanitized_state is not None:
+            sanitized_state = sanitized_state.with_pending_messages(
+                self._sanitize_messages(
+                    self._sanitize_empty_content(sanitized_state.pending_messages)
+                )
+            )
+        instructions, input_items, replayed = prepare_responses_input(
+            sanitized_messages,
+            state=sanitized_state,
+            provider=self._responses_state_provider(),
+            model=model_name,
+        )
 
         body: dict[str, Any] = {
             "model": model_name,
@@ -1055,13 +1114,29 @@ class OpenAICompatProvider(LLMProvider):
             "store": False,
             "stream": False,
         }
+        compact_threshold = resolve_compact_threshold(
+            (
+                provider_context.context_window_tokens
+                if provider_context is not None
+                else None
+            ),
+            max_tokens,
+        )
+        if self.supports_native_compaction(model_name) and compact_threshold is not None:
+            body["context_management"] = [{
+                "type": "compaction",
+                "compact_threshold": compact_threshold,
+            }]
 
         if self._supports_temperature(model_name, reasoning_effort):
             body["temperature"] = temperature
 
+        if not self._supports_temperature(model_name, reasoning_effort):
+            body["include"] = ["reasoning.encrypted_content"]
         if reasoning_effort and reasoning_effort.lower() != "none":
             body["reasoning"] = {"effort": reasoning_effort}
-            body["include"] = ["reasoning.encrypted_content"]
+        if replayed and "gpt-5.6" in model_name.lower():
+            body.setdefault("reasoning", {})["context"] = "all_turns"
 
         if tools:
             body["tools"] = convert_tools(tools)
@@ -1073,6 +1148,29 @@ class OpenAICompatProvider(LLMProvider):
 
         return body
 
+    async def _create_response_with_compaction_fallback(
+        self,
+        client: Any,
+        body: dict[str, Any],
+    ) -> Any:
+        """Retry Responses once without server compaction on compatibility errors."""
+        try:
+            return await client.responses.create(**body)
+        except Exception as exc:
+            if (
+                "context_management" not in body
+                or not is_compaction_compatibility_error(exc)
+            ):
+                raise
+            self._native_compaction_available = False
+            body.pop("context_management", None)
+            logger.warning(
+                "Responses server compaction unsupported; disabled for this provider instance "
+                "(status={})",
+                getattr(exc, "status_code", None),
+            )
+            return await client.responses.create(**body)
+
     # ------------------------------------------------------------------
     # Response parsing
     # ------------------------------------------------------------------
@@ -1599,6 +1697,28 @@ class OpenAICompatProvider(LLMProvider):
     # Public API
     # ------------------------------------------------------------------
 
+    async def chat_with_context(
+        self,
+        *,
+        provider_context: ProviderCallContext,
+        **kwargs: Any,
+    ) -> LLMResponse:
+        return await self.chat(
+            **kwargs,
+            provider_context=provider_context,
+        )
+
+    async def chat_stream_with_context(
+        self,
+        *,
+        provider_context: ProviderCallContext,
+        **kwargs: Any,
+    ) -> LLMResponse:
+        return await self.chat_stream(
+            **kwargs,
+            provider_context=provider_context,
+        )
+
     async def chat(
         self,
         messages: list[dict[str, Any]],
@@ -1608,6 +1728,7 @@ class OpenAICompatProvider(LLMProvider):
         temperature: float = 0.7,
         reasoning_effort: str | None = None,
         tool_choice: str | dict[str, Any] | None = None,
+        provider_context: ProviderCallContext | None = None,
     ) -> LLMResponse:
         client = await self._ensure_client()
         try:
@@ -1616,12 +1737,18 @@ class OpenAICompatProvider(LLMProvider):
                     body = self._build_responses_body(
                         messages, tools, model, max_tokens, temperature,
                         reasoning_effort, tool_choice,
+                        provider_context,
                     )
-                    responses_raw = cast(
+                    responses_raw = await self._create_response_with_compaction_fallback(
-                        Any,
+                        client,
-                        await client.responses.create(**body),
+                        body,
+                    )
+                    result = parse_response_output(
+                        responses_raw,
+                        state_provider=self._responses_state_provider(),
+                        state_model=str(body["model"]),
+                        state_input_items=cast(list[dict[str, Any]], body["input"]),
                     )
-                    result = parse_response_output(responses_raw)
                     self._record_responses_success(model, reasoning_effort)
                     return result
                 except Exception as responses_error:
@@ -1660,6 +1787,7 @@ class OpenAICompatProvider(LLMProvider):
         on_content_delta: Callable[[str], Awaitable[None]] | None = None,
         on_thinking_delta: Callable[[str], Awaitable[None]] | None = None,
         on_tool_call_delta: Callable[[dict[str, Any]], Awaitable[None]] | None = None,
+        provider_context: ProviderCallContext | None = None,
     ) -> LLMResponse:
         client = await self._ensure_client()
         idle_timeout_s = resolve_stream_idle_timeout_s()
@@ -1669,11 +1797,12 @@ class OpenAICompatProvider(LLMProvider):
                     body = self._build_responses_body(
                         messages, tools, model, max_tokens, temperature,
                         reasoning_effort, tool_choice,
+                        provider_context,
                     )
                     body["stream"] = True
-                    responses_stream = cast(
+                    responses_stream = await self._create_response_with_compaction_fallback(
-                        Any,
+                        client,
-                        await client.responses.create(**body),
+                        body,
                     )
 
                     async def _timed_stream() -> AsyncIterator[Any]:
@@ -1687,6 +1816,7 @@ class OpenAICompatProvider(LLMProvider):
                             except StopAsyncIteration:
                                 break
 
+                    capture = ResponsesStreamCapture()
                     (
                         content,
                         tool_calls,
@@ -1697,15 +1827,25 @@ class OpenAICompatProvider(LLMProvider):
                         _timed_stream(),
                         on_content_delta,
                         on_tool_call_delta=on_tool_call_delta,
+                        capture=capture,
                     )
                     self._record_responses_success(model, reasoning_effort)
-                    return LLMResponse(
+                    result = LLMResponse(
                         content=content or None,
                         tool_calls=tool_calls,
                         finish_reason=finish_reason,
                         usage=usage,
                         reasoning_content=reasoning_content,
                     )
+                    if capture.completed and is_replayable_finish_reason(finish_reason):
+                        result.provider_state = build_responses_state(
+                            provider=self._responses_state_provider(),
+                            model=str(body["model"]),
+                            input_items=cast(list[dict[str, Any]], body["input"]),
+                            output_items=capture.output_items,
+                            usage=usage,
+                        )
+                    return result
                 except Exception as responses_error:
                     if self._spec and self._spec.name == "github_copilot":
                         # Copilot gateway exposes GPT-5/o-series only via /responses;

nanobot/providers/openai_responses/__init__.py

@@ -1,4 +1,4 @@
-"""Shared helpers for OpenAI Responses API providers (Codex, Azure OpenAI)."""
+"""Shared helpers for provider backends that implement the OpenAI Responses protocol."""
 
 from nanobot.providers.openai_responses.converters import (
     convert_messages,
@@ -8,13 +8,24 @@ from nanobot.providers.openai_responses.converters import (
 )
 from nanobot.providers.openai_responses.parsing import (
     FINISH_REASON_MAP,
+    ResponsesStreamCapture,
     consume_sdk_stream,
     consume_sse,
     consume_sse_with_reasoning,
+    is_replayable_finish_reason,
     iter_sse,
     map_finish_reason,
     parse_response_output,
 )
+from nanobot.providers.openai_responses.state import (
+    build_responses_state,
+    is_compaction_compatibility_error,
+    prepare_responses_input,
+    resolve_compact_threshold,
+    responses_state_context_tokens,
+    responses_state_items,
+    responses_state_matches,
+)
 
 __all__ = [
     "convert_messages",
@@ -25,7 +36,16 @@ __all__ = [
     "consume_sse",
     "consume_sse_with_reasoning",
     "consume_sdk_stream",
+    "ResponsesStreamCapture",
+    "is_replayable_finish_reason",
     "map_finish_reason",
     "parse_response_output",
+    "build_responses_state",
+    "is_compaction_compatibility_error",
+    "prepare_responses_input",
+    "resolve_compact_threshold",
+    "responses_state_context_tokens",
+    "responses_state_items",
+    "responses_state_matches",
     "FINISH_REASON_MAP",
 ]

nanobot/providers/openai_responses/parsing.py

@@ -4,12 +4,14 @@ from __future__ import annotations
 
 import json
 from collections.abc import Awaitable, Callable
+from dataclasses import dataclass, field
 from typing import Any, AsyncGenerator, cast
 
 import httpx
 from loguru import logger
 
 from nanobot.providers.base import LLMResponse, ToolCallRequest, parse_tool_arguments
+from nanobot.providers.openai_responses.state import build_responses_state
 
 FINISH_REASON_MAP = {
     "completed": "stop",
@@ -17,6 +19,42 @@ FINISH_REASON_MAP = {
     "failed": "error",
     "cancelled": "error",
 }
+REPLAYABLE_FINISH_REASONS = frozenset({"stop", "tool_calls", "function_call"})
+
+
+@dataclass(slots=True)
+class ResponsesStreamCapture:
+    """Losslessly capture terminal output items without changing stream results."""
+
+    completed: bool = False
+    response: dict[str, Any] | None = field(default=None, repr=False)
+    _items_by_index: dict[int, dict[str, Any]] = field(default_factory=dict, repr=False)
+
+    def record_output_item(self, index: object, item: object) -> None:
+        item_object = _response_object(item)
+        if item_object is None:
+            return
+        output_index = (
+            index
+            if isinstance(index, int) and not isinstance(index, bool)
+            else len(self._items_by_index)
+        )
+        self._items_by_index[output_index] = item_object
+
+    def record_completed(self, response: object) -> None:
+        response_object = _response_object(response)
+        if response_object is None:
+            return
+        self.completed = True
+        self.response = response_object
+
+    @property
+    def output_items(self) -> list[dict[str, Any]]:
+        if self.response is not None:
+            output = _response_object_list(self.response.get("output"))
+            if output:
+                return output
+        return [self._items_by_index[index] for index in sorted(self._items_by_index)]
 
 
 def _as_json_object(value: object) -> dict[str, Any] | None:
@@ -54,6 +92,27 @@ def map_finish_reason(status: str | None) -> str:
     return FINISH_REASON_MAP.get(status or "completed", "stop")
 
 
+def is_replayable_finish_reason(finish_reason: str) -> bool:
+    """Return whether a response can safely advance opaque conversation state."""
+    return finish_reason in REPLAYABLE_FINISH_REASONS
+
+
+def _response_finish_reason(
+    response: object,
+    *,
+    fallback_status: str | None = None,
+) -> str:
+    """Map terminal response details without treating content filtering as truncation."""
+    response_object = _response_object(response) or {}
+    status = response_object.get("status")
+    terminal_status = status if isinstance(status, str) else fallback_status
+    if terminal_status == "incomplete":
+        details = _response_object(response_object.get("incomplete_details"))
+        if details is not None and details.get("reason") == "content_filter":
+            return "content_filter"
+    return map_finish_reason(terminal_status)
+
+
 def _usage_from_response_obj(response: object) -> dict[str, int]:
     response_object = _response_object(response)
     usage_raw: object = (
@@ -99,6 +158,47 @@ def _tool_arguments_source(*values: Any) -> Any:
     return "{}"
 
 
+def _refusal_event_key(
+    item_id: object,
+    content_index: object,
+) -> tuple[str | None, int | None]:
+    """Identify one streamed refusal content part across delta/done events."""
+    return (
+        item_id if isinstance(item_id, str) else None,
+        (
+            content_index
+            if isinstance(content_index, int) and not isinstance(content_index, bool)
+            else None
+        ),
+    )
+
+
+def _remaining_refusal_text(streamed_text: str, refusal_text: str) -> str:
+    """Return only text not already surfaced by refusal deltas."""
+    if not streamed_text:
+        return refusal_text
+    if refusal_text.startswith(streamed_text):
+        return refusal_text[len(streamed_text):]
+    return ""
+
+
+def _extract_refusal_text_from_output(output: object) -> tuple[bool, str]:
+    """Extract refusal content from terminal Responses output items."""
+    refusal_seen = False
+    parts: list[str] = []
+    for item in _response_object_list(output):
+        if item.get("type") != "message":
+            continue
+        for block in _response_object_list(item.get("content")):
+            if block.get("type") != "refusal":
+                continue
+            refusal_seen = True
+            refusal_text = block.get("refusal")
+            if isinstance(refusal_text, str):
+                parts.append(refusal_text)
+    return refusal_seen, "".join(parts)
+
+
 async def iter_sse(response: httpx.Response) -> AsyncGenerator[dict[str, Any], None]:
     """Yield parsed JSON events from a Responses API SSE stream."""
     buffer: list[str] = []
@@ -153,6 +253,7 @@ async def consume_sse_with_reasoning(
     on_tool_call_delta: Callable[[dict[str, Any]], Awaitable[None]] | None = None,
     on_reasoning_delta: Callable[[str], Awaitable[None]] | None = None,
     on_response_event: Callable[[dict[str, Any]], Awaitable[None]] | None = None,
+    capture: ResponsesStreamCapture | None = None,
 ) -> tuple[str, list[ToolCallRequest], str, dict[str, int], str | None]:
     """Consume a Responses API SSE stream, including visible reasoning summaries."""
     content = ""
@@ -163,6 +264,9 @@ async def consume_sse_with_reasoning(
     usage: dict[str, int] = {}
     reasoning_content: str | None = None
     streamed_reasoning = False
+    refusal_seen = False
+    refusal_deltas: dict[tuple[str | None, int | None], str] = {}
+    emitted_refusal_text = ""
 
     async for event in iter_sse(response):
         if on_response_event:
@@ -191,6 +295,33 @@ async def consume_sse_with_reasoning(
             content += delta_text
             if on_content_delta and delta_text:
                 await on_content_delta(delta_text)
+        elif event_type == "response.refusal.delta":
+            refusal_seen = True
+            delta_text = event.get("delta")
+            if isinstance(delta_text, str) and delta_text:
+                key = _refusal_event_key(
+                    event.get("item_id"),
+                    event.get("content_index"),
+                )
+                refusal_deltas[key] = refusal_deltas.get(key, "") + delta_text
+                content += delta_text
+                emitted_refusal_text += delta_text
+                if on_content_delta:
+                    await on_content_delta(delta_text)
+        elif event_type == "response.refusal.done":
+            refusal_seen = True
+            refusal_text = event.get("refusal")
+            key = _refusal_event_key(
+                event.get("item_id"),
+                event.get("content_index"),
+            )
+            streamed_text = refusal_deltas.pop(key, "")
+            if isinstance(refusal_text, str) and refusal_text:
+                remaining_text = _remaining_refusal_text(streamed_text, refusal_text)
+                content += remaining_text
+                emitted_refusal_text += remaining_text
+                if on_content_delta and remaining_text:
+                    await on_content_delta(remaining_text)
         elif event_type == "response.reasoning_summary_text.delta":
             delta_text = event.get("delta") or ""
             if delta_text:
@@ -239,6 +370,8 @@ async def consume_sse_with_reasoning(
                     })
         elif event_type == "response.output_item.done":
             item = _as_json_object(event.get("item")) or {}
+            if capture is not None:
+                capture.record_output_item(event.get("output_index"), item)
             if item.get("type") == "function_call":
                 call_id = item.get("call_id")
                 if not call_id:
@@ -269,11 +402,28 @@ async def consume_sse_with_reasoning(
                     reasoning_content = summary
                     if on_reasoning_delta:
                         await on_reasoning_delta(summary)
-        elif event_type == "response.completed":
+        elif event_type in {"response.completed", "response.incomplete"}:
             response_obj = _response_object(event.get("response")) or {}
-            status = response_obj.get("status")
+            if capture is not None:
-            finish_reason = map_finish_reason(status)
+                capture.record_completed(response_obj)
+            finish_reason = _response_finish_reason(
+                response_obj,
+                fallback_status=event_type.removeprefix("response."),
+            )
             usage = _usage_from_response_obj(response_obj) or usage
+            terminal_refusal, terminal_refusal_text = _extract_refusal_text_from_output(
+                response_obj.get("output")
+            )
+            if terminal_refusal:
+                refusal_seen = True
+                remaining_text = _remaining_refusal_text(
+                    emitted_refusal_text,
+                    terminal_refusal_text,
+                )
+                content += remaining_text
+                emitted_refusal_text += remaining_text
+                if on_content_delta and remaining_text:
+                    await on_content_delta(remaining_text)
             if not reasoning_content:
                 summary = _extract_reasoning_summary_from_output(response_obj.get("output"))
                 if summary:
@@ -284,6 +434,8 @@ async def consume_sse_with_reasoning(
             detail = event.get("error") or event.get("message") or event
             raise RuntimeError(f"Response failed: {str(detail)[:500]}")
 
+    if refusal_seen:
+        finish_reason = "refusal"
     return content, tool_calls, finish_reason, usage, reasoning_content
 
 
@@ -300,7 +452,13 @@ def _extract_reasoning_summary_from_output(output: object) -> str | None:
     return "".join(parts) or None
 
 
-def parse_response_output(response: object) -> LLMResponse:
+def parse_response_output(
+    response: object,
+    *,
+    state_provider: str | None = None,
+    state_model: str | None = None,
+    state_input_items: list[dict[str, Any]] | None = None,
+) -> LLMResponse:
     """Parse an SDK ``Response`` object into an ``LLMResponse``."""
     response_object = _response_object(response) or {}
 
@@ -308,15 +466,22 @@ def parse_response_output(response: object) -> LLMResponse:
     content_parts: list[str] = []
     tool_calls: list[ToolCallRequest] = []
     reasoning_content: str | None = None
+    refusal_seen = False
 
     for item in output:
         item_type = item.get("type")
         if item_type == "message":
             for block in _response_object_list(item.get("content")):
-                if block.get("type") == "output_text":
+                block_type = block.get("type")
+                if block_type == "output_text":
                     text = block.get("text")
                     if isinstance(text, str):
                         content_parts.append(text)
+                elif block_type == "refusal":
+                    refusal_seen = True
+                    refusal = block.get("refusal")
+                    if isinstance(refusal, str):
+                        content_parts.append(refusal)
         elif item_type == "reasoning":
             for s in _response_object_list(item.get("summary")):
                 if s.get("type") == "summary_text" and s.get("text"):
@@ -337,21 +502,37 @@ def parse_response_output(response: object) -> LLMResponse:
     usage = _usage_from_response_obj(response_object)
 
     status = response_object.get("status")
-    finish_reason = map_finish_reason(status if isinstance(status, str) else None)
+    finish_reason = "refusal" if refusal_seen else _response_finish_reason(response_object)
 
-    return LLMResponse(
+    result = LLMResponse(
         content="".join(content_parts) or None,
         tool_calls=tool_calls,
         finish_reason=finish_reason,
         usage=usage,
         reasoning_content=reasoning_content if isinstance(reasoning_content, str) else None,
     )
+    if (
+        state_provider is not None
+        and state_model is not None
+        and state_input_items is not None
+        and (status is None or status == "completed")
+        and is_replayable_finish_reason(finish_reason)
+    ):
+        result.provider_state = build_responses_state(
+            provider=state_provider,
+            model=state_model,
+            input_items=state_input_items,
+            output_items=output,
+            usage=usage,
+        )
+    return result
 
 
 async def consume_sdk_stream(
     stream: Any,
     on_content_delta: Callable[[str], Awaitable[None]] | None = None,
     on_tool_call_delta: Callable[[dict[str, Any]], Awaitable[None]] | None = None,
+    capture: ResponsesStreamCapture | None = None,
 ) -> tuple[str, list[ToolCallRequest], str, dict[str, int], str | None]:
     """Consume an SDK async stream from ``client.responses.create(stream=True)``."""
     content = ""
@@ -361,6 +542,9 @@ async def consume_sdk_stream(
     finish_reason = "stop"
     usage: dict[str, int] = {}
     reasoning_content: str | None = None
+    refusal_seen = False
+    refusal_deltas: dict[tuple[str | None, int | None], str] = {}
+    emitted_refusal_text = ""
 
     async for raw_event in stream:
         event: Any = raw_event
@@ -388,6 +572,33 @@ async def consume_sdk_stream(
             content += delta_text
             if on_content_delta and delta_text:
                 await on_content_delta(delta_text)
+        elif event_type == "response.refusal.delta":
+            refusal_seen = True
+            delta_text = getattr(event, "delta", None)
+            if isinstance(delta_text, str) and delta_text:
+                key = _refusal_event_key(
+                    getattr(event, "item_id", None),
+                    getattr(event, "content_index", None),
+                )
+                refusal_deltas[key] = refusal_deltas.get(key, "") + delta_text
+                content += delta_text
+                emitted_refusal_text += delta_text
+                if on_content_delta:
+                    await on_content_delta(delta_text)
+        elif event_type == "response.refusal.done":
+            refusal_seen = True
+            refusal_text = getattr(event, "refusal", None)
+            key = _refusal_event_key(
+                getattr(event, "item_id", None),
+                getattr(event, "content_index", None),
+            )
+            streamed_text = refusal_deltas.pop(key, "")
+            if isinstance(refusal_text, str) and refusal_text:
+                remaining_text = _remaining_refusal_text(streamed_text, refusal_text)
+                content += remaining_text
+                emitted_refusal_text += remaining_text
+                if on_content_delta and remaining_text:
+                    await on_content_delta(remaining_text)
         elif event_type == "response.function_call_arguments.delta":
             call_id = getattr(event, "call_id", None)
             if call_id and call_id in tool_call_buffers:
@@ -416,6 +627,8 @@ async def consume_sdk_stream(
                     })
         elif event_type == "response.output_item.done":
             item = getattr(event, "item", None)
+            if capture is not None:
+                capture.record_output_item(getattr(event, "output_index", None), item)
             if item and getattr(item, "type", None) == "function_call":
                 call_id = getattr(item, "call_id", None)
                 if not call_id:
@@ -443,10 +656,31 @@ async def consume_sdk_stream(
                         arguments=args,
                     )
                 )
-        elif event_type == "response.completed":
+        elif event_type in {"response.completed", "response.incomplete"}:
             resp = getattr(event, "response", None)
-            status = getattr(resp, "status", None) if resp else None
+            response_obj = _response_object(resp) or {}
-            finish_reason = map_finish_reason(status)
+            if capture is not None:
+                capture.record_completed(resp)
+            finish_reason = _response_finish_reason(
+                resp,
+                fallback_status=event_type.removeprefix("response."),
+            )
+            terminal_output = response_obj.get("output")
+            if terminal_output is None:
+                terminal_output = getattr(resp, "output", None)
+            terminal_refusal, terminal_refusal_text = _extract_refusal_text_from_output(
+                terminal_output
+            )
+            if terminal_refusal:
+                refusal_seen = True
+                remaining_text = _remaining_refusal_text(
+                    emitted_refusal_text,
+                    terminal_refusal_text,
+                )
+                content += remaining_text
+                emitted_refusal_text += remaining_text
+                if on_content_delta and remaining_text:
+                    await on_content_delta(remaining_text)
             if resp:
                 usage_obj = getattr(resp, "usage", None)
                 if usage_obj:
@@ -466,4 +700,6 @@ async def consume_sdk_stream(
             detail = getattr(event, "error", None) or getattr(event, "message", None) or event
             raise RuntimeError(f"Response failed: {str(detail)[:500]}")
 
+    if refusal_seen:
+        finish_reason = "refusal"
     return content, tool_calls, finish_reason, usage, reasoning_content

nanobot/providers/openai_responses/state.py

@@ -0,0 +1,197 @@
+"""Opaque conversation state for Responses API item replay."""
+
+from __future__ import annotations
+
+from copy import deepcopy
+from typing import Any, cast
+
+from loguru import logger
+
+from nanobot.providers.base import ProviderConversationState
+from nanobot.providers.openai_responses.converters import convert_messages
+
+RESPONSES_STATE_KIND = "openai_responses"
+RESPONSES_STATE_VERSION = 1
+_ITEMS_KEY = "items"
+_CONTEXT_TOKENS_KEY = "context_tokens"
+_COMPACTION_ITEM_TYPES = frozenset({
+    "compaction",
+    "compaction_summary",
+    "context_compaction",
+})
+
+
+def responses_state_matches(
+    state: ProviderConversationState,
+    *,
+    provider: str,
+    model: str,
+) -> bool:
+    """Return whether *state* belongs to this exact Responses endpoint/model."""
+    return (
+        state.kind == RESPONSES_STATE_KIND
+        and state.version == RESPONSES_STATE_VERSION
+        and state.provider == provider
+        and state.model == model
+        and _state_items(state) is not None
+    )
+
+
+def prepare_responses_input(
+    messages: list[dict[str, Any]],
+    *,
+    state: ProviderConversationState | None,
+    provider: str,
+    model: str,
+) -> tuple[str, list[dict[str, Any]], bool]:
+    """Build a request from exact prior items plus only newly appended messages.
+
+    The full Chat transcript remains the source for the current instructions.
+    When no compatible state exists, it is converted normally as a safe
+    fallback.
+    """
+    instructions, fallback_items = convert_messages(messages)
+    if state is None or not responses_state_matches(
+        state,
+        provider=provider,
+        model=model,
+    ):
+        return instructions, fallback_items, False
+
+    prior_items = _state_items(state)
+    if prior_items is None:
+        return instructions, fallback_items, False
+
+    _, delta_items = convert_messages(state.pending_messages)
+    logger.debug(
+        "Replaying Responses state: prior_items={} pending_messages={}",
+        len(prior_items),
+        len(state.pending_messages),
+    )
+    return instructions, [*deepcopy(prior_items), *delta_items], True
+
+
+def build_responses_state(
+    *,
+    provider: str,
+    model: str,
+    input_items: list[dict[str, Any]],
+    output_items: list[dict[str, Any]],
+    usage: dict[str, int] | None = None,
+) -> ProviderConversationState:
+    """Create the canonical next state from request input and every output item."""
+    unpruned_items = [*input_items, *output_items]
+    items = _prune_before_latest_output_compaction(input_items, output_items)
+    if len(items) < len(unpruned_items):
+        logger.info(
+            "Installed Responses compaction: dropped_items={} retained_items={}",
+            len(unpruned_items) - len(items),
+            len(items),
+        )
+    payload: dict[str, Any] = {_ITEMS_KEY: deepcopy(items)}
+    context_tokens = _context_tokens_from_usage(usage)
+    if context_tokens > 0:
+        payload[_CONTEXT_TOKENS_KEY] = context_tokens
+    return ProviderConversationState(
+        kind=RESPONSES_STATE_KIND,
+        provider=provider,
+        model=model,
+        version=RESPONSES_STATE_VERSION,
+        payload=payload,
+    )
+
+
+def responses_state_items(
+    state: ProviderConversationState,
+) -> list[dict[str, Any]] | None:
+    """Return an isolated copy of canonical input items for tests/consumers."""
+    items = _state_items(state)
+    return deepcopy(items) if items is not None else None
+
+
+def responses_state_context_tokens(state: ProviderConversationState) -> int:
+    """Return the last server-reported active context size."""
+    value = state.payload.get(_CONTEXT_TOKENS_KEY)
+    if isinstance(value, bool) or not isinstance(value, int):
+        return 0
+    return max(0, value)
+
+
+def resolve_compact_threshold(
+    context_window_tokens: int | None,
+    max_output_tokens: int,
+) -> int | None:
+    """Derive Codex-compatible 90% compaction headroom for a model window."""
+    if context_window_tokens is None or context_window_tokens <= 0:
+        return None
+    ninety_percent = max(1, context_window_tokens * 9 // 10)
+    output_headroom = max(1, context_window_tokens - max(1, max_output_tokens))
+    return min(ninety_percent, output_headroom)
+
+
+def is_compaction_compatibility_error(exc: Exception) -> bool:
+    """Recognize endpoints that reject native Responses compaction fields."""
+    if getattr(exc, "compaction_unsupported", False) is True:
+        return True
+    response = getattr(exc, "response", None)
+    status_code = getattr(exc, "status_code", None)
+    if status_code is None and response is not None:
+        status_code = getattr(response, "status_code", None)
+    body = (
+        getattr(exc, "body", None)
+        or getattr(exc, "doc", None)
+        or getattr(response, "text", None)
+        or str(exc)
+    )
+    text = str(body).lower()
+    has_compaction_marker = any(
+        marker in text
+        for marker in ("context_management", "compact_threshold", "compaction_trigger")
+    )
+    if not has_compaction_marker:
+        return False
+    return isinstance(exc, TypeError) or status_code in {400, 404, 422}
+
+
+def _prune_before_latest_output_compaction(
+    input_items: list[dict[str, Any]],
+    output_items: list[dict[str, Any]],
+) -> list[dict[str, Any]]:
+    """Drop old input only when this response emits a new compaction item.
+
+    A canonical compacted input may intentionally retain messages before its
+    compaction item. Those messages must survive ordinary subsequent responses.
+    """
+    latest = None
+    for index, item in enumerate(output_items):
+        if item.get("type") in _COMPACTION_ITEM_TYPES:
+            latest = index
+    if latest is None:
+        return [*input_items, *output_items]
+    return output_items[latest:]
+
+
+def _context_tokens_from_usage(usage: dict[str, int] | None) -> int:
+    if not usage:
+        return 0
+    prompt_tokens = usage.get("prompt_tokens", 0)
+    completion_tokens = usage.get("completion_tokens", 0)
+    total_tokens = usage.get("total_tokens", 0)
+    values = (prompt_tokens, completion_tokens, total_tokens)
+    if any(isinstance(value, bool) for value in values):
+        return 0
+    return max(0, total_tokens or prompt_tokens + completion_tokens)
+
+
+def _state_items(
+    state: ProviderConversationState,
+) -> list[dict[str, Any]] | None:
+    raw_items = state.payload.get(_ITEMS_KEY)
+    if not isinstance(raw_items, list):
+        return None
+    items: list[dict[str, Any]] = []
+    for raw in cast(list[object], raw_items):
+        if not isinstance(raw, dict):
+            return None
+        items.append(cast(dict[str, Any], raw))
+    return items

nanobot/session/manager.py

@@ -17,6 +17,7 @@ from weakref import WeakValueDictionary
 from loguru import logger
 
 from nanobot.config.paths import get_legacy_sessions_dir
+from nanobot.providers.base import ProviderConversationState
 from nanobot.runtime_context import (
     RUNTIME_CONTEXT_HISTORY_META,
     public_history_message,
@@ -43,6 +44,10 @@ _SESSION_PREVIEW_MAX_CHARS = 120
 _SESSION_LIST_PREVIEW_MAX_RECORDS = 200
 _SESSION_LIST_PREVIEW_MAX_CHARS = 1_000_000
 _SESSION_DATA_ERRORS = (ValueError, TypeError, AttributeError, KeyError)
+_PROVIDER_STATE_RECORD_TYPE = "provider_state"
+_PROVIDER_STATE_RECORD_PREFIX_RE = re.compile(
+    r'^\s*\{\s*"_type"\s*:\s*"provider_state"\s*(?:,|\})'
+)
 _FORK_VOLATILE_METADATA_KEYS = {
     "goal_state",
     "pending_user_turn",
@@ -60,6 +65,11 @@ def _json_object(value: object) -> dict[str, Any]:
     return cast(dict[str, Any], value)
 
 
+def _is_provider_state_record_line(line: str) -> bool:
+    """Recognize the canonical private record without decoding its opaque payload."""
+    return _PROVIDER_STATE_RECORD_PREFIX_RE.match(line) is not None
+
+
 def replay_max_messages_for_context(context_window_tokens: int | None) -> int:
     if not context_window_tokens or context_window_tokens <= 0:
         return FILE_MAX_MESSAGES
@@ -146,10 +156,13 @@ class Session:
     updated_at: datetime = field(default_factory=datetime.now)
     metadata: dict[str, Any] = field(default_factory=dict)
     last_consolidated: int = 0  # Number of messages already consolidated to files
+    provider_state: ProviderConversationState | None = field(default=None, repr=False)
 
     def __post_init__(self) -> None:
         if not isinstance(cast(object, self.metadata), dict):
             self.metadata = {}
+        if not isinstance(cast(object, self.provider_state), ProviderConversationState):
+            self.provider_state = None
         # An out-of-range offset (corrupt metadata) would hide all history; reset it.
         last_consolidated = cast(object, self.last_consolidated)
         if (
@@ -304,6 +317,7 @@ class Session:
         """Clear all messages and reset session to initial state."""
         self.messages = []
         self.last_consolidated = 0
+        self.provider_state = None
         self.updated_at = datetime.now()
         self.metadata.pop("_last_summary", None)
 
@@ -396,6 +410,8 @@ class Session:
 
         self.messages = retained
         self.last_consolidated = new_lc
+        if dropped:
+            self.provider_state = None
         self.updated_at = datetime.now()
         return RetentionResult(
             dropped=dropped,
@@ -517,6 +533,7 @@ class JsonlSessionStore:
             created_at: datetime | None = None
             updated_at: datetime | None = None
             last_consolidated = 0
+            provider_state: ProviderConversationState | None = None
 
             with open(path, encoding="utf-8") as f:
                 for line in f:
@@ -527,7 +544,8 @@ class JsonlSessionStore:
                     raw_data: object = json.loads(line)
                     data = _json_object(raw_data)
 
-                    if data.get("_type") == "metadata":
+                    record_type = data.get("_type")
+                    if record_type == "metadata":
                         metadata_value = cast(object, data.get("metadata", {}))
                         metadata = (
                             cast(dict[str, Any], metadata_value)
@@ -552,6 +570,10 @@ class JsonlSessionStore:
                             if isinstance(offset, int) and not isinstance(offset, bool)
                             else 0
                         )
+                    elif record_type == _PROVIDER_STATE_RECORD_TYPE:
+                        provider_state = ProviderConversationState.from_private_record(
+                            data.get("state")
+                        )
                     else:
                         messages.append(data)
 
@@ -562,6 +584,7 @@ class JsonlSessionStore:
                 updated_at=updated_at or datetime.now(),
                 metadata=metadata,
                 last_consolidated=last_consolidated,
+                provider_state=provider_state,
             )
         except _SESSION_DATA_ERRORS as e:
             logger.warning("Failed to load session {}: {}", key, e)
@@ -586,6 +609,7 @@ class JsonlSessionStore:
             created_at: datetime | None = None
             updated_at: datetime | None = None
             last_consolidated = 0
+            provider_state: ProviderConversationState | None = None
             skipped = 0
 
             with open(path, encoding="utf-8") as f:
@@ -603,7 +627,8 @@ class JsonlSessionStore:
                         continue
                     data = cast(dict[str, Any], raw_data)
 
-                    if data.get("_type") == "metadata":
+                    record_type = data.get("_type")
+                    if record_type == "metadata":
                         metadata_value = cast(object, data.get("metadata", {}))
                         metadata = (
                             cast(dict[str, Any], metadata_value)
@@ -624,13 +649,21 @@ class JsonlSessionStore:
                             if isinstance(offset, int) and not isinstance(offset, bool)
                             else 0
                         )
+                    elif record_type == _PROVIDER_STATE_RECORD_TYPE:
+                        candidate = ProviderConversationState.from_private_record(
+                            data.get("state")
+                        )
+                        if candidate is None:
+                            skipped += 1
+                        else:
+                            provider_state = candidate
                     else:
                         messages.append(data)
 
             if skipped:
                 logger.warning("Skipped {} corrupt lines in session {}", skipped, key)
 
-            if not messages and not metadata:
+            if not messages and not metadata and provider_state is None:
                 return None
 
             return Session(
@@ -640,6 +673,7 @@ class JsonlSessionStore:
                 updated_at=updated_at or datetime.now(),
                 metadata=metadata,
                 last_consolidated=last_consolidated,
+                provider_state=provider_state,
             )
         except _SESSION_DATA_ERRORS as e:
             logger.warning("Repair failed for session {}: {}", key, e)
@@ -670,6 +704,12 @@ class JsonlSessionStore:
                     "last_consolidated": session.last_consolidated,
                 }
                 f.write(json.dumps(metadata_line, ensure_ascii=False) + "\n")
+                if session.provider_state is not None:
+                    provider_state_line = {
+                        "_type": _PROVIDER_STATE_RECORD_TYPE,
+                        "state": session.provider_state.to_private_record(),
+                    }
+                    f.write(json.dumps(provider_state_line, ensure_ascii=False) + "\n")
                 for msg in session.messages:
                     f.write(json.dumps(msg, ensure_ascii=False) + "\n")
                 if fsync:
@@ -726,7 +766,8 @@ class JsonlSessionStore:
                         continue
                     raw_data: object = json.loads(line)
                     data = _json_object(raw_data)
-                    if data.get("_type") == "metadata":
+                    record_type = data.get("_type")
+                    if record_type == "metadata":
                         metadata_value = cast(object, data.get("metadata", {}))
                         metadata = (
                             cast(dict[str, Any], metadata_value)
@@ -745,6 +786,8 @@ class JsonlSessionStore:
                         stored_key = (
                             stored_key_value if isinstance(stored_key_value, str) else None
                         )
+                    elif record_type == _PROVIDER_STATE_RECORD_TYPE:
+                        continue
                     else:
                         messages.append(data)
             return {
@@ -837,6 +880,8 @@ class JsonlSessionStore:
                             for line in f:
                                 if not line.strip():
                                     continue
+                                if _is_provider_state_record_line(line):
+                                    continue
                                 scanned_records += 1
                                 scanned_chars += len(line)
                                 if (
@@ -846,7 +891,10 @@ class JsonlSessionStore:
                                     break
                                 raw_item: object = json.loads(line)
                                 item = _json_object(raw_item)
-                                if item.get("_type") == "metadata":
+                                if item.get("_type") in {
+                                    "metadata",
+                                    _PROVIDER_STATE_RECORD_TYPE,
+                                }:
                                     continue
                                 text = _message_preview_text(item)
                                 if not text:

nanobot/utils/gitstore.py

@@ -176,7 +176,10 @@ class GitStore:
             )
             if cast(object, sha_bytes) is None:
                 return None
-            sha = sha_bytes.hex()[:8]
+            # porcelain.commit returns the id as a 40-char hex string that is
+            # already encoded to bytes; .hex() would encode those ASCII bytes
+            # again and produce an id no git command can resolve.
+            sha = sha_bytes.decode()[:8]
             logger.debug("Git auto-commit: {} ({})", sha, message)
             return sha
         except Exception as exc:
@@ -200,7 +203,7 @@ class GitStore:
                     return None
 
                 while sha:
-                    if sha.hex().startswith(short_sha):
+                    if sha.decode().startswith(short_sha):
                         return sha
                     commit_obj = repo[sha]
                     if commit_obj.type_name != b"commit":
@@ -280,7 +283,7 @@ class GitStore:
                     msg = commit.message.decode("utf-8", errors="replace").strip()
                     if message_prefix is None or msg.startswith(message_prefix):
                         entries.append(CommitInfo(
-                            sha=sha.hex()[:8],
+                            sha=sha.decode()[:8],
                             message=msg,
                             timestamp=ts,
                         ))
@@ -484,7 +487,7 @@ class GitStore:
                     with Repo(str(self._workspace)) as repo:
                         commit = cast("Commit", repo[full_sha])
                         parent = commit.parents[0] if commit.parents else None
-                    diff = self.diff_commits(parent.hex()[:8], c.sha) if parent else ""
+                    diff = self.diff_commits(parent.decode()[:8], c.sha) if parent else ""
                     return c, diff
             return None
         except Exception as exc:

nanobot/webui/session_list_index.py

@@ -18,10 +18,12 @@ from loguru import logger
 from nanobot.config.paths import get_webui_dir
 from nanobot.session.history_visibility import is_hidden_history_message
 from nanobot.session.manager import (
+    _PROVIDER_STATE_RECORD_TYPE,  # pyright: ignore[reportPrivateUsage]
     _SESSION_LIST_PREVIEW_MAX_CHARS,  # pyright: ignore[reportPrivateUsage]
     _SESSION_LIST_PREVIEW_MAX_RECORDS,  # pyright: ignore[reportPrivateUsage]
     Session,
     SessionManager,
+    _is_provider_state_record_line,  # pyright: ignore[reportPrivateUsage]
     _message_preview_text,  # pyright: ignore[reportPrivateUsage]
     _metadata_title,  # pyright: ignore[reportPrivateUsage]
 )
@@ -298,7 +300,11 @@ def _scan_session_row(session_manager: SessionManager, path: Path) -> dict[str,
             for line in f:
                 if not line.strip():
                     continue
+                if _is_provider_state_record_line(line):
+                    continue
                 item = json.loads(line)
+                if item.get("_type") == _PROVIDER_STATE_RECORD_TYPE:
+                    continue
                 timestamp = _visible_message_timestamp(item)
                 if timestamp is not None:
                     visible_message_at = _latest_updated_at(visible_message_at, timestamp)

nanobot/webui/token_usage.py

@@ -159,6 +159,13 @@ def normalize_token_usage_state(raw: Any) -> dict[str, Any]:
         if not isinstance(date, str) or len(date) != 10 or not isinstance(row_value, dict):
             continue
         row = cast(dict[str, Any], row_value)
+        try:
+            datetime.fromisoformat(date)
+        except ValueError:
+            # A hand-edited or foreign day key that is not a real date would
+            # otherwise reach token_usage_payload's date parsing and fail every
+            # settings request; drop it like any other malformed row.
+            continue
         normalized = _normalize_usage_row(row)
         if normalized["total_tokens"] <= 0 and normalized["requests"] <= 0:
             continue

nanobot/webui/transcript.py

@@ -2026,6 +2026,7 @@ def replay_transcript_to_ui_messages(
                 continue
             close_activity_for_answer()
             turn_fields = _turn_fields(rec, "answer")
+            source_fields = _source_fields(rec)
             adopted = find_active_placeholder(messages, turn_fields) if buffer_message_id is None else None
             if buffer_message_id is None:
                 if adopted:
@@ -2038,7 +2039,8 @@ def replay_transcript_to_ui_messages(
                             "role": "assistant",
                             "content": "",
                             "isStreaming": True,
-                            **_turn_fields(rec, "answer"),
+                            **turn_fields,
+                            **source_fields,
                             "createdAt": _created_at_ms(rec, idx),
                         },
                     )
@@ -2050,7 +2052,8 @@ def replay_transcript_to_ui_messages(
                         **m,
                         "content": combined,
                         "isStreaming": True,
-                        **_turn_fields(rec, "answer"),
+                        **turn_fields,
+                        **source_fields,
                     }
                     break
             continue
@@ -2062,6 +2065,8 @@ def replay_transcript_to_ui_messages(
                 continue
             merge_next = rec.get("resuming") is True and rec.get("merge_next") is True
             final_text = rec.get("text")
+            turn_fields = _turn_fields(rec, "answer")
+            source_fields = _source_fields(rec)
             if isinstance(final_text, str):
                 if buffer_message_id is None:
                     buffer_message_id = _new_id("buf", idx)
@@ -2071,7 +2076,8 @@ def replay_transcript_to_ui_messages(
                             "role": "assistant",
                             "content": final_text,
                             "isStreaming": True,
-                            **_turn_fields(rec, "answer"),
+                            **turn_fields,
+                            **source_fields,
                             "createdAt": _created_at_ms(rec, idx),
                         },
                     )
@@ -2082,11 +2088,21 @@ def replay_transcript_to_ui_messages(
                                 **m,
                                 "content": final_text,
                                 "isStreaming": True,
-                                **_turn_fields(rec, "answer"),
+                                **turn_fields,
+                                **source_fields,
                             }
                             break
                 if merge_next:
                     buffer_parts = [final_text]
+            elif source_fields and buffer_message_id is not None:
+                for i, m in enumerate(messages):
+                    if m.get("id") == buffer_message_id:
+                        messages[i] = {
+                            **m,
+                            **turn_fields,
+                            **source_fields,
+                        }
+                        break
             if not merge_next:
                 buffer_message_id = None
                 buffer_parts = []

tests/agent/test_autocompact_unit.py

@@ -154,6 +154,26 @@ class TestIsExpired:
         now_over = datetime(2026, 1, 1, 10, 10, 0)
         assert ac._is_expired(ts, now=now_over) is True
 
+    def test_unparseable_string_timestamp_returns_false(self):
+        """A persisted timestamp that no longer parses must not raise.
+
+        list_sessions() forwards the raw persisted updated_at string, and
+        SessionManager._load already tolerates a malformed value through its
+        recovery path. The idle scan must mirror that tolerance instead of crashing.
+        """
+        ac = _make_autocompact(ttl=15)
+        assert ac._is_expired("not-a-timestamp") is False
+
+    def test_tz_aware_string_timestamp_is_compared_by_instant(self):
+        """A valid timestamp with an offset remains eligible for expiry."""
+        ac = _make_autocompact(ttl=15)
+        now = datetime(2026, 1, 1, 12, 0, 0)
+        recent = (now - timedelta(minutes=10)).astimezone().isoformat()
+        expired = (now - timedelta(minutes=20)).astimezone().isoformat()
+
+        assert ac._is_expired(recent, now=now) is False
+        assert ac._is_expired(expired, now=now) is True
+
 
 # ---------------------------------------------------------------------------
 # _format_summary
@@ -221,6 +241,36 @@ class TestCheckExpired:
         assert len(scheduled) == 1
         assert "cli:old" in ac._archiving
 
+    def test_unparseable_updated_at_does_not_stop_scan(self):
+        """A malformed timestamp is skipped without hiding later sessions.
+
+        The idle scan runs from the agent loop's inbound-timeout branch, so a
+        raised exception here would tear down the loop. list_sessions() forwards
+        the raw string, so check_expired must tolerate it like SessionManager
+        does when loading.
+        """
+        ac = _make_autocompact(ttl=15)
+        mock_sm = MagicMock(spec=SessionManager)
+        old_dt = datetime.now() - timedelta(minutes=20)
+        session = _make_session("cli:old", updated_at=old_dt)
+        _add_turns(session, 5)
+        mock_sm.list_sessions.return_value = [
+            {"key": "cli:corrupt", "updated_at": "not-a-timestamp"},
+            {"key": "cli:old", "updated_at": old_dt.isoformat()},
+        ]
+        mock_sm.get_or_create.return_value = session
+        ac.sessions = mock_sm
+        scheduled = []
+
+        def scheduler(coro):
+            scheduled.append(coro)
+            coro.close()
+
+        ac.check_expired(scheduler, _runtime)
+
+        assert len(scheduled) == 1
+        assert ac._archiving == {"cli:old"}
+
     @pytest.mark.asyncio
     async def test_runtime_is_captured_before_background_starts(self):
         ac = _make_autocompact(ttl=15)

tests/agent/test_consolidator.py

@@ -10,7 +10,11 @@ from nanobot.agent.memory import (
     Consolidator,
     MemoryStore,
 )
-from nanobot.providers.base import GenerationSettings, LLMResponse
+from nanobot.providers.base import (
+    GenerationSettings,
+    LLMResponse,
+    ProviderConversationState,
+)
 from nanobot.runtime_context import (
     RUNTIME_CONTEXT_HISTORY_META,
     RuntimeContextBlock,
@@ -74,6 +78,16 @@ def _tool_round(call_id: str) -> list[dict]:
     ]
 
 
+def _provider_state() -> ProviderConversationState:
+    return ProviderConversationState(
+        kind="openai_responses",
+        provider="openai:test",
+        model="test-model",
+        version=1,
+        payload={"items": []},
+    )
+
+
 class TestConsolidatorSummarize:
     async def test_archive_prompt_includes_media_breadcrumb(
         self, consolidator, mock_provider, store, runtime
@@ -385,6 +399,7 @@ class TestConsolidatorTokenBudget:
         """Old messages that cannot be replayed should be materialized first."""
         consolidator._SAFETY_BUFFER = 0
         session = Session(key="test:replay-overflow")
+        session.provider_state = _provider_state()
         for i in range(10):
             session.add_message("user", f"u{i}")
             session.add_message("assistant", f"a{i}")
@@ -404,6 +419,7 @@ class TestConsolidatorTokenBudget:
         assert archived_chunk[-1]["content"] == "a6"
         assert session.last_consolidated == 14
         assert session.metadata["_last_summary"]["text"] == "old conversation summary"
+        assert session.provider_state is None
         consolidator.sessions.save.assert_called()
 
     async def test_replay_window_overflow_extends_to_long_recent_user_turn(
@@ -479,6 +495,7 @@ class TestConsolidatorTokenBudget:
         session = MagicMock()
         session.last_consolidated = 0
         session.key = "test:key"
+        session.provider_state = _provider_state()
         session.messages = [
             {
                 "role": "user" if i in {0, 50, 61} else "assistant",
@@ -500,6 +517,7 @@ class TestConsolidatorTokenBudget:
         # pick_consolidation_boundary returns (50, tokens) — user turn at idx 50
         assert archived_chunk[0]["content"] == "m0"
         assert session.last_consolidated > 0
+        assert session.provider_state is None
 
     async def test_raw_archive_fallback_advances_last_consolidated(
         self, consolidator, runtime
@@ -610,6 +628,7 @@ class TestCompactIdleSession:
         )
         sessions = real_consolidator.sessions
         session = sessions.get_or_create("cli:test")
+        session.provider_state = _provider_state()
         old_ts = session.updated_at
         for i in range(20):
             session.add_message("user", f"user msg {i}")
@@ -627,6 +646,7 @@ class TestCompactIdleSession:
         assert len(reloaded.messages) == 40
         assert reloaded.messages[0]["content"] == "user msg 0"
         assert reloaded.last_consolidated == 32
+        assert reloaded.provider_state is None
         visible = reloaded.get_history(max_messages=40)
         assert len(visible) == 8
         assert visible[0]["content"] == "user msg 16"

tests/agent/test_context_builder.py

@@ -452,6 +452,20 @@ class TestBuildMessages:
         assert "previous user message" in str(messages[1]["content"])
         assert "new message" in str(messages[1]["content"])
 
+    def test_current_message_can_be_built_without_history_merge(self, tmp_path):
+        builder = _builder(tmp_path)
+        current = builder.build_current_message(
+            "new message",
+            runtime_context_blocks=[
+                RuntimeContextBlock(source="test", content="fresh context"),
+            ],
+        )
+
+        assert current["role"] == "user"
+        assert "new message" in current["content"]
+        assert "fresh context" in current["content"]
+        assert current["_meta"]["runtime_context"]["sources"] == ["test"]
+
     def test_different_role_appended(self, tmp_path):
         builder = _builder(tmp_path)
         history = [{"role": "assistant", "content": "previous response"}]

tests/agent/test_loop_save_turn.py

@@ -1,4 +1,5 @@
 import asyncio
+import json
 from pathlib import Path
 from types import SimpleNamespace
 from unittest.mock import AsyncMock, MagicMock
@@ -19,7 +20,7 @@ from nanobot.bus.outbound_events import (
 )
 from nanobot.bus.queue import MessageBus
 from nanobot.cron.session_turns import CRON_HISTORY_META, CRON_TRIGGER_META
-from nanobot.providers.base import LLMResponse
+from nanobot.providers.base import LLMProvider, LLMResponse, ProviderConversationState
 from nanobot.providers.factory import ProviderSnapshot
 from nanobot.runtime_context import (
     RUNTIME_CONTEXT_HISTORY_META,
@@ -59,6 +60,16 @@ def _mk_loop() -> AgentLoop:
     return loop
 
 
+def _provider_state() -> ProviderConversationState:
+    return ProviderConversationState(
+        kind="openai_responses",
+        provider="openai:test",
+        model="test-model",
+        version=1,
+        payload={"items": []},
+    )
+
+
 def _runtime_message(content, blocks: list[RuntimeContextBlock]) -> dict:
     merged, marker = append_runtime_context(content, blocks)
     assert marker is not None
@@ -494,6 +505,7 @@ def test_restore_runtime_checkpoint_rehydrates_completed_and_pending_tools() ->
     loop = _mk_loop()
     session = Session(
         key="test:checkpoint",
+        provider_state=_provider_state(),
         metadata={
             AgentLoop._RUNTIME_CHECKPOINT_KEY: {
                 "assistant_message": {
@@ -539,6 +551,104 @@ def test_restore_runtime_checkpoint_rehydrates_completed_and_pending_tools() ->
     assert session.messages[1]["tool_call_id"] == "call_done"
     assert session.messages[2]["tool_call_id"] == "call_pending"
     assert "interrupted before this tool finished" in session.messages[2]["content"].lower()
+    assert session.provider_state is None
+
+
+def test_restore_final_response_checkpoint_preserves_matching_provider_state() -> None:
+    loop = _mk_loop()
+    state = _provider_state()
+    session = Session(
+        key="test:final-checkpoint",
+        provider_state=state,
+        metadata={
+            AgentLoop._RUNTIME_CHECKPOINT_KEY: {
+                "phase": "final_response",
+                AgentLoop._PROVIDER_STATE_CHECKPOINT_VERSION_KEY: (
+                    AgentLoop._PROVIDER_STATE_CHECKPOINT_VERSION
+                ),
+                "assistant_message": {
+                    "role": "assistant",
+                    "content": "finished",
+                },
+                "completed_tool_results": [],
+                "pending_tool_calls": [],
+            }
+        },
+    )
+
+    restored = loop._restore_runtime_checkpoint(session)
+
+    assert restored is True
+    assert session.messages[-1]["content"] == "finished"
+    assert session.provider_state is state
+    assert session.metadata.get(AgentLoop._RUNTIME_CHECKPOINT_KEY) is None
+
+
+def test_restore_legacy_final_checkpoint_discards_unproven_provider_state() -> None:
+    loop = _mk_loop()
+    session = Session(
+        key="test:legacy-final-checkpoint",
+        provider_state=_provider_state(),
+        metadata={
+            AgentLoop._RUNTIME_CHECKPOINT_KEY: {
+                "phase": "final_response",
+                "assistant_message": {
+                    "role": "assistant",
+                    "content": "finished",
+                },
+                "completed_tool_results": [],
+                "pending_tool_calls": [],
+            }
+        },
+    )
+
+    restored = loop._restore_runtime_checkpoint(session)
+
+    assert restored is True
+    assert session.messages[-1]["content"] == "finished"
+    assert session.provider_state is None
+
+
+def test_restore_completed_tools_checkpoint_preserves_matching_provider_state() -> None:
+    loop = _mk_loop()
+    tool_result = {
+        "role": "tool",
+        "tool_call_id": "call_done",
+        "name": "read_file",
+        "content": "compacted result",
+    }
+    state = _provider_state().with_pending_messages([tool_result])
+    session = Session(
+        key="test:completed-tools-checkpoint",
+        provider_state=state,
+        metadata={
+            AgentLoop._RUNTIME_CHECKPOINT_KEY: {
+                "phase": "tools_completed",
+                AgentLoop._PROVIDER_STATE_CHECKPOINT_VERSION_KEY: (
+                    AgentLoop._PROVIDER_STATE_CHECKPOINT_VERSION
+                ),
+                "assistant_message": {
+                    "role": "assistant",
+                    "content": None,
+                    "tool_calls": [
+                        {
+                            "id": "call_done",
+                            "type": "function",
+                            "function": {"name": "read_file", "arguments": "{}"},
+                        }
+                    ],
+                },
+                "completed_tool_results": [tool_result],
+                "pending_tool_calls": [],
+            }
+        },
+    )
+
+    restored = loop._restore_runtime_checkpoint(session)
+
+    assert restored is True
+    assert session.messages[-1]["content"] == "compacted result"
+    assert session.provider_state is state
 
 
 def test_restore_runtime_checkpoint_dedupes_overlapping_tail() -> None:
@@ -616,6 +726,55 @@ def test_restore_runtime_checkpoint_dedupes_overlapping_tail() -> None:
     assert session.messages[2]["tool_call_id"] == "call_pending"
 
 
+@pytest.mark.asyncio
+async def test_runtime_checkpoint_keeps_provider_state_out_of_public_metadata(
+    tmp_path: Path,
+) -> None:
+    loop = _make_full_loop(tmp_path)
+    state = ProviderConversationState(
+        kind="openai_responses",
+        provider="openai:test",
+        model="test-model",
+        version=1,
+        payload={
+            "items": [
+                {
+                    "type": "reasoning",
+                    "encrypted_content": "private-checkpoint-blob",
+                }
+            ]
+        },
+    )
+    loop.provider.can_resume_conversation_state.return_value = True
+    loop.provider.chat_with_retry = AsyncMock(
+        return_value=LLMResponse(content="done", provider_state=state)
+    )
+    session = loop.sessions.get_or_create("cli:private-checkpoint")
+
+    await loop._run_agent_loop(
+        [
+            {"role": "system", "content": "system"},
+            {"role": "user", "content": "question"},
+        ],
+        runtime=loop.llm_runtime(),
+        session=session,
+    )
+
+    assert session.provider_state is not None
+    checkpoint = session.metadata[AgentLoop._RUNTIME_CHECKPOINT_KEY]
+    assert "provider_state" not in checkpoint
+    assert checkpoint[AgentLoop._PROVIDER_STATE_CHECKPOINT_VERSION_KEY] == (
+        AgentLoop._PROVIDER_STATE_CHECKPOINT_VERSION
+    )
+    assert "private-checkpoint-blob" not in json.dumps(session.metadata)
+
+    public_payload = loop.sessions.read_session_file(session.key)
+    assert public_payload is not None
+    assert "private-checkpoint-blob" not in json.dumps(public_payload)
+    raw = loop.sessions._get_session_path(session.key).read_text(encoding="utf-8")
+    assert "private-checkpoint-blob" in raw
+
+
 @pytest.mark.asyncio
 async def test_process_message_persists_user_message_before_turn_completes(tmp_path: Path) -> None:
     loop = _make_full_loop(tmp_path)
@@ -634,6 +793,150 @@ async def test_process_message_persists_user_message_before_turn_completes(tmp_p
     assert persisted.updated_at >= persisted.created_at
 
 
+@pytest.mark.asyncio
+async def test_subagent_followup_stages_provider_state_before_turn_runs(
+    tmp_path: Path,
+) -> None:
+    loop = _make_full_loop(tmp_path)
+    loop.consolidator.maybe_consolidate_by_tokens = AsyncMock(return_value=False)  # type: ignore[method-assign]
+    loop._run_agent_loop = AsyncMock(side_effect=RuntimeError("boom"))  # type: ignore[method-assign]
+    loop.provider.can_resume_conversation_state.return_value = True
+    session = loop.sessions.get_or_create("cli:subagent-crash")
+    session.provider_state = _provider_state()
+    loop.sessions.save(session)
+
+    msg = InboundMessage(
+        channel="system",
+        sender_id="subagent",
+        chat_id="cli:subagent-crash",
+        content="subagent result",
+        metadata={"subagent_task_id": "sub-1"},
+    )
+    with pytest.raises(RuntimeError, match="boom"):
+        await loop._process_message(msg)
+
+    loop.sessions.invalidate("cli:subagent-crash")
+    persisted = loop.sessions.get_or_create("cli:subagent-crash")
+    assert persisted.messages[-1]["content"] == "subagent result"
+    assert persisted.provider_state is not None
+    assert persisted.provider_state.pending_messages[-1]["role"] == "user"
+    assert persisted.provider_state.pending_messages[-1]["content"] == "subagent result"
+
+
+@pytest.mark.asyncio
+async def test_subagent_followup_state_is_durable_before_prompt_assembly(
+    tmp_path: Path,
+) -> None:
+    loop = _make_full_loop(tmp_path)
+    loop.consolidator.maybe_consolidate_by_tokens = AsyncMock(return_value=False)  # type: ignore[method-assign]
+    loop.provider.can_resume_conversation_state.return_value = True
+    loop._build_initial_messages = MagicMock(  # type: ignore[method-assign]
+        side_effect=RuntimeError("prompt boom"),
+    )
+    session = loop.sessions.get_or_create("cli:subagent-prompt-crash")
+    session.provider_state = _provider_state()
+    loop.sessions.save(session)
+
+    msg = InboundMessage(
+        channel="system",
+        sender_id="subagent",
+        chat_id="cli:subagent-prompt-crash",
+        content="subagent result",
+        metadata={"subagent_task_id": "sub-1"},
+    )
+    with pytest.raises(RuntimeError, match="prompt boom"):
+        await loop._process_message(msg)
+
+    loop.sessions.invalidate("cli:subagent-prompt-crash")
+    persisted = loop.sessions.get_or_create("cli:subagent-prompt-crash")
+    assert persisted.messages[-1]["content"] == "subagent result"
+    assert persisted.provider_state is not None
+    assert persisted.provider_state.pending_messages[-1]["content"] == (
+        "subagent result"
+    )
+
+
+@pytest.mark.asyncio
+async def test_subagent_redelivery_does_not_duplicate_staged_provider_input(
+    tmp_path: Path,
+) -> None:
+    loop = _make_full_loop(tmp_path)
+    loop.consolidator.maybe_consolidate_by_tokens = AsyncMock(return_value=False)  # type: ignore[method-assign]
+    loop.provider.can_resume_conversation_state.return_value = True
+    build_initial_messages = loop._build_initial_messages
+    loop._build_initial_messages = MagicMock(  # type: ignore[method-assign]
+        side_effect=RuntimeError("prompt boom"),
+    )
+    session = loop.sessions.get_or_create("cli:subagent-redelivery")
+    session.provider_state = _provider_state()
+    loop.sessions.save(session)
+    msg = InboundMessage(
+        channel="system",
+        sender_id="subagent",
+        chat_id="cli:subagent-redelivery",
+        content="subagent result",
+        metadata={"subagent_task_id": "sub-1"},
+    )
+
+    with pytest.raises(RuntimeError, match="prompt boom"):
+        await loop._process_message(msg)
+
+    loop.sessions.invalidate("cli:subagent-redelivery")
+    persisted = loop.sessions.get_or_create("cli:subagent-redelivery")
+    assert persisted.provider_state is not None
+    assert [
+        message.get("content")
+        for message in persisted.provider_state.pending_messages
+    ].count("subagent result") == 1
+    loop._build_initial_messages = build_initial_messages  # type: ignore[method-assign]
+    loop._run_agent_loop = AsyncMock(  # type: ignore[method-assign]
+        side_effect=RuntimeError("provider boom"),
+    )
+    with pytest.raises(RuntimeError, match="provider boom"):
+        await loop._process_message(msg)
+
+    provider_state = loop._run_agent_loop.await_args.kwargs["provider_state"]
+    assert provider_state is not None
+    pending_results = [
+        message
+        for message in provider_state.pending_messages
+        if message.get("content") == "subagent result"
+    ]
+    assert len(pending_results) == 1
+    assert LLMProvider._sanitize_empty_content(pending_results) == [
+        {"role": "user", "content": "subagent result"},
+    ]
+
+
+@pytest.mark.asyncio
+async def test_subagent_followup_clears_state_before_compatibility_failure(
+    tmp_path: Path,
+) -> None:
+    loop = _make_full_loop(tmp_path)
+    loop.consolidator.maybe_consolidate_by_tokens = AsyncMock(return_value=False)  # type: ignore[method-assign]
+    loop.provider.can_resume_conversation_state.side_effect = RuntimeError(
+        "compatibility boom"
+    )
+    session = loop.sessions.get_or_create("cli:subagent-compat-crash")
+    session.provider_state = _provider_state()
+    loop.sessions.save(session)
+
+    msg = InboundMessage(
+        channel="system",
+        sender_id="subagent",
+        chat_id="cli:subagent-compat-crash",
+        content="subagent result",
+        metadata={"subagent_task_id": "sub-1"},
+    )
+    with pytest.raises(RuntimeError, match="compatibility boom"):
+        await loop._process_message(msg)
+
+    loop.sessions.invalidate("cli:subagent-compat-crash")
+    persisted = loop.sessions.get_or_create("cli:subagent-compat-crash")
+    assert persisted.messages[-1]["content"] == "subagent result"
+    assert persisted.provider_state is None
+
+
 @pytest.mark.asyncio
 async def test_process_message_persists_unified_session_delivery_route(tmp_path: Path) -> None:
     loop = _make_full_loop(tmp_path)
@@ -1245,6 +1548,9 @@ async def test_next_turn_after_crash_closes_pending_user_turn_before_new_input(t
     session = loop.sessions.get_or_create("feishu:c3")
     session.add_message("user", "old question")
     session.metadata[AgentLoop._PENDING_USER_TURN_KEY] = True
+    session.provider_state = _provider_state().with_pending_messages([
+        {"role": "user", "content": "old question"},
+    ])
     loop.sessions.save(session)
 
     loop._run_agent_loop = AsyncMock(return_value=(
@@ -1278,6 +1584,7 @@ async def test_next_turn_after_crash_closes_pending_user_turn_before_new_input(t
         {"role": "assistant", "content": "new answer"},
     ]
     assert AgentLoop._PENDING_USER_TURN_KEY not in session.metadata
+    assert session.provider_state is None
 
 
 @pytest.mark.asyncio

tests/agent/test_mcp_reconnect_crash.py

@@ -27,8 +27,10 @@ from nanobot.bus.queue import MessageBus
 from nanobot.config.schema import MCPServerConfig
 from nanobot.security import network as security_network
 
-_IDLE_TIMEOUT_SECONDS = 0.25
+# Leave enough headroom for reconnect handshakes on slower CI hosts; each test
-_IDLE_EXPIRY_GRACE_SECONDS = 0.25
+# still waits beyond this deadline explicitly before exercising recovery.
+_IDLE_TIMEOUT_SECONDS = 1.0
+_IDLE_EXPIRY_GRACE_SECONDS = 0.5
 _TOOL_TIMEOUT_SECONDS = 10
 
 

tests/agent/test_runner_core.py

@@ -11,7 +11,13 @@ import pytest
 
 from agent.runner_helpers import make_run_spec
 from nanobot.config.schema import AgentDefaults
-from nanobot.providers.base import LLMProvider, LLMResponse, ToolCallRequest
+from nanobot.providers.base import (
+    LLMProvider,
+    LLMResponse,
+    ProviderCallContext,
+    ProviderConversationState,
+    ToolCallRequest,
+)
 
 _MAX_TOOL_RESULT_CHARS = AgentDefaults().max_tool_result_chars
 
@@ -73,6 +79,311 @@ async def test_runner_preserves_reasoning_fields_and_tool_results():
     )
 
 
+@pytest.mark.asyncio
+async def test_runner_replays_provider_state_without_chat_projection_duplicates():
+    from nanobot.agent.runner import AgentRunner
+
+    provider = MagicMock(spec=LLMProvider)
+    provider.can_resume_conversation_state.return_value = True
+    provider.supports_native_compaction.return_value = False
+    captured_second_kwargs: dict = {}
+    checkpoints: list[dict] = []
+    calls = 0
+
+    async def checkpoint(payload: dict) -> None:
+        checkpoints.append(payload)
+
+    first_state = ProviderConversationState(
+        kind="openai_responses",
+        provider="openai:test",
+        model="gpt-5.6",
+        version=1,
+        payload={"items": [{"type": "reasoning", "encrypted_content": "opaque"}]},
+    )
+    second_state = ProviderConversationState(
+        kind="openai_responses",
+        provider="openai:test",
+        model="gpt-5.6",
+        version=1,
+        payload={"items": [{"type": "message", "role": "assistant"}]},
+    )
+
+    async def chat_with_retry(**kwargs):
+        nonlocal calls
+        calls += 1
+        if calls == 1:
+            provider_context = kwargs["provider_context"]
+            assert isinstance(provider_context, ProviderCallContext)
+            assert provider_context.conversation_state is None
+            return LLMResponse(
+                content=None,
+                tool_calls=[
+                    ToolCallRequest(
+                        id="call_1|fc_1",
+                        name="list_dir",
+                        arguments={"path": "."},
+                    ),
+                ],
+                provider_state=first_state,
+            )
+        captured_second_kwargs.update(kwargs)
+        return LLMResponse(content="done", provider_state=second_state)
+
+    provider.chat_with_retry = chat_with_retry
+    tools = MagicMock()
+    tools.get_definitions.return_value = []
+    tools.execute = AsyncMock(return_value="tool result")
+
+    result = await AgentRunner().run(make_run_spec(
+        provider,
+        initial_messages=[
+            {"role": "system", "content": "system"},
+            {"role": "user", "content": "do task"},
+        ],
+        tools=tools,
+        model="gpt-5.6",
+        max_iterations=3,
+        max_tool_result_chars=_MAX_TOOL_RESULT_CHARS,
+        checkpoint_callback=checkpoint,
+    ))
+
+    provider_context = captured_second_kwargs["provider_context"]
+    assert isinstance(provider_context, ProviderCallContext)
+    assert provider_context.conversation_state is not None
+    assert provider_context.conversation_state.payload == first_state.payload
+    assert provider_context.conversation_state.pending_messages == [{
+        "role": "tool",
+        "tool_call_id": "call_1|fc_1",
+        "name": "list_dir",
+        "content": "tool result",
+    }]
+    assert not any(
+        message.get("role") == "assistant"
+        for message in provider_context.conversation_state.pending_messages
+    )
+    assert result.provider_state is not None
+    assert result.provider_state.payload == second_state.payload
+    assert result.provider_state.pending_messages == []
+    assert checkpoints[0]["phase"] == "awaiting_tools"
+    assert "provider_state" not in checkpoints[0]
+    assert checkpoints[1]["phase"] == "tools_completed"
+    assert checkpoints[1]["provider_state"].pending_messages == [{
+        "role": "tool",
+        "tool_call_id": "call_1|fc_1",
+        "name": "list_dir",
+        "content": "tool result",
+    }]
+    assert checkpoints[2]["phase"] == "final_response"
+    assert checkpoints[2]["provider_state"].payload == second_state.payload
+
+
+@pytest.mark.asyncio
+async def test_runner_governs_tool_result_before_adding_it_to_provider_state():
+    from nanobot.agent.runner import AgentRunner
+
+    provider = MagicMock(spec=LLMProvider)
+    provider.can_resume_conversation_state.return_value = True
+    provider.supports_native_compaction.return_value = False
+    calls = 0
+    captured_context: ProviderCallContext | None = None
+    checkpoints: list[dict] = []
+    state = ProviderConversationState(
+        kind="openai_responses",
+        provider="openai:test",
+        model="gpt-5.6",
+        version=1,
+        payload={"items": [{"type": "reasoning", "encrypted_content": "opaque"}]},
+    )
+
+    async def chat_with_retry(**kwargs):
+        nonlocal calls, captured_context
+        calls += 1
+        if calls == 1:
+            return LLMResponse(
+                content=None,
+                tool_calls=[
+                    ToolCallRequest(
+                        id="call_1",
+                        name="read_file",
+                        arguments={"path": "large.txt"},
+                    ),
+                ],
+                provider_state=state,
+            )
+        captured_context = kwargs["provider_context"]
+        return LLMResponse(content="done")
+
+    provider.chat_with_retry = chat_with_retry
+    tools = MagicMock()
+    tools.get_definitions.return_value = []
+    tools.execute = AsyncMock(return_value="x" * 5_000)
+
+    async def checkpoint(payload: dict) -> None:
+        checkpoints.append(payload)
+
+    await AgentRunner().run(make_run_spec(
+        provider,
+        initial_messages=[
+            {"role": "system", "content": "system"},
+            {"role": "user", "content": "read the file"},
+        ],
+        tools=tools,
+        model="gpt-5.6",
+        context_window_tokens=3_000,
+        context_block_limit=200,
+        max_tokens=1_000,
+        max_iterations=3,
+        max_tool_result_chars=10_000,
+        checkpoint_callback=checkpoint,
+    ))
+
+    assert captured_context is not None
+    assert captured_context.conversation_state is not None
+    pending = captured_context.conversation_state.pending_messages
+    assert len(pending) == 1
+    assert pending[0]["role"] == "tool"
+    assert "compacted to fit context" in pending[0]["content"]
+    assert pending[0]["content"] != "x" * 5_000
+    completed_checkpoint = next(
+        checkpoint
+        for checkpoint in checkpoints
+        if checkpoint["phase"] == "tools_completed"
+    )
+    checkpoint_pending = completed_checkpoint["provider_state"].pending_messages
+    assert "compacted to fit context" in checkpoint_pending[0]["content"]
+    assert checkpoint_pending[0]["content"] != "x" * 5_000
+
+
+@pytest.mark.asyncio
+async def test_injected_final_response_checkpoint_includes_provider_state():
+    from nanobot.agent.runner import AgentRunner
+
+    provider = MagicMock(spec=LLMProvider)
+    provider.can_resume_conversation_state.return_value = True
+    provider.supports_native_compaction.return_value = False
+    first_state = ProviderConversationState(
+        kind="openai_responses",
+        provider="openai:test",
+        model="gpt-5.6",
+        version=1,
+        payload={"items": [{"type": "message", "content": "first answer"}]},
+    )
+    second_state = ProviderConversationState(
+        kind="openai_responses",
+        provider="openai:test",
+        model="gpt-5.6",
+        version=1,
+        payload={"items": [{"type": "message", "content": "second answer"}]},
+    )
+    provider.chat_with_retry = AsyncMock(side_effect=[
+        LLMResponse(content="first answer", provider_state=first_state),
+        LLMResponse(content="second answer", provider_state=second_state),
+    ])
+    tools = MagicMock()
+    tools.get_definitions.return_value = []
+    checkpoints: list[dict] = []
+    injections = [[{"role": "user", "content": "follow up"}], []]
+
+    async def checkpoint(payload: dict) -> None:
+        checkpoints.append(payload)
+
+    async def inject() -> list[dict]:
+        return injections.pop(0)
+
+    await AgentRunner().run(make_run_spec(
+        provider,
+        initial_messages=[{"role": "user", "content": "start"}],
+        tools=tools,
+        model="gpt-5.6",
+        max_iterations=3,
+        max_tool_result_chars=_MAX_TOOL_RESULT_CHARS,
+        checkpoint_callback=checkpoint,
+        injection_callback=inject,
+    ))
+
+    assert checkpoints[0]["phase"] == "final_response"
+    assert checkpoints[0]["provider_state"].payload == first_state.payload
+
+
+@pytest.mark.asyncio
+async def test_runner_preserves_last_completed_provider_state_on_model_error():
+    from nanobot.agent.runner import AgentRunner
+
+    provider = MagicMock(spec=LLMProvider)
+    provider.can_resume_conversation_state.return_value = True
+    provider.chat_with_retry = AsyncMock(return_value=LLMResponse(
+        content="temporary upstream failure",
+        finish_reason="error",
+        error_kind="timeout",
+    ))
+    tools = MagicMock()
+    tools.get_definitions.return_value = []
+    state = ProviderConversationState(
+        kind="openai_responses",
+        provider="openai:test",
+        model="gpt-5.6",
+        version=1,
+        payload={"items": [{"type": "reasoning", "encrypted_content": "opaque"}]},
+    )
+    unsaved_input = {"role": "user", "content": "ephemeral follow-up"}
+
+    result = await AgentRunner().run(make_run_spec(
+        provider,
+        initial_messages=[
+            {"role": "system", "content": "system"},
+            unsaved_input,
+        ],
+        tools=tools,
+        model="gpt-5.6",
+        max_iterations=1,
+        max_tool_result_chars=_MAX_TOOL_RESULT_CHARS,
+        provider_state=state.with_pending_messages([unsaved_input]),
+    ))
+
+    assert result.stop_reason == "error"
+    assert result.provider_state is not None
+    assert result.provider_state.payload == state.payload
+    assert result.provider_state.pending_messages[0] == unsaved_input
+    assert result.provider_state.pending_messages[1]["role"] == "assistant"
+    assert "model error" in result.provider_state.pending_messages[1]["content"]
+
+
+@pytest.mark.asyncio
+async def test_runner_discards_provider_state_on_non_retryable_model_error():
+    from nanobot.agent.runner import AgentRunner
+
+    provider = MagicMock(spec=LLMProvider)
+    provider.can_resume_conversation_state.return_value = True
+    provider.chat_with_retry = AsyncMock(return_value=LLMResponse(
+        content="context length exceeded",
+        finish_reason="error",
+        error_status_code=400,
+        error_should_retry=False,
+    ))
+    tools = MagicMock()
+    tools.get_definitions.return_value = []
+    state = ProviderConversationState(
+        kind="openai_responses",
+        provider="openai:test",
+        model="gpt-5.6",
+        version=1,
+        payload={"items": [{"type": "reasoning", "encrypted_content": "opaque"}]},
+    )
+
+    result = await AgentRunner().run(make_run_spec(
+        provider,
+        initial_messages=[{"role": "user", "content": "continue"}],
+        tools=tools,
+        model="gpt-5.6",
+        max_iterations=1,
+        max_tool_result_chars=_MAX_TOOL_RESULT_CHARS,
+        provider_state=state,
+    ))
+
+    assert result.stop_reason == "error"
+    assert result.provider_state is None
+
+
 @pytest.mark.asyncio
 async def test_runner_returns_max_iterations_fallback():
     from nanobot.agent.runner import AgentRunner
@@ -422,6 +733,66 @@ async def test_runner_retries_empty_final_response_with_summary_prompt():
     assert result.usage["completion_tokens"] == 9
 
 
+@pytest.mark.asyncio
+@pytest.mark.parametrize("finish_reason", ["refusal", "content_filter"])
+async def test_runner_does_not_retry_blank_policy_terminal(
+    finish_reason: str,
+) -> None:
+    from nanobot.agent.runner import AgentRunner
+    from nanobot.utils.runtime import EMPTY_FINAL_RESPONSE_MESSAGE
+
+    provider = MagicMock(spec=LLMProvider)
+    provider.chat_with_retry = AsyncMock(return_value=LLMResponse(
+        content=None,
+        finish_reason=finish_reason,
+    ))
+    tools = MagicMock()
+    tools.get_definitions.return_value = []
+
+    result = await AgentRunner().run(make_run_spec(
+        provider,
+        initial_messages=[{"role": "user", "content": "do task"}],
+        tools=tools,
+        model="test-model",
+        max_iterations=3,
+        max_tool_result_chars=_MAX_TOOL_RESULT_CHARS,
+    ))
+
+    assert provider.chat_with_retry.await_count == 1
+    assert result.final_content == EMPTY_FINAL_RESPONSE_MESSAGE
+    assert result.stop_reason == "empty_final_response"
+
+
+@pytest.mark.asyncio
+@pytest.mark.parametrize("finish_reason", ["refusal", "content_filter"])
+async def test_runner_does_not_auto_continue_goal_after_policy_terminal(
+    finish_reason: str,
+) -> None:
+    from nanobot.agent.runner import AgentRunner
+
+    provider = MagicMock(spec=LLMProvider)
+    provider.chat_with_retry = AsyncMock(return_value=LLMResponse(
+        content="Request blocked by provider policy.",
+        finish_reason=finish_reason,
+    ))
+    tools = MagicMock()
+    tools.get_definitions.return_value = []
+
+    result = await AgentRunner().run(make_run_spec(
+        provider,
+        initial_messages=[{"role": "user", "content": "do task"}],
+        tools=tools,
+        model="test-model",
+        max_iterations=3,
+        max_tool_result_chars=_MAX_TOOL_RESULT_CHARS,
+        goal_active_predicate=lambda: True,
+    ))
+
+    assert provider.chat_with_retry.await_count == 1
+    assert result.final_content == "Request blocked by provider policy."
+    assert result.stop_reason == "completed"
+
+
 @pytest.mark.asyncio
 async def test_runner_uses_specific_message_after_empty_finalization_retry():
     """After silent retries + finalization all return empty, stop_reason is empty_final_response."""
@@ -450,6 +821,56 @@ async def test_runner_uses_specific_message_after_empty_finalization_retry():
     assert result.stop_reason == "empty_final_response"
 
 
+@pytest.mark.asyncio
+async def test_empty_finalization_retry_discards_candidate_provider_state():
+    from nanobot.agent.runner import AgentRunner
+
+    candidate = ProviderConversationState(
+        kind="openai_responses",
+        provider="openai:test",
+        model="test-model",
+        version=1,
+        payload={
+            "items": [{
+                "type": "function_call",
+                "call_id": "call_1",
+                "name": "exec",
+                "arguments": "{}",
+            }],
+        },
+    )
+    provider = MagicMock(spec=LLMProvider)
+    provider.can_resume_conversation_state.return_value = True
+    provider.chat_with_retry = AsyncMock(side_effect=[
+        LLMResponse(content=None, tool_calls=[], usage={}),
+        LLMResponse(content=None, tool_calls=[], usage={}),
+        LLMResponse(
+            content="finalized without tools",
+            tool_calls=[ToolCallRequest(id="call_1", name="exec", arguments={})],
+            finish_reason="stop",
+            provider_state=candidate,
+            usage={},
+        ),
+    ])
+    tools = MagicMock()
+    tools.get_definitions.return_value = []
+    tools.execute = AsyncMock(return_value="must not run")
+
+    runner = AgentRunner()
+    result = await runner.run(make_run_spec(
+        provider,
+        initial_messages=[{"role": "user", "content": "do task"}],
+        tools=tools,
+        model="test-model",
+        max_iterations=3,
+        max_tool_result_chars=_MAX_TOOL_RESULT_CHARS,
+    ))
+
+    tools.execute.assert_not_awaited()
+    assert result.final_content == "finalized without tools"
+    assert result.provider_state is None
+
+
 @pytest.mark.asyncio
 async def test_runner_length_recovery_returns_all_segments():
     """Recovered output segments are returned together instead of only the tail."""

tests/agent/test_runner_errors.py

@@ -310,3 +310,50 @@ async def test_runner_tool_error_preserves_tool_results_in_messages():
         i for i, m in enumerate(result.messages) if m.get("role") == "tool"
     ]
     assert all(ti > asst_tc_idx for ti in tool_indices)
+
+
+@pytest.mark.asyncio
+async def test_length_finish_with_blank_content_routes_to_length_recovery():
+    """Regression test for #5133.
+
+    A response with finish_reason='length' and blank content (e.g. the model
+    spent its whole output budget on a tool call whose closing tag was
+    truncated) must take the length-recovery path, not the empty-response
+    retry path. Retrying the same prompt cannot recover from output-budget
+    exhaustion.
+    """
+    from nanobot.agent.runner import AgentRunner
+    from nanobot.utils.runtime import LENGTH_RECOVERY_PROMPT
+
+    provider = MagicMock(spec=LLMProvider)
+    # First call: truncated (length) with blank content and a dropped tool call.
+    # Second call: normal completion so the loop can terminate.
+    provider.chat_with_retry = AsyncMock(side_effect=[
+        LLMResponse(
+            content="",
+            finish_reason="length",
+            tool_calls=[ToolCallRequest(id="call_1", name="exec", arguments={})],
+            usage={},
+        ),
+        LLMResponse(content="done", finish_reason="stop", tool_calls=[], usage={}),
+    ])
+    tools = MagicMock()
+    tools.get_definitions.return_value = []
+
+    runner = AgentRunner()
+    result = await runner.run(make_run_spec(provider,
+        initial_messages=[{"role": "user", "content": "do a long task"}],
+        tools=tools,
+        model="test-model",
+        max_iterations=5,
+        max_tool_result_chars=_MAX_TOOL_RESULT_CHARS,
+    ))
+
+    # The runner must have injected a length-recovery prompt and continued,
+    # rather than exhausting empty-response retries into a generic apology.
+    user_msgs = [m.get("content") or "" for m in result.messages if m.get("role") == "user"]
+    assert any(LENGTH_RECOVERY_PROMPT in c for c in user_msgs), (
+        "expected a length-recovery message to be appended for a "
+        "finish_reason='length' response with blank content"
+    )
+    assert result.final_content == "done"

tests/agent/test_runner_fallback.py

@@ -9,8 +9,15 @@ import pytest
 from loguru import logger
 
 from nanobot.config.schema import ModelPresetConfig
-from nanobot.providers.base import LLMProvider, LLMResponse
+from nanobot.providers.base import (
+    LLMProvider,
+    LLMResponse,
+    ProviderCallContext,
+    ProviderConversationState,
+)
+from nanobot.providers.conversation_state import ProviderConversationStateController
 from nanobot.providers.fallback_provider import FallbackProvider
+from nanobot.providers.openai_responses import resolve_compact_threshold
 
 
 def _make_response(
@@ -66,6 +73,9 @@ class _FakeProvider(LLMProvider):
         self._response = response or _make_response()
         self.chat_calls: list[dict[str, Any]] = []
         self.chat_stream_calls: list[dict[str, Any]] = []
+        self.context_calls: list[ProviderCallContext | None] = []
+        self.resumable = False
+        self.compact = False
 
     def get_default_model(self) -> str:
         return f"{self.name}/model"
@@ -81,6 +91,26 @@ class _FakeProvider(LLMProvider):
             await on_delta(self._response.content)
         return self._response
 
+    async def chat_with_context(
+        self,
+        provider_context: ProviderCallContext | None = None,
+        **kwargs: Any,
+    ) -> LLMResponse:
+        self.context_calls.append(provider_context)
+        return await self.chat(**kwargs)
+
+    def can_resume_conversation_state(
+        self,
+        state: ProviderConversationState,
+        model: str | None = None,
+    ) -> bool:
+        _ = state, model
+        return self.resumable
+
+    def supports_native_compaction(self, model: str | None = None) -> bool:
+        _ = model
+        return self.compact
+
 
 # -- config-level tests --
 
@@ -211,6 +241,8 @@ def test_provider_snapshot_uses_smallest_fallback_context_window() -> None:
         snapshot = build_provider_snapshot(config)
 
     assert snapshot.context_window_tokens == 64000
+    assert isinstance(snapshot.provider, FallbackProvider)
+    assert snapshot.provider._primary_context_window_tokens == 128000
 
 
 def test_inline_fallback_reasoning_effort_does_not_inherit_primary() -> None:
@@ -285,6 +317,257 @@ class TestFallbackOnPrimaryError:
         assert primary.chat_calls[0]["model"] == "primary-model"
         assert fallback.chat_calls[0]["model"] == "fallback-a"
 
+    @pytest.mark.asyncio
+    async def test_primary_compaction_uses_primary_context_window(self) -> None:
+        primary = _FakeProvider("primary", _make_response("primary ok"))
+        primary.compact = True
+        fb = FallbackProvider(
+            primary=primary,
+            fallback_presets=[
+                _fallback("small-chat", context_window_tokens=50_000),
+            ],
+            provider_factory=MagicMock(),
+            primary_context_window_tokens=200_000,
+        )
+
+        await fb.chat_with_context(
+            messages=[{"role": "user", "content": "hi"}],
+            model="gpt-5.6",
+            max_tokens=10_000,
+            provider_context=ProviderCallContext(context_window_tokens=50_000),
+        )
+
+        primary_context = primary.context_calls[0]
+        assert primary_context is not None
+        assert primary_context.context_window_tokens == 200_000
+        assert resolve_compact_threshold(
+            primary_context.context_window_tokens,
+            10_000,
+        ) == 180_000
+
+    @pytest.mark.asyncio
+    async def test_native_fallback_compaction_uses_its_own_context_window(self) -> None:
+        primary = _FakeProvider("primary", _error_response())
+        primary.compact = True
+        fallback = _FakeProvider("fallback", _make_response("fallback ok"))
+        fallback.compact = True
+        fb = FallbackProvider(
+            primary=primary,
+            fallback_presets=[
+                _fallback("fallback-a", context_window_tokens=120_000),
+            ],
+            provider_factory=MagicMock(return_value=fallback),
+            primary_context_window_tokens=200_000,
+        )
+
+        result = await fb.chat_with_context(
+            messages=[{"role": "user", "content": "hi"}],
+            model="gpt-5.6",
+            provider_context=ProviderCallContext(context_window_tokens=50_000),
+        )
+
+        assert result.content == "fallback ok"
+        assert primary.context_calls == [
+            ProviderCallContext(context_window_tokens=200_000)
+        ]
+        assert fallback.context_calls == [
+            ProviderCallContext(context_window_tokens=120_000)
+        ]
+
+    @pytest.mark.asyncio
+    async def test_native_fallback_gets_context_when_primary_does_not_use_it(self) -> None:
+        primary = _FakeProvider("primary", _error_response())
+        fallback = _FakeProvider("fallback", _make_response("fallback ok"))
+        fallback.compact = True
+        fb = FallbackProvider(
+            primary=primary,
+            fallback_presets=[
+                _fallback("fallback-a", context_window_tokens=120_000),
+            ],
+            provider_factory=MagicMock(return_value=fallback),
+            primary_context_window_tokens=200_000,
+        )
+        messages = [{"role": "user", "content": "hi"}]
+        controller = ProviderConversationStateController(
+            provider=fb,
+            model="primary-model",
+            messages=messages,
+        )
+        assert fb.supports_native_compaction("primary-model") is False
+        provider_context = controller.prepare_request(
+            messages,
+            context_window_tokens=50_000,
+        )
+
+        assert provider_context == ProviderCallContext(
+            context_window_tokens=50_000
+        )
+        result = await fb.chat_with_context(
+            messages=messages,
+            model="primary-model",
+            provider_context=provider_context,
+        )
+
+        assert result.content == "fallback ok"
+        assert primary.context_calls == [ProviderCallContext()]
+        assert fallback.context_calls == [
+            ProviderCallContext(context_window_tokens=120_000)
+        ]
+
+    @pytest.mark.asyncio
+    async def test_responses_chat_fallback_responses_rebuilds_state(self) -> None:
+        primary = _FakeProvider("primary", _error_response())
+        primary.resumable = True
+        primary.compact = True
+        fallback = _FakeProvider("fallback", _make_response("fallback ok"))
+        messages = [{"role": "user", "content": "hi"}]
+        state = ProviderConversationState(
+            kind="openai_responses",
+            provider="openai:test",
+            model="gpt-5.6",
+            version=1,
+            payload={"items": [{"type": "reasoning", "encrypted_content": "opaque"}]},
+            pending_messages=list(messages),
+        )
+        fb = FallbackProvider(
+            primary=primary,
+            fallback_presets=[_fallback("fallback-a")],
+            provider_factory=MagicMock(return_value=fallback),
+        )
+        controller = ProviderConversationStateController(
+            provider=fb,
+            model="gpt-5.6",
+            messages=messages,
+            state=state,
+        )
+        provider_context = controller.prepare_request(
+            messages,
+            context_window_tokens=200_000,
+        )
+        assert provider_context is not None
+
+        result = await fb.chat_with_context(
+            messages=messages,
+            model="gpt-5.6",
+            provider_context=provider_context,
+        )
+
+        assert result.content == "fallback ok"
+        assert primary.context_calls == [provider_context]
+        assert fallback.context_calls == [ProviderCallContext()]
+        assert fallback.chat_calls[0]["messages"] == messages
+
+        controller.observe_response(result, messages)
+        messages.append({"role": "assistant", "content": result.content})
+        assert controller.finish(messages) is None
+
+        recovered_state = ProviderConversationState(
+            kind="openai_responses",
+            provider="openai:test",
+            model="gpt-5.6",
+            version=1,
+            payload={"items": [{"type": "reasoning", "encrypted_content": "recovered"}]},
+        )
+        primary._response = LLMResponse(
+            content="primary recovered",
+            provider_state=recovered_state,
+        )
+        next_turn = ProviderConversationStateController(
+            provider=fb,
+            model="gpt-5.6",
+            messages=messages,
+        )
+        next_context = next_turn.prepare_request(
+            messages,
+            context_window_tokens=200_000,
+        )
+        assert next_context == ProviderCallContext(context_window_tokens=200_000)
+
+        recovered = await fb.chat_with_context(
+            messages=messages,
+            model="gpt-5.6",
+            provider_context=next_context,
+        )
+
+        assert recovered.provider_state is recovered_state
+        assert primary.context_calls[-1] == next_context
+        assert primary.chat_calls[-1]["messages"] == messages
+
+    @pytest.mark.asyncio
+    @pytest.mark.parametrize(
+        ("primary_error_kind", "primary_status", "primary_should_retry"),
+        [
+            ("server_error", 503, True),
+            ("authentication", 401, False),
+        ],
+        ids=["transient", "authentication"],
+    )
+    async def test_final_fallback_error_uses_primary_state_disposition(
+        self,
+        primary_error_kind: str,
+        primary_status: int,
+        primary_should_retry: bool,
+    ) -> None:
+        primary = _FakeProvider(
+            "primary",
+            _make_response(
+                "primary unavailable",
+                finish_reason="error",
+                error_kind=primary_error_kind,
+                error_status_code=primary_status,
+                error_should_retry=primary_should_retry,
+            ),
+        )
+        primary.resumable = True
+        fallback = _FakeProvider(
+            "fallback",
+            _make_response(
+                "fallback invalid request",
+                finish_reason="error",
+                error_kind="invalid_request",
+                error_status_code=400,
+                error_should_retry=False,
+            ),
+        )
+        messages = [{"role": "user", "content": "continue"}]
+        state = ProviderConversationState(
+            kind="openai_responses",
+            provider="openai:test",
+            model="gpt-5.6",
+            version=1,
+            payload={"items": [{"type": "reasoning", "encrypted_content": "opaque"}]},
+            pending_messages=list(messages),
+        )
+        provider = FallbackProvider(
+            primary=primary,
+            fallback_presets=[_fallback("fallback-a")],
+            provider_factory=MagicMock(return_value=fallback),
+        )
+        controller = ProviderConversationStateController(
+            provider=provider,
+            model="gpt-5.6",
+            messages=messages,
+            state=state,
+        )
+        provider_context = controller.prepare_request(
+            messages,
+            context_window_tokens=200_000,
+        )
+        assert provider_context is not None
+
+        response = await provider.chat_with_context(
+            messages=messages,
+            model="gpt-5.6",
+            provider_context=provider_context,
+        )
+        controller.observe_response(response, messages)
+
+        assert response.content == "fallback invalid request"
+        assert response.preserve_provider_state_on_error is True
+        restored = controller.finish(messages)
+        assert restored is not None
+        assert restored.payload == state.payload
+
     @pytest.mark.asyncio
     async def test_reports_the_fallback_model_before_its_request(self) -> None:
         primary = _FakeProvider("primary", _error_response())

tests/agent/test_runner_governance.py

@@ -15,7 +15,11 @@ from nanobot.agent.context_governance import (
 )
 from nanobot.agent.runner import AgentRunSpec
 from nanobot.config.schema import AgentDefaults
-from nanobot.providers.base import LLMResponse, ToolCallRequest
+from nanobot.providers.base import (
+    LLMResponse,
+    ProviderConversationState,
+    ToolCallRequest,
+)
 
 _MAX_TOOL_RESULT_CHARS = AgentDefaults().max_tool_result_chars
 
@@ -886,6 +890,13 @@ def test_drop_malformed_tool_calls_trims_response():
     """LLM response tool_calls with a missing/empty name are dropped in place."""
     from nanobot.agent.runner import AgentRunner
 
+    candidate_state = ProviderConversationState(
+        kind="openai_responses",
+        provider="openai:test",
+        model="gpt-5.6",
+        version=1,
+        payload={"items": [{"type": "function_call", "name": None}]},
+    )
     response = LLMResponse(
         content=None,
         tool_calls=[
@@ -895,9 +906,11 @@ def test_drop_malformed_tool_calls_trims_response():
             ToolCallRequest(id="4", name="read_file", arguments={}),
         ],
         finish_reason="tool_calls",
+        provider_state=candidate_state,
     )
     dropped, all_dropped, orig = AgentRunner._drop_malformed_tool_calls(response)
     assert [tc.name for tc in response.tool_calls] == ["read_file"]
+    assert response.provider_state is None
     assert response.finish_reason == "tool_calls"
     assert response.should_execute_tools is True
     assert dropped == 3

tests/agent/test_session_atomic.py

@@ -4,6 +4,7 @@ import json
 from datetime import datetime
 from pathlib import Path
 
+from nanobot.providers.base import ProviderConversationState
 from nanobot.session.manager import Session, SessionManager
 
 
@@ -101,6 +102,137 @@ class TestAtomicSave:
         for i in range(5):
             assert loaded.messages[i]["content"] == f"msg{i}"
 
+    def test_provider_state_round_trips_in_private_record_only(self, tmp_path: Path):
+        mgr = SessionManager(tmp_path)
+        secret = "encrypted-reasoning-blob"
+        session = Session(
+            key="test:provider-state",
+            provider_state=ProviderConversationState(
+                kind="openai_responses",
+                provider="openai:https://api.openai.com/v1",
+                model="gpt-5.6",
+                version=1,
+                payload={
+                    "items": [
+                        {
+                            "type": "reasoning",
+                            "encrypted_content": secret,
+                        }
+                    ]
+                },
+                pending_messages=[{"role": "user", "content": "continue"}],
+            ),
+        )
+        session.add_message("user", "hello")
+        mgr.save(session)
+
+        records = [
+            json.loads(line)
+            for line in mgr._get_session_path(session.key)
+            .read_text(encoding="utf-8")
+            .splitlines()
+        ]
+        assert [record.get("_type") for record in records] == [
+            "metadata",
+            "provider_state",
+            None,
+        ]
+        assert secret in records[1]["state"]["payload"]["items"][0]["encrypted_content"]
+
+        mgr.invalidate(session.key)
+        loaded = mgr.get_or_create(session.key)
+        assert loaded.provider_state is not None
+        assert loaded.provider_state.to_private_record() == session.provider_state.to_private_record()
+
+        public_payload = mgr.read_session_file(session.key)
+        assert public_payload is not None
+        assert public_payload["messages"] == [session.messages[0]]
+        assert secret not in json.dumps(public_payload)
+        assert secret not in json.dumps(mgr.list_sessions())
+
+    def test_provider_state_does_not_consume_list_preview_budget(
+        self,
+        tmp_path: Path,
+        monkeypatch,
+    ):
+        import nanobot.session.manager as session_manager
+
+        monkeypatch.setattr(session_manager, "_SESSION_LIST_PREVIEW_MAX_CHARS", 100)
+        mgr = SessionManager(tmp_path)
+        session = Session(
+            key="test:provider-state-preview",
+            provider_state=ProviderConversationState(
+                kind="openai_responses",
+                provider="openai:test",
+                model="test-model",
+                version=1,
+                payload={"items": [{"encrypted_content": "x" * 200}]},
+            ),
+        )
+        session.add_message("user", "visible preview")
+        mgr.save(session)
+
+        assert mgr.list_sessions()[0]["preview"] == "visible preview"
+
+    def test_clear_and_fork_discard_provider_state(self, tmp_path: Path):
+        mgr = SessionManager(tmp_path)
+        state = ProviderConversationState(
+            kind="openai_responses",
+            provider="openai:test",
+            model="gpt-5.6",
+            version=1,
+            payload={"items": []},
+        )
+        source = Session(key="test:state-source", provider_state=state)
+        source.add_message("user", "hello")
+        mgr.save(source)
+
+        fork = mgr.fork_session_before_user_index(
+            source.key,
+            "test:state-fork",
+            1,
+        )
+        assert fork is not None
+        assert fork.provider_state is None
+
+        source.clear()
+        assert source.provider_state is None
+
+    def test_invalid_provider_state_record_is_not_public_history(self, tmp_path: Path):
+        mgr = SessionManager(tmp_path)
+        path = mgr._get_session_path("test:bad-provider-state")
+        path.parent.mkdir(parents=True, exist_ok=True)
+        path.write_text(
+            "\n".join(
+                [
+                    json.dumps(
+                        {
+                            "_type": "metadata",
+                            "key": "test:bad-provider-state",
+                            "created_at": datetime.now().isoformat(),
+                            "updated_at": datetime.now().isoformat(),
+                            "metadata": {},
+                            "last_consolidated": 0,
+                        }
+                    ),
+                    json.dumps(
+                        {
+                            "_type": "provider_state",
+                            "state": {"kind": "openai_responses"},
+                        }
+                    ),
+                    json.dumps({"role": "user", "content": "safe"}),
+                ]
+            )
+            + "\n",
+            encoding="utf-8",
+        )
+
+        loaded = mgr._load("test:bad-provider-state")
+        assert loaded is not None
+        assert loaded.provider_state is None
+        assert loaded.messages == [{"role": "user", "content": "safe"}]
+
 
 class TestRepairCorruptFile:
     def _write_corrupt_jsonl(self, path: Path, lines: list[str]) -> None:

tests/agent/test_session_lock_lifecycle.py

@@ -0,0 +1,53 @@
+from __future__ import annotations
+
+import asyncio
+import gc
+from unittest.mock import MagicMock
+
+import pytest
+
+
+def _make_loop(loop_factory):
+    provider = MagicMock()
+    provider.get_default_model.return_value = "test-model"
+    return loop_factory(provider=provider)
+
+
+def test_idle_agent_session_locks_are_released(loop_factory):
+    loop = _make_loop(loop_factory)
+
+    for index in range(1000):
+        lock = loop._get_session_lock(f"api:temporary-{index}")
+
+    del lock
+    gc.collect()
+
+    assert len(loop._session_locks) == 0
+
+
+@pytest.mark.asyncio
+async def test_waiter_keeps_agent_session_lock_alive(loop_factory):
+    loop = _make_loop(loop_factory)
+    owner_lock = loop._get_session_lock("api:shared")
+    await owner_lock.acquire()
+    waiter_started = asyncio.Event()
+    waiter_entered = asyncio.Event()
+
+    async def wait_for_lock() -> None:
+        lock = loop._get_session_lock("api:shared")
+        waiter_started.set()
+        async with lock:
+            waiter_entered.set()
+
+    waiter = asyncio.create_task(wait_for_lock())
+    await waiter_started.wait()
+
+    assert loop._get_session_lock("api:shared") is owner_lock
+    assert not waiter_entered.is_set()
+
+    owner_lock.release()
+    await waiter
+    del owner_lock
+    gc.collect()
+
+    assert "api:shared" not in loop._session_locks

tests/agent/test_session_manager_history.py

@@ -1,3 +1,4 @@
+from nanobot.providers.base import ProviderConversationState
 from nanobot.runtime_context import (
     RUNTIME_CONTEXT_HISTORY_META,
     RuntimeContextBlock,
@@ -769,7 +770,16 @@ def test_get_history_extend_to_user_keeps_newer_user_inside_window():
 
 def test_retain_recent_legal_suffix_returns_dropped_messages():
     """retain_recent_legal_suffix returns the actually-dropped messages."""
-    session = Session(key="test:return-dropped")
+    session = Session(
+        key="test:return-dropped",
+        provider_state=ProviderConversationState(
+            kind="openai_responses",
+            provider="openai:test",
+            model="test-model",
+            version=1,
+            payload={"items": []},
+        ),
+    )
     for i in range(10):
         session.messages.append({"role": "user", "content": f"msg{i}"})
 
@@ -779,11 +789,19 @@ def test_retain_recent_legal_suffix_returns_dropped_messages():
     assert [m["content"] for m in result.dropped] == [f"msg{i}" for i in range(6)]
     assert len(session.messages) == 4
     assert result.already_consolidated_count == 0
+    assert session.provider_state is None
 
 
 def test_retain_recent_legal_suffix_returns_empty_when_no_drop():
     """No messages dropped → empty list returned."""
-    session = Session(key="test:no-drop")
+    state = ProviderConversationState(
+        kind="openai_responses",
+        provider="openai:test",
+        model="test-model",
+        version=1,
+        payload={"items": []},
+    )
+    session = Session(key="test:no-drop", provider_state=state)
     for i in range(3):
         session.messages.append({"role": "user", "content": f"msg{i}"})
 
@@ -792,6 +810,7 @@ def test_retain_recent_legal_suffix_returns_empty_when_no_drop():
     assert result.dropped == []
     assert result.already_consolidated_count == 0
     assert len(session.messages) == 3
+    assert session.provider_state is state
 
 
 def test_retain_recent_legal_suffix_returns_all_on_zero():

tests/agent/tools/test_subagent_tools.py

@@ -504,6 +504,7 @@ async def test_drain_pending_blocks_while_subagents_running(tmp_path):
             usage={},
             had_injections=False,
             tools_used=[],
+            provider_state=None,
         )
 
     loop.runner.run = AsyncMock(side_effect=fake_runner_run)
@@ -589,6 +590,7 @@ async def test_drain_pending_no_block_when_no_subagents(tmp_path):
             usage={},
             had_injections=False,
             tools_used=[],
+            provider_state=None,
         )
 
     loop.runner.run = AsyncMock(side_effect=fake_runner_run)
@@ -638,6 +640,7 @@ async def test_drain_pending_timeout(tmp_path):
             usage={},
             had_injections=False,
             tools_used=[],
+            provider_state=None,
         )
 
     loop.runner.run = AsyncMock(side_effect=fake_runner_run)

tests/channels/test_base_channel.py

@@ -83,6 +83,49 @@ async def test_handle_message_dm_sends_pairing_code(monkeypatch) -> None:
     assert msg.metadata.get("_pairing_code") == "ABCD-EFGH"
 
 
+@pytest.mark.asyncio
+async def test_dm_during_transient_store_failure_keeps_approvals(
+    tmp_path, monkeypatch
+) -> None:
+    """An unapproved DM while pairing.json is unreadable must not wipe approvals.
+
+    The pairing store treated a transient OSError like corruption and returned
+    an empty store; the DM pairing path then persisted that empty view,
+    erasing every approved sender.
+    """
+    import builtins
+    from pathlib import Path
+
+    from nanobot.pairing import store
+
+    path = tmp_path / "pairing.json"
+    monkeypatch.setattr(store, "_store_path", lambda: path)
+    code = store.generate_code("dummy", "friend")
+    store.approve_code(code)
+
+    channel = _DummyChannel({"allowFrom": []}, MessageBus())
+
+    real_open = builtins.open
+
+    def flaky_open(file, mode="r", *args, **kwargs):
+        try:
+            same = Path(file) == path
+        except TypeError:
+            same = False
+        if same and "r" in mode and "+" not in mode:
+            raise PermissionError(13, "temporarily locked", str(path))
+        return real_open(file, mode, *args, **kwargs)
+
+    with monkeypatch.context() as m:
+        m.setattr(builtins, "open", flaky_open)
+        await channel._handle_message(
+            sender_id="stranger", chat_id="chat1", content="hello", is_dm=True
+        )
+
+    assert channel._sent == []
+    assert store.is_approved("dummy", "friend") is True
+
+
 @pytest.mark.asyncio
 async def test_handle_message_group_ignores_unknown() -> None:
     channel = _DummyChannel({"allowFrom": []}, MessageBus())

tests/pairing/test_store.py

@@ -323,3 +323,66 @@ def test_pending_gc_drops_malformed_entries(tmp_path, monkeypatch):
     )
     monkeypatch.setattr(store, "_store_path", lambda: path)
     assert store.list_pending() == []
+
+
+def _fail_reads_of(monkeypatch, path):
+    """Make reads of *path* raise like a transiently locked/busy file."""
+    import builtins
+    from pathlib import Path
+
+    real_open = builtins.open
+
+    def flaky_open(file, mode="r", *args, **kwargs):
+        try:
+            same = Path(file) == path
+        except TypeError:
+            same = False
+        if same and "r" in mode and "+" not in mode:
+            raise PermissionError(13, "temporarily locked", str(path))
+        return real_open(file, mode, *args, **kwargs)
+
+    monkeypatch.setattr(builtins, "open", flaky_open)
+
+
+class TestTransientReadFailure:
+    """A transient I/O failure is not corruption and must never wipe the store."""
+
+    def test_generate_code_does_not_wipe_approvals(self, tmp_path, monkeypatch):
+        """An unapproved DM during a read blip previously erased every approval.
+
+        _load treated OSError like corruption and returned an empty store;
+        generate_code then unconditionally saved it, overwriting pairing.json
+        with no approved senders.
+        """
+        code = store.generate_code("telegram", "123")
+        store.approve_code(code)
+
+        with monkeypatch.context() as m:
+            _fail_reads_of(m, store._store_path())
+            with pytest.raises(OSError):
+                store.generate_code("telegram", "stranger")
+
+        assert store.is_approved("telegram", "123") is True
+
+    def test_reads_fail_closed_without_crashing(self, tmp_path, monkeypatch):
+        code = store.generate_code("telegram", "123")
+        store.approve_code(code)
+
+        with monkeypatch.context() as m:
+            _fail_reads_of(m, store._store_path())
+            assert store.is_approved("telegram", "123") is False
+            assert store.list_pending() == []
+            assert store.get_approved("telegram") == []
+
+        assert store.is_approved("telegram", "123") is True
+
+    def test_approve_command_reports_store_unavailable(self, tmp_path, monkeypatch):
+        """/pairing approve must fail loudly instead of claiming the code is invalid."""
+        code = store.generate_code("telegram", "123")
+
+        with monkeypatch.context() as m:
+            _fail_reads_of(m, store._store_path())
+            reply = store.handle_pairing_command("telegram", f"approve {code}")
+
+        assert "unavailable" in reply.lower()
+        assert store.approve_code(code) == ("telegram", "123")

tests/providers/test_azure_openai_provider.py

@@ -11,7 +11,7 @@ from nanobot.providers.azure_openai_provider import (
     AzureOpenAIProvider,
     _AzureTokenProvider,
 )
-from nanobot.providers.base import LLMResponse
+from nanobot.providers.base import LLMResponse, ProviderCallContext
 
 # ---------------------------------------------------------------------------
 # Init & validation
@@ -234,6 +234,7 @@ def test_build_body_basic():
     assert body["max_output_tokens"] == 4096
     assert body["store"] is False
     assert "reasoning" not in body
+    assert "include" not in body
     # input should contain the converted user message only (system extracted)
     assert any(
         item.get("role") == "user"
@@ -241,6 +242,30 @@ def test_build_body_basic():
     )
 
 
+def test_build_body_enables_server_compaction():
+    provider = AzureOpenAIProvider(
+        api_key="k",
+        api_base="https://res.openai.azure.com",
+        default_model="gpt-5.6",
+    )
+
+    body = provider._build_body(
+        [{"role": "user", "content": "hello"}],
+        None,
+        None,
+        10_000,
+        0.1,
+        "high",
+        None,
+        provider_context=ProviderCallContext(context_window_tokens=200_000),
+    )
+
+    assert body["context_management"] == [{
+        "type": "compaction",
+        "compact_threshold": 180_000,
+    }]
+
+
 def test_build_body_max_tokens_minimum():
     """max_output_tokens should never be less than 1."""
     provider = AzureOpenAIProvider(api_key="k", api_base="https://r.com", default_model="gpt-4o")
@@ -358,6 +383,38 @@ async def test_chat_success():
     assert result.usage["prompt_tokens"] == 10
 
 
+@pytest.mark.asyncio
+async def test_chat_retries_without_unsupported_server_compaction():
+    provider = AzureOpenAIProvider(
+        api_key="test-key",
+        api_base="https://test.openai.azure.com",
+        default_model="gpt-5.6",
+    )
+
+    class UnsupportedCompactionError(Exception):
+        status_code = 400
+        body = {"error": {"message": "Unknown parameter: context_management"}}
+
+    provider._client.responses = MagicMock()
+    provider._client.responses.create = AsyncMock(side_effect=[
+        UnsupportedCompactionError(),
+        _make_sdk_response(content="compaction fallback"),
+    ])
+
+    result = await provider.chat(
+        [{"role": "user", "content": "Hi"}],
+        provider_context=ProviderCallContext(context_window_tokens=200_000),
+    )
+
+    create = provider._client.responses.create
+    assert result.content == "compaction fallback"
+    assert result.provider_state is not None
+    assert create.await_count == 2
+    assert "context_management" in create.call_args_list[0].kwargs
+    assert "context_management" not in create.call_args_list[1].kwargs
+    assert provider.supports_native_compaction() is False
+
+
 @pytest.mark.asyncio
 async def test_chat_uses_default_model():
     provider = AzureOpenAIProvider(
@@ -411,6 +468,7 @@ async def test_chat_with_tool_calls():
     assert len(result.tool_calls) == 1
     assert result.tool_calls[0].name == "get_weather"
     assert result.tool_calls[0].arguments == {"location": "SF"}
+    assert result.provider_state is not None
 
 
 @pytest.mark.asyncio
@@ -510,6 +568,7 @@ async def test_chat_stream_with_tool_calls():
     item_done.name = "get_weather"
     ev_item_done = MagicMock(type="response.output_item.done", item=item_done)
     resp_obj = MagicMock(status="completed")
+    resp_obj.model_dump.return_value = {"status": "completed", "output": []}
     ev_completed = MagicMock(type="response.completed", response=resp_obj)
 
     async def mock_stream():
@@ -527,6 +586,7 @@ async def test_chat_stream_with_tool_calls():
     assert len(result.tool_calls) == 1
     assert result.tool_calls[0].name == "get_weather"
     assert result.tool_calls[0].arguments == {"location": "SF"}
+    assert result.provider_state is not None
 
 
 @pytest.mark.asyncio

tests/providers/test_conversation_state.py

@@ -0,0 +1,291 @@
+"""Tests for provider-owned conversation-state lifecycle coordination."""
+
+from __future__ import annotations
+
+from unittest.mock import MagicMock
+
+import pytest
+
+from nanobot.providers.base import (
+    LLMProvider,
+    LLMResponse,
+    ProviderConversationState,
+    ToolCallRequest,
+)
+from nanobot.providers.conversation_state import (
+    ProviderConversationStateController,
+    allows_conversation_message_merge,
+)
+
+
+def _provider(*, resumable: bool = True, compact: bool = False) -> MagicMock:
+    provider = MagicMock(spec=LLMProvider)
+    provider.can_resume_conversation_state.return_value = resumable
+    provider.supports_native_compaction.return_value = compact
+    return provider
+
+
+def _state(label: str, *, pending: list[dict] | None = None) -> ProviderConversationState:
+    return ProviderConversationState(
+        kind="openai_responses",
+        provider="openai:test",
+        model="gpt-5.6",
+        version=1,
+        payload={"items": [{"type": "reasoning", "encrypted_content": label}]},
+        pending_messages=pending or [],
+    )
+
+
+def test_controller_replays_only_messages_after_provider_output() -> None:
+    provider = _provider()
+    messages = [
+        {"role": "system", "content": "system"},
+        {"role": "user", "content": "run a tool"},
+    ]
+    controller = ProviderConversationStateController(
+        provider=provider,
+        model="gpt-5.6",
+        messages=messages,
+    )
+    state = _state("first")
+
+    controller.prepare_request(messages, context_window_tokens=200_000)
+    response = LLMResponse(content=None, provider_state=state)
+    controller.observe_response(response, messages)
+    assert allows_conversation_message_merge(messages[-1]) is False
+
+    messages.append(controller.project_response_message(
+        {
+            "role": "assistant",
+            "content": None,
+            "tool_calls": [{"id": "call_1", "type": "function"}],
+        },
+        response,
+    ))
+    tool_message = {
+        "role": "tool",
+        "tool_call_id": "call_1",
+        "content": "tool result",
+    }
+    messages.append(tool_message)
+
+    provider_context = controller.prepare_request(
+        messages,
+        context_window_tokens=200_000,
+    )
+
+    assert provider_context is not None
+    assert provider_context.conversation_state is not None
+    assert provider_context.conversation_state.payload == state.payload
+    assert provider_context.conversation_state.pending_messages == [tool_message]
+    assert controller.checkpoint(messages).pending_messages == [tool_message]
+
+
+def test_controller_uses_governed_messages_for_provider_state_delta() -> None:
+    provider = _provider()
+    messages = [
+        {"role": "user", "content": "run a tool"},
+    ]
+    controller = ProviderConversationStateController(
+        provider=provider,
+        model="gpt-5.6",
+        messages=messages,
+    )
+    state = _state("first")
+
+    controller.prepare_request(messages, context_window_tokens=200_000)
+    response = LLMResponse(content=None, provider_state=state)
+    controller.observe_response(response, messages)
+    messages.extend([
+        controller.project_response_message(
+            {
+                "role": "assistant",
+                "content": None,
+                "tool_calls": [{"id": "call_1", "type": "function"}],
+            },
+            response,
+        ),
+        {
+            "role": "tool",
+            "tool_call_id": "call_1",
+            "content": "raw oversized result",
+        },
+    ])
+    governed_messages = [
+        messages[0],
+        messages[1],
+        {
+            "role": "tool",
+            "tool_call_id": "call_1",
+            "content": "compacted result",
+        },
+    ]
+
+    provider_context = controller.prepare_request(
+        messages,
+        context_window_tokens=200_000,
+        model_messages=governed_messages,
+    )
+
+    assert provider_context is not None
+    assert provider_context.conversation_state is not None
+    assert provider_context.conversation_state.pending_messages == [{
+        "role": "tool",
+        "tool_call_id": "call_1",
+        "content": "compacted result",
+    }]
+    assert controller.checkpoint(messages).pending_messages[-1]["content"] == (
+        "raw oversized result"
+    )
+    governed_checkpoint = controller.checkpoint(
+        messages,
+        model_messages=governed_messages,
+    )
+    assert governed_checkpoint is not None
+    assert governed_checkpoint.pending_messages[-1]["content"] == "compacted result"
+
+
+def test_transient_response_preserves_only_durable_request_messages() -> None:
+    provider = _provider()
+    current_message = {"role": "user", "content": "continue"}
+    supplemental = {"role": "user", "content": "internal finalization retry"}
+    messages = [{"role": "system", "content": "system"}, current_message]
+    controller = ProviderConversationStateController(
+        provider=provider,
+        model="gpt-5.6",
+        messages=messages,
+        state=_state("saved", pending=[
+            {"role": "tool", "content": "prior"},
+            current_message,
+        ]),
+    )
+
+    provider_context = controller.prepare_request(
+        messages,
+        context_window_tokens=200_000,
+        supplemental_messages=[supplemental],
+    )
+    assert provider_context is not None
+    assert provider_context.conversation_state is not None
+    assert provider_context.conversation_state.pending_messages == [
+        {"role": "tool", "content": "prior"},
+        current_message,
+        supplemental,
+    ]
+
+    controller.observe_response(
+        LLMResponse(
+            content="temporary failure",
+            finish_reason="error",
+            error_kind="timeout",
+        ),
+        messages,
+    )
+    placeholder = {"role": "assistant", "content": "model error"}
+    messages.append(placeholder)
+
+    state = controller.finish(messages)
+    assert state is not None
+    assert state.pending_messages == [
+        {"role": "tool", "content": "prior"},
+        current_message,
+        placeholder,
+    ]
+
+
+def test_non_retryable_response_discards_saved_state() -> None:
+    provider = _provider()
+    messages = [{"role": "user", "content": "continue"}]
+    controller = ProviderConversationStateController(
+        provider=provider,
+        model="gpt-5.6",
+        messages=messages,
+        state=_state("saved"),
+    )
+
+    controller.prepare_request(messages, context_window_tokens=200_000)
+    controller.observe_response(
+        LLMResponse(
+            content="invalid request",
+            finish_reason="error",
+            error_status_code=400,
+            error_should_retry=False,
+        ),
+        messages,
+    )
+
+    assert controller.finish(messages) is None
+
+
+@pytest.mark.parametrize(
+    ("finish_reason", "exposes_tool_call"),
+    [
+        ("length", False),
+        ("length", True),
+        ("refusal", True),
+        ("content_filter", True),
+    ],
+)
+def test_terminal_response_discards_candidate_state(
+    finish_reason: str,
+    exposes_tool_call: bool,
+) -> None:
+    provider = _provider()
+    messages = [{"role": "user", "content": "continue"}]
+    controller = ProviderConversationStateController(
+        provider=provider,
+        model="gpt-5.6",
+        messages=messages,
+        state=_state("saved"),
+    )
+
+    controller.prepare_request(messages, context_window_tokens=200_000)
+    candidate = ProviderConversationState(
+        kind="openai_responses",
+        provider="openai:test",
+        model="gpt-5.6",
+        version=1,
+        payload={
+            "items": [{
+                "type": "function_call",
+                "call_id": "call_1",
+                "name": "exec",
+                "arguments": "{}",
+            }],
+        },
+    )
+    response = LLMResponse(
+        content="terminal response",
+        tool_calls=(
+            [ToolCallRequest(id="call_1", name="exec", arguments={})]
+            if exposes_tool_call
+            else []
+        ),
+        finish_reason=finish_reason,
+        provider_state=candidate,
+    )
+    assert response.has_tool_calls is exposes_tool_call
+    assert response.should_execute_tools is False
+
+    controller.observe_response(response, messages)
+
+    assert controller.finish(messages) is None
+
+
+def test_independent_request_exposes_context_without_capability_check() -> None:
+    provider = _provider(compact=False)
+    messages = [{"role": "user", "content": "hello"}]
+    controller = ProviderConversationStateController(
+        provider=provider,
+        model="gpt-5.6",
+        messages=messages,
+        state=_state("saved"),
+    )
+
+    provider_context = controller.independent_request_context(
+        context_window_tokens=200_000,
+    )
+    assert provider_context is not None
+    assert provider_context.conversation_state is None
+    assert provider_context.context_window_tokens == 200_000
+    provider.supports_native_compaction.assert_not_called()

tests/providers/test_github_copilot_routing.py

@@ -10,6 +10,7 @@ from unittest.mock import AsyncMock, MagicMock, patch
 
 import pytest
 
+from nanobot.providers.base import ProviderCallContext
 from nanobot.providers.openai_compat_provider import OpenAICompatProvider
 from nanobot.providers.registry import find_by_name
 
@@ -44,8 +45,10 @@ def test_build_responses_body_strips_github_copilot_prefix():
         temperature=0.1,
         reasoning_effort=None,
         tool_choice=None,
+        provider_context=ProviderCallContext(context_window_tokens=128_000),
     )
     assert body["model"] == "gpt-5.4-mini"
+    assert "context_management" not in body
 
 
 @pytest.mark.asyncio

tests/providers/test_litellm_kwargs.py

@@ -14,6 +14,7 @@ from unittest.mock import AsyncMock, MagicMock, patch
 
 import pytest
 
+from nanobot.providers.base import ProviderCallContext
 from nanobot.providers.openai_compat_provider import OpenAICompatProvider
 from nanobot.providers.registry import find_by_name
 
@@ -679,6 +680,7 @@ async def test_direct_openai_gpt5_uses_responses_api() -> None:
     assert call_kwargs["max_output_tokens"] == 4096
     assert "input" in call_kwargs
     assert "messages" not in call_kwargs
+    assert call_kwargs["include"] == ["reasoning.encrypted_content"]
 
 
 @pytest.mark.asyncio
@@ -710,6 +712,40 @@ async def test_direct_openai_reasoning_prefers_responses_api() -> None:
     assert call_kwargs["include"] == ["reasoning.encrypted_content"]
 
 
+@pytest.mark.asyncio
+async def test_direct_openai_retries_without_unsupported_server_compaction() -> None:
+    mock_chat = AsyncMock(return_value=_fake_chat_response())
+    mock_responses = AsyncMock(side_effect=[
+        _FakeResponsesError(400, "Unknown parameter: context_management"),
+        _fake_responses_response("compaction fallback"),
+    ])
+    spec = find_by_name("openai")
+
+    with patch("nanobot.providers.openai_compat_provider.AsyncOpenAI") as mock_client_class:
+        client_instance = mock_client_class.return_value
+        client_instance.chat.completions.create = mock_chat
+        client_instance.responses.create = mock_responses
+        provider = OpenAICompatProvider(
+            api_key="sk-test-key",
+            default_model="gpt-5.6",
+            spec=spec,
+        )
+
+        result = await provider.chat_with_context(
+            messages=[{"role": "user", "content": "hello"}],
+            model="gpt-5.6",
+            provider_context=ProviderCallContext(context_window_tokens=200_000),
+        )
+
+    assert result.content == "compaction fallback"
+    assert result.provider_state is not None
+    assert mock_responses.await_count == 2
+    assert "context_management" in mock_responses.call_args_list[0].kwargs
+    assert "context_management" not in mock_responses.call_args_list[1].kwargs
+    assert provider.supports_native_compaction("gpt-5.6") is False
+    mock_chat.assert_not_awaited()
+
+
 @pytest.mark.asyncio
 async def test_direct_openai_gpt4o_stays_on_chat_completions() -> None:
     mock_chat = AsyncMock(return_value=_fake_chat_response())

tests/providers/test_openai_codex_provider.py

@@ -20,6 +20,7 @@ from nanobot.providers.openai_codex_provider import (
     _request_codex,
     _should_retry_status,
 )
+from nanobot.providers.openai_responses import build_responses_state
 from nanobot.providers.registry import find_by_name
 
 
@@ -115,6 +116,48 @@ async def test_codex_request_non_200_populates_http_metadata(monkeypatch) -> Non
     assert error.should_retry is True
 
 
+@pytest.mark.asyncio
+async def test_codex_request_marks_rejected_compaction_without_retaining_raw_body(
+    monkeypatch,
+) -> None:
+    original_client = httpx.AsyncClient
+    secret = "PRIVATE PROMPT MUST NOT BE RETAINED"
+
+    def handler(request: httpx.Request) -> httpx.Response:
+        return httpx.Response(
+            400,
+            json={
+                "error": {
+                    "message": f"Unknown input type compaction_trigger; {secret}",
+                },
+            },
+            request=request,
+        )
+
+    def fake_client(
+        *,
+        timeout: int,
+        verify: bool,
+        **_kwargs: object,
+    ) -> httpx.AsyncClient:
+        return original_client(transport=httpx.MockTransport(handler), timeout=timeout)
+
+    monkeypatch.setattr("nanobot.providers.openai_codex_provider.httpx.AsyncClient", fake_client)
+
+    with pytest.raises(_CodexHTTPError) as caught:
+        await _request_codex(
+            "https://codex.example/responses",
+            {},
+            {"input": [{"type": "compaction_trigger"}]},
+            verify=True,
+        )
+
+    error = caught.value
+    assert error.compaction_unsupported is True
+    assert secret not in str(error)
+    assert not hasattr(error, "body")
+
+
 @pytest.mark.asyncio
 async def test_codex_request_honors_stream_idle_timeout_env(monkeypatch) -> None:
     """NANOBOT_STREAM_IDLE_TIMEOUT_S overrides the default Codex stream timeout."""
@@ -192,7 +235,7 @@ async def test_codex_prompt_cache_key_uses_stable_conversation_prefix(monkeypatc
     ):
         _ = proxy, on_thinking_delta, on_tool_call_delta
         bodies.append(body)
-        return "ok", [], "stop", {}, None
+        return provider_base.LLMResponse(content="ok")
 
     monkeypatch.setattr("nanobot.providers.openai_codex_provider._request_codex", fake_request)
 
@@ -232,7 +275,7 @@ async def test_codex_provider_applies_extra_body_from_config(monkeypatch) -> Non
 
     async def fake_request(_url, _headers, body, **_kwargs):
         bodies.append(body)
-        return "ok", [], "stop", {}, None
+        return provider_base.LLMResponse(content="ok")
 
     monkeypatch.setattr("nanobot.providers.openai_codex_provider._request_codex", fake_request)
     config = Config.model_validate({
@@ -297,7 +340,7 @@ async def test_codex_provider_passes_proxy_to_oauth_and_response_request(monkeyp
     ):
         _ = url, headers, body, verify, on_content_delta, on_thinking_delta, on_tool_call_delta
         seen["request_proxy"] = proxy
-        return "ok", [], "stop", {}, None
+        return provider_base.LLMResponse(content="ok")
 
     monkeypatch.setattr("nanobot.providers.openai_codex_provider.get_codex_token", fake_token)
     monkeypatch.setattr("nanobot.providers.openai_codex_provider._request_codex", fake_request)
@@ -384,7 +427,7 @@ async def test_codex_retry_uses_structured_timeout_metadata(monkeypatch) -> None
         calls += 1
         if calls == 1:
             raise httpx.ReadTimeout("")
-        return "ok", [], "stop", {}, None
+        return provider_base.LLMResponse(content="ok")
 
     async def fake_sleep(delay: float) -> None:
         delays.append(delay)
@@ -533,6 +576,254 @@ def test_codex_reasoning_options_request_summary_without_forcing_effort() -> Non
     assert _build_reasoning_options("none") == {"effort": "none"}
 
 
+@pytest.mark.asyncio
+async def test_codex_replayed_tool_turn_omits_server_item_ids(monkeypatch) -> None:
+    _mock_codex_token(monkeypatch)
+    provider = OpenAICodexProvider(default_model="openai-codex/gpt-5.6-sol")
+    state = build_responses_state(
+        provider=provider._responses_state_provider(),
+        model="gpt-5.6-sol",
+        input_items=[{
+            "id": "msg_user",
+            "type": "message",
+            "role": "user",
+            "content": [{"type": "input_text", "text": "Check the weather"}],
+        }],
+        output_items=[
+            {
+                "id": "rs_reasoning",
+                "type": "reasoning",
+                "encrypted_content": "opaque reasoning",
+                "summary": [],
+            },
+            {
+                "id": "fc_read",
+                "type": "function_call",
+                "call_id": "call_read",
+                "name": "read_file",
+                "arguments": '{"path":"weather/SKILL.md"}',
+                "status": "completed",
+            },
+        ],
+    )
+    bodies: list[dict[str, Any]] = []
+
+    async def fake_request(
+        url,
+        headers,
+        body,
+        verify,
+        proxy=None,
+        on_content_delta=None,
+        on_thinking_delta=None,
+        on_tool_call_delta=None,
+    ):
+        bodies.append(body)
+        return provider_base.LLMResponse(content="done")
+
+    monkeypatch.setattr(
+        "nanobot.providers.openai_codex_provider._request_codex",
+        fake_request,
+    )
+
+    response = await provider.chat(
+        [{"role": "user", "content": "Check the weather"}],
+        provider_context=provider_base.ProviderCallContext(
+            conversation_state=state.with_pending_messages([{
+                "role": "tool",
+                "tool_call_id": "call_read|fc_read",
+                "content": "weather skill contents",
+            }]),
+        ),
+    )
+
+    assert response.content == "done"
+    assert len(bodies) == 1
+    input_items = bodies[0]["input"]
+    assert [item.get("type") for item in input_items] == [
+        "message",
+        "reasoning",
+        "function_call",
+        "function_call_output",
+    ]
+    assert all("id" not in item for item in input_items)
+    assert input_items[1]["encrypted_content"] == "opaque reasoning"
+    assert input_items[2]["call_id"] == "call_read"
+    assert input_items[3]["call_id"] == "call_read"
+
+
+@pytest.mark.asyncio
+async def test_codex_compacts_state_at_ninety_percent_before_next_request(
+    monkeypatch,
+) -> None:
+    _mock_codex_token(monkeypatch)
+    provider = OpenAICodexProvider(default_model="openai-codex/gpt-5.6-sol")
+    state_provider = provider._responses_state_provider()
+    state = build_responses_state(
+        provider=state_provider,
+        model="gpt-5.6-sol",
+        input_items=[{"type": "message", "role": "user", "content": "old question"}],
+        output_items=[
+            {"type": "reasoning", "encrypted_content": "old opaque reasoning"},
+            {
+                "type": "message",
+                "role": "assistant",
+                "content": [{"type": "output_text", "text": "old answer"}],
+            },
+        ],
+        usage={
+            "prompt_tokens": 90,
+            "completion_tokens": 5,
+            "total_tokens": 95,
+        },
+    )
+    bodies: list[dict[str, Any]] = []
+
+    async def fake_request(
+        url,
+        headers,
+        body,
+        verify,
+        proxy=None,
+        on_content_delta=None,
+        on_thinking_delta=None,
+        on_tool_call_delta=None,
+    ):
+        _ = (
+            url,
+            headers,
+            verify,
+            proxy,
+            on_content_delta,
+            on_thinking_delta,
+            on_tool_call_delta,
+        )
+        bodies.append(body)
+        if body["input"][-1].get("type") == "compaction_trigger":
+            compact_item = {
+                "type": "compaction",
+                "encrypted_content": "compacted opaque state",
+            }
+            return provider_base.LLMResponse(
+                content=None,
+                provider_state=build_responses_state(
+                    provider=state_provider,
+                    model="gpt-5.6-sol",
+                    input_items=body["input"],
+                    output_items=[compact_item],
+                    usage={
+                        "prompt_tokens": 95,
+                        "completion_tokens": 2,
+                        "total_tokens": 97,
+                    },
+                ),
+            )
+        return provider_base.LLMResponse(content="done")
+
+    monkeypatch.setattr(
+        "nanobot.providers.openai_codex_provider._request_codex",
+        fake_request,
+    )
+
+    response = await provider.chat_with_retry(
+        [
+            {"role": "system", "content": "system"},
+            {"role": "user", "content": "new question"},
+        ],
+        max_tokens=5,
+        provider_context=provider_base.ProviderCallContext(
+            conversation_state=state.with_pending_messages([
+                {"role": "user", "content": "new question"},
+            ]),
+            context_window_tokens=100,
+        ),
+    )
+
+    assert response.content == "done"
+    assert len(bodies) == 2
+    assert bodies[0]["input"][-1] == {"type": "compaction_trigger"}
+    assert bodies[1]["input"][-1] == {
+        "type": "compaction",
+        "encrypted_content": "compacted opaque state",
+    }
+    assert not any(
+        item.get("type") == "reasoning"
+        for item in bodies[1]["input"]
+    )
+    assert any(
+        item.get("role") == "user"
+        and "new question" in str(item.get("content"))
+        for item in bodies[1]["input"]
+    )
+
+
+@pytest.mark.asyncio
+async def test_codex_disables_unsupported_native_compaction_and_continues(
+    monkeypatch,
+) -> None:
+    _mock_codex_token(monkeypatch)
+    provider = OpenAICodexProvider(default_model="openai-codex/gpt-5.6-sol")
+    state_provider = provider._responses_state_provider()
+    state = build_responses_state(
+        provider=state_provider,
+        model="gpt-5.6-sol",
+        input_items=[{"type": "message", "role": "user", "content": "old"}],
+        output_items=[{"type": "reasoning", "encrypted_content": "opaque"}],
+        usage={"prompt_tokens": 90, "completion_tokens": 5, "total_tokens": 95},
+    )
+    bodies: list[dict[str, Any]] = []
+
+    async def fake_request(
+        url,
+        headers,
+        body,
+        verify,
+        proxy=None,
+        on_content_delta=None,
+        on_thinking_delta=None,
+        on_tool_call_delta=None,
+    ):
+        _ = (
+            url,
+            headers,
+            verify,
+            proxy,
+            on_content_delta,
+            on_thinking_delta,
+            on_tool_call_delta,
+        )
+        bodies.append(body)
+        if body["input"][-1].get("type") == "compaction_trigger":
+            raise _CodexHTTPError(
+                "HTTP 400: Codex API request failed",
+                status_code=400,
+                compaction_unsupported=True,
+            )
+        return provider_base.LLMResponse(content="done")
+
+    monkeypatch.setattr(
+        "nanobot.providers.openai_codex_provider._request_codex",
+        fake_request,
+    )
+
+    response = await provider.chat(
+        [{"role": "user", "content": "new"}],
+        max_tokens=5,
+        provider_context=provider_base.ProviderCallContext(
+            conversation_state=state.with_pending_messages([
+                {"role": "user", "content": "new"},
+            ]),
+            context_window_tokens=100,
+        ),
+    )
+
+    assert response.content == "done"
+    assert len(bodies) == 2
+    assert bodies[0]["input"][-1] == {"type": "compaction_trigger"}
+    assert bodies[1]["input"][-1] != {"type": "compaction_trigger"}
+    assert provider.supports_native_compaction() is False
+
+
 @pytest.mark.asyncio
 async def test_codex_stream_surfaces_reasoning_summary(monkeypatch) -> None:
     def fake_token(**_kwargs):
@@ -559,7 +850,12 @@ async def test_codex_stream_surfaces_reasoning_summary(monkeypatch) -> None:
             await on_content_delta("answer")
         if on_thinking_delta:
             await on_thinking_delta("summary")
-        return "answer", [], "stop", {"prompt_tokens": 10, "completion_tokens": 5}, "summary"
+        return provider_base.LLMResponse(
+            content="answer",
+            finish_reason="stop",
+            usage={"prompt_tokens": 10, "completion_tokens": 5},
+            reasoning_content="summary",
+        )
 
     monkeypatch.setattr("nanobot.providers.openai_codex_provider._request_codex", fake_request)
 

tests/providers/test_openai_responses.py

@@ -1,9 +1,11 @@
 """Tests for the shared openai_responses converters and parsers."""
 
 import json
+from io import StringIO
 from unittest.mock import MagicMock, patch
 
 import pytest
+from loguru import logger
 
 from nanobot.providers.openai_responses.converters import (
     convert_messages,
@@ -12,12 +14,22 @@ from nanobot.providers.openai_responses.converters import (
     split_tool_call_id,
 )
 from nanobot.providers.openai_responses.parsing import (
+    ResponsesStreamCapture,
     consume_sdk_stream,
     consume_sse,
     consume_sse_with_reasoning,
+    is_replayable_finish_reason,
     map_finish_reason,
     parse_response_output,
 )
+from nanobot.providers.openai_responses.state import (
+    build_responses_state,
+    is_compaction_compatibility_error,
+    prepare_responses_input,
+    resolve_compact_threshold,
+    responses_state_context_tokens,
+    responses_state_items,
+)
 
 # ======================================================================
 # converters - split_tool_call_id
@@ -398,6 +410,17 @@ class TestMapFinishReason:
     def test_unknown_defaults_to_stop(self):
         assert map_finish_reason("some_new_status") == "stop"
 
+    @pytest.mark.parametrize("finish_reason", ["stop", "tool_calls", "function_call"])
+    def test_replayable_finish_reasons(self, finish_reason):
+        assert is_replayable_finish_reason(finish_reason) is True
+
+    @pytest.mark.parametrize(
+        "finish_reason",
+        ["length", "refusal", "content_filter", "error"],
+    )
+    def test_non_replayable_finish_reasons(self, finish_reason):
+        assert is_replayable_finish_reason(finish_reason) is False
+
 
 # ======================================================================
 # parsing - parse_response_output
@@ -418,6 +441,29 @@ class TestParseResponseOutput:
         assert result.usage == {"prompt_tokens": 10, "completion_tokens": 5, "total_tokens": 15}
         assert result.tool_calls == []
 
+    def test_refusal_response_surfaces_text_without_advancing_state(self):
+        refusal = "I can’t help with that request."
+        resp = {
+            "output": [{
+                "type": "message",
+                "role": "assistant",
+                "content": [{"type": "refusal", "refusal": refusal}],
+            }],
+            "status": "completed",
+            "usage": {},
+        }
+
+        result = parse_response_output(
+            resp,
+            state_provider="openai:test",
+            state_model="gpt-5.6",
+            state_input_items=[{"role": "user", "content": "request"}],
+        )
+
+        assert result.content == refusal
+        assert result.finish_reason == "refusal"
+        assert result.provider_state is None
+
     def test_tool_call_response(self):
         resp = {
             "output": [{
@@ -429,12 +475,18 @@ class TestParseResponseOutput:
             "status": "completed",
             "usage": {},
         }
-        result = parse_response_output(resp)
+        result = parse_response_output(
+            resp,
+            state_provider="openai:test",
+            state_model="gpt-5.6",
+            state_input_items=[{"role": "user", "content": "weather?"}],
+        )
         assert result.content is None
         assert len(result.tool_calls) == 1
         assert result.tool_calls[0].name == "get_weather"
         assert result.tool_calls[0].arguments == {"city": "SF"}
         assert result.tool_calls[0].id == "call_1|fc_1"
+        assert result.provider_state is not None
 
     def test_malformed_tool_arguments_logged(self):
         """Malformed JSON arguments should log a warning and remain non-object."""
@@ -493,10 +545,39 @@ class TestParseResponseOutput:
         assert result.content is None
         assert result.tool_calls == []
 
-    def test_incomplete_status(self):
+    @pytest.mark.parametrize(
-        resp = {"output": [], "status": "incomplete", "usage": {}}
+        ("reason", "expected_finish_reason"),
-        result = parse_response_output(resp)
+        [
-        assert result.finish_reason == "length"
+            ("max_output_tokens", "length"),
+            ("content_filter", "content_filter"),
+        ],
+    )
+    def test_incomplete_status(self, reason, expected_finish_reason):
+        resp = {
+            "output": [],
+            "status": "incomplete",
+            "incomplete_details": {"reason": reason},
+            "usage": {},
+        }
+        result = parse_response_output(
+            resp,
+            state_provider="openai:test",
+            state_model="gpt-5.6",
+            state_input_items=[{"role": "user", "content": "prompt"}],
+        )
+        assert result.finish_reason == expected_finish_reason
+        assert result.provider_state is None
+
+    def test_unknown_status_does_not_advance_provider_state(self):
+        result = parse_response_output(
+            {"output": [], "status": "future_terminal_status", "usage": {}},
+            state_provider="openai:test",
+            state_model="gpt-5.6",
+            state_input_items=[{"role": "user", "content": "prompt"}],
+        )
+
+        assert result.finish_reason == "stop"
+        assert result.provider_state is None
 
     def test_sdk_model_object(self):
         """parse_response_output should handle SDK objects with model_dump()."""
@@ -523,6 +604,194 @@ class TestParseResponseOutput:
         assert result.usage["completion_tokens"] == 50
         assert result.usage["total_tokens"] == 150
 
+    def test_preserves_every_output_item_as_opaque_state(self):
+        input_items = [{"role": "user", "content": "inspect the repo"}]
+        output = [
+            {
+                "id": "rs_1",
+                "type": "reasoning",
+                "encrypted_content": "opaque-secret",
+                "summary": [],
+            },
+            {
+                "id": "future_1",
+                "type": "future_item_type",
+                "provider_field": {"nested": True},
+            },
+            {
+                "id": "msg_1",
+                "type": "message",
+                "role": "assistant",
+                "status": "completed",
+                "content": [{"type": "output_text", "text": "done"}],
+            },
+        ]
+
+        result = parse_response_output(
+            {"output": output, "status": "completed", "usage": {}},
+            state_provider="openai:test",
+            state_model="gpt-5.6",
+            state_input_items=input_items,
+        )
+
+        assert result.provider_state is not None
+        assert responses_state_items(result.provider_state) == [*input_items, *output]
+
+
+class TestResponsesConversationState:
+    def test_server_compaction_prunes_superseded_prefix(self):
+        state = build_responses_state(
+            provider="openai:test",
+            model="gpt-5.6",
+            input_items=[
+                {"type": "message", "role": "user", "content": "old"},
+                {"type": "reasoning", "encrypted_content": "old-reasoning"},
+            ],
+            output_items=[
+                {"type": "compaction", "encrypted_content": "compact"},
+                {"type": "message", "role": "assistant", "content": "new"},
+            ],
+            usage={
+                "prompt_tokens": 90,
+                "completion_tokens": 10,
+                "total_tokens": 100,
+            },
+        )
+
+        assert responses_state_items(state) == [
+            {"type": "compaction", "encrypted_content": "compact"},
+            {"type": "message", "role": "assistant", "content": "new"},
+        ]
+        assert responses_state_context_tokens(state) == 100
+
+    def test_existing_compaction_keeps_canonical_retained_prefix(self):
+        canonical_input = [
+            {"type": "message", "role": "user", "content": "retained"},
+            {"type": "compaction", "encrypted_content": "compact"},
+        ]
+        output = [{"type": "message", "role": "assistant", "content": "new"}]
+
+        state = build_responses_state(
+            provider="openai:test",
+            model="gpt-5.6",
+            input_items=canonical_input,
+            output_items=output,
+        )
+
+        assert responses_state_items(state) == [*canonical_input, *output]
+
+    @pytest.mark.parametrize(
+        ("context_window", "max_output", "expected"),
+        [
+            (200_000, 20_000, 180_000),
+            (100_000, 30_000, 70_000),
+            (0, 4_096, None),
+        ],
+    )
+    def test_compact_threshold_reserves_codex_style_headroom(
+        self,
+        context_window,
+        max_output,
+        expected,
+    ):
+        assert resolve_compact_threshold(context_window, max_output) == expected
+
+    def test_compaction_compatibility_recognizes_old_sdk_signature_error(self):
+        error = TypeError("create() got an unexpected keyword argument 'context_management'")
+        assert is_compaction_compatibility_error(error) is True
+        assert is_compaction_compatibility_error(TypeError("unrelated argument")) is False
+
+    def test_state_observability_logs_counts_without_opaque_content(self):
+        secret = "opaque-secret-that-must-not-be-logged"
+        state = build_responses_state(
+            provider=f"openai:https://example.test/?key={secret}",
+            model=f"secret-model-{secret}",
+            input_items=[{"role": "user", "content": secret}],
+            output_items=[{"type": "reasoning", "encrypted_content": secret}],
+        ).with_pending_messages([{"role": "user", "content": secret}])
+        sink = StringIO()
+        sink_id = logger.add(sink, level="DEBUG", format="{message}")
+        try:
+            prepare_responses_input(
+                [{"role": "user", "content": secret}],
+                state=state,
+                provider=state.provider,
+                model=state.model,
+            )
+            build_responses_state(
+                provider=state.provider,
+                model=state.model,
+                input_items=[
+                    {"role": "user", "content": secret},
+                    {"type": "reasoning", "encrypted_content": secret},
+                ],
+                output_items=[
+                    {"type": "compaction", "encrypted_content": secret},
+                ],
+            )
+        finally:
+            logger.remove(sink_id)
+
+        log_text = sink.getvalue()
+        assert "prior_items=2" in log_text
+        assert "pending_messages=1" in log_text
+        assert "dropped_items=2" in log_text
+        assert secret not in log_text
+
+    def test_replays_exact_items_then_only_pending_and_new_messages(self):
+        prior_items = [
+            {"role": "user", "content": "first"},
+            {
+                "type": "reasoning",
+                "id": "rs_1",
+                "encrypted_content": "opaque-secret",
+            },
+            {
+                "type": "function_call",
+                "id": "fc_1",
+                "call_id": "call_1",
+                "name": "read_file",
+                "arguments": '{"path":"a.py"}',
+            },
+        ]
+        state = build_responses_state(
+            provider="openai:test",
+            model="gpt-5.6",
+            input_items=prior_items[:1],
+            output_items=prior_items[1:],
+        ).with_pending_messages([
+            {
+                "role": "tool",
+                "tool_call_id": "call_1|fc_1",
+                "content": "file contents",
+            },
+            {"role": "user", "content": "continue"},
+        ])
+
+        instructions, items, replayed = prepare_responses_input(
+            [
+                {"role": "system", "content": "current instructions"},
+                {"role": "user", "content": "a lossy public transcript"},
+            ],
+            state=state,
+            provider="openai:test",
+            model="gpt-5.6",
+        )
+
+        assert instructions == "current instructions"
+        assert replayed is True
+        assert items[:3] == prior_items
+        assert items[3] == {
+            "type": "function_call_output",
+            "call_id": "call_1",
+            "output": "file contents",
+        }
+        assert items[4] == {
+            "role": "user",
+            "content": [{"type": "input_text", "text": "continue"}],
+        }
+        assert "lossy public transcript" not in str(items)
+
 
 # ======================================================================
 # parsing - consume_sse
@@ -553,6 +822,122 @@ class TestConsumeSse:
         assert tool_calls == []
         assert finish_reason == "stop"
 
+    @pytest.mark.asyncio
+    async def test_refusal_events_reconcile_parts_and_terminal_output(self):
+        refusal = "First and second sentence. Done-only. Terminal suffix."
+        terminal_response = {
+            "status": "completed",
+            "output": [{
+                "type": "message",
+                "id": "msg_2",
+                "role": "assistant",
+                "content": [{"type": "refusal", "refusal": refusal}],
+            }],
+        }
+        response = _SseResponse([
+            {
+                "type": "response.refusal.delta",
+                "item_id": "msg_1",
+                "content_index": 0,
+                "delta": "First",
+            },
+            {
+                "type": "response.refusal.delta",
+                "item_id": "msg_1",
+                "content_index": 1,
+                "delta": " and second",
+            },
+            {
+                "type": "response.refusal.done",
+                "item_id": "msg_1",
+                "content_index": 0,
+                "refusal": "First",
+            },
+            {
+                "type": "response.refusal.done",
+                "item_id": "msg_1",
+                "content_index": 1,
+                "refusal": " and second sentence.",
+            },
+            {
+                "type": "response.refusal.done",
+                "item_id": "msg_2",
+                "content_index": 0,
+                "refusal": " Done-only.",
+            },
+            {
+                "type": "response.refusal.delta",
+                "item_id": "msg_2",
+                "content_index": 1,
+                "delta": " Terminal",
+            },
+            {"type": "response.completed", "response": terminal_response},
+        ])
+        capture = ResponsesStreamCapture()
+        deltas: list[str] = []
+
+        async def on_content(delta: str) -> None:
+            deltas.append(delta)
+
+        content, _, finish_reason, _, _ = await consume_sse_with_reasoning(
+            response,
+            on_content_delta=on_content,
+            capture=capture,
+        )
+
+        assert content == refusal
+        assert deltas == [
+            "First",
+            " and second",
+            " sentence.",
+            " Done-only.",
+            " Terminal",
+            " suffix.",
+        ]
+        assert finish_reason == "refusal"
+        assert capture.completed is True
+        assert is_replayable_finish_reason(finish_reason) is False
+
+    @pytest.mark.asyncio
+    @pytest.mark.parametrize("source", ["events", "terminal"])
+    async def test_refusal_without_deltas_has_non_replayable_finish(self, source: str):
+        refusal = "I can’t help with that request."
+        terminal_response = {
+            "status": "completed",
+            "output": [{
+                "type": "message",
+                "id": "msg_1",
+                "role": "assistant",
+                "content": [{"type": "refusal", "refusal": refusal}],
+            }],
+        }
+        events = (
+            [
+                {"type": "response.refusal.done", "refusal": refusal},
+                {"type": "response.completed", "response": {"status": "completed"}},
+            ]
+            if source == "events"
+            else [{"type": "response.completed", "response": terminal_response}]
+        )
+        response = _SseResponse(events)
+        capture = ResponsesStreamCapture()
+        deltas: list[str] = []
+
+        async def on_content(delta: str) -> None:
+            deltas.append(delta)
+
+        content, _, finish_reason, _, _ = await consume_sse_with_reasoning(
+            response,
+            on_content_delta=on_content,
+            capture=capture,
+        )
+
+        assert content == refusal
+        assert deltas == [refusal]
+        assert finish_reason == "refusal"
+        assert capture.completed is True
+        assert is_replayable_finish_reason(finish_reason) is False
+
     @pytest.mark.asyncio
     async def test_reasoning_summary_delta_extracted(self):
         response = _SseResponse([
@@ -599,6 +984,139 @@ class TestConsumeSse:
 
         assert reasoning == "cached summary"
 
+    @pytest.mark.asyncio
+    async def test_capture_commits_exact_items_only_after_completed_event(self):
+        output = [
+            {
+                "type": "reasoning",
+                "id": "rs_1",
+                "encrypted_content": "opaque-secret",
+            },
+            {"type": "future_item_type", "id": "future_1", "value": 7},
+        ]
+        capture = ResponsesStreamCapture()
+        response = _SseResponse([
+            {
+                "type": "response.output_item.done",
+                "output_index": 0,
+                "item": output[0],
+            },
+            {
+                "type": "response.output_item.done",
+                "output_index": 1,
+                "item": output[1],
+            },
+            {
+                "type": "response.completed",
+                "response": {"status": "completed", "output": output},
+            },
+        ])
+
+        await consume_sse_with_reasoning(response, capture=capture)
+
+        assert capture.completed is True
+        assert capture.output_items == output
+
+    @pytest.mark.asyncio
+    async def test_capture_keeps_done_items_when_completed_output_is_empty(self):
+        output = [
+            {
+                "type": "reasoning",
+                "id": "rs_1",
+                "encrypted_content": "opaque-secret",
+                "summary": [],
+            },
+            {
+                "type": "function_call",
+                "id": "fc_1",
+                "call_id": "call_1",
+                "name": "read_file",
+                "arguments": '{"path":"weather/SKILL.md"}',
+            },
+        ]
+        capture = ResponsesStreamCapture()
+        response = _SseResponse([
+            {
+                "type": "response.output_item.done",
+                "output_index": index,
+                "item": item,
+            }
+            for index, item in enumerate(output)
+        ] + [{
+            "type": "response.completed",
+            "response": {"status": "completed", "output": []},
+        }])
+
+        await consume_sse_with_reasoning(response, capture=capture)
+
+        assert capture.completed is True
+        assert capture.output_items == output
+
+    @pytest.mark.asyncio
+    @pytest.mark.parametrize(
+        ("reason", "expected_finish_reason"),
+        [
+            ("max_output_tokens", "length"),
+            ("content_filter", "content_filter"),
+        ],
+    )
+    async def test_incomplete_event_commits_capture_usage(
+        self,
+        reason,
+        expected_finish_reason,
+    ):
+        output = [
+            {
+                "type": "message",
+                "id": "msg_1",
+                "status": "incomplete",
+                "content": [{"type": "output_text", "text": "partial"}],
+            },
+        ]
+        terminal_response = {
+            "id": "resp_1",
+            "status": "incomplete",
+            "incomplete_details": {"reason": reason},
+            "output": output,
+            "usage": {"input_tokens": 10, "output_tokens": 5, "total_tokens": 15},
+        }
+        capture = ResponsesStreamCapture()
+        response = _SseResponse([
+            {"type": "response.output_text.delta", "delta": "partial"},
+            {"type": "response.incomplete", "response": terminal_response},
+        ])
+
+        content, _, finish_reason, usage, _ = await consume_sse_with_reasoning(
+            response,
+            capture=capture,
+        )
+
+        assert content == "partial"
+        assert finish_reason == expected_finish_reason
+        assert usage == {"prompt_tokens": 10, "completion_tokens": 5, "total_tokens": 15}
+        assert capture.completed is True
+        assert capture.response == terminal_response
+        assert capture.output_items == output
+
+    @pytest.mark.asyncio
+    async def test_capture_does_not_commit_interrupted_stream(self):
+        capture = ResponsesStreamCapture()
+        response = _SseResponse([
+            {
+                "type": "response.output_item.done",
+                "output_index": 0,
+                "item": {
+                    "type": "reasoning",
+                    "id": "rs_1",
+                    "encrypted_content": "opaque-secret",
+                },
+            },
+        ])
+
+        await consume_sse_with_reasoning(response, capture=capture)
+
+        assert capture.completed is False
+
     @pytest.mark.asyncio
     async def test_reasoning_summary_from_done_item(self):
         response = _SseResponse([
@@ -755,6 +1273,131 @@ class TestConsumeSdkStream:
         assert tool_calls == []
         assert finish_reason == "stop"
 
+    @pytest.mark.asyncio
+    async def test_refusal_events_reconcile_parts_and_terminal_output(self):
+        refusal = "First and second sentence. Done-only. Terminal suffix."
+        terminal_response = {
+            "status": "completed",
+            "output": [{
+                "type": "message",
+                "id": "msg_2",
+                "role": "assistant",
+                "content": [{"type": "refusal", "refusal": refusal}],
+            }],
+        }
+        resp_obj = MagicMock(status="completed", usage=None, output=[])
+        resp_obj.model_dump.return_value = terminal_response
+        events = [
+            MagicMock(
+                type="response.refusal.delta",
+                item_id="msg_1",
+                content_index=0,
+                delta="First",
+            ),
+            MagicMock(
+                type="response.refusal.delta",
+                item_id="msg_1",
+                content_index=1,
+                delta=" and second",
+            ),
+            MagicMock(
+                type="response.refusal.done",
+                item_id="msg_1",
+                content_index=0,
+                refusal="First",
+            ),
+            MagicMock(
+                type="response.refusal.done",
+                item_id="msg_1",
+                content_index=1,
+                refusal=" and second sentence.",
+            ),
+            MagicMock(
+                type="response.refusal.done",
+                item_id="msg_2",
+                content_index=0,
+                refusal=" Done-only.",
+            ),
+            MagicMock(
+                type="response.refusal.delta",
+                item_id="msg_2",
+                content_index=1,
+                delta=" Terminal",
+            ),
+            MagicMock(type="response.completed", response=resp_obj),
+        ]
+        capture = ResponsesStreamCapture()
+        deltas: list[str] = []
+
+        async def on_content(delta: str) -> None:
+            deltas.append(delta)
+
+        async def stream():
+            for event in events:
+                yield event
+
+        content, _, finish_reason, _, _ = await consume_sdk_stream(
+            stream(),
+            on_content_delta=on_content,
+            capture=capture,
+        )
+
+        assert content == refusal
+        assert deltas == [
+            "First",
+            " and second",
+            " sentence.",
+            " Done-only.",
+            " Terminal",
+            " suffix.",
+        ]
+        assert finish_reason == "refusal"
+        assert capture.completed is True
+        assert is_replayable_finish_reason(finish_reason) is False
+
+    @pytest.mark.asyncio
+    @pytest.mark.parametrize("source", ["events", "terminal"])
+    async def test_refusal_without_deltas_has_non_replayable_finish(self, source: str):
+        refusal = "I can’t help with that request."
+        terminal_response = {
+            "status": "completed",
+            "output": [{
+                "type": "message",
+                "id": "msg_1",
+                "role": "assistant",
+                "content": [{"type": "refusal", "refusal": refusal}],
+            }],
+        }
+        resp_obj = MagicMock(status="completed", usage=None, output=[])
+        resp_obj.model_dump.return_value = terminal_response
+        capture = ResponsesStreamCapture()
+        deltas: list[str] = []
+
+        async def on_content(delta: str) -> None:
+            deltas.append(delta)
+
+        async def stream():
+            if source == "events":
+                yield MagicMock(type="response.refusal.done", refusal=refusal)
+                yield MagicMock(
+                    type="response.completed",
+                    response={"status": "completed"},
+                )
+            else:
+                yield MagicMock(type="response.completed", response=resp_obj)
+
+        content, _, finish_reason, _, _ = await consume_sdk_stream(
+            stream(),
+            on_content_delta=on_content,
+            capture=capture,
+        )
+
+        assert content == refusal
+        assert deltas == [refusal]
+        assert finish_reason == "refusal"
+        assert capture.completed is True
+        assert is_replayable_finish_reason(finish_reason) is False
+
     @pytest.mark.asyncio
     async def test_on_content_delta_called(self):
         ev1 = MagicMock(type="response.output_text.delta", delta="hi")
@@ -919,6 +1562,64 @@ class TestConsumeSdkStream:
         _, _, _, usage, _ = await consume_sdk_stream(stream())
         assert usage == {"prompt_tokens": 10, "completion_tokens": 5, "total_tokens": 15}
 
+    @pytest.mark.asyncio
+    @pytest.mark.parametrize(
+        ("reason", "expected_finish_reason"),
+        [
+            ("max_output_tokens", "length"),
+            ("content_filter", "content_filter"),
+        ],
+    )
+    async def test_incomplete_event_commits_capture_usage(
+        self,
+        reason,
+        expected_finish_reason,
+    ):
+        output = [
+            {
+                "type": "message",
+                "id": "msg_1",
+                "status": "incomplete",
+                "content": [{"type": "output_text", "text": "partial"}],
+            },
+        ]
+        usage_obj = MagicMock(input_tokens=10, output_tokens=5, total_tokens=15)
+        output_item = MagicMock(type="message")
+        terminal_response = {
+            "id": "resp_1",
+            "status": "incomplete",
+            "incomplete_details": {"reason": reason},
+            "output": output,
+            "usage": {"input_tokens": 10, "output_tokens": 5, "total_tokens": 15},
+        }
+        resp_obj = MagicMock(
+            status="incomplete",
+            usage=usage_obj,
+            output=[output_item],
+        )
+        resp_obj.model_dump.return_value = terminal_response
+        events = [
+            MagicMock(type="response.output_text.delta", delta="partial"),
+            MagicMock(type="response.incomplete", response=resp_obj),
+        ]
+        capture = ResponsesStreamCapture()
+
+        async def stream():
+            for event in events:
+                yield event
+
+        content, _, finish_reason, usage, _ = await consume_sdk_stream(
+            stream(),
+            capture=capture,
+        )
+
+        assert content == "partial"
+        assert finish_reason == expected_finish_reason
+        assert usage == {"prompt_tokens": 10, "completion_tokens": 5, "total_tokens": 15}
+        assert capture.completed is True
+        assert capture.response == terminal_response
+        assert capture.output_items == output
+
     @pytest.mark.asyncio
     async def test_reasoning_extracted(self):
         summary_item = MagicMock(type="summary_text", text="thinking...")

tests/providers/test_provider_retry.py

@@ -3,7 +3,14 @@ import copy
 
 import pytest
 
-from nanobot.providers.base import RETRY_AFTER_BUFFER, GenerationSettings, LLMProvider, LLMResponse
+from nanobot.providers.base import (
+    RETRY_AFTER_BUFFER,
+    GenerationSettings,
+    LLMProvider,
+    LLMResponse,
+    ProviderCallContext,
+    ProviderConversationState,
+)
 
 
 class ScriptedProvider(LLMProvider):
@@ -330,6 +337,79 @@ async def test_successful_image_retry_mutates_original_messages_in_place() -> No
     assert any("not delivered" in (block.get("text") or "").lower() for block in content)
 
 
+@pytest.mark.asyncio
+@pytest.mark.parametrize(
+    ("messages", "payload", "pending_messages"),
+    [
+        (_IMAGE_MSG, {}, _IMAGE_MSG),
+        (
+            [{"role": "user", "content": "continue"}],
+            {
+                "items": [
+                    {
+                        "type": "message",
+                        "role": "user",
+                        "content": [
+                            {
+                                "type": "input_image",
+                                "image_url": "data:image/png;base64,abc",
+                            }
+                        ],
+                    }
+                ]
+            },
+            [],
+        ),
+    ],
+    ids=["pending-image", "opaque-payload-image"],
+)
+async def test_image_retry_discards_provider_state_with_images(
+    messages,
+    payload,
+    pending_messages,
+) -> None:
+    class ContextScriptedProvider(ScriptedProvider):
+        def __init__(self, responses):
+            super().__init__(responses)
+            self.contexts: list[ProviderCallContext] = []
+
+        async def chat_with_context(
+            self,
+            *,
+            provider_context: ProviderCallContext,
+            **kwargs,
+        ) -> LLMResponse:
+            self.contexts.append(provider_context)
+            return await self.chat(**kwargs)
+
+    provider = ContextScriptedProvider([
+        LLMResponse(content="model does not support images", finish_reason="error"),
+        LLMResponse(content="ok, no image"),
+    ])
+    messages = copy.deepcopy(messages)
+    state = ProviderConversationState(
+        kind="openai_responses",
+        provider="openai:test",
+        model="gpt-5.6",
+        version=1,
+        payload=copy.deepcopy(payload),
+        pending_messages=copy.deepcopy(pending_messages),
+    )
+
+    response = await provider.chat_with_retry(
+        messages=messages,
+        provider_context=ProviderCallContext(conversation_state=state),
+    )
+
+    assert response.content == "ok, no image"
+    retry_context = provider.contexts[-1]
+    assert isinstance(retry_context, ProviderCallContext)
+    assert retry_context.conversation_state is None
+    public_content = messages[0]["content"]
+    if isinstance(public_content, list):
+        assert all(block.get("type") != "image_url" for block in public_content)
+
+
 @pytest.mark.asyncio
 async def test_non_transient_error_without_images_no_retry() -> None:
     """Non-transient errors without image content are returned immediately."""

tests/providers/test_responses_circuit_breaker.py

@@ -4,6 +4,7 @@ import time
 
 import pytest
 
+from nanobot.providers.base import ProviderCallContext
 from nanobot.providers.openai_compat_provider import (
     _RESPONSES_FAILURE_THRESHOLD,
     _RESPONSES_PROBE_INTERVAL_S,
@@ -28,6 +29,26 @@ def test_responses_api_available_by_default(provider):
     assert provider._should_use_responses_api("gpt-5", None) is True
 
 
+def test_direct_openai_enables_server_compaction(provider):
+    provider._extra_body = {}
+
+    body = provider._build_responses_body(
+        messages=[{"role": "user", "content": "hello"}],
+        tools=None,
+        model="gpt-5.6",
+        max_tokens=30_000,
+        temperature=0.1,
+        reasoning_effort="high",
+        tool_choice=None,
+        provider_context=ProviderCallContext(context_window_tokens=100_000),
+    )
+
+    assert body["context_management"] == [{
+        "type": "compaction",
+        "compact_threshold": 70_000,
+    }]
+
+
 def test_api_type_chat_completions_disables_responses(provider):
     provider._api_type = "chat_completions"
     assert provider._should_use_responses_api("gpt-5", None) is False

tests/tools/test_exec_session_tools.py

@@ -19,6 +19,8 @@ from nanobot.agent.tools.exec_session import (
     ExecSessionManager,
     ListExecSessionsTool,
     WriteStdinTool,
+    _BoundedOutputBuffer,
+    _SessionPoll,
 )
 from nanobot.agent.tools.registry import is_tool_error_result
 from nanobot.agent.tools.shell import ExecTool
@@ -143,6 +145,88 @@ def test_exec_session_accepts_max_output_tokens_alias(tmp_path):
     assert "Exit code: 0" in result
 
 
+def test_bounded_output_buffer_keeps_head_tail_and_exact_drop_count():
+    buffer = _BoundedOutputBuffer(10)
+
+    buffer.append("012345")
+    buffer.append("6789ABCDEF")
+
+    assert buffer.retained_chars == 10
+    assert buffer.drain() == ("01234BCDEF", 6)
+    assert buffer.retained_chars == 0
+
+
+def test_exec_session_bounds_unpolled_stdout_and_stderr(tmp_path):
+    async def run() -> tuple[int, int, str, int]:
+        manager = ExecSessionManager()
+        tool = ExecTool(working_dir=str(tmp_path), timeout=5, session_manager=manager)
+        command = _python_command(
+            "import sys,time; time.sleep(0.05); "
+            "sys.stdout.write('OUT_HEAD' + 'o' * 200000 + 'OUT_TAIL'); "
+            "sys.stderr.write('ERR_HEAD' + 'e' * 200000 + 'ERR_TAIL')"
+        )
+
+        initial = await tool.execute(
+            command=command,
+            yield_time_ms=0,
+            max_output_chars=1000,
+        )
+        sid = _session_id(initial)
+        session = manager._sessions[sid]
+        await asyncio.wait_for(session.process.wait(), timeout=5)
+        await asyncio.wait_for(
+            asyncio.gather(session._stdout_task, session._stderr_task),
+            timeout=5,
+        )
+        retained_stdout = session._stdout.retained_chars
+        retained_stderr = session._stderr.retained_chars
+        poll = await manager.write(
+            session_id=sid,
+            chars=None,
+            close_stdin=False,
+            terminate=False,
+            yield_time_ms=0,
+            max_output_chars=1000,
+        )
+        return retained_stdout, retained_stderr, poll.output, poll.truncated_chars
+
+    retained_stdout, retained_stderr, output, truncated_chars = asyncio.run(run())
+
+    assert retained_stdout == 50000
+    assert retained_stderr == 50000
+    assert output.startswith("OUT_HEAD")
+    assert output.endswith("ERR_TAIL")
+    assert truncated_chars > 390000
+
+
+def test_write_stdin_wait_for_keeps_aggregate_within_output_budget():
+    async def run() -> str:
+        manager = SimpleNamespace(
+            write=AsyncMock(side_effect=[
+                _SessionPoll(output="HEAD" + "a" * 596, done=False, exit_code=None),
+                _SessionPoll(output="b" * 600, done=False, exit_code=None),
+                _SessionPoll(output="c" * 590 + "TARGET", done=False, exit_code=None),
+            ])
+        )
+        tool = WriteStdinTool(manager=manager)
+        return await tool._wait_for_output(
+            session_id="session",
+            chars=None,
+            close_stdin=False,
+            terminate=False,
+            wait_for="TARGET",
+            wait_timeout_ms=1000,
+            max_output_chars=1000,
+        )
+
+    result = asyncio.run(run())
+
+    assert result.startswith("HEAD")
+    assert "TARGET" in result
+    assert "(796 chars truncated from output)" in result
+    assert len(result) < 1100
+
+
 def test_exec_one_shot_accepts_max_output_tokens_alias(tmp_path):
     async def run() -> str:
         tool = ExecTool(working_dir=str(tmp_path), timeout=5)

tests/utils/test_gitstore.py

@@ -310,3 +310,25 @@ class TestNestedRepoProtection:
 
         assert result is False
         assert not (workspace / ".git").exists()
+
+
+class TestCommitIdEncoding:
+    """Commit ids must be usable with git, not hex-of-hex."""
+
+    def test_auto_commit_returns_the_real_short_sha(self, git, tmp_path):
+        (tmp_path / "MEMORY.md").write_text("- a fact\n", encoding="utf-8")
+        sha = git.auto_commit("memory update")
+        expected = subprocess.run(
+            ["git", "-C", str(tmp_path), "log", "-1", "--format=%h", "--abbrev=8"],
+            capture_output=True, text=True, check=True,
+        ).stdout.strip()
+        assert sha == expected
+
+    def test_a_real_git_sha_resolves(self, git, tmp_path):
+        (tmp_path / "MEMORY.md").write_text("- a fact\n", encoding="utf-8")
+        git.auto_commit("memory update")
+        real = subprocess.run(
+            ["git", "-C", str(tmp_path), "log", "-1", "--format=%h", "--abbrev=8"],
+            capture_output=True, text=True, check=True,
+        ).stdout.strip()
+        assert git._resolve_sha(real) is not None

tests/utils/test_webui_transcript.py

@@ -696,6 +696,42 @@ def test_replay_preserves_local_trigger_source_metadata(tmp_path, monkeypatch) -
     assert msgs[0]["source"] == {"kind": "local_trigger", "label": "PR review"}
 
 
+def test_replay_preserves_automation_source_metadata_on_streamed_reply(
+    tmp_path,
+    monkeypatch,
+) -> None:
+    monkeypatch.setattr("nanobot.config.paths.get_data_dir", lambda: tmp_path)
+    key = "websocket:t-streamed-cron-source"
+    source = {"kind": "cron", "label": "Repo check"}
+
+    for record in (
+        {
+            "event": "delta",
+            "chat_id": "t-streamed-cron-source",
+            "text": "Repo ",
+            "source": source,
+        },
+        {
+            "event": "delta",
+            "chat_id": "t-streamed-cron-source",
+            "text": "clean.",
+            "source": source,
+        },
+        {
+            "event": "stream_end",
+            "chat_id": "t-streamed-cron-source",
+            "source": source,
+        },
+        {"event": "turn_end", "chat_id": "t-streamed-cron-source"},
+    ):
+        append_transcript_object(key, record)
+
+    msgs = replay_transcript_to_ui_messages(read_transcript_lines(key))
+
+    assert msgs[0]["content"] == "Repo clean."
+    assert msgs[0]["source"] == source
+
+
 def test_replay_preserves_legacy_trigger_source_metadata(tmp_path, monkeypatch) -> None:
     monkeypatch.setattr("nanobot.config.paths.get_data_dir", lambda: tmp_path)
     key = "websocket:t-trigger-source"

tests/webui/test_session_list_index.py

@@ -8,6 +8,7 @@ import pytest
 
 import nanobot.webui.session_list_index as session_list_index
 from nanobot.cron.session_turns import CRON_HISTORY_META
+from nanobot.providers.base import ProviderConversationState
 from nanobot.session.automation_turns import AUTOMATION_HISTORY_META
 from nanobot.session.history_visibility import HIDDEN_HISTORY_META
 from nanobot.session.manager import SessionManager
@@ -85,6 +86,26 @@ def test_webui_session_list_rescans_only_changed_file(tmp_path: Path, monkeypatc
     assert {row["preview"] for row in rows} == {"first", "second after"}
 
 
+def test_webui_session_list_skips_provider_state_before_preview_budget(
+    tmp_path: Path,
+    monkeypatch,
+) -> None:
+    monkeypatch.setattr(session_list_index, "_SESSION_LIST_PREVIEW_MAX_CHARS", 100)
+    manager = SessionManager(tmp_path)
+    session = manager.get_or_create("websocket:private-state")
+    session.provider_state = ProviderConversationState(
+        kind="openai_responses",
+        provider="openai:test",
+        model="test-model",
+        version=1,
+        payload={"items": [{"encrypted_content": "x" * 200}]},
+    )
+    session.add_message("user", "visible preview")
+    manager.save(session)
+
+    assert list_webui_sessions(manager)[0]["preview"] == "visible preview"
+
+
 def test_webui_session_list_drops_deleted_index_rows(tmp_path: Path) -> None:
     manager = SessionManager(tmp_path)
     session = manager.get_or_create("websocket:deleted")

tests/webui/test_token_usage.py

@@ -1,5 +1,6 @@
 from __future__ import annotations
 
+import json
 from datetime import datetime, timezone
 from types import SimpleNamespace
 
@@ -14,6 +15,60 @@ from nanobot.webui.token_usage import (
 )
 
 
+def _write_state(tmp_path, days: dict) -> None:
+    state_dir = tmp_path / "webui"
+    state_dir.mkdir(parents=True, exist_ok=True)
+    (state_dir / "token-usage.json").write_text(
+        json.dumps({"days": days}), encoding="utf-8"
+    )
+
+
+def test_payload_tolerates_malformed_persisted_day_keys(tmp_path, monkeypatch) -> None:
+    """Day keys that are not real dates must not break settings payloads.
+
+    normalize_token_usage_state only length-checks day keys, so a hand-edited
+    10-char key survives reads and atomic rewrites; token_usage_payload then
+    parsed it with an unguarded fromisoformat, failing every /api/settings and
+    /api/settings/usage request until the file was fixed by hand.
+    """
+    monkeypatch.setattr("nanobot.webui.token_usage.get_webui_dir", lambda: tmp_path / "webui")
+    _write_state(tmp_path, {
+        "not-a-dat3": {"total_tokens": 7, "requests": 1},
+        "2026-13-01": {"total_tokens": 9, "requests": 1},
+        "2026-06-02": {"total_tokens": 5, "requests": 1},
+    })
+
+    payload = token_usage_payload(
+        timezone_name="UTC",
+        now=datetime(2026, 6, 3, 12, 0, tzinfo=timezone.utc),
+    )
+
+    assert payload["total_tokens"] == 5
+    assert payload["total_tokens_30d"] == 5
+    assert payload["requests_30d"] == 1
+    assert payload["active_days_30d"] == 1
+
+
+def test_record_scrubs_malformed_day_keys(tmp_path, monkeypatch) -> None:
+    """Rewrites drop malformed day keys instead of persisting them forever."""
+    monkeypatch.setattr("nanobot.webui.token_usage.get_webui_dir", lambda: tmp_path / "webui")
+    _write_state(tmp_path, {
+        "not-a-dat3": {"total_tokens": 7, "requests": 1},
+        "2026-06-02": {"total_tokens": 5, "requests": 1},
+    })
+
+    record_token_usage(
+        {"prompt_tokens": 1, "completion_tokens": 1},
+        timezone_name="UTC",
+        now=datetime(2026, 6, 3, 12, 0, tzinfo=timezone.utc),
+    )
+
+    raw = json.loads((tmp_path / "webui" / "token-usage.json").read_text(encoding="utf-8"))
+    assert "not-a-dat3" not in raw["days"]
+    assert "2026-06-02" in raw["days"]
+    assert "2026-06-03" in raw["days"]
+
+
 def test_record_token_usage_aggregates_by_local_day(tmp_path, monkeypatch) -> None:
     monkeypatch.setattr("nanobot.webui.token_usage.get_webui_dir", lambda: tmp_path / "webui")
 

webui/src/App.tsx

@@ -37,6 +37,12 @@ import {
 import { displayTitle } from "@/lib/chat-groups";
 import { deriveTitle } from "@/lib/format";
 import { NanobotClient } from "@/lib/nanobot-client";
+import {
+  isQuickChatKey,
+  QUICK_CHAT_ID,
+  QUICK_CHAT_KEY,
+  quickChatSession,
+} from "@/lib/quick-chat";
 import { ClientProvider, useClient } from "@/providers/ClientProvider";
 import type {
   BootstrapResponse,
@@ -225,6 +231,9 @@ function readShellRoute(): ShellRoute {
   if (path === "/skills") {
     return { view: "skills", activeKey, settingsSection: "skills" };
   }
+  if (path === "/quick-chat") {
+    return { view: "chat", activeKey: QUICK_CHAT_KEY, settingsSection: "overview" };
+  }
   if (path.startsWith("/chat/")) {
     const encoded = path.slice("/chat/".length);
     try {
@@ -241,6 +250,7 @@ function readShellRoute(): ShellRoute {
 
 function shellRouteHash(route: ShellRoute): string {
   if (route.view === "chat") {
+    if (isQuickChatKey(route.activeKey)) return "#/quick-chat";
     return route.activeKey
       ? `#/chat/${encodeURIComponent(route.activeKey)}`
       : "#/new";
@@ -947,8 +957,16 @@ function Shell({
     deleteChat,
     getSessionAutomations,
   } = useSessions();
+  const regularSessions = useMemo(
+    () => sessions.filter((session) => !isQuickChatKey(session.key)),
+    [sessions],
+  );
+  const quickSession = useMemo(
+    () => quickChatSession(sessions.find((session) => isQuickChatKey(session.key))),
+    [sessions],
+  );
   const { state: sidebarState, update: updateSidebarState } =
-    useSidebarState(sessions, !loading);
+    useSidebarState(regularSessions, !loading);
   const initialRouteRef = useRef<ShellRoute | null>(null);
   if (!initialRouteRef.current) initialRouteRef.current = readShellRoute();
   const [activeKey, setActiveKey] = useState<string | null>(
@@ -1114,8 +1132,10 @@ function Shell({
 
   const activeSession = useMemo<ChatSummary | null>(() => {
     if (!activeKey) return null;
+    if (isQuickChatKey(activeKey)) return quickSession;
     return sessions.find((s) => s.key === activeKey) ?? null;
-  }, [sessions, activeKey]);
+  }, [sessions, activeKey, quickSession]);
+  const quickChatActive = isQuickChatKey(activeKey);
   const runningChatIdList = useMemo(() => Array.from(runningChatIds), [runningChatIds]);
   const updatedChatIdList = useMemo(() => Array.from(updatedChatIds), [updatedChatIds]);
   const activeChatId = activeSession?.chatId ?? null;
@@ -1130,6 +1150,9 @@ function Shell({
     });
   }, [activeChatId]);
   const activeWorkspaceScope = useMemo<WorkspaceScopePayload | null>(() => {
+    if (quickChatActive) {
+      return workspaces?.default_scope ?? null;
+    }
     if (activeChatId && workspaceOverrides[activeChatId]) {
       return workspaceOverrides[activeChatId];
     }
@@ -1141,6 +1164,7 @@ function Shell({
     activeChatId,
     activeSession?.workspaceScope,
     draftWorkspaceScope,
+    quickChatActive,
     workspaceOverrides,
     workspaces?.default_scope,
   ]);
@@ -1161,7 +1185,10 @@ function Shell({
 
   useEffect(() => {
     if (loading) return;
-    const knownChatIds = new Set(sessions.map((session) => session.chatId));
+    const knownChatIds = new Set([
+      QUICK_CHAT_ID,
+      ...sessions.map((session) => session.chatId),
+    ]);
     setUpdatedChatIds((current) => {
       const next = new Set(
         Array.from(current).filter((chatId) => knownChatIds.has(chatId)),
@@ -1176,6 +1203,7 @@ function Shell({
 
   useEffect(() => {
     if (loading || !activeKey) return;
+    if (isQuickChatKey(activeKey)) return;
     if (sessions.some((session) => session.key === activeKey)) return;
     const currentRoute = readShellRoute();
     navigate(
@@ -1417,6 +1445,18 @@ function Shell({
     setMobileSidebarOpen(false);
   }, [navigate]);
 
+  const onOpenQuickChat = useCallback(() => {
+    setDraftWorkspaceScope(null);
+    setWorkspaceError(null);
+    setSessionSearchOpen(false);
+    navigate({
+      view: "chat",
+      activeKey: QUICK_CHAT_KEY,
+      settingsSection: "overview",
+    });
+    setMobileSidebarOpen(false);
+  }, [navigate]);
+
   const onNewChatInProject = useCallback(
     (projectPath: string, projectName: string) => {
       const base = workspaces?.default_scope ?? activeWorkspaceScope;
@@ -1682,6 +1722,7 @@ function Shell({
     setMobileSidebarOpen(false);
     const nextKey = (() => {
       if (!activeKey) return null;
+      if (isQuickChatKey(activeKey)) return activeKey;
       if (sessions.some((session) => session.key === activeKey)) return activeKey;
       return sessions[0]?.key ?? null;
     })();
@@ -1773,7 +1814,10 @@ function Shell({
     });
   }, [client, t]);
 
-  const onTurnEnd = useDeferredTitleRefresh(activeSession, refresh);
+  const onTurnEnd = useDeferredTitleRefresh(
+    quickChatActive ? null : activeSession,
+    refresh,
+  );
 
   const onConfirmDelete = useCallback(async () => {
     if (!pendingDelete) return;
@@ -1863,7 +1907,9 @@ function Shell({
     });
   }, []);
 
-  const headerTitle = activeSession
+  const headerTitle = quickChatActive
+    ? t("sidebar.quickChat")
+    : activeSession
     ? sidebarState.title_overrides[activeSession.key] ||
       activeSession.title ||
       deriveTitle(activeSession.preview, t("chat.newChat"))
@@ -1900,9 +1946,12 @@ function Shell({
   }, [activeSession, headerTitle, i18n.resolvedLanguage, t, view]);
 
   const sidebarProps = {
-    sessions,
+    sessions: regularSessions,
-    activeKey,
+    activeKey: view === "chat" ? activeKey : null,
     loading,
+    quickChatActive: view === "chat" && quickChatActive,
+    newChatActive: view === "chat" && activeKey === null,
+    onOpenQuickChat,
     onNewChat,
     onSelect: onSelectChat,
     onRequestDelete,
@@ -2065,7 +2114,7 @@ function Shell({
               <SessionSearchDialog
                 open
                 onOpenChange={setSessionSearchOpen}
-                sessions={sessions}
+                sessions={regularSessions}
                 activeKey={activeKey}
                 loading={loading}
                 titleOverrides={sidebarState.title_overrides}
@@ -2090,7 +2139,7 @@ function Shell({
                 onToggleSidebar={toggleSidebar}
                 onNewChat={onNewChat}
                 onCreateChat={onCreateChat}
-                onForkChat={onForkChat}
+                onForkChat={quickChatActive ? undefined : onForkChat}
                 onTurnEnd={onTurnEnd}
                 theme={theme}
                 onToggleTheme={toggle}
@@ -2099,13 +2148,20 @@ function Shell({
                 hideHeader={false}
                 workspaceScope={activeWorkspaceScope}
                 workspaceDefaultScope={workspaces?.default_scope ?? null}
-                workspaceControls={workspaces?.controls ?? null}
+                workspaceControls={
+                  quickChatActive ? null : (workspaces?.controls ?? null)
+                }
                 workspaceScopeDisabled={activeChatRunning}
                 workspaceError={workspaceError}
                 onWorkspaceScopeChange={applyWorkspaceScope}
                 settingsSnapshot={settingsSnapshot}
                 onOpenModelSettings={onOpenModelSettings}
                 skills={skills}
+                allowConversationReset={!quickChatActive}
+                showSessionInfo={!quickChatActive}
+                emptyStateGreeting={
+                  quickChatActive ? t("quickChat.greeting") : undefined
+                }
               />
             </div>
             {view !== "chat" && (

webui/src/components/ChatList.tsx

@@ -1,7 +1,6 @@
 import {
   memo,
   useEffect,
-  useLayoutEffect,
   useMemo,
   useRef,
   useState,
@@ -25,6 +24,10 @@ import {
   DropdownMenuItem,
   DropdownMenuTrigger,
 } from "@/components/ui/dropdown-menu";
+import {
+  SIDEBAR_SELECTION_ITEM_CLASS,
+  SidebarSelectionHighlight,
+} from "@/components/SidebarSelectionHighlight";
 import { deriveTitle, relativeTime, visibleSessionPreview } from "@/lib/format";
 import {
   COLLAPSED_CHATS_VISIBLE_COUNT,
@@ -106,9 +109,6 @@ export const ChatList = memo(function ChatList({
   const [visibleLimit, setVisibleLimit] = useState(INITIAL_VISIBLE_SESSIONS);
   const listContentRef = useRef<HTMLDivElement>(null);
   const activeRowRef = useRef<HTMLDivElement>(null);
-  const activeHighlightRef = useRef<HTMLDivElement>(null);
-  const activeHighlightSurfaceRef = useRef<HTMLDivElement>(null);
-  const highlightVisibleRef = useRef(false);
   const labels = useMemo<ChatGroupLabels>(() => ({
     pinned: t("chat.groups.pinned"),
     all: t("chat.groups.all"),
@@ -163,74 +163,6 @@ export const ChatList = memo(function ChatList({
     setVisibleLimit(INITIAL_VISIBLE_SESSIONS);
   }, [showArchived, sort]);
 
-  useLayoutEffect(() => {
-    let resetTransitionFrame: number | null = null;
-
-    const updateHighlight = () => {
-      const content = listContentRef.current;
-      const row = activeRowRef.current;
-      const highlight = activeHighlightRef.current;
-      const surface = activeHighlightSurfaceRef.current;
-
-      if (!highlight || !surface) return;
-      if (!content || !row) {
-        surface.style.opacity = "0";
-        surface.style.transform = "scale(0.97)";
-        highlightVisibleRef.current = false;
-        return;
-      }
-
-      const shouldFloatIn = !highlightVisibleRef.current;
-      if (shouldFloatIn) {
-        highlight.style.transitionProperty = "none";
-      }
-
-      const contentRect = content.getBoundingClientRect();
-      const rowRect = row.getBoundingClientRect();
-      highlight.style.width = `${rowRect.width}px`;
-      highlight.style.height = `${rowRect.height}px`;
-      highlight.style.transform = `translate3d(${rowRect.left - contentRect.left}px, ${
-        rowRect.top - contentRect.top
-      }px, 0)`;
-
-      if (shouldFloatIn) {
-        void highlight.offsetWidth;
-      }
-
-      surface.style.opacity = "1";
-      surface.style.transform = "scale(1)";
-      highlightVisibleRef.current = true;
-
-      if (shouldFloatIn) {
-        resetTransitionFrame = window.requestAnimationFrame(() => {
-          highlight.style.removeProperty("transition-property");
-          resetTransitionFrame = null;
-        });
-      }
-    };
-
-    updateHighlight();
-
-    const resizeObserver =
-      typeof ResizeObserver === "undefined"
-        ? null
-        : new ResizeObserver(updateHighlight);
-    if (resizeObserver) {
-      if (listContentRef.current) resizeObserver.observe(listContentRef.current);
-      if (activeRowRef.current) resizeObserver.observe(activeRowRef.current);
-    }
-    window.addEventListener("resize", updateHighlight);
-
-    return () => {
-      if (resetTransitionFrame !== null) {
-        window.cancelAnimationFrame(resetTransitionFrame);
-      }
-      activeHighlightRef.current?.style.removeProperty("transition-property");
-      resizeObserver?.disconnect();
-      window.removeEventListener("resize", updateHighlight);
-    };
-  }, [activeKey, density, limitedGroups, showPreviews, showTimestamps]);
-
   if (loading && sessions.length === 0) {
     return (
       <div className="px-3 py-6 text-[12px] text-muted-foreground">
@@ -333,7 +265,8 @@ export const ChatList = memo(function ChatList({
                           ref={active ? activeRowRef : undefined}
                           data-chat-row={s.key}
                           className={cn(
-                            "group flex min-w-0 max-w-full items-center gap-2 rounded-xl px-2 text-[13px] transition-colors",
+                            "group flex min-w-0 max-w-full items-center gap-2 rounded-xl px-2 text-[13px]",
+                            SIDEBAR_SELECTION_ITEM_CLASS,
                             compact ? "min-h-7" : "min-h-8",
                             active
                               ? "text-sidebar-accent-foreground"
@@ -475,18 +408,12 @@ export const ChatList = memo(function ChatList({
             </button>
           </div>
         ) : null}
-        <div
+        <SidebarSelectionHighlight
-          ref={activeHighlightRef}
+          containerRef={listContentRef}
-          data-testid="active-chat-highlight"
+          targetRef={activeRowRef}
-          aria-hidden="true"
+          activeId={activeKey}
-          className="pointer-events-none absolute left-0 top-0 z-0 !mt-0 transition-[transform,width,height] duration-300 ease-out will-change-transform motion-reduce:transition-none"
+          scope="sessions"
-        >
+        />
-          <div
-            ref={activeHighlightSurfaceRef}
-            data-testid="active-chat-highlight-surface"
-            className="h-full w-full scale-[0.97] rounded-xl bg-sidebar-foreground/[0.055] opacity-0 transition-[opacity,transform] duration-200 ease-out motion-reduce:transition-none dark:bg-white/[0.07]"
-          />
-        </div>
       </div>
     </div>
   );

webui/src/components/Sidebar.tsx

@@ -1,8 +1,14 @@
-import { useState, type ReactNode } from "react";
+import {
+  type ReactNode,
+  type RefObject,
+  useRef,
+  useState,
+} from "react";
 import {
   Archive,
   Brain,
   CalendarClock,
+  MessageCircle,
   Menu,
   Search,
   Settings,
@@ -13,6 +19,10 @@ import { useTranslation } from "react-i18next";
 
 import { ChatList } from "@/components/ChatList";
 import { ConnectionBadge } from "@/components/ConnectionBadge";
+import {
+  SIDEBAR_SELECTION_ACTION_ITEM_CLASS,
+  SidebarSelectionHighlight,
+} from "@/components/SidebarSelectionHighlight";
 import { Button } from "@/components/ui/button";
 import type {
   ChatSummary,
@@ -24,6 +34,9 @@ interface SidebarProps {
   sessions: ChatSummary[];
   activeKey: string | null;
   loading: boolean;
+  quickChatActive: boolean;
+  newChatActive: boolean;
+  onOpenQuickChat: () => void;
   onNewChat: () => void;
   onSelect: (key: string) => void;
   onRequestDelete: (key: string, label: string) => void;
@@ -82,6 +95,15 @@ export function Sidebar(props: SidebarProps) {
   const collapsed = Boolean(props.collapsed);
   const toggleLabel = t("thread.header.toggleSidebar");
   const newChatShortcut = newChatShortcutLabel();
+  const actionListRef = useRef<HTMLDivElement>(null);
+  const activeActionRef = useRef<HTMLButtonElement>(null);
+  const activeActionId = props.quickChatActive
+    ? "quick-chat"
+    : props.newChatActive
+      ? "new-chat"
+      : props.activeUtility
+        ? `utility:${props.activeUtility}`
+        : null;
 
   return (
     <nav
@@ -134,15 +156,26 @@ export function Sidebar(props: SidebarProps) {
       </div>
 
       <div
+        ref={actionListRef}
         className={cn(
-          "space-y-1.5 px-2 pb-2",
+          "relative space-y-1.5 px-2 pb-2",
           collapsed && "flex w-14 flex-col items-center px-0",
         )}
       >
+        <SidebarActionButton
+          collapsed={collapsed}
+          label={t("sidebar.quickChat")}
+          onClick={props.onOpenQuickChat}
+          active={props.quickChatActive}
+          selectionRef={activeActionRef}
+          icon={<MessageCircle className="h-4 w-4" />}
+        />
         <SidebarActionButton
           collapsed={collapsed}
           label={t("sidebar.newChat")}
           onClick={props.onNewChat}
+          active={props.newChatActive}
+          selectionRef={activeActionRef}
           icon={<SquarePen className="h-4 w-4" />}
           shortcut={newChatShortcut}
           ariaKeyShortcuts="Meta+Shift+O Control+Shift+O"
@@ -159,6 +192,7 @@ export function Sidebar(props: SidebarProps) {
           onClick={props.onOpenApps}
           onIntent={props.onSettingsIntent}
           active={props.activeUtility === "apps"}
+          selectionRef={activeActionRef}
           icon={<Blocks className="h-4 w-4" />}
         />
         <SidebarActionButton
@@ -167,6 +201,7 @@ export function Sidebar(props: SidebarProps) {
           onClick={props.onOpenSkills}
           onIntent={props.onSettingsIntent}
           active={props.activeUtility === "skills"}
+          selectionRef={activeActionRef}
           icon={<Brain className="h-4 w-4" />}
         />
         <SidebarActionButton
@@ -175,6 +210,7 @@ export function Sidebar(props: SidebarProps) {
           onClick={props.onOpenAutomations}
           onIntent={props.onSettingsIntent}
           active={props.activeUtility === "automations"}
+          selectionRef={activeActionRef}
           icon={<CalendarClock className="h-4 w-4" />}
         />
         {props.archivedCount ? (
@@ -185,6 +221,12 @@ export function Sidebar(props: SidebarProps) {
             icon={<Archive className="h-4 w-4" />}
           />
         ) : null}
+        <SidebarSelectionHighlight
+          containerRef={actionListRef}
+          targetRef={activeActionRef}
+          activeId={activeActionId}
+          scope="actions"
+        />
       </div>
       <div
         className={cn(
@@ -255,6 +297,7 @@ function SidebarActionButton({
   shortcut,
   ariaKeyShortcuts,
   onIntent,
+  selectionRef,
 }: {
   collapsed: boolean;
   label: string;
@@ -265,13 +308,15 @@ function SidebarActionButton({
   shortcut?: string;
   ariaKeyShortcuts?: string;
   onIntent?: () => void;
+  selectionRef?: RefObject<HTMLButtonElement>;
 }) {
   const title = shortcut ? `${label} (${shortcut})` : collapsed ? label : undefined;
 
   return (
     <Button
+      ref={active ? selectionRef : undefined}
       type="button"
-      variant="ghost"
+      variant={null}
       aria-label={label}
       aria-current={active ? "page" : undefined}
       aria-keyshortcuts={ariaKeyShortcuts}
@@ -280,12 +325,14 @@ function SidebarActionButton({
       onFocus={onIntent}
       onPointerEnter={onIntent}
       className={cn(
-        "touch-target group h-8 min-w-0 gap-2 overflow-hidden rounded-full font-medium text-sidebar-foreground/85 hover:bg-sidebar-accent/75 hover:text-sidebar-foreground",
+        "touch-target group h-8 min-w-0 gap-2 overflow-hidden rounded-xl font-medium",
-        "transition-[width,padding,border-radius,color,background-color] duration-300 ease-out",
+        SIDEBAR_SELECTION_ACTION_ITEM_CLASS,
         collapsed
-          ? "w-9 justify-center gap-0 rounded-xl px-0"
+          ? "w-9 justify-center gap-0 px-0"
           : "w-full justify-start gap-2 px-3 text-[12.5px]",
-        active && "bg-sidebar-accent text-sidebar-foreground shadow-[inset_0_0_0_1px_hsl(var(--sidebar-border)/0.55)]",
+        active
+          ? "text-sidebar-accent-foreground"
+          : "text-sidebar-foreground/85 hover:bg-sidebar-foreground/[0.035] hover:text-sidebar-foreground dark:hover:bg-white/[0.05]",
         className,
       )}
     >

webui/src/components/SidebarSelectionHighlight.tsx

@@ -0,0 +1,90 @@
+import {
+  type RefObject,
+  useLayoutEffect,
+  useRef,
+} from "react";
+
+interface SidebarSelectionHighlightProps {
+  containerRef: RefObject<HTMLElement>;
+  targetRef: RefObject<HTMLElement>;
+  activeId: string | null;
+  scope: string;
+}
+
+export const SIDEBAR_SELECTION_ITEM_CLASS =
+  "relative z-[1] transition-[color] duration-150 ease-out motion-reduce:transition-none";
+
+export const SIDEBAR_SELECTION_ACTION_ITEM_CLASS =
+  "relative z-[1] transition-[width,padding,color] [transition-duration:300ms,300ms,150ms] ease-out motion-reduce:transition-none";
+
+export function SidebarSelectionHighlight({
+  containerRef,
+  targetRef,
+  activeId,
+  scope,
+}: SidebarSelectionHighlightProps) {
+  const highlightRef = useRef<HTMLDivElement>(null);
+  const positionedRef = useRef(false);
+
+  useLayoutEffect(() => {
+    const highlight = highlightRef.current;
+    const container = containerRef.current;
+    const target = targetRef.current;
+    let restoreTransitionFrame: number | null = null;
+
+    const position = () => {
+      if (!highlight) return;
+      if (!activeId || !container || !target) {
+        highlight.style.opacity = "0";
+        positionedRef.current = false;
+        return;
+      }
+
+      const firstPosition = !positionedRef.current;
+      if (firstPosition) highlight.style.transitionProperty = "none";
+
+      const containerRect = container.getBoundingClientRect();
+      const targetRect = target.getBoundingClientRect();
+      highlight.style.width = `${targetRect.width}px`;
+      highlight.style.height = `${targetRect.height}px`;
+      highlight.style.transform = `translate3d(${targetRect.left - containerRect.left}px, ${
+        targetRect.top - containerRect.top
+      }px, 0)`;
+      highlight.style.opacity = "1";
+      positionedRef.current = true;
+
+      if (firstPosition) {
+        restoreTransitionFrame = window.requestAnimationFrame(() => {
+          highlight.style.removeProperty("transition-property");
+          restoreTransitionFrame = null;
+        });
+      }
+    };
+
+    position();
+    const resizeObserver =
+      typeof ResizeObserver === "undefined" ? null : new ResizeObserver(position);
+    if (container) resizeObserver?.observe(container);
+    if (target) resizeObserver?.observe(target);
+    window.addEventListener("resize", position);
+
+    return () => {
+      if (restoreTransitionFrame !== null) {
+        window.cancelAnimationFrame(restoreTransitionFrame);
+      }
+      highlight?.style.removeProperty("transition-property");
+      resizeObserver?.disconnect();
+      window.removeEventListener("resize", position);
+    };
+  });
+
+  return (
+    <div
+      ref={highlightRef}
+      data-testid={`${scope}-selection-highlight`}
+      data-active-id={activeId ?? undefined}
+      aria-hidden="true"
+      className="pointer-events-none absolute left-0 top-0 z-0 !mt-0 rounded-xl bg-sidebar-foreground/[0.055] opacity-0 transition-[transform,width,height] duration-300 ease-out will-change-transform motion-reduce:transition-none dark:bg-white/[0.07]"
+    />
+  );
+}

webui/src/components/settings/SettingsView.tsx

@@ -65,6 +65,10 @@ import { useTranslation } from "react-i18next";
 
 import { channelUiPresentation } from "@/channel-plugins/registry";
 import { LanguageSwitcher } from "@/components/LanguageSwitcher";
+import {
+  SIDEBAR_SELECTION_ITEM_CLASS,
+  SidebarSelectionHighlight,
+} from "@/components/SidebarSelectionHighlight";
 import { SkillsCatalogSettings } from "@/components/settings/SkillsCatalogSettings";
 import { TokenUsageHeatmap } from "@/components/settings/TokenUsageHeatmap";
 import { ToggleButton } from "@/components/settings/ToggleButton";
@@ -2497,6 +2501,8 @@ function SettingsSidebar({
   hostChromeInset?: boolean;
 }) {
   const { t } = useTranslation();
+  const desktopNavRef = useRef<HTMLDivElement>(null);
+  const activeNavItemRef = useRef<HTMLButtonElement>(null);
   const activeItem = SETTINGS_NAV_ITEMS.find((item) => item.key === activeSection)
     ?? SETTINGS_NAV_ITEMS[0];
   const ActiveIcon = activeItem.icon;
@@ -2569,19 +2575,21 @@ function SettingsSidebar({
           </DropdownMenuContent>
         </DropdownMenu>
 
-        <div className="hidden space-y-1 lg:block">
+        <div ref={desktopNavRef} className="relative hidden space-y-1 lg:block">
           {SETTINGS_NAV_ITEMS.map(({ key, icon: Icon, fallback }) => {
             const active = key === activeSection;
             return (
               <button
+                ref={active ? activeNavItemRef : undefined}
                 key={key}
                 type="button"
                 aria-current={active ? "page" : undefined}
                 onClick={() => onSelectSection(key)}
                 className={cn(
-                  "touch-target flex h-9 w-full items-center gap-2 rounded-[10px] px-2.5 text-left text-[13px] font-medium transition-colors",
+                  "touch-target flex h-9 w-full items-center gap-2 rounded-xl px-2.5 text-left text-[13px] font-medium",
+                  SIDEBAR_SELECTION_ITEM_CLASS,
                   active
-                    ? "bg-sidebar-accent text-foreground"
+                    ? "text-sidebar-accent-foreground"
                     : "text-muted-foreground/78 hover:bg-muted/45 hover:text-foreground",
                 )}
               >
@@ -2592,6 +2600,12 @@ function SettingsSidebar({
               </button>
             );
           })}
+          <SidebarSelectionHighlight
+            containerRef={desktopNavRef}
+            targetRef={activeNavItemRef}
+            activeId={activeSection}
+            scope="settings"
+          />
         </div>
       </nav>
 

webui/src/components/thread/ThreadShell.tsx

@@ -315,6 +315,9 @@ interface ThreadShellProps {
   settingsSnapshot?: SettingsPayload | null;
   onOpenModelSettings?: () => void;
   skills?: SkillSummary[];
+  allowConversationReset?: boolean;
+  showSessionInfo?: boolean;
+  emptyStateGreeting?: string;
 }
 
 function toModelBadgeLabel(modelName: string | null): string | null {
@@ -597,6 +600,9 @@ export function ThreadShell({
   settingsSnapshot = null,
   onOpenModelSettings,
   skills = [],
+  allowConversationReset = true,
+  showSessionInfo = true,
+  emptyStateGreeting,
 }: ThreadShellProps) {
   const { t } = useTranslation();
   const chatId = session?.chatId ?? null;
@@ -622,6 +628,12 @@ export function ThreadShell({
   const [fallbackModelName, setFallbackModelName] = useState<string | null>(null);
   const [booting, setBooting] = useState(false);
   const [slashCommands, setSlashCommands] = useState<SlashCommand[]>([]);
+  const availableSlashCommands = useMemo(
+    () => allowConversationReset
+      ? slashCommands
+      : slashCommands.filter((command) => command.command !== "/new"),
+    [allowConversationReset, slashCommands],
+  );
   const cliApps = useInstalledSettingItems({
     getToken,
     eventName: CLI_APPS_CHANGED_EVENT,
@@ -1374,7 +1386,7 @@ export function ThreadShell({
           fallbackModelName={fallbackModelName}
           onModelBadgeClick={modelBadge.needsSetup ? onOpenModelSettings : undefined}
           variant={showHeroComposer ? "hero" : "thread"}
-          slashCommands={slashCommands}
+          slashCommands={availableSlashCommands}
           cliApps={cliApps}
           mcpPresets={mcpPresets}
           skills={skills}
@@ -1416,7 +1428,7 @@ export function ThreadShell({
           fallbackModelName={fallbackModelName}
           onModelBadgeClick={modelBadge.needsSetup ? onOpenModelSettings : undefined}
           variant="hero"
-          slashCommands={slashCommands}
+          slashCommands={availableSlashCommands}
           cliApps={cliApps}
           mcpPresets={mcpPresets}
           skills={skills}
@@ -1442,10 +1454,10 @@ export function ThreadShell({
     </div>
   ) : (
     <div className="flex w-full flex-col items-center text-center animate-in fade-in-0 slide-in-from-bottom-2 duration-500">
-      <HeroGreeting text={t(heroGreetingKey)} />
+      <HeroGreeting text={emptyStateGreeting ?? t(heroGreetingKey)} />
     </div>
   );
-  const sessionInfoAction = historyKey ? (
+  const sessionInfoAction = historyKey && showSessionInfo ? (
     <SessionInfoPopover sessionKey={historyKey} token={token} title={title} />
   ) : undefined;
   const promptNavigatorAction = historyKey ? (
@@ -1488,7 +1500,7 @@ export function ThreadShell({
             showScrollToBottomButton={!!session}
             cliApps={cliApps}
             mcpPresets={mcpPresets}
-            slashCommands={slashCommands}
+            slashCommands={availableSlashCommands}
             forkBoundaryMessageCount={forkBoundaryMessageCount}
             hasMoreBefore={hasMoreBefore}
             loadingOlder={loadingOlder}

webui/src/hooks/useNanobotStream.ts

@@ -37,7 +37,7 @@ interface ActiveAssistantCursor {
 }
 
 type PendingStreamEvent =
-  | { kind: "delta"; text: string; turn: UIMessageTurnFields }
+  | { kind: "delta"; text: string; turn: UIMessageTurnFields; source?: UIMessage["source"] }
   | { kind: "reasoning"; text: string; turn: UIMessageTurnFields };
 
 type UIMessageTurnFields = Pick<UIMessage, "turnId" | "turnPhase" | "turnSeq">;
@@ -778,7 +778,12 @@ export function useNanobotStream(
   }, []);
 
   const appendAnswerChunk = useCallback(
-    (prev: UIMessage[], chunk: string, turn: UIMessageTurnFields = {}): UIMessage[] => {
+    (
+      prev: UIMessage[],
+      chunk: string,
+      turn: UIMessageTurnFields = {},
+      source?: UIMessage["source"],
+    ): UIMessage[] => {
       let next = prev;
       let targetIndex = resolveActiveAssistantIndex(next, turn);
 
@@ -809,6 +814,7 @@ export function useNanobotStream(
         content: target.content + chunk,
         isStreaming: true,
         ...turn,
+        ...(source ? { source } : {}),
       };
       closedAssistantStreamIdsRef.current.delete(merged.id);
       activeAssistantRef.current = { id: merged.id, index: targetIndex };
@@ -823,7 +829,7 @@ export function useNanobotStream(
       let next = prev;
       for (const event of events) {
         if (event.kind === "delta") {
-          next = appendAnswerChunk(next, event.text, event.turn);
+          next = appendAnswerChunk(next, event.text, event.turn, event.source);
         } else {
           if (closeActiveAssistantStream()) clearActivitySegment();
           next = attachReasoningChunk(
@@ -843,6 +849,7 @@ export function useNanobotStream(
     closeAnswerSegment?: boolean;
     finalAnswerText?: string;
     turn?: UIMessageTurnFields;
+    source?: UIMessage["source"];
   }) => {
     if (streamFrameRef.current !== null) {
       window.cancelAnimationFrame(streamFrameRef.current);
@@ -855,7 +862,8 @@ export function useNanobotStream(
     const events = pendingStreamEventsRef.current;
     const finalAnswerText = options?.finalAnswerText;
     const turn = options?.turn ?? {};
-    if (events.length === 0 && finalAnswerText === undefined) {
+    const source = options?.source;
+    if (events.length === 0 && finalAnswerText === undefined && source === undefined) {
       if (options?.closeAnswerSegment) closeActiveAssistantStream();
       return;
     }
@@ -873,6 +881,7 @@ export function useNanobotStream(
             content: finalAnswerText,
             isStreaming: true,
             ...turn,
+            ...(source ? { source } : {}),
           };
           next = replaceMessageAt(next, targetIndex, merged);
           if (!options?.closeAnswerSegment) {
@@ -890,6 +899,7 @@ export function useNanobotStream(
               content: finalAnswerText,
               isStreaming: true,
               ...turn,
+              ...(source ? { source } : {}),
               createdAt: Date.now(),
             },
           ];
@@ -900,6 +910,18 @@ export function useNanobotStream(
             buffer.current = { messageId: id };
           }
         }
+      } else if (source) {
+        const targetIndex =
+          resolveActiveAssistantIndex(next, turn)
+          ?? findStreamingAssistantIndex(next, closedAssistantStreamIdsRef.current, turn);
+        if (targetIndex !== null) {
+          const target = next[targetIndex];
+          next = replaceMessageAt(next, targetIndex, {
+            ...target,
+            ...turn,
+            source,
+          });
+        }
       }
       if (options?.closeAnswerSegment) closeActiveAssistantStream();
       return next;
@@ -1027,6 +1049,7 @@ export function useNanobotStream(
           kind: "delta",
           text: chunk,
           turn: turnFieldsFromEvent(ev, "answer"),
+          source: ev.source,
         });
         schedulePendingStreamFlush();
         return;
@@ -1054,6 +1077,7 @@ export function useNanobotStream(
           closeAnswerSegment: !mergeNext,
           ...(typeof ev.text === "string" ? { finalAnswerText: ev.text } : {}),
           turn,
+          source: ev.source,
         });
         if (suppressStreamUntilTurnEndRef.current) return;
         if (ev.resuming) {

webui/src/i18n/locales/en/common.json

@@ -43,6 +43,7 @@
   "sidebar": {
     "navigation": "Sidebar navigation",
     "collapse": "Collapse sidebar",
+    "quickChat": "Quick Chat",
     "newChat": "New topic",
     "searchAria": "Search",
     "searchPlaceholder": "Search",
@@ -60,6 +61,9 @@
       "title": "Skills"
     }
   },
+  "quickChat": {
+    "greeting": "What's on your mind?"
+  },
   "settings": {
     "backToChat": "Back to chat",
     "sidebar": {

webui/src/i18n/locales/es/common.json

@@ -43,6 +43,7 @@
   "sidebar": {
     "navigation": "Navegación de la barra lateral",
     "collapse": "Contraer barra lateral",
+    "quickChat": "Chat rápido",
     "newChat": "Nuevo tema",
     "searchAria": "Buscar",
     "searchPlaceholder": "Buscar",
@@ -60,6 +61,9 @@
       "title": "Habilidades"
     }
   },
+  "quickChat": {
+    "greeting": "¿Qué tienes en mente?"
+  },
   "settings": {
     "backToChat": "Volver al chat",
     "sidebar": {

webui/src/i18n/locales/fr/common.json

@@ -43,6 +43,7 @@
   "sidebar": {
     "navigation": "Navigation de la barre latérale",
     "collapse": "Réduire la barre latérale",
+    "quickChat": "Discussion rapide",
     "newChat": "Nouveau sujet",
     "searchAria": "Rechercher",
     "searchPlaceholder": "Rechercher",
@@ -60,6 +61,9 @@
       "title": "Compétences"
     }
   },
+  "quickChat": {
+    "greeting": "De quoi avez-vous envie de parler ?"
+  },
   "settings": {
     "backToChat": "Retour au chat",
     "sidebar": {

webui/src/i18n/locales/id/common.json

@@ -43,6 +43,7 @@
   "sidebar": {
     "navigation": "Navigasi bilah samping",
     "collapse": "Ciutkan sidebar",
+    "quickChat": "Obrolan cepat",
     "newChat": "Topik baru",
     "searchAria": "Cari",
     "searchPlaceholder": "Cari",
@@ -60,6 +61,9 @@
       "title": "Skill"
     }
   },
+  "quickChat": {
+    "greeting": "Apa yang sedang kamu pikirkan?"
+  },
   "settings": {
     "backToChat": "Kembali ke chat",
     "sidebar": {

webui/src/i18n/locales/ja/common.json

@@ -43,6 +43,7 @@
   "sidebar": {
     "navigation": "サイドバーのナビゲーション",
     "collapse": "サイドバーを閉じる",
+    "quickChat": "クイックチャット",
     "newChat": "新しいトピック",
     "searchAria": "検索",
     "searchPlaceholder": "検索",
@@ -60,6 +61,9 @@
       "title": "スキル"
     }
   },
+  "quickChat": {
+    "greeting": "何について話しますか？"
+  },
   "settings": {
     "backToChat": "チャットに戻る",
     "sidebar": {

webui/src/i18n/locales/ko/common.json

@@ -43,6 +43,7 @@
   "sidebar": {
     "navigation": "사이드바 탐색",
     "collapse": "사이드바 접기",
+    "quickChat": "빠른 채팅",
     "newChat": "새 주제",
     "searchAria": "검색",
     "searchPlaceholder": "검색",
@@ -60,6 +61,9 @@
       "title": "스킬"
     }
   },
+  "quickChat": {
+    "greeting": "무슨 이야기를 나눠볼까요?"
+  },
   "settings": {
     "backToChat": "채팅으로 돌아가기",
     "sidebar": {

webui/src/i18n/locales/pt-BR/common.json

@@ -43,6 +43,7 @@
   "sidebar": {
     "navigation": "Navegação da barra lateral",
     "collapse": "Recolher barra lateral",
+    "quickChat": "Chat rápido",
     "newChat": "Novo tópico",
     "searchAria": "Buscar",
     "searchPlaceholder": "Buscar",
@@ -60,6 +61,9 @@
       "title": "Skills"
     }
   },
+  "quickChat": {
+    "greeting": "O que você está pensando?"
+  },
   "settings": {
     "backToChat": "Voltar para a conversa",
     "sidebar": {

webui/src/i18n/locales/vi/common.json

@@ -43,6 +43,7 @@
   "sidebar": {
     "navigation": "Điều hướng thanh bên",
     "collapse": "Thu gọn thanh bên",
+    "quickChat": "Trò chuyện nhanh",
     "newChat": "Chủ đề mới",
     "searchAria": "Tìm kiếm",
     "searchPlaceholder": "Tìm kiếm",
@@ -60,6 +61,9 @@
       "title": "Kỹ năng"
     }
   },
+  "quickChat": {
+    "greeting": "Bạn đang nghĩ gì?"
+  },
   "settings": {
     "backToChat": "Quay lại chat",
     "sidebar": {

webui/src/i18n/locales/zh-CN/common.json

@@ -43,6 +43,7 @@
   "sidebar": {
     "navigation": "侧边栏导航",
     "collapse": "收起侧边栏",
+    "quickChat": "随便聊聊",
     "newChat": "新建话题",
     "searchAria": "搜索",
     "searchPlaceholder": "搜索",
@@ -60,6 +61,9 @@
       "title": "技能"
     }
   },
+  "quickChat": {
+    "greeting": "想聊点什么？"
+  },
   "settings": {
     "backToChat": "返回聊天",
     "sidebar": {

webui/src/i18n/locales/zh-TW/common.json

@@ -43,6 +43,7 @@
   "sidebar": {
     "navigation": "側邊欄導覽",
     "collapse": "收合側邊欄",
+    "quickChat": "輕鬆聊聊",
     "newChat": "新增話題",
     "searchAria": "搜尋",
     "searchPlaceholder": "搜尋",
@@ -60,6 +61,9 @@
       "title": "技能"
     }
   },
+  "quickChat": {
+    "greeting": "想聊點什麼？"
+  },
   "settings": {
     "backToChat": "返回聊天",
     "sidebar": {

webui/src/lib/quick-chat.ts

@@ -0,0 +1,22 @@
+import type { ChatSummary } from "@/lib/types";
+
+export const QUICK_CHAT_ID = "quick-chat";
+export const QUICK_CHAT_KEY = `websocket:${QUICK_CHAT_ID}`;
+
+export function isQuickChatKey(key: string | null): boolean {
+  return key === QUICK_CHAT_KEY;
+}
+
+export function quickChatSession(persisted?: ChatSummary): ChatSummary {
+  return {
+    key: QUICK_CHAT_KEY,
+    channel: "websocket",
+    chatId: QUICK_CHAT_ID,
+    createdAt: persisted?.createdAt ?? null,
+    updatedAt: persisted?.updatedAt ?? null,
+    preview: persisted?.preview ?? "",
+    modelPreset: persisted?.modelPreset ?? null,
+    runStartedAt: persisted?.runStartedAt ?? null,
+    workspaceScope: persisted?.workspaceScope ?? null,
+  };
+}

webui/src/lib/types.ts

@@ -1185,12 +1185,16 @@ export type InboundEvent =
       chat_id: string;
       text: string;
       stream_id?: string;
+      /** Lightweight provenance for proactive streamed assistant messages. */
+      source?: UIMessageSource;
     } & InboundTurnMetadata)
   | ({
       event: "stream_end";
       chat_id: string;
       stream_id?: string;
       text?: string;
+      /** Lightweight provenance for proactive streamed assistant messages. */
+      source?: UIMessageSource;
       /** This answer segment ended, but the active agent turn will continue. */
       resuming?: boolean;
       /** The next answer segment continues this same assistant message. */

webui/src/tests/app-layout.test.tsx

@@ -349,6 +349,107 @@ describe("App layout", () => {
     ).toBeTruthy();
   });
 
+  it("opens a single fixed Quick Chat without provisioning a new session", async () => {
+    render(<App />);
+
+    await waitFor(() => expect(connectSpy).toHaveBeenCalled());
+    const sidebar = screen.getByRole("navigation", { name: "Sidebar navigation" });
+    const quickChatButton = within(sidebar).getByRole("button", {
+      name: "Quick Chat",
+    });
+    const newTopicButton = within(sidebar).getByRole("button", {
+      name: "New topic",
+    });
+    const actionHighlight = within(sidebar).getByTestId(
+      "actions-selection-highlight",
+    );
+
+    fireEvent.click(quickChatButton);
+
+    expect(window.location.hash).toBe("#/quick-chat");
+    expect(quickChatButton).toHaveAttribute("aria-current", "page");
+    expect(newTopicButton).not.toHaveAttribute("aria-current");
+    expect(quickChatButton).not.toHaveClass("bg-sidebar-accent");
+    expect(quickChatButton).toHaveClass("transition-[width,padding,color]");
+    expect(actionHighlight).toHaveAttribute("data-active-id", "quick-chat");
+    expect(
+      within(sidebar).queryByTestId("actions-selection-highlight-surface"),
+    ).not.toBeInTheDocument();
+    await waitFor(() =>
+      expect(fetch).toHaveBeenCalledWith(
+        expect.stringContaining(
+          "/api/sessions/websocket%3Aquick-chat/webui-thread",
+        ),
+        expect.anything(),
+      ),
+    );
+    expect(createChatSpy).not.toHaveBeenCalled();
+    expect(document.title).toBe("Quick Chat · nanobot");
+    expect(screen.getByText("What's on your mind?")).toBeInTheDocument();
+
+    fireEvent.click(newTopicButton);
+
+    expect(window.location.hash).toBe("#/new");
+    expect(newTopicButton).toHaveAttribute("aria-current", "page");
+    expect(quickChatButton).not.toHaveAttribute("aria-current");
+    expect(actionHighlight).toHaveAttribute("data-active-id", "new-chat");
+    expect(within(sidebar).queryAllByRole("button", { current: "page" })).toHaveLength(1);
+  });
+
+  it("restores Quick Chat before it has a persisted session", async () => {
+    window.history.replaceState(null, "", "/#/quick-chat");
+
+    render(<App />);
+
+    await waitFor(() => expect(connectSpy).toHaveBeenCalled());
+    expect(window.location.hash).toBe("#/quick-chat");
+    await waitFor(() =>
+      expect(fetch).toHaveBeenCalledWith(
+        expect.stringContaining(
+          "/api/sessions/websocket%3Aquick-chat/webui-thread",
+        ),
+        expect.anything(),
+      ),
+    );
+    expect(
+      within(screen.getByRole("navigation", { name: "Sidebar navigation" }))
+        .getByRole("button", { name: "Quick Chat" }),
+    ).toHaveAttribute("aria-current", "page");
+  });
+
+  it("keeps persisted Quick Chat out of the topic list and topic search", async () => {
+    mockSessions = [
+      {
+        key: "websocket:quick-chat",
+        channel: "websocket",
+        chatId: "quick-chat",
+        createdAt: "2026-07-30T08:00:00Z",
+        updatedAt: "2026-07-30T08:05:00Z",
+        preview: "A private casual message",
+      },
+      {
+        key: "websocket:project-chat",
+        channel: "websocket",
+        chatId: "project-chat",
+        createdAt: "2026-07-30T08:00:00Z",
+        updatedAt: "2026-07-30T08:05:00Z",
+        preview: "Project roadmap",
+      },
+    ];
+
+    render(<App />);
+
+    await waitFor(() => expect(connectSpy).toHaveBeenCalled());
+    const sidebar = screen.getByRole("navigation", { name: "Sidebar navigation" });
+    expect(within(sidebar).getByText("Project roadmap")).toBeInTheDocument();
+    expect(within(sidebar).queryByText("A private casual message")).not.toBeInTheDocument();
+
+    fireEvent.click(within(sidebar).getByRole("button", { name: "Search" }));
+    const dialog = await screen.findByRole("dialog", { name: "Search" });
+    expect(within(dialog).getByText("Project roadmap")).toBeInTheDocument();
+    expect(within(dialog).queryByText("A private casual message")).not.toBeInTheDocument();
+  });
+
   it("restores the Settings route after a restart fallback hash", async () => {
     localStorage.setItem("nanobot-webui.restartStartedAt", String(Date.now()));
     localStorage.setItem("nanobot-webui.restartRoute", "#/settings?section=channels");
@@ -2128,16 +2229,41 @@ describe("App layout", () => {
     expect(window.location.hash).toBe("#/settings");
 
     const settingsNav = screen.getByRole("navigation", { name: "Settings sections" });
-    fireEvent.click(within(settingsNav).getByRole("button", { name: "Models" }));
+    const overviewButton = within(settingsNav).getByRole("button", {
+      name: "Overview",
+      exact: true,
+    });
+    const modelsButton = within(settingsNav).getByRole("button", {
+      name: "Models",
+      exact: true,
+    });
+    const settingsHighlight = within(settingsNav).getByTestId(
+      "settings-selection-highlight",
+    );
+
+    expect(overviewButton).toHaveAttribute("aria-current", "page");
+    expect(overviewButton).not.toHaveClass("bg-sidebar-accent");
+    expect(overviewButton).toHaveClass("transition-[color]");
+    expect(settingsHighlight).toHaveAttribute("data-active-id", "overview");
+
+    fireEvent.click(modelsButton);
 
     expect(await screen.findByText("Model presets")).toBeInTheDocument();
     expect(screen.queryByRole("heading", { name: "Models" })).not.toBeInTheDocument();
     expect(window.location.hash).toBe("#/settings?section=models");
+    expect(modelsButton).toHaveAttribute("aria-current", "page");
+    expect(settingsHighlight).toHaveAttribute("data-active-id", "models");
 
-    fireEvent.click(within(settingsNav).getByRole("button", { name: "Voice" }));
+    const voiceButton = within(settingsNav).getByRole("button", {
+      name: "Voice",
+      exact: true,
+    });
+    fireEvent.click(voiceButton);
 
     expect(await screen.findByRole("heading", { name: "Voice input" })).toBeInTheDocument();
     expect(window.location.hash).toBe("#/settings?section=voice");
+    expect(voiceButton).toHaveAttribute("aria-current", "page");
+    expect(settingsHighlight).toHaveAttribute("data-active-id", "voice");
   });
 
   it("transitions between Apps and Skills without replacing the sidebar", async () => {
@@ -2163,6 +2289,11 @@ describe("App layout", () => {
       "aria-current",
       "page",
     );
+    expect(within(sidebar).getByTestId("actions-selection-highlight")).toHaveAttribute(
+      "data-active-id",
+      "utility:apps",
+    );
+    expect(within(sidebar).queryAllByRole("button", { current: "page" })).toHaveLength(1);
     expect(screen.getByTestId("settings-section-transition")).toHaveAttribute(
       "data-settings-section",
       "apps",
@@ -2190,6 +2321,10 @@ describe("App layout", () => {
       "aria-current",
       "page",
     );
+    expect(within(sidebar).getByTestId("actions-selection-highlight")).toHaveAttribute(
+      "data-active-id",
+      "utility:skills",
+    );
     expect(document.title).toBe("Skills · nanobot");
   });
 

webui/src/tests/chat-list.test.tsx

@@ -220,7 +220,7 @@ describe("ChatList", () => {
     expect(within(chatsSection).queryByText("Project chat")).not.toBeInTheDocument();
   });
 
-  it("floats a borderless highlight in, then slides it between selected topics", () => {
+  it("positions one background highlight, then slides it between selected topics", () => {
     let revealFrame: FrameRequestCallback | null = null;
     vi.spyOn(window, "requestAnimationFrame").mockImplementation((callback) => {
       revealFrame = callback;
@@ -259,14 +259,15 @@ describe("ChatList", () => {
       />,
     );
 
-    const highlight = screen.getByTestId("active-chat-highlight");
+    const highlight = screen.getByTestId("sessions-selection-highlight");
-    const surface = screen.getByTestId("active-chat-highlight-surface");
+    expect(highlight).toHaveClass(
-    expect(surface).toHaveClass(
       "bg-sidebar-foreground/[0.055]",
-      "transition-[opacity,transform]",
+      "transition-[transform,width,height]",
       "motion-reduce:transition-none",
     );
-    expect(surface).toHaveStyle("opacity: 0; transform: scale(0.97)");
+    expect(highlight).toHaveStyle("opacity: 0");
+    expect(screen.queryByTestId("sessions-selection-highlight-surface"))
+      .not.toBeInTheDocument();
 
     rerender(
       <ChatList
@@ -277,18 +278,15 @@ describe("ChatList", () => {
 
     const activeButton = screen.getByTitle("Active topic");
     expect(activeButton).toHaveAttribute("aria-current", "page");
+    expect(activeButton.parentElement).toHaveClass("transition-[color]");
+    expect(activeButton.parentElement).not.toHaveClass("transition-colors");
     expect(activeButton.parentElement).not.toHaveClass(
       "bg-sidebar-accent",
       "shadow-[inset_0_0_0_1px_hsl(var(--sidebar-border)/0.55)]",
     );
-    expect(highlight).toHaveClass(
-      "transition-[transform,width,height]",
-      "motion-reduce:transition-none",
-    );
     expect(highlight).toHaveStyle(
-      "width: 284px; height: 32px; transform: translate3d(8px, 12px, 0); transition-property: none",
+      "width: 284px; height: 32px; transform: translate3d(8px, 12px, 0); opacity: 1; transition-property: none",
     );
-    expect(surface).toHaveStyle("opacity: 1; transform: scale(1)");
 
     revealFrame?.(0);
     expect(highlight.style.transitionProperty).toBe("");

webui/src/tests/quick-chat.test.ts

@@ -0,0 +1,37 @@
+import { describe, expect, it } from "vitest";
+
+import {
+  isQuickChatKey,
+  QUICK_CHAT_ID,
+  QUICK_CHAT_KEY,
+  quickChatSession,
+} from "@/lib/quick-chat";
+
+describe("Quick Chat identity", () => {
+  it("uses one stable websocket session", () => {
+    expect(QUICK_CHAT_ID).toBe("quick-chat");
+    expect(QUICK_CHAT_KEY).toBe("websocket:quick-chat");
+    expect(isQuickChatKey(QUICK_CHAT_KEY)).toBe(true);
+    expect(isQuickChatKey("websocket:another-chat")).toBe(false);
+  });
+
+  it("keeps persisted metadata behind the fixed identity", () => {
+    expect(quickChatSession({
+      key: "websocket:quick-chat",
+      channel: "websocket",
+      chatId: "quick-chat",
+      createdAt: "2026-07-30T08:00:00Z",
+      updatedAt: "2026-07-30T08:05:00Z",
+      preview: "hello",
+      modelPreset: "fast",
+    })).toMatchObject({
+      key: QUICK_CHAT_KEY,
+      channel: "websocket",
+      chatId: QUICK_CHAT_ID,
+      createdAt: "2026-07-30T08:00:00Z",
+      updatedAt: "2026-07-30T08:05:00Z",
+      preview: "hello",
+      modelPreset: "fast",
+    });
+  });
+});

webui/src/tests/thread-shell.test.tsx

@@ -3369,6 +3369,74 @@ describe("ThreadShell", () => {
     expect(screen.getByRole("option", { name: /\/history/i })).toBeInTheDocument();
   });
 
+  it("removes session-management affordances from a fixed conversation", async () => {
+    const client = makeClient();
+    vi.stubGlobal(
+      "fetch",
+      vi.fn(async (input: RequestInfo | URL) => {
+        const url = String(input);
+        if (url.endsWith("/api/commands")) {
+          return httpJson({
+            commands: [
+              {
+                command: "/new",
+                title: "New chat",
+                description: "Reset this chat and start a fresh conversation.",
+                icon: "square-pen",
+                lifecycle: "finalize_active_turn",
+                accepts_args: false,
+              },
+              {
+                command: "/history",
+                title: "Show conversation history",
+                description: "Print the last N persisted messages.",
+                icon: "history",
+                arg_hint: "[n]",
+                lifecycle: "side_channel",
+                accepts_args: true,
+              },
+            ],
+          });
+        }
+        return {
+          ok: false,
+          status: 404,
+          json: async () => ({}),
+        };
+      }),
+    );
+
+    render(
+      wrap(
+        client,
+        <ThreadShell
+          session={session("quick-chat")}
+          title="Quick Chat"
+          onToggleSidebar={() => {}}
+          allowConversationReset={false}
+          showSessionInfo={false}
+        />,
+      ),
+    );
+
+    await waitFor(() => expect(fetch).toHaveBeenCalledWith(
+      "/api/commands",
+      expect.objectContaining({
+        headers: { Authorization: "Bearer tok" },
+      }),
+    ));
+
+    fireEvent.change(screen.getByLabelText("Message input"), {
+      target: { value: "/" },
+    });
+
+    expect(screen.getByRole("option", { name: /\/history/i })).toBeInTheDocument();
+    expect(screen.queryByRole("option", { name: /\/new/i })).not.toBeInTheDocument();
+    expect(
+      screen.queryByRole("button", { name: "Session details" }),
+    ).not.toBeInTheDocument();
+  });
+
   it("does not bring back welcome cards when image mode is enabled", async () => {
     const client = makeClient();
     const settings = modelSettings("deepseek-v4-pro", "deepseek");

webui/src/tests/useNanobotStream.test.tsx

@@ -314,6 +314,63 @@ describe("useNanobotStream", () => {
     });
   });
 
+  it("preserves proactive automation source metadata on streamed assistant messages", () => {
+    const fake = fakeClient();
+    const { result } = renderHook(() => useNanobotStream("chat-cron-stream", EMPTY_MESSAGES), {
+      wrapper: wrap(fake.client),
+    });
+    const source = { kind: "cron", label: "Repo check" };
+
+    act(() => {
+      fake.emit("chat-cron-stream", {
+        event: "delta",
+        chat_id: "chat-cron-stream",
+        text: "Repo ",
+        source,
+      });
+      fake.emit("chat-cron-stream", {
+        event: "stream_end",
+        chat_id: "chat-cron-stream",
+        source,
+      });
+      fake.emit("chat-cron-stream", {
+        event: "turn_end",
+        chat_id: "chat-cron-stream",
+      });
+    });
+
+    expect(result.current.messages[0]).toMatchObject({
+      role: "assistant",
+      content: "Repo ",
+      isStreaming: false,
+      source,
+    });
+  });
+
+  it("preserves proactive automation source metadata on stream_end final text", () => {
+    const fake = fakeClient();
+    const { result } = renderHook(() => useNanobotStream("chat-cron-stream-end", EMPTY_MESSAGES), {
+      wrapper: wrap(fake.client),
+    });
+    const source = { kind: "cron", label: "Repo check" };
+
+    act(() => {
+      fake.emit("chat-cron-stream-end", {
+        event: "stream_end",
+        chat_id: "chat-cron-stream-end",
+        text: "Repo clean.",
+        source,
+      });
+    });
+
+    expect(result.current.messages[0]).toMatchObject({
+      role: "assistant",
+      content: "Repo clean.",
+      isStreaming: true,
+      source,
+    });
+  });
+
   it("does not start streaming from completed trailing activity after an answer", () => {
     const fake = fakeClient();
     const initialMessages = [