* fix(dream): ground commit messages and cursor advance in the real git diff
Dream consolidation could emit a /dream-log audit record that did not match
the actual file changes: build_dream_commit_message appended the LLM's
unverified resp.content, dream_run_completed only checked the stop reason, and
file contents were deliberately omitted from the prompt. The combination let a
single-turn self-report become the durable audit record.
- gitstore: add summarize_working_tree() — a structured, machine-derived
summary (per-file +N/-M, totals, capped unified diff) of working-tree
changes vs HEAD. Pure filesystem/git ground truth, never LLM narrative.
- memory: build_dream_commit_message now takes the diff body instead of resp;
dream_content_diff() exposes the real delta over SOUL/USER/MEMORY.md only
(excludes .dream_cursor so cursor writes aren't mistaken for edits);
build_dream_prompt embeds current file contents so the model edits reality,
not a stale mental model.
- builtin/cli: both Dream paths now compute the diff, gate cursor advance on
a non-empty delta (no-op runs no longer swallow history), and commit with
the diff-grounded message. Non-git workspaces fall back to the completion
check.
- dream.md: document that contents are embedded, and add a chain-of-
verification guardrail so the model's summary cannot claim unmade edits.
A regression test proves a lying resp.content never reaches the audit log
while the real diff does.
* fix(dream): mark non-UTF-8 memory files as binary in diff summary
Address review feedback (Q1 on PR #4673): summarize_working_tree read
working-tree files with errors="replace", which would emit U+FFFD
replacement chars into the audit record if a memory file ever held
invalid UTF-8 — misrepresenting the diff it is meant to make truthful.
Switch to errors="strict" and catch UnicodeDecodeError: a non-UTF-8
(or binary/corrupt) file is now recorded as "{path}: binary or
non-UTF-8 file changed" and omitted from the unified diff, so the
audit record stays honest. An empty diff block is also suppressed when
all changes are binary.
Adds a defensive regression test asserting no replacement char leaks.
apply_final_call_ids iterated over all final tool calls, including
non-file-edit tools like read_file. The greedy path-match in
matches_final_tool_call could overwrite a correct unique id with a
stale one from a different streaming state, producing duplicate
tool_use ids that poison the persisted session.
Guard the loop with is_file_edit_tool() so only tracked file-edit
tools (write_file, edit_file, apply_patch) are subject to canonical
id remapping. Non-file-edit tools keep their authoritative id from
get_final_message().
Fixes#4595
Combine malformed tool-call handling with placeholder filtering and a
no-tools fallback so a relay that returns tool_use blocks with null
id/name/input can no longer crash a turn or permanently wedge a session.
Adapted to the ContextGovernor architecture (context governance now lives
in nanobot/agent/context_governance.py, not runner.py):
- ToolCallRequest.has_valid_name(): single source of truth for "usable
name" (non-empty string).
- tool_hints.format_tool_hints(): skip tool calls with a non-string/empty
name instead of raising AttributeError on the whole turn.
- ContextGovernor.strip_placeholder_assistant_messages() and
strip_malformed_tool_calls() (plus the _tool_call_name_is_valid helper):
history-cleaning staticmethods invoked at the START of
prepare_for_model() — strip_placeholder, then strip_malformed, then the
existing drop_orphan/backfill chain. Both only repair the model-facing
copy and leave persisted history untouched (return a copy, or the same
list when nothing changes). Also wired into runner's minimal-repair path.
- AgentRunner._drop_malformed_tool_calls(): returns
(dropped, all_dropped, original_finish_reason); clears finish_reason to
"stop" when all calls are dropped.
- AgentRunner._malformed_tool_call_retry_messages() + _request_model
malformed_retry flag: when an all-dropped tool_calls response comes back,
retry once with a corrective note; if the retry STILL comes back
all-dropped, fall back to _request_no_tools for graceful text degradation.
Tests for the history-cleaning methods live with ContextGovernor in
tests/agent/test_runner_governance.py; response-layer and tool-hint tests
stay on AgentRunner / tool_hints.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
maintainer edit: derive thinking tag regexes and streaming partial prefixes from one tag list so future aliases only need one entry while preserving legacy self-closing think/thought behavior.
maintainer edit: native reasoning streams can split <thinking> wrapper tags across chunks. Buffer the stream and emit only cleaned incremental reasoning so raw partial tags do not reach WebUI.
Treat the current live user message as the replay boundary for normal user turns, while keeping user-turn extension for history and consolidation paths that need it. Add regression coverage for the user-triggered long tool-turn case.
Maintainer edit: make token truncation include the suffix within the budget and route the consolidator through the shared helper so recent-history and archive truncation keep the same semantics.
The recent-history section injected into the system prompt was capped by character count (_MAX_HISTORY_CHARS = 32_000). Characters are a poor proxy for tokens: ~32k chars of English is ~8k tokens, but the same char count of CJK text or code can be far more, so the cap could let the section blow well past its intended size on non English/code-heavy histories.
Add a reusable truncate_text_to_tokens() helper (reusing the tiktoken
cl100k_base encoder already used elsewhere, with a char-based fallback) and
switch the digest cap to a token budget (_MAX_HISTORY_TOKENS = 8_000),
matching the previous English-text size while holding regardless of
content.
Move the fenced-code-block-aware splitting logic out of the shared
split_message helper (used by Signal, Slack, Discord, Weixin, etc.)
and into a Telegram-specific _split_telegram_markdown function.
The shared split_message remains a plain-text chunker. The Telegram
channel now uses _split_telegram_markdown for its raw Markdown paths
that feed _markdown_to_telegram_html, preventing broken HTML rendering
when splits fall inside fenced code blocks.
Also fixes a regression where content beginning with whitespace before
a fence could emit a whitespace-only chunk.
Addresses review feedback on #4257.
When split_message splits a long message, it now checks whether the
split point falls inside a fenced code block. If so, it either moves
the split to before the opening fence or closes/reopens the fence
across chunks, preventing broken HTML rendering.
Addresses #4250
Extract is_image_file() and reference_non_image_attachments() from
AgentLoop private static methods into nanobot/utils/document.py where
they belong alongside extract_documents(). Simplify config lookup by
removing dead isinstance(dict) branch.
Remove standalone nanobot/heartbeat/ service and replace it with an
auto-registered system cron job on gateway startup. Key behaviors preserved:
- HeartbeatConfig (enabled, interval_s, keep_recent_messages) remains in
GatewayConfig for backward compatibility.
- On startup, if enabled, a system cron job "heartbeat" is registered with
schedule derived from interval_s.
- HEARTBEAT.md is checked on each tick; empty/template-identical files skip
to avoid wasting LLM calls.
- Post-run evaluate_response and session history truncation
(keep_recent_messages) are retained.
- Delivery target selection, deliverable filtering, and preamble guidance
are preserved.
Files removed:
- nanobot/heartbeat/__init__.py
- nanobot/heartbeat/service.py
- tests/heartbeat/*
- tests/agent/test_heartbeat_service.py
Templates and docs updated to reflect cron-based usage.
`long_task` registers a sustained objective, but `AgentRunner` would
still exit with `stop_reason="completed"` when the LLM produced a final
text response without calling `complete_goal`. This defeated the purpose
of sustained goals.
Add `goal_active_predicate` and `goal_continue_message` to `AgentRunSpec`.
When the predicate returns `True` at the natural completion checkpoint,
inject a continuation message via the existing `_try_drain_injections`
machinery, forcing the runner to continue looping.
Also extract the default continuation prompt to
`nanobot/utils/runtime.py` alongside the existing recovery-message
builders.
Drop the legacy unified-diff patch parameter and all related parsing/
generation logic (_parse_patch, _generate_patch, _apply_hunks, etc.).
The tool now accepts only the structured `edits` array, eliminating the
intermediate diff-string round-trip.
Also update file_edit_events tracking and tests to work exclusively
with edits.
Benchmark (zhipu glm-5.1, edits mode): 15/15 cases passed.
- Remove generated_image_paths_from_messages() and _extract_text_payload() from artifacts.py (no runtime callers)
- Remove session_attachments.py entirely (merge_turn_media_into_last_assistant and stage_media_paths_for_session_replay had no runtime callers)
- Remove test_session_media_persist.py and the orphaned test in test_artifacts.py
The runtime media-attachment mechanism was broken for streaming channels
(e.g. WebSocket): the _streamed flag caused _send_once to skip the final
OutboundMessage that carried generated media, so images were never delivered.
Rather than adding complex coordination between streaming and media delivery,
delegate image delivery to the LLM: after generate_image returns artifact
paths, the next_step prompt now instructs the LLM to call the message tool
with the paths in the media parameter. This works uniformly across all
channels, streaming or not.
Remove generated_media from TurnContext, _assemble_outbound, and _state_save.
Update prompts in identity.md, SKILL.md, message tool description, and
artifacts.py to reflect the new flow.
