convert_messages() emitted reasoning items with ``content`` as a plain
string whenever preserve_reasoning was enabled (the DeepSeek spec).
DeepSeek's Responses gateway rejects that shape with a serde error
("input: invalid type: string ..., expected a sequence"), which surfaced
only after token consolidation cleared provider_state and forced the
full-history conversion path; replayed server items already carry list
content, which is why normal multi-turn requests never failed. Serialize
reasoning content as a list of output_text parts, matching the OpenAI
Responses schema and DeepSeek's accepted wire shape (verified live against
api.deepseek.com/responses).
The serde fallback classifier introduced in the previous commit remains as
a last-resort safeguard for any remaining wire incompatibility.
Tests: extend test_preserves_deepseek_reasoning_content to the array shape;
add a full-history regression with the observed failing item, a
replay/consolidation regression covering both replayed and converted
reasoning items, and provider-level request fixtures for both paths.
Full suite: 5773 passed, 22 skipped (only the known local-only
channels/sms packaging failure remains).
DeepSeek's new Responses endpoint (deepseek-v4-flash) intermittently rejects valid request bodies with serde deserialization errors such as 'input: invalid type: string ..., expected a sequence'. These were not classified as compatibility errors, so affected conversations died instead of falling back to Chat Completions.
The wire format is correct (input serializes as a list), so this is a server-side Responses compatibility issue; Chat Completions is strictly more permissive, making fallback safe. Extend the fallback classifier to recognize serde body-parsing markers. Repeated failures still trip the existing circuit breaker.
The helper never waits on the runtime-tasks gather after cancelling it
(its children are bounded individually), so the finished-gather test must
hand the helper an already-complete gather to exercise the bounded
retrieval path, and the cancelled-gather test must settle the gather
itself instead of expecting the helper to await a still-pending future.
Use a pre-completed child for the finished case and suppress(await) for
the cancelled case; both now assert done() and a single close.
Covers the lifecycle contract of _close_gateway_runtime: runtime tasks are
cancelled before shared resources close, pending background work is drained
before the close returns, cancellation-swallowing tasks and hanging cleanup
are bounded by their timeouts, a failing close is logged without blocking the
stop, duplicate cleanup is idempotent, and the runtime_tasks gather await path
is exercised for both completed and cancelled gathers.
The live v1beta API rejects the legacy responseFormat.image block
(enum-based aspectRatio/imageSize fields) for gemini-3.1-flash-lite-image
with INVALID_ARGUMENT, even for documented plain-string values. Gemini
Flash image models accept plain-string hints under
generationConfig.imageConfig instead (e.g. aspectRatio 16:9, imageSize
1K), which the API accepts. Switch the flash path to imageConfig and
update the provider tests accordingly. Other providers (aihubmix,
ollama, imagen) are untouched.
Ollama's spec keeps "nemotron" as a keyword so bare `nemotron-3-nano`
auto-routes to a configured Ollama install (PR #1863). NVIDIA NIM was
later registered with the same "nemotron" keyword (commit 046d0831),
creating the only keyword collision in the registry.
In `_match_provider`, the keyword loop accepted any local provider on
`spec.is_local` alone — no api_base check. Models like
`nvidia/nemotron-3-super-120b-a12b` (intended for OpenRouter or NVIDIA
NIM) were therefore hijacked to http://localhost:11434/v1 even when the
user had never configured Ollama, causing silent connection errors at
runtime.
Add the same api_base gate the local-fallback loop already uses: a local
provider only wins by keyword when the user has actually set its
api_base. Preserves PR #1863's intent for users who configured Ollama;
fixes the silent hijack for everyone else.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
AutoCompact.prepare_session runs on the turn hot path
(AgentLoop._compact_session) and read the persisted _last_summary metadata
with an unguarded meta['text'] and datetime.fromisoformat(meta['last_active']).
A _last_summary dict that was hand-edited or written by another version
(missing text/last_active, or a non-ISO last_active) raised KeyError/ValueError
out of the turn.
Sibling readers already tolerate the same data: estimate_session_prompt_tokens
uses .get('text') and _archive parses inside try/except. Mirror that tolerance:
skip when text is unusable, and fall back to the session's own updated_at (the
value the writer persists) when last_active is missing or unparseable, so the
archived summary is preserved instead of crashing the turn.
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
_load() treated any OSError like corruption and returned an empty store. When pairing.json was transiently unreadable, an unapproved DM could deny the sender, generate a pairing code from the empty view, and overwrite the store without its approved senders.
Keep the existing JSONDecodeError reset behavior, but propagate OSError so mutations cannot persist unreadable state. Read-only checks fail closed without writing; mutating /pairing subcommands report temporary unavailability; and the DM pairing path skips one reply instead of crashing the handler.
This mirrors the refuse-to-overwrite strategy used by the cron and trigger stores.
chengyongru
Xubin Renandchengyongru
ziuusandchengyongru
arcdrake22andchengyongru
KDBandXubin Ren
santhrealandXubin Ren
yu-xin-candXubin Ren
shixi-liandchengyongru
Solaris-starandXubin Ren