The title call sends reasoning_effort "none". Endpoints whose reasoning is
mandatory refuse it (OpenRouter GLM 5.3: 400 "Reasoning is mandatory for this
endpoint and cannot be disabled"), so those sessions never got a title. When
the refusal is about reasoning or effort, the title is asked once more without
the effort and with room for the reasoning. Other failures are not retried.
The native Anthropic KeyError first reported here is already fixed on main
(effort.NO_EFFORT handling).
The titling instruction asked for a '4-5 word title' in English without
pinning an output language, so non-English sessions (zh/ja/ko/…) got
English titles. Tell the model to write the title in the same language
as the user's opening message, and treat the word count as an
English-scale hint only — the natural equivalent brevity elsewhere
(CJK titles are a few characters, not 4-5 words).
The small-talk sentinel stays the fixed English token it always was
(normalization already tolerates casing/punctuation riffs).
Tests: the title call's system prompt pins the same-language rule; a
CJK opener keeps a CJK title end to end (sanitizer whitespace/quote
handling and the 80-char cap are script-agnostic); the sentinel still
fires for a non-English small-talk opener and earns its single retry.
Specialists offer starting points as an ask_user (free text stays available) so a picked option briefs the reviewer; title generation adds the agent first reply as evidence with a third attempt window.
The title rides the user message the moment it lands — a long agentic turn no longer holds the session name hostage; an opener signature keeps the completion hook (background turns) from burning duplicate attempts.
Imported from andrewyng/aisuite@1b4bbf303e
(contents of its platform/ directory, hoisted to the repo root).
Development history prior to this commit lives in that repository.
Co-authored-by: Devika <devikaverma11@gmail.com>