From ten independent read-only audits (11 Oct 2026), verified against the code:
- a permanent gateway error (no tools, 401/402/403, bare upstream 400) ends the turn with a visible
assistant message instead of a silent retry and an empty turn;
- "modeli ... yapma", "X modeli için test yap" and general "hangi model ..." questions are no longer
commands; a name whose version does not fit, or that several models share, is asked back;
- the stream is forwarded as it arrives (read1) and an upstream drop is no longer logged as a client
stop nor saved as an interrupted answer;
- a lone "." joining a running turn is removed without the hook messages after it;
- the interrupted-answer note finds the user message behind hook messages and detects completion
by event type; a per-minute rate limit is not learned as a context window.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
A 607 KB Codex request was estimated at 189776 tokens (window 200000), sent untrimmed and rejected by
Claude; the gateway only said "upstream 400", so Codex saw a dropped stream. Claude targets now use
2.6 bytes/token, and that bare 400 on a request filling most of the window becomes context_length_exceeded.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Measured with the unmodified Codex binary on 10 Oct 2026:
- real context windows per model (NanoGPT list, local table, combo = its
smallest member, capped at the summarizer's window) so Codex compacts by
itself when switching to a smaller model; an overflow guard trims
oversized old parts and reports what still does not fit
- classic tools instead of code mode for external models: GLM 0/3 -> 3/3,
Qwen 0/3 -> 2/3 on simple tool tasks, Claude/cx/Gemini unchanged
- phase (commentary / final_answer) on external assistant messages
- an answer cut by the client is carried to the next turn once
- images stay for vision-capable targets
- summary brake stops tight loops, not frequent compaction
- permanent gateway errors and tool-less targets fail at once with a
readable message instead of five reconnects
- a lone "." (used to apply a model change in voice) is answered locally
- the bridge listens within seconds; the IDE waits 45 s for it
Release: AITURK IDE 1.0.0-beta.26.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
