fix: strip OpenAI-compatible replay reasoning

This commit is contained in:
Peter Steinberger
2026-05-10 10:08:26 +01:00
parent 0235040840
commit 9444b2ad9b
11 changed files with 60 additions and 31 deletions
+6 -3
View File
@@ -124,12 +124,15 @@ inter-session user turns that only have provenance metadata.
- Missing OpenAI Responses-family tool outputs are synthesized as `aborted` to match Codex replay normalization.
- No thought signature stripping.
**OpenAI-compatible Gemma 4**
**OpenAI-compatible Chat Completions**
- Historical assistant thinking/reasoning blocks are stripped before replay so local
OpenAI-compatible Gemma 4 servers do not receive prior-turn reasoning content.
- Historical assistant thinking/reasoning blocks are stripped before replay so
local and proxy-style OpenAI-compatible servers do not receive prior-turn
reasoning fields such as `reasoning` or `reasoning_content`.
- Current same-turn tool-call continuations keep the assistant reasoning block
attached to the tool call until the tool result has been replayed.
- Provider-owned exceptions can opt out when their wire protocol requires
replayed reasoning metadata.
**Google (Generative AI / Gemini CLI / Antigravity)**