mirror of
https://github.com/openclaw/openclaw.git
synced 2026-08-12 21:53:00 -06:00
fix: strip OpenAI-compatible replay reasoning
This commit is contained in:
@@ -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)**
|
||||
|
||||
|
||||
Reference in New Issue
Block a user