fix(llm): correctly handle reasoning=off in Anthropic and Google adapters#100938
fix(llm): correctly handle reasoning=off in Anthropic and Google adapters#100938aniruddhaadak80 wants to merge 2 commits into
Conversation
|
Codex review: stale review; fresh review needed. Summary Next step Review history (1 earlier review cycle)
|
1fad2ff to
c565c30
Compare
|
Thanks. After tracing the runtime contract, this patch targets a state the public provider API does not permit: The latest head also edits only Anthropic despite the PR title claiming Google; Fable 5's off-to-low result already matches its always-on adaptive contract. Closing rather than widening the plugin/provider API in a small fix. If OpenClaw should expose |
Fixes #100926
When users pass
reasoning: "off"toSimpleStreamOptions, the adapters incorrectly evaluated it as truthy since"off"is a non-empty string. This PR explicitly checks for"off"to disable reasoning in the Anthropic and Google providers. For models with mandatory adaptive thinking (like Claude Fable 5),"off"now explicitly maps to"low"effort.