lsm / HyperNeo / 32338456026

80%

Build Type

push

github

Committed by web-flow

Commit Message

fix(daemon): enforce session thinking config in anthropic-messages bridge (#2595)

* fix(daemon): enforce session thinking config in anthropic-messages bridge

The claude-agent-sdk CLI encodes thinking per its built-in model catalog:
for unrecognized model IDs (custom endpoints) any budgeted thinking config
is silently rewritten to {type:"adaptive"} on the wire, dropping
budget_tokens — and "adaptive" is even injected when no thinking was
requested. Verified by probing the binary directly:

  --max-thinking-tokens 31999 + swe-1-7   -> {"type":"adaptive"}
  --max-thinking-tokens 31999 + claude-sonnet-4-5 -> {"type":"enabled","budget_tokens":31999}

Port the openai-chat-bridge sessionThinkingConfigs pattern to the
anthropic-messages bridge: store the session's intended thinking via
setSessionThinkingConfig (already called by CustomEndpointProvider) and
rewrite the request body before forwarding — enabled+budget for a level,
no thinking field for off. Unknown sessions pass through unchanged.

Strengthened vs the openai-chat version: it only fills in thinking when
absent, which would not fix this case (adaptive is truthy); here the
mismatched value is overridden.

* fix(daemon): guard anthropic thinking rewrites

* fix(daemon): respect anthropic thinking capability

* fix(daemon): strip unsupported anthropic thinking

Coverage Stats

43679 of 59119 branches covered (73.88%)

Branch coverage included in aggregate %.

0 of 24 new or added lines in 1 file covered. (0.0%)

1 existing line in 1 file now uncovered.

59527 of 71316 relevant lines covered (83.47%)

3852.86 hits per line