Disagreement · 8ff8c1af-openai-3

Anthropic provider uses different default output_config.effort between streaming and non-streaming

mismatch

repo 56f59a0d·PR #2·reviewed 4 days ago

Primary finding

Anthropic provider uses different default output_config.effort between streaming and non-streaming

mediummaintainabilityhigh

src/providers/anthropic.ts:92-95
src/providers/anthropic.ts:181-185

Defaulting to 'high' for streaming and 'low' for non-streaming creates surprising differences in latency/cost/quality between otherwise identical calls when only the delivery mode changes. This is non-obvious to callers and complicates tuning.

Recommendation

Choose a single default (e.g., 'high' or 'medium') for both paths, or centralize defaulting logic (e.g., resolveEffort(options.effort ?? 'high')) so behavior is consistent across streaming/non-streaming.

Counterpart finding

AnthropicProvider passes output_config.effort that has no validation against actual model support

lowapi-contractmedium

src/providers/anthropic.ts:81-90
src/providers/anthropic.ts:156-165

options.effort is typed as `string` in StreamOptions/SendOptions (types.ts line 64). The provider blindly casts it to a tri-state literal and passes through, allowing arbitrary strings to reach the Anthropic API and causing 400 errors for unknown values. Additionally, the default differs between streaming ('high') and non-streaming ('low'), which is a surprising inconsistency — the same call with the same options will request different effort tiers depending on whether streaming is used.

Recommendation

Validate effort against allowed set; align defaults across stream/send; reconsider whether `output_config.effort` is a real Anthropic field (it is not in the public API — possible bug).

Why this didn't post

This finding didn't meet AntFleet's unanimous agreement threshold. Both frontier models review every PR independently; only findings they both flag with the same severity and category are posted to the PR. This one fell through.

read the methodology →

From the same review

These findings passed the unanimous gate on the same PR review. The disagreement above was filtered out; the findings below were posted.

← back to all disagreements view public receipts see unanimous findings + anatomies →

Tweet ↗