[hakarrr]

Summary

• Retry classification in both API providers missed transient statuses, so a single provider-side blip failed the request immediately instead of backing off:
  • is_retryable_status matched 408/409/429/500/502/503/504 but not 529 (Anthropic's documented overloaded_error status, also emitted by several OpenAI-compatible gateways).
  • is_retryable_400 (gateway-flake HTTP 400 bodies: "no parseable body", "connection reset", "broken pipe", "empty reply from server") was defined with a doc comment saying these should be retried, but was dead code - never called by expect_success in either provider.
• expect_success in anthropic.rs and openai_compat.rs now computes retryable = is_retryable_status(status) || is_retryable_400(status, &body), and 529 is retryable in both matchers. Real 400s (e.g. invalid request bodies) remain non-retryable.

Anti-slop triage

• Classification: actionable-fix
• Evidence: failing tests on unpatched main, then green after the fix:
  • With new tests applied to unpatched sources: cargo test -p api --lib providers:: -> 2 failures (retryable_statuses_are_detected in both providers, on the 529 assertion).
  • With only the 529 matcher fix applied: 2 failures remain (expect_success_marks_gateway_flake_400_retryable in both providers), proving the dead-code wiring gap through the real expect_success path.
  • With the full fix: cargo test -p api --lib -> 180 passed, 0 failed.
• Non-destructive review result: merge candidate

Verification

• Targeted tests/docs checks ran, or the gap is explicitly recorded. (cargo test -p api --lib: 180 passed incl. 6 new tests; cargo clippy -p api --lib: clean; cargo fmt --check -p api: clean. Full-workspace cargo test --workspace not run locally - change is confined to the api crate's retry classification.)
• git diff --check passes.
• No live secrets, tokens, private logs, or unrelated generated churn are included.

Resolution gate

• If this PR resolves an issue, the issue number and fix evidence are linked. (No tracking issue found - behavior-anchored fix per CONTRIBUTING "tied to a concrete issue or behavior"; evidence above.)
• If this PR should not merge, the rejection/defer rationale is evidence-backed and does not rely on vibes.
• I did not merge/close remote PRs or issues from an automation lane without owner approval.

[Yeachan-Heo]

Reviewed at 758960d. The 400 retry rule is limited to the four known gateway-flake body phrases; ordinary invalid-request 400s remain non-retryable. HTTP 529 is added to the retryable statuses without changing retry limits or backoff. Both provider test suites exercise expect_success directly, and local formatting, clippy, and workspace tests pass.
—
[repo owner's gaebal-gajae (clawdbot) 🦞]