Best possible solution:

Merge the narrow fallback-loop fail-fast guard after maintainer review and green required checks; handle any broader queue, retry, or backoff policy separately.

Do we have a high-confidence way to reproduce the issue?

Yes. A high-confidence source-level path exists: configure a primary plus fallback and have the first candidate throw SessionWriteLockTimeoutError; current main leaves it unclassified and continues because non-final unknown errors advance through fallback.

Is this the best way to solve the issue?

Yes for this PR's scope. The explicit session-lock escape at the fallback-loop boundary is the narrowest maintainable fix for stopping model substitution while preserving generic unknown-error fallback behavior.

Acceptance criteria:

• pnpm test src/agents/model-fallback.test.ts src/agents/failover-error.test.ts
• pnpm check:changed

What I checked:

• current-main-fallback-gap: Current main normalizes candidate errors, then only rethrows an unrecognized error on the final candidate; unclassified non-final errors are recorded and fallback continues. (src/agents/model-fallback.ts:1096, 372e270871a2)
• typed-lock-error-source: Session write-lock acquisition timeouts are converted into SessionWriteLockTimeoutError, giving the fallback layer a typed local contention signal. (src/agents/session-write-lock.ts:540, 372e270871a2)
• classifier-suppresses-lock-timeout: Current main recursively detects session write-lock timeout errors and returns no failover classification when no explicit provider failover metadata is present. (src/agents/failover-error.ts:408, 372e270871a2)
• pr-fix-boundary: The PR head checks hasSessionWriteLockTimeout(normalized) immediately after normalization and throws the original error before the generic fallback path can advance. (src/agents/model-fallback.ts:1052, e3143af07868)
• regression-coverage: The PR adds a runWithModelFallback test that throws SessionWriteLockTimeoutError, expects that same error, and asserts the fallback runner was called once. (src/agents/model-fallback.test.ts:559, e3143af07868)
• changelog-entry: The PR adds a user-facing changelog bullet for failing fast on session write-lock timeouts and links the related report. (CHANGELOG.md:134, e3143af07868)

Likely related people:

• MonkeyLeeT: Authored the merged prior fix fix(agents): stop treating session lock waits as timeout #68700, which added the session-lock timeout suppression and adjacent failover coverage this PR builds on. (role: introduced partial classifier fix; confidence: high; commits: 8cc38c1b86f4; files: src/agents/failover-error.ts, src/agents/failover-error.test.ts, src/agents/session-write-lock-error.ts)
• obviyus: Co-authored the merged prior fix that introduced typed session write-lock timeout behavior at the lock boundary and failover classifier suppression. (role: adjacent session-lock maintainer; confidence: medium; commits: 8cc38c1b86f4; files: src/agents/session-write-lock.ts, src/agents/session-write-lock-error.ts, src/agents/failover-error.ts)
• steipete: Current checkout blame and commit metadata show recent maintenance touching the central fallback, failover, session-lock, and test files immediately before this PR. (role: recent fallback/session maintainer; confidence: medium; commits: 0f9f956bbd04; files: src/agents/model-fallback.ts, src/agents/model-fallback.test.ts, src/agents/failover-error.ts)
• vincentkoc: Prior ClawSweeper review context identifies recent work on model fallback helper imports and live-session fallback redirects in the same code path, making this a reasonable routing candidate for fallback behavior review. (role: adjacent fallback owner; confidence: medium; commits: 43a003b8a062, 480a3f66c9f7, da3977e6818b; files: src/agents/model-fallback.ts, src/agents/model-fallback.test.ts)

Remaining risk / open question:

• Some latest-head check-runs were still queued at inspection time, so maintainers should wait for required checks before merge.
• No live concurrent-lock scenario was run in this read-only review; the reproduction confidence comes from source tracing and the PR's focused regression path.

Codex review notes: model gpt-5.5, reasoning high; reviewed against 372e270871a2.