Codex review: needs real behavior proof before merge. Reviewed July 20, 2026, 9:19 AM ET / 13:19 UTC.

Summary
The branch permits replay-safe, provider-owned Bedrock prompt aborts to enter existing model fallback while preserving user stops, timeouts, visible replies, and other replay-unsafe outcomes.

PR surface: Source +68, Tests +275. Total +343 across 5 files.

Reproducibility: no. high-confidence live reproduction is currently available: the source-level abort shape is covered by focused tests, but the contributor's 11 real Bedrock attempts did not trigger the provider-owned abort from the report.

Review metrics: none identified.

Merge readiness
Overall: 🦐 gold shrimp
Proof: 🦐 gold shrimp
Patch quality: 🐚 platinum hermit
Result: blocked until stronger real behavior proof is added.

Overall follows the weaker of proof and patch quality, so missing proof can cap an otherwise strong patch.

Rank-up moves:

• Provide a redacted trace showing provider-owned abort classification, configured fallback selection, and successful completion, or obtain explicit maintainer acceptance of the current proof standard.

Proof guidance:

• [P1] Needs stronger real behavior proof before merge: The contributor supplied meaningful real Bedrock reachability and long-run evidence, but it does not show the changed provider-abort classification, fallback attempt, and recovered completion; add a redacted affected-environment trace or obtain a maintainer evidence-standard override. After adding proof, update the PR body; ClawSweeper should re-review automatically. If it does not, the PR author or someone with repository write access can comment @clawsweeper re-review.

Risk before merge

• [P2] Merging changes the terminal outcome for a narrow class of aborted agent runs; if the provider envelope is misclassified as replay-safe, a fallback could repeat an attempt after user-visible delivery or another side effect.
• [P2] The PR's live Bedrock matrix proves reachability and normal long-run behavior, not the intended abort → fallback recovery chain, so green CI does not settle the session-state safety claim.

Maintainer options:

1. Capture the actual recovery chain (recommended)
   Before merging, collect a redacted reporter or production trace that shows the provider-owned abort, replay-safe classification, fallback attempt, and successful terminal result.
2. Accept the bounded replay contract
   A maintainer may accept the source-level guard and regression suite despite the missing real abort trace, explicitly owning the remaining provider-envelope uncertainty.

Next step before merge

• [P1] The remaining blocker is a maintainer decision on whether source-level replay-safety proof can substitute for a rare affected-environment trace; no safe mechanical repair remains.

Maintainer decision needed

• Question: Should this rare Bedrock provider-abort recovery require a real affected-environment abort-to-fallback trace before merge, or may the reviewed replay-safety guards plus focused tests and negative live probes satisfy the evidence bar?
• Rationale: The contributor has exercised Bedrock organically across two platforms but cannot reproduce the provider-owned abort, and choosing whether to accept source-level safety proof is an evidence-standard decision rather than a mechanical repair.
• Likely owner: steipete — Likely recent core-agent routing contact, though the available feature-history evidence is incomplete.
• Options:
  • Require affected-environment trace (recommended): Keep the PR blocked until a redacted production or reporter trace shows provider-owned abort classification, fallback selection, and successful completion.
  • Accept bounded source proof: Accept the current predicate, replay-state coverage, focused tests, and negative live matrix as adequate for a rare envelope that cannot presently be reproduced.

Security
Cleared: The five-file TypeScript diff changes agent recovery and tests only; no dependency, workflow, secret, package-resolution, or supply-chain concern is evident.

Updated the PR body with the live Bedrock evidence I could produce.

Summary: using profile-based AWS auth only (AWS_PROFILE + AWS_SDK_LOAD_CONFIG=1, no exported AWS key/session env), the direct Bedrock model smoke and Gateway Bedrock smoke both passed against amazon-bedrock/global.anthropic.claude-sonnet-4-6. I also ran two longer issue-specific probes with Bedrock primary plus anthropic/claude-haiku-4-5 fallback; both completed normally with status=ok / stopReason=stop, so this AWS setup did not reproduce the provider-originated abort envelope.

I cannot honestly attach the exact after-fix abort-to-fallback log ClawSweeper asked for from this environment. The remaining decision is whether maintainers accept the source-level predicate/tests plus these negative live attempts, or keep this blocked until a reporter/production environment can produce the real Bedrock abort after applying this patch.

@clawsweeper re-review

🦞🧹
ClawSweeper re-review requested.

I asked ClawSweeper to review this item again.
Action: item re-review queued (workflow sweep.yml, event repository_dispatch).
Result: the existing ClawSweeper review comment will be edited in place when the review finishes.

This pull request has been automatically marked as stale due to inactivity.
Please add updates or it will be closed.

Evidence update: expanded organic Bedrock matrix (no qualifying abort)

This supersedes and expands my earlier negative live probes. I completed the planned macOS + Linux organic reproduction matrix, including a refreshable-auth prerequisite for Linux.

Scope

• PR head: ccc9500b014d136864eb41f5ad70627045733e1a
• Tested source: 362f3d48210f8daadbba741d60bb380c241c851d
  • This is the PR head plus one local committed, content-free observability change used to classify replay safety. That extra commit was not pushed.
• Node / pnpm: v24.15.0 / 11.2.2
• Frozen prompt: 72,986 bytes; SHA-256 df02349af9bc34b51bb27420bf0484567889e269fe9f5cfa19045ff230bccc10
• Real Gateway, bedrock-converse-stream, default HTTP/2, tools disabled

Credential-refresh gate

The Linux lane used an isolated AWS login cache through credential_process; static AWS credential environment variables were unset.

• Initial remote STS and Bedrock control-plane calls passed with credentials expiring at 2026-07-20T04:46:05Z.
• I waited beyond that expiration without another login, profile rewrite, or credential-cache transfer.
• Post-expiration STS, Bedrock control-plane, and exact installed-SDK calls passed through the same cache/profile, with refreshed credentials expiring at 2026-07-20T05:02:04Z.
• A short real-Gateway primary smoke then passed with provider amazon-bedrock, model us.anthropic.claude-sonnet-4-6, API bedrock-converse-stream, zero tools, and no ExpiredToken.

No long Linux attempt began until this gate passed.

Accepted organic matrix

All 11 accepted attempts returned status=ok, finishReason=stop, and a one-entry successful primary execution trace. Tools and side effects were zero. The three Linux trajectories additionally recorded every abort and timeout flag as false and had no terminal error, ExpiredToken, replay-safe abort event, failover decision, or fallback decision.

The earlier two Linux attempts that used static exported credentials are excluded. They ended in credential expiry and are diagnostics, not provider-abort evidence.

Honest conclusion

This is stronger negative evidence only. The organic matrix did not reproduce the reported provider-owned abort, so it does not demonstrate the requested provider-abort → replay-safe classification → configured fallback → successful completion chain. I cannot claim that the requested rank-up proof is complete.

The next useful evidence is a redacted trace from the original reporter or a production environment running the instrumented source. I do not think a synthetic fault proxy should be built until that real envelope is available.

I retained a checksum-verified private bundle containing the raw CLI results, Gateway logs, and exported trajectories. It was not uploaded publicly: the available Crabbox artifact publisher has no configured remote broker in this account and resolved only to local storage. The inline evidence above is content-free and was scanned for AWS keys/tokens, account/ARN data, local paths, prompts, generated responses, and raw session/run identifiers.

@clawsweeper re-review

🦞🧹
ClawSweeper re-review requested.

I asked ClawSweeper to review this item again.
Action: item re-review queued (workflow sweep.yml, event repository_dispatch).
Result: the existing ClawSweeper review comment will be edited in place when the review finishes.

Re-review progress:

• State: Failed
• Detail: The review artifact was captured, but publication did not finish cleanly.
• Run: https://github.com/openclaw/clawsweeper/actions/runs/29782155259
• Updated: 2026-07-20T22:03:46.479Z