[vincentkoc]

Summary

Add an unexposed historical native Azure speech repair for prospective Fusion handoff confirmation. The task covers reader/capacity retention during invalid-header cleanup, caller abort and rejected asynchronous cancellation, including reentrant abort. Only the native speech module is editable; the existing retirement interface is available as context.

Verification

• Original visible tests: 9/9 on both original and reference source.
• Historical acceptance: original source fails exactly 3 assertions across 12 tests; reference source passes 12/12. No skips or cancellations.
• Cached pinned Node image; isolated fixture size 1.59 GB.
• Workload harness tests: 18 passed, 3 optional integration cases skipped.
• Independent P0–P2 contract/fairness review: no actionable findings.

Deployment and risk

No model calls or runtime changes. No benchmark efficacy claim. This task remains unexposed and must not be described as held-out confirmation after using its acceptance results to tune a policy. Full hosted corpus qualification is required before landing.

[clawsweeper]

🦞👀
ClawSweeper picked this up.

Pull request received. I will update this pull request when review starts.

ClawSweeper review complete

ClawSweeper finished reviewing this revision. The review result is being finalized.

View the workflow run.

[clawsweeper]

Codex review: needs maintainer review before merge.

What this changes

Adds a historical Azure speech cleanup repair task and documents its acceptance criteria and evaluation limits.

Example: Benchmarking cancellation while an Azure speech response still owns a 32 MiB capacity lease.

• Before: The Fusion corpus has no dedicated task for this historical native speech retirement repair.
• After: The new development task requires retaining the reader and capacity until cancellation settles.

Review scores

Measure	Result	What it means
Overall readiness	🦐 gold shrimp (3/6)	A focused, source-consistent benchmark addition with supporting harness checks; final hosted qualification is not yet available.
Proof confidence	🦐 gold shrimp (3/6)	This benchmark-only change needs no shipped user-entry-point demonstration; historical qualification results are reported, while full hosted corpus qualification remains in progress.
Patch quality	🐚 platinum hermit (4/6)	No actionable review findings were identified.

Product

Kind: Test only · Worth it: Yes · Fix scope: Complete
User problem: Fusion evaluation needs additional unexposed repair tasks that distinguish passing visible tests from satisfying response-cleanup contracts.
Reason: The maintainer-authored addition covers a distinct historical failure with a narrow editable surface and avoids unsupported model-quality claims.

Merge readiness

✅ Ready for maintainer review

Keep open: this maintainer-authored PR adds a distinct, focused benchmark task, and the reviewed patch introduces no actionable defects.

Priority: P3
Reviewed head: 630850c55518fe7b3997105824b44c8b9d26682d

Before merge

None.

Findings

None.

Agent review details

How this fits together

The offline Fusion harness consumes pinned workload manifests, exposes restricted historical source to repair agents, and grades their edits against isolated historical acceptance tests.

flowchart LR
  A[Workload manifest] --> B[Historical source fixture]
  B --> C[Visible tests and source context]
  C --> D[Restricted repair agent]
  D --> E[Isolated acceptance verifier]
  E --> F[Qualification and evaluation receipts]

Loading

Technical review

Best possible solution:

Land the focused workload addition once the existing hosted qualification completes successfully, retaining its development-data and no-efficacy-claim limits.

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

This PR adds evaluation coverage rather than repairing current main; the historical source and acceptance-test diff substantiate the declared cleanup failure.

Is this the best way to solve the issue?

The manifest reuses the existing qualification, source-context, edit restriction, and verifier owners without adding parallel machinery.

AGENTS.md: found and applied where relevant.

Codex review notes: model internal, reasoning medium; reviewed against 43d978e80dfa.

Provenance checked

• Fusion evaluation interpretation keeps the original intent (653447d: Separate submitted failures from unfinished repairs and preserve the predeclared reliability threshold, as explained in fix(fusion): distinguish submitted failures from unfinished repairs #1042.)

Testing

Proof path: in-process harness.

Security

None.

Evidence

What I checked:

• Precisely scoped introduced patch: The pinned introduction adds one development workload and explanatory documentation; no runtime, dependency, permission, or harness implementation changes are introduced. (benchmarks/fusion/azure-retirement-workloads.json:5, 630850c55518)
• Historical repair matches the task: The oracle's parent is exactly the declared base. The native speech repair adopts inputRetirement, and its three added acceptance cases cover invalid headers, abort, and rejected cancellation with retained locks and capacity; the historical work is documented in fix(worker): retain native adapter input through cleanup #868. (worker/test/native-azure-speech.test.mjs:123, 80ecc42d4061)
• Existing harness supports this manifest: Qualification checks the repaired source against original visible tests before installing historical acceptance tests; only declared source edits cross into the fresh verifier, and the requested InputRetirement interface exists at the pinned base. (scripts/fusion-workloads.mjs:296, 630850c55518)
• Hosted qualification remains underway: The exact-head qualification run passed the Fusion test step and pinned-history fetch; corpus qualification was still running, with final isolated checks pending: https://github.com/openclaw/clawrouter/actions/runs/43482443998. (.github/workflows/fusion-workloads.yml:50, 630850c55518)
• Evaluation contract preserved: The new section retains development labeling and explicitly rejects efficacy or production-readiness claims, consistent with the existing reliability limits recorded in fix(fusion): distinguish submitted failures from unfinished repairs #1042. (docs/fusion-small-models.md:1067, 653447d3e58b)
• Feature ownership and current-main distinction: Merged Fusion history consistently identifies Vincent Koc as an area contributor; current main lacks this workload. Following the documentation history encountered an unavailable older blob, so that traversal does not establish its introduction. (docs/fusion-small-models.md:1050, 43d978e80dfa)

Likely related people:

• Vincent Koc: Raw commit 653447d adds docs/fusion-small-models.md:1031 relative to its recorded parents. This identifies author metadata, not feature responsibility or a PR merger. (role: source-line author; confidence: high; commits: 653447d3e58b; files: docs/fusion-small-models.md)

Preflight source

Source unknown. Source comparison at the observed PR snapshot only. Completion is self-reported; body, policy, tests, execution, and report bytes are not verified.

Labels

Label changes:

• add P3: This is an offline benchmark coverage addition with no production behavior change.
• add rating: 🦐 gold shrimp: Overall readiness is 🦐 gold shrimp; proof is 🦐 gold shrimp and patch quality is 🐚 platinum hermit.
• add status: 👀 ready for maintainer look: ClawSweeper has no concrete contributor-facing blocker left for this PR.

Label justifications:

• P3: This is an offline benchmark coverage addition with no production behavior change.
• rating: 🦐 gold shrimp: Overall readiness is 🦐 gold shrimp; proof is 🦐 gold shrimp and patch quality is 🐚 platinum hermit.
• status: 👀 ready for maintainer look: ClawSweeper has no concrete contributor-facing blocker left for this PR.

Rating scale

6/6 🦀 challenger crab · 5/6 🦞 diamond lobster · 4/6 🐚 platinum hermit · 3/6 🦐 gold shrimp · 2/6 🦪 silver shellfish · 1/6 🧂 unranked krab. Overall follows the weaker of proof and patch quality; ✨ marks media proof (a screenshot, video, or linked artifact) that directly shows the changed behavior.

Workflow

ClawSweeper edits this one comment on every review. Comment @clawsweeper re-review for a fresh review only. Repair and merge execution lanes are retired.

Reviewed November 13, 2026, 1:39 PM ET / 18:39 UTC.