Agent Skills well-known proxy fetch has no timeout

[SebTardif]

Summary

The site handler for /{owner}/skills/{slug}/.well-known/agent-skills/index.json fetches the Convex agent-skills index with no AbortSignal. A stalled upstream holds the Node request until the platform limit. Installers that probe this well-known URL hang.

This is a source-level hang. I did not stall production Convex from the site handler.

Verification / reproduction

1. Read src/routes/$owner/skills/$slug/[.]well-known/agent-skills/index[.]json.ts.
2. Confirm GET and HEAD fetch have no signal.
3. Compare with OG helpers that use withOgFetchTimeout.

Actual

• A stalled Convex /api/v1/agent-skills/.../index.json hangs the well-known route.

Expected

• Upstream GET and HEAD use a timeout signal.
• Timeout returns a failed discovery response instead of hanging.

Evidence

Source: the well-known TanStack server handler above. Proposed patch: PR #3621.

Siblings: CLI body timeout #3548; changelog #3615; Hermit #3579. Those cover different fetches.

Security / privacy

Hang / availability only.

Acceptance criteria

• Well-known GET and HEAD pass a timeout to the upstream fetch.
• A stalled origin fails inside that budget.
• Successful upstream responses still proxy status, selected headers, and GET body.

This tracker stays open if PR #3621 is closed without landing on main.

[SebTardif]

Proposed patch: #3621

Linked with Ref only. This issue stays open if that PR is closed without landing on main.

[clawsweeper]

Codex review: keeping this open for maintainer follow-up; there is still a little grit to resolve. Reviewed September 11, 2026, 10:13 AM ET / 14:13 UTC.

Summary
The missing deadline remains on current main and in v0.23.3. The author’s linked fix PR is still open, so this issue remains useful as its tracker.

Reproducibility: yes. at source level: GET and HEAD await an upstream fetch without cancellation, and GET also waits for the complete body. No stalled-server runtime reproduction was executed.

Next step
Continue through #3621; the paired implementation and bulk-filing guard preclude another automated fix PR.

Review details

Best possible solution:

Apply a route-local deadline covering upstream headers and GET body consumption, with a predictable failed-discovery response and unchanged successful proxy semantics.

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

Yes, at source level: GET and HEAD await an upstream fetch without cancellation, and GET also waits for the complete body. No stalled-server runtime reproduction was executed.

Is this the best way to solve the issue?

Yes, a deadline in the shared route helper is the narrow repair; changing the general URL resolver or CLI client would not bound this server request.

AGENTS.md: found and applied where relevant.

Remaining risk / open question:

• Source proves the missing application deadline; actual platform termination timing and installer-visible failure behavior were not exercised.

Codex review notes: model internal, reasoning medium; reviewed against cbfee7343ddc.

Label changes

Label justifications:

• P2: This is a bounded discovery-route availability defect under upstream stalls, without evidence of a current widespread outage.
• impact:crash-loop: The affected requests can hang awaiting upstream headers or body completion.

Evidence reviewed

What I checked:

• Current source establishes the missing deadline: Both handlers call the same helper, whose fetch supplies only method and Accept. GET subsequently awaits arrayBuffer without a deadline. ([src/routes/$owner/skills/$slug/[.]well-known/agent-skills/index[.]json.ts:17](https://github.com/openclaw/clawhub/blob/cbfee7343ddc867316dd9b3de6fa8856730f9f41/src/routes/%24owner/skills/%24slug/%5B.%5Dwell-known/agent-skills/index%5B.%5Djson.ts#L17), cbfee7343ddc867316dd9b3de6fa8856730f9f41)
• Existing proxy contract: The two route tests cover successful body/header forwarding and bodyless HEAD responses; they do not exercise stalled upstream requests. publicApiUrl resolves the server request directly to the configured Convex base. (src/__tests__/agent-skills-discovery-route.test.ts:4, cbfee7343ddc)
• Separate discovery client already bounds body consumption: CLI registry discovery retains its abort timer through JSON parsing. This is a separate client and does not protect the site’s Agent Skills proxy. (packages/clawhub/src/discovery.ts:9, cbfee7343ddc)
• Feature history: The route history identifies the discovery implementation commit. Its raw recorded parent is 97bc586; the route is absent from that parent tree. GitHub verifies the matching merged feature PR at feat: support npx skills installs from skill pages #3233. ([src/routes/$owner/skills/$slug/[.]well-known/agent-skills/index[.]json.ts:17](https://github.com/openclaw/clawhub/blob/7aff40d26a898b934dc9553a35683e52f9aa20c7/src/routes/%24owner/skills/%24slug/%5B.%5Dwell-known/agent-skills/index%5B.%5Djson.ts#L17), 7aff40d26a898b934dc9553a35683e52f9aa20c7)
• Latest release remains affected: GitHub identifies v0.23.3 as the latest release; its route has the same fetch without a signal and unbounded GET body read. ([src/routes/$owner/skills/$slug/[.]well-known/agent-skills/index[.]json.ts:17](https://github.com/openclaw/clawhub/blob/87ca030c30f3cfb78ab15c8e66b5ff1469c8f9c8/src/routes/%24owner/skills/%24slug/%5B.%5Dwell-known/agent-skills/index%5B.%5Djson.ts#L17), 87ca030c30f3cfb78ab15c8e66b5ff1469c8f9c8)
• An existing PR owns the proposed fix: The author explicitly linked fix: bound stalled Agent Skills discovery requests #3621 in the discussion. A live REST read confirms it remains open and unmerged at the supplied head. (b17c30c301b6)

Likely related people:

• Patrick Erichsen: Raw commit 7aff40d adds src/routes/$owner/skills/$slug/[.]well-known/agent-skills/index[.]json.ts:17 relative to its recorded parents. This identifies author metadata, not feature responsibility or a PR merger. (role: source-line author; confidence: high; commits: 7aff40d26a89; files: src/routes/$owner/skills/$slug/[.]well-known/agent-skills/index[.]json.ts)

How this review workflow works

• ClawSweeper keeps one durable marker-backed review comment per issue or PR.
• Re-runs edit this comment so the latest verdict, findings, and automation markers stay together instead of adding duplicate bot comments.
• A fresh review can be triggered by eligible @clawsweeper re-review comments, exact-item GitHub events, scheduled/background review runs, or manual workflow dispatch.
• PR/issue authors and users with repository write access can comment @clawsweeper re-review or @clawsweeper re-run on an open PR or issue to request a fresh review only.
• Maintainers can also comment @clawsweeper review to request a fresh review only.
• Fresh-review commands do not start repair, autofix, rebase, CI repair, or automerge.
• Maintainer-only repair and merge flows require explicit commands such as @clawsweeper autofix, @clawsweeper automerge, @clawsweeper fix ci, or @clawsweeper address review.
• Maintainers can comment @clawsweeper explain to ask for more context, or @clawsweeper stop to stop active automation.

[steipete]

Fixed on main by #3621, squash commit 5bf0026.

The ten-second upstream deadline now covers stalled GET/HEAD headers and the full GET response body. A real loopback regression fails without the deadline and passes with it. The built Nitro server returned HTTP 500 for all three stalled cases in 10,006–10,009 ms; healthy and 404 GET/HEAD responses retained their status, cache policy, and expected bodies.

Full bun run ci:pr passed, including 6,886 coverage tests. Exact-head GitHub CI, including every browser shard, passed at c1bc2ec: https://github.com/openclaw/clawhub/actions/runs/35049843536. The optional Vercel fork preview required separate authorization; built-server and CI browser proof passed independently.

Thanks @SebTardif for the report and original fix.