Best possible solution:

Provider streams should cancel on parser exit and release guarded slots on abandoned wrappers without adding provider APIs, config, or fallback policy.

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

Yes, source-level: current main only releases managed provider responses on normal end, read error, or explicit cancel, and the linked report provides production lsof evidence of leaked sockets. I did not run a current-main multi-hour Gateway soak in this read-only review.

Is this the best way to solve the issue?

Yes, with a maintainer-proof caveat: the shared guarded fetch wrapper is the right boundary for returning slots across provider transports. The remaining decision is whether the supplied local TCP/GC harness is enough or a longer live soak is required.

AGENTS.md: found and applied where relevant.

Codex review notes: model gpt-5.5, reasoning high; reviewed against 63621eead219.

Label justifications:

• P1: The linked bug can exhaust gateway file descriptors and break real agent turns that depend on streamed provider calls.
• merge-risk: 🚨 availability: The PR changes provider stream cancellation/finalization, where missed or premature cleanup can affect long-running Gateway availability.
• rating: 🐚 platinum hermit: Overall readiness is 🐚 platinum hermit; proof is 🦞 diamond lobster and patch quality is 🐚 platinum hermit.
• status: 👀 ready for maintainer look: ClawSweeper has no concrete contributor-facing blocker left for this PR. Sufficient (live_output): The PR body includes after-fix real local TCP streaming output using the actual buildGuardedModelFetch path and showing the abandoned stream socket closed.
• proof: sufficient: Contributor real behavior proof is sufficient. The PR body includes after-fix real local TCP streaming output using the actual buildGuardedModelFetch path and showing the abandoned stream socket closed.

PR surface:

Source +23, Tests +163, Other +1. Total +187 across 9 files.

What I checked:

• Repository policy read: Read the full root AGENTS.md and scoped guides for src/agents, src/gateway, scripts, extensions, and test; provider runtime proof, dependency inspection, and real behavior proof requirements shaped this review. (AGENTS.md:1, 63621eead219)
• Current main leak path: Current main's managed provider response releases the guarded fetch slot on normal completion, read error, or explicit cancel, but has no release path when the wrapper body is abandoned before those callbacks run. (src/agents/provider-transport-fetch.ts:389, 63621eead219)
• PR fallback cleanup: The PR head adds a FinalizationRegistry fallback, unregisters on normal finalize, cancels the locked source reader, and keeps release/local-service cleanup idempotent. (src/agents/provider-transport-fetch.ts:373, 766b36cf9ccb)
• Parser-side deterministic cancellation: The PR also cancels open Anthropic and Google SSE readers when parser iteration exits before a clean end, reducing reliance on GC for known early-exit error paths. (src/agents/anthropic-transport-stream.ts:655, 766b36cf9ccb)
• Focused regression coverage: The PR adds tests for abandoned managed-stream finalization plus Anthropic and Google malformed/open SSE cancellation paths. (src/agents/provider-transport-fetch.test.ts:186, 766b36cf9ccb)
• Dependency contract check: TypeScript's FinalizationRegistry contract allows an unregister token and forbids heldValue being the target; DOM stream types define reader cancel as a Promise-returning cleanup path, matching the PR's approach.

Likely related people:

• steipete: Local blame attributes the current provider transport stream implementation to Peter Steinberger in the grafted main history, and the PR branch includes maintainer follow-up commits from the same area for deterministic reader cancellation and shard/tooling updates. (role: recent area contributor; confidence: medium; commits: 2b61d38a455c, 32f20d552745, 3c8f526d197e; files: src/agents/provider-transport-fetch.ts, src/agents/anthropic-transport-stream.ts, extensions/google/transport-stream.ts)

• 🦀 challenger crab: rare, exceptional readiness with strong proof, clean implementation, and convincing validation.
• 🦞 diamond lobster: very strong readiness with only minor maintainer review expected.
• 🐚 platinum hermit: good normal PR, likely mergeable with ordinary maintainer review.
• 🦐 gold shrimp: useful signal, but proof or patch confidence is still limited.
• 🦪 silver shellfish: thin signal; proof, validation, or implementation needs work.
• 🧂 unranked krab: not merge-ready because proof is missing/unusable or there are serious correctness or safety concerns.
• 🌊 off-meta tidepool: rating does not apply to this item.

Shiny media proof means a screenshot, video, or linked artifact directly shows the changed behavior. Runtime, network, CSP, and security claims still need visible diagnostics.

• 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.