Best possible solution:

Merge after a maintainer accepts the scoped Ollama proof and availability tradeoff, while keeping the broader Windows local-provider starvation issue open until all related paths are proven fixed.

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

Yes for the scoped dense Ollama stream path: current main parses all NDJSON lines from one body chunk without a cooperative wait, and the PR adds a regression test for timer progress before stream completion. The full Windows Gateway starvation scenario remains unverified in this review.

Is this the best way to solve the issue?

Yes for the scoped Ollama native stream problem: bounded cooperative yielding mirrors the existing OpenAI transport cadence and preserves abort classification. It is not a complete solution for every local-provider starvation path in the related issue.

AGENTS.md: found and applied where relevant.

Codex review notes: model gpt-5.5, reasoning high; reviewed against e0e7bae61219.

Label justifications:

• P1: The PR targets a user-facing local-model responsiveness regression that can make Gateway/channel workflows effectively unusable during model calls.
• merge-risk: 🚨 availability: Changing cooperative scheduling in the Ollama stream loop can affect Gateway responsiveness, provider throughput, and cancellation behavior under dense streams.
• rating: 🐚 platinum hermit: Overall readiness is 🐚 platinum hermit; proof is 🐚 platinum hermit and patch quality is 🐚 platinum hermit.
• status: 👀 ready for maintainer look: ClawSweeper has no concrete contributor-facing blocker left for this PR. Sufficient (terminal): The PR body records focused stream tests plus an AWS Crabbox changed gate for the scoped branch, while clearly noting that the full Windows Gateway scenario was not proven here.
• proof: sufficient: Contributor real behavior proof is sufficient. The PR body records focused stream tests plus an AWS Crabbox changed gate for the scoped branch, while clearly noting that the full Windows Gateway scenario was not proven here.

PR surface:

Source +47, Tests +86. Total +133 across 2 files.

What I checked:

• Repository policy read: Root AGENTS.md and extensions/AGENTS.md were read fully enough for this review; the extension boundary and provider hot-path guidance apply to this PR. (AGENTS.md:17, e0e7bae61219)
• Current main lacks the Ollama scheduler: Current main has cooperative stream scheduling in OpenAI transport paths, but no Ollama cooperative scheduler or dense native stream test in extensions/ollama. (extensions/ollama/src/stream.ts:1291, e0e7bae61219)
• Patch changes the dense stream loop: The PR adds a 64-event/12ms cooperative scheduler, checks abort before and after yields, and awaits the scheduler after non-final NDJSON chunks. (extensions/ollama/src/stream.ts:54, 3a478d76384d)
• Stream error contract supports abort classification: The stream function contract requires request/runtime failures to be encoded in the returned stream, with final error messages using stopReason "error" or "aborted". (packages/llm-core/src/types.ts:181, e0e7bae61219)
• Regression coverage added: The PR adds tests proving a timer can fire before dense stream completion and that caller aborts during dense processing surface as aborted. (extensions/ollama/src/stream.test.ts:360, 3a478d76384d)
• Scoped validation evidence: The PR body records focused Vitest, git diff check, and AWS Crabbox pnpm check:changed proof for the branch; this is scoped proof, not a live Windows Gateway reproduction. (3a478d76384d)

Likely related people:

• Jakub Rusz: Commit 8f44bd6 added incremental Ollama streaming events and the partial-snapshot text_delta path this PR now yields around. (role: introduced streaming behavior; confidence: high; commits: 8f44bd642660; files: extensions/ollama/src/stream.ts)
• Sunwoo Yu: Commit 1170229 added the native Ollama /api/chat provider and original stream implementation before it moved under extensions. (role: feature introducer; confidence: medium; commits: 11702290ff80; files: src/agents/ollama-stream.ts, extensions/ollama/src/stream.ts)
• vincentkoc: Vincent has prior merged Ollama stream/custom API history in commit 7e946b3 and is carrying the current maintainer PR. (role: recent area contributor; confidence: high; commits: 7e946b3c6c25, 3a478d76384d; files: extensions/ollama/src/stream.ts, extensions/ollama/src/stream.test.ts)
• steipete: Git shortlog and provider seam commits show repeated Ollama/provider stream maintenance and SDK boundary refactors, including be526d6. (role: adjacent owner; confidence: medium; commits: be526d642312, 87abcfd6a6b8; files: extensions/ollama/src/stream.ts, src/plugin-sdk/provider-stream-shared.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.