fix(turns): skip getReadable on already-cancelled attach

[btipling]

Summary

Fixes F5 on a durable run that was cancelled hours ago: GET/POST no longer call getReadable() when run.status is already cancelled/failed (plain synthetic SSE). Subscribe-fail (Failed to fetch) no longer hot-resumes. Producer Request cancelled. without an abort clears running (operator attach Stop/Esc still keep-running). Live running/completed still use the #917 wrap.

Fixes #919
Fixes #918
Refs #915
Refs #916
Refs #816
Refs #794

Plan locks

• bodyForRun: cancelled/failed → synthetic; else pipeRunReadable(getReadable(…)). Routes do not call getReadable.
• attachOperatorStop requires signal.aborted.
• shouldKickHotResume / streamOpened: no hot-resume unless the GET opened.
• G22 (backend-agents G22: Stop on attached viewport issues a server cancel #816) not this PR.

Tests

Targeted + --changed: 10 files, 494 passed (includes +11 new cases: bodyForRun ×6, shouldKickHotResume ×4, producer-cancel fold ×1). tsc --noEmit green.

Baseline last merge-pr full suite was 3011. Full vitest run is the merge-pr gate, not this PR-open round.

Docs

• docs/agent-stream.md End of turn
• docs/architecture.md attach getReadable

Do not merge

implement-plan stop. Next: adversarial-review, then explicit merge.

[vercel]

The latest updates on your projects. Learn more about Vercel for GitHub.

1 Skipped Deployment

Project	Deployment	Actions	Updated
invincible	Ignored		Sep 1, 2026 10:09pm UTC

[btipling]

implement-plan for #919.

• bodyForRun: already-cancelled/failed → synthetic SSE, no getReadable()
• subscribe-fail does not hot-resume (shouldKickHotResume / streamOpened)
• producer Request cancelled. without abort clears running; 6i/6j still keep-running
• tsc --noEmit + vitest run --changed: 10 files / 494 passed
• not merged; next is adversarial-review

[btipling]

Adversarial review — PR #920

Verdict: PASS WITH NOTES
Repo: btipling/invincible
Scope: main ← plan/f5-cancelled-attach · 13 files · F5 cancelled-attach unstick (bodyForRun + host hot-resume gate)
Lenses run: L1, L2 (API trust), L3, L5, L6, L8 (skip: L4 no workflow/artifact; L7 no config-seam change; L9 no palette/Wasm chrome)
AGENTS.md read: yes

Findings

Sev	Lens	Finding	Break scenario	Refutation attempt	Confidence
Minor	L8	app/api/turns/[runId]/stream/route.ts file header + handler comment still describe attach as always getRun().getReadable(). Plan-review locked only bodyForRun may call getReadable, and only on the wrap branch.	Next implementer reads the route comment, writes const readable = run.getReadable({ startIndex }); return new Response(await bodyForRun(run, readable)) (or calls getReadable then wraps). Cancelled-hours-ago F5 hits the SDK handle again → pre-headers Failed to fetch (#918).	Tests on cancelled (row 13b / POST row 5b / bodyForRun unit) would fail if the current route called getReadable. They do not stop a future agent who “fixes” the comment-shaped code. Living docs/architecture.md prose is already correct; the route comment is what implement-plan reads first.	high
Nit	L8	docs/architecture.md mermaid still labels attach getRun.getReadable. Prose four lines later is the new contract.	Diagram-only readers keep wrapping the SDK readable.	Prose is the operator/agent contract; mermaid is schematic.	high
Nit	L6	Route tests lock cancelled skip-getReadable. failed skip is only in lib/agent/pipeRunReadable.test.ts. GET row 13 is named “completed run” but uses default status: 'running'.	A route-local branch that special-cases only 'cancelled' would stay green. C16 wrap-on-completed is not locked at the route.	Routes do not branch on status; bodyForRun unit tests cover failed + completed wrap.	medium

Residual risk

Wrapped running/completed streams still call getReadable() + wrapper start()/getReader() inside the Function before the Response is returned — same shape as #917. If Preview ever proves every wrapped SDK readable pre-header-fails, this PR does not unwrap those paths (plan lock; follow-up). bodyForRun races run.status for ≤1s; hung/rejected status falls through to wrap (or GET 503) and Goal 2 (shouldKickHotResume) still stops the Busy loop, but Goal 1 (clear running) then depends on a later synthetic attach or G22 (#816). Synthetic cancelled/failed drops any buffered frames (hours-ago cancel does not need replay). Operator attach Stop/Esc still keep-running (6i/6j); dashboard cancel without abort now clears.

Merge guidance

• PASS WITH NOTES: safe to merge from this attack; nits optional (route comment + mermaid + failed/completed route locks).
• Do not merge until an explicit merge ask (merge-pr). Full vitest run is the merge gate, not this review.

What was not attacked

Live DO runner, prod Gateway, Vercel Workflows getReadable() of a real hours-ago cancelled run, Preview network waterfall, G22 cancel POST (#816).

[btipling]

Adversarial follow-up landed on 2c4115b:

• GET stream route comments + architecture mermaid now describe bodyForRun (synthetic vs wrap), not always-getReadable
• GET row 13 actually sets completed; GET row 13c / POST row 5c lock failed skip-getReadable
• Source-lock: GET/POST route code never calls getReadable( — only bodyForRun

Targeted vitest: 3 files / 85 passed. Still PASS WITH NOTES; not merged.

[btipling]

Operator still broken on production (Preview skipped, this PR not merged): #922 — hang + stacked Failed to fetch, not the one-reject case this PR targets.