fix(workflows): hold one writer for durable live SSE

[btipling]

Summary

Durable live SSE was calling getWritable() per token. Each call opens a new Workflows serialize pipe to the same Redis stream; chunks land out of order. The host appends whatever still parses → word-salad assistant, frozen thinking, dead reader.

This PR holds one writer for the model-step burst (withDefaultStreamWriter around generateOneRound). Sparse loop writes (writeTurnSse → writeOnDefaultStream) stay one write per step.

Implements #855 (HANDOFF-READY). Source bug #854.

Fixes #855
Refs #854
Refs #794
Refs #850

What changed

• lib/workflows/turnSseWrite.ts — withDefaultStreamWriter (one getWritable / held getWriter / releaseLock in finally; does not close).
• lib/workflows/modelGenerateStep.ts — live onEvent writes through that held writer. No writeOnDefaultStream in this step.
• docs/agent-stream.md + AGENTS.md — one writer per model round, not per token.

Tests

	Count
Baseline (main @ 2f77992)	2346 passed / 146 files / failed=0
This PR (targeted + --changed)	+5 new rows in lib/workflows/turnSseWrite.test.ts
Expected full suite	2351 passed / 147 files

New rows: N writes → 1 pipe; releaseLock on fn throw; releaseLock on mid-burst write throw; ordered payloads concatenate; sparse writeOnDefaultStream still per-call.

npm run typecheck green. Targeted: turnSseWrite / turnWorkflow / turnLoop / staticGraph / turnSseFormat 68 passed. vitest run --changed 66 passed / failed=0. Full vitest run is the merge-pr gate.

Out of scope

• Host / Wasm / parseSseChunk
• E19 attach (backend-agents E19: attach handshake — Blob hydrate + stream attach #813)
• EOF detach (fix(harness): treat durable stream EOF as detach, not Turn ended #853)
• Seq numbers on AgentStreamEvent

Operator smoke (after Preview)

Long durable Busy turn (/plan_review or reasoning): thinking grows in order, assistant is readable English, turn does not die mid-sentence while Workflows is still working.

[vercel]

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

1 Skipped Deployment

Project	Deployment	Actions	Updated (UTC)
invincible	Ignored		Aug 26, 2026 12:56am

[btipling]

implement-plan for #855.

• Held one Workflows writer for the model-step burst (withDefaultStreamWriter). Per-token getWritable() is gone from modelGenerateStep.
• writeOnDefaultStream remains the sparse writeTurnSse path.
• Living docs: docs/agent-stream.md + AGENTS.md where-to-change.

Verify: npm run typecheck exit 0. Targeted workflow tests 68 passed. vitest run --changed 66 passed / failed=0. Full suite is the merge-pr gate (expected 2346+5 = 2351).

Not merged. Next: adversarial-review on this PR.

[btipling]

Adversarial review — PR #856

Verdict: PASS WITH NOTES
Repo: btipling/invincible
Scope: main ← plan/durable-stream-writer-hold · 8 files · durable live-SSE writer hold
Lenses run: L1, L2, L5, L6, L7, L8 (skip: L3 no host/Wasm/dual-chat; L4 no GHA/artifacts; L9 no UI)
AGENTS.md read: yes (SSE where-to-change, Workflows, DI/test-cost). SECURITY.md + docs/feature-divide.md also read.

Findings

None. Attack did not breach with current evidence.

Candidate breaks that died in self-refutation (not findings): getWritable() throw moving outside onEvent is still closed by runTurnLoop catch → fail() + turnWorkflow finally closeTurnSse; write_error from generateOneRound still releaseLocks because the helper awaits fn and finallys; writeTurnSse cannot race the held lock because it is a later 'use step' and the SDK pushes state.promise onto ctx.ops (flush on releaseLock, poll 100ms, pendingOps === 0); BYOK/assemble stay outside the hold; DI/cost gate clean (mocked workflow, no PGlite/Redis).

Residual risk

Hold-across-streamText is the SDK chat example (streamTextStep keeps one getWriter for the UI-message read loop) and getWritable() in a step really does construct a new serialize pipe every call (@workflow/core step/writable-stream.js). That is the #854 cause this PR removes. It is not proven on a live Workflows Preview: if the VM rejects a writer held across generateOneRound awaits, the named error is Not supported in workflow functions (or equivalent) — plan forbids shipping the per-line fallback unless that error is observed. Replay of modelGenerateStep would re-append tokens to the same Redis stream (pre-existing). Operator close-out for #854 remains a long durable Busy turn after deploy (thinking in order, English assistant, no mid-sentence death while the run is still working). This review did not execute vitest or a live Workflows run.

Merge guidance

• PASS WITH NOTES: safe to merge from this attack; nits optional (none logged)
• Do not treat this as Preview proof of #854. merge-pr still owns the full vitest run gate

What was not attacked

Live Vercel Workflows / Redis stream, prod Gateway, host parseSseChunk / growAssistant / Wasm (out of scope by plan), E19 attach (#813), EOF detach (#853), DO runner, full vitest process.