backend-agents C (plan #789): reserved meta.turnRunId/turnStatus + host detach (unmount ≠ abort run)

[btipling]

Summary

Implements plan #789 (backend-agents slice C, source #766): stop treating
the tab as the turn lifetime.

• Reserved meta.turnRunId session carrier (Redis-safe opaque via the
  existing isRedisSafeOpaqueId rule, ^[A-Za-z0-9_-]{1,512}$) + optional
  meta.turnStatus (idle|running|cancelling, a cached hint only —
  getRun(runId) is the authority when an id is present). Both follow the
  meta.selectedModel drop-to-unset-on-poison, never 400 discipline.
• Host detach: HarnessHost unmount / session switch / New/Clear now call a
  single detachTurn() seam (lib/detachTurn.ts decideDetach) that keys on
  turnRunId presence:
  • Durable run (id present, post-E): close-reader only — never abort the
    run, never a server cancel. The run identity survives on the envelope.
  • Legacy tab-owned turn (no id, slice-C reality): still aborts the
    /api/agent fetch so a detached busy tab never leaves an unpersisted 1800 s
    Function burning with no writer (the plan-review Major guard).
  • No run id ⇒ idle (today's Ready); Stop/Esc (takePendingCancel) still
    cancels, isolated from the detach path.
• No Wasm/protocol change, no queue persist, no Workflow port (E/G/H own
  those). This slice ships the carrier + detach seam only.

Files

• lib/sessionCloudCaps.ts — client-safe sanitizeTurnRunId / sanitizeTurnStatus,
  TurnStatus type, NEW cap TURN_STATUS_MAX_BYTES = 32.
• lib/sessions/sessionStore.ts — both keys in RESERVED_META_KEYS + dedicated
  drop-to-unset branches in validateMeta (never 400; STRICT unknown-key intact).
• lib/sessionStore.ts (local snapshot) — fields + sanitize-on-load.
• lib/sessionRepository.ts — fold into cloudMetaFor, restore on pull/adopt
  (parseCloudSessionSnapshot, overlayEnvelopeMeta); emit-when-set, omit=clear.
• app/harness/HarnessHost.tsx — detachTurn() seam wired to unmount/switch/
  Clear/New; Stop/Esc (takePendingCancel) unchanged.
• lib/detachTurn.ts (+ test) — pure detach-vs-abort guard.
• Docs: docs/session-model.md (reserved-meta table + schema list),
  docs/feature-divide.md (unmount = detach row). AGENTS/README/SECURITY/.env.example N/A (one-line whys in plan).

Verification

• npm run typecheck → exit 0.
• Targeted + full suite run directly with vitest (local binary):
  • baseline full vitest run → 2046 passed (merge feat(backend-agents): turns runtime spike — resumable AgentStreamEvent stream (plan #787) #788); now 2061 passed — +15 new test cases across the added/changed files (rows 1–9 of the plan's test matrix).
  • Full vitest run: 133 files / 2061 tests passed, failed=0, exit 0.
  • npm run test:di-gate → OK (no new I/O construction).
• Cap governance: no existing cap raised/lowered → no human gate. TURN_STATUS_MAX_BYTES=32 is a NEW generous cap (plan Caps table); turnRunId reuses REDIS_SAFE_OPAQUE_ID_MAX=512. Both ride the existing 1 MiB whole-meta budget / 4.5 MB Function ceiling.
• Cloud ops: N/A — no Production mutate (reserved-meta carriers + host detach only). No workflow_dispatch change.

Closes #789 · Refs #766 (source) · #764 (parent) · #710 (umbrella) · #765 / #767 (registry B/D).

[vercel]

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

1 Skipped Deployment

Project	Deployment	Actions	Updated (UTC)
invincible	Ignored		Aug 23, 2026 5:04am

[btipling]

Implemented plan #789 (backend-agents slice C). Adds reserved meta.turnRunId/meta.turnStatus carriers with drop-to-unset-on-poison, and a host detachTurn() seam (unmount/switch/New-Clear = detach; abort only a legacy no-run-id turn; Stop/Esc still cancels). No Wasm/protocol change.

Verification: typecheck exit 0; full vitest run (direct local binary) 133 files / 2061 tests passed, failed=0, exit 0 (baseline 2046 → +15 new); test:di-gate OK. Cap governance: no existing cap raised/lowered; TURN_STATUS_MAX_BYTES=32 NEW tabled; turnRunId reuses REDIS_SAFE_OPAQUE_ID_MAX=512.

[btipling]

Adversarial review — PR #790

Verdict: PASS WITH NOTES
Repo: btipling/invincible
Scope: main ← plan/turns-detach · 13 files · backend-agents C (reserved meta.turnRunId/turnStatus + host detach seam)
Lenses run: L1, L2, L3, L5, L6, L7, L8 (skip: L4 — no GHA/CI/artifact files; L9 — no palette/UX chrome)
AGENTS.md read: yes (docs/feature-divide.md unmount=detach row; SECURITY.md session envelope trust)

Findings

Sev	Lens	Finding	Break scenario	Refutation attempt	Confidence
Minor	L1	Plan edge case + docs/feature-divide.md claim session switch / New / Clear mid-turn = detach. onSwitchSession / onNewSession / onClear still return on inflightRef.current before detachTurn() (app/harness/HarnessHost.tsx). The only tear-down that can fire while a turn is in flight is unmount.	Start a turn. Click another rail session / New / Clear. Host no-ops; viewport stays Busy. Detach-vs-abort never runs. Tab close is the only live-turn detach. Post-E, a durable run still cannot be abandoned by switch — user is stuck until the Function ends or the tab unmounts.	Inflight guards are the #642 leftover-Send lock and pre-exist this PR; step 4 of the plan listed unmount/switch/clear calling detach without saying “lift the guard.” Fails: the plan’s own edge case (“Session switch / New / Clear mid-turn: same detach discipline — abandonment is not cancellation”) and the feature-divide sentence state it as current truth. The new detachTurn() lines on those three paths are dead for a live turn.	high
Minor	L6	No Host wiring proof that unmount uses detachTurn(). Plan row 7 allowed an extracted helper; lib/detachTurn.test.ts covers decideDetach only. Unmount cleanup in HarnessHost.tsx is untested. This repo already source-locks Host call sites (lib/harnessHostModelPersist.test.ts, lib/sessionSummaryLabel.test.ts).	Revert the unmount cleanup from detachTurn() back to abortRef.current?.abort(). decideDetach tests stay green. Default (no turnRunId) still aborts, so C looks the same — the durable-run close-reader path is the untested wiring.	Helper extraction was an allowed DoD option; the four call sites are one-liners. Fails: the #710-class behavior lives in the cleanup, not the helper. A source lock that unmount/switch/new/clear call detachTurn() and that the remaining abortRef.abort() sites are only runPrompt (replace controller) and takePendingCancel (Stop/Esc) would pin the seam.	high

Residual risk

decideDetach keys on presence of a client-writable meta.turnRunId (envelope PUT accepts any Redis-safe opaque; drop-to-unset, never 400). At C the host never mints an id (/api/agent is still the turn owner), so a session that already carries a valid-looking id will not abort on unmount — the plan-review Major (unpersisted 1800s Function) returns for that session. Default snapshots omit the field → abort (today’s behavior). close-reader is a documented no-op until F. Logout relies on unmount after signOut (LogoutButton does not call detachTurn). getRun is not consulted (hint vs authority is docs-only this slice).

Merge guidance

• PASS WITH NOTES: safe to merge from this attack; nits optional
• E/F should lift or justify the inflight guards when switch-away must detach a durable run, and should not trust a client-supplied id without a real Workflow run

What was not attacked

Live Workflow turn owner (E). Viewport attach / reader close (F). Stop→Workflow cancel (H). Prod Gateway. Self-hosted runner. Dual-chat. Wasm protocol (no native/ in diff).

[btipling]

Plan to address feedback on PR #790 (Status: DONE — all round-1 + round-2 items addressed)

Last updated: 2026-08-23

Feedback items (most complex → least)

#	Severity	Review item	Status
1	Minor	L1 — switch/New/Clear mid-turn detach is dead code behind the inflight guard; docs overclaim	🟢 done
2	Minor	L6 — no Host wiring proof unmount uses detachTurn(); add a source lock	🟢 done
3	Nit	L8 (round-2) — docs/feature-divide.md detach row cites "PR #790 review L1" (PR-review archaeology in a living ownership table); optional trim	🟢 done

#1 — switch/New/Clear detachTurn() is dead for a live turn (Minor, L1) (🟢 DONE)

• Review item: onSwitchSession / onNewSession / onClear still return on
  inflightRef.current / switchInFlightRef.current before detachTurn()
  (app/harness/HarnessHost.tsx), so the only tear-down that can fire while a turn
  is in flight is unmount. The plan edge case + docs/feature-divide.md row
  claim "session switch / New / Clear mid-turn = detach" as current truth.
• Grounded problem (verified on head 7a9cb85c): onClear:831 returns on
  inflightRef.current || switchInFlightRef.current before detachTurn() at :834;
  onNewSession:911 returns on the same guard before detachTurn() at :916;
  onSwitchSession:936-937 returns on inflightRef.current /
  switchInFlightRef.current before detachTurn() at :941. The unmount cleanup
  (:739) calls detachTurn() unconditionally — the only live-turn detach site
  at slice C. docs/feature-divide.md:47 overclaimed the switch/New/Clear paths.
• Severity assessment: Minor. No runtime bug at slice C (those detachTurn()
  calls are reachable only when not in flight, where they degrade to today's
  no-op). The defect is a claim-vs-reality gap in living docs plus a seam that
  E/F must lift the guard to actually use for durable-abandon.
• Strategies considered:
  • A) Document honestly + defer the guard lift to E/F — fix the
    docs/feature-divide.md row so it states the seam is wired on all four
    tear-down sites but switch/New/Clear are gated by the pre-existing harness: session list in the Wasm transcript rail (protocol v17) #642
    inflight lock, so during a live turn only unmount fires detach; E/F lifts
    or justifies the guards for durable-abandon (reviewer's own merge guidance).
    (→ CHOSEN — matches the review's "E/F should lift OR JUSTIFY"; lifting the
    guard here would change the harness: session list in the Wasm transcript rail (protocol v17) #642 leftover-Send lock in a slice-C carrier/lock
    PR and at C (no durable run) would make a busy-legacy-turn switch abort the
    turn you're actively watching — out of scope.)
  • B) Lift the inflight guard on the three paths so switch/New/Clear detach a
    durable run — rejected: scope creep into harness: session list in the Wasm transcript rail (protocol v17) #642 behavior; unsafe at C where
    /api/agent is still the turn owner (would cancel the busy turn on switch).
  • C) Remove the dead detachTurn() calls — rejected: would remove the E/F seam
    the plan's step 4 intends; L6 wants the sites pinned, not deleted.
• Open questions: none — resolved by choosing A.
• Fix (locked, applied in commit 31d7f78): edited docs/feature-divide.md:47
  to state honestly: seam wired on unmount + switch + New/Clear; during an
  in-flight turn the switch/New/Clear paths are held by the pre-existing harness: session list in the Wasm transcript rail (protocol v17) #642
  inflight guard, so unmount is the live-turn detach site at slice C; E/F
  lifts/justifies the guard when switch-away must detach a durable run. No Host
  code change (behavior at C is correct). The plan-issue edge case is superseded
  by this living-doc truth + the E/F follow-up.
• Tests: none needed for the doc fix (docs-only); the L6 source lock (§2.2 Install and register GitHub Actions self-hosted runner #2)
  independently pins the four detachTurn() call sites.

#2 — Host wiring source lock (Minor, L6) (🟢 DONE)

• Review item: no proof unmount uses detachTurn(); lib/detachTurn.test.ts
  covers decideDetach only. Revert the unmount cleanup and decideDetach tests
  stay green — the backend-agents: client-independent turns — in-flight + queue survive refresh / tab close #710-class behavior lives in the cleanup, not the helper.
• Grounded problem (verified on head 7a9cb85c): lib/detachTurn.test.ts
  tested only the pure decideDetach decision. Unmount (HarnessHost.tsx:739),
  switch (:941), New (:916), Clear (:834) call detachTurn(). The remaining
  abortRef.current?.abort() sites are detachTurn legacy path (:384),
  runPrompt controller replace (:410), and the Stop/Esc takePendingCancel
  poll (:662).
• Severity assessment: Minor. No defect today (unmount IS wired), but the next
  person could revert it and all tests stay green — the seam would silently vanish.
  Matches the repo's existing source-lock precedent
  (lib/harnessHostModelPersist.test.ts "HarnessHost wiring lock").
• Strategies considered:
  • A) Add a source-lock test (node readFileSync on HarnessHost.tsx, matching
    the existing wiring-lock pattern) asserting the four detachTurn() call sites
    and that the only abortRef.current?.abort() sites are detachTurn,
    runPrompt, and the Stop/Esc takePendingCancel path. (→ CHOSEN)
  • B) Extract + unit test a small host function that wraps the unmount cleanup —
    heavier than warranted; C is pure-cleanup already.
• Fix (locked, applied in commit 31d7f78): added a describe('HarnessHost wiring lock — tear-down uses detachTurn (PR #790 review L6)') block to
  lib/detachTurn.test.ts that (1) asserts detachTurn(); appears exactly 4
  times, each anchored on its // Plan #789 (source #766): … comment; and (2)
  asserts abortRef.current?.abort(); appears exactly 3 times, with the three
  allowed contexts (if (decision.kind === 'abort') {, the runPrompt controller
  replace, and if (b.takePendingCancel()) {).
• Tests: new source-lock row in lib/detachTurn.test.ts — now 5 tests
  (was 4). Gate: npm run typecheck exit 0; node scripts/di-gate.mjs OK;
  vitest run on the file + dependents = 112 passed / exit 0; npm run build
  green.

#3 — feature-divide detach row cites "PR #790 review L1" (Nit, L8 round-2) (🟢 DONE)

• Review item: docs/feature-divide.md detach row ends "E/F lifts or justifies
  those guards when switch-away must detach a durable run (PR backend-agents C (plan #789): reserved meta.turnRunId/turnStatus + host detach (unmount ≠ abort run) #790 review L1)" — a
  PR-review archaeology citation in a living ownership table. A clone-and-run reader
  would have to open this PR to decode the sentence; the inflight-lock fact is
  already stated in the same cell.
• Severity assessment: Nit, optional trim. Not a correctness issue; purely a
  living-docs hygiene point (counts as PR-review citation, not a plan/issue id).
• Fix (applied in commit 4a6f3cf): removed the " (PR backend-agents C (plan #789): reserved meta.turnRunId/turnStatus + host detach (unmount ≠ abort run) #790 review L1)" PR-review
  reference, leaving "E/F lifts or justifies those guards when switch-away must
  detach a durable run." Plan plan: backend-agents C — session meta.turnRunId + detach (unmount ≠ abort run) #789 / source backend-agents C: session meta.turnRunId + detach (unmount must not abort the run) #766 ids and the harness: session list in the Wasm transcript rail (protocol v17) #642 inflight-lock fact
  remain in the cell. Docs-only change.
• Tests: none needed (docs-only). No TS/Zig diff → no test gate per skill scope.

Progress log (append-only)

• 2026-08-23 — created plan; 2.1 Provision DigitalOcean droplet for builds #1 (Minor L1) + 2.2 Install and register GitHub Actions self-hosted runner #2 (Minor L6) queued from round-1
  adversarial-review (PASS WITH NOTES on head 7a9cb85c).
• 2026-08-23 — commit 31d7f78 on plan/turns-detach: 2.1 Provision DigitalOcean droplet for builds #1 doc fix
  (docs/feature-divide.md:47 honest unmount-is-the-live-turn-detach-site,
  harness: session list in the Wasm transcript rail (protocol v17) #642-guard gating + E/F follow-up) and 2.2 Install and register GitHub Actions self-hosted runner #2 source lock
  (lib/detachTurn.test.ts +1 row). Gates green (typecheck, di-gate, vitest file
  • dependents 112 passed, build exit 0). Pushed 7a9cb85..31d7f78.
    → both items 🟢 done.
• 2026-08-23 — round-2 adversarial-review on head 31d7f78 = PASS WITH NOTES with
  one new Nit (L8). Commit 4a6f3cf on plan/turns-detach: 1.1 Create GitHub repo invincible #3 removed the
  "PR backend-agents C (plan #789): reserved meta.turnRunId/turnStatus + host detach (unmount ≠ abort run) #790 review L1" archaeology from the docs/feature-divide.md detach row
  (docs-only). Pushed 31d7f78..4a6f3cf → 1.1 Create GitHub repo invincible #3 🟢 done.

Merge decision

Round-1 items (#1, #2) and round-2's only Nit (#3) all addressed. PR #790 is
merge-ready; per flow, do not self-merge — invoke /merge_pr #790 (mandatory full
vitest run) after a confirming re-review on head 4a6f3cf.

[btipling]

Adversarial review — PR #790 (round 2)

Verdict: PASS WITH NOTES
Repo: btipling/invincible
Scope: main ← plan/turns-detach · 13 files · HEAD 31d7f78 (backend-agents C)
Lenses run: L1, L2, L3, L5, L6, L7, L8 (skip: L4 — no GHA/CI/artifact files; L9 — no palette/UX chrome)
AGENTS.md read: yes (docs/feature-divide.md detach row; SECURITY.md session envelope trust)

Round 1 PASS WITH NOTES: inflight guards make switch/New/Clear detach dead for a live turn (docs overclaim); no Host wiring lock that unmount calls detachTurn().

Prior items

Prior	Status on 31d7f78
Minor L1 switch/New/Clear mid-turn detach is dead; feature-divide overclaim	Closed (docs). Row now states: seam wired on four sites; during an in-flight turn only unmount fires detach; #642 inflight lock holds the other three; E/F lifts or justifies. Host behavior unchanged (correct at C).
Minor L6 no Host proof unmount uses detachTurn()	Closed. lib/detachTurn.test.ts source-locks 4× detachTurn(); with comment anchors + 3× abortRef.current?.abort(); (detachTurn / runPrompt replace / takePendingCancel). Reverting unmount to a direct abort drops the detach count to 3 and bumps abort to 4.

Findings

Sev	Lens	Finding	Break scenario	Refutation attempt	Confidence
Nit	L8	docs/feature-divide.md detach row cites “PR #790 review L1” (issue/PR archaeology in a living ownership table).	A clone-and-run reader has to open this PR to decode the sentence. The inflight-lock fact is already in the same cell.	Plan #789 living-docs rule was timeless / no phase theater; the repo’s tables routinely cite plan numbers. Fails only as a PR-review citation, not a plan id. Optional trim.	high

Residual risk

decideDetach still keys on presence of a client-writable meta.turnRunId (envelope PUT accepts any Redis-safe opaque). At C the host never mints an id, so a session that already carries a valid-looking id will not abort on unmount — the plan-review Major (unpersisted 1800s Function) returns for that session. Default snapshots omit the field → abort (today’s behavior). close-reader is a no-op until F. Logout still relies on unmount after signOut. Source lock is regex-shaped (abortRef.current?.abort();) — an abort() without ?. would not bump the count.

Merge guidance

• PASS WITH NOTES: safe to merge from this attack; nits optional
• E/F: lift or justify the inflight guards when switch-away must detach a durable run; do not trust a client-supplied id without a real Workflow run

What was not attacked

Live Workflow turn owner (E). Viewport attach / reader close (F). Stop→Workflow cancel (H). Prod Gateway. Self-hosted runner. Dual-chat. Wasm protocol (no native/ in this diff).

[btipling]

adversarial-review — PASS (VERDICT: PASS)

Target commit: 4a6f3cf (current head). Fresh review after the head was pushed for the round-2 Nit L8 doc fix.

Head delta reviewed (31d7f78..4a6f3cf)

Single commit, docs-only: docs(feature-divide) removes the (PR #790 review L1) archaeology citation from the detach row (docs/feature-divide.md:47). One line changed; the substantive clause E/F lifts or justifies those guards when switch-away must detach a durable run is retained verbatim. No TS/Zig/code surface touched.

Round findings — all closed, re-confirmed on this head

• L1 (Minor): detach dead-behind-inflight-guard docs overclaim — fixed; doc now states only unmount fires detach during an in-flight turn. ✓
• L6 (Minor): HarnessHost wiring source lock added in lib/detachTurn.test.ts (4 detach / 3 abort sites). ✓
• L8 (Nit): PR archaeology dropped from ownership table. ✓

Independent verification on this head

• npm run typecheck → exit 0
• Full node_modules/vitest/vitest.mjs run → 133 files / 2062 tests passed, exit 0 (direct vitest, no wrapper)
• node scripts/di-gate.mjs → di-gate OK; no new PGlite(/createDbConnection( in any changed test file

Verdict: PASS. No blocking/major/minor findings survive on head 4a6f3cf; the change is a safe, non-behavioral doc clean-up matching the intended round-2 alignment.