chore: sync private v5.5.0 (8bb0ffd)

.cursor-plugin/plugin.json

@@ -15,7 +15,7 @@
     "anti-slop"
   ],
   "license": "PolyForm-Noncommercial-1.0.0",
-  "version": "5.4.0",
+  "version": "5.5.0",
   "homepage": "https://github.com/agent-kit-startup/agent-kit",
   "repository": "https://github.com/agent-kit-startup/agent-kit",
   "logo": "dashboard/logo-marketplace.svg",

.cursor/agent-kit.json

@@ -1,6 +1,6 @@
 {
   "schemaVersion": 1,
-  "version": "5.4.0",
+  "version": "5.5.0",
   "protected": [
     ".cursor/HANDOFF.md",
     ".cursor/agents/test-suites.md",

.cursor/commands/plan-review-triage.md

@@ -16,6 +16,7 @@ Supports **multi-path walk**: iterate multiple monitors in blocking-first then d
 - After Claude external plan review completed (monitor file exists under `.cursor/memory/plan-monitor-*.md`)
 - You want to process findings from the monitor and decide next steps
 - **Daily path:** `/run-plan` (exhaustion) and `/run-plan-all` (queue-end) continue into this Ask after wait exit 0. This slash stays HITL SoT; operators should not need to type `done` or paste triage solely to resume. ADR `2026-08-14_main-command-dogfood-audit-routing.md`.
+- **Owed-row adoption:** an explicit path invocation (`/plan-review-triage <path>`) on a monitor whose slug has a **dead** wait-state (`.cursor/context/audit-wait/<slug>.json` `status: "timeout"`/`"soft-fail"`, or `"armed"` with `now >= deadline`) and an **owed** Field Report row for that plan runs this same Step 1-5 walk unchanged, but Step 4's durable heading additionally records the adoption (see Step 4). Gap-aware skip's "no open residuals / clean Outcome" row does **not** apply to an adoption target: a clean adopted monitor still gets the Ask and the durable heading (only "already triaged" or "not a monitor file" may still skip it) — the owed row needs that heading to close. This is the `Adopt existing monitor` HITL label offered by `/run-plan` / `/run-plan-all`'s owed-close path (`.cursor/memory/errors/2026-08-14_audit-owed-ledger-no-close-path.md`); it is never automatic and never rewrites the earlier exit `3`.
 - **Not for mid-plan reviews** - this command expects `completed` work only
 
 ## Usage
@@ -150,7 +151,8 @@ Rules:
 2. **Ack and stop** must write the heading on the monitor. Updating HANDOFF alone is **not** enough: Field Report uses `isReportTriaged`, which looks for a triage heading (or a follow-up plan reference). Without the heading, the row stays untriaged.
 3. Keep HITL: do not invent a choice; do not write the heading before the user picks an option.
 4. Prefer appending once near the end of the file; do not delete prior review evidence.
-5. **Residuals executors (R15):** when closing Still open items from a residuals plan, **append** a `## Closed by residuals plan` section (ids + evidence). Do **not** rewrite or empty the reviewer's `### Still open` table in place. Prefer the monitor already committed when written so edits have history (ADR `decisions/2026-07-29_plan-monitor-staging-hygiene-r14-r15.md`).
+5. **Owed-row adoption:** when this monitor's slug has a dead wait-state and an owed Field Report row (see "Owed-row adoption" above), append the adoption fact to the same heading — for example `- **Closes owed:** yes (adoption; earlier exit 3 for this slug stays 3)`. Never narrate this as "the timed-out audit completed"; it is "an independent monitor exists and the operator adopted it." Choice `Ack and stop` on an owed row records the row as **acked**, still **unreviewed** (do not write `reviewed` for an Ack close).
+6. **Residuals executors (R15):** when closing Still open items from a residuals plan, **append** a `## Closed by residuals plan` section (ids + evidence). Do **not** rewrite or empty the reviewer's `### Still open` table in place. Prefer the monitor already committed when written so edits have history (ADR `decisions/2026-07-29_plan-monitor-staging-hygiene-r14-r15.md`).
 
 ### Step 5: Execute the choice
 

.cursor/commands/run-plan-all.md

@@ -101,6 +101,8 @@ This step is delegated to a **Task(explore) subagent** using the reusable worker
 
 **Audit session-pile preflight (before the confirm Ask):** count detached workspace-owned `agent-kit-audit-*` sessions via `.cursor/scripts/plan-external-review.sh --reap-audit-sessions --dry-run` (or the arm `--dry-run` line `audit-sessions: N detached owned`). At warn: print the dispose command (or offer `--reap-audit-sessions`) in the Ask/preflight body and continue. At cap: do not arm; surface the cap in the orchestrator Ask/preflight, not only launcher stderr. Attached sessions are operator work and are never counted. This orchestrator check runs once, at queue confirm; per-arm protection is the launcher's own session-cap refusal (exit 4), which the mid-batch and queue-end arms already honour, so no second orchestrator call site is wired.
 
+**Audits unsatisfiable-config preflight (before the confirm Ask):** when `externalPlanReview.preflight` is not `off`, check whether `enabled: true` and `backend` is pinned `"claude"` while this lane's implementer is Claude — that combination can only end owed (`CLAUDE.md` lists Claude external plan-review audits as a non-goal here; implementer≠reviewer same-model skip, ADR `decisions/2026-08-13_audits-atomic-wait-reviewer-fallback.md` point 4). `"auto"`, `"cursor"`, and `"cloud"` stay satisfiable. `midBatchAudits: true` multiplies the cost: every completed plan arms, every arm same-model-skips, every plan in the queue ends owed. `warn`: surface once, naming the combination and the three outs (`backend: "auto"`, `"cursor"`, or `"cloud"`), then continue. `block`: do not start the queue into a config that can only end owed — stop and surface the fix, or continue only on explicit operator deferral; never steal `/git-prod`. This runs before the confirm Ask because that is the one place an operator can still change config cheaply, ahead of a multi-plan run. A growing owed pile is not an acceptable substitute for surfacing this (`.cursor/memory/errors/2026-08-14_audit-owed-ledger-no-close-path.md`).
+
 After synthesis, present the proposal using **Ask questions** tool. Include the ordered list, key overlaps/consolidations, and coherence notes in the question body. Fallback to chat numbered list if the tool is unavailable.
 
 > "Plans synthesized. Proposal: [N] plans in order, [M] consolidations, [K] overlaps. Here is the proposed execution queue..."
@@ -187,14 +189,14 @@ Read `externalPlanReview` before the queue confirm Ask and at each advance:
 
 | Config | Behavior |
 |--------|----------|
-| Audits **pre-flight** (`preflight`: `off` \| `warn` \| `block`) | Before the confirm Ask and before each mid-queue advance: same owed/untriaged check as `/run-plan`. `block` arms or stops; never steals `/git-prod`. |
+| Audits **pre-flight** (`preflight`: `off` \| `warn` \| `block`) | Before the confirm Ask and before each mid-queue advance: same owed/untriaged check as `/run-plan`, plus the unsatisfiable-config check (`enabled: true` + pinned `backend: "claude"` in this lane can only end owed; see "Audits unsatisfiable-config preflight" above). `block` arms or stops; never steals `/git-prod`. |
 | `midBatchAudits: true` and audits enabled | After each plan Task returns `outcome: completed`, the **orchestrator** arms **one** full audit for that plan with `--force --autonomous --wait-monitor` (or one `--batch` + wait_all when batching is intentional) **before** advancing the cursor. No paste Ask between plans. Soft-fail → Field Report owed; still advance. AwaitShell until exit `0|3|4` (chat slice ~90s; remaining budget in `.cursor/context/audit-wait/<slug>.json`). **Exit 3 with remaining `waitTimeoutSeconds`:** do not treat as arm-done. Same orchestrator session resumes wait-state polling (re-arm `--wait-monitor` against leftover budget) before advancing the cursor or skipping triage. Exit 3 with zero leftover budget, or exit 4: Field Report owed, then advance. Wait success requires a **fresh** monitor after arm start. Reviewer cascade: `backend: "auto"` uses Claude (Haiku) when usable, else Cursor Agent. Same-model implementer/reviewer is an honest skip. Do **not** fan out N background sessions without wait. Do **not** insert a mid-queue triage Ask (operator non-stop preserved; record ready path for queue-end). Mid-batch stays findings-only: **never** auto-Write residuals or rewrite the Run queue between plans. |
 | `midBatchAudits` false/missing | **Non-stop** mid-queue: do **not** pause for audit Ask/paste between plans. Mid-queue completed plans stay Field Report **owed** until reviewed. |
 | Queue exhausted | Final HANDOFF; cadence `batch-complete`; then queue-end audit arm covering remaining owed/unreviewed targets (enabled → `--force --autonomous --wait-monitor` or paste per `mode`; else `offerOnExhausted` Ask). Prefer one launcher `--batch` + wait_all when multiple basenames. After wait exit `0`: run `/plan-review-triage` Ask with an **explicit path list** of fresh monitors (batch uniform Ask when outcomes match; sequential fallback when mixed; durable heading per file). **Batch exhaust without conveyor:** when remaining monitors are process-only / depth-capped, prefer uniform **Ack and stop** or **Fix nits only**; do not spawn unbounded `close-*` backlog from Write residuals (ADR `decisions/2026-08-11_plan-audit-residuals-termination.md`). Then suggest `/git-prod` if staging is ahead of `main` (separate HITL). |
 
 Never steal `/git-prod` confirmation. Chat never runs silent headless `--force` / `claude -p` in the agent shell. Spawn-only exit 0 without `--wait-monitor` is **not** review done. Never stop at Final HANDOFF "when monitors exist, run triage" after arming: wait (freshness) then continue (mid-batch waits for file only; queue-end waits then triage Ask with explicit paths). ADR: `2026-07-27_audits-autonomous-plan-review-contract.md` (supersedes queue-end-only); wait freshness: `2026-07-27_audits-wait-freshness-enforce.md`.
 
-**Exit 3 stays timeout-only across the queue.** A mid-queue or queue-end arm that returns `3` reviewed nothing: leave that plan Field Report **owed**, keep its path out of the queue-end triage list, and never narrate it as reviewed. **Same-session resume:** when leftover `waitTimeoutSeconds` remains in `.cursor/context/audit-wait/<slug>.json`, the same orchestrator session must keep polling (re-arm `--wait-monitor`) before advancing the cursor or skipping `/plan-review-triage`. A later session may poll leftover budget; that is fallback, not the default while this session is still open. Do not treat a first-slice exit `3` as "arm done, continue the queue." Monitors that show up later, including monitors written by a different arm or a later queue position, do **not** retroactively upgrade an earlier `3`. Exit `4` covers the launcher soft-fails: no usable reviewer (`backend: "auto"` tried Claude then Cursor; pinned `claude` still tips when Claude is missing), same-model refuse, background spawn unavailable, a **silent PTY** early abort (spawn succeeded but produced no scrollback in the grace window), and a **session-cap refusal** (detached `agent-kit-audit-*` pile at the cap, so nothing spawned). Advance the queue on exit 4 or on exit 3 with zero leftover budget, but record the target as owed, never as reviewed. ADR: `2026-07-30_audits-pty-progress-gate-zombie-policy.md`; wait resume: `2026-08-13_audits-atomic-wait-reviewer-fallback.md`.
+**Exit 3 stays timeout-only across the queue.** A mid-queue or queue-end arm that returns `3` reviewed nothing: leave that plan Field Report **owed**, keep its path out of the queue-end triage list, and never narrate it as reviewed. **Same-session resume:** when leftover `waitTimeoutSeconds` remains in `.cursor/context/audit-wait/<slug>.json`, the same orchestrator session must keep polling (re-arm `--wait-monitor`) before advancing the cursor or skipping `/plan-review-triage`. A later session may poll leftover budget; that is fallback, not the default while this session is still open. Do not treat a first-slice exit `3` as "arm done, continue the queue." Monitors that show up later, including monitors written by a different arm or a later queue position, do **not** retroactively upgrade an earlier `3` — the `3` stays `3` even when a later genuine monitor exists. Exit `4` covers the launcher soft-fails: no usable reviewer (`backend: "auto"` tried Claude then Cursor; pinned `claude` still tips when Claude is missing), same-model refuse, background spawn unavailable, a **silent PTY** early abort (spawn succeeded but produced no scrollback in the grace window), and a **session-cap refusal** (detached `agent-kit-audit-*` pile at the cap, so nothing spawned). Advance the queue on exit 4 or on exit 3 with zero leftover budget, but record the target as owed, never as reviewed. **Leftover budget is `deadline` vs wall clock, never `status: "armed"` alone:** the launcher expires a wait-state file left `armed` past its `deadline` on contact (to `status: "timeout"`, `remainingBudgetSeconds: 0`), so a stale arm from an earlier queue run is never resumed as live budget; sweep all slugs with `.cursor/scripts/plan-external-review.sh --gc-wait-state [--dry-run]`. **Owed close (separate, later event, not an upgrade of the `3`):** once the wait-state for that slug is terminal-and-dead (`status: "timeout"`/`"soft-fail"`, or `"armed"` with `now >= deadline`) and a genuine post-hoc monitor for the slug exists, `/run-plan`'s owed-close HITL applies — `Adopt existing monitor` (into `/plan-review-triage`, closes as reviewed-by-adoption) or `Ack owed without review` (closes as acked/unreviewed); a duplicate re-arm against already-merged work is not the only route. Queue-end triage lists still exclude dead-timeout rows by default; adoption is operator-initiated per row. ADR: `2026-07-30_audits-pty-progress-gate-zombie-policy.md`; wait resume: `2026-08-13_audits-atomic-wait-reviewer-fallback.md`; owed-close: `.cursor/memory/errors/2026-08-14_audit-owed-ledger-no-close-path.md`.
 
 ### External plan review (legacy heading)