agent-kit-startup · agent-kit-startup · Aug 15, 2026 · Aug 15, 2026
diff --git a/.cursor-plugin/plugin.json b/.cursor-plugin/plugin.json
@@ -15,7 +15,7 @@
     "anti-slop"
   ],
   "license": "PolyForm-Noncommercial-1.0.0",
-  "version": "5.2.1",
+  "version": "5.3.0",
   "homepage": "https://github.com/agent-kit-startup/agent-kit",
   "repository": "https://github.com/agent-kit-startup/agent-kit",
   "logo": "dashboard/logo-marketplace.svg",

diff --git a/.cursor/agent-kit.json b/.cursor/agent-kit.json
@@ -1,6 +1,6 @@
 {
   "schemaVersion": 1,
-  "version": "5.2.1",
+  "version": "5.3.0",
   "protected": [
     ".cursor/HANDOFF.md",
     ".cursor/agents/test-suites.md",

diff --git a/.cursor/commands/backlog-add.md b/.cursor/commands/backlog-add.md
@@ -36,6 +36,8 @@ Example: `/backlog-add Polish Mission Control empty-state icons`
 
 Same planning blockers as `/start-project`: unresolved essential readiness → point to `/agent-kit-onboard`. Non-essential pending is advisory only.
 
+**Never-Ask:** `confirm-provider` / `collaboration.provider` must not become an Ask. One-line advisory or silence, then continue Broad Intake → propose → write Ask. Do not halt enqueue. Distinct from an essential-readiness hard stop (that still points to `/agent-kit-onboard`). The Broad Intake worker inherits this: do not return a readiness-gate Ask for that check.
+
 ## What to Do
 
 ### 1. Broad Intake Review (required before plan proposal)
@@ -52,7 +54,7 @@ Before enqueueing a new plan, **scan** these sources (read/skim; do not deep-div
 | **Archived context** | Prior packs for same theme | `.cursor/context/archive/**` (if present; glob by topic) |
 | **Decisions** | ADRs that constrain the goal | `.cursor/memory/decisions/`, `_index.md` Decisions table |
 | **Memory** | Errors, audits, consolidations, review logs, plan-monitors, findings audits | `.cursor/memory/errors/`, `.cursor/memory/plan-monitor-*.md`, theme-matched `plan-review-*.md`, `_index.md` (Audits + Decisions) |
-| **Unprocessed dogfood** | Factory/consumer inbox notes awaiting triage (not sessionStart-only) | `dogfood/README.md` or `.cursor/dogfood/README.md` under `##` or `### Unprocessed Files`; skim titles/summaries only. Missing/empty inbox → no findings. Labels: ignore (owned by open plan), error/include (kit gap), note (inbox evidence only). Never auto-analyze or memory WRITE (ADR `decisions/2026-08-11_dogfood-unprocessed-broad-intake-bucket.md`) |
+| **Unprocessed dogfood** | Factory/consumer inbox notes awaiting triage (not sessionStart-only) | `dogfood/README.md` or `.cursor/dogfood/README.md` under `##` or `### Unprocessed Files`; skim titles/summaries only. Missing/empty inbox → no findings. Labels: ignore (owned by open plan), error/include (kit gap), note (inbox evidence only). Never auto-analyze or memory WRITE (ADRs `decisions/2026-08-11_dogfood-unprocessed-broad-intake-bucket.md`, `decisions/2026-08-14_main-command-dogfood-audit-routing.md`) |
 | **Local docs** | SoT / inventories / getting-started that the goal touches | `docs/**`, especially files named in the payload or related SoT |
 | **Working tree** | Uncommitted local work that would collide | `git status`, `git diff` (staged + unstaged); do not commit |
 | **Recent commits** | What already shipped for this theme | `git log` (short, recent), related PR titles if available |
@@ -83,6 +85,13 @@ The actual scanning and triage is delegated to a **Task(explore) subagent** usin
 
 **Fallback:** If Task dispatch is unavailable, run the Broad Intake Review inline (same as pre-delegation behavior).
 
+**Unprocessed inbox Ask (when non-empty):** after the skim, if Unprocessed has rows, **Ask questions** (one question; chat numbered-list fallback) with labels exactly:
+- `Analyze inbox now` — start the ingest ritual (analyze → memory WRITE → triage). Notes become plans/memory after HITL, never `plan-monitor-*.md`. Then continue this enqueue.
+- `Enqueue Fix now` — fold include/error Unprocessed rows into this proposal (already in Broad Intake). Do not start a nested `/backlog-add`.
+- `Not now` — continue the write Ask. Treat remaining inbox rows as `note` only.
+
+Never auto-analyze. Empty or missing inbox: skip this Ask.
+
 ### 2. Vague goal
 
 If the goal is missing or vague, use **Ask questions** (chat numbered-list fallback):

diff --git a/.cursor/commands/continue-plan.md b/.cursor/commands/continue-plan.md
@@ -19,7 +19,11 @@ Resume a plan from the last handoff. Confirm the next unit, then execute **only
 1. **Read `.cursor/HANDOFF.md` first.** No handoff → say so and suggest `/start-project`. Do not invent progress.
 2. **Pre-flight (API/usage limit):** if `- **Mode:**`, `- **Gaps:**`, or `- **Instruction for the next agent:**` still records an API/usage limit hard stop from a prior tick, **do not** mark a to-do `in_progress`, dispatch a Task, or edit product files until the operator confirms recovery (switch from Auto to a named model such as Claude Opus, Sonnet 4.6, or Composer 2.5 Fast; and/or wait for quota reset). Surface recovery via **Ask questions** when available. **Fallback after Auto→Grok:** Grok 4.5 and Auto often lack Ask questions (see `.cursor/memory/decisions/2026-07-20_ask-questions-model-availability.md`); use the numbered-list chat fallback (reply by number or label; typed answer = "Other"). Pre-flight is HANDOFF check plus operator model choice only; the kit has no remaining-quota API. Align with `context-guardian` Quota-blocked sessions and `/run-plan` Read-state pre-flight.
 3. **Audits pre-flight** (config `externalPlanReview.preflight`: `off` | `warn` | `block`; missing = `off`): before the next-unit confirmation Ask, check owed / untriaged audits for the chosen plan slug (Field Report owed, untriaged `plan-monitor-*.md`, open cadence WARNING). `off`: skip. `warn`: surface once, then continue. `block`: arm via `.cursor/scripts/plan-external-review.sh` (prefer `--autonomous` when `mode: autonomous`, else `--paste-only`) or stop until the operator defers; never steal `/git-prod`. Distinct from the advisory pre-unit monitor skim below (ADR `2026-07-27_audits-autonomous-plan-review-contract.md`).
-4. **Unprocessed dogfood preflight (advisory):** before the next-unit confirmation Ask, skim `##` or `### Unprocessed Files` in factory `dogfood/README.md` or consumer `.cursor/dogfood/README.md`. Non-empty: mention count and top titles once with standard triage labels (`ignore` / `error` / `include` / `note`). Empty or missing inbox: silent OK. Do **not** force analysis, memory WRITE, or Field Reports; do **not** block resume solely because the inbox is non-empty. Runs regardless of `externalPlanReview.preflight`. sessionStart inbox tip stays complementary (ADR `2026-08-11_dogfood-unprocessed-broad-intake-bucket.md`).
+4. **Unprocessed dogfood preflight:** before the next-unit confirmation Ask, skim `##` or `### Unprocessed Files` in factory `dogfood/README.md` or consumer `.cursor/dogfood/README.md`. Empty or missing inbox: silent OK. Non-empty: mention count and top titles once with standard triage labels (`ignore` / `error` / `include` / `note`), then **Ask questions** (one question; chat numbered-list fallback) with labels exactly:
+   - `Analyze inbox now` — start the ingest ritual (analyze → memory WRITE → triage). Notes become plans/memory after HITL, never `plan-monitor-*.md`. `/dogfood` stays file-only.
+   - `Enqueue Fix now` — route include/error notes through the `/backlog-add` contract (Broad Intake + write-confirm; no Gate B, no activate).
+   - `Not now` — continue to the next-unit Ask. Do not start analysis.
+   Never auto-analyze, never invent Field Reports, never block resume solely because the inbox is non-empty. Runs regardless of `externalPlanReview.preflight`. sessionStart inbox tip stays complementary (ADRs `2026-08-11_dogfood-unprocessed-broad-intake-bucket.md`, `2026-08-14_main-command-dogfood-audit-routing.md`).
 5. **Apply Agent Persona chrome.** Read `.cursor/context/config.json` for `agentPersona.modes.continue-plan` (fallback to legacy `workspaceSkin`, then "autopilot"). Use the corresponding persona's `chatHints` from `registry/personas/core/` for tone and confirmations.
 6. **Confirm the next unit with Ask questions** before editing. Use concrete option labels (see What to Do step 5). Do not accept a typed "yes" as the gate.
 7. **One unit per chat** (phase or one heavy to-do) unless the user explicitly ran `/run-plan`.

diff --git a/.cursor/commands/dashboard-broadcast.md b/.cursor/commands/dashboard-broadcast.md
@@ -37,7 +37,9 @@ node dashboard/start-broadcast.mjs
    export MC_PORT=$(node -e 'import("./dashboard/lib/guards.mjs").then(m => console.log(m.preferredPortForRepoRoot(process.cwd()))).catch(e => { console.error(e.message); process.exit(1) })')
    ```
 
-   If you set an explicit `PORT`, use that value instead of the derivation.
+   If you set an explicit `PORT`, use that value instead of the derivation. `MC_PORT` is the *preferred*
+   port: when it is already held the starter walks to the next per-workspace candidate, so always prefer
+   the `Bind:` line it prints over the derivation.
 
 1. **Prefer the starter** (handles token generation, `HOST=0.0.0.0`, detach, LAN URL print):
 
@@ -91,13 +93,19 @@ node dashboard/start-broadcast.mjs
   ```bash
   curl -sf "http://127.0.0.1:$MC_PORT/dashboard-data.json?token=$MISSION_CONTROL_TOKEN" | node -e 'let d="";process.stdin.on("data",c=>d+=c);process.stdin.on("end",()=>console.log(JSON.parse(d).system?.repoRoot))'
   ```
-- If a loopback `/dashboard` already holds the port, stop it before broadcast.
+- **You do not have to stop anything to broadcast.** When the preferred port is held by this workspace's
+  loopback `/dashboard`, by another workspace, or by a token-gated instance the starter cannot identify,
+  it walks to the next per-workspace candidate and prints what it skipped ("left running"). Concurrent
+  Mission Control instances are supported; nothing is killed on your behalf.
+- To **reuse** an existing broadcast for this workspace instead of starting another, export the same
+  `MISSION_CONTROL_TOKEN` it was started with. A freshly generated token cannot match a running instance,
+  so the starter treats it as someone else's and starts beside it.
 - OS firewall may block inbound LAN TCP; allow the chosen port for your local network profile if needed.
 
 ## Notes
 
-- Port: `PORT` env overrides; default is the per-workspace hash allocation (range `3333-3588`). Derive it with the snippet in step 0 or read the printed `system.port`.
-- Log default: `/tmp/mission-control-broadcast.log`
+- Port: `PORT` env overrides; default is the per-workspace hash allocation (range `3333-3588`), walking to the next candidate when one is held. Derive the preferred port with the snippet in step 0, or read the printed `Bind:` line / `system.port`. Explicit `PORT` refuses instead of walking, so a pinned port never silently moves.
+- Log default: `/tmp/mission-control-broadcast-<rootId>.log` (per workspace; `MISSION_CONTROL_LOG` overrides)
 - Loopback UX remains `/dashboard` / `npm run dashboard` / `agent-kit dashboard`
 - Detach lessons match `/dashboard` (error `2026-07-25_dashboard-server-reaped-agent-shell`)
 
@@ -106,6 +114,8 @@ node dashboard/start-broadcast.mjs
 | Symptom | Cause | Fix |
 |---|---|---|
 | Serve exits: non-loopback requires token | `HOST` set without `MISSION_CONTROL_TOKEN` | Use `dashboard:broadcast` or set a ≥16 char token |
-| Port busy / token rejected | Existing instance on the allocated port | Kill LISTEN pid for **this** workspace only (verify `repoRoot`); retry broadcast |
+| Port busy / token rejected | Existing instance on the allocated port | Nothing to do: the starter skips it and binds the next per-workspace candidate. Want that exact port? Kill the LISTEN pid **only** after verifying `repoRoot` is yours |
+| Explicit `PORT` refused | You pinned a `PORT` that another instance holds | Unset `PORT` (auto-pick a free per-workspace port), or free that port yourself if it is this workspace's |
+| A second broadcast appears each run | `MISSION_CONTROL_TOKEN` is regenerated per run, so the running one cannot be identified | Export a stable `MISSION_CONTROL_TOKEN` to reuse the existing broadcast |
 | Phone cannot connect | Firewall or wrong IP | Confirm printed LAN IPv4; allow inbound TCP |
 | Config save 403 from phone | Expected | Config writes are loopback-only |
diff --git a/.cursor/commands/dogfood.md b/.cursor/commands/dogfood.md
@@ -132,13 +132,14 @@ If the operator chooses `Keep local only`, stop. The local file is the record.
 
 ### Step 6: Respond
 
-> Dogfood filed: `dogfood/cursor_<topic>_<date>.md` (factory) or `.cursor/dogfood/cursor_<topic>_<date>.md` (consumer). Next: analyze → memory WRITE → triage, or ask for a public issue if the pattern is upstream-relevant.
+> Dogfood filed: `dogfood/cursor_<topic>_<date>.md` (factory) or `.cursor/dogfood/cursor_<topic>_<date>.md` (consumer). This command stays file-only. Analysis is offered from `/continue-plan`, `/run-plan`, `/run-plan-all`, or `/backlog-add` preflight Ask (`Analyze inbox now` / `Enqueue Fix now` / `Not now`), not from `/dogfood`. Notes become plans/memory after HITL, never `plan-monitor-*.md`.
 
 ## Related
 
 - `dogfood/README.md` — factory inbox and ingest ritual
 - `.cursor/memory/decisions/2026-07-31_dogfood-factory-consumer-lanes.md` — lane decision
 - `.cursor/memory/decisions/2026-07-31_dogfood-ingest-contract.md` — ingest contract
+- `.cursor/memory/decisions/2026-08-14_main-command-dogfood-audit-routing.md` — main-command daily path; this slash stays file-only
 - `.cursor/memory/decisions/2026-07-17_session-origin-not-product-usecase.md` — hygiene
 - Cursor product-update gaps may route here via `/cursor-update-awareness` (Ask → `/dogfood`)
 - Incoming **public** issue triage (factory-only) is `/public-issue-triage`, not this command
diff --git a/.cursor/commands/plan-external-review.md b/.cursor/commands/plan-external-review.md
@@ -15,7 +15,7 @@ Manually arm **optional plan audits** (external plan review via Claude Code or C
 - You want a second-agent check of shipped work vs the plan (gaps, residuals)
 - Auto-arm from the exhausted path was skipped (opt-in off, no `claude` on PATH, soft-fail tip) or you prefer a manual re-run
 
-**Wired path:** when `/run-plan` (orchestrated / in-session) or headless `agent-kit run-plan` stops on plan exhausted, the kit arms or suggests `.cursor/scripts/plan-external-review.sh` (see `/run-plan` "Optional external plan review"). Use this command when you need to re-arm manually. Still not a Cursor `hooks.json` `stop` entry.
+**Daily path:** `/run-plan` and `/run-plan-all` invoke this contract when audits are enabled (arm + `--wait-monitor`). Use this slash for paste, manual re-arm, or when auto-arm was skipped. Still not a Cursor `hooks.json` `stop` entry. ADR `2026-08-14_main-command-dogfood-audit-routing.md`.
 
 Do **not** use this mid-plan for in-flight to-dos; the monitor method only verdicts `completed` work.
 
@@ -34,7 +34,7 @@ If any are missing: stop. Do **not** claim a review ran. Tell the user to run `a
 1. Prefight files above exist.
 2. `.cursor/context/config.json` has `externalPlanReview.enabled: true` (see `config.example.json`), **or** use `--force` for a one-shot arm without persisting opt-in. Missing file = disabled unless `--force`.
 3. Prefer `externalPlanReview.mode: "autonomous"` for background/inspectable auto-launch. Missing `mode` keeps paste-compatible / legacy behavior.
-4. A usable reviewer: `backend: "auto"` (default for new example/docs) uses Claude when present, else Cursor Agent. Pinned `backend: "claude"` still tips + no-op when Claude is missing. `--paste-only` still prints the command without requiring a binary yet. Same-family implementer and reviewer is an honest skip (including Auto/Auto). Claude review uses `reviewerModel` (default `sonnet` so `--permission-mode auto` can run); `advisorModel` (default `opus`) runs only on escalate. An explicit Haiku pin is valid and cannot run auto.
+4. A usable reviewer: `backend: "auto"` (default for new example/docs) uses Claude when present, else Cursor Agent. Pinned `backend: "claude"` still tips + no-op when Claude is missing. Pinned `backend: "cloud"` (Cursor Cloud Agents over REST) needs `curl`, `node`, and `CURSOR_API_KEY`, and reviews the **pushed** branch — it soft-fails rather than auditing unpushed state, and is never reached by `"auto"`. `--paste-only` still prints the command without requiring a binary yet. Same-family implementer and reviewer is an honest skip (including Auto/Auto). Claude review uses `reviewerModel` (default `sonnet` so `--permission-mode auto` can run); `advisorModel` (default `opus`) runs only on escalate. An explicit Haiku pin is valid and cannot run auto.
 
 ## Manual arm
 
@@ -56,7 +56,7 @@ If any are missing: stop. Do **not** claim a review ran. Tell the user to run `a
 - **Exit 3 is timeout-only:** it means the freshness gate was not satisfied inside the budget, never that the review finished. A monitor that appears later, including one written by a different or later arm, does **not** convert a `3` into success. Leave the target Field Report **owed** and re-arm. ADR: `decisions/2026-07-30_audits-pty-progress-gate-zombie-policy.md`.
 - **Paste-only:** clipboard + printed interactive one-liner; review starts only after the operator pastes into their Cursor Terminal. After paste (Claude running), the session still waits for the monitor file then continues into triage Ask when possible.
 - **`--dry-run`:** resolves mode/plan and prints `background-cmd` / `paste-cmd` / `focus-terminal` / `reviewer-backend` / `same-model-refuse` without spawning a reviewer (useful for smoke).
-- **Reviewer cascade:** `--backend auto|claude|cursor` (or config). Claude spawn passes `--model` from `reviewerModel`. Cursor fallback cannot honor a Claude-family name; Auto/Auto is refused. Findings-only until `/plan-review-triage`. Never `/git-prod`.
+- **Reviewer cascade:** `--backend auto|claude|cursor|cloud` (or config). `cloud` is an opt-in pin (Cursor Cloud Agents over REST) that `auto` never reaches. Claude spawn passes `--model` from `reviewerModel`. Cursor fallback cannot honor a Claude-family name; Auto/Auto is refused. Findings-only until `/plan-review-triage`. Never `/git-prod`.
 
 ### A. Script (preferred)
 

diff --git a/.cursor/commands/plan-review-triage.md b/.cursor/commands/plan-review-triage.md
@@ -15,7 +15,7 @@ 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
-- **Automatic after chat arm:** `/run-plan` / `/plan-external-review` (and `/run-plan-all` queue-end) wait for the monitor then continue into this command's Ask; operators should not need to type `done` or paste triage solely to resume
+- **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`.
 - **Not for mid-plan reviews** - this command expects `completed` work only
 
 ## Usage