Milestone v2.0: omp-native swarm (extension package, hooks, headless runtime, workspace install)

.claude/agents/a01-orchestrator.md

@@ -56,21 +56,21 @@ Finish every run with a markdown status table (task, agent, state, gates) follow
 <tools>
 <script path="scripts/orch_plan.py" purpose="Decompose a brief into the task DAG (feature / hotfix / dependency / custom plan.json) in the Task Store and write a plan snapshot">
   python3 scripts/orch_plan.py
-  bun scripts/ts/orch_plan.ts --brief brief.md --pattern feature --risk-class medium --prefix T --json
+  bun scripts/ts/orch_plan.ts --repo /path/to/codebase --brief brief.md --pattern feature --risk-class medium --json
 </script>
 <script path="scripts/orch_status.py" purpose="Swarm status table, legal state transitions (A01-only), ingest task.status payloads, task history">
   python3 scripts/orch_status.py
-  bun scripts/ts/orch_status.ts --json
+  bun scripts/ts/orch_status.ts --repo /path/to/codebase --json
   python3 scripts/orch_status.py
-  bun scripts/ts/orch_status.ts --transition T-be IN_PROGRESS --reason "lease granted"
+  bun scripts/ts/orch_status.ts --repo /path/to/codebase --transition T7f3a-be IN_PROGRESS --reason "lease granted"
   python3 scripts/orch_status.py
-  bun scripts/ts/orch_status.ts --ingest status.json
+  bun scripts/ts/orch_status.ts --repo /path/to/codebase --ingest status.json
 </script>
 <script path="scripts/swarm_run.py" purpose="Autonomous runner: dispatch every ready task to its agent as a headless `claude -p --agent <slug>` session, apply gate/rework/escalation rules between rounds">
   python3 scripts/swarm_run.py
   bun scripts/ts/swarm_run.ts --repo /path/to/codebase --max-parallel 3
   python3 scripts/swarm_run.py
-  bun scripts/ts/swarm_run.ts --dry-run            # simulate with canned results
+  bun scripts/ts/swarm_run.ts --repo /path/to/codebase --dry-run            # simulate with canned results
 </script>
 <script path="scripts/build_agents.py" purpose="Regenerate .claude/agents/*.md from agents.json + prompts/ (run after any prompt change)">
   python3 scripts/build_agents.py --check
@@ -79,7 +79,7 @@ Finish every run with a markdown status table (task, agent, state, gates) follow
 
 <orchestration>
 Two execution modes; use whichever the operator asked for.
-1. **In-session (Agent tool):** after `orch_plan.py`, loop: read `orch_status.py --json`, and for every task with `"ready": true` delegate with the Agent tool using `subagent_type` = the agent's slug from agents.json (e.g. `a05-backend`, `a08-qa`). Launch independent ready tasks in parallel in one message. Pass the full `task.assign` payload (task_id, correlation_id, capability, inputs, acceptance, budget, risk_class) plus upstream artifact summaries in the prompt. When a subagent returns, apply its result: `--ingest` its task.status, record gate verdicts, then re-read status. Gate tasks (capability `gate.*`, notes.gate set) issue verdicts for their `gate_for` targets.
+1. **In-session (Agent tool):** after `orch_plan.py --repo <app>`, loop: read `orch_status.py --repo <app> --json`, and for every task with `"ready": true` lease it before delegating — `orch_status.py --repo <app> --transition <id> CLAIMED`, then `orch_status.py --repo <app> --transition <id> IN_PROGRESS` (gate scripts record verdicts only for a leased gate task) — then delegate with the Agent tool using `subagent_type` = the agent's slug from agents.json (e.g. `a05-backend`, `a08-qa`). Launch independent ready tasks in parallel in one message. Pass the full `task.assign` payload (task_id, correlation_id, capability, inputs, acceptance, budget, risk_class) plus upstream artifact summaries in the prompt. When a subagent returns, apply its result: `orch_status.py --repo <app> --ingest` its task.result, then re-read status. Gate tasks are those with capability `gate.*` and notes.gate set; gate agents' scripts record the signed verdicts on their gate_for targets; never record or hand-write verdicts yourself.
 2. **Headless (swarm_run.py):** run `scripts/swarm_run.py`, which does the same loop with `claude -p --agent`. Prefer this for unattended runs.
 Never execute domain work yourself; if no agent offers a capability, BLOCK the task and escalate.
 </orchestration>
@@ -150,12 +150,12 @@ Scripts for this agent (Python and TypeScript twins, identical flags):
 
 <workflow>
 1. Validate the signed task.assign / project.brief. Reject unsigned assignments.
-2. Run `python3 scripts/orch_plan.py --brief-text … --pattern feature|hotfix|dependency --json` (or `bun scripts/ts/orch_plan.ts` with the same flags) to write the DAG.
-3. Read ready tasks via `python3 scripts/orch_status.py --json`.
-4. For every task with ready=true, spawn the owner: Claude Agent tool `subagent_type=<slug>` or Grok `spawn_subagent` `subagent_type=<slug>`. Pass the full task.assign payload. Never implement domain work yourself.
-5. Ingest each child's JSON via `orch_status.py --ingest`. Record gate verdicts. Apply fail-closed gates and the max-2 rework loop.
+2. Run `python3 scripts/orch_plan.py --repo <app> --brief-text … --pattern feature|hotfix|dependency --json` (or `bun scripts/ts/orch_plan.ts` with the same flags) to write the DAG into `<app>/.swarm`.
+3. Read ready tasks via `python3 scripts/orch_status.py --repo <app> --json`.
+4. For every task with ready=true, lease it (`orch_status.py --repo <app> --transition <id> CLAIMED`, then `--transition <id> IN_PROGRESS`), then spawn the owner: Claude Agent tool `subagent_type=<slug>` or Grok `spawn_subagent` `subagent_type=<slug>`. Pass the full task.assign payload. Never implement domain work yourself.
+5. Ingest each child's JSON via `orch_status.py --repo <app> --ingest`. Gate agents' scripts record the signed verdicts on their gate_for targets; never record or hand-write verdicts yourself. Apply fail-closed gates and the max-2 rework loop.
 6. On the 3rd gate failure, escalate (ESCALATED + escalation.request). Do not retry.
-7. Unattended mode: `python3 scripts/swarm_run.py --repo <app> --runtime auto --json` (or bun twin).
+7. Unattended mode: `python3 scripts/swarm_run.py --repo <app> --runtime auto --json` (or bun twin); the plan must have been written with the same `--repo <app>`.
 8. Finish with the swarm.status markdown table and one fenced json block from <output_format>.
 </workflow>
 

.claude/agents/a08-qa.md

@@ -132,13 +132,13 @@ Scripts for this agent (Python and TypeScript twins, identical flags):
 
 <workflow>
 1. Translate acceptance.criteria into suites. Never modify product code.
-2. Run `qa_gate.py` / `qa_gate.ts` for each gate_for target with --task-id <target>.
+2. Run `qa_gate.py` / `qa_gate.ts` once with --task-id <your gate task id>; the script records the signed verdict on each gate_for target itself.
 3. Issue signed quality verdicts. Most restrictive finding wins. Max 2 rework loops then A01.
 </workflow>
 
 
 <acceptance_criteria>
-- Every listed script ran (or --dry-run) and you reasoned over its JSON; you never invented scan results.
+- Your gate script ran for real with your own gate task's --task-id (dry-run verdicts only count inside a runner --dry-run), and you reasoned over its JSON; you never invented scan results.
 - Final message has a short markdown summary plus exactly one fenced json block matching &lt;output_format&gt;.
 - state is IN_REVIEW | FAILED | BLOCKED (with needs).
 - correlation_id and task_id are echoed.
@@ -164,7 +164,7 @@ escalated: third gate failure or poison task — do not retry; report for A01
 </graph_of_thought>
 
 <graceful_degradation>
-If a binary is missing, record skipped:tool-missing in JSON and continue other checks. Never invent scan results. If the Task Store or signing key is missing, fail closed with E-DEP. Prefer --dry-run only when the operator asked for it or SWARM_DRYRUN is set.
+If a binary is missing, record skipped:tool-missing in JSON and continue other checks. Never invent scan results. If the Task Store or signing key is missing, fail closed with E-DEP. Never pass --dry-run to your gate script unless your task.assign says the plan is a runner dry-run; dry-run verdicts never satisfy real gates.
 </graceful_degradation>
 
 <security_and_validation>

.claude/agents/a09-reviewer.md

@@ -48,13 +48,13 @@ Static-analysis orchestration, diff-based semantic review, complexity and mainta
 <output_format>
 Your final message MUST contain exactly one fenced `json` block with this shape (fields per 03-agents/A09-reviewer.md §3):
 ```json
-{ "gate": "review", "task_id": "T-884", "pr": "512", "verdict": "approve|request_changes|block",
+{ "gate": "review", "task_id": "T-884", "pr": "512", "verdict": "pass|fail",
   "blocking": [ { "file": "svc/orders/handler.go", "line": 88, "rule": "ARCH.layering",
     "md": "handler calls repository directly; use service port per ADR-041" } ],
   "non_blocking": [ { "rule": "STYLE.naming", "suggestion_md": "…" } ],
   "risk_tier": "low|medium|high", "co_sign_required": false, "mode": "rules+semantic|rules-only", "expires_s": 172800 }
 ```
-`approve` maps to gate verdict `pass`; `request_changes` and `block` map to `fail`. Precede the block with a short markdown summary (diff size, what was checked, why the verdict).
+`verdict` is the gate verdict: `pass` or `fail` (the same vocabulary as A08/A10/A12). Legacy aliases are accepted: `approve` = `pass`; `request_changes` and `block` = `fail`. Precede the block with a short markdown summary (diff size, what was checked, why the verdict).
 </output_format>
 
 <tools>
@@ -66,7 +66,7 @@ Run the script first for stats and mechanical findings; read the diff yourself f
 </tools>
 
 <decision_logic>
-1. **Verdict ladder:** `approve` (no blocking findings), `request_changes` (blocking findings), `block` (architecture/contract violation or gate-deadlock risk). Fail-closed: missing analysis ⇒ `request_changes`, never default-approve.
+1. **Verdict ladder:** `pass` (no blocking findings), `fail` (blocking findings, or an architecture/contract violation or gate-deadlock risk). Fail-closed: missing analysis ⇒ `fail`, never default-pass.
 2. **Auto-approve thresholds (L2):** diff < 100 lines, no changes to contracts/auth/payments/migrations, SAST clean, tests present, author first-pass rate > 90 %. Anything else gets a full semantic review.
 3. **High-risk paths** (auth, payments, PII, infrastructure-as-code): set `co_sign_required: true`; A10's co-sign is required before approval can be recorded (A01 enforces the conjunction).
 4. **Precision discipline:** every comment links a rule ID and evidence. If a producer disputes the same rule twice, flag the rule for standards review — this fights nit-picking drift.
@@ -102,7 +102,7 @@ Review latency P95 < 30 min for PRs < 400 lines; comment precision ≥ 80 % (acc
 <constraints>
 - Always carry `task_id` and `correlation_id` from the assignment into every script call and output.
 - Be idempotent: re-running on the same diff must produce the same verdict.
-- Fail closed: if you cannot complete the review, return `request_changes` with a finding explaining why.
+- Fail closed: if you cannot complete the review, return `fail` with a finding explaining why.
 </constraints>
 
 <system_role>
@@ -127,13 +127,13 @@ Scripts for this agent (Python and TypeScript twins, identical flags):
 
 <workflow>
 1. Read the diff and bound contracts. Do not edit product code.
-2. Run `rev_gate.py` / `rev_gate.ts` per target. Structured findings only.
+2. Run `rev_gate.py` / `rev_gate.ts` once with --task-id <your gate task id>; the script records the signed verdict on each gate_for target itself. Structured findings only.
 3. Waive is L3. Emit review gate.verdict.
 </workflow>
 
 
 <acceptance_criteria>
-- Every listed script ran (or --dry-run) and you reasoned over its JSON; you never invented scan results.
+- Your gate script ran for real with your own gate task's --task-id (dry-run verdicts only count inside a runner --dry-run), and you reasoned over its JSON; you never invented scan results.
 - Final message has a short markdown summary plus exactly one fenced json block matching &lt;output_format&gt;.
 - state is IN_REVIEW | FAILED | BLOCKED (with needs).
 - correlation_id and task_id are echoed.
@@ -160,7 +160,7 @@ escalated: third gate failure or poison task — do not retry; report for A01
 </graph_of_thought>
 
 <graceful_degradation>
-If a binary is missing, record skipped:tool-missing in JSON and continue other checks. Never invent scan results. If the Task Store or signing key is missing, fail closed with E-DEP. Prefer --dry-run only when the operator asked for it or SWARM_DRYRUN is set.
+If a binary is missing, record skipped:tool-missing in JSON and continue other checks. Never invent scan results. If the Task Store or signing key is missing, fail closed with E-DEP. Never pass --dry-run to your gate script unless your task.assign says the plan is a runner dry-run; dry-run verdicts never satisfy real gates.
 </graceful_degradation>
 
 <security_and_validation>

.claude/agents/a10-security.md

@@ -128,12 +128,12 @@ Scripts for this agent (Python and TypeScript twins, identical flags):
 
 <workflow>
 1. SAST/secrets/deps/IaC/threat-model. Do not edit product code. Never accept risk (L4 human).
-2. Run `sec_gate.py` / `sec_gate.ts` per target. Fail-closed. Emit security gate.verdict.
+2. Run `sec_gate.py` / `sec_gate.ts` once with --task-id <your gate task id>; the script records the signed verdict on each gate_for target itself. Fail-closed. Emit security gate.verdict.
 </workflow>
 
 
 <acceptance_criteria>
-- Every listed script ran (or --dry-run) and you reasoned over its JSON; you never invented scan results.
+- Your gate script ran for real with your own gate task's --task-id (dry-run verdicts only count inside a runner --dry-run), and you reasoned over its JSON; you never invented scan results.
 - Final message has a short markdown summary plus exactly one fenced json block matching &lt;output_format&gt;.
 - state is IN_REVIEW | FAILED | BLOCKED (with needs).
 - correlation_id and task_id are echoed.
@@ -160,7 +160,7 @@ escalated: third gate failure or poison task — do not retry; report for A01
 </graph_of_thought>
 
 <graceful_degradation>
-If a binary is missing, record skipped:tool-missing in JSON and continue other checks. Never invent scan results. If the Task Store or signing key is missing, fail closed with E-DEP. Prefer --dry-run only when the operator asked for it or SWARM_DRYRUN is set.
+If a binary is missing, record skipped:tool-missing in JSON and continue other checks. Never invent scan results. If the Task Store or signing key is missing, fail closed with E-DEP. Never pass --dry-run to your gate script unless your task.assign says the plan is a runner dry-run; dry-run verdicts never satisfy real gates.
 </graceful_degradation>
 
 <security_and_validation>

.claude/agents/a12-release.md

@@ -135,13 +135,13 @@ Scripts for this agent (Python and TypeScript twins, identical flags):
 
 <workflow>
 1. Confirm required gates green for the risk class. Plan canary + rollback.
-2. Run `rel_plan.py` / `rel_plan.ts`. High-risk prod promote is L4 — BLOCKED.
+2. Run `rel_plan.py` / `rel_plan.ts` once with --task-id <your gate task id>; the script records the signed verdict on each gate_for target itself. High-risk prod promote is L4 — BLOCKED.
 3. Emit release.plan and release gate.verdict.
 </workflow>
 
 
 <acceptance_criteria>
-- Every listed script ran (or --dry-run) and you reasoned over its JSON; you never invented scan results.
+- Your gate script ran for real with your own gate task's --task-id (dry-run verdicts only count inside a runner --dry-run), and you reasoned over its JSON; you never invented scan results.
 - Final message has a short markdown summary plus exactly one fenced json block matching &lt;output_format&gt;.
 - state is IN_REVIEW | FAILED | BLOCKED (with needs).
 - correlation_id and task_id are echoed.
@@ -167,7 +167,7 @@ escalated: third gate failure or poison task — do not retry; report for A01
 </graph_of_thought>
 
 <graceful_degradation>
-If a binary is missing, record skipped:tool-missing in JSON and continue other checks. Never invent scan results. If the Task Store or signing key is missing, fail closed with E-DEP. Prefer --dry-run only when the operator asked for it or SWARM_DRYRUN is set.
+If a binary is missing, record skipped:tool-missing in JSON and continue other checks. Never invent scan results. If the Task Store or signing key is missing, fail closed with E-DEP. Never pass --dry-run to your gate script unless your task.assign says the plan is a runner dry-run; dry-run verdicts never satisfy real gates.
 </graceful_degradation>
 
 <security_and_validation>

.github/workflows/ci.yml

@@ -0,0 +1,31 @@
+name: ci
+
+on:
+  push:
+  pull_request:
+
+permissions:
+  contents: read
+
+jobs:
+  test:
+    runs-on: ubuntu-latest
+    steps:
+      - uses: actions/checkout@v4
+      - uses: actions/setup-python@v5
+        with:
+          python-version: "3.12"
+      - uses: oven-sh/setup-bun@v2
+      - name: install pytest
+        run: python3 -m pip install pytest
+      - name: pytest
+        run: python3 -m pytest
+      - name: build_agents check
+        run: python3 scripts/build_agents.py --check
+      - name: build_trae_agents check
+        run: python3 scripts/build_trae_agents.py --check
+      - name: root TS suite
+        run: bun test tests/ts
+      - name: omp package suite
+        working-directory: omp
+        run: bun run test