/cs-pm-loop

disable-model-invocation: true
description: Drive a project-delivery goal through a bounded agentic loop — Jira MCP snapshot → flow/sprint analytics bridge → routed sub-skill execution → machine-verified gates → close refused until everything is verified or human-waived. The PM-domain adapter over engineering/agent-harness.
argument-hint: "<delivery goal, e.g. 'get sprint 14 to a verified close with health >= 70'>"

/cs-pm-loop — run a delivery goal to a verified close

Goal:

**$ARGUMENTS**

Sequence (gates are blocking — never skip forward)

1. **Intake gate** — the goal must name an observable outcome and its proof. If vague, run the `/cs-grill-pm` branches first (one question per turn). Do not loop on fuzz. 2. **Observe** — pull fresh data: `mcp__atlassian__getAccessibleAtlassianResources` (get cloudId) → `mcp__atlassian__searchJiraIssuesUsingJql` → save `snapshot.json`, then:

   python3 ${CLAUDE_PLUGIN_ROOT}/skills/pm-skills/scripts/jira_snapshot_bridge.py --input snapshot.json --to flow
   python3 ${CLAUDE_PLUGIN_ROOT}/skills/pm-skills/scripts/jira_snapshot_bridge.py --input snapshot.json --to sprint > sprint_data.json

3. **Plan** — write the task plan (owners, executors, reviewers, machine-checkable acceptance per task; shape via `delivery_loop_gate.py --sample`), then gate it:

   python3 ${CLAUDE_PLUGIN_ROOT}/skills/pm-skills/scripts/delivery_loop_gate.py --plan plan.json --mode plan

Exit 2 → fix the listed G1–G4 violations before executing. For multi-task goals, compile through the repo harness instead (`goal_compiler.py` with the `project-management.json` manifest) and drive it with `loop_controller.py`. 4. **Execute** — one task at a time: route with `pm_goal_router.py`, run the routed sub-skill's own tools, record real exit codes and evidence. Retry means a changed approach; max 3 attempts per task. 5. **Verify** — the task's acceptance command must exit 0; sub-skill gates apply (scrum-master's ≥3-sprints rule, atlassian-admin's VERIFY steps). Never adjudicate your own verification; never edit a gate to make it pass. 6. **Close** —

   python3 ${CLAUDE_PLUGIN_ROOT}/skills/pm-skills/scripts/delivery_loop_gate.py --plan plan.json --mode close

Exit 4 → close refused: finish, escalate, or get a human waiver (with reason). Exit 0 → report the handoff: tasks, statuses, evidence, waivers, and the flow-metrics before/after.

Rules

• Terminal states: success · clean no-op · blocked · approval-required · exhausted ·

stagnated. Exhausted budgets escalate to the named human — never reported as success.

• Jira writes are auditable: no `transitionJiraIssue` to Done without verify evidence;

admin/destructive actions are approval-required, full stop.

• Max 12 loop iterations per goal; 3 attempts per task.