Pm Skills

Project Management — Domain Orchestrator & Delivery Loop

This orchestrator does two jobs. Routing: fork context, classify a PM inquiry with scripts/pm_goal_router.py, run exactly one of the 8 sub-skills, return a digest. Looping: turn a delivery goal into a bounded agentic loop — pull live Jira data via the bundled Atlassian MCP, bridge it into the domain's deterministic analytics tools, verify every step with machine-run gates, and refuse to close until everything is verified or a human waives it. The bundled .mcp.json wires the Atlassian Remote MCP (https://mcp.atlassian.com/v1/sse, OAuth handled by Claude Code).

When to invoke

Routing logic (deterministic)

Run the router — do not eyeball the table when a script can decide:

python3 scripts/pm_goal_router.py --text "<the goal>" --output json

Exit 0 → route_to names the sub-skill: load its SKILL.md and follow its workflow. Exit 2 → ask ONE clarifying question naming the listed candidates, with a recommended answer. Exit 3 → no signal: ask the user to restate the goal with the deliverable named. Never guess silently; never silently chain a second sub-skill — digest first, confirm, then chain.

The delivery loop (agentic)

For goals (not questions) — "get sprint 14 to a verified close", "produce a portfolio health report from live Jira", "make our flow metrics visible weekly" — run the loop-library contract (Observe → Choose → Act → Verify → Record → Repeat-or-stop):

1. Observe — pull fresh state: mcp__atlassian__searchJiraIssuesUsingJql (get cloudId via getAccessibleAtlassianResources first), save the result JSON, then bridge it:

   python3 scripts/jira_snapshot_bridge.py --input snapshot.json --to flow            # WIP, throughput, cycle time p50/85/95, work-item age, SLE, aging alerts
   python3 scripts/jira_snapshot_bridge.py --input snapshot.json --to sprint > s.json # scrum-master schema
   python3 ../scrum-master/scripts/velocity_analyzer.py s.json                        # velocity + volatility + forecast

   Add --forecast N for a seeded Monte Carlo "when will N items be done" answer (refuses on < 10 completed items — thin history forecasts are lies).
2. Choose — route the next task with pm_goal_router.py; one task at a time.
3. Act — execute with the routed sub-skill's own tools per its SKILL.md.
4. Verify — gate the plan and every close with:

   python3 scripts/delivery_loop_gate.py --plan plan.json --mode plan    # exit 2 = blocked
   python3 scripts/delivery_loop_gate.py --plan plan.json --mode close   # exit 4 = close refused

   Plus each sub-skill's own gates (scrum-master's ≥ 3-sprints rule, atlassian-admin's VERIFY steps). Never adjudicate your own verification.
5. Record / Repeat-or-stop — for multi-task goals, run the state through the repo-wide harness (it enforces attempt caps, iteration budgets, and evidence logging):

   python3 engineering/agent-harness/skills/agent-harness/scripts/goal_compiler.py \
     --goal "<goal>" --manifest engineering/agent-harness/skills/agent-harness/assets/harnesses/project-management.json \
     --out .agent-harness/plan.json
   python3 engineering/agent-harness/skills/agent-harness/scripts/loop_controller.py init|next|record|verify|close ...

   Terminal states: success, clean no-op, blocked, approval-required, exhausted, stagnated. An exhausted budget is an escalation — never a success report.

Hard rules (agentic delegation governance)

1. Agents are contributors, never owners (Linear model): every loop task carries a named human owner; agent-executed tasks also carry a named human reviewer. delivery_loop_gate.py enforces this (G1/G2).
2. Acceptance must be machine-checkable — a command, or a criterion with a threshold. "Looks good" is not a gate (G3).
3. Every Jira/Confluence write is auditable and reversible-first (Rovo discipline): never transitionJiraIssue to Done without verify evidence; destructive/irreversible actions (deletes, permission changes, org-wide admin) are approval-required terminal states, not loop steps.
4. Never modify a gate you are judged by — same locked-evaluator invariant as autoresearch-agent.
5. Forecasts are ranges with confidence, never dates — Monte Carlo percentiles (p50/p70/p85/p95), per Vacanti. Single-date promises are the anti-pattern.
6. Max 3 attempts per task, 12 loop iterations per goal — then escalate to the named human with the evidence log.