pm

---

description: Invoke the Product Manager to marshal and orchestrate work across all councils

Session-scoped PM lock

The PM command is session-scoped. PM mode state is stored in .claude/sessions/<SESSION_ID>.pm-mode (contains "on" or "off").

Special subcommands — handle these directly without spawning an agent:

If $ARGUMENTS is "on":

• The SESSION_ID is injected into the system prompt as SESSION_ID = <value>. Read it from there; if not present, use "default".
• Run mkdir -p .claude/sessions && echo "on" > ".claude/sessions/<SESSION_ID>.pm-mode" via Bash.
• Confirm: "PM mode enabled. Every prompt will now be routed through the PM."
• Stop here.

If $ARGUMENTS is "off":

• The SESSION_ID is injected into the system prompt as SESSION_ID = <value>. Read it from there; if not present, use "default".
• Run rm -f ".claude/sessions/<SESSION_ID>.pm-mode" && echo "off" via Bash.
• Confirm: "PM mode disabled."
• Stop here.

If $ARGUMENTS is "status" or "list":

• The SESSION_ID is injected into the system prompt as SESSION_ID = <value>. Read it from there; if not present, use "default".
• Run ls .claude/sessions/*.pm-mode 2>/dev/null || echo "(none)" via Bash to find all sessions with PM mode enabled.
• For each .pm-mode file found, print the session ID (filename without .pm-mode) and whether mode is on/off.
• Print current session PM mode status: ON or OFF.
• Stop here.

---

Otherwise — invoke the PM agent

Use the Agent tool to spawn the product-manager subagent with the following task:

Task: $ARGUMENTS

Instructions for the PM agent:

You are being invoked to lead and orchestrate the following work: $ARGUMENTS

Execute the full automated workflow:

1. Assess the request — Is this intake (new project), a sprint (ongoing work), or a targeted delegation (specific epic/feature)?

2. Load context — Read the active client lock at .claude/sessions/<SESSION_ID>.client if present (contents: <client> or <client>/<project>). Use it to locate the correct workspace under clients/<client>/projects/<project>/. If no lock is set, ask once and run /use-client before proceeding — without a lock the client boundary guard is inactive.

3. Produce a delegation plan — Break the work into epics, assign each to the correct council and agent, map dependencies, and identify the critical path.

4. Execute immediately — Spawn the required agents in parallel where dependencies allow. Do not wait for manual prompting between stages. Collect each agent's output, verify the gate is satisfied, and advance.

5. Surface blockers — If a gate is blocked, say so explicitly with the blocker, impact, and your recommended resolution. Do not stall silently.

6. Write a summary — When all stages complete, write a PM summary to: clients/{{CLIENT_SLUG}}/projects/{{PROJECT_SLUG}}/specs/pm-summary-{{topic}}-{{date}}.md

Proceed now without asking for clarification unless a critical piece of information is entirely absent.