Imported from fadhly-permata/AI-Gate (.opencode/skills/pm-orchestration/SKILL.md). Install upstream with npx skills add fadhly-permata/AI-Gate --skill pm-orchestration. Copyright stays with the author.

PM Orchestration Playbook

1. Memory Bank (persist context across agents)

Keep in documents/pm/:

• memory-bank.md — brief, decisions, progress, risks.
• OPERATING_RULES.md — append R# rules when the user corrects the PM.
• state.md — mode / delay_seconds / checkpoint / updated / rules_ref.
• status.md — spawned sub-agents + outcomes. Sub-agents READ documents/pm/ (handover source) but never WRITE it. Only PM writes.

2. Handover Protocol (give sub-agents full context)

Every spawn prompt must include:

• Goal — what success looks like.
• Context — links/summary from documents/pm/ (decisions, prior outputs).
• Scope — exact write roots + read roots (from the roster).
• Definition-of-done — concrete acceptance criteria.
• Constraints — deadlines, tech limits, forbidden actions. The sub-agent replies with a receipt (files changed, decisions, open questions). PM integrates receipts; never duplicates the sub-agent's work.

3. Delegation matrix

| Design / trade-offs | tech-architect | documents/architecture/** | | Public copy / README / wiki / docs-site / release notes | public-writer | documents/pm/wiki-drafts/, README.md, documents/readme-variants/ |

Spawn only when the need appears. Never pre-create. Generate agent + skill together. Reuse (do not delete) once generated.

4. Parallel vs Sequential decision

When a task is large / long / multi-part, STOP and ask the user (Indonesian casual): run paralel (faster, independent pieces, needs merge review) or sekuensial (ordered, each step depends on the previous). See .opencode/rules/parallel-sequential.md. Use parallel only when pieces are truly independent and scopes don't overlap; otherwise sequential.

5. File-boundary enforcement

Each sub-agent writes only its scope. PM owns documents/pm/ and the final merge. Cross-scope writes are violations — reject the receipt and ask for fix. See .opencode/rules/agent-boundaries.md.

6. Record Protocol (user correction → durable rule)

Trigger: user corrects a PM decision, OR PM acted outside a granted file scope, OR a rule was violated.

1. Append block to documents/pm/status.md (newest first):

## <yyyymmdd-HHMM> — PM decided wrong → user correction recorded (ProjectManager)

### Violation
- Rule broken: <rule id + path:line>
- What PM did: <one factual line>

### Correction
- Durable rule captured: <R#> — <title>
- Decision: <user ruling + rationale>

### Prevention
- Mechanism: <guardrail / gate added>
- Verification: <gate or check that catches a recurrence>

2. Append the R# to documents/pm/OPERATING_RULES.md (heading + lesson + rule).
3. Update documents/pm/memory-bank.md Decisions/Progress.
4. Update documents/pm/state.md: mode, delay_seconds, checkpoint, updated, rules_ref.
5. Report to user: violation + correction + prevention.
6. Never delete a recorded violation. Never edit a past block. No extra file per event — append to status.md (R38).