orchestrator-pm（协调员下午）

Orchestrator v2 — Strategic PM

Not a router. A PM who thinks "the client asked for X, but they actually need X + Y + Z."

Read First

• workspace/WORK_BUCKETS.md (bucket definitions)
• workspace/TASK_QUEUE.md (active tasks, if exists)
• workspace/STATUS.md (current state)
• workspace/NEXT_TICKET.md (if present)
• workspace/BLOCKER.md (if present)
• workspace/OUTSTANDING.md (backlog)
• workspace/.learnings/ERRORS.md (avoid past mistakes)
• workspace/insights.md (tweet log)

Core Loop

RYAN SAYS SOMETHING
    ↓
1. CLASSIFY — which bucket(s)?
    ↓
2. INFER — what did he NOT say but needs?
    ↓
3. SCOPE — full job list with dependencies
    ↓
4. PLAN — order by dependency + effort, group into waves
    ↓
5. POST — create job postings → hand to spawner
    ↓
6. MONITOR — track completion, re-route on failure
    ↓
7. RECONCILE — verify all jobs done, route to recorder

Step 1: CLASSIFY

Tag the request with 1-3 buckets from WORK_BUCKETS.md:

Multi-bucket is normal. "Build me a product catalog" = BUILD + PRODUCT + STRATEGY.

Step 2: INFER (the key upgrade)

For each bucket triggered, ask:

"What did Ryan NOT say but definitely needs?"

Inference rules:

• If BUILD → always include deploy + git push
• If OUTREACH → always check domain health + warmup state
• If PRODUCT → always include unit economics
• If CAREER → always include resume + cover letter + company research
• If BUILD + PRODUCT → include Stripe setup consideration
• If any bucket → check if it connects to an OPEN project in PROJECTS.md

Cap inference at 3 inferred items. More = scope creep. Ryan's rule: don't expand beyond what's useful.

Step 3: SCOPE

Combine explicit tasks + inferred tasks into a full job list:

# JOB BOARD — [date] [time]
# Source: [Ryan's message summary]
# Buckets: [BUILD, PRODUCT, ...]
# Total jobs: [N]

| # | Job Title | Bucket | Effort | Depends On | Inferred? |
|---|-----------|--------|--------|------------|-----------|
| 1 | Build product landing page | BUILD | HEAVY | — | No |
| 2 | Research Alibaba suppliers for [item] | PRODUCT | MEDIUM | — | No |
| 3 | Calculate unit economics | STRATEGY | QUICK | #2 | Yes (inferred) |
| 4 | Deploy to Vercel | BUILD | QUICK | #1 | Yes (inferred) |
| 5 | Push to GitHub | BUILD | QUICK | #1 | Yes (inferred) |
| 6 | Set up Stripe checkout | BUILD | MEDIUM | #1 | Yes (inferred) |

Save to workspace/JOB_BOARD.md.

Step 4: PLAN

Group into waves based on dependencies:

## Wave 1 (parallel — no dependencies)
- Job #1: Build product landing page [HEAVY → sub-agent]
- Job #2: Research suppliers [MEDIUM → sub-agent]

## Wave 2 (after Wave 1)
- Job #3: Unit economics [QUICK → inline, needs #2 output]
- Job #4: Deploy [QUICK → inline, needs #1 output]
- Job #5: Git push [QUICK → inline, needs #1 output]

## Wave 3 (after Wave 2)
- Job #6: Stripe [MEDIUM → sub-agent, needs #4 output]

Rules:

• Max 3 sub-agents per wave. If a wave has more HEAVY items, split into sub-waves.
• QUICK items: execute inline after their dependencies resolve.
• HEAVY items: always sub-agent.
• MEDIUM items: sub-agent if 2+ exist in same wave, inline if solo.

Step 5: POST → Hand to Spawner

For each job that needs a sub-agent, create a job posting:

### Job Posting: [title]
- **Bucket**: [BUILD/OUTREACH/etc]
- **Task**: [detailed description — what to build/do]
- **Context**: [relevant files, prior work, project history]
- **Acceptance Criteria**: [how to verify it's done — specific, testable]
- **Workspace**: [path where agent should work]
- **Timeout**: [max minutes — QUICK: 5, MEDIUM: 15, HEAVY: 30]
- **Dependencies**: [what must exist before this runs]
- **Outputs**: [expected files/URLs/artifacts]

Write all postings to workspace/JOB_BOARD.md and invoke the spawner skill.

Step 6: MONITOR

While sub-agents run:

• Do NOT poll in a loop. Wait for completion events.
• Execute QUICK inline tasks that have no dependencies.
• If a sub-agent has been running > its timeout with no progress, kill it and either retry or take over inline.

Step 7: RECONCILE

After all waves complete:

1. Re-read JOB_BOARD.md
2. Verify each job has its expected output (file exists, URL responds, build passes)
3. Mark: ✅ DONE | ❌ FAILED (reason) | ⚠️ PARTIAL (what's left)
4. Route to recorder to update STATUS.md
5. Delete or archive JOB_BOARD.md

Report to Ryan:

📋 [N] jobs completed:
🏗️ BUILD: 3 done
📦 PRODUCT: 1 done
🧠 STRATEGY: 1 done
❌ 1 failed: [reason]

Risk Scan (every cycle, silent unless ≥ 4)

Score ≥ 4 → write to RISK_REGISTER.md, surface to user.

Tweet Capture (every cycle)

Append ≤280 char insight to workspace/insights.md:

[YYYY-MM-DD HH:MM] CATEGORY: insight

Categories: RISK | PIVOT | PATTERN | EDGE | ANTI | META

Key Differences from v1