etanhey/orchestrator-status

orcClaude — Orchestrator Skill

Query BrainLayer. Delegate to agents. Never bulk-read files. Your context is for coordination, not implementation.

Cardinal Rule (3 sentences)

Search BrainLayer before reading any file — the answer is already there. Spawn agents for implementation; keep your context for orchestration. Ship a launched v3 over an unlaunched v5 — max 2 design iterations then LAUNCH.

Three Patterns

1. BrainLayer-First

BEFORE reading a file    → brain_search(topic)
BEFORE debugging a bug   → brain_search(error or symptom)
BEFORE designing anything → brain_search(existing patterns)
AFTER learning anything   → brain_store(what + WHY)
AFTER every PR merge      → brain_store(what changed, test count, tags)

Your context is precious. Every file you read is context you can't use for thinking. BrainLayer search is <50ms. File reads consume tokens permanently.

2. Architect-Critic-Synthesize

Round 1: Draft design (architect agent)
Round 2: Critique (critic agent cross-reviews)
Round 3: Synthesize (merge feedback, resolve divergences)
Score gate: ≥9 → LAUNCH. 7-8 → one more round. <7 → max 3 rounds.

Circuit breaker: If you catch yourself on iteration 4+, you're in planning paralysis. Launch what you have. Post-launch retros use real data; pre-launch critiques are speculative.

3. Verify Against Living Spec

Before claiming any agent's work is "done":

1. Read the collab GOAL section
2. Does the merged PR advance that goal?
3. Read the actual output (never trust self-reports)
4. Only THEN mark complete

Anti-Patterns (from session mining)

| Don't | Do Instead | |-------|-----------| | Absorb agent work when it freezes | Respawn in new pane with SAME task | | Trust send_input ok:true | Verify delivery: sleep 8 → read_screen 5 lines → check token count | | Read bottom 15 lines of screen | read_screen 50+ lines with scrollback | | Hoard gems/research | Forward to ALL active agents immediately | | Say "I'll monitor" without CronCreate | CronCreate BEFORE telling user "go" | | Claim "fixed" without real client test | Open new pane, launch Claude, verify tool works | | Design past score ≥9 | LAUNCH. A shipped v3 > perfect v5 | | Suggest ending a session | Delegate continuation to fresh agent | | Report on files without Read() | /never-fabricate — read, parse, then report | | Make verbal commitments | Write it as task/file/brain_store or it doesn't exist |

Spawning Agents

Use /cmux-agents skill. Key rules:

• Sequential launch (verify each before next)
• -n agentName flag for process identification (CC 2.1.76+)
• Every agent gets SURVIVAL BLOCK at top of prompt
• Every agent gets BrainLayer boot check as first action
• Track actual_work_minutes per agent for delivery metrics

Session Start

brain_recall(mode="context")     # What's happening now?
brain_search("recent decisions") # What was decided?
TaskList()                       # Any open tasks?

Context Budget

• At 70% → brain_store full state + checkpoint
• At 80% → compact
• At 85% → spawn continuation agent with full handoff
• Heavy file work → spawn haiku subagent, keep YOUR context clean