Orchestra router

Orchestra Router

Routes requests to the right orchestra(s). v3 doctrine: route silently — announce only delegation, in one line. This is the runtime companion to the reference doctrine at ~/.claude/rules/orchestra-system.md.

Routing algorithm (run on every task)

1. Check NEXUS first. If the message signals ideation or business planning ("I have an idea", "thinking of building", "plan a business", "what if I built", "is this a good idea", "I've been thinking about", "should I build") → invoke the nexus-strategy skill. NEXUS becomes the meta-conductor and sequences orchestras across its 7 phases. This overrides single-orchestra routing — err toward firing it.

2. Match intent to orchestra(s). Use the trigger table below. Match on meaning, not just exact words. A request can match several — that's a stack.

3. Detect stacks. Compound requests activate multiple orchestras in sequence. Examples:

   • "make a landing page" → PLANNING → DESIGN → BUILD (QA inside) → optionally SEO/GROWTH
   • "launch my app" → MARKETING → CONTENT → SEO → PAID ADS → GROWTH → ANALYTICS
   • "find clients and email them" → LEAD GEN → CONTENT (email) → (ASK before sending)

4. Announce delegation only, in one line. No ensemble blocks. Format: → <conductor/agent>: <task> (e.g. → search-agent: competitor pricing sweep). Routing that doesn't delegate stays silent — the work leads the reply.

5. Let the conductor lead. Hand the work to the orchestra's conductor agent, which sequences its First Chair + Section players. One conductor per orchestra — no free-for-all.

6. Respect allowances + gates. Honor each orchestra's ASK-before rules and quality gate (see constitution). Reserve Bench agents never auto-fire — explicit name only.

Trigger → Orchestra quick table

If the message is about…	Activate
ideation, business idea, "what if", "should I build"	NEXUS (meta) → orchestras
build, fix, refactor, implement, deploy, code review, hard bug, diagnose, flaky test, can't reproduce, intermittent bug, investigate this error, debug this	🏗️ BUILD
UI, UX, design, mockup, Figma, layout, brand visuals, component, 21st.dev, design inspiration, "give me UI ideas", build a component	🎨 DESIGN — reach for Magic / 21st.dev MCP FIRST (component inspiration + generation)
research, investigate, competitive analysis, deep dive	🔍 RESEARCH
marketing strategy, social, campaign, GTM, awareness	📣 MARKETING
copy, blog, newsletter, carousel, email sequence	✍️ CONTENT
SEO, rank, schema, AI search, GEO, AEO	📈 SEO+GEO
find leads/clients, prospect, outreach, sales pipeline	💼 LEAD GEN
PRD, product strategy, roadmap, prioritization	🧩 PRODUCT
video, reels, image gen, thumbnail, Higgsfield	🎬 VIDEO+MEDIA
analytics, metrics, funnel, cohort, A/B results, demand forecasting, trend prediction	📊 ANALYTICS
remember, knowledge graph, Obsidian, map codebase, query NotebookLM, ask my notebooks (Python), notebooklm-scripts	🧠 KNOWLEDGE
report, deck, PDF, doc, slides, save to Notion/Drive	📄 DOCUMENTS
Google/Meta ads, PPC, paid media, ad creative	💰 PAID ADS
automate, workflow, n8n, "when X do Y", recurring	⚙️ AUTOMATION
stealth browser, cloakbrowser, bypass bot detection, bypass Cloudflare, bypass DataDome, fingerprint evasion, scrape without detection, AI agent browsing undetected	⚙️ AUTOMATION
AI pipeline, multi-agent, build MCP, RAG, fine-tune, time-series forecast, TimesFM, predict future values, forecasting model	🤖 AI/ML
iOS, Swift, SwiftUI, Xcode	📱 iOS/SWIFT
plan, sprint, OKRs, user stories, pre-mortem	📋 PLANNING
conversion, churn, paywall, ASO, pricing, retention	🚀 GROWTH
invest, stock, valuation, DCF, earnings, portfolio	💵 PERSONAL FINANCE
as CEO/CFO/CTO, founder advice, board deck, M&A, scenario plan, org health, strategic decision, devil's advocate, stress-test this	👔 EXECUTIVE ADVISORY
RevOps, deal desk, pricing strategy, partnerships, RFP, customer success	💼 LEAD GEN & SALES (RevOps section)
process mapping, vendor mgmt, procurement, capacity planning	⚙️ AUTOMATION & OPS (Business Ops)
security incident, pentest, malware, forensics, DFIR, threat intel, CTI, vulnerability, CVE, exploit, ransomware, phishing investigation, red team, blue team, purple team, zero trust, OSINT, IOC, TTPs, MITRE ATT&CK, compliance audit, NIST, ISO 27001, PCI-DSS, OT/ICS/SCADA security, cloud breach, lateral movement, C2, threat hunting, detection engineering, Sigma, Volatility, Ghidra, BloodHound	🛡️ CYBERSECURITY

Full rosters (conductor, first chair, section, harness, gates) → ~/.claude/rules/orchestra-system.md.

---

Announcement format v3 (one-line delegation only)

The v2.3 mandatory-ensemble-block format is retired (2026-07-03, CLAUDE.md v3). Rules:

• Delegating to an agent/conductor → one line: → <conductor>: <task>.
• Search-ladder hit/miss → one line: 🔍 Found in <rung>: <path> / 🔍 Internal miss → external.
• Everything else (skill picks, MCP usage, tool combinations) → silent. Use the right tool; don't narrate it. The work leads the reply.

Historical v2.3 format: rules/orchestra-system.md (reference only — do not revive).

---

Internal-first search ladder (research tasks)

When 🔍 RESEARCH activates — alone, as part of a stack, or as a research sub-step inside another orchestra (e.g. ✍️ CONTENT researching positioning, 📣 MARKETING checking competitor moves) — run this ladder BEFORE spawning any external research agent (search-agent, research-manager, WebSearch, WebFetch, brightdata, nimble, firecrawl, etc.).

The ladder:

1. Obsidian brain → qmd MCP. Query with both type: 'lex' (BM25 keyword) AND type: 'vec' (semantic). Provide intent: <one-sentence task summary> for snippet quality.
2. ~/.claude/ → grep across CLAUDE.md, rules/*.md, workflows/*.md, docs/learnings/*.md, project memory/*.md. Use Grep tool, not find + read.
3. Score Archive → Skill(score-archive) with args: "search: <task description>" — curated, high-signal namespace inside claude-mem of ✅ AUDIT-verified outcomes. Returns entries with reusable_pattern field for direct application. (Phase 1.6 shipped 2026-05-26.)
4. claude-mem → search via the claude-mem MCP if connected. Captures every prior session automatically — high recall on "what did we decide about X last week."

Only if rungs 1–4 return no useful hits → escalate to external.

Announce which rung hit (always):

• On hit: 🔍 Found in Obsidian: notes/decisions/orchestra-routing.md (cite path + 1-line snippet)
• On full miss: 🔍 Internal ladder: 0 hits across 4 rungs → going external

Skip the ladder when:

• Task is "fix this stack trace" / "debug this error message" (paste = self-contained)
• the user explicitly says "just web search" / "skip the ladder"
• Topic provably post-dates the last Obsidian sync (e.g. a tool released this week)

Doctrine source: ~/.claude/CLAUDE.md § MEMORY & SEARCH — INTERNAL FIRST (non-negotiable).

---

Chain protocol (reference only)

When multiple orchestras chain (e.g. DESIGN → BUILD → AUDIT), the full v2 protocol — chain-shape selection, structured handoff payloads, quality gates, cycle detection, ASK_HUMAN pause/resume — lives in ~/.claude/rules/orchestra-system.md § Chain protocol. For routine stacks (single conductor, linear order) it runs implicitly; read the reference only when orchestrating 3+ phase transitions or debugging multi-agent handoffs. The AUDIT gate for high-stakes work (defaults NEEDS WORK, max 3 bounce-backs, then ask the user) always applies regardless.

---

Notes

• iOS axiom- skills* don't auto-fire (disable-model-invocation). When 📱 activates, invoke them by name as needed (e.g. "use axiom-audit-swiftui-architecture").
• Creative/feature work still runs superpowers:brainstorming first (HARD-GATE) before BUILD.
• "From now on when X" automation requests need a HOOK (update-config skill), not memory.
• If a request matches nothing cleanly, say so and ask — don't force-fit.

v2.1 primitives (added 2026-05-25)

• Session start? Read HANDOFF.md at project root FIRST, before any other action — that's the Handoff Standard (see ~/.claude/rules/orchestra-system.md v2.1 Orchestra Primitives §1).
• Context filling / about to compact? The PreCompact hook + Context-Compaction Protocol preserves handoff_registry, pending ASK_HUMAN, and active chain state to ~/.claude/.compact-state.md for rehydration after compaction (orchestra-system.md §2).
• Caught a failure or false-positive? Run the Calibration Loop — log → name the past mistake → embed as calibration anchor in doctrine → verify next time (orchestra-system.md §3).
• Two skills could do the job? skill-selector will pick a STACK (up to 3) if they're complementary, or a single winner if they're substitutes. Stack mode added in v2.1.