@takuma-hirai/hirai-method 0.1.0

package/.claude/skills/agent-router/README.md

@@ -0,0 +1,137 @@
+# agent-router
+
+Skill that recommends a specialist subagent for ambiguous prompts that would otherwise default to `general-purpose`. The skill establishes a habit of naming the right specialist at dispatch time, instead of falling back to general-purpose by default.
+
+> **Note on justification**: An earlier draft cited "66% (1,178/1,749) of historical subagent invocations went to general-purpose while named specialists sat dormant" as the activation rationale. That figure was measured before the 100 named agents existed in their current form — they were imported on 2026-05-04 in a single commit, so the 90-day pre-import window contains structurally unreliable dormant percentages (see `docs/INVENTORY-stocktake-2026-05-04.md` §8). The router is justified primarily by **habit formation** at dispatch time, not by the dormant figure. The empirical 84.1% dispatch rate against the 1,178 historical `general-purpose` prompts remains a meaningful coverage indicator and will be re-measured against post-import transcripts in 30 / 90 days.
+
+## Files
+
+| Path | Purpose |
+|---|---|
+| `SKILL.md` | Skill manifest (name, description, when-to-use) |
+| `router.py` | Dispatch implementation (stdlib only) |
+| `dispatch-table.yml` | Single source of truth: keyword → agent map |
+| `tests/test_router.py` | 40 unit + integration tests (19 keyword + 21 Phase 2) |
+| `samples/representative_prompts.txt` | Verification fixtures sampled from real history |
+| `README.md` | This file (design notes & extension guide) |
+
+## Design
+
+### Why scoring rather than ML?
+
+- Stdlib only, zero dependencies — runs anywhere Python 3 runs.
+- Auditable: every routing decision is reproducible from the dispatch table.
+- Easy to extend: add a keyword line, no retraining.
+- Threshold-controlled fallback prevents false-positive specialist routing.
+
+### Algorithm (Phase 1: keyword)
+
+1. **Tokenize** the prompt: lowercase, alphanumeric + `-` + `_` + `.`, drop a small stop-word set (~50 entries — small to avoid filtering domain terms).
+2. **Match keywords**:
+   - Single-word keywords: O(1) lookup against the token set.
+   - Multi-word phrases (`"app router"`, `"sql injection"`): word-boundary regex search.
+3. **Score**: each keyword hit contributes `weight` (default 1.0) to its target agent.
+4. **Specialty boost**: when the same agent matches N>1 distinct keywords, add `(N-1) * specialty_boost` (default 0.6). This rewards prompts with strong domain signal.
+5. **Confidence** = `min(score / saturation, 1.0)`. Saturation 3.0 means a single hit yields ~0.33 confidence; two hits ~0.87.
+6. **Threshold**: if best confidence < 0.25, return `general-purpose` fallback.
+
+### Phase 2: Hybrid (LLM fallback)
+
+When `--use-llm-fallback` is on AND keyword confidence < `--llm-threshold` (default `0.5`):
+
+1. Build a candidate list: top-N keyword candidates first, padded to 30 agents from the active set for breadth.
+2. Invoke the Claude CLI selector via subprocess (`claude -p --model claude-haiku-4-5 ...`) with timeout `30s` and per-call budget cap `$0.05`.
+3. Parse the JSON output `{"agent": "...", "confidence": 0.0-1.0, "reason": "..."}`. The model may wrap in `\\\`\\\`\\\`json` fences — those are stripped.
+4. **Name normalization** maps loose responses (`Code Reviewer`, `code_reviewer`) to canonical agent names. Unknown agents trigger a retry.
+5. Up to **3 retries** on timeout / parse failure / unknown agent. After 3 failures, the keyword result is kept and the failure is surfaced in `reason`.
+6. **Cycle detection** reads `~/.claude/agent-router-history.json` (last 100 entries). When the same prompt hash dispatched to the same agent twice in a row, the router forces a switch to the next-best alternative.
+
+### Tuning levers
+
+| Field | Default | Effect |
+|---|---:|---|
+| `threshold` (table) | 0.25 | Lower = route more aggressively to specialists |
+| `saturation` (table) | 3.0 | Lower = single keyword hits more confident |
+| `specialty_boost` (table) | 0.6 | Higher = stack-rank prompts that hit multiple kw on same agent |
+| `--llm-threshold` | 0.5 | Confidence under this triggers LLM selector (Hybrid mode) |
+| `--llm-max-attempts` | 3 | Selector retries before keyword fallback |
+| `--llm-budget-usd` | 0.05 | Per-call cost cap (forwarded to `claude -p --max-budget-usd`) |
+
+Empirical tuning on 1,178 historical `general-purpose` prompts achieves **~84% routing to named agents** at the current keyword settings. Phase 2 Hybrid mode (mock-evaluated) routes the residual ~16% low-confidence prompts to a named agent in the majority of cases — see [`docs/AGENT-ROUTER.md`](../../../docs/AGENT-ROUTER.md) for the measured uplift.
+
+## Adding / removing agents
+
+When a new agent is added under `.claude/agents/<category>/<name>.md`:
+
+```yaml
+# dispatch-table.yml
+keywords:
+  ...
+  <new-keyword>: <new-agent-name>
+```
+
+When an agent is **archived** (moved to `docs/archive/agents/`):
+
+1. Open `dispatch-table.yml`.
+2. Search for the archived agent name and **delete every keyword** pointing at it.
+3. Run `python3 -m unittest discover tests -v` to confirm nothing breaks.
+
+If an archived agent is left in the table, the router will return its name and the caller will fail at `Agent(subagent_type=...)` invocation. The skill never validates that recommended agents still exist on disk — keep the table clean.
+
+## Coordinating with the archive subagent
+
+The agent inventory was reduced from 144 → 100 active agents in a parallel task (see `docs/draft/agent-router-handoff-2026-05-05.md`). The dispatch table here was authored against that final 100-agent active set:
+
+- Archived language specialists (PowerShell, .NET, Rails, Java, Kotlin, Swift, C++, Vue, Angular, FastAPI, Django, Elixir, Flutter, PHP, Rust, etc.) have **no entries** in the table; their typical keywords route to general specialists (e.g. `django` → `python-pro`, `vue` → `frontend-developer`).
+- Archived infra/security/data agents (`azure-infra-engineer`, `chaos-engineer`, `ai-engineer`, `reinforcement-learning-engineer`, etc.) have no entries either.
+- The `09-meta-orchestration/agent-router.md` agent itself (if added) is not currently in the dispatch table — the router skill is meant to be invoked, not subagented.
+
+## Tests
+
+```bash
+cd .claude/skills/agent-router
+python3 -m unittest discover tests -v
+# 40 tests: parse + tokenize + dispatch + coverage + LLM fallback + cycle detection + name normalization + history recording
+```
+
+The `CoverageTests` suite verifies that ≥70% of representative prompts route correctly. The implementation currently achieves 100% on the 20-prompt fixture (keyword-only) and ~84% on the unfiltered 1,178-prompt 90-day historical sample. Phase 2 Hybrid mode further lifts low-confidence prompts; see the benchmark table below.
+
+## Benchmark
+
+| Mode | Sample | Dispatch rate | LLM calls | Cost / prompt |
+|---|---|---:|---:|---:|
+| Phase 1 (keyword only) | 1,178 historical prompts | 84.1% | 0 | $0.00 |
+| Phase 1 (keyword only) | 20 representative prompts | 100% | 0 | $0.00 |
+| Phase 2 (Hybrid mock) | 15 low-/subtle-signal prompts | 100% | 13 | ~$0.0009 |
+
+The Phase 2 row uses a deterministic mock selector to validate the wiring; live measurement against the historical 188 low-confidence sample is gated on cost — set `AGENT_ROUTER_LLM_FALLBACK=on` and re-run the harness audit to collect live numbers.
+
+## CLI examples
+
+```bash
+# Basic
+python3 router.py "review the auth module"
+# → {"agent": "security-auditor", "confidence": 1.0, "reason": "matched: auth, review", ...}
+
+# With per-candidate breakdown
+python3 router.py --explain "set up kubernetes ingress with helm"
+
+# From stdin
+echo "write tdd unit tests with pytest" | python3 router.py --stdin
+
+# With alternative table
+python3 router.py --table /tmp/experimental-table.yml "<prompt>"
+
+# Hybrid mode (LLM selector for low-confidence prompts)
+python3 router.py --use-llm-fallback --llm-threshold 0.5 "<vague prompt>"
+
+# Hybrid via env (recommended for hooks)
+AGENT_ROUTER_LLM_FALLBACK=on python3 router.py "<vague prompt>"
+```
+
+## Limitations & future work
+
+- **No semantic understanding** in keyword mode: a prompt like "I need help with Python\'s payment processing" routes on `payment` (→ payment-integration), not on `python` (→ python-pro). Specialty boost mitigates this when both keywords appear, but truly ambiguous prompts may pick the wrong specialist. Hybrid mode addresses this when enabled.
+- **No `SubagentStart` hook**: see `SKILL.md#why-a-skill-rather-than-a-hook-today`. The optional `agent-router-suggest.sh` `UserPromptSubmit` hook offers a workaround by injecting routing hints, but the main agent must still consult the result.
+- **No invocation feedback loop**: if an agent recommendation produces poor downstream results, the table doesn\'t learn. Periodically re-run the keyword-frequency analysis (see `docs/AGENT-ROUTER.md`) and tune.
+- **Hybrid cost discipline**: each LLM selector call is bounded by `--llm-budget-usd` (default $0.05) and timeout 30s. Aggregate cost still scales with the rate of low-confidence prompts; monitor `~/.claude/agent-router-history.json` and the `llm_cost_usd` field for runaway usage.

package/.claude/skills/agent-router/SKILL.md

@@ -0,0 +1,74 @@
+---
+name: agent-router
+description: Route ambiguous "general-purpose" subagent prompts to the most relevant named agent. Use when about to launch Agent/Task with subagent_type=general-purpose and unsure which named agent fits, when reviewing past general-purpose flows for downgrade to a specialist, or when the user asks "which agent should handle X?". Inputs a prompt string, scores against the dispatch table (~100 active agents under .claude/agents/), and returns the best-matching agent + confidence + reason. Hybrid mode (--use-llm-fallback / AGENT_ROUTER_LLM_FALLBACK=on) invokes a Claude CLI selector for low-confidence prompts to lift dispatch rate further. Three-stage fallback (keyword → LLM → previous → general-purpose) inherits the prior dispatch on the same prompt before falling back to general-purpose.
+---
+
+# Agent Router
+
+Establishes a hook-driven habit for the main agent to choose the right `subagent_type` at dispatch time. Recommends a specialist named agent based on prompt keywords so newly-imported specialists do not sit structurally idle.
+
+> **Note on justification**: The original framing cited "66% of historical flow went to general-purpose" as the activation rationale. That figure was measured before the 100 named agents existed in their current form — they were imported on 2026-05-04 in a single commit, so the 90-day pre-import window contains structurally unreliable dormant percentages (see `docs/INVENTORY-stocktake-2026-05-04.md` §8). The router\'s primary justification is **habit formation**: making "name the right specialist" the default behaviour at dispatch time. Dormant numbers are at best a secondary indicator and will be re-measured against post-import transcripts in 30 / 90 days. The empirical 84.1% dispatch rate against the historical prompt sample remains a meaningful coverage signal.
+
+## When to invoke
+
+The main agent (or another skill) should consult agent-router **before** launching `Agent(subagent_type="general-purpose", ...)` when:
+
+1. The prompt mentions a specific domain (review / debug / docs / infra / data / etc.).
+2. The user explicitly asks "which agent should handle this?".
+3. A previous general-purpose run produced low-confidence output and a re-run with a specialist might do better.
+
+Skip when:
+
+- The prompt is genuinely cross-cutting research with no dominant domain (router will return `general-purpose` as fallback).
+- A `subagent_type` is already explicitly specified by the user.
+
+## How to use
+
+```bash
+# CLI form (returns JSON)
+python3 .claude/skills/agent-router/router.py "review the auth module for SQL injection"
+# → {"agent": "security-auditor", "confidence": 0.71, "reason": "matched: security, auth, injection, review", "fallback": false}
+
+# Hybrid mode — invokes the Claude CLI selector when keyword confidence < 0.5
+AGENT_ROUTER_LLM_FALLBACK=on python3 .claude/skills/agent-router/router.py "investigate the supply chain dependencies"
+# → {"agent": "security-auditor", "confidence": 0.75, "reason": "llm-selector: ...", "llm_used": true, "llm_attempts": 1, "llm_cost_usd": 0.0009}
+```
+
+The CLI exits 0 with the JSON result on stdout. Exit 2 means the router itself errored (file missing, table unparseable). On exit 2 the caller should fall back to `general-purpose` and continue.
+
+## Dispatch table
+
+`dispatch-table.yml` is the single source of truth. Each entry maps a keyword (case-insensitive, word-boundary match) to the preferred named agent under `.claude/agents/<category>/<name>.md`. Multiple keywords can point at the same agent. Specialty boost is applied when several keywords from the same agent fire together.
+
+Routing model:
+
+1. Tokenize prompt (lowercase, alphanumeric / `-` / `_`, stop-word filtered).
+2. For every keyword in the table, look for a word-boundary match. Each hit contributes `weight` (default 1.0) to the target agent\'s score.
+3. Apply specialty boost: when two or more keywords resolve to the same agent, add `0.5` per extra hit.
+4. Normalize against a saturation cap (`6.0`) to produce a confidence in `[0, 1]`.
+5. If best confidence < `0.30`, return the fallback `general-purpose` with the original tokens as reason.
+6. **Hybrid mode (Phase 2)**: when `--use-llm-fallback` (or `AGENT_ROUTER_LLM_FALLBACK=on`) is set and best confidence is below `--llm-threshold` (default `0.5`), invoke a Claude CLI selector with the top 30 candidate agents. Up to 3 retries on parse / timeout / unknown-agent errors. The selector\'s recommendation overrides the keyword fallback when valid; otherwise keyword result is preserved.
+7. **Cycle detection (Phase 2)**: same prompt hash dispatched to the same agent twice in a row triggers a forced switch to the next-best agent (or `general-purpose`).
+8. **Name normalization (Phase 2)**: LLM agent names get normalized — `Code Reviewer`, `code_reviewer`, `code-reviewer` all map to the canonical `code-reviewer`.
+
+The threshold is conservative on purpose — false-positive routing wastes tokens reframing the prompt for an unsuitable specialist.
+
+## Updating the dispatch table
+
+When a new agent is added under `.claude/agents/`, update `dispatch-table.yml` with at least one keyword. When an agent is archived to `docs/archive/agents/`, **remove** its keywords from the table — routing to an archived agent fails silently and degrades to general-purpose.
+
+Validation:
+
+```bash
+python3 -m unittest discover .claude/skills/agent-router/tests -v
+```
+
+## Why a skill rather than a hook (today)
+
+Subagent dispatch happens at `Agent(subagent_type=...)` invocation time. There is no native `SubagentStart` PreToolUse hook that can rewrite `subagent_type` (PreToolUse can block but cannot transparently rewrite tool inputs in a way the calling model accepts). Until that lands, the router is invoked manually by the orchestrator (main agent) via this skill or via the optional `agent-router-suggest.sh` `UserPromptSubmit` hook, which injects routing hints into freshly submitted prompts.
+
+## Related
+
+- Inventory & rationale: `docs/INVENTORY-stocktake-2026-05-04.md`
+- Handoff context: `docs/draft/agent-router-handoff-2026-05-05.md`
+- Router architecture doc: `docs/AGENT-ROUTER.md`

package/.claude/skills/agent-router/dispatch-table.yml

@@ -0,0 +1,352 @@
+# agent-router dispatch table
+# version: 1
+# Last updated: 2026-05-04
+#
+# Format:
+#   threshold: <float>           # Minimum confidence to return a non-fallback agent (0.0–1.0)
+#   saturation: <float>          # Score above this saturates confidence at 1.0
+#   specialty_boost: <float>     # Bonus per extra keyword that hits the same agent
+#   fallback: <agent_name>       # Returned when no candidate clears threshold
+#   keywords:
+#     <keyword>: <agent_name>    # Word-boundary match, lowercased
+#     <keyword>: { agent: <name>, weight: <float> }
+#
+# Keywords are lowercased and matched on word boundaries (\b<keyword>\b).
+# Multi-word keywords (e.g. "code review") are matched as the literal phrase.
+# Hyphens / underscores are treated as part of the token.
+#
+# When you ARCHIVE an agent under docs/archive/agents/, REMOVE its keywords here
+# so the router degrades to general-purpose instead of returning a dead agent.
+
+threshold: 0.25
+saturation: 3.0
+specialty_boost: 0.6
+fallback: general-purpose
+
+keywords:
+  # ---------- 04-quality-security: review / audit / test ----------
+  review: code-reviewer
+  reviewer: code-reviewer
+  "code review": code-reviewer
+  "pr review": code-reviewer
+  refactor: refactoring-specialist
+  refactoring: refactoring-specialist
+  cleanup: refactoring-specialist
+  audit: security-auditor
+  auditor: security-auditor
+  security: security-auditor
+  vulnerability: security-auditor
+  vulnerabilities: security-auditor
+  exploit: security-auditor
+  pentest: penetration-tester
+  pen-test: penetration-tester
+  penetration: penetration-tester
+  owasp: penetration-tester
+  csrf: penetration-tester
+  xss: penetration-tester
+  injection: security-auditor
+  "sql injection": security-auditor
+  auth: security-auditor
+  authentication: security-auditor
+  authorization: security-auditor
+  oauth: security-auditor
+  jwt: security-auditor
+  rbac: security-auditor
+  secrets: security-auditor
+  credentials: security-auditor
+  compliance: compliance-auditor
+  gdpr: compliance-auditor
+  hipaa: compliance-auditor
+  pci: compliance-auditor
+  test: test-automator
+  tests: test-automator
+  testing: test-automator
+  tdd: test-automator
+  unittest: test-automator
+  pytest: test-automator
+  jest: test-automator
+  vitest: test-automator
+  qa: qa-expert
+  e2e: ui-ux-tester
+  playwright: ui-ux-tester
+  cypress: ui-ux-tester
+  accessibility: accessibility-tester
+  a11y: accessibility-tester
+  wcag: accessibility-tester
+  aria: accessibility-tester
+  debug: debugger
+  debugger: debugger
+  debugging: debugger
+  trace: debugger
+  stacktrace: debugger
+  exception: error-detective
+  error: error-detective
+  failure: error-detective
+  crash: error-detective
+  performance: performance-engineer
+  optimization: performance-engineer
+  profiling: performance-engineer
+  bottleneck: performance-engineer
+  architecture: architect-reviewer
+  architect: architect-reviewer
+  "architecture review": architect-reviewer
+  "ai writing": ai-writing-auditor
+  "ai-generated": ai-writing-auditor
+
+  # ---------- 01-core-development ----------
+  api: api-designer
+  endpoint: api-designer
+  rest: api-designer
+  openapi: api-designer
+  swagger: api-designer
+  graphql: graphql-architect
+  microservice: microservices-architect
+  microservices: microservices-architect
+  websocket: websocket-engineer
+  socket.io: websocket-engineer
+  realtime: websocket-engineer
+  fullstack: fullstack-developer
+  full-stack: fullstack-developer
+  backend: backend-developer
+  frontend: frontend-developer
+  ui: ui-designer
+  "ui design": ui-designer
+  "user interface": ui-designer
+  mobile: mobile-developer
+  electron: electron-pro
+
+  # ---------- 02-language-specialists (active subset only) ----------
+  python: python-pro
+  py: python-pro
+  django: python-pro          # django-developer is archived; route to python-pro
+  fastapi: python-pro         # fastapi-developer is archived; route to python-pro
+  typescript: typescript-pro
+  ts: typescript-pro
+  tsx: typescript-pro
+  javascript: javascript-pro
+  js: javascript-pro
+  esm: javascript-pro
+  node: node-specialist
+  node.js: node-specialist
+  nodejs: node-specialist
+  npm: node-specialist
+  pnpm: node-specialist
+  golang: golang-pro
+  go: golang-pro
+  goroutine: golang-pro
+  next: nextjs-developer
+  next.js: nextjs-developer
+  nextjs: nextjs-developer
+  "app router": nextjs-developer
+  rsc: nextjs-developer
+  "server component": nextjs-developer
+  react: react-specialist
+  jsx: react-specialist
+  hook: react-specialist
+  hooks: react-specialist
+  usestate: react-specialist
+  useeffect: react-specialist
+  sql: sql-pro
+  query: sql-pro
+
+  # ---------- 03-infrastructure ----------
+  cloud: cloud-architect
+  aws: cloud-architect
+  gcp: cloud-architect
+  s3: cloud-architect
+  ec2: cloud-architect
+  lambda: cloud-architect
+  cloudfront: cloud-architect
+  database: database-administrator
+  db: database-administrator
+  schema: database-administrator
+  migration: database-administrator
+  deploy: deployment-engineer
+  deployment: deployment-engineer
+  release: deployment-engineer
+  rollout: deployment-engineer
+  devops: devops-engineer
+  ci: devops-engineer
+  cd: devops-engineer
+  "ci/cd": devops-engineer
+  pipeline: devops-engineer
+  github-actions: devops-engineer
+  docker: docker-expert
+  dockerfile: docker-expert
+  container: docker-expert
+  compose: docker-expert
+  kubernetes: kubernetes-specialist
+  k8s: kubernetes-specialist
+  kubectl: kubernetes-specialist
+  helm: kubernetes-specialist
+  pod: kubernetes-specialist
+  network: network-engineer
+  dns: network-engineer
+  vpn: network-engineer
+  tcp: network-engineer
+  platform: platform-engineer
+  internal-tooling: platform-engineer
+  sre: sre-engineer
+  reliability: sre-engineer
+  slo: sre-engineer
+  sli: sre-engineer
+  terraform: terraform-engineer
+  hcl: terraform-engineer
+  iac: terraform-engineer
+  pulumi: terraform-engineer
+  "security engineering": security-engineer
+  hardening: security-engineer
+
+  # ---------- 05-data-ai ----------
+  llm: llm-architect
+  "llm architecture": llm-architect
+  rag: llm-architect
+  embedding: llm-architect
+  embeddings: llm-architect
+  prompt: prompt-engineer
+  "prompt engineering": prompt-engineer
+  "system prompt": prompt-engineer
+  ml: ml-engineer
+  machine-learning: machine-learning-engineer
+  classifier: machine-learning-engineer
+  regression: machine-learning-engineer
+  mlops: mlops-engineer
+  nlp: nlp-engineer
+  tokenizer: nlp-engineer
+  data: data-engineer
+  etl: data-engineer
+  pipeline-ingest: data-engineer
+  airflow: data-engineer
+  dbt: data-engineer
+  parquet: data-engineer
+  pandas: data-analyst
+  dataframe: data-analyst
+  notebook: data-scientist
+  jupyter: data-scientist
+  postgres: postgres-pro
+  postgresql: postgres-pro
+  pgsql: postgres-pro
+  index: database-optimizer
+  "query plan": database-optimizer
+  explain: database-optimizer
+
+  # ---------- 06-developer-experience ----------
+  build: build-engineer
+  webpack: build-engineer
+  vite: build-engineer
+  rollup: build-engineer
+  turbopack: build-engineer
+  cli: cli-developer
+  command-line: cli-developer
+  argparse: cli-developer
+  dependency: dependency-manager
+  dependencies: dependency-manager
+  "package update": dependency-manager
+  documentation: documentation-engineer
+  docs: documentation-engineer
+  doc: documentation-engineer
+  tutorial: documentation-engineer
+  guide: documentation-engineer
+  readme: readme-generator
+  changelog: documentation-engineer
+  legacy: legacy-modernizer
+  modernize: legacy-modernizer
+  "code modernization": legacy-modernizer
+  mcp: mcp-developer
+  "mcp server": mcp-developer
+  slack: slack-expert
+  "slack workflow": slack-expert
+  "slack bot": slack-expert
+  git-workflow: git-workflow-manager
+  rebase: git-workflow-manager
+  cherry-pick: git-workflow-manager
+  branch: git-workflow-manager
+  commit: git-workflow-manager
+  dx: dx-optimizer
+  "developer experience": dx-optimizer
+  tooling: tooling-engineer
+
+  # ---------- 07-specialized-domains ----------
+  "api docs": api-documenter
+  "api reference": api-documenter
+  "openapi spec": api-documenter
+  ios: mobile-app-developer
+  android: mobile-app-developer
+  "react native": mobile-app-developer
+  payment: payment-integration
+  stripe: payment-integration
+  paypal: payment-integration
+  billing: payment-integration
+  seo: seo-specialist
+  sitemap: seo-specialist
+  metadata: seo-specialist
+  opengraph: seo-specialist
+
+  # ---------- 08-business-product ----------
+  "business analysis": business-analyst
+  "requirements analysis": business-analyst
+  "content marketing": content-marketer
+  "blog post": content-marketer
+  newsletter: content-marketer
+  "customer success": customer-success-manager
+  "user retention": customer-success-manager
+  legal: legal-advisor
+  contract: legal-advisor
+  tos: legal-advisor
+  privacy: legal-advisor
+  license: license-engineer
+  "open source license": license-engineer
+  "product manager": product-manager
+  prd: product-manager
+  "product spec": product-manager
+  roadmap: product-manager
+  "project manager": project-manager
+  "project plan": project-manager
+  gantt: project-manager
+  "sales engineering": sales-engineer
+  rfp: sales-engineer
+  scrum: scrum-master
+  sprint: scrum-master
+  retrospective: scrum-master
+  technical-writing: technical-writer
+  "technical writing": technical-writer
+  whitepaper: technical-writer
+  "ux research": ux-researcher
+  "user interview": ux-researcher
+  usability: ux-researcher
+  wordpress: wordpress-master
+  wp-admin: wordpress-master
+  woocommerce: wordpress-master
+
+  # ---------- 09-meta-orchestration ----------
+  "agent install": agent-installer
+  "agent organize": agent-organizer
+  "agent organizer": agent-organizer
+  orchestrate: workflow-orchestrator
+  workflow: workflow-orchestrator
+  multi-agent: multi-agent-coordinator
+  "context manager": context-manager
+  "error coordinator": error-coordinator
+  "error coordination": error-coordinator
+  "knowledge synthesis": knowledge-synthesizer
+  "task distribution": task-distributor
+  it-ops: it-ops-orchestrator
+  "performance monitor": performance-monitor
+  "codebase orchestrator": codebase-orchestrator
+
+  # ---------- 10-research-analysis ----------
+  research: research-analyst
+  "research analyst": research-analyst
+  "deep research": research-analyst
+  competitive: competitive-analyst
+  "competitive analysis": competitive-analyst
+  competitor: competitive-analyst
+  "data research": data-researcher
+  "market research": market-researcher
+  market: market-researcher
+  trend: trend-analyst
+  trends: trend-analyst
+  "idea validation": project-idea-validator
+  validate: project-idea-validator
+  search-specialist: search-specialist
+  "advanced search": search-specialist