@cohortapp/agent-sdk 2.3.0

package/docs/guides/mac-mini.md

@@ -0,0 +1,119 @@
+# Mac mini bring-up — an always-on Cohort agent, end to end
+
+The one-page operator runbook: from a boxed Apple-silicon Mac mini to a fully
+enrolled AI colleague on `os.cohortapp.com` — messaging, board work, and a real
+workspace email address — verified by `cohort doctor`. OS-level hardening and
+fleet mechanics live in the deeper runbook: [Mac Mini Bootstrap](../runbooks/mac-mini-bootstrap.md)
+(this page links it rather than duplicating it).
+
+## 1. Hardware / prereqs
+
+- Apple-silicon Mac mini (M-series), macOS 15+.
+- A dedicated admin macOS account for the agent (keep your own account separate).
+- Always-on power settings:
+
+```bash
+sudo pmset -a sleep 0 displaysleep 0 autorestart 1
+```
+
+## 2. Bootstrap
+
+```bash
+xcode-select --install
+/bin/bash -c "$(curl -fsSL https://raw.githubusercontent.com/Homebrew/install/HEAD/install.sh)"
+brew install node@20 git python@3.12
+npm i -g @anthropic-ai/claude-code && claude login   # Max subscription — the daemon
+                                                     # strips API keys so sub-sessions
+                                                     # ride keychain OAuth
+npm i -g @cohortapp/agent-sdk
+```
+
+## 3. Create the agent repo
+
+```bash
+cohort create <agent-name>
+cd <agent-name>
+```
+
+`create` copies the framework (`lib/`, `scripts/`, `.claude/`, `scaffold` configs —
+including `.mcp.json`, which exposes the `cohort-mcp` org tool surface to
+interactive Claude sessions, and the PreToolUse hooks that keep its outbound
+writes on the CLI lane).
+
+## 4. Enroll against `https://os.cohortapp.com`
+
+Two lanes — pick one:
+
+- **Admin-minted key**: in Cohort → Settings → API keys mint an agent key and
+  pair it to the member, then:
+
+  ```bash
+  export COHORT_API_KEY=nlk_…  COHORT_ORG_ID=<org-slug>  COHORT_AGENT_ID=<member-slug>
+  cohort setup
+  ```
+
+  Pull-enrollment populates `config/agent.json` from the member profile and
+  writes `config/org.yaml` (`org.cohort.{enabled,base,orgId,token}`).
+
+- **Pre-auth pairing**: `cohort pair <code>` → an org admin approves the
+  handshake in Cohort (`pairing.request`/`pairing.approve`).
+
+## 5. Wizard
+
+`cohort setup` walks the sections in order — the org trio is:
+
+| Order | Section | What it does |
+| --- | --- | --- |
+| 75 | `org` | endpoint + token (the enrollment SoT) |
+| 76 | `messaging` | messaging.read/write + calling.write scopes, home channels |
+| 77 | `orgmail` | **workspace mailbox** — writes `config/orgmail.yaml` |
+
+For 77 to go live, an admin must have verified a domain and assigned this
+agent a mailbox in Cohort → Settings → Email first (the wizard's verify step
+probes `email.inbox` and WARNS until then — it never blocks setup).
+
+## 6. Always-on (launchd)
+
+```bash
+scripts/setup/init-agent.sh     # deps → state dirs → generate-plists.sh → launchctl load
+launchctl list | grep ai.maestro
+```
+
+Expect the daemon plus the cadence triggers; the daemon plist is KeepAlive
+`{SuccessfulExit:false, Crashed:true}` so it restarts on crash but respects a
+clean stop.
+
+## 7. Verify — `cohort doctor` is the gate
+
+```bash
+cohort doctor
+```
+
+All green includes the **Cohort connectivity** block: enrollment resolved →
+directory reachable (latency) → protocol version aligned (the server's
+`x-org-protocol` header; a drift warning means run `cohort upgrade`) → **key
+paired to a workforce member** (the `messaging.channels` gate) → **workspace
+mailbox reachable** (when `config/orgmail.yaml` exists).
+
+Then two live probes:
+
+1. **DM the agent in Cohort** — expect a reply (inbound rides the messaging
+   plane; watch `state/inbox/` fill and `logs/` narrate).
+2. **Email the agent's workspace address** — watch `state/inbox/orgmail/`
+   populate within ≤45s (the poll tick), and expect a correctly-threaded reply
+   dispatched by the hq outbox within ≤2 min (post-commit kick ≈1.5s, loop
+   tick ≤60s). First-contact recipients park as approvals in /decisions — by
+   design.
+
+## 8. Operations
+
+- **Kill switch**: touch `.emergency-stop` in the repo root (human-only). The
+  daemon exits clean, so launchd stops the restart treadmill. Resume with
+  `scripts/resume-operations.sh`.
+- **Email-only kill switch**: delete/rename `config/orgmail.yaml` and restart —
+  the gate-file loop skips the platform; server-side an admin can flip the
+  mailbox to Disabled in Settings → Email.
+- **Upgrade**: `cohort upgrade` (framework files), `npm i -g @cohortapp/agent-sdk@latest`
+  (the SDK itself). A doctor "protocol drift" warning is the upgrade cue.
+- **Health**: `scripts/healthcheck.sh`; logs under `logs/` (`logs/audit/` carries
+  the per-action JSONL rows from BOTH the native executor and the MCP plane).

package/docs/guides/media-generation-setup.md

@@ -0,0 +1,349 @@
+# Media Generation Setup Guide
+
+How to generate branded illustrations, diagrams, and video assets using Google Gemini and Veo APIs. Covers API setup, prompt specification authoring, image generation, video generation, brand alignment, and troubleshooting.
+
+**Prerequisites**: Complete the [Mac Mini Bootstrap](../runbooks/mac-mini-bootstrap.md). Node.js 20+ must be installed.
+
+---
+
+## Architecture Overview
+
+```
+┌─────────────────────────────────────────────────────────────────┐
+│  PROMPT SPECS (TypeScript/ESM files)                             │
+│                                                                   │
+│  scripts/media-generation/prompts/                               │
+│    ├── illustrations/   *.ts / *.mjs                             │
+│    ├── diagrams/        *.ts / *.mjs                             │
+│    └── videos/          *.ts / *.mjs                             │
+│                                                                   │
+│  Each spec exports: { id, slug, title, purpose, style,           │
+│    colorPalette, mood, elements, aspectRatio, ... }              │
+├─────────────────────────────────────────────────────────────────┤
+│  GENERATION                                                      │
+│                                                                   │
+│  generate-assets.mjs                                             │
+│    ├── --illustration <slug> → gemini-image-client.mjs           │
+│    │     └── @google/genai SDK → Gemini 3 Pro / Imagen           │
+│    │         └── public/generated/illustrations/<slug>.png       │
+│    │                                                              │
+│    ├── --video <slug> → veo-video-client.mjs                     │
+│    │     └── @google/genai SDK → Veo 3.1                         │
+│    │         └── public/generated/videos/<slug>.mp4              │
+│    │                                                              │
+│    ├── --all-missing → batch generate all missing illustrations  │
+│    └── --all-missing-videos → batch generate all missing videos  │
+│                                                                   │
+│  Rate limit: 15s between requests                                │
+│  Generation log: scripts/media-generation/generation-log.json    │
+├─────────────────────────────────────────────────────────────────┤
+│  OUTPUT                                                          │
+│                                                                   │
+│  public/generated/                                               │
+│    ├── illustrations/   *.png                                    │
+│    ├── diagrams/        *.png                                    │
+│    └── videos/          *.mp4                                    │
+└─────────────────────────────────────────────────────────────────┘
+```
+
+---
+
+## 1. API Setup
+
+### 1.1 Get a Gemini API Key
+
+1. Go to https://aistudio.google.com/apikey
+2. Create a new API key (or use an existing one)
+3. The same key works for both image (Gemini/Imagen) and video (Veo) generation
+
+### 1.2 Configure Environment
+
+Add to `.env`:
+
+```bash
+GEMINI_API_KEY=your-api-key-here
+```
+
+### 1.3 Install Dependencies
+
+```bash
+cd ~/agent-repo && npm install
+```
+
+Required packages (in `package.json`):
+- `@google/genai` — Google AI SDK for Gemini and Veo
+- `dotenv` — Environment variable loading
+
+### 1.4 Verify Setup
+
+```bash
+# Quick test — list available prompt specs
+node scripts/media-generation/generate-assets.mjs --list
+```
+
+---
+
+## 2. Image Generation (Gemini / Imagen)
+
+### 2.1 Generate a Single Illustration
+
+```bash
+node scripts/media-generation/generate-assets.mjs --illustration <slug>
+```
+
+The `<slug>` must match a prompt spec file in `scripts/media-generation/prompts/illustrations/`.
+
+### 2.2 Generate All Missing
+
+```bash
+# Generate illustrations for all specs that don't have output files yet
+node scripts/media-generation/generate-assets.mjs --all-missing
+```
+
+### 2.3 Models
+
+| Model | Quality | Speed | Use Case |
+|---|---|---|---|
+| `gemini-3-pro-image-preview` | Best | Slower | Default — board packs, investor materials |
+| `imagen-*` variants | Good | Fast | High-volume generation, social media |
+
+Override via prompt spec or environment:
+
+```bash
+GEMINI_MODEL=imagen-3.0-generate-002 node scripts/media-generation/generate-assets.mjs --illustration <slug>
+```
+
+### 2.4 Aspect Ratios
+
+Supported: `1:1`, `3:4`, `4:3`, `9:16`, `16:9`
+
+Set per prompt spec in the `aspectRatio` field.
+
+---
+
+## 3. Video Generation (Veo 3.1)
+
+### 3.1 Generate a Video
+
+```bash
+node scripts/media-generation/generate-assets.mjs --video <slug>
+```
+
+### 3.2 Generate All Missing Videos
+
+```bash
+node scripts/media-generation/generate-assets.mjs --all-missing-videos
+```
+
+### 3.3 Configuration
+
+| Setting | Default | Options |
+|---|---|---|
+| Model | `veo-3.1-generate-preview` | — |
+| Duration | 8 seconds | 4, 6, 8 |
+| Resolution | 720p | 720p, 1080p, 4k |
+| Modes | text-to-video | text-to-video, image-to-video |
+
+### 3.4 Polling
+
+Video generation is asynchronous. The client:
+1. Submits the generation request
+2. Polls every 10 seconds for completion
+3. Times out after 10 minutes
+4. Downloads the video file on completion
+
+---
+
+## 4. Prompt Specifications
+
+### 4.1 Structure
+
+Prompts are TypeScript (`.ts`) or ESM (`.mjs`) files in `scripts/media-generation/prompts/`:
+
+```
+prompts/
+  illustrations/    # Board pack covers, section illustrations
+  diagrams/         # Architecture diagrams, flow charts
+  videos/           # Animated intros, presentation backgrounds
+```
+
+### 4.2 Writing a Prompt Spec
+
+```typescript
+// scripts/media-generation/prompts/illustrations/board-pack-cover-q2.ts
+export default {
+  id: "board-pack-cover-q2-2026",
+  slug: "board-pack-cover-q2-2026",
+  title: "Q2 2026 Board Pack Cover Illustration",
+  purpose: "Cover page visual for quarterly board materials",
+  imageCategory: "editorial-hand-drawn",
+  colorPalette: ["#1C1917", "#FAFAF9", "#78716C", "#D6D3D1"],
+  mood: ["institutional", "sophisticated", "premium"],
+  elements: [
+    "abstract representation of global financial network",
+    "interconnected nodes suggesting multi-jurisdiction structure",
+    "subtle reference to algorithmic trading"
+  ],
+  style: "Monochromatic NYT-style hand-drawn editorial illustration. " +
+    "Pen-and-ink technique with cross-hatching and stippling. " +
+    "No color, no gradients, no digital effects.",
+  aspectRatio: "3:4",
+};
+```
+
+### 4.3 Key Fields
+
+| Field | Required | Description |
+|---|---|---|
+| `id` | Yes | Unique identifier |
+| `slug` | Yes | Filename-safe identifier (used for output filename) |
+| `title` | Yes | Human-readable title |
+| `purpose` | Yes | What this asset is used for |
+| `style` | Yes | Detailed style description for the AI model |
+| `colorPalette` | No | Hex colour values (brand-aligned) |
+| `mood` | No | Tone descriptors |
+| `elements` | No | Visual elements to include |
+| `aspectRatio` | No | Output aspect ratio (default: `16:9`) |
+| `imageCategory` | No | Category tag for organisation |
+
+### 4.4 TypeScript vs ESM
+
+- `.ts` files require `tsx` (installed via devDependencies) — the generator shells out to `npx tsx` to evaluate them
+- `.mjs` files are imported directly — slightly faster, no build step
+- Both work identically; use `.ts` for consistency with the rest of the project
+
+---
+
+## 5. Brand Alignment
+
+All generated assets must follow `config/brand-assets.yaml`:
+
+### 5.1 Illustration Style
+
+- **Monochromatic editorial hand-drawn** (NYT-style)
+- Pen-and-ink technique: cross-hatching, stippling
+- No colour, no gradients, no digital effects
+- Restrained and institutional mood
+
+### 5.2 Colour Palette
+
+| Colour | Hex | Usage |
+|---|---|---|
+| Near-black | `#1C1917` | Primary line work |
+| Off-white | `#FAFAF9` | Background/negative space |
+| Warm gray | `#78716C` | Secondary elements |
+| Light gray | `#D6D3D1` | Tertiary/halftone areas |
+| Medium gray | `#A8A29E` | Mid-tones |
+
+### 5.3 What to Avoid
+
+- Cartoon or comic styles
+- Isometric SaaS/tech aesthetics
+- Stock photography look
+- Bright colours or saturated palettes
+- Clip art or flat design icons
+
+---
+
+## 6. Output & Logging
+
+### 6.1 Output Directories
+
+| Type | Path |
+|---|---|
+| Illustrations | `public/generated/illustrations/<slug>.png` |
+| Diagrams | `public/generated/diagrams/<slug>.png` |
+| Videos | `public/generated/videos/<slug>.mp4` |
+
+### 6.2 Generation Log
+
+Every generation is recorded in `scripts/media-generation/generation-log.json`:
+
+```json
+[
+  {
+    "slug": "board-pack-cover-q2-2026",
+    "type": "illustration",
+    "model": "gemini-3-pro-image-preview",
+    "generatedAt": "2026-04-09T14:30:00.000Z",
+    "outputPath": "public/generated/illustrations/board-pack-cover-q2-2026.png"
+  }
+]
+```
+
+### 6.3 Rate Limiting
+
+A 15-second delay is enforced between consecutive API requests to avoid rate limits. For batch generation (`--all-missing`), this means ~4 images per minute.
+
+---
+
+## 7. Testing
+
+| # | Test | How to Verify |
+|---|---|---|
+| 1 | API key valid | `node -e "require('dotenv').config(); console.log(process.env.GEMINI_API_KEY?.slice(0,8))"` |
+| 2 | List specs | `node scripts/media-generation/generate-assets.mjs --list` |
+| 3 | Generate image | Generate one illustration and verify output in `public/generated/` |
+| 4 | Generate video | Generate one video (takes several minutes) |
+| 5 | Brand alignment | Verify output matches monochromatic editorial style |
+| 6 | Batch generation | `--all-missing` — should skip already-generated assets |
+| 7 | Generation log | Check `generation-log.json` for entries |
+
+---
+
+## 8. Troubleshooting
+
+### "GEMINI_API_KEY not set"
+
+1. Check `.env` file exists and contains the key
+2. Verify key format (should be a long alphanumeric string)
+3. Get a key from https://aistudio.google.com/apikey
+
+### Rate limit errors
+
+1. The 15-second delay should prevent most rate limits
+2. If hit, wait 60 seconds and retry
+3. For heavy batch generation, consider running overnight
+4. Check your API quota at https://aistudio.google.com/
+
+### "Prompt spec not found"
+
+1. Check the slug matches a file in `scripts/media-generation/prompts/illustrations/` (or `diagrams/`, `videos/`)
+2. File must end in `.ts` or `.mjs`
+3. List available specs: `node scripts/media-generation/generate-assets.mjs --list`
+
+### Image quality issues
+
+1. Review the `style` field in the prompt spec — be more specific about the desired aesthetic
+2. Add negative prompts ("no cartoon", "no digital effects")
+3. Try a different model (Imagen may produce different results than Gemini)
+4. Adjust aspect ratio to match the intended use
+
+### Video generation timeout
+
+1. Default timeout is 10 minutes — video generation can be slow
+2. Check the Veo API status if consistently timing out
+3. Try shorter duration (4s instead of 8s) for faster generation
+4. Lower resolution (720p) generates faster than 1080p/4k
+
+---
+
+## Key Files
+
+| File | Purpose |
+|---|---|
+| `scripts/media-generation/generate-assets.mjs` | Main generation orchestrator |
+| `scripts/media-generation/gemini-image-client.mjs` | Google Gemini/Imagen image API client |
+| `scripts/media-generation/veo-video-client.mjs` | Veo 3.1 video API client |
+| `scripts/media-generation/prompts/` | Prompt spec directory (illustrations, diagrams, videos) |
+| `scripts/media-generation/generation-log.json` | Generation history log |
+| `public/generated/` | Output directory for all generated assets |
+| `config/brand-assets.yaml` | Brand style guide (colours, typography, illustration style) |
+
+---
+
+## Related Documents
+
+- [PDF Generation Setup](pdf-generation-setup.md) — Using generated assets in branded PDFs
+- [Agent Persona Setup](agent-persona-setup.md) — Brand configuration
+- [Mac Mini Bootstrap](../runbooks/mac-mini-bootstrap.md) — Node.js and dependency installation

package/docs/guides/model-routing.md

@@ -0,0 +1,162 @@
+# Model routing — Anthropic / Moonshot / Qwen via one config
+
+Maestro's daemon historically routed every spawned `claude --print`
+through the Claude CLI's keychain-resident OAuth token (Max
+subscription). That's still the default. **The model router is
+opt-in** — drop a `config/model-routing.yaml` into an agent and the
+daemon will start routing each session to the cheapest backend that
+satisfies the request's declared capability needs.
+
+## Why
+
+Two big shifts in 2026:
+
+1. **Moonshot AI** ships a native `/anthropic` Messages endpoint, so
+   Kimi K2.6 is a literal drop-in for Claude Code — no router, no
+   wire-format translation, no tool-use degradation. K2.6 lands on
+   80.2% SWE-bench Verified (within ~12 points of Sonnet 5) at
+   roughly 30% of Sonnet's per-token cost. The Anthropic endpoint
+   also supports cache-hit pricing (~$0.16/M cache reads).
+2. **Qwen3-Coder-Next** (80B / 3B-active MoE) hits ~70% SWE-bench
+   Verified and is available everywhere — DashScope direct, OpenRouter,
+   DeepInfra, local vLLM/Ollama. At $0.22/M input via OpenRouter it's
+   the cheapest credible coding model.
+
+For a daemon that runs 100+ sessions/day across inbox, backlog, and
+cadence ticks, routing routine work to Moonshot/Qwen and only escalating
+to Anthropic when an agent actually needs `extended-thinking` or
+`prompt-cache-1h` cuts spend by ~70% with no loss on the work that
+matters.
+
+## Quick start
+
+```bash
+# In the agent repo:
+npx @cohortapp/agent-sdk init   # idempotent — places model-routing.yaml.disabled
+
+# Add credentials to .env:
+echo "MOONSHOT_API_KEY=msk-..." >> .env
+echo "OPENROUTER_API_KEY=sk-or-..." >> .env
+
+# Activate by renaming:
+mv config/model-routing.yaml.disabled config/model-routing.yaml
+
+# Restart the daemon — every new spawn now respects the routing policy.
+launchctl unload ~/Library/LaunchAgents/ai.maestro.<agent>-daemon.plist
+launchctl load   ~/Library/LaunchAgents/ai.maestro.<agent>-daemon.plist
+```
+
+## Config reference
+
+`config/model-routing.yaml` has three sections:
+
+### `backends:`
+Each backend declares the wire format, base URL, auth env var,
+capabilities, pricing, and concrete model names per tier.
+
+| Field | Meaning |
+|-------|---------|
+| `transport` | `anthropic-cli` (no env injection — current default) \| `anthropic-native` (point at any Messages endpoint) \| `openai-compat` (OpenAI-compatible like OpenRouter, DashScope) |
+| `base_url` | HTTPS endpoint base |
+| `auth_env` | Env var holding the auth token (omit for keychain OAuth) |
+| `capabilities` | Tagged-set: `thinking`, `vision`, `tool_use`, `parallel_tools`, `long_context_262k`/`_1m`, `prompt_cache_1h`/`_hit`, `degraded_tool_use` |
+| `pricing` | `input_per_m`, `output_per_m`, optional `cache_hit_per_m` |
+| `models` | `classifier` / `fast` / `default` / `premium` slots |
+
+### `routing_policy:`
+List of rules, top-down. First match wins. A `default: true` rule is
+the terminal fallback.
+
+Supported match keys:
+- `agent_role`, `agent_role_in: [...]`
+- `needs_thinking`, `needs_vision`, `needs_tool_use`,
+  `needs_parallel_tools`, `needs_long_context`
+- `tier: classifier | fast | default | premium`
+- `model_hint: haiku | sonnet | opus`
+- `token_estimate_gte: 60000` / `token_estimate_lt: …`
+- `source: inbox | backlog | cadence | voice`
+- `priority: critical | high | normal`
+
+Each rule can also declare a `fallback: [backend_name, …]` list: if
+the primary backend doesn't satisfy the request's capability needs,
+the resolver tries each fallback in order before failing through to
+the global Anthropic safety net.
+
+### Global flags
+- `fallback_to_anthropic` (default `true`) — final safety net when no
+  backend satisfies a request. Set `false` to make the router strict.
+- `strip_attribution_header` (default `true`) — sets
+  `CLAUDE_CODE_ATTRIBUTION_HEADER=0` for non-Anthropic backends so a
+  third-party gateway's prompt-cache key isn't poisoned.
+- `disable_experimental_betas` (default `true`) — sets
+  `CLAUDE_CODE_DISABLE_EXPERIMENTAL_BETAS=1` for non-Anthropic backends
+  so `/advisor`, etc. don't silently degrade.
+
+## Recommended policy template
+
+```yaml
+routing_policy:
+  # Safety — never cheap-route security-critical work.
+  - match: { agent_role_in: [ceo_pre_pass, audit, decision_writer, regulatory] }
+    backend: anthropic
+  # Extended-thinking → Anthropic.
+  - match: { needs_thinking: true }
+    backend: anthropic
+    fallback: [moonshot]
+  # Reliable tool use → Moonshot first (native Anthropic shape).
+  - match: { needs_tool_use: true }
+    backend: moonshot
+    fallback: [anthropic]
+  # Long-context (>60k) → Moonshot 262k window.
+  - match: { token_estimate_gte: 60000 }
+    backend: moonshot
+    fallback: [anthropic]
+  # Classifier / background → cheapest.
+  - match: { tier: classifier }
+    backend: openrouter_qwen
+    fallback: [moonshot, anthropic]
+  - match: { source: backlog }
+    backend: openrouter_qwen
+    fallback: [moonshot, anthropic]
+  - default: true
+    backend: moonshot
+    fallback: [anthropic]
+```
+
+## Observability
+
+Every spawn that picks a non-`anthropic-cli` backend writes a
+`routed` event to `logs/daemon/<date>-sessions.jsonl`:
+
+```json
+{
+  "event": "routed",
+  "sessionId": "s-...",
+  "backend": "moonshot",
+  "transport": "anthropic-native",
+  "model": "kimi-k2.6",
+  "tried": [...],
+  "fallback_reason": null
+}
+```
+
+Use these to compute cost-per-agent, success-rate-per-model, and
+failover counters. A future `scripts/cost/track-claude-usage.mjs`
+patch will surface backend breakdowns in the daily
+`state/dashboards/cost-summary.yaml`.
+
+## Risks + guardrails
+
+| Risk | Symptom | Mitigation |
+|------|---------|------------|
+| Empty `tool_use.input` from OpenRouter+Qwen | Agent loops do nothing, claude logs "no input" | Tag backend with `degraded_tool_use`; router routes `needs_tool_use: true` away. |
+| Lost prompt caching | Cost regression after switching | Only Moonshot supports cache-hit pricing on non-Anthropic backends. Watch `state/cost-tracking/` for >2× spend on previously-cached prompts. |
+| Hallucinated tool names | Agent calls non-existent tools | maestro's existing `tool_not_found` counter catches this; auto-escalate after N consecutive misses by piping the agent through a fallback Anthropic rule. |
+| `/advisor` (Anthropic-only) silently degrading | Subtle quality drop | `disable_experimental_betas: true` is the default; leave it on. |
+| LiteLLM compromised versions | Credential theft (1.82.7/1.82.8) | Pin router/proxy versions in `package.json`; never auto-upgrade router binaries. |
+
+## See also
+- `lib/model-router.mjs` — source
+- `lib/model-router.test.mjs` — behaviour reference
+- `scripts/setup/init-model-router.mjs` — wizard
+- `scaffold/config/model-routing.yaml.example` — template