hail-hydra-cc 2.3.1 → 2.3.2

package/files/references/routing-guide.md

@@ -1,303 +1,303 @@
-# Routing Guide — Mandatory Delegation Examples
-
-This reference provides concrete examples to see mandatory delegation in action.
-Read this when you need to see how delegation rules apply to real tasks.
-
-## Table of Contents
-1. [ALWAYS Delegate — Haiku 4.5 Agents](#always-delegate-haiku)
-2. [ALWAYS Delegate — Sonnet 4.6 Agents](#always-delegate-sonnet)
-3. [ALWAYS Handle Yourself — Opus 4.6](#always-handle-yourself-opus)
-4. [Compound Task Decomposition](#compound-tasks)
-5. [Common Misclassifications](#common-misclassifications)
-
----
-
-## ALWAYS Delegate — Haiku 4.5 Agents
-
-### hydra-scout (Haiku 4.5) examples
-| User says | Route to | Why |
-|-----------|----------|-----|
-| "What framework is this project using?" | hydra-scout (Haiku 4.5) | Read package.json/config files |
-| "Find where the login endpoint is defined" | hydra-scout (Haiku 4.5) | Grep for routes/endpoints |
-| "How many files are in the src directory?" | hydra-scout (Haiku 4.5) | Glob + count |
-| "Show me the database schema" | hydra-scout (Haiku 4.5) | Find and read schema/migration files |
-| "What does the User model look like?" | hydra-scout (Haiku 4.5) | Find and read model definition |
-| "Is there a rate limiter in this project?" | hydra-scout (Haiku 4.5) | Grep for rate limit patterns |
-| "What version of React are we using?" | hydra-scout (Haiku 4.5) | Read package.json |
-
-### hydra-runner (Haiku 4.5) examples
-| User says | Route to | Why |
-|-----------|----------|-----|
-| "Run the tests" | hydra-runner (Haiku 4.5) | Execute test command, report results |
-| "Does the build pass?" | hydra-runner (Haiku 4.5) | Run build, report success/failure |
-| "Check if there are any lint errors" | hydra-runner (Haiku 4.5) | Run linter, report findings |
-| "What's the git status?" | hydra-runner (Haiku 4.5) | Run git status/diff |
-| "Run the migration" | hydra-runner (Haiku 4.5) | Execute migration command |
-| "Check if the server starts" | hydra-runner (Haiku 4.5) | Start server, check for errors |
-
-### hydra-scribe (Haiku 4.5) examples
-| User says | Route to | Why |
-|-----------|----------|-----|
-| "Add docstrings to this file" | hydra-scribe (Haiku 4.5) | Read code, write docstrings |
-| "Update the README with the new API endpoints" | hydra-scribe (Haiku 4.5) | Descriptive writing from code |
-| "Write a changelog entry for these changes" | hydra-scribe (Haiku 4.5) | Summarize changes |
-| "Add comments explaining this function" | hydra-scribe (Haiku 4.5) | Read and annotate |
-
-### hydra-guard (Haiku 4.5) examples
-| User says | Route to | Why |
-|-----------|----------|-----|
-| "Scan my changes for security issues" | hydra-guard (Haiku 4.5) | Security gate scan |
-| "Are there any secrets in this file?" | hydra-guard (Haiku 4.5) | Pattern scan for credentials |
-| "Check this PR for obvious quality issues" | hydra-guard (Haiku 4.5) | Quality gate on code diff |
-| (auto-invoked after hydra-coder) | hydra-guard (Haiku 4.5) | Auto-guard protocol |
-
-### hydra-git (Haiku 4.5) examples
-| User says | Route to | Why |
-|-----------|----------|-----|
-| "Commit these changes" | hydra-git (Haiku 4.5) | Stage + commit with message |
-| "What's the git status?" | hydra-git (Haiku 4.5) | Run git status, summarize |
-| "Create a branch for this feature" | hydra-git (Haiku 4.5) | Branch creation |
-| "Show me the diff for the last commit" | hydra-git (Haiku 4.5) | Diff inspection |
-| "Stash my changes" | hydra-git (Haiku 4.5) | Git stash |
-| "Are there any merge conflicts?" | hydra-git (Haiku 4.5) | Conflict detection (not resolution) |
-| "Cherry-pick commit abc123" | hydra-git (Haiku 4.5) | Cherry-pick execution |
-
----
-
-## ALWAYS Delegate — Sonnet 4.6 Agents
-
-### hydra-coder (Sonnet 4.6) examples
-| User says | Route to | Why |
-|-----------|----------|-----|
-| "Add a password reset endpoint" | hydra-coder (Sonnet 4.6) | Feature implementation from spec |
-| "Refactor this class to use composition" | hydra-coder (Sonnet 4.6) | Code transformation, clear goal |
-| "Write tests for the auth module" | hydra-coder (Sonnet 4.6) | Test creation requires comprehension |
-| "Fix this TypeError: cannot read property of undefined" | hydra-coder (Sonnet 4.6) | Bug fix with clear error |
-| "Add pagination to the list API" | hydra-coder (Sonnet 4.6) | Standard pattern implementation |
-| "Convert this JavaScript to TypeScript" | hydra-coder (Sonnet 4.6) | Mechanical but needs type reasoning |
-| "Add input validation to the form" | hydra-coder (Sonnet 4.6) | Implementation with business logic |
-
-### hydra-analyst (Sonnet 4.6) examples
-| User says | Route to | Why |
-|-----------|----------|-----|
-| "Review this PR for issues" | hydra-analyst (Sonnet 4.6) | Code review |
-| "Why is this test flaky?" | hydra-analyst (Sonnet 4.6) | Debug analysis with test output |
-| "Are there any security issues in the auth code?" | hydra-analyst (Sonnet 4.6) | Focused security review |
-| "What's causing this memory leak?" | hydra-analyst (Sonnet 4.6) | Performance debugging with symptoms |
-| "How well tested is this module?" | hydra-analyst (Sonnet 4.6) | Coverage analysis |
-
----
-
-## ALWAYS Handle Yourself — Opus 4.6
-
-| User says | Why Opus? |
-|-----------|-----------|
-| "Design the database schema for a multi-tenant SaaS" | Novel architecture decisions |
-| "This works in staging but not production and I can't figure out why" | Subtle debugging, no clear clues |
-| "How should we structure the microservices?" | System design requiring tradeoff analysis |
-| "Optimize this algorithm — it's too slow for large inputs" | Algorithmic insight needed |
-| "We need to support both REST and GraphQL — how?" | Architectural decision with implications |
-| "Review the overall architecture and suggest improvements" | Requires deep holistic understanding |
-| "There's a race condition somewhere in the checkout flow" | Subtle concurrency debugging |
-
----
-
-## Compound Tasks
-
-> **Note**: By default, hydra-scout is pre-dispatched in parallel with task classification for
-> any prompt involving codebase context (see Speculative Pre-Dispatch in SKILL.md). The compound
-> task decompositions below show the explicit Wave 1 scout dispatch for clarity, but in practice,
-> scout may already be returning results by the time Wave 1 launches.
-
-Many real user requests contain multiple subtasks at different tiers. Decompose them:
-
-### Example 1: "Fix the bug in auth.py and add tests"
-1. **hydra-scout (Haiku 4.5)** [BLOCKING] → Find auth.py, read it, understand the context
-2. **hydra-coder (Sonnet 4.6)** [BLOCKING] → Fix the bug and write tests
-3. **hydra-sentinel-scan (Haiku 4.5)** [BLOCKING] → Integration sweep on changed files
-   **hydra-guard (Haiku 4.5)** [BLOCKING] → Security scan on changed files
-   **hydra-runner (Haiku 4.5)** [BLOCKING] → Run the tests to verify
-4. *(If sentinel-scan flags issues)* **hydra-sentinel (Sonnet 4.6)** [BLOCKING] → Deep analysis
-
-### Example 2: "Refactor the API module and update the docs"
-1. **hydra-scout (Haiku 4.5)** [BLOCKING] → Map the current API module structure
-2. **hydra-coder (Sonnet 4.6)** [BLOCKING] → Perform the refactoring
-3. **hydra-sentinel-scan (Haiku 4.5)** [BLOCKING] → Integration sweep on refactored code
-   **hydra-guard (Haiku 4.5)** [BLOCKING] → Security scan on changed files
-   **hydra-runner (Haiku 4.5)** [BLOCKING] → Run tests to verify nothing broke
-   **hydra-scribe (Haiku 4.5)** [NON-BLOCKING] → Update documentation (fire & forget)
-4. *(If sentinel-scan flags issues)* **hydra-sentinel (Sonnet 4.6)** [BLOCKING] → Deep analysis
-
-### Example 3: "Review the codebase and suggest architecture improvements"
-1. **hydra-scout (Haiku 4.5)** → Map the project structure and key files
-2. **hydra-analyst (Sonnet 4.6)** → Review code quality and patterns
-3. **Opus 4.6 (you)** → Synthesize findings into architecture recommendations
-
-### Example 4: "Set up CI/CD for this project"
-1. **hydra-scout (Haiku 4.5)** → Understand the project structure, build system, test framework
-2. **Opus 4.6 (you)** → Design the CI/CD pipeline (architectural decision)
-3. **hydra-coder (Sonnet 4.6)** → Implement the config files
-4. **hydra-runner (Haiku 4.5)** → Validate the configuration
-
----
-
-## Common Misclassifications
-
-These are tasks that look like one tier but are actually another:
-
-| Task | Looks like | Actually | Why |
-|------|-----------|----------|-----|
-| "Add a simple button" | Haiku (simple) | hydra-coder (Sonnet) | Needs to match existing component patterns |
-| "Read the logs and find the error" | Haiku (read) | hydra-analyst (Sonnet) | Log analysis requires reasoning |
-| "Fix the typo in line 42" | hydra-coder (Sonnet) | hydra-scribe (Haiku) | Trivial mechanical change |
-| "Add caching" | hydra-coder (Sonnet) | Opus (handle yourself) | Cache invalidation strategy is hard |
-| "Write a migration to add a column" | hydra-coder (Sonnet) | hydra-scribe (Haiku) | Template-level SQL |
-| "Upgrade React from 17 to 18" | Haiku (simple) | Opus (handle yourself) | Breaking changes need careful analysis |
-
----
-
-## Map-Aware Routing Examples
-
-These examples show how the codebase map changes routing decisions by providing
-risk scores and blast radius data upfront.
-
-### "Fix the bug in auth.ts"
-1. Check map: auth.ts has risk=critical, 12 dependents
-2. hydra-scout → verify map is current (incremental update if needed)
-3. hydra-analyst → diagnose the bug
-4. hydra-coder → implement the fix
-5. hydra-sentinel-scan → map shows blast radius of 12 files, check all 12
-   (without map, would have to grep the entire codebase)
-6. hydra-sentinel → deep analysis (auto-escalated because risk=critical)
-
-### "Add a new utility function"
-1. Check map: new file, risk=low (zero dependents initially)
-2. hydra-coder → write the function
-3. hydra-sentinel-scan → low risk, quick scan, auto-accept if clean
-   (without map, would run the same expensive scan as a critical file)
-
-### "Refactor the database connection module"
-1. Check map: src/db/connection.ts has risk=critical, 15 dependents
-2. Plan execution with full blast radius awareness
-3. Dispatch parallel hydra-coders for each affected file
-4. Sentinel deep analysis is MANDATORY (critical risk)
-
----
-
-## Quick Decision Flowchart
-
-```
-Task arrives
-│
-├── In ALWAYS Delegate table? ── YES ──→ Route to specified agent
-│
-├── In ALWAYS Handle Yourself table? ── YES ──→ Opus handles directly
-│
-└── Neither? → JUDGMENT CALLS:
-    ├── Requires conversation context? ── YES ──→ Opus
-    ├── Haiku/Sonnet can do equally well? ── NO ──→ Opus
-    └── Delegation takes LONGER? ── YES (and truly trivial) ──→ Opus
-        └── Otherwise ──→ Delegate to best-fit agent
-```
-
-## Routing Examples — Mandatory Delegation
-
-These examples show the RIGHT and WRONG way to handle common requests under the mandatory delegation rules.
-
-### 1. "Fix the bug in auth.ts"
-
-**WRONG** (Opus does it all):
-```
-Read auth.ts yourself → find the bug → fix it → run tests
-```
-
-**RIGHT** (mandatory delegation):
-```
-1. hydra-scout → explore auth module, find the bug location
-2. hydra-analyst → analyze the bug, identify root cause
-3. hydra-coder → implement the fix
-4. hydra-sentinel-scan + hydra-guard → verify changes [parallel]
-5. hydra-runner → run tests to confirm fix
-```
-
-### 2. "What's the project structure?"
-
-**WRONG** (Opus runs find/ls itself):
-```
-Run `find . -type f` or `ls -R` yourself
-```
-
-**RIGHT** (mandatory delegation):
-```
-1. hydra-scout → map project structure, report back
-```
-
-### 3. "Run the tests"
-
-**WRONG** (Opus runs npm test itself):
-```
-Run `npm test` yourself
-```
-
-**RIGHT** (mandatory delegation):
-```
-1. hydra-runner → execute test suite, report results
-```
-
-### 4. "Commit these changes"
-
-**WRONG** (Opus runs git add/commit itself):
-```
-Run `git add . && git commit -m "..."` yourself
-```
-
-**RIGHT** (mandatory delegation):
-```
-1. hydra-git → stage and commit with well-crafted message
-```
-
-### 5. "This function is slow, figure out why"
-
-**RIGHT** (Opus orchestrates, delegates execution):
-```
-1. hydra-analyst → profile and diagnose the performance issue
-2. YOU → decide the optimization approach (architectural judgment)
-3. hydra-coder → implement the optimization
-4. hydra-runner → benchmark before and after
-```
-
-### 6. "Redesign the auth module to use OAuth2"
-
-**RIGHT** (Opus architects, delegates implementation):
-```
-1. hydra-scout → map current auth module structure [parallel with Step 2]
-2. hydra-scout → research OAuth2 patterns in codebase [parallel with Step 1]
-3. YOU → design the new architecture (this is YOUR job)
-4. hydra-coder #1 → implement OAuth2 provider [parallel with Steps 5-6]
-5. hydra-coder #2 → implement token management [parallel with Steps 4, 6]
-6. hydra-coder #3 → update route handlers [parallel with Steps 4-5]
-7. hydra-sentinel-scan → integration sweep on all changes
-8. hydra-runner → run full test suite
-```
-
-### 7. "Quick — add a console.log to line 5"
-
-**ACCEPTABLE** (trivial <5s edit — uses overhead budget):
-```
-YOU → add the console.log directly (counts as 1 of max 2-3 exceptions per session)
-```
-
-## Sentinel Routing
-
-Sentinel is NOT manually routed. It triggers AUTOMATICALLY after code changes:
-
-| Event | Triggers |
-|-------|----------|
-| hydra-coder writes/edits code | → sentinel-scan (always) |
-| hydra-analyst suggests code changes | → sentinel-scan (always) |
-| Orchestrator makes direct code changes | → sentinel-scan (always) |
-| sentinel-scan finds issues | → sentinel deep analysis |
-| sentinel-scan is clean | → nothing (done) |
-| hydra-scribe writes docs | → nothing (skip) |
-| hydra-git commits | → nothing (skip) |
-| hydra-runner runs tests | → nothing (skip) |
+# Routing Guide — Mandatory Delegation Examples
+
+This reference provides concrete examples to see mandatory delegation in action.
+Read this when you need to see how delegation rules apply to real tasks.
+
+## Table of Contents
+1. [ALWAYS Delegate — Haiku 4.5 Agents](#always-delegate-haiku)
+2. [ALWAYS Delegate — Sonnet 4.6 Agents](#always-delegate-sonnet)
+3. [ALWAYS Handle Yourself — Opus 4.6](#always-handle-yourself-opus)
+4. [Compound Task Decomposition](#compound-tasks)
+5. [Common Misclassifications](#common-misclassifications)
+
+---
+
+## ALWAYS Delegate — Haiku 4.5 Agents
+
+### hydra-scout (Haiku 4.5) examples
+| User says | Route to | Why |
+|-----------|----------|-----|
+| "What framework is this project using?" | hydra-scout (Haiku 4.5) | Read package.json/config files |
+| "Find where the login endpoint is defined" | hydra-scout (Haiku 4.5) | Grep for routes/endpoints |
+| "How many files are in the src directory?" | hydra-scout (Haiku 4.5) | Glob + count |
+| "Show me the database schema" | hydra-scout (Haiku 4.5) | Find and read schema/migration files |
+| "What does the User model look like?" | hydra-scout (Haiku 4.5) | Find and read model definition |
+| "Is there a rate limiter in this project?" | hydra-scout (Haiku 4.5) | Grep for rate limit patterns |
+| "What version of React are we using?" | hydra-scout (Haiku 4.5) | Read package.json |
+
+### hydra-runner (Haiku 4.5) examples
+| User says | Route to | Why |
+|-----------|----------|-----|
+| "Run the tests" | hydra-runner (Haiku 4.5) | Execute test command, report results |
+| "Does the build pass?" | hydra-runner (Haiku 4.5) | Run build, report success/failure |
+| "Check if there are any lint errors" | hydra-runner (Haiku 4.5) | Run linter, report findings |
+| "What's the git status?" | hydra-runner (Haiku 4.5) | Run git status/diff |
+| "Run the migration" | hydra-runner (Haiku 4.5) | Execute migration command |
+| "Check if the server starts" | hydra-runner (Haiku 4.5) | Start server, check for errors |
+
+### hydra-scribe (Haiku 4.5) examples
+| User says | Route to | Why |
+|-----------|----------|-----|
+| "Add docstrings to this file" | hydra-scribe (Haiku 4.5) | Read code, write docstrings |
+| "Update the README with the new API endpoints" | hydra-scribe (Haiku 4.5) | Descriptive writing from code |
+| "Write a changelog entry for these changes" | hydra-scribe (Haiku 4.5) | Summarize changes |
+| "Add comments explaining this function" | hydra-scribe (Haiku 4.5) | Read and annotate |
+
+### hydra-guard (Haiku 4.5) examples
+| User says | Route to | Why |
+|-----------|----------|-----|
+| "Scan my changes for security issues" | hydra-guard (Haiku 4.5) | Security gate scan |
+| "Are there any secrets in this file?" | hydra-guard (Haiku 4.5) | Pattern scan for credentials |
+| "Check this PR for obvious quality issues" | hydra-guard (Haiku 4.5) | Quality gate on code diff |
+| (auto-invoked after hydra-coder) | hydra-guard (Haiku 4.5) | Auto-guard protocol |
+
+### hydra-git (Haiku 4.5) examples
+| User says | Route to | Why |
+|-----------|----------|-----|
+| "Commit these changes" | hydra-git (Haiku 4.5) | Stage + commit with message |
+| "What's the git status?" | hydra-git (Haiku 4.5) | Run git status, summarize |
+| "Create a branch for this feature" | hydra-git (Haiku 4.5) | Branch creation |
+| "Show me the diff for the last commit" | hydra-git (Haiku 4.5) | Diff inspection |
+| "Stash my changes" | hydra-git (Haiku 4.5) | Git stash |
+| "Are there any merge conflicts?" | hydra-git (Haiku 4.5) | Conflict detection (not resolution) |
+| "Cherry-pick commit abc123" | hydra-git (Haiku 4.5) | Cherry-pick execution |
+
+---
+
+## ALWAYS Delegate — Sonnet 4.6 Agents
+
+### hydra-coder (Sonnet 4.6) examples
+| User says | Route to | Why |
+|-----------|----------|-----|
+| "Add a password reset endpoint" | hydra-coder (Sonnet 4.6) | Feature implementation from spec |
+| "Refactor this class to use composition" | hydra-coder (Sonnet 4.6) | Code transformation, clear goal |
+| "Write tests for the auth module" | hydra-coder (Sonnet 4.6) | Test creation requires comprehension |
+| "Fix this TypeError: cannot read property of undefined" | hydra-coder (Sonnet 4.6) | Bug fix with clear error |
+| "Add pagination to the list API" | hydra-coder (Sonnet 4.6) | Standard pattern implementation |
+| "Convert this JavaScript to TypeScript" | hydra-coder (Sonnet 4.6) | Mechanical but needs type reasoning |
+| "Add input validation to the form" | hydra-coder (Sonnet 4.6) | Implementation with business logic |
+
+### hydra-analyst (Sonnet 4.6) examples
+| User says | Route to | Why |
+|-----------|----------|-----|
+| "Review this PR for issues" | hydra-analyst (Sonnet 4.6) | Code review |
+| "Why is this test flaky?" | hydra-analyst (Sonnet 4.6) | Debug analysis with test output |
+| "Are there any security issues in the auth code?" | hydra-analyst (Sonnet 4.6) | Focused security review |
+| "What's causing this memory leak?" | hydra-analyst (Sonnet 4.6) | Performance debugging with symptoms |
+| "How well tested is this module?" | hydra-analyst (Sonnet 4.6) | Coverage analysis |
+
+---
+
+## ALWAYS Handle Yourself — Opus 4.6
+
+| User says | Why Opus? |
+|-----------|-----------|
+| "Design the database schema for a multi-tenant SaaS" | Novel architecture decisions |
+| "This works in staging but not production and I can't figure out why" | Subtle debugging, no clear clues |
+| "How should we structure the microservices?" | System design requiring tradeoff analysis |
+| "Optimize this algorithm — it's too slow for large inputs" | Algorithmic insight needed |
+| "We need to support both REST and GraphQL — how?" | Architectural decision with implications |
+| "Review the overall architecture and suggest improvements" | Requires deep holistic understanding |
+| "There's a race condition somewhere in the checkout flow" | Subtle concurrency debugging |
+
+---
+
+## Compound Tasks
+
+> **Note**: By default, hydra-scout is pre-dispatched in parallel with task classification for
+> any prompt involving codebase context (see Speculative Pre-Dispatch in SKILL.md). The compound
+> task decompositions below show the explicit Wave 1 scout dispatch for clarity, but in practice,
+> scout may already be returning results by the time Wave 1 launches.
+
+Many real user requests contain multiple subtasks at different tiers. Decompose them:
+
+### Example 1: "Fix the bug in auth.py and add tests"
+1. **hydra-scout (Haiku 4.5)** [BLOCKING] → Find auth.py, read it, understand the context
+2. **hydra-coder (Sonnet 4.6)** [BLOCKING] → Fix the bug and write tests
+3. **hydra-sentinel-scan (Haiku 4.5)** [BLOCKING] → Integration sweep on changed files
+   **hydra-guard (Haiku 4.5)** [BLOCKING] → Security scan on changed files
+   **hydra-runner (Haiku 4.5)** [BLOCKING] → Run the tests to verify
+4. *(If sentinel-scan flags issues)* **hydra-sentinel (Sonnet 4.6)** [BLOCKING] → Deep analysis
+
+### Example 2: "Refactor the API module and update the docs"
+1. **hydra-scout (Haiku 4.5)** [BLOCKING] → Map the current API module structure
+2. **hydra-coder (Sonnet 4.6)** [BLOCKING] → Perform the refactoring
+3. **hydra-sentinel-scan (Haiku 4.5)** [BLOCKING] → Integration sweep on refactored code
+   **hydra-guard (Haiku 4.5)** [BLOCKING] → Security scan on changed files
+   **hydra-runner (Haiku 4.5)** [BLOCKING] → Run tests to verify nothing broke
+   **hydra-scribe (Haiku 4.5)** [NON-BLOCKING] → Update documentation (fire & forget)
+4. *(If sentinel-scan flags issues)* **hydra-sentinel (Sonnet 4.6)** [BLOCKING] → Deep analysis
+
+### Example 3: "Review the codebase and suggest architecture improvements"
+1. **hydra-scout (Haiku 4.5)** → Map the project structure and key files
+2. **hydra-analyst (Sonnet 4.6)** → Review code quality and patterns
+3. **Opus 4.6 (you)** → Synthesize findings into architecture recommendations
+
+### Example 4: "Set up CI/CD for this project"
+1. **hydra-scout (Haiku 4.5)** → Understand the project structure, build system, test framework
+2. **Opus 4.6 (you)** → Design the CI/CD pipeline (architectural decision)
+3. **hydra-coder (Sonnet 4.6)** → Implement the config files
+4. **hydra-runner (Haiku 4.5)** → Validate the configuration
+
+---
+
+## Common Misclassifications
+
+These are tasks that look like one tier but are actually another:
+
+| Task | Looks like | Actually | Why |
+|------|-----------|----------|-----|
+| "Add a simple button" | Haiku (simple) | hydra-coder (Sonnet) | Needs to match existing component patterns |
+| "Read the logs and find the error" | Haiku (read) | hydra-analyst (Sonnet) | Log analysis requires reasoning |
+| "Fix the typo in line 42" | hydra-coder (Sonnet) | hydra-scribe (Haiku) | Trivial mechanical change |
+| "Add caching" | hydra-coder (Sonnet) | Opus (handle yourself) | Cache invalidation strategy is hard |
+| "Write a migration to add a column" | hydra-coder (Sonnet) | hydra-scribe (Haiku) | Template-level SQL |
+| "Upgrade React from 17 to 18" | Haiku (simple) | Opus (handle yourself) | Breaking changes need careful analysis |
+
+---
+
+## Map-Aware Routing Examples
+
+These examples show how the codebase map changes routing decisions by providing
+risk scores and blast radius data upfront.
+
+### "Fix the bug in auth.ts"
+1. Check map: auth.ts has risk=critical, 12 dependents
+2. hydra-scout → verify map is current (incremental update if needed)
+3. hydra-analyst → diagnose the bug
+4. hydra-coder → implement the fix
+5. hydra-sentinel-scan → map shows blast radius of 12 files, check all 12
+   (without map, would have to grep the entire codebase)
+6. hydra-sentinel → deep analysis (auto-escalated because risk=critical)
+
+### "Add a new utility function"
+1. Check map: new file, risk=low (zero dependents initially)
+2. hydra-coder → write the function
+3. hydra-sentinel-scan → low risk, quick scan, auto-accept if clean
+   (without map, would run the same expensive scan as a critical file)
+
+### "Refactor the database connection module"
+1. Check map: src/db/connection.ts has risk=critical, 15 dependents
+2. Plan execution with full blast radius awareness
+3. Dispatch parallel hydra-coders for each affected file
+4. Sentinel deep analysis is MANDATORY (critical risk)
+
+---
+
+## Quick Decision Flowchart
+
+```
+Task arrives
+│
+├── In ALWAYS Delegate table? ── YES ──→ Route to specified agent
+│
+├── In ALWAYS Handle Yourself table? ── YES ──→ Opus handles directly
+│
+└── Neither? → JUDGMENT CALLS:
+    ├── Requires conversation context? ── YES ──→ Opus
+    ├── Haiku/Sonnet can do equally well? ── NO ──→ Opus
+    └── Delegation takes LONGER? ── YES (and truly trivial) ──→ Opus
+        └── Otherwise ──→ Delegate to best-fit agent
+```
+
+## Routing Examples — Mandatory Delegation
+
+These examples show the RIGHT and WRONG way to handle common requests under the mandatory delegation rules.
+
+### 1. "Fix the bug in auth.ts"
+
+**WRONG** (Opus does it all):
+```
+Read auth.ts yourself → find the bug → fix it → run tests
+```
+
+**RIGHT** (mandatory delegation):
+```
+1. hydra-scout → explore auth module, find the bug location
+2. hydra-analyst → analyze the bug, identify root cause
+3. hydra-coder → implement the fix
+4. hydra-sentinel-scan + hydra-guard → verify changes [parallel]
+5. hydra-runner → run tests to confirm fix
+```
+
+### 2. "What's the project structure?"
+
+**WRONG** (Opus runs find/ls itself):
+```
+Run `find . -type f` or `ls -R` yourself
+```
+
+**RIGHT** (mandatory delegation):
+```
+1. hydra-scout → map project structure, report back
+```
+
+### 3. "Run the tests"
+
+**WRONG** (Opus runs npm test itself):
+```
+Run `npm test` yourself
+```
+
+**RIGHT** (mandatory delegation):
+```
+1. hydra-runner → execute test suite, report results
+```
+
+### 4. "Commit these changes"
+
+**WRONG** (Opus runs git add/commit itself):
+```
+Run `git add . && git commit -m "..."` yourself
+```
+
+**RIGHT** (mandatory delegation):
+```
+1. hydra-git → stage and commit with well-crafted message
+```
+
+### 5. "This function is slow, figure out why"
+
+**RIGHT** (Opus orchestrates, delegates execution):
+```
+1. hydra-analyst → profile and diagnose the performance issue
+2. YOU → decide the optimization approach (architectural judgment)
+3. hydra-coder → implement the optimization
+4. hydra-runner → benchmark before and after
+```
+
+### 6. "Redesign the auth module to use OAuth2"
+
+**RIGHT** (Opus architects, delegates implementation):
+```
+1. hydra-scout → map current auth module structure [parallel with Step 2]
+2. hydra-scout → research OAuth2 patterns in codebase [parallel with Step 1]
+3. YOU → design the new architecture (this is YOUR job)
+4. hydra-coder #1 → implement OAuth2 provider [parallel with Steps 5-6]
+5. hydra-coder #2 → implement token management [parallel with Steps 4, 6]
+6. hydra-coder #3 → update route handlers [parallel with Steps 4-5]
+7. hydra-sentinel-scan → integration sweep on all changes
+8. hydra-runner → run full test suite
+```
+
+### 7. "Quick — add a console.log to line 5"
+
+**ACCEPTABLE** (trivial <5s edit — uses overhead budget):
+```
+YOU → add the console.log directly (counts as 1 of max 2-3 exceptions per session)
+```
+
+## Sentinel Routing
+
+Sentinel is NOT manually routed. It triggers AUTOMATICALLY after code changes:
+
+| Event | Triggers |
+|-------|----------|
+| hydra-coder writes/edits code | → sentinel-scan (always) |
+| hydra-analyst suggests code changes | → sentinel-scan (always) |
+| Orchestrator makes direct code changes | → sentinel-scan (always) |
+| sentinel-scan finds issues | → sentinel deep analysis |
+| sentinel-scan is clean | → nothing (done) |
+| hydra-scribe writes docs | → nothing (skip) |
+| hydra-git commits | → nothing (skip) |
+| hydra-runner runs tests | → nothing (skip) |