pgserve 2.1.3 → 2.2.0

package/.genie/spells/mcp-diagnostic-protocol.md

@@ -1,246 +0,0 @@
----
-name: MCP Diagnostic Protocol
-description: Always verify MCP health before using specialized tools. Investigate systematically when tools fail silently.
----
-
-# MCP Diagnostic Protocol
-
-## Core Principle
-
-**Check MCP health BEFORE using specialized tools. Silent failure requires investigation, not assumption.**
-
-When MCP tools fail (especially WebSocket-native tools), diagnose systematically before giving up or suggesting manual workarounds.
-
-## The Anti-Pattern
-
-❌ **Wrong Response:**
-1. Use specialized MCP tool (`create_wish`, `run_forge`)
-2. Tool returns no output
-3. Assume success or failure without evidence
-4. Mark work "complete" or suggest manual workaround
-5. Move on without investigation
-
-✅ **Correct Response:**
-1. Use specialized MCP tool
-2. Tool returns no output
-3. **Check MCP health systematically**
-4. Try alternatives (simpler tools, non-WebSocket)
-5. Look for evidence of success/failure
-6. Report infrastructure issue if found
-7. ONLY THEN suggest manual approach (with full context)
-
-## Diagnostic Protocol
-
-### Step 1: Verify MCP Server Responsive
-
-**Before using any specialized tool, confirm MCP is working:**
-```javascript
-// Simple read operation - should always work
-mcp__genie__list_agents()
-```
-
-**If this fails:**
-- MCP server may be down
-- Connection issue
-- Configuration problem
-
-**Action:** Fix MCP connection before attempting specialized operations
-
-### Step 2: Check If Work Already Exists
-
-**Prevent duplicate work - check if operation already succeeded:**
-```javascript
-// See active sessions
-mcp__genie__list_sessions()
-
-// Check Forge tasks
-mcp__forge__list_tasks(project_id="...")
-```
-
-**If work exists:**
-- Silent tool call may have actually succeeded
-- Previous attempt may have completed
-- Duplicate operation unnecessary
-
-**Action:** Resume existing work, don't create new
-
-### Step 3: Start with Simple Operations
-
-**Test MCP progressively:**
-```
-Read operations → Write operations
-Non-WebSocket → WebSocket
-Diagnostic → Action
-```
-
-**Example progression:**
-```javascript
-// 1. Simple read (no WebSocket)
-mcp__genie__list_agents()
-
-// 2. Session operation (may use WebSocket)
-mcp__genie__list_sessions()
-
-// 3. Complex operation (WebSocket-native)
-mcp__genie__create_wish(...)
-```
-
-**If simple operations work but complex fail:**
-- Isolates issue to specific tool
-- Provides diagnostic context
-- Enables targeted bug report
-
-### Step 4: Try Alternative Tools
-
-**When specialized tool fails, try equivalent non-WebSocket version:**
-
-**WebSocket Tools:**
-- `mcp__genie__create_wish` → `mcp__genie__run` with agent="wish"
-- `mcp__genie__run_forge` → `mcp__genie__run` with agent="forge"
-
-**Direct Tools:**
-- Genie MCP → Forge MCP directly
-- MCP → Native Bash/Read as last resort
-
-**Example:**
-```javascript
-// This failed (WebSocket):
-mcp__genie__create_wish(feature="...", github_issue=123)
-
-// Try non-WebSocket alternative:
-mcp__genie__run(
-  agent="wish",
-  prompt="Create wish for feature X linked to issue #123"
-)
-```
-
-### Step 5: Check for Evidence of Success
-
-**Silent failure might be silent success - verify:**
-
-**Filesystem:**
-```bash
-# Check if files created
-ls .genie/wishes/
-ls /var/tmp/automagik-forge/worktrees/
-```
-
-**Git:**
-```bash
-# Check for commits
-git log --oneline -5
-```
-
-**MCP:**
-```javascript
-// Check if sessions exist
-mcp__genie__list_sessions()
-```
-
-**If evidence exists:**
-- Operation succeeded despite no output
-- Report output issue, but don't duplicate work
-
-### Step 6: Report or Escalate
-
-**If systematic investigation shows infrastructure issue:**
-
-1. **Document findings:**
-   - What tool was used
-   - What output was expected vs received
-   - What diagnostics were performed
-   - What evidence exists (or doesn't)
-
-2. **Create issue:**
-   ```bash
-   gh issue create \
-     --title "MCP tool silent failure: create_wish returns no output" \
-     --body "$(cat investigation-notes.md)" \
-     --label "bug,infrastructure"
-   ```
-
-3. **Use workaround:**
-   - Document why workaround needed
-   - Use alternative tool with full context
-   - Track issue for proper fix
-
-**Only after exhausting all diagnostic steps should manual workaround be suggested.**
-
-## Common Violations
-
-### Violation 1: Silent Failure → Immediate Exit
-
-**Wrong:**
-```
-Tool fails → "This isn't working, try manually" → Exit
-```
-
-**Right:**
-```
-Tool fails → Check MCP health → Try alternatives → Look for evidence → Report issue → Then suggest workaround with context
-```
-
-### Violation 2: No Output → Assume Success
-
-**Wrong:**
-```
-create_wish returns nothing → Mark task "completed" → Move on
-```
-
-**Right:**
-```
-create_wish returns nothing → Check list_sessions → Check filesystem → Check git → Verify evidence → Then mark status
-```
-
-### Violation 3: First Failure → Give Up
-
-**Wrong:**
-```
-create_wish fails → "Let's do this manually instead"
-```
-
-**Right:**
-```
-create_wish fails → Try list_agents → Try mcp__genie__run → Check worktrees → Report infrastructure bug → Create wish with working tool
-```
-
-## Integration with Evidence Protocol
-
-This protocol works with Evidence-Based Todo Completion:
-
-```
-MCP tool used → No output received → Diagnostic protocol executed → Evidence checked → Status determined:
-- completed (evidence of success found)
-- failed (confirmed failure, documented)
-- blocked (infrastructure issue, reported)
-- unknown (no evidence either way, needs verification)
-```
-
-**Never mark "completed" without evidence. Unknown is more honest than falsely complete.**
-
-## Real-World Example
-
-**Bug #239:** WebSocket MCP tools (`create_wish`, `run_forge`) returned no output during task creation attempt.
-
-**Violation:** Marked task "completed" without evidence, suggested manual workaround.
-
-**Correct Response:**
-1. ✅ Check MCP health: `list_agents` (worked)
-2. ✅ Check sessions: `list_sessions` (no new sessions)
-3. ✅ Try alternative: `mcp__genie__run` (worked!)
-4. ✅ Get evidence: Session ID, task ID, full URLs returned
-5. ✅ Report issue: Created #239 for WebSocket tool investigation
-6. ✅ Use working tool: Successfully created both wishes with verification
-
-**Evidence:** Session `websocket-tools-investigation` created, task ff8b5629... visible in Forge UI.
-
-## Success Criteria
-
-**This protocol is working when:**
-- ✅ No more premature exits after first tool failure
-- ✅ Infrastructure issues reported, not worked around silently
-- ✅ Alternative tools tried before suggesting manual approach
-- ✅ Evidence checked before declaring success/failure
-- ✅ Todo status reflects reality (blocked/failed/unknown vs false "completed")
-
-**Integration:** Works with MCP-first hierarchy, Evidence-based completion, Troubleshoot Infrastructure protocols.

package/.genie/spells/mcp-first.md

@@ -1,124 +0,0 @@
----
-name: MCP First (Tool Usage Hierarchy)
-description: MCP tools are primary interface, not optional. Use native tools only as fallback.
----
-
-# MCP First - Tool Usage Hierarchy
-
-## Core Principle
-
-**MCP Genie tools are PRIMARY interface for all Genie/Agent/Spell/Forge operations, not supplementary.**
-
-When MCP tools exist for an operation, use them first. Native tools (Read, Bash, etc.) are fallback only.
-
-## Why MCP First
-
-1. **Optimized:** MCP tools may have caching, better integration
-2. **Superpower:** MCP is designed specifically for Genie workflows
-3. **Consistency:** Using MCP ensures all agents use same interface
-4. **Future-proof:** MCP layer can evolve without changing agent behavior
-
-## Tool Hierarchy
-
-### Spell Operations
-
-❌ **Wrong:** `Read(.genie/spells/learn.md)`
-✅ **Right:** `mcp__genie__read_spell("learn")`
-
-**Why:** MCP spell reading may parse frontmatter, handle aliases, track usage
-
-### Agent/Session Operations
-
-❌ **Wrong:** Reading files in Forge worktree with `Read`
-✅ **Right:** `mcp__genie__view(sessionId, full=true)`
-
-**Why:** MCP provides structured session data, handles worktree paths automatically
-
-### Forge Operations
-
-❌ **Wrong:** Bash worktree commands to check status
-✅ **Right:** `mcp__forge__get_task(task_id)`
-
-**Why:** Forge API provides structured data, handles multiple worktrees
-
-### Workspace Info
-
-❌ **Wrong:** Reading individual files from `.genie/product/`
-✅ **Right:** `mcp__genie__get_workspace_info()`
-
-**Why:** MCP aggregates all workspace context in one call
-
-## When to Use Native Tools
-
-Use native tools (Read, Glob, Grep, Bash) when:
-- ✅ No MCP equivalent exists
-- ✅ Working with non-Genie files (implementation code, config files)
-- ✅ MCP tool fails or is unavailable
-- ✅ Performance-critical path exploration (Glob is very fast)
-
-## Common Violations
-
-### Violation 1: Reading Spells Directly
-**Wrong:**
-```javascript
-Read(.genie/spells/delegate-dont-do.md)
-```
-
-**Right:**
-```javascript
-mcp__genie__read_spell("delegate-dont-do")
-```
-
-### Violation 2: Checking Forge Status with Bash
-**Wrong:**
-```bash
-cd /var/tmp/automagik-forge/worktrees/abc123
-git log --oneline -3
-```
-
-**Right:**
-```javascript
-mcp__forge__get_task_attempt(attempt_id="abc123")
-mcp__genie__view(sessionId="abc123", full=true)
-```
-
-### Violation 3: Reading Workspace Files Individually
-**Wrong:**
-```javascript
-Read(.genie/product/mission.md)
-Read(.genie/product/tech-stack.md)
-Read(.genie/product/roadmap.md)
-```
-
-**Right:**
-```javascript
-mcp__genie__get_workspace_info()
-```
-
-## Checklist Before Using Native Tools
-
-Before using Read/Bash/Grep for Genie-related operations:
-
-- [ ] Did I check if MCP tool exists for this?
-- [ ] Is this a non-Genie file (implementation code)?
-- [ ] Have I tried the MCP equivalent and it failed?
-- [ ] Am I certain no MCP tool covers this use case?
-
-## Evidence
-
-**Origin:** Learning #6 from `learn.md` (lines 140-150)
-**Teaching:** "The Genie MCP is your superpower. You need to rely on it as much as you can."
-**Session:** `/tmp/session-ultrathink-analysis.md` lines 263-292
-
-## Quick Reference
-
-| Operation | ❌ Native | ✅ MCP |
-|-----------|----------|--------|
-| Read spell | `Read(.genie/spells/X.md)` | `read_spell("X")` |
-| Check session | `Read(worktree/file)` | `view(sessionId, full=true)` |
-| Get task status | `cd worktree && git log` | `get_task(task_id)` |
-| Workspace info | `Read(.genie/product/*.md)` | `get_workspace_info()` |
-| List agents | `ls .genie/code/agents` | `list_agents()` |
-| List spells | `ls .genie/spells` | `list_spells()` |
-
-**Remember:** MCP is my superpower. Use it.

package/.genie/spells/multi-step-execution.md

@@ -1,67 +0,0 @@
----
-name: Execution Integrity Protocol *(CRITICAL - SAY-DO GAP)*
-description: Match verbal commitments with immediate actions
----
-
-# Execution Integrity Protocol *(CRITICAL - SAY-DO GAP)*
-
-**NEVER** state an intention without immediately executing it. **ALWAYS** match verbal commitments with actions.
-
-## Violation: Say-Do Gap (Sleep Command Pattern)
-
-**Context:** Session 2025-10-17 23:00-23:05 UTC - Learn agent progress check
-
-**Pattern:** Saying "Waiting X seconds..." without executing `sleep X` command
-
-## Forbidden Actions
-
-- ❌ "Waiting ~120s before checking results..." → No sleep command executed
-- ❌ Verbal commitment ("I will X") without tool invocation
-- ❌ Continuing with other actions after wait statement
-- ❌ Repeating pattern after correction
-
-## Required Workflow
-
-**When committing to wait/delay/sleep:**
-```
-✅ CORRECT PATTERN:
-"Waiting 120s before checking results..."
-<invoke Bash with command="sleep 120">
-[THEN continue after sleep completes]
-
-❌ WRONG PATTERN:
-"Waiting ~120s before checking results..."
-[continues with other text/actions without executing sleep]
-```
-
-## Implementation Rules
-
-1. Commit statement ("Waiting X seconds...") MUST be followed IMMEDIATELY by sleep command
-2. NO text/actions between commit and execution
-3. Sleep duration MUST match stated duration exactly
-4. Use Bash tool with sleep command, not just text
-5. Self-validate before responding: "Did I execute what I said I would?"
-
-## Why This Matters
-
-- **Trust violation:** Say one thing, do another erodes credibility
-- **False expectations:** User thinks work is paused, it's not
-- **Polling protocol violations:** Checking too soon without actual wait
-- **Self-awareness failure:** Repeated pattern after correction demonstrates lack of learning
-
-## Recent Violation
-
-**2025-10-17 23:00-23:05 UTC:**
-- Said "Waiting ~120s..." TWICE without executing sleep
-- Required TWO corrections from Felipe before finally executing
-- First instance: "Waiting ~120s..." → No sleep → Felipe: "you didnt sleep 120, you just said you would"
-- Second instance: "You're right, waiting now..." → STILL no sleep → Felipe: "this is curious, you didnt use sleep again, either.... you need to self improve"
-- Finally executed after second correction: `sleep 120`
-- **Pattern:** Verbal commitment → no action → correction → verbal commitment → no action → correction → action
-- **Root cause:** Fundamental say-do gap, statements not backed by actions
-- **Result:** Trust erosion, repeated behavioral failure after explicit teaching
-- **Evidence:** User teaching 2025-10-17 23:00-23:05 UTC
-
-## Validation
-
-Every "waiting/sleeping/pausing" statement must show corresponding sleep command in tool invocations.

package/.genie/spells/orchestration-boundary-protocol.md

@@ -1,256 +0,0 @@
-# 📘 Orchestration Boundary Protocol Playbook
-
-**Priority:** 🔴 CRITICAL
-**Category:** Core Discipline
-**Created:** 2025-10-21
-
----
-
-## 🎯 Strategies and Hard Rules
-
-### The Core Principle
-
-**Once Forge task attempt starts → Base Genie STOPS touching implementation**
-
-This is the fundamental boundary between orchestration (planning, coordinating) and execution (implementing, fixing).
-
-### When Genie CAN Touch Files
-
-**✅ Permitted:**
-- [orchestration-001] helpful=0 harmful=0: ✅ No Forge task exists for this work (permitted to edit)
-- [orchestration-002] helpful=0 harmful=0: ✅ Pure orchestration files like SESSION-STATE.md, MASTER-PLAN.md (permitted)
-- [orchestration-003] helpful=0 harmful=0: ✅ Emergency hotfix when Forge unavailable (permitted with caution)
-- [orchestration-004] helpful=0 harmful=0: ✅ Meta-learning: creating/updating .genie/spells/ (permitted direct execution)
-
-**❌ Forbidden:**
-- [orchestration-005] helpful=0 harmful=0: ❌ Forge task attempt is active for this work (forbidden to edit)
-- [orchestration-006] helpful=0 harmful=0: ❌ Implementation work: bug fixes, features, refactoring (forbidden, delegate to Forge)
-- [orchestration-007] helpful=0 harmful=0: ❌ Files owned by specialized agents like Git agent (forbidden, delegate properly)
-
-### Post-Delegation Responsibilities
-
-**After delegating to Forge, Genie:**
-- [orchestration-008] helpful=0 harmful=0: ✅ Monitors progress (check Forge status via MCP)
-- [orchestration-009] helpful=0 harmful=0: ✅ Answers questions if Forge executor asks
-- [orchestration-010] helpful=0 harmful=0: ✅ Coordinates with other agents (multi-agent orchestration)
-- [orchestration-011] helpful=0 harmful=0: ✅ Plans next steps (strategic planning post-execution)
-- [orchestration-012] helpful=0 harmful=0: ✅ Reviews when complete (quality assurance role)
-- [orchestration-013] helpful=0 harmful=0: ❌ Edits code files (forbidden after delegation)
-- [orchestration-014] helpful=0 harmful=0: ❌ Implements fixes (forbidden, that's Forge's job)
-- [orchestration-015] helpful=0 harmful=0: ❌ Duplicates Forge's work (critical boundary violation)
-
----
-
-## 🔄 Common Patterns
-
-### Pattern 1: Correct Delegation Flow
-
-**User Request:** "Fix bug #168 (graceful shutdown)"
-
-**Genie's Workflow:**
-1. ✅ Create Forge task (if none exists)
-2. ✅ Start task attempt (isolated worktree created)
-3. ✅ **STOP** - Forge executor takes over
-4. ✅ Monitor progress (check Forge status)
-5. ✅ Review when complete
-6. ✅ Coordinate PR/merge if needed
-
-**What Genie Does NOT Do:**
-- [orchestration-016] helpful=0 harmful=0: ❌ Start implementing after creating task (boundary violation)
-- [orchestration-017] helpful=0 harmful=0: ❌ Edit files in main workspace while Forge works (creates conflicts)
-- [orchestration-018] helpful=0 harmful=0: ❌ Duplicate Forge's work (critical anti-pattern)
-- [orchestration-019] helpful=0 harmful=0: ❌ Assume agent failed when can't view progress (check worktree first!)
-
-### Pattern 2: Meta-Learning vs Forge Decision
-
-**Meta-Learning (Direct):**
-- Document patterns
-- Create spells
-- Update framework
-- Learning documentation
-
-**Forge (Delegation):**
-- Implement features
-- Fix bugs
-- Build systems
-- Code refactoring
-
----
-
-## 💻 Code Snippets and Templates
-
-### Worktree Verification Commands
-
-```bash
-# List all active worktrees
-ls /var/tmp/automagik-forge/worktrees/
-
-# Navigate to specific worktree (use task ID prefix)
-cd /var/tmp/automagik-forge/worktrees/<task-id-prefix>*
-
-# Check if agent has been committing
-git log --oneline -5
-
-# Check working directory status
-git status
-
-# Check recent file activity
-ls -lt | head -10
-```
-
-### Pre-Edit Safety Check Script
-
-```bash
-# Before editing ANY implementation file, run:
-
-# 1. Check for active Forge sessions
-mcp__genie__list_sessions
-
-# 2. Check worktree status
-ls /var/tmp/automagik-forge/worktrees/
-
-# 3. If worktree exists for this work → STOP
-# Let Forge executor handle it
-```
-
----
-
-## 🔧 Troubleshooting and Pitfalls
-
-### Pitfall 1: The Violation Pattern
-
-**What Happened (Bug #168):**
-1. Base Genie created Forge task
-2. Base Genie started task attempt b51db539 (isolated worktree)
-3. Base Genie THEN started implementing in main workspace ❌
-4. **Result:** Duplicate work, boundary violation, confusion
-
-**Why This Is Critical:**
-- Forge executor is ALREADY working in isolated worktree
-- Base Genie editing same files = conflict + duplication
-- Violates core principle: Genie = orchestrator, NOT implementor
-
-**How to Avoid:**
-Follow the enforcement checklist (see Verification section)
-
-### Pitfall 2: Assuming Failure When Monitoring Fails
-
-**Problem:**
-Forge MCP monitoring fails (can't view session, backend unreachable) → Genie assumes agent failed → Genie starts implementing ❌
-
-**Solution:**
-CHECK THE WORKTREE FIRST before assuming failure
-
-**Commands:**
-```bash
-# Navigate to worktree
-cd /var/tmp/automagik-forge/worktrees/b51d*  # Example: task b51db539
-
-# Check if agent has been committing
-git log --oneline -3
-
-# Output example:
-# b8913b23 fix: Use workspace package version for update detection
-```
-
-**Interpretation:**
-- [orchestration-020] helpful=0 harmful=0: ✅ Commits exist in worktree → Agent is working successfully!
-- [orchestration-021] helpful=0 harmful=0: ✅ Recent changes in worktree → Agent progressing normally
-- [orchestration-022] helpful=0 harmful=0: ❌ No commits AND task old → Might be stalled (investigate)
-- [orchestration-023] helpful=0 harmful=0: ❌ Worktree doesn't exist → Task not started (safe to work)
-
-**Real-World Example (Bug #168):**
-While Base Genie was implementing duplicate work, fix agent had ALREADY completed the fix in its worktree. Monitoring failed, but agent succeeded.
-
-**Lesson:** Infrastructure issues ≠ Agent failures. Always check worktree before assuming failure.
-
-**Related:** `@.genie/spells/troubleshoot-infrastructure.md` (5-step diagnostic protocol)
-
----
-
-## 📚 Domain-Specific Knowledge
-
-### Worktree Isolation Architecture
-
-**How Forge Works:**
-- Each task attempt = isolated git worktree
-- Location: `/var/tmp/automagik-forge/worktrees/<task-id-prefix>*/`
-- Independent workspace (doesn't affect main repo)
-- Agent commits to worktree branch
-- Success → merge to target branch
-- Failure → discard worktree
-
-**Why This Matters:**
-- [orchestration-024] helpful=0 harmful=0: Genie's main workspace ≠ Forge executor's worktree (different locations)
-- [orchestration-025] helpful=0 harmful=0: Editing main workspace while Forge works = boundary violation (creates conflicts)
-- [orchestration-026] helpful=0 harmful=0: Always check worktree to see actual progress (monitoring infrastructure may fail)
-
-### Agent Responsibility Matrix
-
-- [orchestration-027] helpful=0 harmful=0: Implementation work → Forge executor (Base Genie: orchestrate, monitor only)
-- [orchestration-028] helpful=0 harmful=0: Git operations → Git agent (Base Genie: delegate, don't execute)
-- [orchestration-029] helpful=0 harmful=0: Meta-learning → Base Genie direct execution (update spells, patterns)
-- [orchestration-030] helpful=0 harmful=0: Orchestration files → Base Genie direct execution (SESSION-STATE.md, plans)
-- [orchestration-031] helpful=0 harmful=0: Emergency hotfix → Base Genie if no Forge available (execute with caution)
-
-### Historical Context
-
-**First Documented Violation:**
-- **Date:** 2025-10-21 ~05:45 UTC
-- **Bug:** #168 (graceful shutdown)
-- **Task attempt:** b51db539
-- **Files affected:** src/cli/genie-cli.ts (historical - old path structure)
-- **Root cause:** Unclear boundaries between orchestration vs execution
-- **Pattern recognized by:** Felipe (user feedback)
-- **Learning applied:** This playbook created
-
-**Amendment Status:**
-This protocol became Amendment #4 in AGENTS.md: "Orchestration Boundary - Once Delegated, Never Duplicated"
-
----
-
-## ✅ Verification Checklist
-
-Before editing ANY implementation file, Base Genie must verify:
-
-### 1. Active Task Check
-- [orchestration-032] helpful=0 harmful=0: [ ] Check SESSION-STATE.md for active attempts before editing
-- [orchestration-033] helpful=0 harmful=0: [ ] Run `mcp__genie__list_sessions` to see running tasks
-- [orchestration-034] helpful=0 harmful=0: [ ] **If active task exists for this work → STOP** (let executor handle it)
-
-### 2. Worktree Status Check
-- [orchestration-035] helpful=0 harmful=0: [ ] List worktrees: `ls /var/tmp/automagik-forge/worktrees/`
-- [orchestration-036] helpful=0 harmful=0: [ ] Navigate to relevant worktree (if exists)
-- [orchestration-037] helpful=0 harmful=0: [ ] Check commits: `git log --oneline -5` (verify agent activity)
-- [orchestration-038] helpful=0 harmful=0: [ ] Check status: `git status` (see working changes)
-- [orchestration-039] helpful=0 harmful=0: [ ] **If commits exist → Agent is working! DO NOT DUPLICATE**
-
-### 3. Agent Responsibility Check
-- [orchestration-040] helpful=0 harmful=0: [ ] Is this implementation work? → Forge executor (delegate)
-- [orchestration-041] helpful=0 harmful=0: [ ] Is this orchestration? → Base Genie OK (proceed)
-- [orchestration-042] helpful=0 harmful=0: [ ] Is this learning? → Meta-learn protocol (use learn spell)
-- [orchestration-043] helpful=0 harmful=0: [ ] Is this git operations? → Git agent (delegate properly)
-
-### 4. Work Type Classification
-- [orchestration-044] helpful=0 harmful=0: [ ] Is this exploration (reading, analyzing)? → OK for Genie (proceed)
-- [orchestration-045] helpful=0 harmful=0: [ ] Is this execution (editing, implementing)? → Delegate (boundary rule)
-
-### 5. Quick Decision Tree
-
-```
-Are you about to edit an implementation file?
-│
-├─ YES → Is there an active Forge task for this work?
-│         │
-│         ├─ YES → STOP ❌ (Let Forge handle it)
-│         │
-│         └─ NO → Are you the right agent for this?
-│                  │
-│                  ├─ Implementation → Delegate to Forge
-│                  │
-│                  └─ Orchestration/Learning → Proceed ✅
-│
-└─ NO → Proceed ✅
-```
-
-**Remember:** Once delegated, never duplicated.