phase-4-heal

Phase 4: Systematic Healing

Goal

Systematically investigate and fix implementation errors using root cause analysis. NO FIXES WITHOUT INVESTIGATION FIRST.

The Iron Law

NO FIXES WITHOUT ROOT CAUSE INVESTIGATION FIRST

If you haven't completed Phase 1 (Root Cause Investigation), you CANNOT propose fixes.

Input

• Task ID: {task_id}
• Error message: {error_msg}
• Error context: Stack trace, test output, build logs

---

Workflow

Step 0: Initialize CLI (Automatic)

IMPORTANT: This skill requires the Ralph-dev CLI. It will build automatically on first use.

# Bootstrap CLI - runs automatically, builds if needed
source ${CLAUDE_PLUGIN_ROOT}/shared/bootstrap-cli.sh

# Verify CLI is ready
ralph-dev --version

# Context-compression resilience: Verify current phase
CURRENT_PHASE=$(ralph-dev state get --json 2>/dev/null | jq -r '.phase // "none"')
echo "Current phase: $CURRENT_PHASE"
# Expected: implement or heal (invoked by phase-3)

Step 1: Load Context

# Get task details (context-compression safe)
TASK_JSON=$(ralph-dev tasks get "$TASK_ID" --json)
TASK_DESC=$(echo "$TASK_JSON" | jq -r '.data.description // "No description"')

# Get test command from language config
TEST_CMD=$(ralph-dev detect --json | jq -r '.verifyCommands[] | select(contains("test"))' | head -1)

echo "Task: $TASK_ID"
echo "Description: $TASK_DESC"
echo "Test command: $TEST_CMD"

---

The Four Phases

MANDATORY: Complete each phase before proceeding to the next.

Phase 1: Root Cause Investigation

1. Read error completely - Extract line number, file path, error code
2. Reproduce consistently - Run CI=true {test_command} to confirm
3. Check recent changes - git log -5, git diff HEAD~1
4. Trace data flow - Find where the bad value originates

Phase 2: Pattern Analysis

1. Find working examples - Search codebase for similar working patterns
2. Compare working vs broken - Identify key differences
3. Check dependencies - Verify modules are installed and versions match

Phase 3: Hypothesis and Testing

1. Form hypothesis - Based on investigation, classify error type
2. Use WebSearch - Confirm hypothesis with proven solutions
3. Validate approach - Ensure fix addresses root cause, not symptoms

Error Classification:

Error Type	Indicators	Fix Type
Missing dependency	"Module not found", "Cannot find module"	npm install
Type error	"TypeError", type mismatch	Code fix
Undefined reference	"is not defined", "ReferenceError"	Import/declaration
Test failure	"Expected X but got Y"	Implementation logic
Build error	"compilation error", "build failed"	Config fix

Phase 4: Implementation with Verification

Max 3 attempts. After 3 failures → mark task as failed.

MAX_ATTEMPTS=3
ATTEMPT=1

while [ $ATTEMPT -le $MAX_ATTEMPTS ]; do
  # Apply single fix based on hypothesis
  # Run tests: CI=true {test_command}

  if [ $? -eq 0 ]; then
    # SUCCESS - return result
    break
  fi

  ATTEMPT=$((ATTEMPT + 1))
  # If still failing, refine hypothesis and retry
done

CRITICAL: Always show full test output as verification evidence.

---

Return Result

REQUIRED Output Format:

---HEALING RESULT---
task_id: {task_id}
status: success | failed
verification_passed: true | false
attempts: {N}
fix_type: dependency | code | implementation | config
hypothesis: {what was wrong}
notes: {brief summary}
---END HEALING RESULT---

---

Safety Limits

Limit	Value	Purpose
Max attempts	3	Prevent infinite retry loops
Timeout	1 hour	Prevent stuck healing
Single fix per attempt	1	Isolate variables

---

Constraints

• NEVER apply fixes without completing Phase 1 investigation
• NEVER apply multiple fixes at once (isolate variables)
• NEVER retry more than 3 times
• ALWAYS use CI=true when running tests
• ALWAYS show full test output as verification evidence
• ALWAYS return structured HEALING RESULT block

---

Error Handling

Error	Action
Cannot reproduce	Gather more data, check environment
No working examples	Use WebSearch for reference implementations
All 3 attempts fail	Mark task failed, require manual intervention
Hypothesis unclear	Return to Phase 1 with more investigation

---

Key Principles

1. Investigation before action - Understand the problem first
2. One fix at a time - Isolate variables for debugging
3. Evidence-based - Show test output, don't just assert success
4. WebSearch to confirm - Don't guess, find proven solutions
5. Circuit breaker - 3 attempts max, then escalate