Ralph Execute — Autonomous Epic Runner

Process a Beads epic autonomously, dispatching fresh subagents per task with two-stage review gates.

Arguments

• First argument: Beads epic ID (required)
• --dry-run: Show what would execute without doing it (optional)

Pre-Flight Checks

Before starting the loop:

1. Verify epic exists: bd show <epic-id>
2. Check ready tasks: bd ready --parent <epic-id> — must have at least one
3. Verify worktree: Must be in a worktree, not main repo (enforced by block-work-on-develop hook)
4. Initialize circuit breaker: .claude/hooks/circuit-breaker.sh reset
5. Create session marker: mkdir -p .ralph-session/questions .ralph-session/answers .ralph-session/messages
6. Notify start: Send ntfy notification "Ralph starting epic "
7. Set phase to IMPLEMENT: If .claude/phases/output/current-phase.json exists and current_phase is not implement, update it:
   bashCopy

   PHASE_FILE=".claude/phases/output/current-phase.json"
   if [[ -f "$PHASE_FILE" ]]; then
     python3 -c "
   import json
   with open('$PHASE_FILE') as f: state = json.load(f)
   if state.get('current_phase') != 'implement':
       completed = state.get('completed_phases', [])
       if state.get('current_phase') and state['current_phase'] not in completed:
           completed.append(state['current_phase'])
       state['current_phase'] = 'implement'
       state['completed_phases'] = completed
       with open('$PHASE_FILE', 'w') as f: json.dump(state, f, indent=2)
   "
   fi
   

   If no phase file exists, create one:
   bashCopy

   mkdir -p .claude/phases/output
   echo '{"topic":"<epic-name>","current_phase":"implement","completed_phases":["research","plan"],"output_files":{}}' > "$PHASE_FILE"
   

Execution Loop

For each task from bd ready --parent <epic-id>:

Step A: Claim Task

bd update <task-id> --status in_progress

Log activity and update monitor:

echo "[$(date +%H:%M)] Task <task-id> started: <title>" >> .ralph-session/ralph.log
.claude/hooks/circuit-breaker.sh write-status "<task-id>" "<title>" <total> <completed> <mode>

Step A.5: Plan Pass (First Iteration Only)

For the FIRST iteration of each task, run a gap analysis:

.claude/hooks/generate-prompt.sh <task-id> .ralph-session/PROMPT.md plan

Dispatch a read-only subagent (model: haiku) with the plan-mode PROMPT.md. The subagent writes findings to .ralph-session/gap-analysis.md.

Validate plan pass output: After the plan subagent completes, verify .ralph-session/gap-analysis.md exists and has content:

if [[ ! -f ".ralph-session/gap-analysis.md" ]] || [[ $(wc -w < ".ralph-session/gap-analysis.md") -lt 20 ]]; then
  echo "[Ralph] Plan pass failed — no gap analysis produced" >> .ralph-session/ralph.log
  .claude/hooks/circuit-breaker.sh record-failure
  # Skip to next task
  continue
fi

Read the gap analysis before proceeding to build mode. Use it to inform the implementation subagent's context.

Step B: Generate PROMPT.md

.claude/hooks/generate-prompt.sh <task-id> .ralph-session/PROMPT.md build

Step C: Check Circuit Breaker

.claude/hooks/circuit-breaker.sh check

If PAUSED — stop loop, notify human, exit. If DEGRADED — wait 30 seconds (sleep 30) as a backoff, then continue.

Step D: Dispatch Subagent

Launch a Task subagent with the PROMPT.md content. The subagent:

• Reads the generated PROMPT.md
• Implements the task following the rules
• Runs verification commands (test, lint, typecheck)
• Outputs RALPH_COMPLETE when done
• Writes completion signal: echo "RALPH_COMPLETE" > .ralph-session/completion-signal

Step D.5: Handle Escalation

After the subagent returns, check if it escalated:

If result contains ESCALATE::

1. Read the question file:
   bashCopy

   cat .ralph-session/questions/<task-id>.json
   

2. Log the escalation:
   bashCopy

   echo "[$(date +%H:%M)] Task <task-id> ESCALATED: <summary>" >> .ralph-session/ralph.log
   

3. Ask the user the question directly in your response text. Include the full question, context, and options from the JSON file.
4. Write the user's answer to .ralph-session/answers/<task-id>.json:
   jsonCopy

   {"task_id":"<task-id>","answer":"<user's response>","timestamp":"<iso>"}
   

5. Dispatch a new subagent with a Task call that includes the user's answer in the prompt context, along with the original task description and any partial work from .ralph-session/.
6. After the new subagent completes, proceed to Step E as normal.

If result contains RALPH_COMPLETE: Proceed to Step E.

If result contains neither: The subagent may have failed silently. Record failure and proceed:

.claude/hooks/circuit-breaker.sh record-failure
echo "[$(date +%H:%M)] Task <task-id> returned without completion signal" >> .ralph-session/ralph.log

Step E: Two-Stage Review

Stage 1 — Spec Compliance: Launch a review subagent that compares the code changes against the task description. Check: Does the implementation match what was requested? Any missing requirements?

Stage 2 — Code Quality: Launch the code-reviewer agent (.claude/agents/code-reviewer.md). Check: Anti-patterns, security, test coverage, conventions.

Step F: Handle Review Results

If both pass:

.claude/hooks/circuit-breaker.sh record-success
bd close <task-id>
curl -s -X POST "https://ntfy.sh/property-tracker-claude" \
  -d "Task <task-id> complete" -H "Title: Ralph Progress"

Clean up: rm -f .ralph-session/completion-signal

Log completion:

echo "[$(date +%H:%M)] Task <task-id> completed (pass@1)" >> .ralph-session/ralph.log

If either fails:

.claude/hooks/circuit-breaker.sh record-failure

Dispatch a fix subagent with the review feedback. Then re-run Step E. Max 3 fix attempts per task (enforced by circuit breaker).

Step F.5: Record to eval-tracker

After each task resolution (pass or max-attempts-exceeded), append to .claude/instincts/eval-tracker.jsonl:

echo "{\"task_id\":\"<task-id>\",\"epic_id\":\"<epic-id>\",\"attempts\":1,\"fix_attempts\":<N>,\"passed\":<true|false>,\"date\":\"$(date -u +%Y-%m-%dT%H:%M:%SZ)\",\"branch\":\"$(git branch --show-current)\",\"duration_sec\":<elapsed>,\"skill_context\":[\"<instinct-id-1>\",\"<instinct-id-2>\"]}" >> .claude/instincts/eval-tracker.jsonl

Where:

• attempts: always 1 (one dispatch per task)
• fix_attempts: 0 if passed on first try, 1-3 for fix iterations
• passed: true if eventually passed review, false if max attempts exceeded
• duration_sec: wall-clock seconds from dispatch to completion
• skill_context: array of instinct IDs that were active (injected via {{LEARNED_PATTERNS}}) when this task was dispatched. Enables /eval-report Step 4 to measure per-instinct effectiveness.

Step G: Next Task

Loop back to Step A with the next bd ready --parent <epic-id> task.

Exit Conditions

The loop exits when ANY of these are true:

1. bd ready --parent <epic-id> returns no tasks (epic complete)
2. Circuit breaker enters PAUSED state
3. User manually creates .ralph-session/STOP file

Cleanup

On exit (success or circuit breaker), save summary before deleting session evidence:

# Preserve session summary for post-mortem analysis
mkdir -p .claude/agent-memory/ralph
if [[ -f .ralph-session/circuit-breaker.json ]]; then
  cp .ralph-session/circuit-breaker.json ".claude/agent-memory/ralph/session-$(date +%Y%m%d-%H%M%S).json"
fi
rm -rf .ralph-session

Phase Advance (auto): After all tasks complete (not on PAUSED exit):

PHASE_FILE=".claude/phases/output/current-phase.json"
TOPIC=$(python3 -c "import json; print(json.load(open('$PHASE_FILE')).get('topic','unknown'))" 2>/dev/null)

# Run checkpoint evaluation
CHECKPOINT_RESULT=$(.claude/hooks/checkpoint-eval.sh implement review "$TOPIC" 2>&1)
CHECKPOINT_EXIT=$?

if [[ "$CHECKPOINT_EXIT" -eq 0 ]]; then
  python3 -c "
import json
with open('$PHASE_FILE') as f: state = json.load(f)
state['completed_phases'] = list(set(state.get('completed_phases', []) + ['implement']))
state['current_phase'] = 'review'
with open('$PHASE_FILE', 'w') as f: json.dump(state, f, indent=2)
"
  echo "Phase advanced to REVIEW"
else
  echo "Phase remains IMPLEMENT — checkpoint failed: $CHECKPOINT_RESULT"
fi

On PAUSED exit, do NOT advance phase — user intervention needed.

Send final notification:

curl -s -X POST "https://ntfy.sh/property-tracker-claude" \
  -d "Ralph finished epic <id>. Status: <complete|paused>" \
  -H "Title: Ralph Done" -H "Priority: high"
osascript -e 'display notification "Ralph finished" with title "Claude Code"'

Safety Guarantees

• All 17 existing hooks remain active (lint, anti-patterns, security, auth checks)
• Circuit breaker prevents infinite loops (max 3 failures per task, max 5 iterations without progress)
• Fresh subagent per task prevents context pollution
• Two-stage review catches both spec drift and quality issues
• ntfy notifications on every state change

Language Patterns (From Ghuntley Research)

When generating PROMPT.md, these language patterns improve subagent performance:

• "Study" not "read" or "look at" — triggers deeper analysis
• "Don't assume not implemented" — prevents hallucinated code
• "Using parallel subagents" — explicit parallelism instruction
• "Only 1 subagent for build/tests" — enforces backpressure
• "Capture the why" — documentation emphasis