$crank - Autonomous Epic Execution (Codex Native)

Quick Ref: Execute every open issue in an epic via wave-based workers using spawn_agent, wait_agent, send_input, and close_agent. Output: closed issues + final validation.

You must execute this workflow. Do not just describe it.

Architecture

Crank (lead agent)
    |
    +-> bd ready (current wave)
    |
    +-> Build a wave task packet
    |
    +-> spawn_agent per issue (worker or explorer role)
    |
    +-> wait_agent for all worker ids
    |
    +-> Validate results + bd update
    |
    +-> Loop until epic DONE

Backend Rules

1. Prefer Codex session agents when spawn_agent is available.
2. Use agent_type=worker for implementation agents and agent_type=explorer for discovery agents when the runtime exposes roles.
3. Use send_input only for short steering or retry prompts.
4. Use close_agent for stalled or unnecessary agents.
5. Never depend on legacy CSV fan-out or host-task result polling. Use spawn_agent, wait_agent, send_input, and close_agent instead.

Codex Lifecycle Guard

When this skill runs in Codex hookless mode (CODEX_THREAD_ID is set or CODEX_INTERNAL_ORIGINATOR_OVERRIDE is Codex Desktop), ensure startup context before the first wave:

ao codex ensure-start 2>/dev/null || true

ao codex ensure-start is the single startup guard for Codex skills. It records startup once per thread and skips duplicate startup automatically. Leave ao codex ensure-stop to closeout skills after the implementation wave ends.

Flags

Global Limits

MAX_EPIC_WAVES = 50 (hard limit). Typical epics use 5-10 waves.

Completion Enforcement (Sisyphus Rule)

After each wave, output one of:

• <promise>DONE</promise> - epic complete, all issues closed
• <promise>BLOCKED</promise> - cannot proceed, with reason
• <promise>PARTIAL</promise> - incomplete, with remaining items

Never claim completion without the marker.

Node Repair Operator

When a task fails during wave execution, classify as RETRY (transient — re-add with adjustment, max 2), DECOMPOSE (too complex — split into sub-issues, terminal), or PRUNE (blocked — escalate immediately). Budget: 2 per task.

Mutation logging on failure: DECOMPOSE logs task_removed + task_added per sub-task. PRUNE logs task_removed. RETRY logs nothing (task identity unchanged).

Execution Steps

Given $crank [epic-id | .agents/rpi/execution-packet.json | plan-file.md | "description"]:

Step 0: Load Knowledge Context

if command -v ao &>/dev/null; then
    ao lookup --query "<epic-title>" --limit 5 2>/dev/null || true
    ao ratchet status 2>/dev/null || true
fi

Apply retrieved knowledge: If learnings are returned, check each for applicability to this epic. For applicable learnings, treat as implementation constraints and cite by filename. Record citations with the correct type: ao metrics cite "<path>" --type applied when the learning influenced a decision, or --type retrieved when loaded but not referenced.

Section evidence: When lookup results include section_heading, matched_snippet, or match_confidence fields, prefer the matched section over the whole file — it pinpoints the relevant portion. Higher match_confidence (>0.7) means the section is a strong match; lower values (<0.4) are weaker signals. Use the matched_snippet as the primary context rather than reading the full file.

Step 0.5: Detect Tracking Mode

if bd ready --json >/dev/null 2>&1 && bd list --type epic --status open --json >/dev/null 2>&1; then
    TRACKING_MODE="beads"
else
    TRACKING_MODE="tasklist"
fi

Step 0.6: Initialize Shared Task Notes

Create the shared notes file for cross-wave context persistence. See references/shared-task-notes.md for the full pattern.

mkdir -p .agents/crank
cat > .agents/crank/SHARED_TASK_NOTES.md <<EOF
# Shared Task Notes — Epic ${EPIC_ID:-unknown}

> Cross-wave context for workers. Read before starting. Report discoveries in task output.
> Maintained by the crank orchestrator — workers do NOT write to this file directly.

EOF

Step 0.7: Initialize Plan Mutation Audit Trail

Create the JSONL file that tracks every plan mutation during execution. See references/plan-mutations.md for the full schema and mutation budget.

mkdir -p .agents/rpi
: > .agents/rpi/plan-mutations.jsonl

# Budget counters
MUTATION_TASK_ADDED=0
MUTATION_TASK_ADDED_LIMIT=5
MUTATION_TASK_REORDERED=0
MUTATION_TASK_REORDERED_LIMIT=3

Helper function:

log_plan_mutation() {
    local mutation_type="$1" task_id="$2" before="$3" after="$4"
    local ts
    ts=$(date -Iseconds)

    if [[ "$mutation_type" == "task_added" ]]; then
        MUTATION_TASK_ADDED=$((MUTATION_TASK_ADDED + 1))
        if [[ $MUTATION_TASK_ADDED -gt $MUTATION_TASK_ADDED_LIMIT ]]; then
            echo "WARN: task_added budget exceeded ($MUTATION_TASK_ADDED/$MUTATION_TASK_ADDED_LIMIT). Consider re-running $plan."
        fi
    elif [[ "$mutation_type" == "task_reordered" ]]; then
        MUTATION_TASK_REORDERED=$((MUTATION_TASK_REORDERED + 1))
        if [[ $MUTATION_TASK_REORDERED -gt $MUTATION_TASK_REORDERED_LIMIT ]]; then
            echo "WARN: task_reordered budget exceeded ($MUTATION_TASK_REORDERED/$MUTATION_TASK_REORDERED_LIMIT)."
        fi
    fi

    echo "{\"timestamp\":\"$ts\",\"wave\":$wave,\"task_id\":\"$task_id\",\"mutation_type\":\"$mutation_type\",\"before\":$before,\"after\":$after}" \
        >> .agents/rpi/plan-mutations.jsonl
}

Mutation types: task_added, task_removed, task_reordered, scope_changed, dependency_changed.

Step 1: Identify the Execution Target

Beads mode:

• If epic ID provided: use it directly
• If no epic ID: bd list --type epic --status open 2>/dev/null | head -5

Execution-packet/file mode:

• If the input is .agents/rpi/execution-packet.json, read objective, epic_id, tracker_mode, done_criteria, and validation_commands
• If epic_id exists inside the execution packet, keep that epic as the execution spine
• If epic_id is absent, keep the packet objective as the execution spine and continue in file-backed mode instead of inventing an epic ID
• For other plan files, read the plan file and extract tasks

Step 2: Load Execution Details

Beads mode:

bd show <epic-id> 2>/dev/null

Execution-packet/file mode:

• Read the packet or plan file into local state for the current objective
• Preserve the same objective across retries; do not narrow to one slice from bd ready

Step 3: List Ready Work for the Current Wave

Beads mode:

bd ready 2>/dev/null

bd ready returns all unblocked issues - these can run in parallel.

Execution-packet/file mode:

• Read remaining tasks from .agents/rpi/execution-packet.json or the plan file
• Execute against the packet objective until the plan-backed work is done, blocked, or the retry budget is exhausted

Step 3a: Pre-flight Checks

1. Verify there are ready issues. Empty list is an error unless the epic is already complete.
2. If 3+ issues are ready, check .agents/council/ for pre-mortem evidence.
3. If tracking mode is beads and scripts/bd-audit.sh exists, run the backlog audit before spawning workers.
4. If bd-audit flags backlog hygiene issues, stop and clean them up before continuing. Use --skip-audit only when you intentionally want to bypass that gate.
5. For every string being modified, grep the codebase for stale cross-references.

Step 3b: Language Standards Injection

Detect project language (go.mod -> Go, pyproject.toml -> Python, etc.) and read applicable standards from $standards. Include a Testing section in worker prompts.

Step 4: Execute the Wave with Codex Session Agents

Crank follows the FIRE loop for each wave:

• FIND: locate the next ready set
• IGNITE: spawn workers
• REAP: wait, validate, and merge results
• ESCALATE: retry or block when needed

4a: Load Shared Task Notes

Read cross-wave context to include in worker prompts:

SHARED_NOTES=""
if [ -f .agents/crank/SHARED_TASK_NOTES.md ]; then
    SHARED_NOTES=$(cat .agents/crank/SHARED_TASK_NOTES.md)
fi

If SHARED_NOTES exceeds ~50 lines, summarize older waves (keep last 3 in full detail, preserve [CRITICAL] entries).

4b: Build a Wave Task Packet

Create one packet per ready issue. Do not use CSV fan-out.

mkdir -p .agents/crank
cat > ".agents/crank/wave-${wave}-tasks.json" << EOF
{
  "wave": $wave,
  "epic_id": "$EPIC_ID",
  "tasks": [
    {
      "issue_id": "bd-123",
      "subject": "Short issue summary",
      "description": "Issue details and acceptance criteria",
      "files": ["path/to/file.go"],
      "validation_cmd": "go test ./...",
      "metadata": {
        "issue_type": "feature"
      }
    }
  ]
}
EOF

Each task packet must include metadata.issue_type.

4c: Pre-spawn File Conflict Check

wave_tasks = [tasks from packet]
all_files = {}
for task in wave_tasks:
    for f in task.files:
        if f in all_files:
            CONFLICT -> serialize into sub-waves
        all_files[f] = task.id

Display an ownership table before spawning workers. If conflicts exist, split into sub-waves and keep file ownership disjoint.

4d: Spawn Workers

Spawn one agent per issue. Prefer worker roles for implementation and explorer roles for file discovery when the runtime exposes agent_type.

spawn_agent(
  agent_type="worker",
  message="You are worker-<issue-id>.

Assignment: <subject>

<description>

---
Context from prior waves (read before starting):
<SHARED_NOTES content, or 'First wave — no prior context.' if empty>

---

FILE MANIFEST (files you are permitted to modify):
<list of files>

Rules:
1. Stay within your assigned files
2. Run validation: <validation_cmd>
3. Keep your response short
4. Write any durable notes to .agents/crank/results/<issue-id>.md or .agents/crank/results/<issue-id>.json
5. DISCOVERY REPORTING: If you discover codebase quirks, failed approaches,
   convention requirements, or dependency constraints, include a section in your
   output titled '## Discoveries' with one bullet per finding.

Use the repo's current Codex primitives only."
)

If a task is missing its file manifest, spawn a short-lived explorer agent first:

spawn_agent(
  agent_type="explorer",
  message="You are explorer-<issue-id>.

Task: identify the files that must be created or modified for this issue.
Return a JSON array of paths only."
)

4e: Wait for Workers

wait_agent(ids=["agent-id-1", "agent-id-2"])

If a worker needs a short correction, use send_input(id=..., message=...).

If a worker stalls or is no longer needed, use close_agent(id=...).

Step 5: Verify and Sync

External Gate Enforcement: After each worker completes, the orchestrator (not the worker) runs the gate command. Workers must not declare their own completion. See references/external-gate-protocol.md.

For each completed worker:

1. PASS -> close the issue.
2. FAIL -> log the failure, keep the issue open, and retry only if the issue is still within the retry budget.
3. BLOCKED -> mark blocked with the reason and continue the wave.

Update beads:

bd close "$issue_id" 2>/dev/null
bd update "$issue_id" --status blocked --append-notes "Wave $wave FAIL: $reason" 2>/dev/null

Step 5.5: Wave Acceptance Check

After all workers complete:

1. Compute git diff for the wave.
2. Run project-level tests appropriate to the wave.
3. If tests fail, identify which worker's changes broke things and requeue only that work.

Step 5.7: Wave Checkpoint

cat > ".agents/crank/wave-${wave}-checkpoint.json" << EOF
{
  "wave": $wave,
  "epic_id": "$EPIC_ID",
  "completed": $COMPLETED_COUNT,
  "failed": $FAILED_COUNT,
  "timestamp": "$(date -u +%Y-%m-%dT%H:%M:%SZ)",
  "mutations_this_wave": $(grep -c "\"wave\":${wave}" .agents/rpi/plan-mutations.jsonl 2>/dev/null || echo 0),
  "total_mutations": $(wc -l < .agents/rpi/plan-mutations.jsonl 2>/dev/null | tr -d ' '),
  "mutation_budget": {
    "task_added": {"used": ${MUTATION_TASK_ADDED:-0}, "limit": 5},
    "task_reordered": {"used": ${MUTATION_TASK_REORDERED:-0}, "limit": 3}
  }
}
EOF

Step 5.8: Update Shared Task Notes

Harvest discoveries from completed workers and append to the shared notes file:

WAVE_DISCOVERIES=""
for result_file in .agents/crank/results/*; do
    if [ -f "$result_file" ]; then
        DISCOVERIES=$(sed -n '/^## Discoveries/,/^## /{ /^## Discoveries/d; /^## /d; p; }' "$result_file" 2>/dev/null)
        if [ -n "$DISCOVERIES" ]; then
            WAVE_DISCOVERIES="${WAVE_DISCOVERIES}${DISCOVERIES}\n"
        fi
    fi
done

if [ -n "$WAVE_DISCOVERIES" ]; then
    cat >> .agents/crank/SHARED_TASK_NOTES.md <<EOF

## Wave ${wave} ($(date -Iseconds))
$(echo -e "$WAVE_DISCOVERIES")
EOF
fi

Capture: Failed approaches, codebase quirks, convention discoveries, dependency notes. Skip: Full error logs, implementation details, task status.

Step 5.9: Log Plan Mutations

After processing wave results, log mutations for any plan changes. Call log_plan_mutation for each:

• DECOMPOSE: task_removed for original, task_added for each sub-task
• PRUNE: task_removed with block reason
• Scope change: scope_changed when file manifest updated after exploration
• Dependency discovered: dependency_changed when blocked-by list modified
• Wave reassignment: task_reordered when task moves between waves

# Example: task decomposed into sub-tasks
log_plan_mutation "task_removed" "$decomposed_id" \
    "{\"subject\":\"$ORIGINAL_SUBJECT\",\"status\":\"decomposed\"}" "null"
log_plan_mutation "task_added" "$sub_id" "null" \
    "{\"subject\":\"$SUB_SUBJECT\",\"reason\":\"Split from $decomposed_id\"}"

# Example: scope change after exploration
log_plan_mutation "scope_changed" "$task_id" \
    "{\"files\":$ORIGINAL_FILES}" \
    "{\"files\":$UPDATED_FILES,\"reason\":\"$REASON\"}"

Mutations are append-only to .agents/rpi/plan-mutations.jsonl. Read by $post-mortem for drift analysis.

Step 6: Commit Wave Results

Lead-only commit - workers write files, lead validates and commits once per wave:

for f in $WORKER_FILES_CHANGED; do
    git add -- "$f"
done
git commit -m "feat(<scope>): wave $wave - $COMPLETED_COUNT issues completed"

Step 7: Loop or Complete

wave=$((wave + 1))

if [[ $wave -ge 50 ]]; then
    echo "<promise>BLOCKED</promise>"
    echo "Global wave limit (50) reached."
    exit 1
fi

REMAINING=$(bd ready 2>/dev/null | wc -l)
if [[ $REMAINING -eq 0 ]]; then
    ALL_CLOSED=$(bd children "$EPIC_ID" 2>/dev/null | grep -c "CLOSED" || echo 0)
    ALL_TOTAL=$(bd children "$EPIC_ID" 2>/dev/null | wc -l || echo 0)

    if [[ $ALL_CLOSED -eq $ALL_TOTAL ]]; then
        echo "<promise>DONE</promise>"
    else
        echo "<promise>BLOCKED</promise>"
        echo "No ready issues but $((ALL_TOTAL - ALL_CLOSED)) issues remain unclosed."
    fi
else
    # Continue to next wave - return to Step 3
fi

Step 8: Final Validation

When the epic is DONE:

$vibe validate the completed epic

Step 8.5: Archive Shared Task Notes

Move the shared notes to an archive after epic completion:

if [ -f .agents/crank/SHARED_TASK_NOTES.md ]; then
    mkdir -p .agents/crank/archives
    mv .agents/crank/SHARED_TASK_NOTES.md \
       ".agents/crank/archives/SHARED_TASK_NOTES-${EPIC_ID:-unknown}-$(date +%Y%m%d-%H%M%S).md"
fi

Retry Policy

• Max 2 retries per issue across all waves
• On third failure: mark BLOCKED and continue with remaining issues
• Track retries with bd comments add "$issue_id" "retry $N: $reason"

Failure Recovery

Reference Documents

• references/de-sloppify.md - cleanup pass after implementation waves
• references/plan-mutations.md - plan mutation audit trail for drift analysis
• references/shared-task-notes.md - cross-wave context persistence
• references/commit-strategies.md - per-task vs wave-batch commits
• references/contract-template.md - contract template for worker specs
• references/failure-recovery.md - escalation and retry logic
• references/failure-taxonomy.md - failure classification
• references/fire.md - FIRE loop specification
• references/ralph-loop-contract.md - Ralph Wiggum loop contract
• references/taskcreate-examples.md - task creation examples
• references/team-coordination.md - worker coordination details
• references/external-gate-protocol.md - external gate protocol for wave validation
• references/test-first-mode.md - test-first wave sequence
• references/troubleshooting.md - common issues and fixes
• references/uat-integration-wave.md - UAT integration wave patterns
• references/wave-patterns.md - acceptance checks and checkpoints
• references/gc-pool-dispatch.md - gc pool worker dispatch
• references/wave1-spec-consistency-checklist.md - Wave 1 spec consistency checklist
• references/worktree-per-worker.md - worktree isolation pattern