Autonomous Loop

Autonomous Loop

The loop lets a single /ck:make invocation run dozens of agent iterations without user intervention, while still respecting per-task budgets, session budgets, and explicit user-approval gates.

Architecture

┌─────────────────┐  Stop event   ┌────────────────────┐
│  Claude Code    │──────────────▶│   stop-hook.sh     │
│   session       │               │  (.cavekit/loop    │
└─────────────────┘               │   active?)         │
        ▲                         └─────────┬──────────┘
        │                                   │ route()
        │ next prompt injected              ▼
        │                         ┌────────────────────┐
        │                         │ cavekit-tools.cjs  │
        │                         │  routeDecision()   │
        │                         │  + status-block    │
        │                         │  + backprop-       │
        │                         │    directive       │
        │                         └─────────┬──────────┘
        │                                   │
        └───────── {decision:"block",       │
                    reason:<next prompt>} ◀─┘

Files the loop touches

All under <project>/.cavekit/:

Lifecycle

1. Setup — /ck:make calls:

   node "${CLAUDE_PLUGIN_ROOT}/scripts/cavekit-tools.cjs" setup-loop
   

   This writes .loop.json (activating the stop hook) and resets state.md.

2. Work — the agent does one wave of task execution per iteration.

3. Stop fires — Claude Code's Stop event triggers stop-hook.sh. The hook:

   • reads stdin for session_id and transcript_path
   • acquires / refreshes the lock
   • scans the last 20 transcript lines for <promise>CAVEKIT COMPLETE</promise>
   • if sentinel found → teardown + exit silently
   • else → asks routeDecision() for the next prompt
   • prepends the backprop directive if the flag file exists
   • returns {"decision":"block","reason":<next prompt>}

4. Repeat — Claude Code treats decision:block + reason:... as a new user message, so the session continues.

5. Teardown — one of:

   • completion sentinel detected
   • max-iterations reached (CAVEKIT_MAX_ITERATIONS)
   • session budget exhausted (CAVEKIT_BUDGET_EXHAUSTED)
   • lock stolen by another session (CAVEKIT_LOCK_CONFLICT)
   • user interrupt (hook returns no output; session stops normally)

Completion sentinel

To end the loop cleanly, emit exactly:

<promise>CAVEKIT COMPLETE</promise>

The hook searches for this literal in the last 20 transcript lines. Put it on its own line at the very end of the final message. Do not wrap it in code fences or paraphrase it — the search is a literal substring match.

Terminal sentinels

When routeDecision() cannot safely continue, it returns one of these strings instead of a prompt:

The hook translates each into a short user-facing message before returning.

Lock protocol

The lock is a JSON file (.loop.lock) with {owner, pid, host, heartbeat_at}.

• Owner tag: session:<session_id>.
• Heartbeat: every Stop invocation refreshes heartbeat_at.
• Stale: > 5 minutes since last heartbeat. A new session may steal a stale lock.
• Conflict: non-owner with fresh lock → returns CAVEKIT_LOCK_CONFLICT.

Never delete .loop.lock while a session might be active. Use cavekit-tools release-lock --owner <tag> instead.

Debugging a stuck loop

# Inspect current state
node "${CLAUDE_PLUGIN_ROOT}/scripts/cavekit-tools.cjs" status

# Who holds the lock?
cat .cavekit/.loop.lock

# What would the router do right now?
node "${CLAUDE_PLUGIN_ROOT}/scripts/cavekit-tools.cjs" route

# Drop the loop entirely (safe after a crash)
node "${CLAUDE_PLUGIN_ROOT}/scripts/cavekit-tools.cjs" teardown-loop

Turn on debug logging by exporting CAVEKIT_DEBUG=1. The hook will write to .cavekit/.debug.log.

What not to do

• Do not hand-edit state.md's phase field while a loop is active. The hook assumes single-writer semantics and will overwrite you.
• Do not set your own stop hook that also blocks. Multiple blocking hooks race, and Claude Code picks one arbitrarily.
• Do not emit the completion sentinel unless every task in the registry is truly complete. The hook trusts the sentinel and tears down immediately.

Source: hashgraph-online/awesome-codex-plugins → plugins/JuliusBrussee/blueprint/skills/autonomous-loop/SKILL.md