Agentic Loop Skill

A small, portable skill for turning a substantial objective into a bounded, resumable, proof-backed execution loop.

It started as a project-specific Claude Code proof. This edition keeps the useful ideas and removes the hard-coded machine paths, WhatsApp dependency, permission bypass, immortal-process claim, and single-project assumptions.

What it enforces

• frozen objective and acceptance criteria
• one active task at a time
• dependency-ordered work
• iteration and no-progress limits
• one retry per task
• a successful no-shell verifier run plus an existing artifact before a task can be marked done
• fail-closed completion
• resumable JSON state

The same agentic-loop/ folder follows the Agent Skills format used by both Codex and Claude Code.

Install

Codex, user scope:

python3 agentic-loop/scripts/install.py --target codex --scope user

Claude Code, user scope:

python3 agentic-loop/scripts/install.py --target claude --scope user

For a repository-only install, add --scope project --project /path/to/repo.

The installer refuses to overwrite an existing skill unless --force is supplied. A forced install moves the old directory to a timestamped backup first.

Important boundary

This skill does not create a 24/7 daemon. It preserves a loop while an agent is working and makes later sessions resumable. True unattended operation still needs a separately approved host scheduler, credential policy, stop switch, and runtime supervision.

Quick check

python3 -m unittest discover -s tests -v
python3 agentic-loop/scripts/loop_state.py --help

The test suite covers contract freezing, dependency order, one-task ownership, verifier receipts, finish-time oracle replay, file and directory artifact hashing, blocked-task recovery, retry limits, no-progress pauses, state tampering, missing or changed artifacts, terminal-state integrity, resume integrity, and fail-closed completion. CI runs the suite on Python 3.11 through 3.14 and exercises project-scoped installs for both Codex and Claude Code.

Security

Loop state can contain private objectives and local paths. Keep .agentic-loop/ out of public source, and only run verifier commands that you wrote or reviewed. See SECURITY.md for the full boundary.

License

MIT