ToolsAgent

One repo for how autonomous agents behave and find things.

Pillar	Folder	Origin
How to behave	how_to_behave/	Submind — completion contracts, anti–fake-done, subconscious habits
What to find	what_to_find/search_as_code/	Search as Code — YAML profiles, discover → validate → rank → JSON
DREAM (nightly)	Dream/	Diary audit, reflect, hermes update, backup, Drive upload (scripts only)
ANT Loop (coding)	skills/ant-loop/	Autonomous /spec → /build → /review → Merge loop (self-contained: bundles backend-developer + code-review-and-quality)
Backend playbook	skills/backend-developer/	/build rails — REST, auth, SQL/NoSQL, caching, deploy
Code review gate	skills/code-review-and-quality/	/review multi-axis gate — correctness, security, perf, readability

Designed for Hermes Agent (and any Python agent loop).

Hermes wiring (default profile)

State lives in ~/.hermes/toolsagent/:

• active_contract.json — current task done-checks + evidence
• subconscious_memory.json + subconscious_procedures.md — soft habits from successful runs

Scripts:

• ~/.hermes/scripts/toolsagent_sac_hawks.sh — job scout pipeline
• toolsagent/hermes/behave_gate.py — must pass before claiming “done” on build/fix tasks
• toolsagent/hermes/reflect_run.py — optional post-run reflection

Load skill: toolsagent (skill_view toolsagent).

Behave: finish gate

python3 ~/toolsagent/hermes/behave_gate.py create \
  --goal "Ship feature X" \
  --checks "tests pass,artifact exists,user verified"

python3 ~/toolsagent/hermes/behave_gate.py evidence --check "tests pass" --output "9 passed"
python3 ~/toolsagent/hermes/behave_gate.py finish --message "Done, all tests pass"

Exit code 2 = not allowed to report complete (missing evidence or lazy flags).

Find: search pipeline

export PYTHONPATH=~/toolsagent/what_to_find
python3 -m search_as_code --profile ~/toolsagent/profiles/generic_web.yaml --pretty

Tests

cd ~/toolsagent
pip install -r requirements.txt pytest
PYTHONPATH=how_to_behave:what_to_find pytest -q

ANT Loop (autonomous coding)

skills/ant-loop/ is the orchestration layer: a self-driving coding loop where the only human steps are /spec (describe a feature) and 🚀 approve.

/spec ──▶ /build ⇄ /review ──▶ Merge (🚀)
          └───────────┘  (review loops back to build on blocking issues)

• /spec — writes a tracked issue (spec, acceptance, edge cases, tech).
• /build — cron build-monitor polls spec-ready issues, spawns a delegate_task subagent to implement (follows backend-developer playbook).
• /review — cron review-monitor tests, opens a PR, deploys a preview, notifies; waits for 🚀 before merging.
• Board — GitHub Issues (default) or a local Markdown file (board.example.md ships as the template).
• Daily improvement — a cron patches ant-loop + backend-developer with lessons from the day's runs.

Dependencies (bundled): the loop delegates to two sibling skills that ship in this repo — backend-developer (build rails) and code-review-and-quality (review gate). Load all three together: skill_view ant-loop, skill_view backend-developer, skill_view code-review-and-quality. The branded flow diagram is skills/ant-loop/references/loop.jpg.

License

MIT — Gowrishkar.