{"data":{"kind":"file","path":"README.md","version_id":"f6og8ordoqwhm1o2t4f6q5we","entry":{"name":"README.md","path":"README.md","is_directory":false,"size":2746,"modified_at":"2026-10-18T02:37:20.402000","content_hash":"9cbe9e9e6975c84fa3b5cdeb15d5e7da1410ba2715ebab980e18d2a1624fb805"},"entries":[],"content":"# Production Fire Drill — Autonomous SRE Agent (V0.0.1)\n\nStateful Prime Intellect Verifiers environment for incident response. Default `backend=\"sim\"` is a deterministic Python simulator with 24 scenarios (six failure families × four telemetry/load levels). Optional `backend=\"docker\"` provides one real connection-leak drill with FastAPI, PostgreSQL 16 and Redis 7 in a per-rollout Compose project. The Docker backend is opt-in, requires a working Docker daemon and Compose v2, and is not guaranteed to run on hosted Prime Intellect workers. Never enable Docker on a host with production containers or untrusted co-tenants; this integration uses the host Docker daemon and a bounded allowlist, not a security-grade sandbox.\n\nThe agent gets three constrained tools: `inspect(area)`, `apply_change(path,field,value)`, `operate(action)`. In Docker mode it can inspect API logs, live pool metrics and `app.py`, then apply the allowlisted connection-lifecycle edit. No raw shell or arbitrary Docker command is exposed. The verifier probes real requests, repeats requests beyond the connection-pool capacity, and compares PostgreSQL records to an independent seeded-data oracle. Docker projects and volumes are removed after scoring or on model-generation errors; process crashes can leave residual projects. The Docker path has one case, not 24 Docker scenarios.\n\n## Scoring\n\nDiagnosis/evidence 20, real recovery plus smoke check 30, safety/data integrity 20, regression and integrity checks 15, action efficiency 10, accurate report 5. Reward is normalized to [0,1]. Destructive tool actions are blocked and penalized. The Docker mode verifies liveness and records from the running containers; the default mode checks simulator state.\n\n## Run locally\n\nInstall Python 3.12 and Docker Compose v2. `uv venv --python python3.12 .venv && uv pip install --python .venv/bin/python -e . pytest`.\n\n- Simulator: `.venv/bin/python -m pytest -q tests/test_env.py` and `.venv/bin/python tests/rollout_smoke.py`.\n- Docker drill: `.venv/bin/python -m pytest -q tests/test_docker.py` (runs containers; the integration test creates its own Compose project).\n- Loader: `from production_fire_drill import load_environment; env = load_environment(backend=\"docker\")`; default is `load_environment()` for all 24 simulator cases.\n\n`split=\"train\"` and `split=\"eval\"` share the same public templates; these are not unseen cases. No independent private grading service is available: reward logic is published in the wheel and visible to anyone who inspects the package. A true hidden verifier requires a separate, access-controlled grading service and unseen incident families. Do not call the included tests a model evaluation. Display label V0.0.1; wheel version 0.0.1.\n","encoding":"utf-8","truncated":false,"total_bytes":2746},"status":null}