agent-test-loop

Your agent says "done". This makes it prove it.

agent-test-loop is a deterministic critic: it runs the checks you declare (tests, build, lint, a health endpoint, a secrets scan), captures the real output as evidence, and prints a score from 0 to 100. Below the threshold, delivery is blocked. No model calls, no network, no telemetry — stdlib only.

Install

pip install agent-test-loop

Use

agent-test-loop --config loop.json

loop.json:

{
  "name": "my-project",
  "threshold": 85,
  "max_rounds": 8,
  "checks": [
    {"name": "unit tests", "cmd": "pytest -q", "weight": 40, "timeout_s": 900},
    {"name": "build", "cmd": "npm run build", "weight": 25, "timeout_s": 600},
    {"name": "lint", "cmd": "ruff check .", "weight": 10},
    {"name": "no secrets", "cmd": "gitleaks detect --no-banner", "weight": 15,
     "fail_if_output_matches": "leaks found: [1-9]"},
    {"name": "api responds", "cmd": "curl -sS localhost:8000/health", "weight": 10,
     "must_match": "\"status\"\\s*:\\s*\"ok\""}
  ]
}

Output:

SCORE=100
GAPS=[]
EVIDENCE=.loop/round-1.json

Useful flags: --dry-run (validate the config without running), --quiet (only the three lines above), --json (full report), --show-state (last round from .loop/).

What it does not do

It does not judge code quality, style, or intent — it measures whether the commands you declared actually pass. That is the point: the agent that wrote the code is not allowed to grade itself.

Need the full gate?

Visual diff that does not produce false positives, MCP handshake preflight, GPU training guard and the delivery-proof ebook:

• Full pack — https://skills.seu.pet/en/p/pack-completo?utm_source=pypi&utm_medium=readme&utm_campaign=agent-test-loop
• Guided setup, first cycle run together — https://skills.seu.pet/en/service?utm_source=pypi&utm_medium=readme&utm_campaign=agent-test-loop

MIT licensed.