agent-output

agent-output

Drastically reduce token consumption when running pytest or unittest inside an AI agent.

Heavily inspired by Nuno Maduro's Pao for PHP.

---

The problem

When an AI agent (Claude, Copilot, Cursor, …) runs your test suite, it reads the full terminal output to understand what happened. A typical pytest run for a medium-sized project can easily produce hundreds of lines:

============================= test session starts ==============================
platform linux -- Python 3.13.0, pytest-8.3.0, pluggy-1.5.0
rootdir: /home/user/myproject
collected 42 items

tests/test_auth.py .......                                               [ 16%]
tests/test_api.py ....F..                                                [ 33%]
...
=========================== short test summary info ============================
FAILED tests/test_api.py::test_create_user - AssertionError: expected 201, got 400
========================= 1 failed, 41 passed in 3.42s =========================

Every character of that output consumes tokens from the agent's context window. On large test suites this waste is significant, and it crowds out space for the actual code the agent needs to read and write.

The solution

agent-output replaces the entire human-readable terminal output with a single compact JSON line — but only when a test run is happening inside an AI agent environment. Regular developer sessions are completely unaffected.

When tests fail, the agent gets the overall result, counts, and precise failure locations — nothing else:

{
  "result": "failed",
  "tests": 42,
  "passed": 41,
  "failed": 1,
  "failures": [
    {
      "test": "tests/test_api.py::test_create_user",
      "file": "tests/test_api.py",
      "line": 14,
      "message": "AssertionError: expected 201, got 400"
    }
  ],
  "duration_ms": 3421
}

When everything passes, there's nothing to report beyond the essentials:

{
  "result": "passed",
  "tests": 42,
  "passed": 42,
  "duration_ms": 2187
}

Only fields that carry information are included — zero-count fields are omitted entirely.

How it works

Agent detection

agent-output uses ai-agent-detector to determine whether the current process is running inside an AI agent. Detection is automatic; no configuration is required.

You can also force agent mode by setting AI_AGENT=1 in your environment, or disable it entirely with AGENT_OUTPUT_DISABLE=1.

Zero-config bootstrap

The package installs a .pth file into your Python environment's site-packages. Python processes this file at interpreter startup, which imports agent_output.bootstrap before any user code runs — so the plugin activates without any changes to your project.

pytest integration

When pytest is detected in sys.argv, agent-output registers itself as a pytest plugin and unregisters the default terminal reporter. It collects every TestReport event and, on process exit, serialises the full summary to a single JSON line on stdout.

Captured fields: result, tests, passed, failed, failures (with file, line, message), errors, error_details, skipped, warnings, xfailed, xpassed, duration_ms.

unittest integration

When unittest is detected, agent-output monkey-patches unittest.TextTestRunner with a silent runner backed by a custom TestResult subclass that accumulates the same statistics. The JSON summary is emitted on process exit via atexit.

Installation

pip install agent-output

That's it. No conftest.py changes, no pytest.ini options, no import statements.

Output format

All fields except result, tests, passed, and duration_ms are omitted when their count is zero.

Disabling

Set AGENT_OUTPUT_DISABLE=1 to suppress agent-output even when an agent is detected. Useful in CI or any context where you want normal output regardless.

Requirements

• Python 3.11+
• pytest 8.0+ (optional — unittest works without pytest)

Credits

This project is heavily inspired by Pao by Nuno Maduro, which applies the same idea to PHP's Artisan test runner.

Author

Yann Rabiller (@einenlum) | blog | From PHP to Python

License

MIT