pxx: Instructions file for Codex

AGENTS.md — working agreements for AI agents in this repo

What this is

pxx 2.0: a local-first, async AI coding-agent runtime. DESIGN.md is the authoritative architecture contract — read it before changing anything. If a contract must change, change DESIGN.md in the same commit.

Commands

• Test: uv run pytest (or .venv/bin/python -m pytest) — must stay green. Tests require no network, no Ollama, no aider, no git repo.
• Lint: uv run ruff check pxx tests — must stay clean.
• Run locally: .venv/bin/pxx doctor

Hard rules

1. Gates raise; telemetry suppresses. Safety/scope/hook/budget decisions raise from pxx.errors and propagate. Audit and memory writes are best-effort and must never crash a session.
2. Paths are untrusted. Canonicalize (pxx.safety.canonicalize, symlinks resolved) before any gate decision. Never derive the trust boundary from model output.
3. Audit is metadata-only. No prompt bodies, file contents, diffs, or secrets in events/audit. Tool-event previews stay truncated.
4. Async core. New long-running operations are async def; tests drive them with asyncio.run() in sync functions (no pytest-asyncio).
5. Pure core, thin I/O edges. Policy modules take data and return decisions; subprocess/network/FS I/O stays injectable. Every module ships with tests in tests/test_<module>.py.
6. Memory is context, never policy. Hooks are deterministic gates — never let model output override them.
7. Keep the tool surface small (~8 built-ins). Small local models degrade past ~10 tools.

Conventions

• Python >= 3.11, from __future__ import annotations everywhere, dataclasses over dicts, StrEnum for enums, logging.getLogger("pxx") (no print outside cli.py/doctor.py).
• Core dependency budget: stdlib + httpx. New core deps need a strong reason; optional functionality goes in extras.
• Follow the existing module layout in DESIGN.md.