agents-md-debugger

See exactly which AGENTS.md instructions Codex is expected to discover for any file in a repository—without starting a model session.

$ agents-md-debugger explain src/api/user.py
Project root: /work/my-app
Project instruction budget: 1840/32768 bytes

Effective project instruction chain (root -> working directory):
  1. AGENTS.md [scope ./**, loaded]
  2. src/AGENTS.md [scope src/**, loaded]
  3. src/api/AGENTS.override.md [scope src/api/**, loaded]

The tool is local-only, deterministic, and key-free. It never sends instruction contents anywhere.

Status: early alpha. This is an independent community tool, not an OpenAI product. It models the public Codex documentation and may lag behind Codex internals.

Why

Layered AGENTS.md files are useful, but discovery problems are hard to inspect: which project root was selected, which file won in a directory, whether a fallback was active, and where the byte budget was exhausted.

agents-md-debugger turns those questions into three small commands:

• scan: inventory instruction files and show same-directory selection/shadowing;
• explain: resolve the expected global and project instruction chain for a target;
• doctor: flag empty, shadowed, oversized, duplicated, and empty-scope files.

Install

Python 3.11+:

python -m pip install "git+https://github.com/2015709544-cell/agents-md-debugger.git@v0.1.1"
agents-md-debugger --version

Development install:

python -m pip install -e ".[dev]"

Uninstall:

python -m pip uninstall agents-md-debugger

Usage

agents-md-debugger scan .
agents-md-debugger explain src/api/user.py
agents-md-debugger explain src/api/user.py --json
agents-md-debugger doctor . --fail-on warning

The shorter amd-debug alias provides the same commands.

Try the included example

From this repository's root:

agents-md-debugger explain examples/demo-repo/src/api/user.py --project-root examples/demo-repo

The checked-in fixture produces this three-level chain (absolute paths above it are omitted):

Project instruction budget: 263/32768 bytes

Effective project instruction chain (root -> working directory):
  1. AGENTS.md [scope ./**, 113/113 bytes, loaded]
  2. src/AGENTS.md [scope src/**, 74/74 bytes, loaded]
  3. src/api/AGENTS.override.md [scope src/api/**, 76/76 bytes, loaded]

By default the tool reads these public Codex settings from $CODEX_HOME/config.toml (or ~/.codex/config.toml):

• project_doc_max_bytes;
• project_doc_fallback_filenames;
• project_root_markers.

It reports global instruction file metadata from $CODEX_HOME, but never prints file contents. Override settings without editing your configuration:

agents-md-debugger explain src/api/user.py \
  --max-bytes 65536 \
  --fallback TEAM_GUIDE.md \
  --root-marker .git

Discovery model

The implementation follows the current public Codex documentation:

1. Global scope selects the first non-empty AGENTS.override.md or AGENTS.md in the Codex home.
2. Project scope starts at the detected project root and walks to the target's directory.
3. Each directory selects at most one non-empty file in this order: AGENTS.override.md, AGENTS.md, configured fallback names.
4. Files merge root-to-target; later, more local instructions have higher precedence.
5. Empty files are skipped and the project byte budget can truncate the chain.

explain path/to/file models starting Codex from that file's parent directory. It does not claim to inspect a running or resumed Codex session.

Doctor diagnostics

Diagnostics are structural hints. The tool does not decide whether Markdown instructions conflict, whether they are good instructions, or whether Codex will follow them.

GitHub Action

Use the tagged release in another repository:

name: AGENTS.md doctor
on: [push, pull_request]

permissions:
  contents: read

jobs:
  doctor:
    runs-on: ubuntu-latest
    steps:
      - uses: actions/checkout@v4
      - uses: 2015709544-cell/agents-md-debugger@v0.1.1
        with:
          fail-on: warning

Known limitations

• This is a documentation-based preview, not Codex's internal resolver.
• It does not inspect session rehydration, TUI state, prompt assembly, symlink cycles, or I/O behavior inside Codex.
• explain uses the target's parent as the modeled working directory.
• Duplicate detection is intentionally conservative and line-based.
• Project-local configuration layers and managed configuration are not yet resolved; pass the effective config explicitly when needed.

Development

ruff check .
pytest
python -m build

Every behavior change should include a minimal filesystem fixture in a test. See CONTRIBUTING.md.

Roadmap

• compare preview output with a redacted Codex session log;
• explain project-local and managed configuration provenance;
• report I/O and symlink edge cases;
• emit shell-completion metadata and richer CI annotations.

Roadmap items are intentions, not shipped features.

References

• OpenAI: Custom instructions with AGENTS.md
• OpenAI: Advanced configuration
• AGENTS.md open format

License

MIT. See LICENSE.