agents-doctor 0.2.2

Show what an AI coding agent actually loads from your AGENTS.md files - including the parts silently truncated away.

agents-doctor

See what a coding agent actually loads from your AGENTS.md files — including the parts it silently throws away.

Status: beta. The project is release-tested and intended for real repositories, but its Codex compatibility is explicitly versioned and should be verified when upstream loader behaviour changes.

Agents concatenate AGENTS.md files from the repository root down to your working directory, then cut the result off at a byte budget. The cut is not announced in the UI. Because root-level instructions are concatenated first, the instructions closest to the code you are editing — the specific ones — are the first to disappear.

agents-doctor reproduces that loader and tells you exactly what survives.

---

Why this matters

The failure mode is structural: a large root-level instruction file can spend the entire budget before the agent reaches the instructions closest to the code being changed. This is hard to spot by reading files individually, especially in a monorepo.

Run agents-doctor explain path/to/package in the repository you maintain to see the exact load order, retained bytes, and any instructions that never reach the agent. The report is intentionally based on the checked-out files rather than on fixed claims about third-party repositories, so it remains reproducible as those repositories evolve.

Install

Install the published package from PyPI with an isolated application environment:

pipx install agents-doctor       # or: uv tool install agents-doctor

# Development checkout
git clone https://github.com/satoissei/agents-doctor.git
cd agents-doctor
pipx install .                  # or: uv tool install .

Try without installing

Run the latest PyPI release in a temporary pipx environment from the repository you want to check:

pipx run agents-doctor check

Use

agents-doctor check      # report problems, exit 1 if any        (for CI)
agents-doctor explain    # what gets loaded in this directory     (for humans)
agents-doctor budget     # budget pressure across the repository  (for monorepos)
agents-doctor explain --format json  # machine-readable load plan
agents-doctor --codex-config ~/.codex/config.toml explain  # use Codex's loader settings
agents-doctor check --format sarif > agents-doctor.sarif  # GitHub Code Scanning

In CI

For a local checkout of this repository, use the composite action directly:

- uses: actions/checkout@3d3c42e5aac5ba805825da76410c181273ba90b1 # v7.0.1
- uses: ./

Other repositories should use the published release tag rather than a development branch:

- uses: satoissei/agents-doctor@v0.2.2

As a pre-commit hook

Use the published release tag as rev:

repos:
  - repo: https://github.com/satoissei/agents-doctor
    rev: v0.2.2
    hooks:
      - id: agents-doctor

Using or evaluating the project in a real repository? Please submit a short public adoption report, including failed or incomplete evaluations. The project collects no telemetry and counts only verifiable, voluntarily shared evidence; see docs/adoption.md.

If you write instructions in Japanese, Chinese or Korean

The budget is counted in bytes, not characters, and CJK text is three bytes per character in UTF-8. The same 32 KiB that holds roughly 30,000 characters of English holds only about 10,000 of Japanese — you hit the ceiling three times sooner.

Worse, the cut happens at a byte offset and the result is then decoded leniently, so a cut landing mid-character replaces it with U+FFFD. agents-doctor reports character counts alongside byte counts, tells you how many characters your budget actually holds, and flags cuts that corrupt a character.

Rules

ID	Name	Default	Detects
AD001	lost-instructions	error	Instructions cut short or never loaded because the budget ran out
AD002	broken-path-reference	error	A reference to a file or directory that no longer exists — a rename or deletion the instructions never caught up with
AD004	empty-instructions	warning	A file that looks like guidance but provides none

More rules are planned — see ROADMAP.md. Rule numbering is stable and reserved, so gaps are intentional.

Configuration

Everything is optional. Put settings in .agents-doctor.toml or under [tool.agents-doctor] in pyproject.toml:

max_bytes = 32768          # match your agent's project_doc_max_bytes
exclude = ["node_modules"] # directories to skip while discovering files
ignore_paths = ["dist/*"]  # references that are allowed not to exist
fallback_filenames = []    # extra per-directory filenames your agent reads

[rules]
AD002 = "warning"          # or "off"

Use --root-marker NAME when your agent's project root is identified by a marker other than .git. Repeat the option to provide more than one marker:

agents-doctor --root-marker .hg explain path/to/package

max_bytes = 0 is accepted and models a Codex configuration that disables project instructions entirely.

To avoid copying Codex settings into a second configuration file, pass the Codex config.toml directly with --codex-config. The tool reads project_doc_max_bytes, project_doc_fallback_filenames, and project_root_markers; unrelated Codex settings are ignored. Command-line options such as --max-bytes and --root-marker take precedence.

For uploading SARIF to GitHub Code Scanning, see docs/github-code-scanning.md.

Privacy and safety

The CLI runs locally: it makes no network requests, sends no telemetry, and does not execute instructions it reads. Its reports contain repository-relative paths, file sizes, and path references, so review output before sharing it publicly. The loader model follows symlinks; do not run it against an untrusted checkout because an instruction-file symlink can point outside that repository.

See PRIVACY.md for the data-handling boundary and SECURITY.md for private vulnerability reporting.

What it mirrors

The simulation is a behavioural reimplementation of codex-rs/core/src/agents_md.rs from openai/codex, verified against commit 2b5bdcf:

1. The project root is the nearest ancestor holding .git.
2. Directories are visited root-first, down to the working directory.
3. Each directory contributes one file: AGENTS.override.md, then AGENTS.md, then any configured fallbacks.
4. Files are appended until project_doc_max_bytes (default 32,768) is spent. A file that does not fit is cut at a byte offset; once nothing remains, later files are never read.
5. A file that is blank after cutting is skipped without consuming budget.

Symlinks are followed, because the loader follows them. This matters: several well-known repositories ship AGENTS.md as a symlink, where measuring the link reports a dozen bytes instead of the kilobytes actually loaded.

Scope of verification: these behaviours were read from source, not observed by instrumenting a running agent. Byte sizes quoted above are measured; load outcomes are computed from the algorithm. If you find a divergence from real behaviour, please open an issue — that is the most valuable bug report this project can receive.

The supported behaviour, its upstream reference, and the update process are recorded in docs/codex-compatibility.md.

Project operations

See CONTRIBUTING.md for development, SUPPORT.md for safe help requests, GOVERNANCE.md for maintainer decisions, and docs/maintainer-playbook.md for the operating cadence. Bug reports that show a real repository where the simulation is wrong are especially welcome.

日本語の概要は README.ja.md にあります。

License

MIT. See NOTICE for upstream compatibility attribution.