agent_context_doctor is an offline, read-only CLI that explains how
repository-local instructions, settings, and skills are discovered for Codex,
Claude Code, and GitHub Copilot.
Use it when an agent appears to ignore repository guidance and you need a privacy-safe inventory of files that are active, shadowed, ignored, on demand, trust gated, or outside the selected host's scope.
Agent hosts use different filenames, directories, ancestry rules, and loading conditions. A valid file in the wrong location can silently have no effect, while a project settings file may be gated by workspace trust. This tool makes that static discovery model reviewable without starting a host or sending any repository data to a model.
Download a release archive for your platform, or build with Rust 1.85 or newer:
cargo install --path . --lockedagent_context_doctor audit . --host all
agent_context_doctor audit crates/core --host codex --format jsonThe audit is limited to the explicitly selected Git repository. Reports contain normalized relative paths and stable rule identifiers, never file contents or absolute paths.
Exit codes:
0: resolution completed without diagnostic findings;1: resolution completed with ignored, shadowed, trust-gated, or out-of-scope artifacts;2: the audit could not be completed safely.
| Host | Instructions | Settings | Skills |
|---|---|---|---|
| Codex | ancestry AGENTS.override.md / AGENTS.md |
ancestry .codex/config.toml |
ancestry .agents/skills/*/SKILL.md |
| Claude Code | ancestry and nested CLAUDE.md variants |
selected working directory settings | ancestry and nested .claude/skills/*/SKILL.md |
| GitHub Copilot code review | repository, path-scoped, and AGENTS.md instructions |
not modeled | not modeled |
--host copilot is deliberately limited to the GitHub Copilot code review
surface. The report identifies its rules as repository-v0.1. Host behavior
evolves; open an issue with an authoritative source when a rule needs updating.
The tool does not parse, retain, or report instruction or settings contents. Its
only content access is a bounded whitespace probe because Codex skips empty
AGENTS.override.md files. It does not inspect home-directory configuration,
run a host or model, infer semantic conflicts or runtime prompt order, decide
workspace trust, repair files, use the network, or emit telemetry.
See the product specification for the exact contract.
If this project saves you time, you can support continued maintenance through the funding link shown by GitHub.