ashishkaloge/agent-policy-map

Última atividade de origem registrada: 19 de julho de 2026 às 09:12

Catálogo do SkillsMP atualizado: 2 de novembro de 2026 às 18:02

skills coletadas: 11

Estrelas no GitHub: 0
Forks no GitHub: 0

Instalar com um assistente de IA

Copie este prompt para o assistente de IA que você usa.

Follow https://skillsmp.com/skill-install/prompt.md to install Agent Skills from https://github.com/ashishkaloge/agent-policy-map.

README do repositório

Arquivo de origem do README: README.md

Ler o README original

Expandir README completo

# Agent Policy Map [![CI](https://github.com/ashishkaloge/agent-policy-map/actions/workflows/ci.yml/badge.svg)](https://github.com/ashishkaloge/agent-policy-map/actions/workflows/ci.yml) [![Node.js 20+](https://img.shields.io/badge/Node.js-20%2B-339933)](https://nodejs.org/) [![MIT License](https://img.shields.io/badge/license-MIT-blue.svg)](LICENSE) **Compare what Codex CLI, Claude Code, Cursor IDE, and GitHub Copilot CLI can see for one file, side by side.** `agent-policy-map` is a local, read-only debugger for coding-agent instruction discovery, activation, ordering, and possible conflicts. It needs no account, API key, telemetry service, network service, or LLM. ![Agent Policy Map compares one target across Codex CLI, Claude Code, Cursor IDE, and GitHub Copilot CLI](docs/assets/agent-policy-map-demo.svg) ## Install Requires Node.js 20 or newer. Run from your project root, replacing `src/app.ts` with an existing file: ```bash npx agent-policy-map@0.1.1 compare src/app.ts --cwd . --root . ``` Or install the preview as a development dependency: ```bash npm install --save-dev agent-policy-map@0.1.1 npx agent-policy-map --help ``` Version 0.1.1 is a public preview. Review the supported surfaces and limitations below before relying on its results. ## 60-second demo from source ```bash git clone https://github.com/ashishkaloge/agent-policy-map.git cd agent-policy-map npm ci npm run build node dist/cli.js compare fixtures/copilot/basic/src/app.ts \ --cwd fixtures/copilot/basic \ --root fixtures/copilot/basic ``` The result is a four-column policy matrix. It shows, for example, that `AGENTS.md` is active for Codex, Cursor, and Copilot in this fixture. Claude loads its existing `CLAUDE.md` files instead. It also gives the documented reason behind every meaningful difference and, where an agent's documented discovery rules allow it, a suggested change with a link to that documentation, such as adding `@AGENTS.md` to `CLAUDE.md` so Claude Code imports it. Discovered instruction content is treated as data and is never printed, executed, or modified by `compare`. ## Support at a glance | Surface | What is modeled | Current evidence | Explicit boundary | | --- | --- | --- | --- | | Codex CLI | Project documents, same-directory overrides, configured fallbacks, and supported skill catalogs | Fixtures, CLI tests, and public-repository scan | Project `.codex/config.toml` controls, trust settings, and session overrides are not fully modeled; other Codex surfaces have no separate live-host parity proof | | Claude Code | Memory layers, AGENTS.md fallback and user instruction modes, imports, scoped rules, exclusions, and visible managed/user locations | Fixtures, CLI tests, and public-repository scan | Automatic memory, Claude Desktop, plugin availability, session/managed settings overrides, and files outside the selected root are not modeled | | Cursor IDE | `.cursor/rules/**/*.mdc`, references, activation modes, and nested `AGENTS.md` | Fixtures, CLI tests, and public-repository scan | Cursor CLI-specific `CLAUDE.md`, legacy `.cursorrules`, cloud rules, and account rules are not modeled | | GitHub Copilot CLI | Repository, path-specific, user, agent, and imported instructions | Fixtures, CLI tests, and public-repository scan | Copilot for VS Code, GitHub.com, cloud agent, and code review are not modeled | | JSON and TypeScript library | Versioned `PolicyMap`, diagnostics, assumptions, and context estimates | Built CLI and package smoke checks | Consumers must evaluate diagnostic levels; exit `0` does not mean conflict-free | “Current evidence” means deterministic tests pass for the modeled file formats and a real public checkout can be scanned. It does not prove that an agent loaded or followed an instruction in a live session. The tool is a standalone scanner, not an agent plugin or hook. You do not need to install Claude Code, Codex, Cursor, or Copilot to inspect their local policy files. ## Why this exists Coding-agent instructions are spread across formats such as `AGENTS.md`, `SKILL.md`, `CLAUDE.md`, `.claude/rules`, Cursor `.mdc` rules, and GitHub Copilot instructions. Each agent has different discovery, matching, import, and activation behavior. Finding the files is the easy part. The useful part is explaining: - which sources are active for one target and launch context; - which sources are conditional or manual; - which sources are excluded and why; - which user, managed, account, or session state is not visible locally; - which sources duplicate or possibly contradict each other. The result is a static projection of documented local policy. It is not a runtime oracle and cannot prove that a model followed an instruction. ## Real use cases Use `agent-policy-map` to: - debug why a nested `AGENTS.md` applies to one package but not another; - check whether a Cursor or Copilot path rule matches a target file; - find missing imports, import cycles, duplicate guidance, and possible package-manager conflicts; - compare how Codex, Claude Code, Cursor IDE, and Copilot CLI see the same project; - estimate how much instruction context could be loaded; - produce versioned JSON for an audit script or CI report; - explain one source by its stable ID without printing its full content. ## Compare four agents Run one command after completing the source setup above: ```bash node dist/cli.js compare src/cli.ts --cwd . --root . ``` The matrix separates meaningful local differences from unavailable account, managed, and session state. `—` means that the modeled adapter did not discover that source path; it does not mean the source was explicitly excluded. `compare` models the exact supported surfaces shown in its header: Codex CLI, Claude Code, Cursor IDE, and GitHub Copilot CLI. It does not claim parity with the Codex app, Claude Desktop, Cursor CLI, or Copilot for VS Code. ### Suggested changes For each meaningful difference in a project file, `compare` suggests a change for every agent that does not load the file, based only on that agent's documented discovery rules. Each suggestion is printed under the agent it is for and cites its source. This fixture has `CLAUDE.md` files, so Claude's default mode does not load `AGENTS.md` directly: ```bash node dist/cli.js compare fixtures/copilot/basic/src/app.ts \ --cwd fixtures/copilot/basic \ --root fixtures/copilot/basic ``` ```text WHY THEY DIFFER ... ./AGENTS.md codex active [codex-agents-md]: AGENTS.md discovered on the root→cwd path (./) claude —: not discovered by this modeled adapter suggestion: add `@AGENTS.md` to ./CLAUDE.md, creating it if needed [5] cursor active [cursor-agents-md]: AGENTS.md applies to its directory subtree (./) copilot active [copilot-agents-md]: AGENTS.md found on the repository-root-to-target search path ... Suggestions apply documented discovery rules, for when you want that agent to read the file. They do not show that an agent will load or follow it. ... [5] https://code.claude.com/docs/en/memory#share-one-file-with-other-coding-tools ``` Suggestions cover `@` imports for Claude Code and Copilot CLI, the AGENTS.md launch path, `project_doc_fallback_filenames`, overrides, and the project-doc byte cap for Codex CLI, `.mdc` rules for Cursor IDE, and each agent's own path-scoped rule format. A suggestion applies only if you want that agent to read the file; some differences are intentional. Files outside the project, manual-only rules, personal files such as `CLAUDE.local.md`, Codex `AGENTS.override.md`, and skills get no suggestion. ### Claude Code instruction modes The adapter models current Claude Code (v2.1.285+) with its built-in AGENTS.md plugin available. Automatic AGENTS.md support began in v2.1.277. In the default `claude-md-or-agents-md` mode, it reads `AGENTS.md` and `.claude/AGENTS.md` on the root-to-launch path when that path contains no `CLAUDE.md`, `.claude/CLAUDE.md`, or `CLAUDE.local.md`. Nested AGENTS.md files are conditional and yield to Claude files in their own directory. Existing user memory and `.claude/rules/` do not disable this fallback. With `--include-user`, the adapter reads `instructionFiles` from `~/.claude/settings.json`, under the `cc-plugin-agents-md@builtin` or legacy `agents-md@builtin` entry in `pluginConfigs`. It models the default, `claude-md-and-agents-md`, `claude-md`, and `managed-only` modes. Project and local settings cannot change this option. Without user-scope opt-in, the documented default is assumed. Plugin availability, session `--settings`, managed settings overrides, and instruction files above `--root` remain unknown. In particular, a Claude file above that boundary can suppress automatic AGENTS.md fallback. See [Claude's instruction-file documentation](https://code.claude.com/docs/en/memory#agentsmd) and the inspection assumptions before relying on the result. ## Commands ### `compare` Compare documented instruction discovery for one target across all four supported surfaces. ```text agent-policy-map compare <target> [options] ``` ```bash npm run cli -- compare apps/api/src/auth.ts --cwd . --root . ``` `compare` does not accept `--agent` because all supported adapters are always checked. It also does not accept `--show-content`; the comparison contains metadata, evidence-backed reasons, and documented suggestions, never instruction bodies. The target resolves from the current directory, not from `--cwd`, and must be an existing file. When a relative target exists only under `--cwd`, the error names the path to pass instead. For example, `compare src/auth.ts --cwd apps/api` run from `fixtures/codex/nested` fails with: ```text error: target "src/auth.ts" was not found from the current directory. Targets resolve from the current directory, not from --cwd; pass "apps/api/src/auth.ts" instead ``` ### `inspect` Resolve instruction state for one target file. ```text agent-policy-map inspect <target> --agent <codex|claude|cursor|copilot> [options] ``` Local repository form: ```bash npm run cli -- inspect apps/api/src/auth.ts \ --agent codex \ --cwd apps/api \ --root . ``` `--cwd` models the agent's launch working directory. This is especially important for Codex. If omitted, `inspect` uses the target's directory and records that assumption. ### `discover` List instruction sources discoverable from a launch context without choosing one concrete target file. ```text agent-policy-map discover --agent <codex|claude|cursor|copilot> [options] ``` ```bash npm run cli -- discover \ --agent cursor \ --cwd fixtures/cursor/modes \ --root fixtures/cursor/modes ``` This example deliberately limits discovery to the public Cursor fixture so the result demonstrates activation states without mixing in unrelated repository fixtures. Rules that need a target path are reported as `conditional` in discovery mode. `discover` does not claim that every listed source is active for every file. ### `explain` Re-run an inspection context and explain one source from its stable ID. ```text agent-policy-map explain <source-id> --agent <agent> --target <file> [options] ``` For example, after inspecting this repository: ```bash npm run cli -- explain codex:codex-agents-md:AGENTS.md \ --agent codex \ --target src/cli.ts \ --cwd . \ --root . ``` Use the same agent, target, cwd, root, and user-scope options that produced the ID. A source ID only resolves within its matching inspection context. ## Options | Option | Meaning | | --- | --- | | `--agent <name>` | Required for `inspect`, `discover`, and `explain`. `compare` checks all four agents. | | `--cwd <path>` | Simulated launch working directory. It does not change how the target resolves. | | `--root <path>` | Containment root. Defaults to the nearest `.git` ancestor, otherwise cwd. | | `--include-user` | Also inspect supported user, home, managed, and configured custom locations. Off by default. | | `--show-content` | For `inspect` and `discover`, include available bounded local content after best-effort credential redaction. Not supported by `compare`. | | `--json` | Emit machine-readable JSON. | | `-h`, `--help` | Show CLI help. | `--include-user` expands the read scope. For Codex, it is also required to read user configuration that changes fallback instruction names or the project-doc byte cap, or disables a discovered skill. Review the reported assumptions when it is omitted. `explain` intentionally returns source metadata and related diagnostics without full instruction content. Use `inspect` or `discover` with `--show-content` when content is required. ## Exit behavior - Help and a completed comparison, inspection, discovery, or explanation exit with code `0`. - Invalid commands, missing required arguments, unknown agents, and source IDs that do not exist in the supplied context exit non-zero. - A target for `compare`, `inspect`, or `explain` that does not exist, or that is a directory, exits non-zero. Targets resolve from the current directory. - Diagnostics do not change the process exit code, including diagnostics whose level is `error`. Automation should inspect `diagnostics[].level` and `diagnostics[].code` in JSON instead of treating exit code `0` as a clean policy result. ## Honest output states There is not always one universal final instruction set. Some agents load nested rules only after reading a matching file. Some let the model select a rule or skill. Some use account, organization, managed, or session state that a local scanner cannot inspect. Every source uses one of these states: | State | Meaning | | --- | --- | | `active` | Documented discovery and matching make it active for the simulated context. | | `conditional` | The agent, model, target, or later file access decides whether it loads. | | `manual` | Explicit invocation or selection is required. | | `excluded` | A documented path, scope, setting, or replacement rule excludes it. | | `unknown-external` | Relevant account, organization, managed, approval, or session state is unavailable locally. | Each source has an ID, path, kind, state, and evidence-backed match reason. On-disk sources also normally include byte and line counts plus a short content hash. Synthetic or unavailable external sources do not contain file metadata. ## Example output Codex discovers project documents along the path from the root to the launch cwd: ```bash npm run cli -- inspect fixtures/codex/nested/apps/api/src/auth.ts \ --agent codex \ --cwd fixtures/codex/nested/apps/api \ --root fixtures/codex/nested ``` Relevant output: ```text Agent: codex (Codex CLI) Target: ./apps/api/src/auth.ts Launch cwd: ./apps/api ACTIVE 1. ./AGENTS.md [codex-agents-md] reason: AGENTS.md discovered on the root→cwd path (./) 2. ./apps/api/AGENTS.md [codex-agents-md] reason: AGENTS.md discovered on the root→cwd path (./apps/api) UNKNOWN-EXTERNAL - /etc/codex/skills [codex-skill-catalog] reason: Administrator skills are outside the project and were not inspected without --include-user - [Codex bundled system skills] [codex-skill-catalog] reason: Bundled system skills come from the running Codex host and cannot be inspected as a stable local catalog DIAGNOSTICS - [warn] possible-conflict: Possible package manager conflict: ./AGENTS.md endorses npm; ./apps/api/AGENTS.md endorses pnpm ASSUMPTIONS - Global Codex instructions ($CODEX_HOME/AGENTS.override.md or AGENTS.md) were not inspected. Pass --include-user to include them. - User skills ($HOME/.agents/skills), administrator skills (/etc/codex/skills), and user skill enable/disable settings were not inspected. Pass --include-user to include locally visible catalogs and $CODEX_HOME/config.toml. ``` ## JSON output Add `--json` to `compare`, `inspect`, `discover`, or `explain`. When invoking the local CLI through an npm script, pass `--silent` so npm's script banner does not precede the JSON on standard output: ```bash npm --silent run cli -- compare src/cli.ts --cwd . --root . --json ``` `compare` emits a versioned `PolicyComparison` containing: - the shared target, launch cwd, containment root, and user-scope choice; - a summary count of meaningful local differences; - source paths grouped into per-agent arrays so duplicate entries are not lost; - a `suggestions` array on each source row, empty unless a documented change applies, with the `agent`, a short `text`, and the documentation `reference`; - each surface's assumptions, diagnostics, and context estimate. `suggestions` was added to version 1 without changing any existing field. `inspect` and `discover` emit a versioned `PolicyMap` containing: - `version`, `agent`, `surface`, `target`, and `cwd`; - explicit `assumptions`; - normalized `sources`; - structured `diagnostics`; - byte, line, and approximate-token context totals. Full instruction content is omitted by default. `compare` never returns it. For `inspect` and `discover`, `--show-content --json` adds a `contents` object for loadable local sources after best-effort redaction. `explain` does not return full content. ## Redaction and sensitive content Redaction is a defense-in-depth heuristic, not a secret scanner. It recognizes common credential formats and sensitive key/value assignments, but it cannot guarantee that arbitrary private text or every secret format is removed. - Do not use `--show-content` unless full instruction text is necessary. - Do not publish or attach output created with `--show-content` without reviewing it. - Keep `--include-user` off when project-only evidence is sufficient. - Treat paths, hashes, excerpts, and diagnostics as potentially sensitive even when full content is hidden. ## Supported surfaces Version 1 models: - Codex CLI project documents, overrides, configured fallbacks, and supported skill catalogs; - Claude Code memory, imports, scoped rules, and visible managed/user layers; - Cursor IDE `.cursor/rules/**/*.mdc`, references, and nested `AGENTS.md`; - GitHub Copilot CLI repository, path-specific, user, agent, and imported instructions. Current gaps are reported instead of guessed: - Codex project `.codex/config.toml` controls, trust settings, and session overrides are not fully modeled. - Claude automatic memory and Claude Desktop are not modeled. - Cursor CLI-specific root `CLAUDE.md`, legacy `.cursorrules`, cloud rules, and account rules are not modeled by the Cursor IDE adapter. - Copilot for VS Code, GitHub.com, cloud agent, and code review are separate surfaces and are not modeled by the Copilot CLI adapter. - Gemini CLI, Antigravity, OpenCode, Windsurf, and other agents do not yet have adapters. A new adapter must be grounded in official documentation for its exact surface. Official behavior references: - [Codex AGENTS.md configuration](https://learn.chatgpt.com/docs/agent-configuration/agents-md) - [Codex skills](https://learn.chatgpt.com/docs/build-skills) - [Claude Code memory](https://code.claude.com/docs/en/memory) - [Cursor rules](https://cursor.com/docs/rules) - [Copilot CLI custom instructions](https://docs.github.com/en/copilot/how-tos/copilot-cli/customize-copilot/add-custom-instructions) - [Copilot custom instruction precedence](https://docs.github.com/en/copilot/concepts/prompting/response-customization) ## Safety properties - Inspection is read-only. Discovered files are never rewritten or executed. - Instruction content is treated as untrusted data, not as commands for this program. - Inspection makes no network requests. - Project reads stay inside the resolved containment root. Supported user and managed locations require `--include-user`. - Direct reads, imports, directory traversal, symlinks, file size, total files, total bytes, and recursion are bounded or containment-checked. - Full content is hidden by default and possible credentials are redacted from excerpts and diagnostics. - Uncertain or externally controlled state is reported instead of inventing precedence. ## Library API The package also exports the comparison and inspection pipelines, renderers, adapter registry, redaction helpers, and public TypeScript types. ```ts import { compare, renderJson, type PolicyComparison } from "agent-policy-map"; const map: PolicyComparison = compare({ target: "src/cli.ts", cwd: ".", root: ".", }); console.log(renderJson(map)); ``` The programmatic API follows the same local, read-only contract as the CLI. Consumers should check `version` before relying on the JSON shape. Unlike the CLI, the library accepts a target that does not exist yet, so callers can check a planned file path. It records that as an assumption. ## Architecture This is one Node.js and TypeScript package, not a monorepo. Agent-specific behavior is isolated behind one normalized policy-map contract. `compare` composes those same results and does not add separate discovery logic. Shared safety, matching, diagnostics, and rendering stay agent-independent. See [ARCHITECTURE.md](ARCHITECTURE.md) for the data flow, module boundaries, extension rules, and test strategy. ## Develop ```bash npm run typecheck npm run lint npm run build npm run test npm run smoke npm run smoke:pack npm run check ``` `npm run check` runs typecheck, lint, build, tests, and the built-CLI smoke test in order. `npm run smoke:pack` separately packs the project, installs the tarball in a temporary consumer, and verifies the installed executable and library export. ## Non-goals Version 1 does not: - modify, synchronize, or generate instruction files; - guarantee that an agent or model followed an instruction; - read private account or organization state through remote APIs; - claim complete semantic understanding of natural-language contradictions; - calculate an exact model token count; - model every editor, coding agent, or Copilot surface; - publish, upload, or send inspection results anywhere. ## Contributing Read [CONTRIBUTING.md](CONTRIBUTING.md) before opening a pull request. New discovery or precedence behavior must cite current official documentation and include public fixtures for matches, exclusions, ambiguity, and failure cases. ## Security Report unintended reads, path escapes, content leakage, content execution, or security-relevant policy results privately. See [SECURITY.md](SECURITY.md). ## Related projects - [coding-agent-guidelines](https://github.com/ashishkaloge/coding-agent-guidelines) - [production-launch-prompts](https://github.com/ashishkaloge/production-launch-prompts) - [awesome-agentic-engineering](https://github.com/ashishkaloge/awesome-agentic-engineering) - [awesome-agent-instructions](https://github.com/ashishkaloge/awesome-agent-instructions) Maintainers: see [npm releases](RELEASING.md) for the trusted publishing workflow. ## License [MIT](LICENSE)

Carregando o README do repositório…

Skills neste repositório

classificação pendente