agent-doc-system 1.0.0 → 1.2.0

package/CHANGELOG.md

@@ -0,0 +1,269 @@
+# Changelog
+
+All notable changes to AgentDoc are documented here.
+
+The format follows [Keep a Changelog](https://keepachangelog.com/en/1.1.0/), and
+the project uses [Semantic Versioning](https://semver.org/spec/v2.0.0.html). The
+compiler version is the package version: a graph compiled by one release is
+refused by the next with `AGENTDOC_COMPILER_MISMATCH` until it is recompiled.
+
+## [Unreleased]
+
+## [1.2.0] - 2026-10-27
+
+Rust and Python are now supported at the same level as Node and Go. Before
+this release, a Cargo crate or a Python package was invisible to the compiler:
+`agentdoc audit` reported zero units, and a hand-written descriptor produced a
+component with no language, no dependency edges and no verification commands.
+
+The release also fixes how a change is routed. `impact` and `query` now follow
+the components that depend on the changed one, not the ones it depends on, and
+recommend those dependents' checks.
+
+### Added
+
+- **`rust` adapter** ([#17]). Reads `Cargo.toml` and never runs Cargo or reads
+  `Cargo.lock`.
+  - A `[package]` is a unit. `[[bin]]`, `src/main.rs` or `src/bin/` make it
+    deployable; `[lib]` or `src/lib.rs` make it importable.
+  - Workspace members come from every `[workspace]` manifest, including one
+    below the repository root (`rust/Cargo.toml` next to other languages).
+  - Sibling crates become `buildDependsOn` / `testDependsOn` edges when the
+    manifest says so: a `path`, `workspace = true` resolved through the nearest
+    `[workspace.dependencies]`, or a `[patch.<registry>]` / `[replace]` path
+    redirect inside the repository. `use`, `extern crate` and `<crate>::` paths
+    add provenance to those edges but never create one, because Rust cannot
+    use a crate the manifest does not declare.
+  - Verification commands: `cargo test -p <name>` and `cargo clippy -p <name>`.
+  - Resource vocabulary: sqlx, diesel (`diesel.toml`), tokio-postgres,
+    deadpool, sea-orm, the `redis` crate and `aws-sdk-s3`.
+- **`python` adapter** ([#18]). Reads `pyproject.toml` (PEP 621 and Poetry),
+  `setup.cfg`, `setup.py` and `requirements*.txt`. It never runs Python,
+  resolves a lockfile or inspects a virtualenv.
+  - Shape comes from evidence. An entrypoint (`[project.scripts]`,
+    `__main__.py`, `manage.py`, an ASGI/WSGI app object) makes a unit
+    deployable; a build backend plus a package makes it importable.
+  - uv workspace members come from the repository root and from workspaces
+    below it. A member with no build backend is still a unit.
+  - Sibling packages become edges from `[project] dependencies`,
+    `[tool.uv.sources]`, Poetry `path` dependencies, `-e <dir>` requirements,
+    and `import` / `from` statements. Dev groups and dev/test extras give
+    `testDependsOn`. An import that the importing unit provides itself, such
+    as a local `utils.py`, is not read as a sibling.
+  - Verification commands are derived only from evidence: `pytest`,
+    `python -m unittest discover` for unittest-only suites, `ruff check` and
+    `mypy`. `uv run` / `poetry run` is prefixed only when the lockfile names
+    the tool.
+  - Also derived: `requiresPython` when declared, and Alembic migrations as
+    database evidence.
+  - Resource vocabulary: SQLAlchemy, asyncpg, psycopg, Django `DATABASES`,
+    `redis.Redis` and boto3.
+- `agentdoc init` enables `rust` and `python` when it finds their manifests.
+  `agentdoc scaffold` names crates and distributions from their manifests and
+  types them from their artifact shape: `service`, `library`, or `other` when
+  neither is proven. A virtual Cargo or uv workspace manifest is not scaffolded.
+- A reproducible benchmark, `scripts/bench/run.mjs` ([docs/BENCHMARK.md]).
+  It clones four public repositories at pinned commits, runs a fresh
+  `init` → `scaffold` → `compile`, counts every hand edit the setup still
+  needs (zero for all four), checks that the graph is byte-identical across
+  compiles, and measures `query` answers.
+- The routing evaluation reports **check selectivity**: the share of the
+  repository's checks that the answer for a change contains ([#43]). The
+  baseline's verification recall is now scored with the same rule as the
+  catalog's. It is 1.0 by construction, because the baseline runs every check.
+- Rust and Python sample repositories (`fixtures/e-rust-workspace`,
+  `fixtures/f-python-workspace`), checked in CI and used to test context
+  routing ([#23]).
+
+### Changed
+
+- **Change routing follows dependents** ([#30], [#25]).
+  - A change to a component reaches the components that depend on it,
+    transitively, through reverse `buildDependsOn` / `testDependsOn` edges.
+    Before, `impact` also listed the components it depends on as affected,
+    and did not follow transitive dependents.
+  - Dependencies stay visible as typed relations in `query`. They are no
+    longer reported as affected components.
+  - `query` and `impact` recommend the dependents' checks after the changed
+    component's own: every tier for build dependents, unit checks only for test
+    dependents.
+  - `impact` adds `recommendedVerificationTotal` and
+    `recommendedVerificationTruncated`, so a capped list is never mistaken for
+    the whole.
+- **`query` markdown shows an entity's context** ([#28]). Resources and
+  Systems now list their `docs`, `CONSTRAINTS` and `runbooks`, as components
+  already did. Constraints are never capped by default. Docs and runbooks show
+  12 paths each, and `--budget docs=…,constraints=…,runbooks=…` sets explicit
+  caps with a `[truncated: x of y]` marker.
+- **`discovery.notUnitRoots` excludes a whole subtree from inferred units**
+  ([#27]). An adapter marker or a scaffold proposal at or below an excluded
+  root is no longer a unit. A hand-written descriptor below an excluded root
+  is still honoured. A descriptor exactly at an excluded root is refused with
+  a message that names the exclusion.
+- **GitHub Actions checks are scoped to the components they run for**
+  ([#29]). A step becomes a component check only when its command is a
+  recognised test, lint or build invocation, scoped by `working-directory`,
+  `cd <dir> &&`, a path argument, a package selector (`--filter`, turbo, nx,
+  `make -C`) or a `paths:` filter. Repository scripts whose names say what
+  they check (`./scripts/test.sh`, `bash scripts/lint.sh`) are recognised
+  with the same scoping ([#40]). Anything else is a repository gate.
+  - Before, every step was attached to every component whose path or package
+    name appeared anywhere in the workflow, including multi-line ops scripts.
+    A shell `test -n` was also recorded as a unit test.
+  - `impact` on a workflow file now names the components and gates derived
+    from it, plus their dependents' checks.
+- **The TOML reader conforms to TOML 1.0** ([#18], [#31]). It reads
+  `Cargo.toml`, `pyproject.toml` and `wrangler.toml`, and passes every case of
+  the official [toml-test](https://github.com/toml-lang/toml-test) 1.0.0
+  suite: 208 valid and 501 invalid documents. Before, it rejected valid
+  documents, such as multi-line strings and lowercase date-time delimiters,
+  which lost every fact in the file. It also accepted invalid ones, such as a
+  trailing comma in an inline table or a table declared twice. Invalid TOML
+  is rejected with `AGENTDOC_YAML_PARSE`.
+
+### Fixed
+
+- **A unit nested inside another unit no longer leaks its facts into the
+  parent** ([#21]). Every per-unit scan stops at nested units, so each file
+  belongs to the deepest unit that contains it. Before, an enclosing unit also
+  claimed its nested unit's health routes, ports, capabilities, resource and
+  migration evidence, imports, contract consumers, TypeScript routes, Ansible
+  playbooks and test markers. This affects any layout with nested units, such
+  as a service containing a tool package, a Cargo root package with member
+  crates below it, or a uv workspace root with its own `[project]`.
+- An Alembic `alembic/versions` directory was listed as a migration location
+  but never recognised. It now counts as migration evidence. The evidence file
+  is a qualifying `alembic.ini` or `[tool.alembic]` configuration when one
+  exists, and otherwise a revision file.
+- A changed migration, including an Alembic revision, now lists the logical
+  database its component uses in `impact`'s `affectedResources` ([#30]).
+  Before, the list stayed empty.
+- The audit no longer reports a document as orphaned when it is linked only
+  from a Resource's or System's `context` ([#28]).
+- `agentdoc scaffold --write` no longer writes descriptors into directories
+  listed in `discovery.notUnitRoots` ([#27]).
+- **A fresh setup compiles without hand edits** ([#37], [#38]).
+  - `scaffold` now types every unit from its artifact evidence: `service`,
+    `library`, or `other` when neither is proven. Before, a Node app without
+    `exports` or `main`, or a Go module, was written as `library` and failed
+    `validate` straight away.
+  - `init` now discovers descriptors with the `*/**/agentdoc.yaml` glob
+    instead of an explicit list, and `scaffold` covers every unit an adapter
+    declares, including workspace members under `examples/`.
+  - A `setup.py` unit is named after the distribution it declares, and the
+    `python` adapter reads `name`, `python_requires` and `install_requires`
+    inside a one-line `setup(…)` or `setuptools.setup(…)` call ([#41]).
+  - On the four public repositories used for benchmarking (tokio-rs/axum,
+    pydantic/pydantic-ai, fastapi/full-stack-fastapi-template and
+    t3-oss/create-t3-turbo), `init` + `scaffold --write` + `compile` now pass
+    with no edits other than choosing each component's placement.
+
+### Upgrading
+
+- The compiler version is now `1.2.0`. A graph committed with 1.1.0 fails
+  `agentdoc check` with `AGENTDOC_COMPILER_MISMATCH`. Run `agentdoc compile`
+  and commit the graph.
+- An existing catalog keeps its adapter list. To use the new adapters, add
+  `rust` and/or `python` to `adapters:` in `agentdoc/agentdoc.config.yaml`,
+  then run `agentdoc audit`: units that were invisible appear as
+  `AGENTDOC_COVERAGE_ZERO` until they have a descriptor.
+  `agentdoc scaffold --write` writes those descriptors.
+- If your repository has nested units, recompiling removes facts that were
+  wrongly attributed to an enclosing component. Review the affected facts,
+  dependency edges and verification commands, and any descriptor or document
+  that relied on them.
+- Tools that read `impact --json` will see fewer `affectedComponents`
+  (dependencies are gone) and more `recommendedVerification` entries
+  (dependents' checks).
+- A unit inferred below a `notUnitRoots` entry, with no descriptor, is no
+  longer reported as `AGENTDOC_COVERAGE_ZERO`. If it should be a component,
+  write its descriptor, or remove the directory from `notUnitRoots`.
+- If your workflows run checks for specific components, recompiling changes
+  which checks are attached to which component. Review the diff: checks that
+  were attached only because a path was mentioned elsewhere in the workflow
+  move to repository gates.
+- Recompiling always updates the compiler version, source commit and input
+  hash. Review the other derived changes, especially in repositories with
+  nested units, Alembic migrations or Wrangler TOML manifests: a manifest the
+  old TOML reader could not parse now yields facts.
+
+## [1.1.0] - 2026-10-18
+
+### Added
+
+- **MIGRATE gates normative requirements and section coverage** ([#5]).
+  Documentation checks now state their real scope, and the strict gate refuses
+  unverified requirements.
+- **`ansible` adapter** ([#6]). Reads playbooks without running anything and
+  derives installed paths and managed services. A requirement with
+  `satisfiedBy: [ansible]` and the new `expects` field
+  (`installs: <path>` | `service: <name>`) is checked against them, so a missing
+  install ends in a conflict instead of `AGENTDOC_REQUIREMENT_UNVERIFIED`.
+- `agentdoc <command> --help` lists each command's flags and exit codes ([#7]).
+
+### Changed
+
+- **The CLI rejects unknown flags** ([#7]) and names the closest valid flag. A
+  typo such as `check --strcit` no longer silently runs a non-strict check.
+  Scripts that passed flags a command does not accept now exit 2.
+
+### Fixed
+
+- A committed graph passes CI ([#11]). `check` no longer fails with
+  `AGENTDOC_GRAPH_FRESHNESS` after `compile` on a clean checkout, where the
+  graph file itself was counted as a changed input.
+- A nested `.github/workflows` directory no longer fails `validate` ([#12]).
+  Only the repository-root directory is read, as GitHub does.
+
+### Upgrading
+
+- A graph committed with 1.0.0 fails `agentdoc check` with
+  `AGENTDOC_COMPILER_MISMATCH` until it is recompiled.
+
+## [1.0.0] - 2026-09-29
+
+First public release.
+
+### Added
+
+- The machine-installed `agentdoc` CLI for Node 22, 24 and 26, with zero
+  runtime dependencies.
+- A deterministic repository graph: repeated compiles of the same tree are
+  byte-identical.
+- Evidence separation between AUTHORED documentation, DERIVED facts extracted
+  from code, and OBSERVED runtime evidence.
+- A fail-closed authority model: contradictions surface as conflicts and are
+  never silently overwritten.
+- `agentdoc query`, which routes minimal, budgeted context for a task, and
+  `agentdoc impact`, which reports what a change forces you to keep consistent.
+- CREATE and MIGRATE workflows through `init`, `audit`, `scaffold`,
+  `validate`, `compile` and `check`.
+- `agentdoc doctor` for machine diagnostics, and `agentdoc doctor --project`
+  for project health.
+- An Agent Skill under `skills/agent-doc-system/`.
+
+[Unreleased]: https://github.com/lbframe/agent-doc-system/compare/v1.2.0...HEAD
+[1.2.0]: https://github.com/lbframe/agent-doc-system/compare/v1.1.0...v1.2.0
+[1.1.0]: https://github.com/lbframe/agent-doc-system/compare/v1.0.0...v1.1.0
+[1.0.0]: https://github.com/lbframe/agent-doc-system/releases/tag/v1.0.0
+[docs/BENCHMARK.md]: docs/BENCHMARK.md
+[#5]: https://github.com/lbframe/agent-doc-system/issues/5
+[#6]: https://github.com/lbframe/agent-doc-system/issues/6
+[#7]: https://github.com/lbframe/agent-doc-system/issues/7
+[#11]: https://github.com/lbframe/agent-doc-system/issues/11
+[#12]: https://github.com/lbframe/agent-doc-system/issues/12
+[#17]: https://github.com/lbframe/agent-doc-system/issues/17
+[#18]: https://github.com/lbframe/agent-doc-system/issues/18
+[#21]: https://github.com/lbframe/agent-doc-system/issues/21
+[#23]: https://github.com/lbframe/agent-doc-system/issues/23
+[#25]: https://github.com/lbframe/agent-doc-system/issues/25
+[#27]: https://github.com/lbframe/agent-doc-system/issues/27
+[#28]: https://github.com/lbframe/agent-doc-system/issues/28
+[#29]: https://github.com/lbframe/agent-doc-system/issues/29
+[#30]: https://github.com/lbframe/agent-doc-system/issues/30
+[#31]: https://github.com/lbframe/agent-doc-system/issues/31
+[#37]: https://github.com/lbframe/agent-doc-system/issues/37
+[#38]: https://github.com/lbframe/agent-doc-system/issues/38
+[#40]: https://github.com/lbframe/agent-doc-system/issues/40
+[#41]: https://github.com/lbframe/agent-doc-system/issues/41
+[#43]: https://github.com/lbframe/agent-doc-system/issues/43

package/README.md

@@ -1,152 +1,137 @@
-# agentdoc — portable documentation & software-catalog system
+# AgentDoc
 
-Zero runtime dependencies. One Node runtime. AgentDoc compiles a repository
-into one deterministic graph and routes a path or an entity to the minimum
-context needed to change it safely.
+![AgentDoc — Give coding agents the context they actually need](docs/assets/readme/agentdoc-hero.webp)
 
-It exists to answer one question cheaply and correctly, for an AI coding agent
-or a human reviewer:
+**Give coding agents the context they actually need.**
 
-> I am about to change *this*. What do I need to know, what must I keep
-> consistent, and which command proves it?
+AgentDoc turns your codebase, docs, contracts, and runtime evidence into a context layer your coding agent can query before it changes anything.
 
-It does that while keeping three kinds of claim apart — what the repository
-*claims*, what can be *derived* from code, and what was *observed* running —
-and surfacing their disagreements as conflicts instead of guessing.
+[![npm](https://img.shields.io/npm/v/agent-doc-system.svg)](https://www.npmjs.com/package/agent-doc-system) [![license](https://img.shields.io/badge/license-MIT-blue.svg)](LICENSE) [![node](https://img.shields.io/badge/node-22%20%7C%2024%20%7C%2026-43853d.svg)](https://nodejs.org/)
 
-## Install the CLI
+## Give this to your coding agent
 
-The CLI is installed once per machine. Requires Node 22 or newer (22, 24 and
-26 are the tested baselines).
+Paste this into your agent **from the repository you want to use AgentDoc on**:
 
-```bash
-npm install -g github:lbframe/agent-doc-system
-
-agentdoc --version
-agentdoc doctor        # machine self-check
+```text
+Set up AgentDoc for this repository and use it as the context layer for your work.
+
+AgentDoc: https://github.com/lbframe/agent-doc-system
+
+1. Read INSTALL_FOR_AGENTS.md and install or upgrade the AgentDoc CLI.
+2. Read skills/agent-doc-system/SKILL.md.
+3. Run agentdoc doctor --project and agentdoc audit.
+4. Use the appropriate CREATE or MIGRATE workflow for this repository.
+5. Do not invent ownership, architecture, authority rules, or product intent. Ask me only for decisions AgentDoc cannot safely derive.
+6. Finish setup with agentdoc validate, agentdoc compile, and agentdoc check.
+7. From now on, use agentdoc query and agentdoc impact before changes, then agentdoc check after changes.
+8. Report anything unresolved or conflicting instead of guessing.
 ```
 
-Unpinned installs track `main`. For reproducible CI, pin a tag or commit:
-`github:lbframe/agent-doc-system#v1.0.0`.
+That is the intended onboarding path: **give the prompt to the agent, let AgentDoc build the context layer, and keep using it during development.**
 
-The install footprint is the CLI only — `bin/`, `core/`, `adapters/`,
-`schemas/`, `templates/`, `evals/`, docs and the agent skill. Nothing is ever
-copied into a project that uses AgentDoc.
+## Why AgentDoc?
 
-### Have an AI agent install it for you
+Coding agents can read code. The hard part is knowing **what matters for this change**.
 
-Paste this into a capable coding agent:
+Important context is scattered across code, documentation, API contracts, architectural decisions, CI rules, and production reality. Agents either load too much or miss something important.
 
-```text
-Install AgentDoc on this machine from:
-https://github.com/lbframe/agent-doc-system
+AgentDoc gives them a map.
+
+- **Query the right context** for a file or component.
+- **See the blast radius** before changing code.
+- **Detect drift** between code and documentation.
+- **Surface conflicts** instead of silently guessing.
+- **Verify constraints in CI** before stale knowledge becomes trusted context.
 
-Read INSTALL_FOR_AGENTS.md from that repository and follow it completely.
+![AgentDoc in action — query, impact, check](docs/assets/readme/agentdoc-in-action.webp)
 
-Do not initialize, inspect, migrate, or modify any software project as part of
-this installation task.
+## Three commands, one workflow
 
-When finished, verify the AgentDoc CLI and report the installed version.
+```bash
+agentdoc query src/payments/webhook.ts
+agentdoc impact src/payments/webhook.ts
+agentdoc check
 ```
 
-[`INSTALL_FOR_AGENTS.md`](INSTALL_FOR_AGENTS.md) contains the procedure.
+`query` tells the agent what it needs to know. `impact` shows what the change can affect. `check` verifies that the repository still agrees with itself.
 
-## Use it in a project
+## Why not just AGENTS.md or RAG?
 
-From the repository root:
+![Why AgentDoc](docs/assets/readme/why-agentdoc.webp)
 
-```bash
-cd my-project
-agentdoc doctor --project   # is there a catalog here, and is it healthy
-agentdoc audit              # inventory the repo, classify every fact
-```
+AgentDoc is not another documentation file and it is not just semantic search.
 
-Then follow the workflow the audit implies:
+It combines **authored intent**, **facts derived from code**, and **runtime observations** without pretending they are the same thing. When sources disagree, AgentDoc surfaces the conflict instead of choosing a convenient answer.
 
-| workflow | when | entry point |
-|---|---|---|
-| CREATE | no catalog exists | `agentdoc init`, then `agentdoc scaffold --write` |
-| MIGRATE | established repo, scattered or stale docs | `agentdoc audit`, work the report |
-| VALIDATE | before committing, in CI | `agentdoc validate` → `compile` → `check` |
-| QUERY | before and during a change | `agentdoc query <path-or-ref>` |
-| MAINTAIN | after a change | `agentdoc impact <paths>` |
+## From repository knowledge to agent-ready context
 
-AgentDoc writes **project data only**: `agentdoc/` (configuration and
-descriptors you own), `docs/` skeleton files, `.agentdoc/graph.json` (the
-compiled artifact), and a CI workflow. The CLI source is never vendored into
-your repository.
+AgentDoc turns scattered repository knowledge into task-specific context your coding agent can query, verify, and act on.
 
-The two directories differ in ownership, not just punctuation:
+![How AgentDoc works](docs/assets/readme/how-agentdoc-works.webp)
 
-- `agentdoc/` is **authored** — configuration, descriptors and observation
-  sets. Everything in it is a claim you wrote and a reviewer can diff.
-- `.agentdoc/` is **generated** — the compiled graph and other build
-  artifacts. Never edit it by hand; `compile` rewrites it and `check` rejects
-  a stale or edited one.
+## What it understands
 
-## Use it with an AI agent
+AgentDoc reads manifests, contracts and configuration. It never runs your toolchain.
 
-The agent skill is [`skills/agent-doc-system/`](skills/agent-doc-system/SKILL.md).
-Install that directory as a skill in your agent harness. It teaches the agent
-when to use the CLI, the evidence-class model, the working loop, and the
-integrity rules — with deeper material under `skills/agent-doc-system/references/`.
+Support comes from adapters and from built-in contract validation (AsyncAPI needs no adapter).
+- `agentdoc init` picks `node`, `go`, `rust`, `python`, `dockerfile` and `wrangler` from the files it finds.
+- It always includes `database`, `openapi`, `events` and `github-actions`.
+- Add the others in `agentdoc/agentdoc.config.yaml`.
 
-## What it is built on
+| area | supported |
+|---|---|
+| languages & packages | Node (npm, pnpm workspaces, Turborepo), TypeScript, Go (modules, `go.work`), Rust (Cargo packages and workspaces), Python (`pyproject.toml` with PEP 621 or Poetry, uv workspaces, `setup.cfg` / `setup.py`, `requirements*.txt`) |
+| contracts | OpenAPI, protobuf, GraphQL, AsyncAPI, OIDC discovery |
+| deployment & CI | Dockerfile / Containerfile, `fly.toml`, `Procfile` and similar deploy configs, Cloudflare Wrangler, Ansible playbooks, GitHub Actions |
+| resources | databases (migrations, ORM config, connection evidence), caches, object stores, event subjects |
 
-Most repositories conflate three different kinds of claim, and the confusion
-is the single largest source of wrong agent behaviour:
+For each language, AgentDoc derives:
 
-| what | who says it | how it fails |
-|---|---|---|
-| what the repository **claims** | authored docs, config, manifests | drifts, is aspirational, is not tested |
-| what can be **derived** from code | imports, manifests, contracts | heuristic, incomplete, wrong about intent |
-| what was **observed** in a running system | schedulers, bindings, live APIs | goes stale, disagrees with intent |
+- the units;
+- whether each one is deployable or importable;
+- dependency edges between sibling units;
+- verification commands taken from the repository's own evidence (`go test`, `cargo test -p`, `pytest`, `python -m unittest`, npm scripts, …).
 
-A claim carries an evidence class, a confidence, and a pointer to its evidence.
-When two classes disagree about the same fact, the disagreement is **surfaced
-as a conflict**, not reconciled. There is no global precedence order, because
-authority depends on the fact: the committed manifest is authoritative for
-*intent*, the live scheduler is authoritative for *what fires*, and the
-canonical contract is authoritative for the *desired wire surface* while a
-live probe is authoritative for *what is deployed today*.
+What every adapter derives, and what it deliberately cannot see, is in [docs/ADAPTERS.md](docs/ADAPTERS.md).
 
-The default is fail-closed. A fact with no governing authority rule and
-disagreeing evidence is an error, and no gate passes while it stands.
+## By the numbers
 
-## Documentation
+Measured, reproducible, and with no model in the loop. Method and caveats are in [docs/BENCHMARK.md](docs/BENCHMARK.md).
 
-| file | covers |
-|---|---|
-| [`skills/agent-doc-system/SKILL.md`](skills/agent-doc-system/SKILL.md) | how an agent uses the system |
-| [INSTALL_FOR_AGENTS.md](INSTALL_FOR_AGENTS.md) | machine installation, for an AI agent |
-| [docs/SPEC.md](docs/SPEC.md) | the normative specification: entities, graph, determinism |
-| [docs/AUTHORITY.md](docs/AUTHORITY.md) | evidence classes, conflict semantics, elections |
-| [docs/PROVENANCE.md](docs/PROVENANCE.md) | evidence records, pointers, secret handling |
-| [docs/QUERY.md](docs/QUERY.md) | the context router, budgets, freshness refusal |
-| [docs/MIGRATION.md](docs/MIGRATION.md) | CREATE and MIGRATE workflows |
-| [docs/CI.md](docs/CI.md) | the drift gate and its failure modes |
-| [docs/ADAPTERS.md](docs/ADAPTERS.md) | writing an adapter, and what must never live in one |
-| [docs/EVALS.md](docs/EVALS.md) | the routing benchmark and pre-registered thresholds |
-| [docs/COMPARISON-KODA.md](docs/COMPARISON-KODA.md) | measured comparison against the reference corpus |
-
-## Fixtures and the golden corpus
-
-- `fixtures/a-node-ts-postgres` — single Node/TypeScript service, REST, Postgres, GitHub Actions.
-- `fixtures/b-go-multi-service` — multi-service Go repository, protobuf contracts, shared library, resources.
-- `fixtures/c-alt-monorepo` — deliberately *not* `apps/`+`packages/`+`workers/`, mixed runtimes, inconsistent pre-existing docs.
-- `examples/koda` — the profile that represents a large, mature, TypeScript+Go monorepo, including the runtime-observation conflicts that motivated the authority model.
-
-## Development
+- **The right context: 100% vs. 58%.** Across 30 routing scenarios (Node, Go, Rust, Python), `query` + `impact` retrieves every fact the change needs. Reading only the touched component retrieves 58%. Noise stays under 9%, against 40%. Every check the change requires is in the answer, and that answer holds about half of the repository's checks (29% on the largest corpus), not all of them. The thresholds were fixed before the first run and have not moved since.
+- **Seconds to map a real repository.** A fresh `init` → `scaffold` → `compile` takes 1.0–4.6 s on four public repositories, from 141 to 5,142 files (axum, pydantic-ai, full-stack-fastapi-template, create-t3-turbo). No edit was needed beyond choosing each component's placement.
+- **A briefing, not a dump.** A `query` answer is about 100–260 tokens: what the change reaches, what to read, and which checks to run.
+- **The same graph every time.** Compiles are byte-identical, and CI can verify it.
+- **Small and strict.** Zero runtime dependencies, a package of about 285 kB packed, 21 adapters, and 709/709 cases on the official TOML 1.0 conformance suite.
+
+## Install manually
+
+Requires Node.js 22+.
+
+```bash
+npm install -g agent-doc-system
+agentdoc doctor
+```
+
+Then, inside a repository:
 
 ```bash
-node --test tests/*.test.mjs   # unit, adversarial and mode tests
-node evals/run-all.mjs         # aggregate routing evaluation (gate)
-npm run selfcheck              # agentdoc doctor, from the source checkout
+agentdoc audit
 ```
 
-## Licence of the design
+AgentDoc also ships with an Agent Skill at [`skills/agent-doc-system/`](skills/agent-doc-system/SKILL.md).
+
+## Documentation
+
+- [Getting started & migration](docs/MIGRATION.md)
+- [Authority & conflicts](docs/AUTHORITY.md)
+- [CI & verification](docs/CI.md)
+- [Specification](docs/SPEC.md)
+- [Adapters](docs/ADAPTERS.md)
+- [Benchmark](docs/BENCHMARK.md)
+- [Agent Skill](skills/agent-doc-system/SKILL.md)
+- [Changelog](CHANGELOG.md)
+
+## License
 
-The architecture, the fail-closed discipline, the provenance model and the
-warning-acceptance mechanism are extracted from a production catalog system
-that proved them at scale. The namespace, schemas, adapters, CLI and authority
-model here are new and neutral.
+[MIT](LICENSE) © 2026 lbframe