library-wiki

The wiki at docs/graph/libraries/ is this project's source of truth for every external dependency. Each library used directly by the project has a page; each page is local, version-pinned, sourced, and compounding.

This skill is the discipline of building and maintaining those pages. It is invoked from the ingest-library protocol and any time an agent uses a library and realizes the page is missing or stale.

When to apply this skill

• A new dependency is being added → create a page.
• A research-scout draft is staged → file and finalize it (Workflow below), at the close-out of the pass that drafted it.
• A pinned version is bumped → refresh the page.
• A new idiom for using a library was just adopted → record it.
• An agent ran into a behavior the page doesn't cover → add a pitfall entry, dated.
• A security advisory affects a wikified library → update §7 and notify grill.md §11.
• The page is older than the project's review cadence → refresh.

The discipline

1. The wiki is the project's distillation

A wiki page is the narrow slice of the library that this project uses, plus the project-local idioms, pitfalls, and history. The whole upstream goes in docs/graph/sources/raw/ and docs/graph/sources/normalized/; the page in docs/graph/libraries/ is the project's distillation.

If a wiki page reads like a tutorial for someone who has never used the library, it has drifted from the discipline: tutorials belong upstream.

2. Pin to a version

Every page pins the exact version the project depends on. "Latest" is not a version. When the project upgrades, the page updates the pin and adds an §8 (Upgrade path) entry summarizing what changed between the previous and current pin.

A pin is not by itself a currency claim. Where the page says the dependency is current, it names the version and the support phase of the line that version belongs to. "On the latest release" and "supported" are two claims: a line that has left active support satisfies the first while failing the second, and a pin recorded without its phase reads as currency to every later reader. Nor is a pin what runs: it is what the project declares. Where the project deploys to more than one environment, a bump the page records stays intended until the running version is measured in each one, and the page says which environments it measured; a declared pin, and even a version endpoint the platform exposes, can disagree with what is deployed. When two measurements disagree, both stay on the page with their dates (protocol.verify-disagreement, mutable subjects).

3. Cite every claim

Every claim on the page has a citation in §10 (References): a URL, a retrieval date, and (when the source is paywalled or transient) a local snapshot path in docs/graph/sources/raw/.

Use upstream text as a short quote, in quotes, with its citation, or rewrite the idea in the project's own words; close paraphrase is neither: it copies upstream text without attribution. Most pages need very little upstream text.

4. Compound

§3 (Used API surface) lists the names the codebase actually uses, one line added when a new file starts using a new name. §4 records idioms once they are adopted, and §5 records pitfalls once they are hit, each dated. Each section grows when the project meets the thing.

The page grows with the project. A bare page that says only "we use library X for Y, install with Z" is a perfectly acceptable starting point. A page left static diverges from the code within weeks: the docs-librarian keeps each page current at close-out, and the implementer and reviewer name wiki drift in their handback.

5. Validate before publishing

A wiki page becomes authoritative once a smoke test confirms the pinned version works. The smoke test:

• Imports the library at the pin.
• Calls one or two of the names in §3.
• Runs in the project's normal test harness.

If the smoke test fails, the page is wrong (or the install is wrong); fix one of them before promoting the page.

Workflow

Creating and refreshing a page (which phase, which owner, which spawn waits for which) are ingest-library.flow and ingest-library.refresh in docs/graph/protocols/ingest-library.md. What this skill adds to the scout's draft, section by section: §0 (Pin) and §1 (Role) from the lockfile and the architect's brief; §2 (Install) from a command actually run in a clean environment, recorded exactly as it worked; §3 (Used API surface) from the current code that uses the library, or, when the page precedes the code, the names from the brief marked "planned"; §10 (References) from the sources research-scout staged in docs/graph/sources/. The smoke test that makes the page authoritative is the tester's phase of the same pass.

Depth when adopting an existing project

Index every direct dependency from manifests. Create a detailed library page during initialization when the dependency is architecturally significant, security/operations critical, unusual, or used across several subsystems. Each page answers how this project uses it, where configuration lives, relevant constraints, sharp edges, and verification status.

Installed versions come from manifests/locks. Upstream lifecycle, compatibility, or current guidance requires a primary upstream source; otherwise mark it pending rather than guessing. Record consulted sources in docs/graph/sources/.

When to also create a best-practices/ page

The library wiki is per-dependency. When a concern spans multiple libraries and needs a cohesive guidance (e.g. "how this project handles HTTP errors across our two HTTP clients"), that's a docs/graph/best-practices/<concern>.md. It links to the relevant wiki pages.

Reference files

• docs/graph/templates/library-page.template.md: the page template.
• docs/graph/protocols/ingest-library.md: the ingest workflow.
• docs/graph/agents/09-docs-librarian.md: the agent that owns the wiki.
• docs/graph/agents/10-research-scout.md: the agent that fetches sources.