Native code graph

bun:graph combines a native SQLite registry with Rust source extraction and graph queries. bun:graph-native runs extraction and graph algorithms in workers. Use the qualified fork for these APIs.

Source scopes

Each indexed domain has a source and profile. A runtime domain can use src/runtime, source graph:bun:src:runtime, and profile bun. A package domain can use packages/buv and source graph:bun:packages:buv. The registry keeps these domain values in repository metadata. The snapshot repeats the scope and records coverage.

File coverage and AST coverage measure different work. The file index can retain a path for an unsupported format. An unsupported or metadata-only input has UNRESOLVED provenance. A metadata-only record has no content SHA-256. The export reports these gaps explicitly.

Snapshot evidence

Select the repository and snapshot IDs before generating documentation. The exporter verifies the requested domain, source and profile against both records. It reads nodes and edges only within that snapshot's namespace. Historical input hashes come from immutable file nodes in the snapshot. The current files table does not supply historical hashes.

The export uses bounded SQL reads and streams Markdown and JSON files. Each section defaults to 256 rows. Each output defaults to an 8 MiB limit. You can set --max-rows up to 10,000 and --max-bytes up to 64 MiB. The manifest gives exact snapshot counts and marks truncated sections. Benchmark samples require matching domain metadata on a completed run.

Use scripts/aphrody/graph-docs.ts --help for the exporter options. The command requires --repository, --snapshot, --domain, --source, and --profile. --workspace selects the checkout. --db selects its SQLite registry. --out selects a local staging root.

The output directory contains:

• graph.md, graph.mdx, and graph.html for source evidence.
• inputs.json, nodes.json, edges.json, and samples.json for bounded evidence rows.
• manifest.json for snapshot, producer, generation-source and artifact hashes.
• SKILL.md for a scoped local skill.

The exporter keys the directory by its generation inputs. Repeated exports verify existing artifact hashes before reusing that directory. Changes produce a new directory.

Docs and skills

Bun's docs use MDX pages under docs/ and navigation in docs/docs.json. Repository skills use .claude/skills/<name>/SKILL.md with name and description frontmatter. Generated graph evidence stays under local staging, which defaults to tmp/graph-docs. Generating evidence does not publish docs or install a skill into a private home.

The PyJS plugin retains its API source hashes. Pass all five scope options to its docs generator to include graph evidence. An export identifies the indexed producer separately from the sources used during generation. A successful export establishes the selected evidence and its hashes. Runtime compatibility and full parser coverage require their own native tests.