llmdoc / docs / workflows

Workflows

Daily retrieval is lightweight and read-only. Knowledge mutations are explicit workflows with validation, Git boundaries, and clear result states.

Daily retrieval

1. Choose one entry point: tree, search, context --files, index, or show.
2. Let llmdoc narrow the working set.
3. Use native tools for exact, live source facts.
4. Reapply the gate only when entering a new subsystem.

Initialize

Use the init workflow only when a repository has no valid V3 knowledge surface. It designs domain and topic boundaries, writes the smallest sufficient set of owner documents, and classifies every first-class subsystem as documented, intentionally reconstructable, or a blocking gap.

After drafting the initial documents, an unborn repository first creates a real HEAD commit. Then init-state seeds the ledger, validate checks structure, Context Floor probes verify routing, and commit --all completes the Git-backed bootstrap.

Update

Update semantically verifies existing knowledge against code deltas and authorized reflection candidates. A document that is still correct is verified unchanged; it does not accumulate a change log.

Prose and routing are reviewed independently: true prose may still need description, code.paths, or relations repaired.

Prune

Prune converges the knowledge surface by removing duplication, fragmentation, and cheaply reconstructable inventory. Mechanical reports provide signals; a recorder still makes the semantic judgment.

Upgrade

Upgrade is only for an explicitly requested V2-to-V3 migration. It is never an implicit part of update or prune.

Safety and completion

• Mutation workflows require confirmation and a clean llmdoc write boundary.
• Validation happens before completion; failed writes are reverted through Git.
• commit excludes unrelated user-staged files and refreshes validity separately.
• Workflows report one result: success, no_change, dry_run, incomplete, or failed.