/llmdoc:update

Use this command when repository changes require tracked llmdoc/ knowledge to be verified or synchronized.

Load the llmdoc skill before broad exploration. CLI commands below run as npx -y @tokenroll/llmdoc <cmd>.

Authorization

An explicit /llmdoc:update invocation authorizes this run to:

• read repository state and existing llmdoc docs
• write impacted llmdoc/ documents and llmdoc/meta.json
• write temporary investigation reports under .llmdoc-tmp/investigations/
• read and resolve declared reflection candidates under .llmdoc-tmp/reflections/

This command does not authorize source-code edits.

Preconditions

• git status -- llmdoc/ must be clean before the first formal write.
• Rollback means git checkout -- llmdoc/ (plus deleting any newly created files under llmdoc/); never hand-edit files back.
• If validate fails after writes and cannot be repaired in-run, roll back the doc write-set before reporting failure.

Workflow

1. Measure the current doc state.

   • Run status.
   • Run delta with any explicit scope flags.
   • Read each explicitly declared --reflection candidate. When the flag has no paths, read all Markdown candidates directly under .llmdoc-tmp/reflections/pending/. A pending candidate is an update signal even when code delta is empty.

2. Gate reflection candidates before promotion.

   • Require a verifiable trigger, wrong assumption or action, root cause, preventive rule, scope, confidence, and existing-doc match.
   • Verify repository claims against code, tests, or stable docs. User corrections prove intent, not repository facts.
   • Reject transient tool failures, one-task preferences, and unverified speculation. Use search and targeted show to find an existing owner before creating a document.

3. Choose the lightest sufficient path from delta plus any qualified candidates.

   • Light: owners are mapped and facts are clear.
   • Deep: files are unmapped, owner/root cause is unclear, boundaries changed, facts conflict, or impact is broad.
   • This choice controls evidence gathering only. It does not decide whether prose must change.
   • For unmapped or moved code and boundary changes, read Knowledge Topology and Context Floor and classify each surface as missing mapping, missing owner, or intentional no-doc.

4. Decide the semantic outcome with recorder.

   • Treat delta and candidates as review evidence, not a write list or prose to copy.
   • Rewrite only false/incomplete claims or new conclusions that pass the Stable Knowledge Gate. Reflection candidates must pass both gates.
   • Gate prose and routing independently; true prose may still need routing metadata repair.
   • If the document remains true and the change adds only reconstructable evidence, mark it verified unchanged.
   • Light: recorder decides from targeted CLI reads. Deep: investigator reports evidence, then recorder applies the gate.
   • Scaffold brand-new docs with new; register docs that already exist as files with adopt <path...> — never hand-edit meta.json or recreate the file through new.

5. Finalize.

   • If document identities changed, read Startup Configuration. mv syncs renames; sync manual merges or deletions before validation.
   • Run validate; after mapping or boundary changes, also run the reference's scoped routing acceptance.
   • If prose changed, run commit -m "<message>", adding --verified <path...> for reviewed unchanged docs. If all stayed unchanged, run commit --verified <path...>. Full verification uses --all, never with --verified.
   • commit validates, commits prose, refreshes fingerprints, and lands meta.json separately. Never reconstruct this sequence manually or --amend it.
   • Dirty mapped source makes commit fail closed; commit or clean that source, then retry.
   • After success, N commits behind HEAD, metadata-only; knowledge clean reflects the meta follow-up commit, not staleness.

6. Fold durable lessons into stable docs directly.

   • Put reusable cautions, invariants, and workflow fixes into the relevant architecture or guide docs.
   • Reflection candidates are a temporary evidence queue, not a tracked reflection kind or a second knowledge tree.
   • On success, resolve consumed candidates. On no_change, resolve only already-covered or rejected candidates. Leave them pending after incomplete or failed.

State Invariants

• Full successful updates may advance the repository baseline.
• Scoped updates must update only per-document fingerprints, not the global baseline.
• Update never changes convergence state. If growth requires convergence, finish update and ask once before running the separate prune workflow.

Result Contract

• success: the declared scope was semantically verified, any necessary prose changes were committed, and applicable revisions advanced.
• no_change: the declared scope was already current, so neither prose nor revision state needed to change.
• dry_run: the user asked for a dry run, or only status/delta/investigation/planning output was produced without writing llmdoc/; do not advance state.
• incomplete: evidence was insufficient, user input is required, or the scope belongs to a different explicit maintenance workflow; roll back writes and do not advance state.
• failed: update failed and doc writes were rolled back.

Always report:

• the chosen path (light or deep) and why
• the status and delta signals used
• any investigation report path
• each reflection candidate and its disposition (promoted, already_covered, dismissed, or pending)
• the stable docs changed and the docs verified unchanged by recorder
• routing classifications and checks when topology changed
• the commit result — validate gate, fingerprint refresh, and the meta follow-up commit — or why finalization was skipped