log

Keep project instructions useful for new agent sessions. Log reviews the project and proposes lasting context, working agreements and lessons.

Start the command

log

Use the prefix for your app or terminal from Setup.

Git-backed updates require a clean main or master branch. Without Git history, log can create instructions and leave them uncommitted.

Flow

flowchart TD
  Start(["Start log"]) --> Inspect["Check instruction files<br/>and the saved review point"]
  Inspect --> Layout{"Shared setup needed?"}
  Layout -->|Yes| Connect["Review layout changes<br/>for the selected work"]
  Layout -->|No| Scope{"Source review selected?"}
  Connect --> Scope
  Scope -->|No| Current(["Report the local result"])
  Scope -->|Yes| Scan["Investigate the selected<br/>Full scan or Gap scan"]
  Scan --> Review["Review findings and<br/>proposed instruction changes"]
  Review --> Choice{"Apply the proposal?"}
  Choice -->|Revise| Review
  Choice -->|Skip| Finish(["Finish locally"])
  Choice -->|All or selected changes| Write["Update and verify<br/>the accepted instructions"]
  Write --> Finish

Shared project instructions

AGENTS.md holds shared knowledge and Helm settings; CLAUDE.md imports it with @AGENTS.md and keeps any Claude-specific additions.

Existing files	Proposed setup
Neither	Create both during the first Full scan.
Only AGENTS.md	Keep it and add the Claude import.
Only CLAUDE.md	Review moving shared content into AGENTS.md.
Both	Keep a working connection or review reconciliation.

Custom content, imports, settings and adopted rules are preserved unless you approve a specific change. Conflicting or shadowed instructions need reconciliation first.

What the review covers

• Full scan: trace important workflows, setup, configuration and tests across every first-party package. This gathers project knowledge rather than auditing every line for code quality.
• Gap scan: review commits and relevant patches since the saved review point, including affected code. If a rewrite changed that commit’s ID, log first tries to recover its verified equivalent automatically. Only an unrecoverable baseline or incomplete required history prevents Gap scan.

Recovery tries preserved normalize maps, Git-filter-repo’s native map, then verified local object/history correspondence. Success adds a short note before the normal scan choices, without another approval gate. Any marker repair joins the existing approved output; Skip and read-only/no-op outcomes leave files unchanged. This restores the earlier review, not a new one. Missing local evidence in another clone may mean Full scan is still needed.

The review looks for useful identity, configuration, development commands, architecture, domain rules, working agreements, safety hazards and traps. Relevant corrections from the current conversation can become proposed lessons too.

For existing instructions, review one proposal and apply all, select changes, revise or cancel. New safety instructions need explicit approval. Custom layouts receive focused edits.

Your result

You get updated instructions, unresolved findings and a size report. Missing Helm preferences are saved only through approved changes. New configuration uses Markdown bullets; existing supported formats stay intact.

The last-reviewed comment advances only after a complete source review and acceptance of all substantive changes. Partial acceptance and layout-only work keep the earlier marker.

Changes follow your local commit policy and stay on the same branch. Use adopt for copied rules and ship for publication.