Agent-Skills.md

Agent Skills: wrap

Pre-commit session analysis. Triages trivial vs substantive changes, checks commit scope, and for substantive sessions proposes docs/permissions/automations updates. Use when user is finishing work, about to commit, or says /wrap.

UncategorizedID: keenranger/dotfiles/wrap

Repository

keenranger/dotfiles
keenranger

Install this agent skill to your local

pnpm dlx add-skill https://github.com/keenranger/dotfiles/tree/HEAD/agent/skills/wrap

Skill Files

Browse the full folder contents for wrap.

Download Skill

Loading file tree…

agent/skills/wrap/SKILL.md

Skill Metadata

Name
wrap
Description
Pre-commit session analysis. Triages trivial vs substantive changes, checks commit scope, and for substantive sessions proposes docs/permissions/automations updates. Use when user is finishing work, about to commit, or says /wrap.

Pre-commit session analysis: $ARGUMENTS

Step 1: Triage (always)

Assess the upcoming commit. Use git diff --staged (or unstaged if nothing staged).

Trivial (skip agent fanout):

  • Single file, roughly under 20 lines changed
  • No new dependencies, no new exported symbols, no new commands/skills/agents
  • No new patterns worth persisting (typo fix, single value tweak, comment update, dependency bump)

If trivial, output one line: Trivial change; nothing to wrap. Retry with WRAP_OK=1 prefix. Skip steps 2-3.

Substantive proceeds to step 2.

Step 2: Scope check (always)

Verify the diff matches a single intent:

  • Flag unrelated reformatting, opportunistic cleanup, or edits in files not required by the stated task
  • Migration / lint / rename PRs are especially prone to scope drift (see CLAUDE.md "Scope and Context")
  • If scope drift suspected, surface Diff touches X which seems unrelated to <stated intent>. Split into a follow-up commit? and wait for confirmation before proceeding

Step 3: Analysis (substantive only)

Run enhance agent (CLAUDE.md updates, automation opportunities, follow-ups). Pass session context.

After enhance completes, run permission analysis via the built-in fewer-permission-prompts skill (not the custom permission-analyzer agent -- the built-in is automode-aware and writes to project settings automatically).

Note: session-level retrospective analysis (course-corrections, ambiguity, missed efficiency) is not done here. The user's CLAUDE.md captures repeat-correction patterns in-session at the moment they occur; that obviates a post-hoc retrospective pass.

Permission scope split

fewer-permission-prompts only touches project settings (.claude/settings.json). For commands that are generally useful across projects (git, pnpm, ruff, etc.), surface them as candidates for user settings (~/.claude/settings.json) manually -- ask the user before adding.

Auto-mode permission framing

User runs in defaultMode: "auto". Most commands execute silently, so a blanket "what ran in this session" list is no longer a useful whitelist signal. Meaningful signals only:

  • Commands that still triggered a prompt despite automode
  • Commands that required dangerouslyDisableSandbox
  • Commands the user explicitly denied, then re-ran with adjustments
  • New WebFetch domains that required approval

The built-in skill already filters for these. Skip proposals for commands that simply ran in auto.

Focus areas

  • Learnings: Technical discoveries, mistakes made, new patterns
  • CLAUDE.md: Preferences or patterns worth persisting
  • Permissions: Commands that bypassed normal flow (see auto-mode framing)
  • Automation: Repetitive patterns that could become skills/commands/agents
  • Follow-ups: Incomplete work, TODO markers, next session priorities
  • Retrospective: User guidance patterns that could be improved
  • Scaffolding cleanup: After exploration sessions, identify experiments whose purpose has resolved -- probe UIs, debug bridges, broadcast receivers, ad-hoc forward-synthesis helpers, intermediate analysis docs. Verify the final diff contains only what serves the stated request; flag scaffolding for removal in a follow-up commit before sending to reviewers.

Duplicate check

  • Before proposing, check both project (.claude/) and user (~/.claude/) locations
  • Skip proposals that duplicate existing content

Permission placement

  • Project settings (.claude/settings.json) for project-specific permissions
  • User settings (~/.claude/settings.json) for personal/global tools

Output

  • Triage result (trivial / substantive)
  • Scope check result (clean / split suggested)
  • Summary of session activity (substantive only)
  • Categorized proposals (docs, permissions, automations) (substantive only)
  • Follow-up tasks for next session
  • Workflow suggestions (from retrospective)
  • User selection: which proposals to apply

User selection

  • Present options via AskUserQuestion
  • Options: update CLAUDE.md, add permissions, create automations, skip
  • Execute only selected actions

Goal: Enrich configuration before each substantive commit; do not burn cycles on trivial ones.