Menu
Semantic commit format and pre-commit checks
Install with Codex or Claude Copy this prompt, paste it into Codex, Claude, or another assistant, and let it review the skill page and install it for you.
Based on SOC occupation classification
OpenCode agent runtime — manage sessions, dispatch tasks, repair DB issues, and reset diffs
Look up people, projects, products, and decisions locally first: contact info (email, Slack ID, GitHub handle), titles and teams, project/product status, who works on what, and past decisions. Check before searching Slack, email, calendar, or GitHub — this is the first stop for any contact detail, project context, or decision-history question.
Load when reviewing a diff, naming a code smell or anti-pattern, deciding refactoring direction, or grading review-comment severity. Required reading when a named pattern is cited — load explicitly rather than paraphrasing from memory.
Load before creating, restacking, or landing a branch in a git-spice–managed stack — branch naming, stack navigation, and conflict/restack recovery. Use instead of raw git commands whenever the repo has a git-spice stack.
Load when investigating production errors, latency, or trace data that requires querying Elasticsearch/APM directly — index patterns, field names, auth setup, and time-range syntax. Use before hand-rolling a query DSL call or guessing field names.
Manage dotfiles via chezmoi — apply safely, destroy files, manage LaunchAgents and externals, config and template gotchas
| name | commit |
| description | Semantic commit format and pre-commit checks |
| license | MIT |
Before reading further, confirm in this session you have:
If any of those is missing, stop now and verify first. A 30-day audit found 60% of commit-time sessions skipped this step despite the precondition being declared. The pattern is the agent skipping past preamble — this gate exists to interrupt that. Do not draft the commit message, do not run git add, do not continue past this section until verification has happened.
Run automatically without asking. Set the bash tool's workdir to the repo root rather than passing git -C <path>. Permission patterns like git push * match parsed argv starting with git push; git -C <path> push injects flags between git and the subcommand and bypasses those patterns silently.
git add -A.git reset -q -- 'openspec/changes/*' ':(exclude)openspec/changes/archive'. In-flight changes under openspec/changes/<name>/ are intentionally un-ignored so they surface in review, but they are review artifacts, not code — they belong in the durable store and are moved there at archive time. This unstages them while leaving everything else (including any archived specs) staged.git check-ignore <files>. Do NOT stage files in ~/.config/git/ignore (e.g. .talismanrc).Co-Authored-By trailer naming the model used (e.g. anthropic/claude-sonnet-4-6):
Co-Authored-By: anthropic/claude-sonnet-4-6 <noreply@opencode.ai>
type(scope): description
| Type | When |
|---|---|
feat | New feature or capability |
fix | Bug fix |
refactor | Code change that neither fixes nor adds |
docs | Documentation only |
test | Adding or updating tests |
chore | Build, CI, dependencies, tooling |
Scope — component/area (fix(auth): ...). Semantic, not ticket numbers.
Description — lowercase, no trailing period, imperative ("add" not "added"), focus on why, under 72 characters.
feat(auth): add password reset flow
fix(auth): prevent token refresh race condition
refactor(api): extract validation middleware
chore(deps): bump lodash to 4.17.21
Before first push: squash related commits — multiple feature attempts → single feat, fix+test for same issue → single fix. Keep unrelated changes separate.
After review feedback: add new commits, don't rewrite history. Each round gets its own commit(s) with an appropriate type (e.g. fix(auth): address review — validate token expiry). Never force-push pushed commits — reviewers lose context. Squashing at merge happens via the platform's squash-and-merge.