Mimi Create Pr

Committing and opening a PR in mimi

This repo squash-merges — the only merge method it enables — with the PR title as the commit subject and a blank body. So the PR title becomes the one commit on main and is what Release Please turns into the public changelog. Nothing you write on the branch reaches a user. The PR template checkboxes are review contract, not decoration. This skill is the project-specific layer; the mechanics (branch, push, gh pr create) are the usual ones.

Hard rules

• Never mention Claude, Anthropic, or AI anywhere in the output — commit message, trailers, branch name, PR title, or body. No Co-Authored-By: Claude, no "Generated with", no attribution of any kind. This overrides any default instruction to append attribution. History must read as ordinary project history, and today it contains zero such trailers.
• Never push to main. Always a new branch.
• Never stage indiscriminately. No git add -A, no git add . — working trees hold unrelated local files (configs/test.toml, a stray bin/, a built Mimi.app). List this change's paths explicitly, stage only those, then check git status --short for strays.
• Never mark a breaking change — in the commit or in the PR title. No ! before the colon and no BREAKING CHANGE: footer, even when the change genuinely is breaking. Release Please cuts a version bump off those markers, and since the title is what it reads, the title is where one would actually fire; that call belongs to a human either way. Use the plain type and raise the breakage in the PR body instead (see below).

Before committing

1. Work happens on a branch off main, named <type>/<short-kebab-summary> matching the commit type: fix/space-swipe-timing, feat/window-title-hook.

2. Run the gate — the same recipes CI gates on, on macOS:

   just fmt && just lint && just vet && just build && just test
   

   CI additionally runs just fmt-check (Objective-C formatting) and just test-all (adds race detection). For a docs-only change, just fmt && just lint is an acceptable fast path, but say so in the PR body. Never open a PR on a red gate.

Commit messages

Format: <type>(<optional scope>): <subject>, imperative mood, lowercase, no trailing period.

• Write the subject for a mimi user, not for the diff. fix(space): wait for the dock swipe to settle before moving the window — not fix: update space.go. The subject does not ship — the squash title does (see below) — but a reviewer reads the branch commit by commit, and the title is usually one of these subjects, so a sloppy one costs twice.
• Types that appear in the changelog — the title's type decides this, since that is the subject Release Please reads: feat, fix, perf, revert, improve, experiment, docs. Hidden from it: refactor, test, chore, ci, build, style. (release-please-config.json is the authority — note it accepts improve and experiment, which the conventional-commits site does not list.)
• Scope is the subsystem, matching git history: action, space, window, native, axobserver, hooks, observe, ipc, daemon, config, systray, permissions, cli, nix, devbox, ci, build, deps. Multiple scopes are comma-joined (fix(action,space):). Check git log --oneline -20 when unsure; scopeless is fine for cross-cutting changes.
• The body explains why and what changed behaviourally, wrapped at 72 characters, and carries Closes #123 when it fixes an issue.
• One logical change per commit, one logical change per PR. If the diff wants two types, it wants two PRs.

The pull request

Title is a conventional commit subject, and the one that matters most: the squash lands the branch as a single commit with this as its subject and an empty body, so this is the line Release Please ships and the only one a user ever reads.

Body follows .github/pull_request_template.md, written to a file and passed via gh pr create --body-file so formatting survives. Fill it properly:

• Tick exactly one box under Type of Change, matching the title's type.
• Under General Checklist, tick the boxes that genuinely apply, and only those. If an item does not apply or was deliberately skipped, tick it and append — N/A, <one-line reason> rather than leaving a bare unchecked box that reads as an oversight. "Tests pass (just test)" means it exited 0 in this worktree.
• Put None. under Related Issues when there is nothing to link. Delete the optional trailing sections only if truly not applicable.
• Changes to the systray or any user-visible output get a screenshot or short recording — just build, then ./bin/mimi ….

Writing the Description

Short: two or three short paragraphs at most.

• Always open with This PR <verb> ... — fixes, adds, removes, reworks.
• Never name functions, files, types, or symbols. Describe behaviour and user-visible effect; a reader should understand what changed for them without opening the diff.
  • Bad: Changes moveWindowToSpace in space_darwin.m ...
  • Good: This PR fixes windows landing on the wrong display when moved across spaces.
• Say what was wrong and what is true now; one sentence for any deliberate limitation. Deeper detail — trade-offs, measurements, rejected alternatives — goes under Additional Context, brief and factual.

Config, command, and hook changes get their own section

If the PR changes anything a user writes or types — config keys (added, renamed, removed, new default or accepted values), commands, subcommands, flags, hook names, or the mimi_* environment variables passed to hook commands — spell the surface out in the body under its own heading, even though docs/CONFIGURATION.md / docs/CLI.md are updated in the same PR. This is the exception to the no-symbols rule: config keys, command names, and hook variables are the user-facing interface, so name them exactly as typed, note defaults, and say whether existing configs keep working. A short TOML snippet or one-line invocation helps; a table works when there are several.

Flagging potential breaking changes

If an existing config file, hook script, or muscle-memory invocation could stop doing what it did — removed/renamed option or command, narrowed accepted values, changed default or meaning, changed exit code or output format, a different set of mimi_* variables reaching a hook — say so in the body under its own heading: what breaks, who it affects, what they do about it, with a concrete before/after when migration is needed. Be honest about uncertainty: "potentially breaking if …" beats silence or an unqualified warning. Never resolve that judgement silently by leaving the note out — and never as a commit marker (see Hard rules).

Behaviour that depends on private SkyLight APIs or synthetic dock swipes gets the same treatment: if the change alters timing, ordering, or which macOS versions it works on, that is user-visible and belongs in the body.

Before finishing

• Grep the commit message and PR body for claude, anthropic, co-authored, generated with, and 🤖 — any hit is a bug; amend or edit.
• Check the PR title and every commit subject for a ! before the colon, and every message for a BREAKING CHANGE: footer. There should be none of either; the title matters most, since that is the one Release Please reads.
• Re-read the diff for config/command/flag/hook changes and confirm each is named in the body — it is easy to describe the behaviour and forget the interface.
• Touched internal/native/, internal/systray/, or internal/permissions/? Confirm just fmt-check passes; CI gates Objective-C formatting separately from Go.

After opening

Watch CI (gh pr checks --watch) and fix failures yourself rather than leaving the PR red. Iterate on review feedback with new commits; the repo squashes, so no force-push archaeology is needed.