yukimemi

Imported from yukimemi/yukimemi (AGENTS.md). Install upstream with npx skills add yukimemi/yukimemi. Copyright stays with the author.

Shared conventions

This file is the agent-agnostic source of truth (per the agents.md convention). Codex and Claude Code (>=2.1.277) read it directly; the GEMINI.md file is a thin shim for tools that don't yet auto-load AGENTS.md. Edit AGENTS.md, not the shim.

Git workflow

• No direct push to main. Open a PR.
  • Exception: trivial typo / whitespace / docs wording fixes.
• Branch names: feat/..., fix/..., chore/....
• PR titles + bodies in English. Commit messages in English.
• Releases are PR-driven and tagging is automatic — in repos that ship a release pipeline. Bump the version in the project's own manifest in a chore/release-vX.Y.Z PR; on merge to main the language layer's auto-tag.yml detects the bump, pushes the vX.Y.Z tag, and that tag is what fires release.yml. Do not run git tag by hand — the bot tag will collide and the manual push fails. The specifics belong to the layers shipping those two workflows, which are not the same layer: kata:agents:rust:* for which file holds the version and for auto-tag.yml, kata:agents:rust-{cli,lib}:* for what release.yml builds and publishes. A repo with no auto-tag.yml has no release pipeline at all: nothing tags, and the version field in its manifest may well be decoration.

Pre-merge review

Review happens before the pull request, on the operator's machine, via magi. This layer no longer ships PR-side review bots: claude-review.yml and claude.yml were removed from it. Their scope was human-authored PRs — their own job-level if: already excluded chore/release-*, kata-apply/auto, apm-bump/auto and Renovate / Dependabot — which is exactly the set magi reviews, so keeping them meant reviewing the same diff twice, a CLAUDE_CODE_OAUTH_TOKEN secret per repository, Actions minutes on private repos, and one trap that silently cost reviews: a PR editing either workflow was skipped by claude-code-action's workflow-validation check and merged with a green check and no review attached.

"Removed" is a statement about this template layer, not about every repo's current state. Dropping a [[file]] entry stops kata from managing the rendered file — it does not delete it. A repo that had these workflows before this change keeps claude-review.yml / claude.yml (and the CLAUDE_CODE_OAUTH_TOKEN secret) under .github/workflows/ until someone deletes them by hand, and until then they still fire on every human-authored PR. Check .github/workflows/ before treating a PR as unreviewed-except-magi: if either file is still there, its comments are a real review, not noise to ignore.

• magi review <branch> --merge none runs only the review + verification + gate half of magi's graph: nothing competes, no implementation, no judging, no vote. That is the mode for hand-written work. magi run "<task>" is the full competition, for work handed over whole. Both end at the same gate.
• Always pass --merge none on a magi review you invoke by hand. [merge] mode in config defaults to "pr" for the queue loop (magi serve / magi web) that drains magi task add — there, ending in an actual PR is the point. magi review run directly inherits that same config default when no --merge flag is given, which would open a PR nobody asked for. --merge none overrides it for that one invocation and leaves the branch for the operator to turn into a PR themselves.
• What the loop actually does: each reviewer gets its own detached worktree pinned at the commit under review (no reviewer can perturb the tree, and the fixer never races one); verify.e2e runs in the branch's worktree and its output is fed to the fixer; finding ids (R2-1-3) are assigned by magi, not by the agent, so the fixer's adoption report can be matched against them; the loop is bounded by review_rounds; verify.gate must exit 0 before any merge is attempted.
• magi.toml is repo-owned, not kata-managed. Point verify.gate at the exact command CI runs, so a local pass means a green PR, and point verify.e2e at the invocation that actually covers the repo — feature flags included. A gate that differs from CI turns a clean magi run into a red PR, which is the one failure this arrangement cannot absorb.
• If you did not run magi, the change was not reviewed, and nothing will tell you. Do not open a PR for a hand-written change before magi review comes back clean; if you must, say so in the PR body and say why. What does not count as a substitute: a green CI run (it compiles and tests, it does not review), and CodeRabbit's silence.
• CodeRabbit stays installed and is not part of the gate. It does not auto-review repositories under 10 stars — the common case here — so treat it as absent unless it posts. When it does post, its findings are a real review: address them, reply in the inline thread with an @coderabbitai mention (the review-comment replies endpoint, gh api repos/<owner>/<repo>/pulls/<N>/comments/<id>/replies -f body=…), and reply even when declining — say why, because a silent skip reads as overlooked. A "review limit reached" quota notice carries no findings and counts as quiet; re-trigger with @coderabbitai review when the quota refills if you want a real pass.
• Read the report, not the exit status. A reviewer seat that times out is logged as WARN agent timed out seat=review-2 and then summarised as "raised 0 finding(s)" — indistinguishable from a genuinely clean pass in both the summary and magi stats. Check for timeouts before believing a clean round: a round where half the panel never answered is not a clean round.
• Review artifacts stay local. magi comments on a pull request only when it stops landing one. Findings, the fixer's adoption report and reviewer precision live in the run directory (magi show, magi stats). When the PR needs a record — a non-obvious fix, a finding declined with an argument — paste that part into the PR body or a comment yourself.
• With merge = "pr", magi opens the pull request and keeps going: watches the checks, reads the review comments (human and bot), runs a bounded fix round when either is unhappy, pushes, and asks before merging. land_approval is on by default and silence is a hold — nothing merges unanswered. magi answer (or the web UI) is where it asks. Out of rounds leaves the PR open with a comment saying what still fails; checks: unknown never merges.
• Merge gate: magi's gate green — or CI green for a change magi never touched — and every review that did post resolved (a leftover claude-review.yml, CodeRabbit, a human) and the owner's explicit approval. The irreversible step stays a human decision.
• No review-monitoring poll loop for bots this layer no longer ships. The old loop existed to wait on them. Where a repo still has claude-review.yml (see above) the old cadence still applies until it is deleted; otherwise, after opening a PR wait for CI and report the wait state to the owner. When magi is landing the PR (land = true), magi does the watching.
• Bot-authored PRs (Renovate / Dependabot) need no review pass at all: CI green + owner approval.
• Version-bump-only PRs — a single chore/release-vX.Y.Z branch whose entire diff is [workspace.package].version / [package].version plus the matching inter-crate refs and the lockfile — likewise. There is nothing in a version bump for a reviewer to find, and the release pipeline downstream of merge (auto-tag → release.yml) is time-sensitive.

Worktree workflow

Before your FIRST edit to any file, run renri add — NEVER edit the main checkout. Read-only inspection (Read / Grep / Glob) stays on the main checkout; the instant you intend to change a file, you must already be in a worktree. The trap that keeps catching agents: diving into a fix the moment the diagnosis lands and editing in place. A concurrent agent shares the main checkout — your in-place edits will clobber theirs or be clobbered, and in a jj-colocated repo a stray working-copy commit entangles unrelated WIP into your branch. If you slip and edit in the main checkout, capture the diff first (jj already snapshotted it into the working-copy commit, so jj diff > patch; for git, git stash or save a patch — if you got as far as committing on a branch, just push it). Then reset the main checkout to pristine main (jj new main@origin, or git switch -), renri add a worktree, and re-apply the captured diff there.

Use renri for any commit-bound change. From the main checkout:

renri add <branch-name> --from main@origin            # create a worktree (jj-first), off latest upstream main
renri --vcs git add <branch-name> --from origin/main  # force a git worktree, off latest upstream main
renri remove <branch-name> -y --non-interactive  # cleanup after merge (agent-safe; see note)
renri prune                        # GC stale worktrees

Read-only inspection can stay on the main checkout.

Always pass --from <upstream main> (main@origin for jj, origin/main for git). Without it, renri add forks off the cwd worktree's current HEAD — in a long-lived main checkout that often lags upstream, so the PR later shows up CONFLICTING against a main that had already moved (e.g. a refactor merged upstream before the branch was cut), forcing a manual re-port of the whole change. renri add does fetch first, but fetching only updates main@origin — it never moves the checkout's HEAD, so an explicit --from is what guarantees a fresh base.

Agents / non-interactive shells: renri remove prints a details panel and waits for a confirmation prompt — without -y it hangs, and --non-interactive alone errors asking for -y. Always pass -y, and add --non-interactive so a mistyped/omitted name fails instead of opening a fuzzy picker (the same picker-fallback applies to remove / cd / exec with no name). Use -f/--force to remove a worktree that still has uncommitted changes or conflicts. To sweep every merged-PR worktree in one shot: renri remove --merged -y.

kata-managed sections

Several files in this repo are managed by kata apply from the yukimemi/pj-presets templates — the bytes between <!-- kata:*:begin --> and <!-- kata:*:end --> markers, plus the overwrite-always files listed in .kata/applied.toml. Editing those bytes locally won't survive the next kata apply — push the change to the upstream template repo (yukimemi/pj-base / yukimemi/pj-rust / …) instead.

The marker scopes are layered, one per applied layer: kata:agents:base:* is this section, and each layer adds its own (kata:agents:rust:*, kata:agents:rust-cli:*, kata:agents:pnpm:*, kata:agents:firebase:*, …). Which ones apply here is a grep away: <!-- kata: in this file.

This project's own conventions

Everything a layer ships is generic by construction: it describes the stack the template assumed, not what this repo grew into. Bytes outside every marker pair are yours and survive kata apply — so project-specific conventions belong in a section of their own, outside the markers (conventionally at the end of the file; if a later layer appends its block below yours, no matter — kata only ever rewrites between its own markers). Same mechanism as the .gitignore / .gitattributes blocks.

Write those conventions down there rather than leaving them in one agent's head, in commit archaeology, or in a README the agent will not read. What earns a line:

• Any layer default that does not hold here. A layer states its assumption flatly ("Hosting is the primary target", "these rules are a placeholder to replace"). When the project has diverged, say so and say why — the layer's text keeps asserting the opposite on every apply, and an agent that only reads the blocks will act on it.
• Facts duplicated across files with no compiler in between — an address or a path that appears in code and in a rules/config file that cannot import it, a timeout that has to stay inside another timeout. List every copy, so the next edit finds them all.
• kata-shipped files this project deleted on purpose, together with the once_applied = true line in .kata/applied.toml that keeps them deleted. Otherwise someone helpfully restores one.
• Shapes the runtime forces but no tool checks — an export form a platform requires, import specifiers that must (or must not) carry a file extension, a directory whose contents are reachable by URL.
• Invariants that money or access rest on, naming the file and line that actually enforces them.
• Which language the code speaks versus what a user reads, when the two differ.

A repo whose AGENTS.md is nothing but kata blocks is a repo where every agent re-derives all of that from scratch — and gets the layer defaults wrong the same way each time.