Ship Fix

Ship a fix

One pass from report to released version. This skill is the order of operations; the rules live elsewhere and are linked, not repeated. Read them when a phase points at them.

• AGENTS.md: Post-Task Checks, "Choosing the right test", Release Pipeline.
• .agents/memory/MEMORY.md index, especially coderabbit-review-workflow.md, finish-before-pr.md, no-ai-attribution.md.
• scripts/release.sh: what a release refuses to do.

Model tiers

If your harness supports subagents with a model choice, delegate mechanical work to the cheaper/faster tier (for example Sonnet in Claude Code): implementation from a precise spec, test runs, CI and CodeRabbit polling, changelog drafting. Keep diagnosis, design decisions, review of subagent output, and the merge/release go/no-go with the main agent. Give a subagent the exact files, the failing test, and the acceptance command; never "figure out the bug". Without subagents, do everything inline in the same order.

Phases

1. Intake

Reproduce or confirm the root cause before touching code: the user report plus logs; for YouTube or extractor issues, run yt-dlp locally against the failing URL. State the chosen fix design (what changes, which layer, what is deliberately out of scope) before coding. verify: root cause is named with evidence, and the design is written down in one short paragraph.

2. Worktree

Create an isolated worktree on fix/<slug> (or feat/<slug>) from a fresh origin/main:

git fetch origin main
git worktree add -b fix/<slug> <path> origin/main
cd <path> && bun run bootstrap

Work only inside it. Never use bare git stash (the stash stack is shared across worktrees); use a WIP commit instead. verify: git status is clean, git rev-parse HEAD equals origin/main, bun run doctor is healthy.

3. Implement, test first

Write the failing test at the layer that owns the behavior ("Choosing the right test" in AGENTS.md). A bug found by using the app belongs in Fixture Product E2E: bun run build, then run with ARROXY_E2E_HEADLESS=1 so no window steals focus. Make it pass with the smallest change. Finish the whole change per finish-before-pr.md: locales and README sources are not follow-ups. When the release follows straight after the merge (this flow), the CHANGELOG entry is written in the release commit (phase 9) rather than in the PR. Delegate the implementation to a cheaper subagent with a precise spec; review its diff yourself. verify: the new test failed before the fix and passes after; the diff contains only what the design called for.

4. Gate

bun run check

Fix failures your change caused. A failure that is clearly unrelated is reported, not fixed (parallel agents may own it). verify: bun run check exits 0, or every remaining failure is documented as unrelated.

5. Commit and PR

Conventional commit message (fix: ..., feat: ...), body explains what and why. No AI attribution anywhere (no-ai-attribution.md). Before pushing:

git log origin/main..HEAD --format="%B" | grep -inE "claude|anthropic|co-authored|generated with"   # must print nothing
git log origin/main..HEAD --format="%an <%ae> | %cn <%ce>" | sort -u                               # only Antonio
gh auth status   # antonio-orionus must be the active account

If antonio-orionus is not active: gh auth switch --user antonio-orionus. The wrong account makes gh fail with misleading errors such as "must be a collaborator". Push, then gh pr create with a body of: problem, root cause, fix, tests. verify: both greps are clean, the PR is open, and its body has the four parts.

6. CodeRabbit

Follow .agents/memory/coderabbit-review-workflow.md exactly. In short: auto reviews are off, so comment @coderabbitai review right after opening; poll issues/<N>/comments for all three terminal outcomes (findings, clean, refused); check each finding against the code before acting; fix valid ones, push back with evidence on wrong ones; reply in the inline thread. Re-trigger once only if the post-review commits are substantial, then confirm coveredCommitId equals the pushed head. verify: every finding is fixed or answered, and the review covers the head commit.