rollback-runbook
Use before and after calling rollback_release - what a rollback restores per delivery mode, when a schema change makes a code rollback unsafe, how to work and report manual_steps, and what to do when rollback_release refuses
- License 许可
- Apache-2.0
- 最近更新
- 2026-11-15
- Stars
- ★ 122
- Forks
- 18
01 前言 Front Matter
- name
- rollback-runbook
- source
- ankane/strong_migrations (MIT), wshobson/agents (MIT), adapted
- category
- release
- description
- Use before and after calling rollback_release - what a rollback restores per delivery mode, when a schema change makes a code rollback unsafe, how to work and report manual_steps, and what to do when rollback_release refuses
03 SKILL.md 正文
Rollback runbook
rollback_release always reverts the release's task merge commits on the default branch and pushes that revert — production can never ship the bad change again, whatever else happens. What restores the RUNNING service depends on the component's delivery mode:
dispatch— the previous good release's commit is tagged and the deploy workflow is dispatched at it (Mechanism: workflow_dispatch). No previous release: it dispatches at the revert commit instead.on_merge— the revert push itself is the redeploy (Mechanism: revert_push); there is nothing else to trigger.batch(desktop, mobile) — nothing is redeployed.Statusgoes straight torolled_back: a published desktop build or a store build cannot be unpublished by reverting a git commit, so there is no redeploy forwatch_releaseto follow. Seebatch-artifact-verificationfor the yank step.
For dispatch/on_merge, Status becomes rolling_back and you are expected to call watch_release again to follow the redeploy through to rolled_back.
Before you roll back: check what schema the release changed
A code revert is not automatically safe. For each task in release.tasks, get_task_pull_request({ task_id, include_diff: false }) to list files, then look for migration paths: migrations/, db/migrate/, prisma/migrations/, alembic/versions/, V*__*.sql, *.up.sql. Read the up file (get_task_pull_request with the diff, or read_file) for any match.
- Additive —
CREATE TABLE/CREATE INDEX,ADD COLUMNthat is nullable or has a constant default: reverting the code is safe. The previous version simply ignores the new objects — leave the schema in place; do not drop what the forward migration added. - Destructive or renaming —
DROP COLUMN/DROP TABLE,RENAME COLUMN/RENAME TABLE,ALTER COLUMN TYPE,SET NOT NULLon an existing column: the previous version may not run against this schema at all. If the task'srollback_plancovers it, follow the plan exactly. Otherwise do not callrollback_release— comment the evidence plus "a code rollback would run the previous version against a schema it does not know — needs a human: forward fix or a down-migration first", and stop.
This is the parallel-change pattern (DORA): decouple a database change from the application change that depends on it, so either one can revert on its own. A migration that only adds is reversible by the code revert alone; one that removes or reshapes is not, unless a plan already accounted for it.
What a revert cannot undo
git revert undoes code. It does not:
- reverse a database migration,
- turn a feature flag back off,
- purge a CDN or cache,
- roll back a third-party configuration change,
- un-send a webhook, an email, or anything else the code triggered while it ran.
rollback.manual_steps is exactly this list, built from each rolled-back task's own rollback_plan/before_deploy/after_deploy fields — the developer who made the change is the only one who knew which of these applied.
Batch releases add one more, first
For a batch release, manual_steps puts one entry ahead of the tasks' own: unpublish or halt the artifact the revert cannot touch. See batch-artifact-verification for the exact commands. Treat it like any other manual step: perform it if you can, report it if you cannot, and put it first in your report — it is the step that actually stops the bad build reaching more users, not the code revert.
Work through it
- Read every item in
manual_steps. - Perform every one you can with the tools you have — most manual steps are outside this system's reach (a third-party console, a human-run command) and are for a human;
release-terminal-scopenames the one write you may make yourself. - Say explicitly, on the card, which ones you performed and which you could not, and who has to finish them.
✅ "Reverted the code (revert a1b2c3d). Left for a human: migration 043 added tasks.priority (NOT NULL DEFAULT 'medium'); the previous version ignores it, so it stays — dropping it would discard values written since the deploy. T-12's after_deploy enabled flag new_pricing: turn it off in the flag console. T-14 recorded no rollback plan."
❌ "Rolled back; the column added by migration 042 needs a human to drop it." — additive columns are not dropped; see "Before you roll back" above.
A rollback reported as complete when half of it was not is worse than one that admits what it did not do — the next release will ship on top of whatever was left half-undone.
No plan recorded
If a rolled-back task recorded no rollback plan at all, say that too. The absence is a finding for whoever reviews this release next, not something to paper over with "nothing else to do here".
When rollback_release refuses
Do not retry — a refusal here is final the same way a merge refusal is (board_rollback_release_refused):
- Too old (more than 24h since it was released) or a newer release already shipped — production now runs something newer than what you'd be rolling back to. One comment saying so, and that this needs a forward fix as a new task (a human or PM call), then stop.
- A newer release is open — resolve that one first; leave this one alone.
verifyingordeploying— the sweeper will hand it back to you at the right moment; stop and let it.
04 同仓库更多 Skills
visual-craft
暂无简介
ux-copy
Use when you write any words a user reads in a mockup or a hand-off - content language and voice, buttons, labels, errors, empty states, confirmations, numbers and dates, and the longest-string check
screen-mockup-html
Use when you draw screens as the design task's HTML mockup - ONE canvas document holding every screen and every variant side by side, the sanitizer-safe recipe, phone and wide frames, the states grid, static control states, and the loopback self-check with screenshots at 375, 768 and 1440 attached to the task
performance-awareness
How the performance score moves for your role. Use when the score context shows a declining trend, or before handing off work you have not self-checked against every criterion.
design-variants
Use when a screen design needs variants - how many, variants as side-by-side rows on the one canvas, what makes two variants genuinely different, recommending one, how the human's choice is read, and narrowing to the chosen one in revision
design-system-authoring
Use when you derive, change or extend a design system - where tokens live per stack, the DTCG token tree, role-based names for both themes, base vs layer vs drift, the DESIGN.md section order, the inventory format, the contrast check and the lint every proposal returns