| name | finalize |
| description | End-of-branch workflow. Dispatches the design-doc, context-doc, and user-docs agents to update every documentation layer, then creates a changeset, squashes commits, pushes, and opens a PR. Use when finishing work on a branch before merge. |
| when_to_use | Trigger phrases — "finalize", "finalize this branch", "wrap up", "wrap up this branch", "ship it", "ship this branch", "close out this branch", "end-of-branch workflow", "I'm done with this work, prep it for merge", "ready to merge — run the prep", "update docs and open a PR for this branch". Do NOT trigger for: mid-task WIP commits (use a normal commit), addressing PR review comments (use /design-docs:review), or the final post-approval squash (use /design-docs:merge-prep).
|
| allowed-tools | Skill, Agent, Read, Glob, Grep, Bash(git *), Bash(gh *), Bash(bun *), Bash(ls *), Write, Edit, TaskCreate, TaskUpdate |
| argument-hint | [--no-push] [--no-pr] [--no-squash] [--split-docs] [--no-context-docs] [--no-user-docs] [--dry-run] |
Branch Finalization
Orchestrates the end-of-branch workflow: analyze what changed, dispatch each documentation agent to update its domain, create a changeset, squash commits into a single clean commit, push, and open a PR.
Argument Parsing
Parse $ARGUMENTS for flags:
--no-push — skip pushing to the remote in step 8 (implies no PR, since a PR requires a pushed branch)--no-pr — push the branch in step 8 but skip PR creation (useful when the PR will be opened separately, e.g. by CI)--no-squash — skip the squash step (step 7), commit docs/changeset normally--split-docs — in step 7, squash into two commits instead of one: functional code changes (review-focus: primary) and ancillary docs/changeset changes (review-focus: ancillary), as a review-time focus signal for agent reviewers. Opt-in; default is a single commit. Ignored when --no-squash is set (nothing is squashed, so there is nothing to split).--no-context-docs — skip step 4 (Update CLAUDE.md Files / design-docs:context-doc-agent dispatch)--no-user-docs — skip step 5 (Update User Docs / design-docs:user-docs agent dispatch)--dry-run — preview what each step would do without modifying any files
If no arguments are provided, run the full workflow.
Task Tracking
Before starting Step 1, use TaskCreate to add one task per step below. Omit any step that the parsed flags disable so the user sees the real plan, not a list with skipped items still on it.
Default task list (no flags):
- Preflight checks
- Analyze branch changes
- Update design docs
- Update CLAUDE.md files
- Update user docs
- Create changeset
- Squash commits
- Push and open PR
Use TaskUpdate to flip each task to in_progress when you start the step and completed when the step finishes successfully. If a step fails or is aborted, leave the task in its current state and stop — do not flip it to completed.
In --dry-run mode, still create the task list so the user sees the plan, but only the first two tasks will run.
Agent Dispatch Convention
Steps 3-6 dispatch domain agents via the Agent tool rather than invoking individual sub-skills. Each agent has its full domain skill suite wired (including the matching *-docs-style or changesets:style skill that enforces project conventions), so the work stays inside an isolated context with the right toolset and rules already in scope.
When dispatching a documentation agent (Steps 3-5), pass a prompt that includes:
- The branch change summary from Step 2 (diff stat + commit list)
- The list of files modified by previous documentation steps in this finalize run (so the next agent has the full picture)
- A directive: "Review the documentation in your domain against these branch changes and update what needs updating. Use whichever of your skills apply."
- An instruction to report back which files were modified (or that no changes were needed)
When dispatching the changeset-manager (Step 6), pass a prompt that includes:
- The branch change summary
- A concise description of what changed in shipped surface (features, breaking changes, behavior shifts, fixes) — not the file list
- The reports from Steps 3-5 about which doc files were modified, so the agent can decide which of those are user-facing changelog material and which are internal-only
- A bump-type recommendation if you have one (the agent makes the final call)
Do not enumerate which specific skills the agent should run — the agent decides based on its skill suite.
Step 1: Preflight Checks
Run these checks before any work. If any check fails, stop and report.
GitHub call convention
Every gh invocation in this workflow must be prefixed with:
GH_TOKEN="${DESIGN_DOCS_GH_TOKEN:-}" GH_PAGER=cat gh …
Why: the session-start hook sets DESIGN_DOCS_GH_TOKEN from GITHUB_PERSONAL_ACCESS_TOKEN. If it's set, that token authenticates the call. If it's unset, the empty assignment scrubs any inherited GH_TOKEN in the shell so gh falls back to the keyring credentials that gh auth status reports — keeping the auth check and the writes on the same identity. GH_PAGER=cat prevents gh from blocking on a pager.
Branch check
Verify you are NOT on the default branch:
!git rev-parse --abbrev-ref HEAD
If on the default branch, stop:
"You are on the default branch. Finalize is for feature branches only."
Dirty working tree
Run git status --porcelain. If there are uncommitted changes, list them and ask the user:
"There are uncommitted changes in your working tree: [list files]. These will be included in the finalize commit. Continue or abort so you can commit/stash first?"
Wait for the user's response. If they say abort, stop the workflow.
Base branch detection
Detect the default branch:
!GH_TOKEN="${DESIGN_DOCS_GH_TOKEN:-}" GH_PAGER=cat gh repo view --json defaultBranchRef -q '.defaultBranchRef.name' 2>/dev/null || git rev-parse --verify main 2>/dev/null && echo main || echo master
Store the result as BASE_BRANCH for use in subsequent steps.
Empty diff check
Run git log $BASE_BRANCH..HEAD --oneline. The branch is empty only when there are no commits ahead of the base branch AND the working tree is clean — an all-uncommitted branch still has work to finalize (the dirty-tree check above already staged it for the finalize commit). So:
-
Commits ahead, or a dirty working tree (per the dirty-tree check above): there is work — proceed.
-
Zero commits ahead AND git status --porcelain is empty: report and stop:
"No changes detected on this branch compared to $BASE_BRANCH (no commits ahead, clean working tree). Nothing to finalize."
When the only changes are uncommitted, Step 7's soft reset to the merge base is a no-op and the staged working tree becomes the finalize commit(s) directly.
Session tag check
Check if the session/start tag exists:
git rev-parse session/start 2>/dev/null
If it does not exist, create it at the merge-base:
git tag session/start $(git merge-base HEAD $BASE_BRANCH)
Report: "Created session/start tag at merge-base."
GitHub auth check
Run the auth check using the same env hygiene as every other gh call in this workflow — the check site and the use sites in step 8 must agree on which credential is in play, otherwise the probe passes while the write later fails or posts to the wrong account:
GH_TOKEN="${DESIGN_DOCS_GH_TOKEN:-}" GH_PAGER=cat gh auth status
If not authenticated, warn the user:
"GitHub CLI is not authenticated. PR creation and the existing-PR check will fail without it. Run gh auth login to fix this, continue with --no-pr (push the branch but skip PR creation — git push uses git's own credentials, not gh), or continue with --no-push to skip the remote entirely."
If the user wants to continue with --no-pr, treat the rest of the workflow as if --no-pr were set. Same for --no-push.
Step 2: Analyze Branch Changes
Build a summary of what changed on this branch:
git diff $BASE_BRANCH...HEAD --stat
git log $BASE_BRANCH..HEAD --oneline
Present a brief summary to the user:
"This branch has N commits ahead of $BASE_BRANCH, touching N files. Key changes: [summarize from the diff stat]"
This summary is passed as the context payload to every agent dispatch in Steps 3-5.
In --dry-run mode: Show the summary and describe what each subsequent step would do, then stop.
Step 3: Update Design Docs
Dispatch the design-docs:design-doc-agent via the Agent tool. Pass a prompt containing the Step 2 branch summary and the list of changed files. Tell the agent to review design docs in .claude/design/ against the branch changes and update or create design docs where the architecture, data flows, design decisions, or implementation-plan outcomes have changed. The agent has its full design skill suite wired — let it decide which skills apply.
When the agent returns, capture its report (which files it modified, or "no changes needed") for the Step 4 dispatch context.
No-op handling: If the agent reports no design doc updates are needed, that is success. Report "No design doc updates needed" and proceed to step 4.
On failure: Report what the agent reported and stop. Tell the user which files the agent modified so they can review or revert.
Step 4: Update CLAUDE.md Files
Skip if --no-context-docs flag is set. Proceed to step 5.
Dispatch the design-docs:context-doc-agent via the Agent tool. Pass a prompt containing the Step 2 branch summary plus whatever the design-docs:design-doc-agent reported in Step 3 — the context agent needs to know which design docs were touched so it can update @ pointers, references, and CLAUDE.md sections that index those docs. The agent also re-records pointer content hashes in .claude/design/refs.json, so any design doc edited in Step 3 leaves its pointers drift-clean.
When the agent returns, capture its report for the Step 5 dispatch context.
No-op handling: Same as step 3.
On failure: Report and stop.
Step 5: Update User Docs
Skip if --no-user-docs flag is set. Proceed to step 6.
Dispatch the design-docs:user-docs agent via the Agent tool. Pass a prompt containing the Step 2 branch summary plus the Step 3 design doc updates and the Step 4 CLAUDE.md updates — the user-docs agent needs the full picture to decide whether READMEs, contributing docs, security docs, or other user-facing pages should be touched.
No-op handling: Same as step 3.
On failure: Report and stop.
Step 6: Create Changeset
Dispatch the changesets:changeset-manager agent via the Agent tool. Pass a prompt following the Agent Dispatch Convention for Step 6 — the branch change summary, a concise description of what changed in shipped surface, the Step 3-5 reports, and a bump-type recommendation. The agent has the full changesets skill suite wired (changesets:create, changesets:update, changesets:delete, changesets:merge, changesets:style, changesets:status, changesets:dependencies, changesets:config) and decides whether to create a new changeset, update an existing one, or report that the diff has no changelog-worthy items.
If the Agent call fails because the changesets:changeset-manager agent type is unknown (the changesets plugin is not installed), fall back to creating the changeset manually:
- Ask the user: "What type of change is this? (major/minor/patch)"
- Ask the user: "Describe the user-facing changes for the changelog:"
- Read the package name from
package.json (or the plugin's package.json)
- Write the changeset file to
.changeset/ in the standard format:
---
"package-name": patch
---
User's description here
Generate a random changeset filename (lowercase adjective-noun pattern).
On failure: Report and stop.
Step 7: Squash Commits
Skip if --no-squash flag is set. Proceed directly to step 8 with a normal commit of just the docs/changeset changes. (When --no-squash is set, --split-docs has no effect — there is no squash to split.)
Squash is the recommended default
This project squash-merges into the default branch and uses agent reviewers. Squash cleanly into a single commit (or two, with --split-docs below) — do not editorialize for keeping granular history. The arguments for preserving per-commit history do not apply to this pipeline:
- The history is discarded at merge anyway. Squash-merge collapses every branch commit into one on the default branch, so bisectability / archaeology arguments are about history this workflow throws away.
- Reviewers read the net diff, not the commit sequence. One commit vs. N produces the same
base...head "Files changed" view.
- Granular history is worse for agent reviewers. A commit-walking reviewer flags issues that were already fixed in a later commit (false positives), reasons poorly about "true in commit 3, superseded in commit 5 — which is current?", and burns context reading superseded code.
--no-squash is the escape hatch for the rare case where granular commits genuinely need to survive. Keep the confirmation gate below regardless.
Stage all changes
Stage everything including doc updates and changeset from previous steps:
git add -A
Show squash preview
Show the user what will be squashed:
git log $(git merge-base HEAD $BASE_BRANCH)..HEAD --oneline
"Squashing N commits into {one commit | two commits: functional + ancillary}. [list commits]. Continue? (yes/no)"
Wait for confirmation. If the user says no, stop.
Squash via soft reset
git reset --soft $(git merge-base HEAD $BASE_BRANCH)
All branch changes are now staged at the merge base. Commit them per the mode below.
Default mode (single commit)
Generate one conventional commit message from:
- The changeset content (for the description)
- The branch name (for inferring type and scope)
- The diff stat (for context)
Present the generated message to the user for review/editing.
Format:
type(scope): subject
[Description from changeset and branch changes]
Signed-off-by: [from git config]
git commit -m "<generated message>"
Split mode (--split-docs: two commits)
When --split-docs is set, produce two commits so agent reviewers can focus on the functional surface and treat doc churn as ancillary. Both are collapsed into one at squash-merge — this is purely a review-time signal. Recognizing the review-focus: trailer is the reviewer's job and is out of scope here; finalize only emits it.
-
Classify the staged paths. Read path roots from .claude/design/design.config.json (paths.designDocs, paths.plans, paths.context, paths.localContext, userDocs.readme, userDocs.repoDocs). A file is ancillary if it matches any of those, any package-level CLAUDE.md / CLAUDE.local.md, any README.md, or lives under .changeset/. Everything else is functional.
git diff --cached --name-only
-
Commit the functional changes first (git reset HEAD <ancillary-paths> to unstage ancillary, then commit what remains). Use the conventional message from the default mode, with a review-focus: primary trailer:
type(scope): subject
[Description of the functional changes]
review-focus: primary
Signed-off-by: [from git config]
-
Commit the ancillary changes second (git add <ancillary-paths>) with a docs-type message and a review-focus: ancillary trailer:
docs(scope): update design docs, context, and changeset
review-focus: ancillary
Signed-off-by: [from git config]
One bucket empty: if the branch has no functional changes (docs-only) or no ancillary changes, produce a single commit for the non-empty bucket with its matching review-focus: trailer. Never create an empty commit.
The review-focus: value is always the canonical lowercase form (primary / ancillary). Place it in the commit footer alongside Signed-off-by.
Move session tag
Move the session tag to the new squashed commit (the last commit created above):
git tag -f session/start HEAD
On failure: Report the git error. The user can recover with git reflog to find the pre-squash state.
Step 8: Push and Open PR
Skip entirely if --no-push flag is set. Report: "Changes committed locally on $BRANCH. Not pushed (--no-push). Run git push -u origin HEAD to push manually."
Check for existing PR
GH_TOKEN="${DESIGN_DOCS_GH_TOKEN:-}" GH_PAGER=cat gh pr view --json number,url 2>/dev/null
If a PR already exists, report:
"A PR already exists for this branch: [url]. Pushing latest changes."
Push and stop (don't create a duplicate PR).
Push
git push -u origin HEAD
Stop here if --no-pr flag is set. Report: "Pushed to origin/$BRANCH. No PR opened (--no-pr). Open one manually via gh pr create or your CI workflow."
Create PR
Generate a PR title from the changeset content or branch name. Generate the body from the branch diff summary and changeset description.
GH_TOKEN="${DESIGN_DOCS_GH_TOKEN:-}" GH_PAGER=cat gh pr create --title "[title]" --body "[body]"
Report the PR URL to the user.
Progress Reporting
Between each step, mark the matching task completed via TaskUpdate and report a brief status update:
- "Step 1: Preflight checks passed (session tag at abc1234)"
- "Step 2: Branch has 8 commits touching 12 files"
- "Step 3: Design docs updated (3 files modified)"
- "Step 4: CLAUDE.md files are current (no changes needed)"
- "Step 5: README.md updated with new API documentation"
- "Step 6: Changeset created (.changeset/fuzzy-cats.md)"
- "Step 7: Squashed 8 commits into 1 (feat: add merge commit support)" — or, with
--split-docs: "Squashed 8 commits into 2 (functional + ancillary)"
- "Step 8: PR opened: https://github.com/..."
Error Recovery
If any step fails:
- Stop immediately — do not continue to subsequent steps
- Leave the in-progress task in its current state (do not flip it to completed via
TaskUpdate)
- Report which step failed and why
- List any files that were modified by previous steps or by agents in this finalize run
- Suggest recovery: "You can review the changes with
git diff, revert with git checkout -- ., or fix the issue and re-run /design-docs:finalize"
- If the squash step (7) fails mid-operation, remind the user they can use
git reflog to find the pre-squash state
Steps 3-5 are idempotent — re-running against already-updated docs produces no changes. Steps 6-8 are not idempotent — the skill should detect existing changesets and PRs before creating duplicates.
