Merge Review Skill

Run this immediately after any git merge. Merges — especially from a shared upstream branch — are the highest-risk moment for silent regressions: git checkout --theirs drops logic, auto-merge concatenates duplicate blocks, and renamed symbols go unnoticed.

Step 1 — identify the merge

git log --merges -5 --oneline

Find the most recent merge commit (or the one named in $ARGUMENTS). Extract:

• Which branch was merged in
• Which files were touched by the merge commit

git show --name-status <merge-commit-sha>

Step 2 — scan for incomplete integration

For each file modified by the merge, do the following checks.

Dropped logic

Compare both parent sides of the merge:

git diff <merge-sha>^1 <merge-sha>   # what our branch added / kept
git diff <merge-sha>^2 <merge-sha>   # what the incoming branch added / kept

Look for code paths that existed on one parent but are absent in the merge result — specifically:

• Activity calls, workflow steps, or service method calls that one side introduced and the other did not touch (these should survive the merge).
• Registered Temporal activities or workers listed on one side but missing in the merged result.
• Import statements that one side added and the merge result dropped.

Duplicate blocks

Read each merged Python file and look for:

• Repeated @field_validator decorators targeting the same field.
• Repeated @property definitions with the same name.
• Repeated field declarations inside a Pydantic model.
• Commented-out blocks that appear twice (sometimes both sides' conflict markers were accepted).

Alembic chain

For any alembic/versions/ directory touched by the merge:

uv run alembic history   # run from inside the app directory

Verify the chain is linear. Flag any script whose down_revision points to a script that no longer exists, or any two scripts that share the same down_revision.

Step 3 — run quality gate

Invoke the /quality-root skill (or run make code-quality from the monorepo root directly). Do not skip this even if Step 2 found nothing — formatting drift and type errors are common post-merge side effects.

Step 4 — check registered workers

For any Temporal worker file touched by the merge (typically activities.py, worker.py):

• List every function decorated with @activity.defn in the activities file.
• Verify every such function is registered in the worker's activities=[...] list.
• List every workflow.execute_activity("name", ...) call in workflow files.
• Verify "name" matches an @activity.defn(name="...") registered in the worker.

Unregistered activities are silent runtime crashes — Temporal accepts the workflow start but fails at execution time with no static warning.

Open team decision — blocking scope: Should this skill hard-stop (refuse to close out) if it finds a dropped-logic or unregistered activity issue, or should it always be advisory so the developer decides? The team should agree on this, especially before running it in a CI context.

Output format

## Merge review — <merge-sha> — <date>
Merged: <incoming-branch> → <target-branch>
Files touched by merge: N

### 🔴 Must verify
- <file>:<line> — <description of suspected dropped logic>

### 🟡 Suspicious patterns
- <file>:<line> — duplicate @field_validator for "retry_initial_interval"

### ✅ Clean
- Alembic chain: linear
- Registered activities: all workflow.execute_activity calls matched
- Quality gate: pass (or: see /quality-root output)

Do not auto-fix anything. The developer must confirm that flagged items are genuinely problems before touching code — the skill cannot always distinguish an intentional removal from an accidental drop.