Codex Rebase

Rebase workflow

Move the branch onto its base, then establish that what came out is right.

Codex doesn't provide a skill-scoped command hook. Use only the scripts named here to pin the rebase settings and check conflict markers before continuing. The continuation script also proves an empty skip. Never use a bare stash pop, start the rebase by hand, or force-push without a lease.

Preflight

Run bash .agents/skills/codex-rebase/scripts/preflight.sh.

Step 0: open the task list

Create these with update_plan, keeping exactly one item in_progress at a time.

1. Choose the base
2. Settle the working tree
3. Start the rebase
4. Resolve whatever it stops on
5. Verify the result
6. Report

Preflight reporting a merge or cherry-pick in progress stops everything. Say what's in the way and hand back. A rebase already in progress is the other case, and that one resumes at step 4.

Step 1: choose the base

Preflight computed this under == rebase base ==. Take its recommendation.

• Local default branch carries everything the remote has, so rebase onto the local branch.
• Local default branch sits behind the remote, so rebase onto origin/<default> and skip the stale local copy.

The commit skill's preflight makes the same decision, and both reach the same answer. Where the operator names a different base, use theirs and say why it differs.

Nothing to replay means nothing to do. Say so and stop.

Step 2: settle the working tree

--autostash handles a dirty tree, and start-rebase.sh passes it. It scopes the save to this rebase and never touches the shared stash stack, which is what makes it the sanctioned form here.

The stashed work sits outside the worktree while you resolve a conflict, so nothing you see there reflects it. It can also conflict on the way back once the replay finishes. If that durability isn't enough, stop and let the operator settle the worktree before starting. Don't invent a temporary commit or reset sequence inside this workflow.

Never a bare git stash pop or git stash apply. Worktrees share one global stash stack whose indices shift whenever any worktree pushes, so a bare pop takes whatever sits at stash@{0} right then. The guard hook refuses it.

Step 3: start the rebase

bash .agents/skills/codex-rebase/scripts/start-rebase.sh <base>

Don't type the git rebase yourself. The script pins off rebase.updateRefs, an operator preference that quietly moves other local branches pointing into the replayed range. It pins off rebase.autoSquash as well, which would otherwise collapse a fixup! commit that earned its own review. Recording the pre-rebase tip is the third thing it does. Without that record, step 5 has nothing to compare against.

Exit 0 means the replay finished, so go to step 5. A 1 means it stopped, so go to step 4.

Read the output for a line about rerere replaying a recorded resolution. That's git staging a resolution from an earlier rebase without stopping, and nothing has read it since. Treat every such path as unverified.

Step 4: resolve what it stops on

bash .agents/skills/codex-rebase/scripts/rebase-status.sh

That says which commit stopped, how far through the replay it sits, which paths remain unmerged, and what git staged without asking. When another workflow started the rebase, preflight reconstructs the durable before/base record from git's active rebase state so verification still has both ranges after the replay finishes. Then follow codex-resolve-rebase-conflicts. Keep the resolution in this thread because delegated agents share the filesystem and concurrent conflict edits are unsafe. The skill classifies each path before anything edits it and settles the mechanical ones itself.

Once you have staged every path:

bash .agents/skills/codex-rebase/scripts/continue-rebase.sh

That reads the staged blobs for conflict markers before it advances. git checks only that no path remains unmerged, so a file you staged with <<<<<<< still in it continues straight into a commit.

Repeat from rebase-status.sh at each further stop. Where git calls a commit empty because the change is already in the base, continue-rebase.sh --skip-empty drops it after proving the index matches HEAD.

Giving up stays available, and it costs nothing:

git rebase --abort

Reach for it when the conflicts say somebody cut this branch from the wrong place. Resolving thirty of them to reach that conclusion helps nobody.