repl

The user designs. You run the rest.

After the spec is approved, nothing waits for the user until the run ends. You sit in the user's seat for every question the pipeline would ask, answer it by decide, and log the answer.

1. Brainstorm with the user

Invoke brainstorm with the intent. It runs exactly as it always does: questions, pushback, approaches, design sections, the spec, and the user's approval. Nothing is automated here.

2. Run limits

Right after the user approves the spec, ask once:

Run limits: at most 3 verify rounds, at most $0 on paid runs. Reply with other numbers, or accept these.

Write the answer under the spec's ## Constraints as Run limits: at most N verify rounds, at most $M on paid runs, and commit it with the approval. This is the last point that waits for the user.

A goal whose only criterion is a paid run, with no budget left for it, cannot be measured. The run pauses on it rather than spending.

3. Set up

1. Invoke decide. It stays on for the rest of the run.
2. Create branch repl/<topic> in a worktree at .vibe-worktrees/<topic>, off the current branch. exec refuses to run without one.
3. Add .vibe-worktrees/ to .git/info/exclude if it is not there. Never edit .gitignore.
4. Commit the spec on the run branch. If its path is gitignored, force-add it there: the branch is local, so this is a two-way door.
5. Start the decision log at .vibe-worktrees/<topic>.decisions.tsv with the header row ts phase decision why evidence result.

4. The pipeline

Run plan, exec, verify and debug exactly as written, every one of them, however small the change looks. Making a change yourself instead of through exec is never yours to decide. Where a skill would wait for the user, you answer:

5. Rounds

A round is one verify pass plus the fixes it routes. Count rounds against the cap and money against the budget, and log both as they change.

A round that ends with the same or more findings than it started with did not help. Undo its fix commits with git revert, never a reset, log it, and change approach in the next round.

6. Amending the spec

You may add a goal or a non-goal for behaviour the spec does not cover, or record one reading of an ambiguous requirement. Each amendment is its own commit, spec: amend <topic>: <one line>, with a log row.

Never delete a goal, weaken a goal or its criterion, or move a goal into the non-goals. If weakening is the only way forward, pause.

7. The decision log

One row per decision, with the columns ts, phase, decision, why, evidence and result. Evidence is a pointer, such as a commit, a file:line or a command, never prose. The log is append-only: a wrong call gets a new row that supersedes it.

The log sits beside the worktree, not in it, so verify's clean-tree and scope checks never see it.

8. Endings

• done: verify, invoked in this run, returns ready, no blocker is open, and every warn or nit still open at the cap carries a ruling.
• paused: the run reached something only the user can decide.
• capped: the round cap was reached with a blocker still open.

On every ending, commit the work on the run branch. For paused and capped, write a resume note to .vibe-worktrees/<topic>.resume.md with the intent, what is verified, the current state and the first action on resume. Check that every evidence commit in the log exists with git cat-file -e, and that every row matches something that happened in this session. Then report.

Never push, open a pull request, or merge. Integration belongs to the user, after they read the report.

The report

Ending:     done | paused | capped, and HEAD
Verdict:    verify's last verdict, each goal with its evidence
Amended:    every spec amendment, or none
Set aside:  every finding set aside, with its ruling, or none
Rulings:    how many, and the log path
Spent:      money spent on paid runs, against the budget
Paused on:  the door or dead end and the resume note path, for paused or capped

Handoff

None of your own. retro fires after the report when a debug report in the run carried a Class: other than none. Otherwise the user takes it from the report.