Product Acceptance Scenarios

These scenarios keep development anchored to user-visible value.

A. Core quality loop — primary MVP

1. Open a local Git Project.
2. Create Task with goal and acceptance criteria.
3. Make it Ready.
4. Start the chosen agent strategy.
5. Forge prepares isolated worktree.
6. Agent works while Forge shows useful activity.
7. If ambiguity occurs, a capable agent raises a Decision and Task shows Needs You.
8. Human resolves; work resumes.
9. Candidate revision is captured.
10. Verification runs against that revision.
11. Human reviews diff and criteria.
12. Human requests changes.
13. Agent revises to a new candidate.
14. Old Review/Verification is visibly stale.
15. Human reviews current candidate and approves.
16. Task becomes Done only through the valid approval path.

B. Crash safety

During worktree creation, agent startup, an open Decision, verification, and candidate settling, kill Forge and restart. The system must preserve work and converge to honest state rather than guessing completion.

C. Long-history responsiveness

Create a deterministic fake Run with thousands of activity/raw events. Reopen the Task. Forge must not block the UI by replaying every raw delta one-by-one; history is paged/snapshotted/coalesced.

D. Capability honesty

Select a strategy that cannot interactively escalate. Escalation-dependent Involvement Modes are unavailable with an explanation. Forge does not silently use a weaker behavior while showing the richer promise.

E. Worktree safety

Create uncommitted changes in a managed worktree and request cleanup. Forge refuses automatic destructive removal and explains why.