Your First Run

A worked example: "add a settings page". This shows what actually happens turn by turn.

The brief

After /leopold-brief, your .leopold/PLAN.md might read:

- [ ] Scaffold the settings route and page shell
- [ ] Build the settings form with validation
- [ ] Wire the form to the API
- [ ] Add tests for the settings flow

And your CHARTER.md includes a rule like:

Prefer the simplest thing that works for the MVP. No new infrastructure.

The run, turn by turn

flowchart TD
    T1["Turn 1 · scaffold route"] --> CR1["/code-review → clean"] --> M1["mark [x]"]
    M1 --> T2["Turn 2 · build form"]
    T2 --> F{"fork:<br/>Zustand or Context?"}
    F -- "charter: simplest, no new infra" --> Dec["decide: Context API<br/>log D1"]
    Dec --> M2["mark [x]"]
    M2 --> T3["Turn 3 · wire API"]
    T3 --> F2{"fork:<br/>migrate users table now?"}
    F2 -- "irreversible + ambiguous" --> Stop["stop · notify you"]
    classDef stop fill:#e63946,stroke:#9d0208,color:#fff;
    class Stop stop;

What you would see:

1. Turn 1 finishes the scaffold, runs /code-review, marks the item done.
2. Turn 2 hits a real fork (state library). The charter says "simplest, no new infra", so Leopold decides Context API, writes a block to DECISIONS.md, and continues, without asking you.
3. Turn 3 reaches a database migration against real data: irreversible and ambiguous. Leopold stops and notifies you, because deciding that is not its call.

What you get back

When you return and run /leopold-status:

active: false  turn: 3/50  stopped_reason: escalation
plan: 2 done, 2 open
decisions logged: 1

• DECISIONS.md shows exactly what "you" decided while you were away, including a Reversal: line for each call.
• Everything is staged, nothing committed. You review and commit.

The decision log is the trust mechanism

Autonomy you cannot audit is autonomy you cannot trust. Every non-mechanical decision is logged with the charter rule that drove it and how to undo it.

Read more about how forks are resolved in the Decision Protocol.