bounded-auto-loop

Bounded Auto Loop

Use this skill when the user wants a few autonomous iterations without an open-ended research spiral.

Loop Contract

Before starting, lock these down:

1. Decision to change.
2. Current control or baseline.
3. Live hypotheses.
4. Iteration budget, usually 2-3 iterations.
5. Stop rule.
6. Return condition.

If any of these are missing, infer the narrowest reasonable version and state it briefly.

Design The Loop

• Choose the smallest checks that can change the decision.
• Freeze everything not under test.
• Prefer one-variable changes over sweeps.
• Stay on the cheapest discriminative surface first.
• When a failure mode is easiest to see, prefer one decisive visual check over a larger batch of scalar metrics.
• Do not widen the search space mid-loop unless the current loop is falsified.

Run One Iteration At A Time

For each iteration:

1. State the root-cause or causal hypothesis.
2. Make the minimal code or config change.
3. Run the nearest useful validation.
4. Interpret the result before launching the next step.
5. Update the plan or stop.

Do not queue multiple speculative branches at once unless they are independent and the stop rule still stays sharp.

Preserve The Active Window

• Once a user starts a time- or iteration-bounded loop, keep the turn open until the budget expires, the declared stop rule is met, a genuine user-only blocker appears, or the user explicitly cancels or replaces the loop.
• Treat later steering, priority changes, and status questions as additions to the active loop unless the user clearly says they replace it. Answer them in commentary and continue with the remaining budget.
• Do not send a final response merely because one subrequest finishes, a worker reaches a decision point, or an intermediate result is ready.
• After every interruption or new message, recompute the remaining time or iterations and resume the same loop.

Stop Aggressively

Stop early if:

• the decision is already changed
• the new branch is materially better than control
• all plausible local variants fail in the same regime
• the next move is a scientific or product choice rather than an engineering extension
• the remaining improvements are too small to justify another cycle