Step Gate Execution Ledger

Problem

Long-context agents can drift from multi-step plans. After compression, a 15-step refactor may shrink in memory, and a Sub Agent may report completion while a step was skipped. StepGateSkill does not control how the agent works; it moves completion verification outside the agent context. Each Step is bound to a one-time key, and a task can only be closed by submitting the final taskKey.

How It Works

The skill uses four levels: Program, Node, Task, and Step. The core loop is: decompose work into concrete Steps, show the DAG, get confirmation, call start-plan, checkpoint each step, then finalize.
- Granularity: each Step should be a binary completion check, such as “write 3 unit tests for auth.ts”, not “write tests”.
- Keying: stepKey appears exactly once in a response; current never returns keys, preventing replay-style faking.
- DAG rules: explicit [] means parallel, omission means auto-serial, ["a","b"] means merge; cycles are rejected at plan creation.
- Recovery: interrupted sessions keep completed steps as cryptographic proofs, and skipKey is consumed once with an event record.

Limits

It is aimed at agent hallucination, context loss, and accidental step skipping, not deliberate external attack. When enabled, use at least Task + Step; an empty Task is invalid, and cross-session cancellation requires admin recovery credentials.