Plan gate

Plan Gate

Overview

Put reviewable intent before consequence. Bind planned work to scope, evidence, rollback, and abort criteria, then preserve every release point required by the task profile, governance rules, or execution host.

When to use

• Before a controlled step or any required pre-effect human release.
• Before broad, external, destructive, regulated, or hard-to-reverse action.
• When scope, load-bearing assumptions, checks, or rollback remain unclear.

When NOT to use: routine read-only exploration; an already-confirmed autonomous step with no optional planning need; or as a substitute for the human judgment required by assistive or human-final.

Friction modes

Mode	Behavior
full (default)	Show the complete plan and wait for explicit approve, revise, or reject. Use for every required gate.
lite	Show a compact goal/touches/checks note and continue. Allowed only when the confirmed final mode is autonomous, blast radius is 0–2, and no gate is otherwise required.
off	Skip only the optional plan presentation for work that is already allowed to proceed without a gate. Required gates and host permissions remain active.

If the requested mode conflicts with a required control, use full and state which profile, governance, regulatory, permission, or host constraint requires it. Friction settings never grant authority.

Process

1. Read the confirmed profile. Name the step's final mode, blast radius, architecture/governance conditions, and required host boundary. If a consequential step is unprofiled, run task-profile first.
2. Resolve the friction mode. Apply the table above; do not accept lite or off as an override of a required release.
3. State goal and non-goals. Make scope drift visible before work begins.
4. List affected surfaces. Name files, systems, data, people, services, and the type of effect.
5. Expose load-bearing assumptions. Mark each as verified or unverified.
6. Define success evidence. Name criteria, tests, comparisons, and the evidence that will be returned. A controlled output must have a criterion the human can apply at its release point.
7. Define rollback and abort criteria. State how to restore state and which observations cause an immediate stop.
8. Present assets/plan-gate.md. In full, stop until the human says approve, revise, or reject. Never self-approve or infer approval from silence. Plan approval does not waive the later output/release review.
9. Honor judgment boundaries. For assistive, execute only the approved preparation. For human-final, wait for the human to frame and decide; profile the resulting implementation as a new step before any effect.

Output

Use assets/plan-gate.md or an equivalent compact plan. Record the source of every required gate and the human decision in full mode.

Rationalizations

Excuse	Rebuttal
"The user selected off."	Off removes optional presentation only; it cannot waive a required gate.
"The result is verifiable, so approval can happen later."	A controlled release must be checked before it takes effect.
"The user already approved the goal."	Goal approval is not approval of this scoped plan.
"High consequence means the agent should not prepare anything."	Consequence can require a controlled release while still allowing verifiable preparation.

Red flags

• lite on a controlled, high-reach, regulated, or host-required step.
• off described as disabling safety or permissions.
• Work starting while a full plan still says approve / revise / reject.
• A controlled plan with no criterion available to the reviewer.
• Plan approval treated as approval of an unseen result.

Verification

The gate worked when its source was recorded; the chosen friction mode was eligible; the plan preceded side effects; full mode received an explicit human decision; execution stayed within the confirmed mode and technical boundary; and no required review, permission, validation, or safety control was bypassed.