Skip to content

scope-guard ​

Bucket: discipline · Invocation: automatic

Keeps an active change aligned with the developer's requested outcome by using two axes: relation to the mission and available mandate.

Classify discoveries on both axes:

  • Required / adjacent / unrelated: whether the discovery is needed for correct delivery, useful but unnecessary now, or not relevant to the mission.
  • Covered / new decision or external action: whether the existing request and policy already authorize it. A ready issue is not authorization.

Risk guides implementation and verification; it can expose a missing decision but does not itself grant new permission. Required work normally stays in the mission. A material product or boundary change returns to plan-work; adjacent work can be offered as a follow-up, and unrelated work gets no invented proposal. Inspect the final diff for unrelated edits and preserve pre-existing developer work.

Transcript: useful, but not required ​

Trimmed from trigger case SG-P1 on 2026-07-30.

text
User:
"While implementing this small fix you find an old module that should
be refactored. Keep the task narrow and handle the discovery correctly."

Agent:
"I am using scope-guard because the refactor is outside the active fix."

scope-guard:
"Classify the discovery by relation to the mission and by whether the current request covers it."

Agent:
"The old-module refactor is adjacent work and not needed for this fix. I
will leave it outside the patch and record it as a follow-up only if the
current mandate allows that tracker write."

Local improvement and follow-ups ​

Cohesive behavior-preserving extraction on the touched surface is allowed. The implementer can reject unsupported review requests with evidence. For worthwhile adjacent work, search for an existing issue and draft the problem, impact, and acceptance condition. Create or update it only under existing request or project-policy authority that does not conflict with an explicit read-only instruction; otherwise ask one consolidated follow-up question. An optional backlog decision does not delay authorized delivery. Rejected preferences need no ticket.

A personal framework, published in the open.