Skip to content

No skill owns change-safety: when may an agent act alone, and what counts as proof? #5

Description

@kuyacarlo

The pack covers how to work (operations, git-signed-commit, continuous-improvement) and how to review (thorough-code-review, specification-compliance, code-simplification), but nothing states the two rules an agent needs before it touches infrastructure, CI, a deploy, or a live system:

  1. Am I allowed to just do this, or does a human have to decide? Without an answer, agents either ask permission for everything — which costs more than the mistakes do and leaves infrastructure nobody will touch — or they act on something genuinely irreversible.
  2. What counts as proof that it worked? "The tests pass" is routinely claimed for changes whose user-visible artifact was never exercised, and red CI that predates the branch gets debugged as if it were the diff.

Neither question is repo-specific, so neither belongs in a project AGENTS.md. Both currently get re-derived per session, usually after something breaks.

Also missing: a rule for where a lesson goes once a session learns one. continuous-improvement writes project-scoped lessons to .agents/LEARNINGS.md, which is right for constraints that die with the task — but there is no stated route for a lesson that outlives the task, and no rule that a portable lesson must be generalized rather than copied with its original repo paths attached.

Proposed: one skill holding the change-safety spine, deliberately narrow, routing to the existing skills for everything they already own.

Metadata

Metadata

Assignees

No one assigned

    Labels

    No labels
    No labels

    Projects

    No projects

    Milestone

    No milestone

    Relationships

    None yet

    Development

    No branches or pull requests

    Issue actions