Skip to content

Agent skill: audit-and-fix loop over validation output #117

Open

Description

@mindsocket

Agent skill: audit-and-fix loop over validation output

Summary

A skill that teaches an agent to consume structured-context validation output and drive an audit-and-fix loop — rather than a documentation page. Parse the report, triage by severity/category, fix what's mechanical, consult the user for judgment calls, re-validate, iterate.

Complements #93 (response-ready report format): #93 is the output contract; this skill is the reference consumer.

Behaviours

  • Respect cascade annotations: fix upstream failures before re-evaluating downstream — don't chase suppressed errors that will resolve themselves (see Layered rule evaluation: cascade gating and coverage annotation #115, layered rule evaluation).
  • Breadth-first: address issues at the top of the hierarchy before deeper layers; re-request evaluation of lower layers only after the top is clean.
  • Distinguish deterministic fixes from judgment fixes: dangling links, missing fields, invalid parents are mechanical; framing, coherence and quality issues need user input or semantic evaluation. The skill should know which is which and hand judgment calls to the user.
  • CLI only: validate, dump, show, schemas — no library dependency, so the skill works for any agent runtime.

Relates

Activity

Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

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