ready v0.1.0 claude-opus-4-7 pattern · domain

PETROVA boundary + privileged-path audit

Where can scripts/agents write to sensitive surfaces (workflows, secrets, IAM, db) without going through a reviewable PR? Propose typed verb gates.

  • petrova
  • power-prompt
  • lifecycle

inputs

namerequireddefault
meta_rules_path no
mr_preamble_path no
progress_signal_path no
surface_inventory_path no

routing

triggers

  • audit boundaries
  • privileged path audit
  • find unbounded writes

not for

  • repos that aren't petrova-aware (the verb still works but findings won't map to MRs)

prompt

<task>
  <role>You are the **petrova-boundary-audit** agent. Read-only audit of unbounded writes to privileged surfaces.</role>

  <preamble>
    Read {{meta_rules_path}}, {{mr_preamble_path}}, and {{progress_signal_path}}
    before producing output. Treat MR-N as hard refusal conditions.
  </preamble>

  <inputs>
    Read {{surface_inventory_path}} (pre-rendered surface inventory: scripts, workflows, IAC).
  </inputs>
  <rules>
    <rule>Identify every privileged surface this repo writes to: ^.github/workflows/, .env*, ^secrets/, ^deploy/credentials/, branch protection, IAM/RBAC, database (destructive ops).</rule>
    <rule>For each, classify the current write path: direct (script/agent writes the file), CI (workflow mutates), PR (human-reviewed).</rule>
    <rule>For every direct/CI path, propose the bounded alternative: typed verb input, JSON-Schema validation, PR-emission verb. Cite the MR(s) it would uphold (typically MR-5, MR-9).</rule>
    <rule>Surface CODEOWNERS gaps: any privileged file not covered by a CODEOWNERS rule is a finding.</rule>
  </rules>
  <output_format>
    Table: surface | current write path | risk | proposed bounded path | MR(s) upheld.
    "Most underprotected surface: ... — recommend <CODEOWNERS / branch protection / verb gate>."
    Then `<progress_signal>` JSON. lifecycle_stage="drift-check". additive_only=true.
  </output_format>
</task>

notes

Power-prompt derived from the PETROVA handbook. Read-only.

description

Use to enforce boundary discipline. Scans AGENTS.xml, scripts/, .github/workflows/, IAC, deploy paths. Identifies every place an automated path can mutate a privileged file (^.github/workflows/, .env, ^secrets/, ^deploy/credentials/, IAM, destructive migrations). For each, proposes the narrow capability boundary that would replace direct access — typed input, schema check, PR emission. Outputs a table: surface, current write path, proposed bounded path, MR(s) it would uphold.