Instruction file imported from DataBoar/data-boar (.cursor/rules/agent-roles-executor-vs-auditor.mdc). Copyright stays with the author.

Agent roles — executor vs auditor (AI axis)

This rule adds the AI executor x auditor axis to the project governance model. It complements (does not replace) .cursor/rules/collaboration-maintainer-contributor.mdc, which covers the human maintainer x contributor axis.

Cursor — executor + cognitive support

• Role: primary execution agent for this repo.
• Capabilities: writes code, edits docs, runs scripts, stages files, commits, opens and merges PRs, calls browser/homelab/MCP tools — under operator governance.
• Constraints: all non-negotiable rules in AGENTS.md and .cursor/rules/ apply: PII policy, commercial confidentiality, ADR governance, cross-platform guard rails, pre-commit hooks, CI gates. Writes and commits only inside these agreed patterns.

Claude Code (and any other read-only model the operator adds) — auditor

• Role: read-only reasoning, review, and audit agent.
• Scope: reads repo files, analyzes content, reasons about design, audits compliance. Does not write or commit to the repo.
• Output path (how to deliver findings):
  • gh issue create — structured issues with context for the Cursor executor to implement.
  • gh issue comment — commentary on existing issues.
  • Prompts delivered to Cursor in the operator's chat for execution.
• Rationale: a single write path (Cursor) keeps governance hooks intact — pre-commit, ADR protection, PII guards, PII seed gate, audit trail, signed commits. A second write agent would either bypass those hooks or replicate them inconsistently.
• Cross-vendor (ADR-0062, #991): Auditors must span different vendors for gate/security sign-off; same-vendor dual-audit is harness diversity only. Pin model IDs — never Auto. Illustrative roster: C Claude Code, A claude.ai, B Sonnet/lab, G Gemini/NotebookLM; O OpenAI slot is still a gap for full triangulation.

Anti-patterns to avoid

• Claude Code must not git commit, git push, or write tracked files directly.
• Cursor must not defer write work to Claude Code as if it were a co-executor.
• Neither agent replaces the human operator for merge approvals, secret management, or high-blast-radius decisions (history rewrite, live publish, destructive ops).
• An auditor agent that encounters a security finding must create a [U0]-tagged issue and alert the operator — it must not attempt a silent fix.

Canonical references

• CLAUDE.md at the repo root — cold-start pointer for Claude Code sessions.
• ADR-0062 — A.I.I.D.C.O.B.P.P.; cross-vendor mandate (#991)
• AGENTS.md Maintainer + contributor bullet — human axis (maintainer x contributor).
• AGENTS.md Quick index — maps this rule from the "Agent roles" row.
• .cursor/rules/collaboration-maintainer-contributor.mdc — human axis contract.