Agent audit reporter
Produces the periodic governance digest of all agent and skill activity — actions taken by autonomy level, approvals granted and by whom, guardrail triggers, denials, and anomalies — from GStack's audit logs. Use when the user mentions AI governance reporting, agent audit, "what did the agents do this week," compliance review of automation, or board/CISO reporting on AI usage.
When to run
- Recurring: weekly governance digest; monthly deep report (recommended)
- On demand: "audit the refund skill's activity" / before compliance reviews and audits
Computer scopes
- Audit log: agent/skill actions, autonomy levels, approvals, denials, guardrail events (read-only)
- Directory: approver identities and roles
- Config: the org's autonomy policy (the standard activity is judged against)
Workflow
- Aggregate the period's activity by skill and autonomy level: actions taken, drafts approved/rejected (and by whom), guardrail triggers, halts, and errors.
- Judge against policy: any action executed above its skill's declared ceiling, approvals by non-designated approvers, or bulk-approvals where item-level approval is required. These are findings, reported without euphemism.
- Detect anomalies: volume spikes per skill, unusual hours, repeated guardrail triggers on the same boundary (a skill probing its cage is a design smell), rising rejection rates (drift between what a skill drafts and what humans accept).
- Compute the trust ledger per skill: approval rate, correction rate, incident count — the evidence base for promoting or demoting autonomy levels.
- Write the digest in plain language with every claim linked to log entries; the deep report adds per-skill appendices formatted for external-audit consumption.
Workflow note
This skill audits others; it must be beyond reproach itself. It is read-only, its own runs appear in the logs it reports on, and it never summarizes away a finding.Guardrails
- Autonomy: L0. Strictly read-only; distribution of the digest follows the configured governance list and nothing else.
- Findings name skills and approvers of record, not rank-and-file individuals' performance.
- No log access beyond the audit scope; content of customer data touched by other skills is out of scope — this reports on actions, not payloads.
Output
Weekly digest: activity summary by skill/level → findings → anomalies → trust ledger deltas. Monthly: the deep report with audit-ready appendices.Eval fixtures
- Set: 4 weeks of synthetic audit logs with seeded events: 3 ceiling violations, 2 improper approvals, 1 bulk-approval violation, 2 anomaly patterns, plus clean baseline noise.
- Rubric: finding recall (all seeded violations surfaced); zero false findings on clean weeks; trust-ledger math exact; every claim links to log entries.
- Ship gate: finding recall 100%; false-finding rate 0.
MORE SKILLS
Governance, Security & Compliance · L2 block / L1 release
PII redaction gate
Scans outbound content — messages, exports, attachments, drafts leaving trusted boundaries — for PII/PHI and blocks flagged items pending human approval, with inline redaction suggestions. Use when the user mentions PII protection, data loss prevention, redaction, outbound compliance, or safe data sharing with external parties.
Governance, Security & Compliance · L1
SOC 2 evidence collector
Gathers audit evidence for SOC 2 and similar control frameworks from connected systems — access reviews, change logs, backup confirmations, training records — mapped to the control checklist with timestamps and provenance. Use when the user mentions SOC 2, audit prep, evidence collection, control testing, compliance audit season, or an auditor request list.
