Agency Continuity Audit

A local, read-only Codex plugin that finds evidence gaps in persistent agent workspaces.

Long-running agents can appear healthy while their objective has drifted, their memory database is corrupt, their scheduler status is stale, or their revenue claim is only a boolean. This audit distinguishes durable evidence from reported claims and marks missing proof as unproven.

Install

npx codex-marketplace add revertcreations/agency-continuity-audit --plugin --project --yes

Then ask Codex:

Audit this workspace's agent continuity.

You can also run the deterministic auditor directly:

python3 skills/audit-agency-continuity/scripts/audit_continuity.py \
  --project /path/to/workspace \
  --state-dir /path/to/durable-state

Add --json for machine-readable output.

What it checks

• An explicit objective or boundary source exists.
• Durable state exists and SQLite files pass PRAGMA quick_check.
• A structured correction or supersession path exists.
• Continuity evidence is current rather than stale.
• Declared objective hashes still match the workspace.
• Recovery has actually been observed across different machine boots.
• Human-authority exceptions are represented instead of silently omitted.
• Commercial claims include mature, settled-payout evidence and arithmetic.
• Scheduled-unit health is labeled as reported evidence, not an independent probe.

The report uses pass, warn, and fail. A warning means unproven, not achieved.

Privacy and claim boundary

The plugin runs locally, makes no network requests, installs no dependencies, and does not modify the workspace or state directory. It reports cited paths and compact evidence; inspect output before sharing it.

This is an evidence-surface audit, not a security certification. It does not prove agent quality, safety, demand, or revenue merely because no failure was found. Scheduler and continuity booleans remain self-reported unless backed by independent observations.

Requirements

• Python 3.10 or newer
• Optional SQLite state for integrity and correction-path checks

MIT licensed. Issues and reproducible false-positive reports are welcome.