agency-continuity-audit — Claude Code plugin: audit agent continuity
The agency-continuity-audit plugin for Claude Code and Codex checks persistent agent workspaces, separating evidence of goals, state, and recovery from self-reports.
MIT ★ 1,596 updated 2026-10-29
/plugin marketplace add hashgraph-online/awesome-codex-plugins
/plugin install agency-continuity-audit@awesome-codex-pluginsWhat agency-continuity-audit does
The plugin walks a persistent agent workspace and separates hard evidence from claims. It looks for an explicit objective or boundary, runs PRAGMA quick_check against SQLite state files, verifies that declared objective hashes still match, and checks whether memory corrections have a structured path. Restart recovery counts only when it was observed across different machine boots. Revenue figures need settled payouts and arithmetic behind them.
Everything stays local. No network calls, no dependency installs, and the workspace plus state directory are left untouched. On Codex the install line is npx codex-marketplace add revertcreations/agency-continuity-audit --plugin --project --yes. The deterministic script skills/audit-agency-continuity/scripts/audit_continuity.py takes --project and --state-dir; add --json for machine-readable output. Python 3.10 or newer is required, and SQLite checks apply when there is SQLite state to read.
Results come back as pass, warn or fail, and a warning means unproven rather than achieved. That is the honest reading of a scheduler boolean nobody independently probed. This is an evidence audit, not a security certification: a clean run says nothing about agent quality, safety or demand.
What it helps with
- tell a drifted agent objective from real progress
- verify SQLite memory integrity before trusting it
- spot stale restart evidence after a reboot
- flag revenue claims that lack settled payouts
- separate scheduler self-reports from independent probes
Example prompts for the agent
- Run a continuity audit on this agent workspace.
- Run
python3 skills/audit-agency-continuity/scripts/audit_continuity.py --project /path/to/workspace --state-dir /path/to/durable-state --jsonand walk me through the output. - Which checks came back
warn, and what evidence would clear them?
How to install agency-continuity-audit
Plugin marketplace
/plugin marketplace add hashgraph-online/awesome-codex-plugins/plugin install agency-continuity-audit@awesome-codex-pluginsStep by step, with commands for each platform: Claude Code plugins: marketplaces, installing, and building your own
Questions about agency-continuity-audit
How do I install agency-continuity-audit in Claude Code?
From the marketplace: run /plugin marketplace add hashgraph-online/awesome-codex-plugins first, then /plugin install agency-continuity-audit@awesome-codex-plugins.
What does agency-continuity-audit do?
It reads an agent workspace without changing it and checks the objective, SQLite integrity, the memory correction path, evidence freshness, objective hashes and observed recovery. Anything without proof is marked warn.
What do I need to run it?
Python 3.10 or newer. Integrity and correction-path checks kick in when agent state lives in SQLite; the audit stays offline and pulls no extra packages.
Source text
agency-continuity-audit · MIT
# 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 ```bash 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: ```bash 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.
© Revert Creations · MIT · hashgraph-online/awesome-codex-plugins