Don’t read it. Run it.
The framework and the method ship as two free scans for Claude Code. Each one reads your repository, reports with file-and-line evidence, and closes with one verdict line. One house rule for both: prove, don’t claim. A level without evidence is reported as unknown, never guessed.
/plugin marketplace add abovebeyond-ai/plugins
/agent-scan: what acts on its own, and has it earned that?
Inventories everything in your repository that acts without a human typing: agents, LLM calls, MCP servers and their tool grants, webhooks, auto-deploys, scheduled jobs, and the credentials each one holds. Then it profiles every entry on thePAC framework’s six axes.
Autonomy and blast radius
Suggestion to autonomous, contained to irreversible. And whether the autonomy is earned from a track record or merely configured.
Reliability and thresholds
A lower confidence bound over judged outcomes, not a headline number. Does the bar rise with the blast radius?
Infrastructure and value
Enforced by architecture or only by policy. And what the thing is worth, so the findings rank by risk against value.
/plugin install agent-scan@abovebeyond /agent-scan
The framework’s nineteen questions answered against your own code, findings ranked by risk with the smallest fix that moves the level, and one verdict line to close:
agent-scan: not yet, A3 claimed, A1 earned
/hallucination-scan: what speaks unchecked.
A hallucination is not a model bug you can fix; it is an exposure you can measure. The question is never whether the model makes things up but what happens when it does. This scan finds every path where model output reaches a person or a decision and scores what stands in between: nothing, review, structure, or a verification gate, which is the ProveML bar.
Audit the repository
Run it without an argument and it inventories the speaking paths and scores the gate on each, with the prompt, the config value or the line of code as evidence.
Verify a text
Hand it a document and it applies the method to the text itself: every claim traced to a fact, every number sourced, every judgement held against a declared threshold. Plausible but unchecked is the failure it exists for.
/plugin install hallucination-scan@abovebeyond /hallucination-scan
The verdict names the count and the worst finding:
hallucination-scan: contradicted, 11/14 claims verified, balance has no source