Workos Audit Harness — Security skill for Claude Code

workos-audit-harness

Audit logging for AI coding agents: one event vocabulary, six working integrations, and an honest account of where each place you can emit from tops out.

**If you build an agent harness, emit audit events from your own backend.** Your servers already see the session, the model, and every tool call the agent makes, and a customer's developer cannot rewrite them there. That is the only place an event's *content* can be attested, and it's the architecture we recommend. Start with the guide [Add enterprise-grade audit logs to your AI harness](#add-enterprise-grade-audit-logs-to-your-ai-harness) below and the shared event taxonomy in [`packages/audit-core`](packages/audit-core/src/harness-audit-schemas.mjs).

**If you operate a fleet and your vendors haven't done that yet,** instrument the endpoint instead. The plugins here emit from Claude Code, Codex, OpenClaw, OpenCode, Hermes, and pi today, with no vendor cooperation required. Events are authenticated per device and attributed server-side, but their content is composed on a machine the user administers, so it can be fabricated or withheld; see the [trust model](packages/proxy/README.md#trust-model) before building on the result.

The two compose. A vendor-emitted spine (session, turn, and tool-call ids from the backend) correlated with endpoint-emitted enrichment (repo, branch, `cwd`, device, local approvals) gives you both halves: the backend sees the conversation, the endpoint sees the machine, and claims from one that don't reconcile against the other are detectably false.

This repo contains six agent integrations, a fleet-deployment proxy and a chat console over the resulting audit trail, sharing one CLI harness and one set of audit schemas:

Package	What it does
`packages/audit-core`	Shared core for every inte