Curtis AI Audit

Local-first audit trail for AI agent activity in your Obsidian vault. Every agent inference gets an append-only markdown receipt — a journal you keep on your own agents, not a forensic system, and honest about the difference (see SPEC.md).

Speaks OM-1 (Obsidian Messenger): agents, other plugins, or a shell one-liner can all file receipts — see INTEGRATION.md. The messenger plugin is not required — the audit inbox is just a file.

What it does

• Receipts — one audit.v1 block per agent action: who, what, when, model, duration, tokens, status, plus a rendered human summary.
• Per-day ledger — append-only shards under Audit/ledger/YYYY-MM-DD.md. Envelope = the receipt as canonical JSON; body = the summary. Readable in Obsidian, parseable by anything.
• Ingest — polls the OM-1 audit inbox (Messenger/audit/inbox.md) with proper cursor semantics; junk is rejected, redeliveries are deduped.
• Agent hooks — Install agent audit hooks writes vault-scoped configs (Claude Code, OpenCode) so local agents file their own receipts through the bundled writer (bin/curtis-audit-file.mjs); Codex and Gemini get manual recipes. Metadata only — tool input content is never recorded. Source installs only: the writer ships with the repo, not with plugin releases (Paths A and B work everywhere).
• Direct API — in-process plugins can skip the bus: window.curtisAudit.record({ agent, model, tokens, ... }).
• Query — the read side: window.curtisAudit.query({ agent, status, since, ... }) for status bars, dashboards, and other plugins.
• Verify — detects malformed blocks, duplicate ids/sequences, reordered shards, and per-agent sequence gaps (deletions leave holes; no crypto, no false promises).
• Digest — a rolling Audit/digest.md: counts, statuses, tokens, durations, a per-day breakdown, errors.
• Retention off by default — an audit trail that silently ages out is a policy contradiction. Opt in if you want it; each prune is itself recorded as a receipt, so the gaps it leaves are self-explaining.

Install

From a release: download main.js, manifest.json, and styles.css into <vault>/.obsidian/plugins/curtis-ai-audit/, then enable the plugin in Obsidian (Settings → Community plugins). Once accepted into the community directory it installs from there directly.

From source: clone this repo into <vault>/.obsidian/plugins/curtis-ai-audit/, run npm install && npm run build, and enable it in Obsidian.

Usage

Ribbon icon or commands:

Other plugins (same window):

const receipt = await window.curtisAudit.record({
  agent: "curtis-ai-chat",
  action: "inference",          // optional, default "other"
  model: "gpt-4o",
  provider: "openai",
  durationMs: 1234,
  tokens: { input: 812, output: 341 },
  subject: "Notes/question.md",
  summary: "Answered a question about the spec",
});
// also: .poll(), .verify(), .digest(days?), .query(filter)
// workspace event: "curtis-audit:receipt"

The read side, for dashboards and status bars:

const errors = await window.curtisAudit.query({
  agent: "curtis-ai-chat",
  status: "error",
  since: "2026-10-01",   // bounds apply to `recorded` (local write time)
  order: "desc",         // newest first (the default)
  limit: 20,
});

Anything that can append to a file (other device, script, agent):

<!-- messenger: {"id":"msg_x","from":"research-agent","to":"audit","ts":"2026-10-10T12:00:00.000Z","replyTo":null} -->
{"action":"inference","model":"gemini-2.5-pro","summary":"Summarized a paper"}
<!-- /messenger -->

See INTEGRATION.md for wiring up curtis-ai-chat.

Testing

npm test        # 107 unit tests, in-memory vault, no Obsidian needed
npm run build   # typecheck + bundle to main.js

Status

v0.2.2 — the local-first tier, deliberately. Tamper-evidence, retention enforcement, and reporting are the next tier up and are not pretended here. Requires Obsidian 1.13.0+ (declarative settings).

License

MIT © Jordan Newell