v0.1.3 npm install @kontourai/veritas

Veritas

Bespoke lint for AI agents

You define what your repo considers mandatory. Veritas tells any AI agent exactly which changes violated those rules — before they finish work.

View on GitHub → npm package
veritas shadow run

$ veritas shadow run --format feedback


veritas 0.1.3 — shadow run

adapter: .veritas/repo.adapter.json

policy: default.policy-pack.json


PASS artifacts: CLAUDE.md exists

PASS artifacts: .veritas/repo.adapter.json exists

PASS governance-block: CLAUDE.md contains canonical block

PASS governance-block: AGENTS.md contains canonical block


WARN if-changed: src/api/ changed

      → tests/api/ must appear in the diff

      → missing: tests/api/rate-limit.test.ts


FAIL proof-lane: npm test

      → exit 1 — 3 tests failed


───────────────────────────────

2 pass   1 warn   1 fail

evidence → .veritas/evidence/run-1748220495.json

Three pillars
01

Rules

Repo-local lint in `.veritas/`

Define required files, governance blocks, companion test requirements, and proof lanes. Your repo, your rules — not a cloud platform.

artifactsgovernance-blockif-changedproof-lanes
02

Feedback

`veritas shadow run` → lint output

Agent-facing path that produces lint-style output directly into your agent's context window. Exit codes 0 (pass), 1 (fail), 2 (error) — designed for hooks.

shadow-runexit-codeshook-integrationformat:feedback
03

Improvement

Local evidence + eval history

Evidence artifacts persist per run. Eval records track whether the agent accepted or rewrote. Summaries measure if Veritas is actually helping over time.

eval-drafteval-recordmarkerssummaries
Agent activation

Works with any agent
that reads your repo.

Veritas injects a governance block into every AI instruction file in your repo. Every agent that reads those files — Claude Code, Cursor, Codex, Copilot — sees the same rules and proof requirements.

Stop-hook integration wires veritas shadow run to fire at session end for Claude Code and Cursor — no manual invocation needed. Codex integration uses a native hook file.

CLAUDE.md
Claude Code
AGENTS.md
Codex
.cursorrules
Cursor
copilot-instructions.md
GitHub Copilot
CLI surface
veritas init Bootstrap adapter, policy pack, governance blocks
veritas shadow run Agent-facing lint path with feedback output
veritas report Full policy report (JSON or summary)
veritas runtime status Check hook wiring and activation state
veritas eval draft Record agent evidence for a run
veritas eval record Accept or reject agent output manually
veritas eval summary Aggregate improvement metrics over time
veritas apply stop-hook Wire Claude Code / Cursor stop hooks