Trust Gate

Trajectory-aware regression gate for autonomous coding agents.

Agents touch code all over a codebase, and "run the tests near the changed files" tooling misses regressions with indirect blast radius (a shared util breaks something three layers away through a DI container, and nothing flags it). Trust Gate unions static graph analysis (diff -> affected symbols -> import/call graph) with runtime test coverage to rank the tests actually at risk, runs them, and distinguishes a pass-to-pass regression (passing before, broken by this diff) from a pre-existing failure or a merely-new test — cheaper than a full suite run, and catches more than changed-file-only heuristics. An optional LLM judge (BYO key) scores the diff against the agent's own stated intent as an advisory signal.

Two entry paths onto the same engine, both landing in one dataset:

• Agent path (primary, trajectory-native) — a coding agent calls check_regression inline over MCP, in its own working tree, before it ever opens a PR. Runs on your machine, not any server.
• PR path (safety net) — a GitHub App webhook fires on PR open/sync, a reusable GitHub Actions workflow runs the same engine against the PR, and results post back as a real GitHub Check + to a self-hosted dashboard.

Benchmark

BENCHMARK.md — the union-with-coverage thesis, checked against vitest --changed on purpose-built fixtures modeling the dynamic-import/DI/registry blind spot: Trust Gate 3/3, vitest --changed 0/3 (plus a static-import control case both catch, so it isn't just "vitest is broadly worse"). Reproduce it yourself: bun run benchmark.

Self-hosting

Docker Compose brings up the dashboard, API, Postgres, and Redis. See docs/ for the full walkthrough (GitHub App setup, env vars, access control) — short version:

bash scripts/setup-env.sh   # scaffolds both .env files
# fill in the GitHub App fields -- see docs/01-github-app.md (or automate it:
# scripts/create-github-app.mjs)
docker compose -f deploy/docker-compose.yml up -d --build

Dashboard at http://localhost:3000, API at http://localhost:3001.

Prefer a released image over building from source? See docs/02-docker-compose.md for docker-compose.prod.yml, pulling from GHCR instead.

For a real (non-localhost) deployment, put a reverse proxy in front for TLS — see docs/05-reverse-proxy.md.

Connecting an agent (MCP path)

@trust-gate/cli is on npm — npx -y @trust-gate/cli mcp, no local clone needed. See docs/03-agent-setup.md for MCP config.

Developing on this repo

bun install
docker compose -f deploy/docker-compose.yml up -d postgres redis
bun run db:migrate
bun run --cwd apps/backend dev

bun run lint       # Biome
bun run typecheck  # TypeScript across all workspaces
bun run test       # Bun test across all workspaces
bun run build      # build all workspaces

Monorepo layout

packages/
  core-graph/     diff -> affected symbols -> import/call graph (ts-morph)
  coverage-map/   per-test V8 coverage -> test<->source map
  impact/         union static+coverage -> ranked at-risk tests
  test-runner/    vitest adapter
  orchestrator/   glue: git diff/stash, ts-morph project loading, ties the above together
  judge/          LLM judge (advisory diff-vs-intent scoring, BYO key)
  benchmark/      catch-rate benchmark vs `vitest --changed` -- see BENCHMARK.md
apps/
  cli/            @trust-gate/cli -- CLI + stdio MCP server (agent path)
  backend/        Fastify API, BullMQ workers, GitHub webhooks + Checks API, /mcp over HTTP
  dashboard/      Next.js dashboard -- PR results, judge scores, sign-in via GitHub OAuth
deploy/
  docker-compose.yml
docs/
  self-hosting walkthrough, access control, CLI publishing plan

License

Apache-2.0 — see LICENSE.