First run

This assumes the server and a worker are up (npx azhi doctor is green). If you work on Azhi itself from a checkout, see the contributor sequence at the bottom.

Run the CI digest example

1. Register the example’s tools and store the Slack token.

   npx azhi apply examples/ci-digest/azhi.config.yaml
   npx azhi secret set slack-bot-token --value xoxb-...

2. Run it and wait for the result.

   npx azhi run examples/ci-digest -i team=payments --wait

3. Inspect it: state, attempts, usage and the action ledger.

   npx azhi inspect <run-id>

Without a Slack workspace, point the server at a fake Slack API for local runs: AZHI_SLACK_API_URL=http://127.0.0.1:<port>/api (see src/testing/fake-slack.ts).

Run the flagship: weekly quality report

The flagship needs a knowledge dataset and, for real model calls, a provider key and model.

npx azhi apply examples/quality-report/azhi.config.yaml
npx azhi dataset create quality-guidelines
npx azhi dataset add quality-guidelines examples/quality-report/knowledge
npx azhi dataset publish quality-guidelines --tag approved
npx azhi secret set anthropic-api-key --value sk-ant-...
export AZHI_ANTHROPIC_MODEL=<model id>   # set on the server; profiles say `name: default`
npx azhi plan examples/quality-report     # capabilities, coverage, taint paths, blockers
npx azhi test-node metrics examples/quality-report --fixture fixtures/metrics.json
npx azhi run examples/quality-report -i team=payments --wait
npx azhi inspect <run-id> --context analyse

The example profile uses Anthropic. To use OpenAI, set model: {provider: openai, name: default} in profiles/quality-analyst@1.yaml and use openai-api-key and AZHI_OPENAI_MODEL.

What to look at

Command	What it shows
azhi plan	Every requirement marked, every policy’s coverage, the taint path analyse → report → post and its gate, and any blockers
azhi test-node	One node on fixture outputs with writes mocked (a test run)
azhi inspect <run> --context analyse	The context manifest: what the model saw on each turn
azhi open <run>	A link to the run page with graph, timeline, ledger, policy coverage and usage

See the taint rule fire

Delete the guard: line on the Slack post in the flagship and run azhi plan again. You get tainted_write_ungated. That is the same check that protects you in production.

The alpha demo

export AZHI_DEMO_DATABASE_URL=postgres://azhi:<password>@localhost:5432/azhi   # admin connection
export AZHI_TEMPORAL_ADDRESS=localhost:7233
npm run demo                 # checks 2 to 6, with a scripted model unless ANTHROPIC_API_KEY is set
npm run demo -- --install    # also check 1: runs deploy/install.sh on this host and times it

The six checks: install, plan, scheduled run with citations, crash mid-post (exactly one message), a trust refusal for an untrusted author, and a 30-fixture executor comparison.

Contributor sequence (run the server from a checkout)

npm install
npm run build:web
docker compose -f deploy/docker-compose.yml up -d --wait
export AZHI_DATABASE_URL=postgres://azhi:azhi@localhost:5432/azhi
export AZHI_TEMPORAL_ADDRESS=localhost:7233
npx azhi server start      # terminal 1
npx azhi worker start      # terminal 2

Tests: npm run typecheck, then npm run test:local (vitest on PGlite).