Developing the wizard
Contents
This page covers running the wizard locally and testing a change.
Running it locally
Pick the smallest setup your change needs.
- Wizard only. Run your wizard branch against production PostHog. Reach for this for most wizard changes.
- Workbench. Test new wizard code together with a context mill branch, on the workbench's test apps.
- Fully local. Run PostHog, the AI gateway, the wizard, and the MCP server all on your machine. Reach for this only when a change sweeps across the whole stack.
Wizard only
Terminal
pnpm try runs from source and logs in with OAuth against production. Any app works. The workbench apps are known to work, so copy one out and run git init in it first.
Workbench
Hand this prompt to Claude Code or Codex to set it up:
If you need to wrangle a run by hand, for artisanal reasons, these phrocs panes are useful:
- context-mill serves your context mill branch on port 8765, and rebuilds when you edit a skill.
- wizard-build builds your wizard branch and rebuilds on every change.
- wizard-run runs the wizard on a test app you pick. wizard-tail-run follows its log.
- wizard-ci-local-run runs a CI run and grades it, without opening a PR.
- wizard-ci-snapshots captures every screen of a real TUI run.
- wizard-run-proxy and mitmproxy show every request the wizard makes.
- wizard-build-dev builds against a local PostHog, for the fully local setup.
Fully local
Run everything on your machine when one change sweeps across PostHog, the AI gateway, the wizard, and the MCP server. The full stack recipe in the wizard repo has the steps.
| Service | Port |
|---|---|
| PostHog | 8010 |
| AI gateway | 8080 |
| Context mill | 8765 |
| MCP server | 8787 |
| mitmproxy | 8888 |
Testing your change
| Test | Command | What it proves | Runs in CI |
|---|---|---|---|
| Unit and snapshot | pnpm test | Logic and TUI frames | Every wizard PR |
| Smoke | ./scripts/smoke-test-ci.sh | A real run on a test app | Every merge to main |
| End to end | /wizard-ci <app> PR comment | A real agent run, optionally against a context mill branch | On demand and weekly |
| Benchmarks | See benchmarking docs | Compares models and harnesses | Manual |
Releasing
See releasing the wizard for how the wizard and the context mill ship, and what to test before you release.











