A desktop coding agent with free access to frontier models.
Scenarios
Start with the feature you want to demonstrate or test. Everything for onboarding is in onboarding/.
scenarios/
onboarding/
world.ts environment and synthetic users
workflow.ts ordinary async TypeScript: the user's actions
e2e.test.ts assertions against UI and server witnesses
screenshots.ts documentation captures using DocShot
video.tsx ordinary React/Remotion composition
empty-chat.tsx illustrated ending, specific to this scenario
remotion.tsx standard Remotion entry
render.mjs command to render a completed capture
Files are optional. Reuse a workflow from tests, recordings, or screenshots; do not copy the journey into each output. Workflows return observations, while tests make claims about those observations. No registration wrappers or custom presentation language are needed.
Find and run
Install both workspaces from the repository root:
pnpm install
pnpm --dir evals install
node evals/bin/evals.mjs --list
The scenarios package belongs to the evals workspace so it shares the existing testkit and behavior packages. Presentation components belong to the root workspace.
Run onboarding (Daytona when authenticated; the CLI prints placement):
OMNIRUSH_EVAL_FILM_DIR="$PWD/evals/results/onboarding-capture" pnpm evals:e2e onboarding
pnpm --dir scenarios onboarding:render "$PWD/evals/results/onboarding-capture"
Use a new capture directory for each run. Without the directory override, the world uses its own temporary directory. The journey has isolated Den accounts and an isolated email outbox; it downloads the public Linux installer from GitHub. It does not authorize Notion or Linear accounts or install the desktop app.
Outputs: capture.json, frames/*.jpg, downloads.json,
screenshots/download.png, and (after rendering) onboarding.mp4.
The test's assertion evidence remains under evals/results/test-runs/.
Render a review frame with --still=300. The renderer accepts
REMOTION_BROWSER_EXECUTABLE when a specific Chrome executable is needed.
The browser remains full screen. Frame timestamps control playback; recorded
typing is not accelerated. The last eight seconds hold the completed download
and illustrate opening an empty chat. That illustration is not proof of installation.
Reuse existing pieces
| Need | Start here |
|---|---|
| Existing environment | worlds/ and evals/worlds/ |
| User, probe, and evidence channels | evals/packages/testkit/ |
| Paced field input | typeField, readableTyping from @omnirush/behaviors |
| Browser recording and download events | captureBrowserFilm from @omnirush/cdp |
| Settled documentation screenshots | evals/docs-shots/gate.ts and loop.ts |
| Browser frame and recording playback | packages/presentation/ |
world.ts configures an existing world factory. Keep provisioning machinery in
the existing packages. Use await typeField(user, target, value) for one complete
field at a time; it focuses once, replaces existing text, types graphemes with a
deterministic cadence, and verifies the final value. Set sensitive: true for
password inputs; the driver requires masking and excludes the value from its trace.
Existing DocShot scene IDs remain discoverable through
node evals/docs-shots/run.ts --list. The onboarding screenshot reuses that
engine on its already-open browser, avoiding a second provisioned environment.
Add another scenario
Create a named folder and export plain functions. Add e2e.test.ts only for a new
user journey, using spec.world from the existing testkit. Scenario tests use the
same CLI, channel rules, boundary checks, and evidence publisher as
legacy specs. The secret-bearing Daytona PR workflow retains its existing legacy-spec
selection; it does not automatically select scenario tests. Run a scenario explicitly
with the CLI and publish its evidence on the PR. Tests are discovered by filename without importing workflows.
For an attached live-service check, use live.test.ts and name it explicitly:
pnpm evals:pr ../scenarios/<name>/live.test.ts. Declare the existing testkit's
requirements and consent variables; missing dependencies must report a skip.
Live checks stay out of default suites. No placeholder scenarios are registered.
Keep pure helper tests beside their implementation in shared packages, outside
scenario journey folders. The older evals/specs/ tests continue to work; migrate
them only when their scenario is being worked on.