leonard_grid_assets #

Lenny's grid assets for a the_grid station: the station-driven selfdrive and leonard-e2e circuits, their shared capability registry, and the deterministic E2eCommand. The E2E service preflights a wired iOS device, launches a fresh Flutter app, invokes leonard_cli, and derives one structured verdict from the typed trajectory. Each invocation explicitly pins Leonard's packaged operating guide so dependency-launched CLI snapshots cannot silently fall back to a goal-only prompt. The private sample-suite mode composes four calls to that same operation against packages/leonard_flutter/example/sample_app.

The development runner registers only the Command:

dart run tool/e2e.dart e2e \
  --goal 'Open Settings' \
  --app-dir /absolute/path/to/flutter_app \
  --expect-route settings

dart run tool/e2e.dart e2e --sample-suite --device DEVICE_ID

dart run tool/e2e.dart e2e \
  --sample-suite \
  --sample-scenario login \
  --model qwen-mlx \
  --device DEVICE_ID

--sample-scenario is valid only with --sample-suite; its accepted values are login, navigation, state_change, and scroll. Omitting it preserves the four-scenario order. The service API follows the same composition: E2eService.runSampleSuite defaults scenarios to kLeonardSampleSuite, while direct performE2eSampleSuite callers must provide a non-empty scenario list.

Both modes emit exactly one JSON object. A child's prose and exit status never decide the result; SessionHeader, TurnRecord, and SessionFooter evidence does. The four circuit capability ids are e2e-preflight, e2e-launch, e2e-run, and e2e-inspect. buildLeonardRegistry composes them and the existing selfdrive capabilities over the ordinary code registry.

Every driver run stores Leonard CLI's raw attachment probe as probe.json beside trajectory.jsonl. The probe is diagnostic evidence only; the typed trajectory inspector remains the sole verdict source.

The required corrected device pair is leonard_agent 0.3.1 with leonard_cli 0.3.0. Device probing verified that application registration and VM attachment were healthy under 0.3.0: the attached isolate returned populated core tool descriptors, routes, and semantics. The qwen-mlx provider instead represented schema-declared numeric arguments such as node_id and seconds as JSON strings. Strict pre-dispatch validation exhausted its retries, after which the terminal synthetic failure record carried an empty observation. Agent 0.3.1 fixes that boundary by losslessly normalizing schema-declared numeric strings before applying the unchanged strict validation.

This package does not compose itself into a station and its manifest contains no station grid: block.

Part of the lenny workspace.