leonard_cli 0.4.3-dev.1

CLI frontend for Leonard.

Changelog

0.4.3-dev.1

• The bundled leonard-drive agent, drive-with-leonard skill and AGENTS.md template read observation.core.routeLayers (the routes on screen) in place of the deprecated routeStack (lenny-kyfdc). leonard_agent rises to ^0.3.4-dev.2, which carries the field.

0.4.2

Stable promotion of 0.4.2-dev.1 on Nico's promotion call (2026-10-21). One change on top: the runtime leonard_native constraint rises from ^0.6.0 to ^0.6.1, so depending on leonard_cli brings in the lenny#198 and lenny#200 backend fixes without a separate pin. The prerelease entry below carries the details.

0.4.2-dev.1

Fixes lenny#199: the lenny#160 fix found the native host through the package config, but leonard_native was only a dev dependency, so a consumer that depended on leonard_cli alone had no leonard_native to find and leonard_drive up fell back to paths that exist only inside a lenny checkout.

• leonard_native (^0.6.0) is now a runtime dependency of leonard_cli, so depending on leonard_cli puts it in your package config. You no longer need to add leonard_native yourself to use up's native path.
• When up cannot find the native host and no --native-host was passed, it exits 64 with an error that says leonard_native is missing from the program's package config and how to fix it (add leonard_native to your dependencies, or pass --native-host), instead of the generic "requires --native-host" message. The working-directory fallback still serves development inside a lenny checkout.
• --summary-json now carries the run's verdict: verdict (pass, fail or infra), a one-line reason, and digest, the last turns as one line each (action, result, and the final rationale) under 3000 UTF-8 bytes. A run fails only when Leonard drove the app and the goal was not met (budget exhausted, the app's VM service gone mid-drive, a golden mismatch); every outcome that kept it from judging the app is infra, including goldens that could not be read or written, which now exit 70 rather than 1 (lenny#177).
• Add --gate-exit-codes: exit 0 pass, 75 infra, 1 fail. 75 is EX_TEMPFAIL, Gas City's infrastructure code for checks. A usage error exits 75 too (lenny#177).

0.4.1

Stable promotion of 0.4.1-dev.1 on Nico's promotion call (2026-10-20), on stable leonard_agent 0.3.3 and leonard_native 0.6.0. The prerelease entry below carries the full list. One change on top:

• --model claude now defaults to claude-sonnet-5-5 (was claude-sonnet-4-6), with vision on. On the sample-app goal it finished in 20 s for $0.26, against 35 s and $0.35 for Sonnet 4.6. Pass --model-id claude-sonnet-4-6 for the old default.

0.4.1-dev.1

• Raise the leonard_agent floor to ^0.3.3-dev.2, so a run's usage counts model calls its turn budget abandoned (lenny-c55p6).
• Raise the leonard_native floor to ^0.6.0-dev.1 (lenny#180 to #183).
• Add --summary-json <path>: writes the run's exit code, outcome, turn count, duration, trajectory path and usage totals as one JSON file, so CI can collect token and cost figures across runs (lenny#178).
• Add --price-input-per-mtok / --price-output-per-mtok, and the LEONARD_PRICE_INPUT_PER_MTOK / LEONARD_PRICE_OUTPUT_PER_MTOK environment pair, to override the built-in price table for the cost estimate. The flags outrank the environment.
• Trajectories record per-turn and per-run model usage; the header's harness_version is now 0.6.0.
• Require leonard_agent ^0.3.3-dev.1.

0.4.0

Stable promotion of 0.4.0-dev.1: the same content, published on the stable rung on Nico's promotion call (2026-10-15), after the Android (Pixel 7a), iOS 26.5 simulator and macOS live suites passed. The prerelease entries below carry the full list.

0.4.0-dev.1

• Breaking — leonard_drive up's native dual path takes --app-id instead of --app. The native host now attaches to the app flutter run launched, on iOS and Android alike, instead of being handed the build to install; on Android that relaunch would have dropped the Flutter VM-service channel. Migration: replace --app <Runner.app> with --app-id <bundle id or package>. A stale --app exits 64 with this migration message.
• Requires leonard_native ^0.5.0-dev.1 and leonard_agent ^0.3.2-dev.2.
• Fix: leonard_drive up finds the native host through the package config, so --native-host auto-resolution works outside the lenny repo (lenny#160, #166).

0.3.1-dev.1

• Add ACP harness inference: drive the loop through a leonard_acp harness and select it among the registered agent specs.
• Add --wall-clock-minutes, the whole-session wall-clock budget (default 15).
• Fix: provider-build errors are caught only for configuration failures.

0.3.0

Promotes 0.3.0-rc.1 to stable. The candidate's full change list is under that entry below.

• Breaking: requires leonard_agent ^0.3.0 and adopts its I/O entrypoint. Migration for anything embedding this package's library surface: import package:leonard_agent/leonard_agent_io.dart for owning connections.
• Fix: core tool descriptors are read from the handshake and passed as the driver's coreTools instead of being projected as bare names, so the model sees each core tool's description and input schema (lenny#123).
• Add: tool/live_macos_smoke.dart — the live macOS smoke that gates a leonard_flutter release (lenny#120).

0.3.0-rc.1

• Breaking: requires leonard_agent ^0.3.0-rc.1 and adopts its I/O entrypoint. Migration for anything embedding this package's library surface: import package:leonard_agent/leonard_agent_io.dart for owning connections.
• Add observation frame goldens: leonard_drive persists observation frames and compares them against a stored golden.
• Add the test-with-leonard skill router.
• Provider model ids are configurable, and driver token and effort defaults are exposed as operator overrides.
• The portable pure-Dart mutation runner ships with its Flutter delegate; dry output counts as sizing evidence, and zero uncovered mutation lines is accepted rather than failed.
• Self-drive: completion is gated on observed timeline evidence, a model must be selected before Start, diagnostic evidence survives an observation failure, and panel session markers are durable.
• Raise the sibling floors to the published wave: leonard_contract ^0.2.2, leonard_host ^0.2.2, leonard_native ^0.4.1.

0.2.1

• install --copilot now overlays .github/agents/ — the location GitHub Copilot (CLI, coding agent, VS Code) actually scans — instead of the retired repo-root agents/+skills/ plugin layout. Skills need no overlay: .agents/skills/ is a first-class Copilot skill location, and gh skill install memento-engineering/lenny drive-with-leonard installs straight from this package's vendored assets.
• The vended leonard-drive agent's tools list carries both harness vocabularies (Bash, Read, execute), so the same file drives Claude Code and Copilot unchanged.

0.2.0

• Breaking: leonard_drive speaks the ext.leonard.* VM-service namespace (protocol version 2). Target apps must run leonard_* 0.2.0 — a 0.1.x app still registers ext.exploration.* and will not answer the handshake.

0.1.5

• Bundled AGENTS.md template documents the perception identifier / value fields and the addressing-vs-inference guidance, matching leonard_agent 0.1.5 (kDefaultAgentsMd); depends on leonard_agent ^0.1.5.

0.1.4

• leonard_drive up gains a native dual path: boot a Flutter target AND the leonard_native host against ONE shared device, expose both VM-service endpoints (DualLaunchHandle{flutterWsUri, nativeEndpoint, deviceId}), and tear both down via a single --pid-file. The single-target path is unchanged.
• New drive-dual subcommand (tools / observe / invoke) drives a multi-host MultiHostSession over both endpoints — merged manifest, merged observation, namespace-routed tool calls — so an external brain can perceive and drive a Flutter app and the native channel together.

0.1.3

• leonard_drive tools now prints a capabilities array next to namespaces. It surfaces reachable host features that are not namespaced tools — notably screenshot (use the screenshot subcommand), which is absent from namespaces by design. Stops the recurring "the manifest has no screenshot, so there is no screenshot capability" mistake. Requires leonard_agent ^0.1.3 (the handshake parse that decodes capabilities).
• leonard_drive gains up / down lifecycle subcommands that erase the manual boot-grep-convert dance. up --runner flutter -d <device> -t <entry> (or --runner dart -t <entry> for a pure-Dart target) boots the app, discovers the VM-service URI, prints it machine-readably ({event:"vm_service_ready", ws_uri, …} plus optional --uri-file / --pid-file), then HOLDS the process alive (teeing its log to stderr) until a signal or down. No model, no goal, no loop — the external brain attaches stateless observe/invoke/screenshot calls to ws_uri. down --pid-file <p> stops a held target. New shared launcher.dart primitive (spawn + scrape + hold + teardown) backs it.
• leonard_cli gains --launch for the autonomous loop: instead of a caller-supplied --vm-uri, boot the target (--runner flutter -d <device> -t <entry>, or --runner dart -t <entry>) via the shared launcher, drive the discovered URI with lenny's own LLM toward --goal, then tear the target down. --launch and --vm-uri are mutually exclusive (and the boot-only flags error without --launch) — no dual-mode interface.

0.1.2

• leonard_drive gains a screenshot subcommand: capture a still and write the PNG to --out path.png (decodes core.screenshot; prints width/height/device-pixel-ratio). Just pixels — no settle, no golden compare.

0.1.1

• Provider construction moves to the DartanticModelProvider seam (the agent's unified backend factory).
• Ship consumer coding-agent assets plus an install command, with harness-specific overlays (--claude / --copilot / --all). The bundled assets are target-agnostic (Dart-VM, not Flutter-specific).

0.1.0

Initial release.