Skip to content

v0.15.0

Choose a tag to compare

@github-actions github-actions released this 06 Oct 13:37
· 531 commits to main since this release
4746541

Changed (BREAKING for WiFi auto-discovery only) — FP-15

  • flutter_probe_agent no longer depends on the native bonsoir mDNS plugin. A plugin in an app's
    dependencies is linked into every build of that app, release included, regardless of
    --dart-define=PROBE_AGENT (release size, privacy, possible local-network prompt).
    mDNS advertising is now an optional ProbeAdvertiser hook
    (ProbeAgent.start(advertiser: ...)) that the app implements with its own bonsoir dependency; a
    copy-paste example is in the Studio docs. No new pub.dev package is published. Without it WiFi testing is
    unchanged (--host <ip> --token <token>); Studio's WiFi auto-discovery needs an advertiser in the device build.
    The advertised port is now the actually bound one.

Optional local-LLM support (FP-14)

AI stays strictly opt-in: no ai: block means no AI code runs, and none of the additions below can change a
test result, a report or the exit code (they run after results are final, are time-bounded, make no retries,
and stop after the first provider error).

  • probe ai doctor — checks the configured ai: provider: endpoint, model listed, a text round trip, and
    whether the model accepts images. Verified live against LM Studio (a 0.5B text-only Qwen: text OK, vision
    not accepted) and Ollama (Gemma 4 31B: both OK).
  • Failure triage — probe test --ai-triage (after the run) and probe triage --input results.json
    produce a short advisory probable cause and next step per failure, printed and written to
    <reports>/triage.md. MCP tool triage_failure wraps it (20 tools now; docs previously said 18).
  • ai.vision: false — text-only models: see "..." with ai is judged from the screen's visible
    texts/keys instead of a screenshot. assert no visual defects with ai and read ... with ai report that
    they need a vision model; the mode is refused when ai.redact rules exist (they can't apply to text).
  • probe generate and AI selector suggestions honour ai.provider, including local with no API key.
    With no ai.provider the original Claude-only behavior is unchanged.
  • Docs: new AI & Local Models page, plan in
    docs/prd/optional-local-llm-plan.md. A 0.5B model gives weak triage advice; 7B+ is recommended.
  • READMEs added for cmd/, internal/, scripts/, docker/, plugins/, mcpb/, tests/,
    tools/probe-convert/.

Fixed (docs)

  • scroll down until "X" appears was documented in the VS Code README and snippets but not implemented by the
    parser; it is now (FP-13, below).

Fixes for limitations hit during a release-gate run (FP-13), reported against CLI 0.14.0.

Added

  • scroll [dir] until <target> appears — scrollIntoView for lazily built lists (Maestro's
    scrollUntilVisible). Resolved agent-side: scrolls half a viewport at a time, then brings the
    target fully on screen (Scrollable.ensureVisible), stops early when the list can't move, and
    fails with the visible texts if the target never shows up. Older agents fall back to a CLI-side
    loop (25 attempts).
  • wait for idle (also wait until idle) — route transitions finished and no pending frames,
    animations or HTTP requests. tap also waits (max 2 s) for an in-flight dialog/sheet/page
    transition before firing.
  • probe test --grant <list> — pre-grant OS permissions once before the first test
    (--grant notifications,camera,location). Android: adb shell pm grant incl. POST_NOTIFICATIONS;
    iOS simulator: every simctl privacy service. iOS notifications cannot be pre-granted (no simctl
    service; the prompt is a SpringBoard alert the agent cannot tap) — --grant warns instead of failing,
    and allow permission "notifications" on iOS now fails with that explanation instead of a
    contradictory "unknown permission — available: notifications".
  • --agent-port <n> on probe test / probe record (alias of --port) and the matching
    --dart-define=PROBE_PORT=<n> for the app, so simulators can't fight over 48686.
  • Port-holder report: a failed agent dial on loopback now names the process holding the port
    (agent port 48686 is held by pid 4242 (Runner)) and how to move off it.
  • probe.visible_summary agent RPC (visible texts + string ValueKeys) used for failure diagnostics.

Changed

  • Timeout errors say what timed out. A step that hit its deadline used to fail with a bare
    context deadline exceeded. It now reports the line, the step, the timeout and the visible texts/keys
    at that moment. Agent-side Widget not found / Timed out waiting for errors carry the same
    suffix, and wait hands the agent a timeout 2 s shorter than the CLI's so the agent's descriptive
    error arrives first.
  • type and clear go through the platform text-input path
    (EditableTextState.userUpdateTextEditingValue) instead of assigning controller.text, so
    onChanged, inputFormatters and form validation run as with real keystrokes.

Fixed

  • dont see "X" (no apostrophe) is accepted as an alias of don't see "X"; it previously failed
    with unknown recipe call "dont see <arg>".