v1.6.0

Highlights

• New skill: automate:drive-automation-session — describe what you want to do on a reserved Kobiton device in plain language ("open YouTube and play the first video"), and the agent drives the device for you through an observe-and-act loop. It runs as an automation session, so the result is recorded with video and is ready to turn into a reusable test case via saveTestCase. Sessions stay alive through long human-in-the-loop pauses (30-min idle timeout), and the agent can watch the screen and stop on its own when it gets stuck. This complements the existing run-interactive-cli-session skill — it doesn't replace it.

• Renamed run-interactive-test → run-interactive-cli-session — clearer naming now that there are two ways to drive a device. The skill works exactly the same as before; only the name changed.

• More reliable Chrome/web testing — capability handling now waits longer for Chrome to start up on cold devices, reducing flaky session starts.

Full changelog: 1.5.0 → 1.6.0

v1.5.0

Highlights

• New getAppParsingStatus tool — after confirmAppUpload, app parsing runs asynchronously: the record is created in state PARSING, and appId may be null for a brand-new upload. Agents now poll getAppParsingStatus(versionId) until the state is terminal (OK or a FAILURE_* value) before reserving devices or starting sessions — the response also resolves the real appId. The confirmAppUpload schema documents the async flow, and the run-automation-suite skill's upload step is now a four-step flow ending in the parse-status poll
• New getUserInputEvents tool — surfaces the touch/swipe gestures a human makes on the device-only live view, so an agent-driven session can be redirected mid-run: the user's tap reaches the device in real time AND is reported to the agent as an observation to react to ("the user just tapped Settings → pivot the test plan to Settings"). The run-automation-suite skill polls it between scripted commands. Keystroke / right-click / pinch / drag-off-canvas remain suppressed
• Docs refresh — README now lists all 29 tools across 5 domains (the 14 test-management tools were previously missing from the inventory), and AGENTS.md got a cross-host cleanup: per-domain tool counts dropped (the runtime tools/list is the source of truth), the confirmAppUpload known-limitation rewritten around getAppParsingStatus polling, Claude-Code-specific wording generalized for Gemini / Copilot / Cursor readers, and a concrete Copilot CLI install path

Full changelog: 1.4.3 → 1.5.0

v1.4.2

Highlights

• Cursor support (CLI and IDE) — install on Cursor in addition to Claude Code, GitHub Copilot CLI, Gemini CLI, and Codex CLI. On the CLI, install in-session via /plugin marketplace add github.com/kobiton/automate; in the IDE, via Cursor Settings → Plugins (paste the repo URL, Add to Cursor, Install), then connect the kobiton MCP server from Tool & MCPs. Plugin installs are shared between the CLI and the IDE — install once and the other picks it up. New .cursor-plugin/plugin.json + .cursor-plugin/marketplace.json following the cursor/plugins convention, or drop just .cursor/mcp.json into any project for an MCP-only setup. MCP requests from Cursor carry X-AI-Tool-Name: Cursor for adoption analytics
• Chromeless device window for run-automation-suite — when the session resolves to the device-only view and Chrome is the preferred browser, the skill launches a chromeless Chrome --app window (no tab strip, no URL bar) sized to the device class: phone 540×920, tablet 780×920, fold 880×920 (landscape swaps). Per-OS shims: macOS (osascript resize with fail-open on Automation denial), Windows (PowerShell SetWindowPos), Linux (--window-size hint). Falls back gracefully to the regular browser-open path when Chrome is absent or another browser is preferred
• Broader Chrome detection on macOS — the chromeless launcher now detects Chrome, Chromium, Chrome Canary, and Brave at their standard /Applications/ paths, and validates --width/--height/--x/--y as positive integers with clear errors
• Fixed Copilot CLI command loading — Copilot CLI validates the command name frontmatter and rejects colons, so name: "automate:setup" broke command loading there; commands now use plain name: "setup" / name: "doctor". Claude Code and Copilot CLI keep /automate:setup and /automate:doctor (filename + plugin namespace); on Cursor CLI they appear as /setup and /doctor, distinguishable from Cursor's built-ins by their Kobiton descriptions
• Docs — new Cursor IDE install section and "Other MCP Clients" README subsection (ChatGPT Apps SDK, Continue / Cline / generic Streamable-HTTP clients), refined Cursor CLI install steps (marketplace flow, required agent restart so skills load), and a Cursor CLI troubleshooting section (stale/missing skills and commands, MCP disconnects, missing ~/.kobiton/bin/kobiton wrapper)
• Hardening & tests — URL validation against shell-metacharacter injection at every launcher entry point; 51 new vitest cases across launcher arg parsing, exit-code sentinels, and .codex/ mirror byte-identity

Full changelog: 1.3.0 → 1.4.2

v1.2.2

Highlights

• Updated bundled kobiton CLI binary in run-interactive-test skill to the latest version
• Expanded run-interactive-test adb-shell documentation for AI agents: quoting rules (local vs device shell parsing), platform guard (Android only), 22-row intent-to-command cookbook, big-output redirect pattern (to avoid 25k-token MCP overflow), long-running command guidance, and response parsing gotchas in references/response-shapes.md, notably that adb returns exit code 0 even when the inner command fails
• Added 14 Test Case Management MCP tool schemas covering the full TCM workflow:
  • Test cases: saveTestCase, listTestCases, getTestCase, updateTestCase, deleteTestCase
  • Test runs: createTestRun, listTestRuns, getTestRun, terminateTestRun
  • Test suites: listTestSuites, getTestSuite, createTestSuite, updateTestSuite, deleteTestSuite

v1.2.0

Highlights

• Multi-CLI support: install on GitHub Copilot CLI, Gemini CLI, and Codex CLI in addition to Claude Code
• New run-interactive-test skill: natural-language WebDriver/device/file commands powered by the bundled kobiton CLI wrapper (macOS Apple Silicon binary included)
• New /automate:setup command: bootstraps ~/.kobiton/.credentials from the authenticated MCP session, no manual file editing
• New /automate:doctor command: read-only health checks for CLI install, credentials file, active profile, and required fields
• New getCredential MCP tool: backs /automate:setup; returns the OAuth user's username, API key (existing or freshly generated), and portal URL
• Session attribution: Appium sessions started via run-automation-suite now emit kobiton:aiToolName; MCP requests from Claude Code, Codex CLI, and Gemini CLI carry optional header X-AI-Tool-Name (set to the originating tool) for adoption analytics
• Governance: CodeQL weekly scans + per-PR analysis, security issue routing template

v1.0.0

What's Changed

• refactor: simplify the appium test skill by @d4rkwinz in #1
• doc: update logo, license, and README by @d4rkwinz in #2

Full Changelog: https://github.com/kobiton/automate/commits/v1.0.0