Skip to content
Agents tracked: 501 Releases (7d): 777 Agent status: no reported issuesSince last limit reset: Claude 10d · Codex 8d For AI agents: Markdown Updated Nov 21, 2026

KaneAI changelog: what's new each month

Every stable KaneAI release summarised by month: the highlights, new features, improvements, fixes and anything you need to act on. 8 months covered; the current month updates daily.

November 2026 · so far

6 releases: 0.10.2 → 0.10.7

November brought major improvements to reliability, diagnostics, and remote execution for KaneAI. The release focused on fixing silent failures, improving error messages, and ensuring consistent behavior across local and remote runs.

Action needed
  • Telemetry is now always on with no opt-out option.
  • Test failures are no longer reported as CLI errors; they appear on separate reporting lanes so scripts can distinguish test failures from CLI failures.
  • Scroll steps that move nothing are now reported as failures instead of silently succeeding.

Highlights

  • Scroll steps now properly fail when they don't actually move the page, eliminating silent successes.
  • Commands and sessions now emit lifecycle events and detailed error codes, making run behavior visible and debuggable.
  • Real devices can reach private backends through LT tunnels, expanding remote testing capabilities.
  • Telemetry now tracks healing attempts accurately, and replayed runs report matching stability numbers.
  • New bug tracking system lets you mark known issues against tests and manage organization-specific RCA categories.

New

  • Bug tracking with kane-cli bugs commands to record known bugs against tests and retrieve them during RCA.
  • Organization-specific root-cause analysis categories that show up in failure reports.
  • Tunnel support for real device access to private backends via --tunnel and --tunnel-name flags.
  • Type step delays for fields that ignore normal fill, with delays preserved in replay and code export.
  • Command lifecycle events for run, testmd run, and explore so early failures are visible.
  • Named explorations with kane-cli explore run --name that persist through export.
  • Custom headers now included in kane-cli test runs.
  • Pool variables per remote job member with separate secret management from the vault.

Improved

  • Locator actions now report why they failed instead of bare failure messages.
  • Vision-based select reports actual failure when dropdown selection doesn't take effect.
  • Reworded dropdown labels no longer discard existing selections.
  • Python steps inherit certificate trust from kane-cli, eliminating TLS proxy failures.
  • Android vision taps now map to full-display pixels for accurate targeting.
  • Session tracking shows surface, target, and outcome including exit type.
  • Scroll detection now reaches scrollable regions in cross-origin iframes, dialogs, and locked html elements.
  • Text-mode page output lists scrollable containers so you can see what the agent can work with.

Fixed

  • Broken trust files no longer poison runs; invalid certificates are skipped instead of passed to Python programs.
  • Vision-based select no longer claims false success when dropdown selection fails.
  • Mobile API call timeouts now preserve millisecond values.
  • Session save failures now report the actual reason instead of generic "No test_id available" message.
  • Until-scrolls now fail when the condition is never met instead of scrolling forever.
  • Runs with no result no longer slip through as successes.

October 2026

6 releases: 0.8.20 → 0.10.0

October brought major improvements to mobile testing consistency and introduced real-device capabilities alongside a significant upgrade to the explore feature. The month focused on platform parity, device management, and cloud-scale exploration.

Action needed
  • target replaces target_kind — update device references from target_kind to use target with values like emulator or simulator.
  • explore run is now remote by default and runs on HyperExecute; use local-only flags or check --help to run locally.
  • Mutating is now the default policy for explore run; use --policy read_only to restrict writes.

Highlights

  • Real devices now work from kane-cli and the TUI with live session viewing.
  • Mobile and web platforms now share the same planner, perception, and replay logic for consistent behavior.
  • Explore runs moved to cloud by default with up to 20 concurrent browsers and a live map interface.
  • Mobile exports now generate the correct platform code automatically based on your session target.
  • Smart variables and template references now resolve correctly, including on Windows.

New

  • Real-device runs available directly from kane-cli run and the TUI with live device window.
  • explore view displays a live interactive map of exploration with zoom, search, and findings filters.
  • Live stream view shows browser captures at up to 15 fps during active runs.
  • clear_input tool for web agents to empty input fields in one action.
  • explore run supports scoping with --focus <screen> and depth-based concurrency caps.
  • Runnable POST endpoints can be named with allow-post for read-only policy mode.
  • Smart variables now work on Windows with shipped timezone data.
  • Real-device runs no longer require local device setup.

Improved

  • Mobile sub-steps now use the parent test's planner instead of falling back to an older one.
  • Mobile and web replay now handle hub settings, loops, and empty steps identically.
  • Appium logs appear in the run log alongside web binding logs.
  • Mobile recordings are cleaner with each step keeping its own label.
  • Typed values are redacted before extraction to prevent passwords and tokens in run data.
  • App launches no longer block runs and execute off the main loop.
  • Until loops inside branches now export correctly within that branch.
  • Import bindings now resolve in the importing file's scope properly.

Fixed

  • Screenshot crash when an element is positioned far outside the visible screen.
  • Mobile assertion results no longer get downgraded in saved runs.
  • Device binding rejections now correctly reported as lost device instead of unexplained crash.
  • Grid-hook steps now replay instead of being skipped.
  • Output variables now flow forward between mobile steps during replay.
  • Unresolved references no longer appear in imported files.

September 2026

11 releases: 0.8.8 → 0.8.18

September brought major improvements to mobile testing, variable handling, and test authoring workflows, along with structured event logging for better automation and integration. The month focused on stability, correctness, and making the CLI more scriptable.

Action needed
  • Unresolved variables now emit a single warning per run and continue instead of refusing to start; runs will proceed without all variables defined.
  • JavaScript is no longer exportable for mobile targets; mobile export is Python-only.

Highlights

  • Mobile code export now works end-to-end, generating Python automation from mobile test runs with correct drag gestures and text input.
  • Variables are now validated before commit, with unresolved references caught upfront and stub files created automatically.
  • Test authoring keeps the same test identity across multiple runs instead of creating duplicates with each --author invocation.
  • Non-TTY output is now a versioned NDJSON stream with event logs, enabling reliable automation and CI integration.
  • Context sync setup guides you through sharing over GitHub, S3, or folders in one flow with readable error messages.

New

  • Code export for mobile runs generates exportable Python automation code, with JavaScript export blocked and explicitly marked as mobile-only.
  • Upload app builds directly from your laptop and wire them into runs automatically without a separate dashboard step.
  • Remote execution now supports Android emulators and iOS simulators, with device catalogs trimmed to what actually exists on HyperExecute.
  • Feed tests from Test Manager dataset rows using --dataset-id and --dataset-row, with rows addressed by ID rather than position.
  • Guided context sync setup for sharing over GitHub, S3-compatible buckets, or folders, with scriptable end-to-end flows.
  • Live run pointers let separate processes attach to and monitor runs already in progress from disk.
  • Structured NDJSON event streams with version stamps, progress events, and per-member logs for testrun execution.
  • Design sessions declare variables a test needs with validation before commit and automatic stub file creation.

Improved

  • Mode agent now pauses on every question instead of auto-accepting, requiring explicit input for every action.
  • Variable placeholders no longer append trailing garbage after substitution.
  • If/else steps without nested checks now anchor correctly and behave as authored.
  • Logs are fully flushed before fatal exit to avoid losing remote log data.
  • Drag gestures now export with correct end points and timing values.
  • Until-loop scroll behavior respects the scroll unit by default and scroll unit is only set when applicable.
  • Stored values resolve at run time with actual produced values instead of falling back to author-time values.
  • Scroll-into-view steps now described by target instead of position, surviving page layout changes on replay.

Fixed

  • Package.json engines field removed to prevent npm from silently installing older releases.
  • JavaScript export on mobile now fails immediately with clear messaging instead of producing silent broken output.
  • Typed text no longer has trailing garbage in mobile exports.
  • Mid-run lock conflicts no longer abandon heals; they continue shrinking and re-authoring locally.
  • Locks are released on every early exit type (bad name, tenant mismatch, launch failure, missing runner).
  • Evolve commits now report correctly with testmdsummary events and TTY box in agreement.

August 2026

11 releases: 0.6.11 → 0.8.7

August brought major speed improvements to Autopilot, expanded content ingestion from multiple platforms, and full end-to-end mobile testing support with intelligent gesture handling and device management.

Action needed
  • The --retry flag is retired; adaptive healing is now always on.
  • kane-cli doctor --target is now mandatory and singular, requiring exactly one target.
  • kane-cli no longer requires a separate Node install; Node 20 is now the minimum supported version.
  • queued mode is incompatible with --source, --resume, and --plan flags; kane-cli exits with code 2 if combined.

Highlights

  • Mobile testing now works end-to-end on emulators and simulators, with the AI understanding mobile screens and replaying recorded flows on mobile targets.
  • Autopilot is faster and more efficient, planning entire forms in one pass and using a single ongoing conversation instead of re-uploading page details at each step.
  • Ingest from Jira, Confluence, and Linear directly by URL, with automatic fetching of issues, comments, attachments, and images.
  • Adaptive healing is now on by default for every run, automatically recovering from layout changes without manual retry flags.
  • Remote test execution now available through HyperExecute integration, with automatic setup and streaming progress back to your terminal.

New

  • Mobile testing with device emulators and simulators, including persistent runner sessions and evidence capture.
  • Ingest content directly from Jira, Confluence, and Linear URLs with automatic attachment and image fetching.
  • Queued extraction mode for pre-drawn work items claimed from a durable queue in budget-aware batches.
  • Remote test execution through HyperExecute with kane-cli plugin system for managing remote runners.
  • Native support for browser date, time, color, and range input controls that the AI can set directly.
  • Scroll-until operations that collapse scroll-and-search sequences into single steps.
  • Variable-named targets (${variable}) that are re-probed at replay time for automatic healing.
  • Device addressing by human-readable name and OS version in mobile test runs.

Improved

  • Clicks land more accurately by targeting the element's real on-screen position and checking parent context.
  • Whole forms now fill in one planning pass when fields sit on the same page.
  • Checks are decided on the live page at verification time instead of planned blindly.
  • Exported tests self-heal when UI drifts with stronger auto-healing that relocates elements.
  • Playwright expectations are derived automatically from plain language descriptions including thresholds and unit conversions.
  • Mobile gesture recording is replay-safe with scroll-into-view actions preserved through round-trips.
  • Device reads are faster by overlapping and reusing reads where possible.
  • Evidence steps show human-readable captions instead of raw locators or bare step numbers.

Fixed

  • Click-then-switch-tab sequences no longer race by waiting for tab-open popups to resolve.
  • Frame context uploads correctly under the v4 code-export contract as a bare list.
  • Frame info is correctly scoped to waituntil click and scroll steps.
  • JSON path in sub-check results is now preserved through replay and export.
  • Log and evidence files write correctly on all systems with pinned UTF-8 encoding.
  • App-start detection is more forgiving on Android, accepting any activity as a valid signal.

July 2026

13 releases: 0.4.9 → 0.6.10

July brought major improvements to KaneAI's testing capabilities, with new features for bug detection, evidence collection, requirements coverage tracking, and document handling. The month also refined the user interface and fixed numerous issues across network streaming, visual testing, and authentication.

Action needed
  • Composer and slash palette removed—replaced by a new options-first question panel surface.
  • SET variable initialization—variables declared with SET now get their value from the store instruction at runtime instead of being pre-seeded.

Highlights

  • Automatic bug detection and investigation—runs now determine whether failures are product bugs (result code 740) or test issues, and can be configured to stop or continue on confirmed bugs.
  • Evidence packs bundle all run artifacts—screenshots with annotations, HAR network logs, console output, and results are automatically collected into sealed, self-contained packs after each run.
  • Requirements and coverage tracking—tests now record acceptance criteria they cover and results report per criterion, with coverage gaps surfaced in a side-by-side view.
  • PDF and Word document support—PDFs and .docx files can now drive test runs with embedded images tracked and fingerprint-verified.
  • Redesigned question panel—keyboard-first navigation and a cleaner visual hierarchy replace the composer and slash palette, with recommendations and multi-select actions in one place.

New

  • Network SSE streaming mode with toggleable flag to stream server-sent events through the run pipeline instead of waiting for full responses.
  • Visual step annotations—each visual action produces an annotated screenshot with a crosshair and bounding box showing exactly what was targeted.
  • Per-step network and console capture—HAR files and NDJSON console output are captured and attributed to individual steps.
  • Bundled Node runtime—kane-cli now ships its own Node runtime per platform, eliminating the need for a separately installed Node version.
  • Scroll-into-view as a recorded step—scroll actions are now first-class recorded steps with generated code and correct replay behavior.
  • PDF-driven test runs—PDFs can be supplied as runconfig.source and are converted to structured projections with complete content inventory.
  • Evidence merge functionality—kane-cli evidence merge combines multiple sealed packs into a single artifact.
  • Markdown attachments in generate—.md files can now be passed directly when generating tests.

Improved

  • Navigation performance—back and forward no longer hang for 30 seconds; they complete as soon as the browser commits.
  • Assertion evaluation locality—assertion evaluations now run locally instead of being managed at the server.
  • Empty input handling—system-only API nodes no longer error when there is no user message.
  • Multi-line branch event matching—if/else branch events spanning multiple lines are now captured and routed correctly.
  • SSE response handling—text/event-stream responses are skipped during network capture to avoid stalling the pipeline.
  • Element detection reliability—presence checks that are uncertain now escalate to visual scan instead of returning false confidently.
  • Hover on vision coordinates—hover actions now correctly map to coordinates identified by vision.
  • OAuth session stability—OAuth credentials are automatically exchanged for long-lived credentials, and mid-session token expiry is prevented.

Fixed

  • Special character handling in typed text—literal tokens in type and fill actions are now properly re-escaped.
  • Validation error visibility—422 errors now log the full request body so users can see exactly what was rejected.
  • Status code pass-through—422 errors are only raised for request-body validation; other upstream errors now forward their original status codes.
  • Missing platform binary install—post-install step now degrades gracefully instead of aborting npm install.
  • Condition block expression evaluation—logical and boolean expressions in condition blocks are now handled correctly.
  • Visual checkpoint conditions—textual/visual conditions now correctly inherit from the checkpoint condition type.

June 2026

11 releases: 0.3.6 → 0.4.8

June brought major expansions to KaneAI's testing and generation capabilities, adding DevTools inspection, an interactive test generation mode, and first-class support for APIs and browser storage. The CLI also gained a cleaner foundation with zero-config startup and improved terminal UI.

Action needed
  • Bare-objective shortcut removed: kane-cli "<objective>" no longer runs tests; use kane-cli run "<objective>" instead.
  • config show and list commands now emit JSON when stdout is piped or redirected instead of human tables.
  • Mid-run interactive project/folder picker removed; startup now auto-defaults headlessly and announces the choice.
  • Exit code for TMS credential-exchange failures changed from 1 to 2 to align with other auth failures.

Highlights

  • Inspect and assert on live network traffic, console output, cookies, localStorage, and performance metrics directly in tests.
  • New interactive generate command lets you refine test objectives through a chat interface before saving .testmd files.
  • External APIs are now first-class test steps that can be called, stored, and referenced across flows.
  • DevTools actions export correctly to automation code, including network replay and storage inspection.
  • URL handling is now configurable at multiple levels: defaults, per-run flags, and test file frontmatter.

New

  • Network request and response inspection with assertion and extraction support.
  • Console output capture (logs, warnings, errors) as assertion and extraction targets.
  • Performance traces collected and exposed as assertion targets during runs.
  • Clipboard, cookies, and localStorage are now readable and writable test tools.
  • Interactive kane generate command with chat-like refinement and /view full-screen scenarios.
  • executeapi steps to call external APIs and chain responses through subsequent steps.
  • WebSocket frame capture toggleable from the TUI, appearing alongside HTTP network activity.
  • Live hints footer in TUI with remote-cached tips relevant to current mode.

Improved

  • Navigation now happens in a defined first phase with predictable timing and accurate logging.
  • Navigation and session setup run in parallel, reducing startup dead time.
  • Custom ARIA comboboxes now receive click interactions instead of select, improving compatibility.
  • Option extraction restored for combobox, listbox, and select elements across the board.
  • Viewport size querying fixed to return accurate layout-sensitive dimensions.
  • Step labels stream in with typewriter effect in real time instead of waiting for completion.
  • Run objective display combines start URL and task description into a single stitched view.
  • Assertion path now checks page content via code extraction instead of just screenshots.

Fixed

  • Binary or malformed request bodies no longer crash sessions with UnicodeDecodeError.
  • Chrome no longer hangs on startup; CDP launch now has bounded retry loop and timeout.
  • Child flows now receive their full tool set (tab switching, file upload, etc.) when spawned.
  • Orphan run directories no longer left behind when runs exit early due to pre-flight errors.
  • Cancelling a run now properly releases its TMS playground lock.
  • Conditionals inside multi-step flows no longer produce malformed step sequences.

May 2026

7 releases: 0.2.11 → 0.3.5

May brought major stability improvements and new terminal capabilities to KaneAI. The month focused on making project selection instant, fixing credential and profile handling, and adding powerful CLI tools for managing test files directly from the command line.

Action needed
  • testmd delete requires explicit --yes confirmation flag to prevent accidental removal.

Highlights

  • Manage test files from the terminal with the new testmd subcommand, supporting run, list, status, export, and delete operations.
  • Project picker now instant even with hundreds of projects, with live search and auto-pick defaults.
  • Retry mode replays recorded actions with live re-evaluation of branches, so tests adapt to condition changes.
  • Credentials and profile switching now work reliably without mix-ups or carry-over from previous sessions.
  • Vision-based drift detection and confidence scoring make element matching more robust across UI changes.

New

  • testmd subcommand for terminal-based test file management with interactive or NDJSON output modes.
  • Retry and replay system with --retry and --retry-count flags that re-evaluate branches during playback.
  • Confidence scoring for element matching with distinction between primary descriptors and supporting visual hints.
  • Triage payload and structured reporting for test results.
  • Opt-out flag for auto-generated verification checks in action mode authoring.
  • Automatic retry with backoff for transient network failures on idempotent reads.
  • Session directory zipping for artifact uploads, capturing full context for post-run inspection.
  • per-step variable and secret delivery, allowing mid-run overrides to take effect immediately.

Improved

  • Project search now filters on the server instead of locally, ensuring accuracy across hundreds of projects.
  • Project picker shows results paginated at 10 per page with a visible loading state.
  • if/else branches re-evaluate during replay instead of blindly replaying recorded steps.
  • Click actions use vision-based drift detection as fallback when coordinates shift.
  • Variable names are pushed to TMS unmodified without internal prefixes.
  • Terminal login uses the same guided flow as the rest of the app.
  • Floating badge during test runs now shows what KaneAI is currently thinking.
  • Assertion outcomes are recorded directly instead of re-executed during codegen.

Fixed

  • Profile switching no longer carries previous profile's project across to new account.
  • Headless login no longer hangs and completes cleanly.
  • Credentials validation happens immediately with in-place editing instead of discovering bad credentials later.
  • Retry no longer gets stuck on terminal-step failures.
  • run.log now written to correct location runs/<n>/run.log instead of nested path.
  • OAuth users can now use retry without falling back to username/access-key.

April 2026

11 releases: 0.2.0 → 0.2.10

April saw rapid iteration on KaneAI with nine releases, culminating in version 0.2.9 which brought substantial improvements to action recording, code generation, and user-facing output formatting.

Highlights

  • Click and type actions now record as clean selector-only operations without coordinate fallbacks, producing more portable generated code.
  • Action rationale is now carried through to uploaded operations and test reports, showing meaningful reasoning instead of generic labels.
  • Tool-call steps display real per-call descriptions like "Open a new tab" instead of generic labels.
  • Run-result output formatting improved with better wrapping on narrow terminals and cleaner branch rendering.

New

  • CSS and XPath selectors are tagged with their type for downstream code generation to use the right Playwright API.
  • Tool-call steps now show per-call descriptions instead of generic tool names.

Improved

  • Click and type actions upload as selector-only operations without coordinate fallback when usable DOM selectors are available.
  • Action rationale from the agent is carried into uploaded operations and instructions.
  • Wait-until loop reporting now accurately reflects non-failing loops that exhaust retries.
  • Run-result headers stack vertically and wrap on narrow terminals instead of overflowing.
  • If/else and child runs render under a branch header inline rather than appearing twice in scrollback.
  • Failed Chrome launches show the full per-platform install message standalone instead of being prefixed and truncated.

Summaries are written automatically from the official release notes (full changelog ↗); check the original notes before relying on a detail. KaneAI: pricing, features and alternatives · All changelogs