actionbook__actionbook/packages/cli/CHANGELOG.md

@actionbookdev/cli

1.4.2

Patch Changes

• #518 2d4fc6d Thanks @Senke0x! - Harden extension bridge startup and surface actionable error hints.

  • Retry bridge port bind with exponential backoff (up to ~8.6s) to recover from TIME_WAIT after daemon restart
  • Move bridge bind to a background task so daemon cold start is no longer blocked by extension port contention (local/cloud modes were incorrectly delayed)
  • Add BridgeListenerStatus (Binding / Listening / Failed) so browser start --mode extension can distinguish still-binding from bind-failed
  • Wait up to 5s (polling every 100ms) for the extension to reconnect on browser start, eliminating a race where the bridge bound slightly before the extension completed its WS handshake
  • Surface chrome://, devtools:// and other restricted schemes as RESTRICTED_ACTIVE_TAB with hint pass --open-url <url> instead of an opaque debugger.attach failure
  • Close CdpSession WebSocket gracefully on failed session start (writer sends a Close frame, session awaits the writer task) so the bridge sees EOF and releases its 1:1 gate — previously the next start attempt would instantly fail with SessionClosed
  • Print hint: line for Fatal/Retryable/UserAction error results in text output (previously only --json surfaced the hint field)

• #518 2d4fc6d Thanks @Senke0x! - Add Hermes agent integration.

  • New SetupTarget::Hermes variant invokes hermes skills install directly instead of routing through npx skills add, targeting Hermes's native skill registry at ~/.hermes/skills/
  • Missing-binary error now points users to install Hermes (not Node.js) when the target is Hermes
  • Post-install verification parses hermes skills list table columns exactly (by │ delimiter) to avoid false positives from similarly-named skills
  • skills/actionbook/SKILL.md gains Hermes-compatible frontmatter (metadata.hermes.tags, requires_toolsets: [terminal], prerequisites.commands: [actionbook]) so the skill is discoverable via hermes skills search and hidden on non-terminal platforms

1.4.1

Patch Changes

• #511 07a53a1 Thanks @mcfn! - Fix Windows Chrome process cleanup and improve orphan recovery.

  • Replace NtQueryInformationProcess/ToolHelp with Win32 Job Objects for atomic termination of Chrome helper processes (renderer, GPU, utility)
  • Delete Chrome singleton lock files before orphan kill so new sessions can start even if helper processes linger
  • Add actionable error message when orphan Chrome still holds the user-data-dir lock
  • Fix 54 Windows e2e test cross-platform compatibility issues

1.4.0

Minor Changes

• #507 5bddf13 Thanks @Senke0x! - Add cloud browser provider support via -p / --provider.

  Supported providers: Driver, Hyperbrowser, Browseruse. Each provider reads its own <PROVIDER>_API_KEY from the caller's shell, and browser restart mints a fresh remote session while preserving the local session_id.

1.3.1

Patch Changes

• Fix browser list-tabs in extension mode by using Extension.listTabs for live tab enumeration instead of Target.getTargets, which is not available in extension-level CDP access.

1.3.0

Minor Changes

• #499 3092c61 Thanks @Senke0x! - Improve actionbook setup with a new skills installation step and targeted quick mode.

  • add a fifth setup step to install Actionbook skills during setup
  • add actionbook setup --target <agent> quick mode for one-shot skills installation
  • improve extension-mode setup guidance with Chrome Web Store and GitHub Releases fallback instructions
  • make API key input visible by default during interactive setup
  • tighten setup failure handling so quick mode and JSON flows report skills install failures correctly

1.2.0

Minor Changes

• #486 86905eb Thanks @Senke0x! - Restore extension browser mode with WebSocket bridge relay

  • Add extension bridge server (ws://127.0.0.1:19222) for transparent CDP relay between Chrome extension and CLI daemon
  • Use Extension API (listTabs, attachTab, createTab, detachTab) for tab lifecycle management
  • Read default browser mode from config.toml instead of hardcoding Local
  • Fix build.rs to track git ref files in worktrees for accurate BUILD_VERSION
  • Add Local mode guard to prevent silent fallback from unsupported modes
  • Reject concurrent CDP clients in bridge to prevent response channel hijacking

1.1.0

Minor Changes

• #483 4d46f8d Thanks @4bmis! - support create multi tabs in one shot

1.0.2

Patch Changes

• Remove search and get commands from CLI help output and skill documentation

1.0.1

See 1.0.0 release notes below.

1.0.0

Patch Changes

• #464 cc58aee Thanks @mcfn! - Actionbook v1.0.0 — Browser Engine for AI Agents

  Rebuilt browser automation runtime designed for AI agents. Stateless session model, agent-friendly command interface, expanded command surface, and improved stability.

  Breaking Changes

  • Stateless session model. CLI now requires explicit --session and --tab flags for browser commands. Stateless interface, stateful runtime — agents reason about browser state through explicit addressing, not hidden side effects.

  Design: Agent-First CLI

  This is not a browser automation tool adapted for agents — it is built for agents from the ground up.

  • Structured, parseable output. Supports JSON via --json, with stable text output by default. Both formats are part of the formal contract.
  • Predictable command surface. Consistent argument patterns and return shapes across all 50+ commands. Agents don't need per-command special-casing.
  • Stateless by default. Explicit --session and --tab addressing means agents manage state in their own context window, not in hidden browser-side state. This maps directly to how LLM agents work — every call is self-contained.
  • Snapshot refs as stable handles. Elements are labeled with refs (@e3, @e7) that persist across commands within a snapshot. Agents can plan multi-step interactions without re-observing the page after every action.

  New Features

  • 50+ browser commands. Expanded from the previous command set to cover the full browser surface: sessions, tabs, navigation, observation, interaction, waits, cookies, storage, queries, screenshots, PDF, uploads, and console logs.
  • Multi-tab execution. First-class --tab addressing. Run work across dozens of tabs in parallel.
  • Snapshot refs. snapshot labels every element with a stable ref. Chain multiple commands without re-snapshotting.

  Improvements

  • Expanded and hardened existing browser commands — more consistent argument handling, better error messages, and predictable return values.
  • 10× faster automation on complex sites via action manuals and snapshot ref chaining.
  • Full rebuild of CLI internals around stateless session/tab model.

  Bug Fixes

  • Fixed session cleanup failures causing orphaned browser processes.
  • Fixed tab addressing race conditions under parallel execution.
  • Resolved inconsistent snapshot refs across navigation events.
  • Improved error handling for browser commands on dynamic pages (SPAs, virtual DOMs, streaming components).

  Examples

  Repo ships with end-to-end examples: 192-site tagline collection (3 min), deep research report generation, and more. See examples/.

  Install

  npm install -g @actionbookdev/cli
  npx skills add actionbook/actionbook