…no OpenClaw dependency

Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>

…d" section

Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>

- clawdcursor mcp now checks hasConsent() and exits with clear stderr
  message if consent has not been given — no more silent cold-start
- Added writeConsentFile() export to onboarding.ts
- Added --accept flag to clawdcursor start (consent + start in one shot)
- Added clawdcursor consent subcommand with --accept, --revoke, --status
- runOnboarding() now takes a context param ('start' | 'consent') to
  show accurate warning language for each entry point
- start warning now says 'AI Agent + REST API on localhost:3847 — any
  local process can call tool endpoints'
- consent warning shows all three transport modes covered by consent

- Hero: clear tagline, install command, stats (33 tools / 6 layers / 3 transports / any model)
- 'Where is the AI coming from?' split — human with API key vs agent connecting
- Three connect modes with setup snippets: MCP, REST, CLI agent
- Consent section — all four commands (interactive / --accept / --status / --revoke)
- 6-layer pipeline visualization with cost indicators per layer
- 33 tools by category (Desktop, A11y, CDP, Orchestration)
- 8 provider cards with text/vision model routing
- v0.6.3 vs v0.7.0 comparison table
- Agent-readable HTML comment block at top for LLM crawlers
- Responsive, matches v0.6.3 design language (dark, Inter, green accent)

- Remove comparison table
- Keep same CSS variables, component patterns, AI cursor animation
- Same mode-card / feat-card / code-box / os-tab patterns from v0.6.3
- Three tabs: MCP Client / CLI Agent / REST (replaces OS tabs)
- Who section: human vs agent split as first decision
- Pipeline: 4-card grid (L0/L1/L2/L3 — simplified)
- What's New: 6 feat-cards, reliability-focused
- CTA: 'Give your AI a body'
- Cleaner, simpler, consistent with existing site

doctor.ts:
- Remove 'Registered as OpenClaw skill' console output in v0.7.0
  (clawd-cursor is standalone — external skill link is silent/optional)

website:
- Hero: 'Give your AI eyes and hands' — no transport jargon
- Who section: use cases first (tell it what to do / connect your AI)
  not 'where is the AI coming from?' technical framing
- Mode cards: Claude Code / Give it tasks / Build with it
- Install tabs: Claude Code / Cursor | Give it tasks | Build with it
- Strip MCP/REST/CLI labels from user-facing copy where not needed

…ve use

verifiers.ts (full rewrite):
- Default is FAIL/UNCERTAIN, not PASS — no more silent auto-pass on unrecognized tasks
- Error passthrough is FAIL, not PASS — broken checks are never invisible
- Primary verifier: text LLM reads a11y tree + active window + focused element,
  must cite specific screen evidence to return PASS
- LLM verdict requires confidence >= 0.65 AND explicit evidence citation
- UNCERTAIN verdict → treated as FAIL (conservative)
- Fast-path heuristics (app_open, clipboard, navigation) still run first for
  zero-cost trivial checks, but only trusted at confidence >= 0.75/0.80/0.85
- Every check attempt logged as VerifyAttempt: checkName, pass, confidence,
  detail, durationMs, optional error
- Full attemptLog returned on every VerifyResult for caller to log

a11y-reasoner.ts:
- Logs every individual verify attempt to console with checkName/pass/conf/ms
- Logs evidence string when present
- Verifier errors are BLOCKING (not silent) — thrown exception = FAIL
- Full attemptLog written into logStep verification.detail for JSONL record
- uiStateSummary now includes per-check pass/fail/confidence summary

agent.ts:
- Pass pipelineConfig to TaskVerifier so LLM verifier has text model access

Block characters (█ ╗ ║ etc.) were getting garbled on Windows during
npm install due to code-page encoding issues. Replaced all Unicode
block/box-drawing chars with explicit \\uXXXX escape sequences.
These survive any encoding transform and render correctly in any
Windows terminal with UTF-8 support.

Banner now shows exactly once — during the first-run consent flow
in runOnboarding(). Every new user sees it. Returning users get a
clean single-line status: 'clawd-cursor v0.7.0 — desktop control
active on localhost:3847'.

- onboarding.ts: added printBanner(), called at top of runOnboarding()
  before the consent warning; also replaced raw box-drawing chars with
  unicode escapes in the consent box
- index.ts: removed full banner from start command; replaced with a
  compact one-liner status message

…lti-monitor, macOS perms

server.ts:
- Bearer token auth: generated at startup, saved to ~/.clawd-cursor/token (mode 0o600)
- Token printed to console on start alongside the server URL
- requireAuth middleware applied to all mutating + sensitive endpoints:
  POST /task, /action, /confirm, /abort, /favorites, /report, /stop
  GET  /screenshot
- CORS middleware: blocks cross-origin browser requests (SSRF/localhost-bypass
  prevention); only localhost:3847 origin is allowed; API callers without an
  Origin header (curl, CLI, MCP) pass through unaffected
- Favorites path moved from process.cwd() to ~/.clawd-cursor/ (persists across cwd)
- Imports DATA_DIR from paths.ts

agent.ts:
- Global 10-minute wall-clock timeout on executeTask() via Promise.race
- Timeout sets this.aborted = true so loops can exit cleanly
- Internal pipeline moved to _executeTaskInternal() — public API unchanged

safety.ts:
- type actions in terminal contexts (cmd, powershell, bash, wt, etc.) now
  require Confirm tier instead of Preview — keystrokes in a terminal can
  execute arbitrary shell commands

doctor.ts:
- macOS: added Screen Recording + Accessibility permission checks before the
  a11y bridge test; uses screencapture dry-run and osascript UI elements check;
  clear error messages linking to System Settings location

native-desktop.ts:
- Added MonitorInfo interface
- Added getMonitors(): enumerates all monitors on Windows (PowerShell
  System.Windows.Forms.Screen), macOS (osascript), Linux (xrandr)
- Added captureMonitor(index): captures a specific monitor region using
  nut-js screen.grabRegion() + sharp resize; falls back to primary on error

paths.ts:
- Added FAVORITES_PATH and TOKEN_PATH exports

…dinate scaling

Files:
- src/__tests__/coordinate-scaling.test.ts (14 tests) — pure math, zero deps
  Scale factor computation, LLM→real coord mapping, multi-monitor offsets

- src/__tests__/safety.test.ts (16 tests) — full SafetyLayer coverage
  Terminal type→Confirm tier (powershell/cmd/bash/wt), non-terminal→Preview,
  blocked patterns, confirm patterns, auto tier, isBlocked()

- src/__tests__/verifiers.test.ts (9 tests) — TaskVerifier behavior
  attemptLog always populated, error=FAIL not PASS, appOpen fast-path,
  clipboard fast-path pass/fail, graceful fallback without pipelineConfig

- src/__tests__/action-router.test.ts (16 tests) — ActionRouter routing logic
  Multi-step compound task rejection (5 cases), type routing (typeText call),
  URL navigation detection, write≠type, telemetry counting/reset

- vitest.config.ts — test runner config, node env, src/__tests__ glob

All tests mock nut-js and sharp to avoid native binary requirements.
55/55 passing, 0 failures.

… Groq, Llama-vision, xAI...)

generic-computer-use.ts (new file):
- Full screenshot → action → screenshot loop using OpenAI function-calling format
- Works with any provider that has a vision model + OpenAI-compat API
- DESKTOP_ACTION_TOOL: single structured tool with discriminated union of action types
  (screenshot, click, double_click, right_click, type, key, scroll, drag, move, wait, done)
- tool_choice: 'required' — LLM must always return a tool call, never prose-only
- Anti-loop guard: blocks LLM from taking >3 consecutive screenshots without acting
- Safety check on every action via SafetyLayer (blocked actions returned as tool error)
- Ground-truth verification on 'done' claims — verifier failure feeds back to LLM
- Coordinate scaling: LLM-space → real screen via same scale factor as other layers
- A11y tree (getScreenContext) injected into screenshot result for grounding
- 2-minute per-call timeout, 25 iteration max, graceful error passthrough
- isGenericComputerUseSupported(): checks vision model + API key, excludes Anthropic
  (which has its own native CU implementation)

agent.ts:
- Import GenericComputerUse and isGenericComputerUseSupported
- Add genericComputerUse field alongside computerUse
- Constructor: init genericComputerUse when Anthropic CU not available but vision key exists
- L3 dispatch updated with 3-tier cascade:
  1. this.computerUse   → Anthropic native Computer Use (tool spec, beta headers)
  2. this.genericComputerUse → Generic OpenAI-compat loop (GPT-4o, Gemini, Groq, etc.)
  3. executeLLMFallback → Legacy vision fallback (no structured tool schema, kept for compat)
- Step label updated: 'Layer 3 (Anthropic)' / 'Layer 3 (Generic)' / 'Layer 3 (legacy)'

providers.ts:
- Added Gemini (generativelanguage.googleapis.com/v1beta/openai, gemini-2.0-flash)
- Added Mistral AI (pixtral-large-latest for vision)
- Added xAI/Grok (grok-2-vision-1212)
- Key auto-detection: AIza → gemini, xai- → xai
- All three are openaiCompat: true — they all work with the generic CU loop

…resort

cdp-driver.ts:
- connect(): filter known OEM/vendor widget URLs (Lenovo Vantage, MSN/Bing widgets,
  NTP pages) in addition to edge:// and chrome:// — these are https:// URLs but
  behave as system pages with JS disabled, causing the agent to get stuck
- Among remaining user pages, pick the last one (most recently opened/navigated)
  instead of any arbitrary real page

a11y-reasoner.ts:
- Last-resort tab recovery: instead of window.location.href on the broken page
  (which fails when JS is disabled), open a fresh new tab via context.newPage()
  and navigate there, then attachToPage() the new tab
- This correctly escapes frozen/JS-disabled OEM widget tabs

- Added cdp_scroll handling in the CDP action dispatch block
  alongside cdp_click, cdp_type, cdp_read_text
- Supports direction (up/down/left/right), amount (px), optional selector
- Falls back to key_press ArrowDown on error rather than failing hard
- Added cdp_scroll examples to the L2 system prompt so the LLM knows
  it can scroll web pages natively without using keyboard shortcuts

Tested: successfully scrolled Reddit front page, found and clicked
upvote button via cdp_click by_text='upvote'. Task correctly escalated
to needs_human when Reddit login wall appeared (expected — not a bug).

…M as last resort

…owerShell

Implements the OcrEngine class (src/ocr-engine.ts) that provides OS-level
OCR with bounding boxes in real screen pixels. This is the foundation for
the OCR-first architecture in v0.8.0, replacing the a11y tree as the
primary UI read layer.

- Windows: PowerShell script using Windows.Media.Ocr WinRT API
- macOS/Linux: graceful stub (isAvailable() returns false)
- 300ms result cache with invalidateCache() for action dirty-bit
- 20 new unit tests (all 75 tests pass, 0 TS errors)

Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>

PowerShell's ConvertTo-Json can leave unescaped control characters
(e.g. bell \x07 from OCR'd icon text) that break JSON.parse().
Strips them in both the PS script and the TS layer for defense-in-depth.

Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>

…uts_execute)

Bridges the gap between the internal ActionRouter (which fuzzy-matches
tasks against 68 keyboard shortcuts) and external MCP agents that
previously had to independently know key combos. Two new tools:
  - shortcuts_list: query shortcuts by category and/or app context
  - shortcuts_execute: run a shortcut by intent with fuzzy matching

15 new tests, all 90 pass. Tool count: 33 → 35.

Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>

Complete v0.8.0 architecture — OCR is now the primary read layer:

  L0: LocalTaskParser (regex, no LLM)
  L1: ActionRouter (shortcuts, deterministic)
  L1.5: SmartInteraction (CDP + UIDriver)
  L2: SkillCache — replays learned task paths (grows over time)
  L2.5: OcrReasoner — OCR + a11y tree + text LLM (primary path)
  L3: VisionLLM — fallback only (unchanged)

Key design: OCR snapshot includes BOTH OCR text AND a11y tree, so if
OCR+A11y combined can't handle it, skip straight to vision (per user
requirement — no separate A11y-only fallback step).

New files:
  - src/ocr-reasoner.ts — L2.5 loop (OCR → text LLM → action → verify)
  - src/skill-cache.ts — learns from 2+ successful runs, auto-promotes
  - src/tools/ocr.ts — ocr_read_screen MCP tool (36 tools total)

Modified:
  - src/agent.ts — SkillCache (L2) + OcrReasoner (L2.5) inserted
  - src/providers.ts — ocrEnabled, skillCacheEnabled config flags
  - src/tools/index.ts — registered OCR tools

All 90 tests pass, 0 TypeScript errors. Live OCR tested successfully.

Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>

…rt_read, smart_type, invoke_element)

These 4 tools let MCP clients interact with the desktop WITHOUT screenshots
or coordinate math. They use a11y → CDP → OCR fallback chains:

- smart_read: primary perception tool, reads screen via a11y/CDP/OCR
- smart_click: click by element name, no coordinates needed
- smart_type: type into element by name with auto-focus
- invoke_element: direct UIA invoke with set-value/get-value/focus support

Key finding: a11y coordinates and nut-js mouseClick share the same coordinate
system — no conversion needed. The smart tools pass coords directly, avoiding
the broken a11yToMouse() dpiRatio division.

Also improved ocr_read_screen with dpiRatio hint for manual coordinate math.
Total: 40 MCP tools. 112 tests passing (22 new smart tool tests).

Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>

…lement

smart_read: runs OCR and a11y in parallel via Promise.all(), OCR text
shown first with a11y tree appended as supplement section.

smart_click: OCR scan + a11y invoke run in parallel; a11y invoke wins
if it succeeds (most reliable OS-level click), otherwise OCR coordinate
click, then a11y bounds fallback, then CDP as last resort.

Tests updated for OCR-first expectations (112/112 pass).

Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>

…ort fix

- tool-server.ts: validate request params against tool schema before execution
- safe-json.ts: balanced-brace JSON extraction replaces greedy regex across
  ai-brain.ts, smart-interaction.ts (prevents malformed LLM response crashes)
- ps-runner.ts: cap command queue at 100 (backpressure on long sessions)
- a11y-reasoner.ts, cdp-driver.ts, browser-layer.ts: replace silent catch {}
  blocks with console.debug logging for debuggability
- onboarding.ts + index.ts: honor custom --port in consent text and startup log
- docs/v0.7.0/index.html: update to 40 tools, OCR-first pipeline, smart tools

Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>

- Text model: sends "Reply with exactly: CLAWD_OK", verifies response
  contains the token (catches HTML error pages, quota errors, broken endpoints)
- Vision model: sends 8x8 green PNG image, verifies non-empty response
  (catches text-only endpoints that would fail at runtime in Layer 3)
- Smoke test phase: a11y reads active window title → sends to LLM → verifies
  round-trip (catches pipeline wiring bugs between perception and reasoning)
- Timeout reduced from 15s to 8s for text, 10s for vision
- extractErrorMessage() shared helper for consistent error formatting

Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>

…esults table

SKILL.md rewritten per task brief — explicit tool decision trees,
sensitive app policy, error recovery table, canvas app patterns.
537 lines, framework-agnostic. README test results table removed.
index.ts: deduplicate createToolContext() shared by mcp + serve.

Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>

…RM64)

- macOS OCR via Apple Vision framework (VNRecognizeTextRequest) — Swift script
- Linux OCR via Tesseract — Python script with CLI and pytesseract fallback
- macOS getFocusedElement via JXA — was returning null, now reads focused UI element
- ocr-engine.ts routes to platform-specific implementation at runtime
- All changes additive — zero impact to existing Windows code paths
- Platform support tables updated in README.md and SKILL.md

Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>

- npm package name: clawdcursor (was clawd-cursor)
- bin command: clawdcursor only (removed clawd-cursor alias)
- Data directory: ~/.clawdcursor/ (was ~/.clawd-cursor/)
- Migration: paths.ts auto-migrates from ~/.clawd-cursor/ and
  ~/.openclaw/clawd-cursor/ on first run
- All 22 files updated: src/, docs/, README, SKILL.md, website,
  package.json, CHANGELOG, scripts
- Display name "Clawd Cursor" (with space) unchanged — it's the brand

Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>

- .clawd-config.json → .clawdcursor-config.json (config file name)
- .clawd-favorites.json → .clawdcursor-favorites.json
- clawd-task-* → clawdcursor-task-* (temp file prefixes)
- clawd-ocr-* → clawdcursor-ocr-* (temp file prefixes)
- clawd-edge → clawdcursor-edge (Edge user data dir)
- Git remote updated: AmrDab/clawdcursor.git
- .gitignore covers both old and new config file names
- Tests updated to match new naming

Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>

…legate_to_agent auth

- POST /execute/* now requires Bearer token (was completely unprotected)
- Token generation moved to lazy init — stops CLI commands (stop, task, consent)
  from overwriting the running server's token on import
- delegate_to_agent and abort calls now include auth headers
- Website install instructions updated: npm install → git clone + npm run setup

Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>

…er listeners on disconnect