Remove interact verify endpoint

Add AGENTS playbook and prioritize automation backlog
Align interact verify test interval with schema minimum
2026-05-04 15:59:43 +02:00 · 2026-05-04 12:27:31 +02:00 · 2026-05-03 21:04:39 +02:00 · 2026-05-03 21:01:29 +02:00 · 2026-05-03 20:57:34 +02:00 · 2026-05-03 20:07:12 +02:00
11 changed files with 1361 additions and 2576 deletions
--- a/AGENTS.md
+++ b/AGENTS.md
@@ -0,0 +1,48 @@
 # AGENTS
 ## Purpose
 This file defines how agents and contributors should work in this repository.
 It is a codebase playbook, not a product roadmap.
 ## Repository Map
 - `server/app.py`: FastAPI routes, auth checks, response envelope, exception handling.
 - `server/services.py`: screenshot/OCR/input/window/exec behavior and safety enforcement.
 - `server/models.py`: request schemas and validation rules.
 - `server/config.py`: environment loading and runtime settings.
 - `tests/`: unit and API contract tests (monkeypatched where host behavior is nondeterministic).
 - `docs/API.md`: public API request/response reference.
 - `skill/SKILL.md`: operational method for agent usage of the API.
 ## Local Workflow
 - Install dependencies: `pip install -r requirements.txt`
 - Run server: `python -m server.app`
 - Run tests: `pytest -q`
 - Basic health check: `GET /health` with `x-clickthrough-token` when token auth is enabled.
 ## Non-Negotiable Contracts
 - Keep the response envelope shape stable:
  - `ok`, `request_id`, `time_ms`, `data`, `error`
 - Preserve one-action-per-request semantics in `/interact`.
 - Keep coordinate behavior in global desktop coordinates unless an explicit versioned change is introduced.
 - All new request fields must be represented in `server/models.py` with explicit validation constraints.
 ## Safety and Security Rules
 - Do not weaken `x-clickthrough-token` validation.
 - Do not weaken `/exec` secret validation (`x-clickthrough-exec-secret`).
 - Preserve `dry_run` behavior for non-destructive execution paths.
 - Preserve allowed-region enforcement for pointer-target actions.
 - Keep `/exec` constraints explicit:
  - shell allowlist
  - timeout limits
  - output truncation limits
 ## Testing Expectations
 - Add tests for each new behavior in services/models/routes.
 - Cover success and failure paths, especially validation and ambiguity branches.
 - Prefer deterministic tests with monkeypatches over host/UI-dependent flakiness.
 - If API behavior changes, update tests in the same change.
 ## Documentation Policy
 - When API behavior changes, update `docs/API.md` in the same change.
 - Keep examples aligned with current schema in `docs/API.md` and `examples/quickstart.py`.
 - Keep `skill/SKILL.md` aligned with current safe usage flow.
--- a/README.md
+++ b/README.md
@@ -1,82 +1,37 @@
 # Clickthrough
-Let an Agent interact with your computer over HTTP, with grid-aware screenshots and precise input actions.
+Clickthrough is a lightweight HTTP control layer that lets an AI safely operate a real computer by repeatedly capturing structured screenshots with coordinate-aware grids (`see`), executing precise mouse/keyboard actions from those coordinates (`interact`), and optionally running authenticated shell commands for system-level tasks (`exec`) under a consistent response contract.
-## What this provides
+## Core Methods
- **Visual endpoints**: full-screen capture with optional grid overlay and labeled cells (`asImage=true` can return raw image bytes)
+- `POST /see`: Capture a full screen or region, optionally with a click-ready grid overlay.
- **Zoom endpoint**: crop around a point with denser grid for fine targeting (`asImage=true` supported)
+- `POST /see/zoom`: Capture a tighter crop around a point and draw a denser grid for precise targeting.
- **Multi-display support**: list displays with `GET /displays` and select one with `?screen=0`, `?screen=1`, ...
+- `POST /interact`: Perform one mouse or keyboard action (`click`, `scroll`, `type`, `hotkey`, etc.).
- **Action endpoints**: move/click/right-click/double-click/middle-click/scroll/type/hotkey
+- `POST /exec`: Run PowerShell/Bash/CMD commands when shell-level control is needed.
 - **Window lifecycle endpoints**: list/focus/restore/minimize/maximize/close windows via `GET /windows` + `POST /windows/action`
 - **Structured launch endpoint**: start an app/process without dropping to a shell via `POST /launch`
 - **Wait/sync endpoint**: poll for text, window, or visual state changes via `POST /wait`
 - **Vision helper endpoints**: compare screenshots and measure stability via `POST /vision/diff` and `POST /vision/stability`
 - **OCR endpoints**: extract text blocks or search for matching text via `POST /ocr` and `POST /ocr/find`
 - **Compound verify endpoint**: execute an action and wait for a structured success condition via `POST /action/verify`
 - **Command execution endpoint**: run PowerShell/Bash/CMD commands via `POST /exec`
 - **Coordinate transform metadata** in visual responses so agents can map grid cells to real pixels
 - **Safety knobs**: token auth, dry-run mode, optional allowed-region restriction
-## Quick start
+## Why this works for AI agents
-```bash
+- Agents do not need live vision; they iterate on snapshots.
-cd /root/external-projects/clickthrough
+- Grid metadata bridges image understanding to deterministic click coordinates.
-python3 -m venv .venv
+- Interaction stays explicit and auditable (one action per request).
-. .venv/bin/activate
+- A unified response envelope (`ok`, `data`, `error`) reduces agent-side branching.
 pip install -r requirements.txt
 CLICKTHROUGH_TOKEN=change-me python -m server.app
 ```
-Server defaults to `127.0.0.1:8123`.
+## Minimal Agent Loop
-For OCR support, install the native `tesseract` binary on the host (in addition to Python deps), or point `CLICKTHROUGH_TESSERACT_CMD` at the executable if it lives somewhere weird.
+1. Call `see` with a coarse grid.
 2. If uncertain, call `see/zoom` with a denser grid.
 3. Call `interact` once.
 4. Call `see` again to verify state change.
 5. Use `exec` only for explicit shell/system tasks.
-`python-dotenv` is enabled, so values from a repo-root `.env` file are loaded automatically.
+## Safety and Auth
-## Minimal API flow
+- `x-clickthrough-token` protects API access when enabled.
 - `x-clickthrough-exec-secret` is required for `/exec`.
 - Optional dry-run and allowed-region constraints reduce accidental risk.
-1. `GET /displays` if you need a non-primary monitor
+## Docs
 2. `GET /screen?screen=0` with grid
 3. Decide cell / target
 4. Optional `POST /zoom?screen=0` for finer targeting
 5. `POST /action?screen=0` to execute (or `POST /action/verify?screen=0` for a bundled action+wait flow)
 6. `GET /screen?screen=0` again to verify result, or use `POST /wait`, `POST /vision/diff`, or `POST /ocr/find`
-Important:
+- API: `docs/API.md`
- `POST /action` expects an `action` plus a `target` object; do not send raw top-level `x` / `y` fields.
+- Agent procedure: `skill/SKILL.md`
- Pixel coordinates and OCR bounding boxes are always global desktop coordinates.
+- Coordinate system details: `docs/coordinate-system.md`
 - The agent does **not** inherently see the remote desktop; it reasons from screenshots, OCR, and window metadata.
 - When OCR is not enough, pair Clickthrough screenshots with OpenClaw's `image` tool for explicit screenshot interpretation.
 - Prefer structured GUI interaction first; use `/windows`, `/launch`, `/wait`, and `/action` before reaching for `/exec`.
 See:
 - `docs/API.md`
 - `docs/coordinate-system.md`
 - `skill/SKILL.md`
 ## Configuration
 Environment variables:
 - `CLICKTHROUGH_HOST` (default `127.0.0.1`)
 - `CLICKTHROUGH_PORT` (default `8123`)
 - `CLICKTHROUGH_TOKEN` (optional; if set, require `x-clickthrough-token` header)
 - `CLICKTHROUGH_DRY_RUN` (`true`/`false`; default `false`)
 - `CLICKTHROUGH_GRID_ROWS` (default `12`)
 - `CLICKTHROUGH_GRID_COLS` (default `12`)
 - `CLICKTHROUGH_ALLOWED_REGION` (optional `x,y,width,height`)
 - `CLICKTHROUGH_EXEC_ENABLED` (default `true`)
 - `CLICKTHROUGH_EXEC_SECRET` (**required for `/exec` to run**)
 - `CLICKTHROUGH_EXEC_DEFAULT_SHELL` (default `powershell`; one of `powershell`, `bash`, `cmd`)
 - `CLICKTHROUGH_EXEC_TIMEOUT_S` (default `30`)
 - `CLICKTHROUGH_EXEC_MAX_TIMEOUT_S` (default `120`)
 - `CLICKTHROUGH_EXEC_MAX_OUTPUT_CHARS` (default `20000`)
 - `CLICKTHROUGH_TESSERACT_CMD` (optional path to the `tesseract` executable)
 Window management endpoints currently target Windows hosts. On non-Windows hosts they return `501` instead of guessing.
 ## Gitea CI
 A Gitea Actions workflow is included at `.gitea/workflows/python-syntax.yml`.
 It runs Python syntax checks (`py_compile`) on every push and pull request.
--- a/TODO.md
+++ b/TODO.md
@@ -26,3 +26,10 @@
 - [x] Upgrade skill with verify-before-click rules, confidence thresholds, two-phase risky actions, and Spotify playbook
 - [x] Add top-level skill section for instance setup + mini API docs
 - [x] Clarify user-owned setup responsibilities vs agent responsibilities in skill docs
 ## Deferred Backlog (Prioritized)
 1. [ ] Higher-level automation macros composed from `see` + `interact`
 2. [ ] Reusable workflow templates (for example: find text -> zoom fallback -> click -> verify)
 3. [ ] Batch-safe orchestration primitives with explicit per-step results and auditability
 4. [ ] Additional verify primitives for post-action validation (image diff region, window title/process state, color/pixel checks)
 5. [ ] Broader API simplification pass to reduce payload overlap and consolidate shared OCR options
--- a/docs/API.md
+++ b/docs/API.md
@@ -1,62 +1,96 @@
-# API Reference (v0.1)
+# API Reference
 Base URL: `http://127.0.0.1:8123`
-If `CLICKTHROUGH_TOKEN` is set, include header:
+Auth header when enabled:
 ```http
 x-clickthrough-token: <token>
 ```
-## `GET /health`
+This API is intended for AI computer control through these methods:
 - `see`
 - `interact`
 - `exec`
-Returns status and runtime safety flags, including `exec` capability config.
+All responses use one envelope.
-## `GET /displays`
+## Response Envelope
-Returns detected displays in API screen order.
+Success:
 ```json
 {
  "ok": true,
-  "default_screen": 0,
+  "request_id": "...",
-  "displays": [
+  "time_ms": 1710000000000,
-    {"screen": 0, "mss_index": 1, "primary": true, "x": 0, "y": 0, "width": 1920, "height": 1080},
+  "data": {},
-    {"screen": 1, "mss_index": 2, "primary": false, "x": 1920, "y": 0, "width": 1920, "height": 1080}
+  "error": null
  ]
 }
 ```
-`screen` is zero-based. `screen=0` is the primary display when detectable, falling back to the first monitor reported by the capture backend.
+Error:
 Invalid `screen` values fall back to `0`.
 ## `GET /screen`
 Query params:
 - `screen` (int, default `0`) — zero-based display selector; invalid values fall back to `0`
 - `with_grid` (bool, default `true`)
 - `grid_rows` (int, default env or `12`)
 - `grid_cols` (int, default env or `12`)
 - `include_labels` (bool, default `true`)
 - `image_format` (`png`|`jpeg`, default `png`)
 - `jpeg_quality` (1-100, default `85`)
 - `asImage` (bool, default `false`) - if `true`, return raw image bytes only (`image/png` or `image/jpeg`)
 Default response includes base64 image and metadata (`meta.region`, `meta.screen`, `meta.displays`, optional `meta.grid`).
 `meta.region` uses global desktop coordinates.
 These image-returning endpoints do not magically grant the agent live vision.
 If the caller needs visual interpretation beyond OCR, pass the returned screenshot to OpenClaw's `image` tool and ask a narrow question about the visible UI.
 ## `POST /zoom`
 Body:
 ```json
 {
  "ok": false,
  "request_id": "...",
  "time_ms": 1710000000000,
  "data": null,
  "error": {
    "code": "validation_error",
    "message": "request validation failed",
    "details": []
  }
 }
 ```
 ## 1) See
 ### `POST /see`
 Capture a full screen or a region. Optional grid overlay returns coordinate metadata for click mapping.
 ```json
 {
  "screen": 0,
  "region_x": null,
  "region_y": null,
  "region_width": null,
  "region_height": null,
  "with_grid": true,
  "grid_rows": 12,
  "grid_cols": 12,
  "include_labels": true,
  "image_format": "png",
  "jpeg_quality": 85,
  "ocr": false,
  "ocr_min_confidence": 0,
  "ocr_lang": "eng",
  "ocr_psm": null
 }
 ```
 Returns:
 - `data.image.base64`
 - `data.meta.region` (global desktop coords)
 - `data.meta.grid` (rows/cols/cell size + formula)
 - `data.meta.ocr` (when `ocr=true`)
 OCR item shape:
 - `text`
 - `confidence`
 - `bbox` (global coords)
 - `center`
 - `region_relative_bbox`
 ### `POST /see/zoom`
 Capture a tighter crop around a global point and draw another grid over that crop.
 ```json
 {
  "screen": 0,
  "center_x": 1200,
-  "center_y": 700,
+  "center_y": 720,
  "width": 500,
  "height": 350,
  "with_grid": true,
@@ -68,70 +102,17 @@ Body:
 }
 ```
-Query params:
+Use this for precision before clicking tiny controls.
- `screen` (int, default `0`) - zero-based display selector; invalid values fall back to `0`
+## 2) Interact
 - `asImage` (bool, default `false`) - if `true`, return raw image bytes only (`image/png` or `image/jpeg`)
-Default response returns cropped image + region metadata in global pixel coordinates. `center_x` and `center_y` are also global coordinates; use the selected display's `meta.region` from `/screen?screen=X` as the coordinate base.
+### `POST /interact`
-
+Mouse/keyboard action execution.
 `POST /zoom` is often the best screenshot to hand to the `image` tool when the agent needs help judging a specific button, icon, or dialog layout.
 ## `POST /action`
 Body: one action.
 Important:
 - the request body uses `action` plus an optional `target`
 - pixel coordinates live inside `target` when `target.mode="pixel"`
 - do **not** send top-level `x` / `y` fields
 Query params:
 - `screen` (int, default `0`) - zero-based display selector included in the response metadata; invalid values fall back to `0`
 Pointer coordinates remain global desktop coordinates. For multi-display actions, first capture `/screen?screen=X` and use that response's `meta.region` or grid metadata to compute the target.
 ### Pointer target modes
 #### Pixel target
 ```json
 {
  "mode": "pixel",
  "x": 100,
  "y": 200,
  "dx": 0,
  "dy": 0
 }
 ```
 #### Grid target
 ```json
 {
  "mode": "grid",
  "region_x": 0,
  "region_y": 0,
  "region_width": 1920,
  "region_height": 1080,
  "rows": 12,
  "cols": 12,
  "row": 5,
  "col": 9,
  "dx": 0.0,
  "dy": 0.0
 }
 ```
 `dx`/`dy` are normalized offsets in `[-1, 1]` inside the selected cell.
 ### Action examples
 Click:
 ```json
 {
  "screen": 0,
  "action": {
    "action": "click",
    "target": {
      "mode": "grid",
@@ -143,438 +124,63 @@ Click:
      "cols": 12,
      "row": 7,
      "col": 3,
-    "dx": 0.2,
+      "dx": 0.0,
-    "dy": -0.1
+      "dy": 0.0
    },
-  "clicks": 1,
+    "button": "left",
-  "button": "left"
+    "clicks": 1
 }
 ```
 Scroll:
 ```json
 {
  "action": "scroll",
  "target": {"mode": "pixel", "x": 1300, "y": 740},
  "scroll_amount": -500
 }
 ```
 Type text:
 ```json
 {
  "action": "type",
  "text": "hello world",
  "interval_ms": 20
 }
 ```
 Hotkey:
 ```json
 {
  "action": "hotkey",
  "keys": ["ctrl", "l"]
 }
 ```
 Right click:
 ```json
 {
  "action": "right_click",
  "target": {"mode": "pixel", "x": 1300, "y": 740}
 }
 ```
 Move only:
 ```json
 {
  "action": "move",
  "target": {"mode": "pixel", "x": 1300, "y": 740},
  "duration_ms": 150
 }
 ```
 ## `GET /windows`
 List desktop windows using structured filters instead of shelling out.
 Query params:
 - `title_contains` (optional substring match)
 - `title_regex` (optional case-insensitive regex)
 - `process_name` (optional exact process name, e.g. `explorer.exe`)
 - `hwnd` (optional exact window handle)
 - `visible_only` (bool, default `true`)
 ```json
 {
  "ok": true,
  "count": 1,
  "windows": [
    {
      "hwnd": 132640,
      "title": "WinDirStat",
      "class_name": "WinDirStatMainWindow",
      "pid": 18420,
      "process_name": "windirstat.exe",
      "visible": true,
      "enabled": true,
      "minimized": false,
      "maximized": false,
      "foreground": true,
      "rect": {"x": 194, "y": 116, "width": 1532, "height": 870}
  }
  ]
 }
 ```
 Notes:
 - Currently supported on Windows hosts only.
 - Returns `409` for ambiguous write-target matches when a mutation endpoint would affect multiple windows.
 ## `POST /windows/action`
 Perform a structured window action against exactly one matched window.
 ```json
 {
  "action": "focus",
  "title_contains": "WinDirStat",
  "visible_only": true,
  "timeout_ms": 3000
 }
 ```
 Supported actions:
- `focus`
+- `move`, `click`, `right_click`, `double_click`, `middle_click`
- `restore`
+- `scroll` (`scroll_amount`)
- `minimize`
+- `type` (`text`, `interval_ms`)
- `maximize`
+- `hotkey` (`keys`)
- `close`
+- `click_text` (OCR-driven text click with optional region)
-The response includes the matched pre-action window and the final observed window state (or `closed=true` if it disappeared).
+Target modes:
-
+- `pixel`: absolute global `x,y`
-## `POST /launch`
+- `grid`: grid cell from a `see`/`see/zoom` response
 Start an app/process without invoking a shell.
 ```json
 {
  "executable": "C:/Program Files/WinDirStat/WinDirStat.exe",
  "args": [],
  "cwd": "C:/Program Files/WinDirStat",
  "wait_for_window": true,
  "match": {
    "title_contains": "WinDirStat",
    "visible_only": true
  },
  "timeout_ms": 8000
 }
 ```
 Notes:
 - Launch uses direct process execution (`subprocess.Popen`) rather than PowerShell/CMD.
 - If `wait_for_window=true`, the server polls for a matching window and returns `window_found`.
 - `dry_run=true` returns the resolved argv/cwd without launching.
 ## `POST /vision/diff`
 Measure whether a screen region changed meaningfully between two captures.
 Query params:
 - `screen` (int, default `0`) - used for `mode=screen` and `mode=region`
 Compare live captures:
 ```json
 {
  "mode": "region",
  "region_x": 120,
  "region_y": 80,
  "region_width": 600,
  "region_height": 300,
  "delay_ms": 400,
  "diff_threshold": 0.01
 }
 ```
 Compare provided images:
 ```json
 {
  "mode": "image",
  "before_image_base64": "iVBORw0KGgoAAA...",
  "after_image_base64": "iVBORw0KGgoBBB...",
  "diff_threshold": 0.01
 }
 ```
 Response includes:
 - `diff_ratio` — average normalized pixel difference
 - `changed` — whether `diff_ratio >= diff_threshold`
 - `region` — compared region
 ## `POST /vision/stability`
 Measure whether a screen region stays visually stable over a short interval.
 Query params:
 - `screen` (int, default `0`)
 ```json
 {
  "region_x": 0,
  "region_y": 0,
  "region_width": 1920,
  "region_height": 1080,
  "sample_interval_ms": 250,
  "duration_ms": 1200,
  "diff_threshold": 0.005
 }
 ```
 Response includes:
 - `stable`
 - `sample_count`
 - `max_diff_ratio`
 - `avg_diff_ratio`
 ## `POST /wait`
 Wait on a structured UI condition instead of guessing sleep durations.
 Query params:
 - `screen` (int, default `0`) - used for text and visual waits
 ### Wait for text to appear
 ```json
 {
  "condition": {
    "kind": "text",
    "mode": "screen",
    "text": "Scan complete",
    "match": "contains",
    "present": true,
    "language_hint": "eng",
    "min_confidence": 0.4
  },
  "timeout_ms": 15000,
  "poll_interval_ms": 400
 }
 ```
 ### Wait for a window state
 ```json
 {
  "condition": {
    "kind": "window",
    "title_contains": "WinDirStat",
    "visible_only": true,
    "state": "focused"
  },
  "timeout_ms": 5000,
  "poll_interval_ms": 200
 }
 ```
 Window states:
 - `exists`
 - `focused`
 - `closed`
 ### Wait for visual change or stability
 ```json
 {
  "condition": {
    "kind": "visual",
    "state": "stable",
    "region_x": 0,
    "region_y": 0,
    "region_width": 1920,
    "region_height": 1080,
    "diff_threshold": 0.005,
    "stable_for_ms": 1000
  },
  "timeout_ms": 12000,
  "poll_interval_ms": 300
 }
 ```
 Visual states:
 - `change` — succeeds when the average pixel diff crosses `diff_threshold`
 - `stable` — succeeds when the diff stays at or below `diff_threshold` for `stable_for_ms`
 Notes:
 - Text waits reuse the OCR pipeline and return matching OCR blocks on success.
 - Window waits build on the structured window discovery endpoint.
 - Visual waits compare repeated captures of either the full selected display or an explicit region.
 ## `POST /action/verify`
 Execute one action and wait for a structured success condition.
 Query params:
 - `screen` (int, default `0`)
 ### `click_text` example (full screen OCR)
 ```json
 {
  "screen": 0,
  "action": {
-    "action": "click",
+    "action": "click_text",
-    "target": {"mode": "pixel", "x": 1300, "y": 740}
+    "click_text": {
-  },
+      "text": "Sign in",
  "condition": {
    "kind": "text",
    "mode": "screen",
    "text": "Settings",
      "match": "contains",
-    "present": true,
+      "case_sensitive": false,
-    "language_hint": "eng",
+      "min_confidence": 45,
-    "min_confidence": 0.4
+      "occurrence": "best"
  },
  "retries": 1,
  "timeout_ms": 4000,
  "poll_interval_ms": 250,
  "retry_delay_ms": 250
 }
 ```
 Condition kinds mirror `POST /wait`:
 - `text`
 - `window`
 - `visual`
 The response returns per-attempt action output plus structured verification output.
 ## `POST /ocr`
 Extract visible text from either a full screenshot, a region crop, or caller-provided image bytes.
 Query params:
 - `screen` (int, default `0`) - zero-based display selector for `mode=screen` and `mode=region`; invalid values fall back to `0`
 Body:
 ```json
 {
  "mode": "screen",
  "language_hint": "eng",
  "min_confidence": 0.4
 }
 ```
 Modes:
 - `screen` (default): OCR over full selected monitor
 - `region`: OCR over explicit region (`region_x`, `region_y`, `region_width`, `region_height`)
 - `image`: OCR over provided `image_base64` (supports plain base64 or data URL)
 Region mode example:
 ```json
 {
  "mode": "region",
  "region_x": 220,
  "region_y": 160,
  "region_width": 900,
  "region_height": 400,
  "language_hint": "eng",
  "min_confidence": 0.5
 }
 ```
 Image mode example:
 ```json
 {
  "mode": "image",
  "image_base64": "iVBORw0KGgoAAAANSUhEUgAA...",
  "language_hint": "eng"
 }
 ```
 Response shape:
 ```json
 {
  "ok": true,
  "request_id": "...",
  "time_ms": 1710000000000,
  "result": {
    "mode": "screen",
    "language_hint": "eng",
    "min_confidence": 0.4,
    "region": {"x": 0, "y": 0, "width": 1920, "height": 1080},
    "blocks": [
      {
        "text": "Settings",
        "confidence": 0.9821,
        "bbox": {"x": 144, "y": 92, "width": 96, "height": 21}
    }
    ]
  }
 }
 ```
-Notes:
+### `click_text` example (region OCR)
 - Output is deterministic JSON (stable ordering by top-to-bottom, then left-to-right).
 - `bbox` coordinates are in global screen space for `screen`/`region`, and image-local for `image`.
 - Requires `tesseract` executable plus Python package `pytesseract`.
 - If `tesseract` is not on `PATH`, set `CLICKTHROUGH_TESSERACT_CMD` to the full executable path.
 ## `POST /ocr/find`
 Search OCR output for matching text instead of post-processing raw OCR blocks client-side.
 Query params:
 - `screen` (int, default `0`) - used for `mode=screen` and `mode=region`
 ```json
 {
-  "mode": "screen",
+  "screen": 0,
-  "query": "Settings",
+  "action": {
-  "match": "contains",
+    "action": "click_text",
-  "group_lines": true,
+    "click_text": {
-  "max_results": 10,
+      "text": "Continue",
-  "language_hint": "eng",
+      "match": "exact",
-  "min_confidence": 0.4
+      "region": { "x": 940, "y": 520, "width": 400, "height": 260 },
      "occurrence": "first"
    }
  }
 }
 ```
-Modes:
+## 3) Exec
 - `screen`
 - `region`
 - `image`
-Options:
+### `POST /exec`
- `match`: `contains`, `exact`, or `regex`
+Run host shell commands (PowerShell/Bash/CMD).
 - `group_lines=true`: combine nearby OCR words into line-level candidates before matching
 - `max_results`: result cap after confidence sorting
 Response includes:
 - `matches` — confidence-sorted candidate matches
 - `match_count`
 - `blocks_considered`
 ## `POST /exec`
 Execute a shell command on the host running Clickthrough.
 Requirements:
 - `CLICKTHROUGH_EXEC_SECRET` must be configured on the server
 - send header `x-clickthrough-exec-secret: <secret>`
 ```json
 {
@@ -586,29 +192,16 @@ Requirements:
 }
 ```
-Notes:
+Required header:
 - `shell` supports `powershell`, `bash`, `cmd`
 - if `shell` is omitted, server uses `CLICKTHROUGH_EXEC_DEFAULT_SHELL`
 - output is truncated based on `CLICKTHROUGH_EXEC_MAX_OUTPUT_CHARS`
 - endpoint can be disabled with `CLICKTHROUGH_EXEC_ENABLED=false`
 - if `CLICKTHROUGH_EXEC_SECRET` is missing, `/exec` is blocked (`403`)
-Response includes `stdout`, `stderr`, `exit_code`, timeout state, and execution metadata.
+```http
-
+x-clickthrough-exec-secret: <secret>
 ## `POST /batch`
 Runs multiple `action` payloads sequentially.
 Query params:
 - `screen` (int, default `0`) - zero-based display selector applied to each action response; invalid values fall back to `0`
 ```json
 {
  "actions": [
    {"action": "move", "target": {"mode": "pixel", "x": 100, "y": 100}},
    {"action": "click", "target": {"mode": "pixel", "x": 100, "y": 100}}
  ],
  "stop_on_error": true
 }
 ```
 ## Minimal Procedure for Agents
 1. `see` full screen with coarse grid.
 2. If uncertain, `see/zoom` target area with denser grid.
 3. `interact` one action.
 4. `see` again to confirm state change.
 5. Use `exec` only when GUI interaction is not the right tool.
--- a/examples/quickstart.py
+++ b/examples/quickstart.py
@@ -13,23 +13,52 @@ if TOKEN:
 def main():
-    r = requests.get(f"{BASE_URL}/health", headers=headers, timeout=10)
+    health = requests.get(f"{BASE_URL}/health", headers=headers, timeout=10)
-    r.raise_for_status()
+    health.raise_for_status()
-    print("health:", r.json())
+    print("health:", health.json()["data"])
-    d = requests.get(f"{BASE_URL}/displays", headers=headers, timeout=10)
+    see = requests.post(
-    d.raise_for_status()
+        f"{BASE_URL}/see",
    print("displays:", d.json().get("displays", []))
    s = requests.get(
        f"{BASE_URL}/screen",
        headers=headers,
-        params={"screen": SCREEN, "with_grid": True, "grid_rows": 12, "grid_cols": 12},
+        json={
            "screen": SCREEN,
            "with_grid": True,
            "grid_rows": 12,
            "grid_cols": 12,
            "image_format": "jpeg",
            "jpeg_quality": 70,
        },
        timeout=30,
    )
-    s.raise_for_status()
+    see.raise_for_status()
-    payload = s.json()
+    payload = see.json()["data"]
-    print("screen meta:", payload.get("meta", {}))
+    print("region:", payload["meta"]["region"])
    print("grid:", payload["meta"].get("grid", {}))
    see_ocr = requests.post(
        f"{BASE_URL}/see",
        headers=headers,
        json={"screen": SCREEN, "ocr": True, "with_grid": False, "ocr_min_confidence": 40},
        timeout=30,
    )
    see_ocr.raise_for_status()
    ocr_items = see_ocr.json()["data"]["meta"].get("ocr", [])
    print("ocr_items:", len(ocr_items))
    if ocr_items:
        label = ocr_items[0]["text"]
        click_text = requests.post(
            f"{BASE_URL}/interact",
            headers=headers,
            json={
                "screen": SCREEN,
                "action": {"action": "click_text", "click_text": {"text": label, "match": "exact", "occurrence": "first"}},
            },
            timeout=30,
        )
        click_text.raise_for_status()
        click_data = click_text.json()["data"]
        print("clicked:", click_data["resolved_target"])
 if __name__ == "__main__":
--- a/server/app.py
+++ b/server/app.py
--- a/server/config.py
+++ b/server/config.py
@@ -0,0 +1,42 @@
 import os
 from typing import Optional
 from dotenv import load_dotenv
 load_dotenv(dotenv_path=".env", override=False)
 def _env_bool(name: str, default: bool) -> bool:
    raw = os.getenv(name)
    if raw is None:
        return default
    return raw.strip().lower() in {"1", "true", "yes", "on"}
 def _parse_allowed_region() -> Optional[tuple[int, int, int, int]]:
    raw = os.getenv("CLICKTHROUGH_ALLOWED_REGION")
    if not raw:
        return None
    parts = [p.strip() for p in raw.split(",")]
    if len(parts) != 4:
        raise ValueError("CLICKTHROUGH_ALLOWED_REGION must be x,y,width,height")
    x, y, w, h = (int(p) for p in parts)
    if w <= 0 or h <= 0:
        raise ValueError("CLICKTHROUGH_ALLOWED_REGION width/height must be > 0")
    return x, y, w, h
 SETTINGS = {
    "host": os.getenv("CLICKTHROUGH_HOST", "127.0.0.1"),
    "port": int(os.getenv("CLICKTHROUGH_PORT", "8123")),
    "token": os.getenv("CLICKTHROUGH_TOKEN", "").strip(),
    "dry_run": _env_bool("CLICKTHROUGH_DRY_RUN", False),
    "allowed_region": _parse_allowed_region(),
    "exec_enabled": _env_bool("CLICKTHROUGH_EXEC_ENABLED", True),
    "exec_default_shell": os.getenv("CLICKTHROUGH_EXEC_DEFAULT_SHELL", "powershell").strip().lower(),
    "exec_default_timeout_s": int(os.getenv("CLICKTHROUGH_EXEC_TIMEOUT_S", "30")),
    "exec_max_timeout_s": int(os.getenv("CLICKTHROUGH_EXEC_MAX_TIMEOUT_S", "120")),
    "exec_max_output_chars": int(os.getenv("CLICKTHROUGH_EXEC_MAX_OUTPUT_CHARS", "20000")),
    "exec_secret": os.getenv("CLICKTHROUGH_EXEC_SECRET", "").strip(),
 }
--- a/server/models.py
+++ b/server/models.py
@@ -0,0 +1,167 @@
 from typing import Literal, Optional
 from pydantic import BaseModel, Field, model_validator
 class PixelTarget(BaseModel):
    mode: Literal["pixel"]
    x: int
    y: int
    dx: int = 0
    dy: int = 0
 class GridTarget(BaseModel):
    mode: Literal["grid"]
    region_x: int
    region_y: int
    region_width: int = Field(gt=0)
    region_height: int = Field(gt=0)
    rows: int = Field(gt=0)
    cols: int = Field(gt=0)
    row: int = Field(ge=0)
    col: int = Field(ge=0)
    dx: float = 0.0
    dy: float = 0.0
    @model_validator(mode="after")
    def _validate_indices(self):
        if self.row >= self.rows or self.col >= self.cols:
            raise ValueError("row/col must be inside rows/cols")
        if not -1.0 <= self.dx <= 1.0:
            raise ValueError("dx must be in [-1, 1]")
        if not -1.0 <= self.dy <= 1.0:
            raise ValueError("dy must be in [-1, 1]")
        return self
 Target = PixelTarget | GridTarget
 class ActionRequest(BaseModel):
    action: Literal[
        "move",
        "click",
        "right_click",
        "double_click",
        "middle_click",
        "scroll",
        "type",
        "hotkey",
        "click_text",
    ]
    target: Optional[Target] = None
    duration_ms: int = Field(default=0, ge=0, le=20000)
    button: Literal["left", "right", "middle"] = "left"
    clicks: int = Field(default=1, ge=1, le=10)
    scroll_amount: int = 0
    text: str = ""
    keys: list[str] = Field(default_factory=list)
    interval_ms: int = Field(default=20, ge=0, le=5000)
    dry_run: bool = False
    click_text: "ClickTextAction | None" = None
    @model_validator(mode="after")
    def _validate_click_text(self):
        if self.action == "click_text" and self.click_text is None:
            raise ValueError("click_text payload is required when action=click_text")
        return self
 class ExecRequest(BaseModel):
    command: str = Field(min_length=1, max_length=10000)
    shell: Literal["powershell", "bash", "cmd"] | None = None
    timeout_s: int | None = Field(default=None, ge=1, le=600)
    cwd: str | None = None
    dry_run: bool = False
 class WindowQuery(BaseModel):
    title_contains: str | None = Field(default=None, max_length=512)
    title_regex: str | None = Field(default=None, max_length=512)
    process_name: str | None = Field(default=None, max_length=260)
    hwnd: int | None = Field(default=None, ge=1)
    visible_only: bool = True
 class WindowActionRequest(WindowQuery):
    action: Literal["focus", "restore", "minimize", "maximize", "close"]
    timeout_ms: int = Field(default=3000, ge=0, le=60000)
 class LaunchRequest(BaseModel):
    executable: str = Field(min_length=1, max_length=2048)
    args: list[str] = Field(default_factory=list, max_length=100)
    cwd: str | None = None
    wait_for_window: bool = False
    match: WindowQuery | None = None
    timeout_ms: int = Field(default=5000, ge=0, le=120000)
    dry_run: bool = False
 class SeeRequest(BaseModel):
    screen: int = 0
    region_x: int | None = Field(default=None, ge=0)
    region_y: int | None = Field(default=None, ge=0)
    region_width: int | None = Field(default=None, gt=0)
    region_height: int | None = Field(default=None, gt=0)
    with_grid: bool = True
    grid_rows: int = Field(default=12, ge=1, le=300)
    grid_cols: int = Field(default=12, ge=1, le=300)
    include_labels: bool = True
    image_format: Literal["png", "jpeg"] = "png"
    jpeg_quality: int = Field(default=85, ge=1, le=100)
    ocr: bool = False
    ocr_min_confidence: float = Field(default=0.0, ge=0.0, le=100.0)
    ocr_lang: str = Field(default="eng", min_length=1, max_length=64)
    ocr_psm: int | None = Field(default=None, ge=0, le=13)
 class SeeZoomRequest(BaseModel):
    screen: int = 0
    center_x: int = Field(ge=0)
    center_y: int = Field(ge=0)
    width: int = Field(default=500, ge=10)
    height: int = Field(default=350, ge=10)
    with_grid: bool = True
    grid_rows: int = Field(default=20, ge=1, le=300)
    grid_cols: int = Field(default=20, ge=1, le=300)
    include_labels: bool = True
    image_format: Literal["png", "jpeg"] = "png"
    jpeg_quality: int = Field(default=90, ge=1, le=100)
 class InteractRequest(BaseModel):
    screen: int = 0
    action: ActionRequest
 class OCRRegion(BaseModel):
    x: int = Field(ge=0)
    y: int = Field(ge=0)
    width: int = Field(gt=0)
    height: int = Field(gt=0)
 class ClickTextAction(BaseModel):
    text: str = Field(min_length=1, max_length=1000)
    match: Literal["contains", "exact", "regex"] = "contains"
    region: OCRRegion | None = None
    screen: int | None = None
    case_sensitive: bool = False
    min_confidence: float = Field(default=0.0, ge=0.0, le=100.0)
    occurrence: Literal["first", "best", "nth"] = "first"
    nth: int | None = Field(default=None, ge=1, le=10000)
    ocr_lang: str = Field(default="eng", min_length=1, max_length=64)
    ocr_psm: int | None = Field(default=None, ge=0, le=13)
    @model_validator(mode="after")
    def _validate_nth(self):
        if self.occurrence == "nth" and self.nth is None:
            raise ValueError("nth is required when occurrence=nth")
        if self.occurrence != "nth" and self.nth is not None:
            raise ValueError("nth is only allowed when occurrence=nth")
        return self
 ActionRequest.model_rebuild()
--- a/server/services.py
+++ b/server/services.py
@@ -0,0 +1,602 @@
 import ctypes
 import io
 import os
 import re
 import subprocess
 import sys
 import time
 from typing import Literal
 from fastapi import HTTPException
 from PIL import ImageChops, ImageStat
 from .config import SETTINGS
 from .models import (
    ActionRequest,
    ClickTextAction,
    GridTarget,
    LaunchRequest,
    PixelTarget,
    Target,
    WindowActionRequest,
    WindowQuery,
 )
 def api_error(status_code: int, code: str, message: str, details=None):
    raise HTTPException(status_code=status_code, detail={"code": code, "message": message, "details": details})
 def import_capture_libs():
    try:
        from PIL import Image, ImageDraw
        import mss
        return Image, ImageDraw, mss
    except Exception as exc:
        raise HTTPException(status_code=500, detail=f"capture backend unavailable: {exc}") from exc
 def display_region(mon: dict, screen: int, mss_index: int, primary: bool) -> dict:
    return {
        "screen": screen,
        "mss_index": mss_index,
        "primary": primary,
        "x": mon["left"],
        "y": mon["top"],
        "width": mon["width"],
        "height": mon["height"],
    }
 def ordered_displays(sct) -> list[dict]:
    raw_monitors = list(enumerate(sct.monitors[1:], start=1))
    if not raw_monitors:
        raise HTTPException(status_code=500, detail="no displays detected")
    primary_pos = next((idx for idx, (_, mon) in enumerate(raw_monitors) if mon["left"] == 0 and mon["top"] == 0), 0)
    ordered = [raw_monitors[primary_pos]] + [item for idx, item in enumerate(raw_monitors) if idx != primary_pos]
    return [display_region(mon, screen=index, mss_index=mss_index, primary=(index == 0)) for index, (mss_index, mon) in enumerate(ordered)]
 def get_displays() -> list[dict]:
    _, _, mss = import_capture_libs()
    with mss.mss() as sct:
        return ordered_displays(sct)
 def select_display(screen: int) -> tuple[dict, list[dict], dict]:
    displays = get_displays()
    selected = displays[screen] if 0 <= screen < len(displays) else displays[0]
    return selected, displays, {"requested": screen, "selected": selected["screen"], "fallback": selected["screen"] != screen}
 def capture_screen(screen: int = 0):
    Image, _, mss = import_capture_libs()
    with mss.mss() as sct:
        displays = ordered_displays(sct)
        mon = displays[screen] if 0 <= screen < len(displays) else displays[0]
        shot = sct.grab({"left": mon["x"], "top": mon["y"], "width": mon["width"], "height": mon["height"]})
        image = Image.frombytes("RGB", shot.size, shot.rgb)
        selection = {"requested": screen, "selected": mon["screen"], "fallback": mon["screen"] != screen}
        return image, mon, displays, selection
 def capture_region_image(screen: int, region_x: int | None, region_y: int | None, region_width: int | None, region_height: int | None):
    base_img, mon, displays, screen_selection = capture_screen(screen)
    if None in {region_x, region_y, region_width, region_height}:
        return base_img, {"x": mon["x"], "y": mon["y"], "width": mon["width"], "height": mon["height"]}, mon, displays, screen_selection
    left = region_x - mon["x"]
    top = region_y - mon["y"]
    right = left + region_width
    bottom = top + region_height
    if left < 0 or top < 0 or right > base_img.size[0] or bottom > base_img.size[1]:
        raise HTTPException(status_code=400, detail="requested region is outside the captured monitor")
    crop = base_img.crop((left, top, right, bottom))
    return crop, {"x": region_x, "y": region_y, "width": region_width, "height": region_height}, mon, displays, screen_selection
 def extract_ocr_items(image, origin_x: int, origin_y: int, min_confidence: float, lang: str, psm: int | None) -> list[dict]:
    try:
        import pytesseract
    except Exception as exc:
        api_error(503, "ocr_unavailable", f"pytesseract unavailable: {exc}")
    config = ""
    if psm is not None:
        config = f"--psm {psm}"
    try:
        data = pytesseract.image_to_data(image, lang=lang, config=config, output_type=pytesseract.Output.DICT)
    except Exception as exc:
        api_error(503, "ocr_failed", f"ocr failed: {exc}")
    out: list[dict] = []
    n = len(data.get("text", []))
    for i in range(n):
        text = (data["text"][i] or "").strip()
        if not text:
            continue
        try:
            confidence = float(data["conf"][i])
        except Exception:
            continue
        if confidence < min_confidence:
            continue
        left = int(data["left"][i])
        top = int(data["top"][i])
        width = int(data["width"][i])
        height = int(data["height"][i])
        bbox = {"x": origin_x + left, "y": origin_y + top, "width": width, "height": height}
        center = {"x": bbox["x"] + (width // 2), "y": bbox["y"] + (height // 2)}
        out.append(
            {
                "text": text,
                "confidence": confidence,
                "bbox": bbox,
                "center": center,
                "region_relative_bbox": {"x": left, "y": top, "width": width, "height": height},
            }
        )
    return out
 def serialize_image(image, image_format: str, jpeg_quality: int) -> bytes:
    buf = io.BytesIO()
    if image_format == "jpeg":
        image.save(buf, format="JPEG", quality=jpeg_quality)
    else:
        image.save(buf, format="PNG")
    return buf.getvalue()
 def encode_image(image, image_format: str, jpeg_quality: int) -> str:
    import base64
    return base64.b64encode(serialize_image(image, image_format, jpeg_quality)).decode("ascii")
 def draw_grid(image, region_x: int, region_y: int, rows: int, cols: int, include_labels: bool):
    _, ImageDraw, _ = import_capture_libs()
    out = image.copy()
    draw = ImageDraw.Draw(out)
    w, h = out.size
    cell_w = w / cols
    cell_h = h / rows
    for c in range(1, cols):
        x = int(round(c * cell_w))
        draw.line([(x, 0), (x, h)], fill=(255, 0, 0), width=1)
    for r in range(1, rows):
        y = int(round(r * cell_h))
        draw.line([(0, y), (w, y)], fill=(255, 0, 0), width=1)
    draw.rectangle([(0, 0), (w - 1, h - 1)], outline=(255, 0, 0), width=2)
    if include_labels:
        for r in range(rows):
            for c in range(cols):
                cx = int((c + 0.5) * cell_w)
                cy = int((r + 0.5) * cell_h)
                draw.text((cx - 12, cy - 6), f"{r},{c}", fill=(255, 255, 0))
    meta = {
        "region": {"x": region_x, "y": region_y, "width": w, "height": h},
        "grid": {
            "rows": rows,
            "cols": cols,
            "cell_width": cell_w,
            "cell_height": cell_h,
            "indexing": "zero-based",
            "point_formula": {
                "pixel_x": "region.x + ((col + 0.5 + dx*0.5) * cell_width)",
                "pixel_y": "region.y + ((row + 0.5 + dy*0.5) * cell_height)",
                "dx_range": "[-1,1]",
                "dy_range": "[-1,1]",
            },
        },
    }
    return out, meta
 def resolve_target(target: Target) -> tuple[int, int, dict]:
    if isinstance(target, PixelTarget):
        x = target.x + target.dx
        y = target.y + target.dy
        return x, y, {"mode": "pixel", "source": target.model_dump()}
    cell_w = target.region_width / target.cols
    cell_h = target.region_height / target.rows
    x = target.region_x + int(round((target.col + 0.5 + (target.dx * 0.5)) * cell_w))
    y = target.region_y + int(round((target.row + 0.5 + (target.dy * 0.5)) * cell_h))
    return x, y, {"mode": "grid", "source": target.model_dump(), "derived": {"cell_width": cell_w, "cell_height": cell_h}}
 def enforce_allowed_region(x: int, y: int):
    region = SETTINGS["allowed_region"]
    if region is None:
        return
    rx, ry, rw, rh = region
    if not (rx <= x < rx + rw and ry <= y < ry + rh):
        raise HTTPException(status_code=403, detail="point outside allowed region")
 def _text_matches(candidate: str, needle: str, mode: str, case_sensitive: bool) -> bool:
    hay = candidate if case_sensitive else candidate.lower()
    ndl = needle if case_sensitive else needle.lower()
    if mode == "contains":
        return ndl in hay
    if mode == "exact":
        return hay == ndl
    flags = 0 if case_sensitive else re.IGNORECASE
    return re.search(needle, candidate, flags=flags) is not None
 def _resolve_text_match(click_text: ClickTextAction, items: list[dict]) -> dict:
    matches = [item for item in items if _text_matches(item["text"], click_text.text, click_text.match, click_text.case_sensitive)]
    if not matches:
        candidates = [item["text"] for item in sorted(items, key=lambda v: v["confidence"], reverse=True)[:8]]
        api_error(404, "ocr_text_not_found", "no OCR text matched", {"query": click_text.text, "candidates": candidates})
    if click_text.occurrence == "best":
        return max(matches, key=lambda item: item["confidence"])
    if click_text.occurrence == "nth":
        idx = (click_text.nth or 1) - 1
        if idx >= len(matches):
            api_error(409, "ocr_nth_out_of_range", "requested nth match is out of range", {"match_count": len(matches), "nth": click_text.nth})
        return matches[idx]
    if len(matches) > 1 and click_text.match == "exact":
        api_error(
            409,
            "ocr_text_ambiguous",
            "multiple OCR entries matched",
            {"match_count": len(matches), "candidates": [item["text"] for item in matches[:8]]},
        )
    return matches[0]
 def import_input_lib():
    try:
        import pyautogui
        pyautogui.FAILSAFE = True
        return pyautogui
    except Exception as exc:
        raise HTTPException(status_code=500, detail=f"input backend unavailable: {exc}") from exc
 def exec_action(req: ActionRequest, screen: int = 0) -> dict:
    run_dry = SETTINGS["dry_run"] or req.dry_run
    action_screen = screen
    if req.action == "click_text" and req.click_text and req.click_text.screen is not None:
        action_screen = req.click_text.screen
    selected_display, _, screen_selection = select_display(action_screen)
    pyautogui = None if run_dry else import_input_lib()
    resolved_target = None
    if req.target is not None:
        x, y, info = resolve_target(req.target)
        enforce_allowed_region(x, y)
        resolved_target = {"x": x, "y": y, "target_info": info}
    duration_sec = req.duration_ms / 1000.0
    if req.action in {"move", "click", "right_click", "double_click", "middle_click"} and resolved_target is None:
        raise HTTPException(status_code=400, detail="target is required for pointer actions")
    if req.action == "scroll" and resolved_target is None:
        raise HTTPException(status_code=400, detail="target is required for scroll")
    click_text_match = None
    if req.action == "click_text":
        if req.click_text is None:
            api_error(400, "click_text_payload_required", "click_text payload is required")
        region = req.click_text.region
        img, captured_region, _, _, _ = capture_region_image(
            action_screen,
            None if region is None else region.x,
            None if region is None else region.y,
            None if region is None else region.width,
            None if region is None else region.height,
        )
        items = extract_ocr_items(
            img,
            captured_region["x"],
            captured_region["y"],
            req.click_text.min_confidence,
            req.click_text.ocr_lang,
            req.click_text.ocr_psm,
        )
        matched = _resolve_text_match(req.click_text, items)
        enforce_allowed_region(matched["center"]["x"], matched["center"]["y"])
        click_text_match = {
            "query": req.click_text.model_dump(),
            "matched": matched,
            "capture_region": captured_region,
            "screen": screen_selection,
        }
        resolved_target = {"x": matched["center"]["x"], "y": matched["center"]["y"], "target_info": {"mode": "ocr_text"}}
    if not run_dry:
        if req.action == "move":
            pyautogui.moveTo(resolved_target["x"], resolved_target["y"], duration=duration_sec)
        elif req.action == "click":
            pyautogui.click(x=resolved_target["x"], y=resolved_target["y"], clicks=req.clicks, interval=req.interval_ms / 1000.0, button=req.button, duration=duration_sec)
        elif req.action == "right_click":
            pyautogui.click(x=resolved_target["x"], y=resolved_target["y"], button="right", duration=duration_sec)
        elif req.action == "double_click":
            pyautogui.doubleClick(x=resolved_target["x"], y=resolved_target["y"], interval=req.interval_ms / 1000.0)
        elif req.action == "middle_click":
            pyautogui.click(x=resolved_target["x"], y=resolved_target["y"], button="middle", duration=duration_sec)
        elif req.action == "scroll":
            pyautogui.moveTo(resolved_target["x"], resolved_target["y"], duration=duration_sec)
            pyautogui.scroll(req.scroll_amount)
        elif req.action == "type":
            pyautogui.write(req.text, interval=req.interval_ms / 1000.0)
        elif req.action == "hotkey":
            if len(req.keys) < 1:
                raise HTTPException(status_code=400, detail="keys is required for hotkey")
            pyautogui.hotkey(*req.keys)
        elif req.action == "click_text":
            pyautogui.click(
                x=resolved_target["x"],
                y=resolved_target["y"],
                clicks=req.clicks,
                interval=req.interval_ms / 1000.0,
                button=req.button,
                duration=duration_sec,
            )
    return {
        "action": req.action,
        "executed": not run_dry,
        "dry_run": run_dry,
        "screen": screen_selection,
        "display": selected_display,
        "resolved_target": resolved_target,
        "click_text_match": click_text_match,
    }
 def windows_only(feature: str):
    if sys.platform != "win32":
        raise HTTPException(status_code=501, detail=f"{feature} is currently supported on Windows hosts only")
 def tasklist_process_name(pid: int) -> str | None:
    try:
        completed = subprocess.run(["tasklist", "/FI", f"PID eq {pid}", "/FO", "CSV", "/NH"], capture_output=True, text=True, timeout=5, check=False)
    except Exception:
        return None
    line = (completed.stdout or "").strip().splitlines()
    if not line:
        return None
    row = line[0].strip()
    if not row or row.startswith("INFO:"):
        return None
    if row.startswith('"') and '","' in row:
        return row.split('","', 1)[0].strip('"')
    return None
 def list_windows(query: WindowQuery | None = None) -> list[dict]:
    windows_only("window endpoints")
    query = query or WindowQuery()
    user32 = ctypes.windll.user32
    kernel32 = ctypes.windll.kernel32
    psapi = ctypes.windll.psapi
    user32.GetWindowTextLengthW.argtypes = [ctypes.c_void_p]
    user32.GetWindowTextLengthW.restype = ctypes.c_int
    user32.GetWindowTextW.argtypes = [ctypes.c_void_p, ctypes.c_wchar_p, ctypes.c_int]
    user32.GetWindowTextW.restype = ctypes.c_int
    user32.IsWindowVisible.argtypes = [ctypes.c_void_p]
    user32.IsWindowVisible.restype = ctypes.c_bool
    user32.IsWindowEnabled.argtypes = [ctypes.c_void_p]
    user32.IsWindowEnabled.restype = ctypes.c_bool
    user32.IsIconic.argtypes = [ctypes.c_void_p]
    user32.IsIconic.restype = ctypes.c_bool
    user32.IsZoomed.argtypes = [ctypes.c_void_p]
    user32.IsZoomed.restype = ctypes.c_bool
    user32.GetForegroundWindow.restype = ctypes.c_void_p
    user32.GetWindowRect.argtypes = [ctypes.c_void_p, ctypes.POINTER(ctypes.wintypes.RECT)]
    user32.GetWindowRect.restype = ctypes.c_bool
    user32.GetClassNameW.argtypes = [ctypes.c_void_p, ctypes.c_wchar_p, ctypes.c_int]
    user32.GetClassNameW.restype = ctypes.c_int
    kernel32.OpenProcess.argtypes = [ctypes.wintypes.DWORD, ctypes.wintypes.BOOL, ctypes.wintypes.DWORD]
    kernel32.OpenProcess.restype = ctypes.wintypes.HANDLE
    kernel32.CloseHandle.argtypes = [ctypes.wintypes.HANDLE]
    kernel32.CloseHandle.restype = ctypes.wintypes.BOOL
    psapi.GetModuleBaseNameW.argtypes = [ctypes.wintypes.HANDLE, ctypes.wintypes.HMODULE, ctypes.c_wchar_p, ctypes.wintypes.DWORD]
    psapi.GetModuleBaseNameW.restype = ctypes.wintypes.DWORD
    foreground = int(user32.GetForegroundWindow() or 0)
    results: list[dict] = []
    def callback(hwnd, _lparam):
        hwnd_int = int(hwnd)
        if query.hwnd and hwnd_int != query.hwnd:
            return True
        visible = bool(user32.IsWindowVisible(hwnd))
        if query.visible_only and not visible:
            return True
        length = user32.GetWindowTextLengthW(hwnd)
        title_buf = ctypes.create_unicode_buffer(max(1, length + 1))
        user32.GetWindowTextW(hwnd, title_buf, len(title_buf))
        title = title_buf.value or ""
        if query.title_contains and query.title_contains.lower() not in title.lower():
            return True
        if query.title_regex and re.search(query.title_regex, title, flags=re.IGNORECASE) is None:
            return True
        pid = ctypes.wintypes.DWORD(0)
        user32.GetWindowThreadProcessId(hwnd, ctypes.byref(pid))
        process_name = tasklist_process_name(pid.value)
        if query.process_name and (process_name or "").lower() != query.process_name.lower():
            return True
        class_buf = ctypes.create_unicode_buffer(256)
        user32.GetClassNameW(hwnd, class_buf, len(class_buf))
        rect = ctypes.wintypes.RECT()
        user32.GetWindowRect(hwnd, ctypes.byref(rect))
        results.append(
            {
                "hwnd": hwnd_int,
                "title": title,
                "class_name": class_buf.value,
                "pid": int(pid.value),
                "process_name": process_name,
                "visible": visible,
                "enabled": bool(user32.IsWindowEnabled(hwnd)),
                "minimized": bool(user32.IsIconic(hwnd)),
                "maximized": bool(user32.IsZoomed(hwnd)),
                "foreground": hwnd_int == foreground,
                "rect": {"x": int(rect.left), "y": int(rect.top), "width": int(rect.right - rect.left), "height": int(rect.bottom - rect.top)},
            }
        )
        return True
    enum_proc = ctypes.WINFUNCTYPE(ctypes.c_bool, ctypes.c_void_p, ctypes.c_void_p)(callback)
    user32.EnumWindows(enum_proc, 0)
    results.sort(key=lambda item: (not item["foreground"], item["title"].lower(), item["hwnd"]))
    return results
 def _pick_single_window(query: WindowQuery) -> dict:
    matches = list_windows(query)
    if not matches:
        raise HTTPException(status_code=404, detail="no window matched")
    if len(matches) > 1:
        raise HTTPException(status_code=409, detail={"message": "multiple windows matched", "matches": matches[:10]})
    return matches[0]
 def apply_window_action(req: WindowActionRequest) -> dict:
    windows_only("window endpoints")
    match = _pick_single_window(req)
    hwnd = match["hwnd"]
    user32 = ctypes.windll.user32
    SW_RESTORE, SW_MINIMIZE, SW_MAXIMIZE = 9, 6, 3
    WM_CLOSE = 0x0010
    if req.action == "focus":
        user32.ShowWindow(hwnd, SW_RESTORE)
        ok = bool(user32.SetForegroundWindow(hwnd))
        if not ok:
            raise HTTPException(status_code=500, detail="failed to focus window")
    elif req.action == "restore":
        user32.ShowWindow(hwnd, SW_RESTORE)
    elif req.action == "minimize":
        user32.ShowWindow(hwnd, SW_MINIMIZE)
    elif req.action == "maximize":
        user32.ShowWindow(hwnd, SW_MAXIMIZE)
    elif req.action == "close":
        user32.PostMessageW(hwnd, WM_CLOSE, 0, 0)
    deadline = time.time() + (req.timeout_ms / 1000.0)
    final = None
    while time.time() <= deadline:
        current = list_windows(WindowQuery(hwnd=hwnd, visible_only=False))
        if not current:
            if req.action == "close":
                return {"matched": match, "closed": True, "final": None}
            time.sleep(0.05)
            continue
        final = current[0]
        if req.action == "focus" and final.get("foreground"):
            break
        if req.action in {"restore", "minimize", "maximize"}:
            break
        time.sleep(0.05)
    return {"matched": match, "closed": False, "final": final}
 def launch_app(req: LaunchRequest) -> dict:
    if req.cwd and not os.path.isdir(req.cwd):
        raise HTTPException(status_code=400, detail="cwd does not exist or is not a directory")
    argv = [req.executable, *req.args]
    cwd = req.cwd or None
    if req.dry_run or SETTINGS["dry_run"]:
        return {"executed": False, "dry_run": True, "argv": argv, "cwd": cwd}
    try:
        proc = subprocess.Popen(argv, cwd=cwd)
    except FileNotFoundError as exc:
        raise HTTPException(status_code=400, detail=f"executable not found: {exc}") from exc
    except OSError as exc:
        raise HTTPException(status_code=400, detail=f"failed to launch process: {exc}") from exc
    result = {"executed": True, "dry_run": False, "argv": argv, "cwd": cwd, "pid": proc.pid}
    if req.wait_for_window:
        query = req.match or WindowQuery(process_name=os.path.basename(req.executable), visible_only=True)
        deadline = time.time() + (req.timeout_ms / 1000.0)
        match = None
        while time.time() <= deadline:
            matches = list_windows(query)
            if matches:
                match = matches[0]
                break
            time.sleep(0.2)
        result["window"] = match
        result["window_found"] = match is not None
    return result
 def _truncate_text(text: str, limit: int) -> tuple[str, bool]:
    if len(text) <= limit:
        return text, False
    return text[:limit], True
 def _resolve_exec_program(shell_name: str, command: str) -> list[str]:
    if shell_name == "powershell":
        return ["powershell", "-NoProfile", "-NonInteractive", "-ExecutionPolicy", "Bypass", "-Command", command]
    if shell_name == "bash":
        return ["bash", "-lc", command]
    if shell_name == "cmd":
        return ["cmd", "/c", command]
    raise HTTPException(status_code=400, detail="unsupported shell")
 def exec_command(req):
    if not SETTINGS["exec_enabled"]:
        raise HTTPException(status_code=403, detail="exec endpoint disabled")
    if not SETTINGS["exec_secret"]:
        raise HTTPException(status_code=403, detail="exec secret not configured")
    shell_name = (req.shell or SETTINGS["exec_default_shell"] or "powershell").lower().strip()
    if shell_name not in {"powershell", "bash", "cmd"}:
        raise HTTPException(status_code=400, detail="unsupported shell")
    run_dry = SETTINGS["dry_run"] or req.dry_run
    timeout_s = req.timeout_s if req.timeout_s is not None else SETTINGS["exec_default_timeout_s"]
    timeout_s = min(timeout_s, SETTINGS["exec_max_timeout_s"])
    cwd = None
    if req.cwd:
        cwd = os.path.abspath(req.cwd)
        if not os.path.isdir(cwd):
            raise HTTPException(status_code=400, detail="cwd does not exist or is not a directory")
    argv = _resolve_exec_program(shell_name, req.command)
    if run_dry:
        return {"executed": False, "dry_run": True, "shell": shell_name, "command": req.command, "argv": argv, "timeout_s": timeout_s, "cwd": cwd}
    start = time.time()
    try:
        completed = subprocess.run(argv, cwd=cwd, capture_output=True, text=True, timeout=timeout_s, check=False)
    except subprocess.TimeoutExpired as exc:
        stdout, stdout_truncated = _truncate_text(str(exc.stdout or ""), SETTINGS["exec_max_output_chars"])
        stderr, stderr_truncated = _truncate_text(str(exc.stderr or ""), SETTINGS["exec_max_output_chars"])
        return {"executed": True, "timed_out": True, "shell": shell_name, "command": req.command, "argv": argv, "timeout_s": timeout_s, "cwd": cwd, "duration_ms": int((time.time() - start) * 1000), "exit_code": None, "stdout": stdout, "stderr": stderr, "stdout_truncated": stdout_truncated, "stderr_truncated": stderr_truncated}
    except FileNotFoundError as exc:
        raise HTTPException(status_code=400, detail=f"shell executable not found: {exc}") from exc
    stdout, stdout_truncated = _truncate_text(completed.stdout or "", SETTINGS["exec_max_output_chars"])
    stderr, stderr_truncated = _truncate_text(completed.stderr or "", SETTINGS["exec_max_output_chars"])
    return {"executed": True, "timed_out": False, "shell": shell_name, "command": req.command, "argv": argv, "timeout_s": timeout_s, "cwd": cwd, "duration_ms": int((time.time() - start) * 1000), "exit_code": completed.returncode, "stdout": stdout, "stderr": stderr, "stdout_truncated": stdout_truncated, "stderr_truncated": stderr_truncated}
--- a/skill/SKILL.md
+++ b/skill/SKILL.md
@@ -1,346 +1,64 @@
 ---
 name: clickthrough-http-control
-description: Control a local computer through the Clickthrough HTTP server using screenshot grids, OCR, zoomed grids, and pointer/keyboard actions. Use when an agent must operate GUI apps by repeatedly capturing the screen, reading visible text, refining target coordinates, and executing precise interactions (click/right-click/double-click/scroll/type/hotkey) with verification.
+description: Use 3 methods to control a computer: see (screenshot+grid), interact (mouse/keyboard), and exec (shell).
 ---
-# Clickthrough HTTP Control
+# Clickthrough Computer Control
-Use a strict observe-decide-act-verify loop.
+Use these methods:
 - `see`
 - `interact`
 - `exec`
-## Getting a computer instance (user-owned setup)
+## Method 1: See
-The **user/operator** is responsible for provisioning and exposing the target machine.
+Use `POST /see` to capture full screen or a region with a grid overlay.
-The agent should not assume it can self-install this stack.
+Use `POST /see/zoom` to capture a tighter crop with a denser grid.
-
+Use `POST /see` with `ocr=true` when text localization is needed.
 ### What the user must do
 1. Install dependencies and run Clickthrough on the target computer (default bind: `127.0.0.1:8123`).
 2. Expose access path to the agent (LAN/Tailscale/reverse proxy) and provide the base URL.
 3. Configure secrets on target machine:
   - `CLICKTHROUGH_TOKEN` for general API auth
   - `CLICKTHROUGH_EXEC_SECRET` for `/exec` calls
 4. Share connection details with the agent through a secure channel:
   - `base_url`
   - `x-clickthrough-token`
   - `x-clickthrough-exec-secret` (only when `/exec` is needed)
 ### What the agent should do
 1. Validate connection with `GET /health` using provided headers.
 2. Refuse `/exec` attempts when exec secret is missing/invalid.
 3. Ask user for missing setup inputs instead of guessing infrastructure.
 ## What the agent can actually see
 The agent does **not** inherently see the remote desktop.
 Clickthrough provides screenshots, OCR data, window metadata, and input endpoints — not native live vision.
 That means:
 - `GET /screen` and `POST /zoom` return image data the agent may need to inspect explicitly
 - `POST /ocr` returns machine-readable text blocks when text extraction is enough
 - the OpenClaw `image` tool is the right fallback when the agent needs judgment about visual layout, icons, button styling, dialog structure, or other non-OCR cues
 - every visual conclusion is only as fresh as the last screenshot; after an action, recapture before assuming the UI changed as expected
 Do not write or think as if the agent is directly watching the screen in real time.
 Say what you actually have: screenshots, OCR output, and fresh verification captures.
 ## Mini API map
 - `GET /health` → server status + safety flags
 - `GET /displays` → detected displays in zero-based API order
 - `GET /screen?screen=0` → full screenshot (JSON with base64 by default, or raw image with `asImage=true`)
 - `POST /zoom?screen=0` → cropped screenshot around point/region (also supports `asImage=true`)
 - `GET /windows` → discover visible desktop windows and their handles/processes
 - `POST /windows/action` → focus/restore/minimize/maximize/close a matched window
 - `POST /launch` → start an app/process without dropping to a shell
 - `POST /wait?screen=0` → wait for text, window, or visual state changes
 - `POST /vision/diff?screen=0` → compare screenshots or regions for meaningful visual change
 - `POST /vision/stability?screen=0` → measure short-interval visual stability
 - `POST /ocr` → text extraction with bounding boxes from full screen, region, or provided image bytes
 - `POST /ocr/find?screen=0` → search OCR output for matching text candidates
 - `POST /action?screen=0` → single interaction (`move`, `click`, `scroll`, `type`, `hotkey`, ...)
 - `POST /action/verify?screen=0` → execute one action plus structured success verification
 - `POST /batch?screen=0` → sequential action list
 - `POST /exec` → PowerShell/Bash/CMD command execution (requires configured exec secret + header)
 ### Display selection
 - Use `GET /displays` before operating on multi-monitor systems.
 - Use `?screen=X` on `/screen`, `/zoom`, `/ocr`, `/action`, and `/batch`; invalid values fall back to `screen=0`.
 - Treat returned `region` and OCR bounding boxes as global desktop coordinates, not screen-local coordinates.
 - Do not assume `screen=1` starts at `(0,0)`; it may start at `(1920,0)`, `(-1920,0)`, or another global offset.
 - If a screenshot came from `/screen?screen=1`, keep using that response's `region` metadata when forming later `/action` targets.
 ### OCR usage
 - Prefer `POST /ocr` when targeting text-heavy UI (menus, labels, buttons, dialogs).
 - Use `mode=screen` for discovery, then `mode=region` for precision and speed.
 - Use `language_hint` when known (for example `eng`) to improve consistency.
 - Filter noise with `min_confidence` (start around `0.4` and tune per app).
 - Treat OCR as one signal, not the only signal, before high-impact clicks.
 ### Screenshot + `image` tool usage
 Use the OpenClaw `image` tool when OCR is not enough.
 This is especially useful for:
 - identifying which visible button looks like the primary confirm action
 - understanding dialog layout or pane structure
 - distinguishing similar nearby controls by icon, spacing, or emphasis
 - checking whether a visual state changed after a click
 Good pattern:
 1. capture with `GET /screen` or `POST /zoom`
 2. hand that screenshot to the `image` tool
 3. ask a precise question about the visible UI
 4. convert the answer into a concrete Clickthrough target
 5. act once
 6. recapture and verify again, or use `POST /action/verify` when the action+postcondition loop is simple enough to bundle cleanly
 Ask narrow questions.
 Good:
 - "Which button in this dialog is the primary confirmation action?"
 - "Is the scan still running, or does this look complete?"
 - "Which of these tabs appears selected?"
 Bad:
 - "What should I click?"
 - "Use your eyes and do the task"
 - anything that assumes the model has live continuity without a new screenshot
 ### Header requirements
 - Always send `x-clickthrough-token` when token auth is enabled.
 - For `/exec`, also send `x-clickthrough-exec-secret`.
 ## `POST /action` request shape (important)
 `/action` always expects an `action` plus an optional `target` object.
 Do **not** invent top-level `x` / `y` fields.
 Minimal pixel click:
 ```json
 {
  "action": "click",
  "target": {"mode": "pixel", "x": 100, "y": 200},
  "button": "left",
  "clicks": 1
 }
 ```
 Minimal grid click:
 ```json
 {
  "action": "click",
  "target": {
    "mode": "grid",
    "region_x": 0,
    "region_y": 0,
    "region_width": 1920,
    "region_height": 1080,
    "rows": 12,
    "cols": 12,
    "row": 6,
    "col": 8,
    "dx": 0.0,
    "dy": 0.0
  }
 }
 ```
 Other canonical examples:
 ```json
 {"action": "move", "target": {"mode": "pixel", "x": 100, "y": 200}}
 {"action": "double_click", "target": {"mode": "pixel", "x": 100, "y": 200}}
 {"action": "right_click", "target": {"mode": "pixel", "x": 100, "y": 200}}
 {"action": "scroll", "target": {"mode": "pixel", "x": 100, "y": 200}, "scroll_amount": -500}
 {"action": "type", "text": "hello world", "interval_ms": 20}
 {"action": "hotkey", "keys": ["ctrl", "l"]}
 ```
 Rules:
- `dx` / `dy` belong inside `target`, not beside it.
+- Start with coarse grid (`12x12`).
- `type` and `hotkey` usually do not need a `target`.
+- For precision, zoom and use denser grid (`20x20` or higher).
- For pixel targets, `x` / `y` are global desktop coordinates.
+- Always use returned `meta.region` and `meta.grid` when computing click targets.
- For grid targets, copy the exact `region_*`, `rows`, and `cols` basis from the screenshot/zoom you actually used.
+- Coordinates are global desktop coordinates.
 - OCR results are in `data.meta.ocr` and include confidence, bbox, and center.
-## When to use `/exec`
+## Method 2: Interact
-Prefer structured GUI control first:
+Use `POST /interact` for one action at a time.
 - `/screen`, `/zoom`, `/ocr` to observe
 - `/action` or `/batch` to interact
-Use `/exec` only when it is the cleanest available tool for the job, for example:
+Mouse actions:
- querying machine state that the GUI does not expose well
+- `move`, `click`, `right_click`, `double_click`, `middle_click`, `scroll`
- performing an explicit user-requested shell/system task
+- `click_text` (OCR-driven click; optionally scope with `click_text.region`)
 - recovering from a blocked GUI flow when normal interaction failed
-Prefer `GET /windows`, `POST /windows/action`, and `POST /launch` for app lifecycle tasks before falling back to `/exec`.
+Keyboard actions:
-Avoid using `/exec` for routine in-app clicks, menu navigation, or text entry when the GUI can be driven directly.
+- `type`, `hotkey`
-## Core workflow (mandatory)
+Rules:
 - Prefer `grid` targets derived from fresh `see`/`see/zoom` captures.
 - For text buttons/labels, prefer `click_text` and bound OCR with a region when possible.
 - Use `pixel` only when you already have reliable coordinates.
 - After each important action, call `see` again before continuing.
-1. Call `GET /screen?screen=0` with coarse grid (e.g., 12x12), or another selected display.
+## Method 3: Exec
 2. Identify likely target region and compute an initial confidence score.
 3. If confidence < 0.85, call `POST /zoom` with denser grid (e.g., 20x20) and re-evaluate.
 4. **Before any click**, verify target identity (OCR text/icon/location consistency).
 5. If OCR is insufficient, inspect the screenshot explicitly with the OpenClaw `image` tool instead of pretending you can already see enough.
 6. Execute one minimal action via `POST /action`.
 7. Re-capture with `GET /screen` or use `POST /wait`, `POST /vision/diff`, `POST /vision/stability`, or `POST /action/verify` to verify the expected state change.
 8. Repeat until objective is complete.
-## Verify-before-click rules
+Use `POST /exec` only for shell/system tasks.
- Never click if target identity is ambiguous.
+Rules:
- Require at least two matching signals before click.
+- Requires `x-clickthrough-exec-secret`.
- Good signal pairs include:
+- Do not use exec for normal clicking/typing flows.
-  - OCR text + expected UI region
+- Prefer GUI interaction first; exec is fallback or explicit shell task.
  - OCR text + matching button shape/icon nearby
  - dialog title text + expected button position within that dialog
  - known app/window focus + expected control location
 - If confidence is low, do not "test click"; zoom and re-localize first.
 - For high-impact actions (close/delete/send/purchase), use two-phase flow:
  1) preview intended coordinate + reason
  2) execute only after explicit confirmation.
-## Precision rules
+## Lightweight Procedure
- Prefer grid targets first, then use `dx/dy` for subcell precision.
+1. `see` capture.
- Keep `dx/dy` in `[-1,1]`; start at `0,0` and only offset when needed.
+2. If needed, `see/zoom` refine.
- Use zoom before guessing offsets.
+3. `interact` one step (`click_text` for text UI targets).
- Avoid stale coordinates: re-capture before action if UI moved/scrolled.
+4. `see` verify.
 5. Repeat.
-## Safety rules
+## Quick Safety Rules
- Respect `dry_run` and `allowed_region` restrictions from `/health`.
+- Never click with stale screenshots.
- Respect `/exec` security requirements (`CLICKTHROUGH_EXEC_SECRET` + `x-clickthrough-exec-secret`).
+- Never send multiple uncertain clicks in a row.
- Avoid destructive shortcuts unless explicitly requested.
+- If localization is ambiguous, re-capture with a tighter zoom.
 - Send one action at a time unless deterministic; then use `/batch`.
 ## Reliability rules
 - After every meaningful action, verify with a fresh screenshot.
 - On mismatch, do not spam clicks: zoom, re-localize, and retry once.
 - Prefer short, reversible actions over long macros.
 - If two retries fail, switch strategy (hotkey/window focus/search) instead of repeating the same click.
 ## Fallback ladder for uncertain targeting
 1. Full-screen capture with a coarse grid.
 2. Zoom into the candidate area with a denser grid.
 3. OCR the full screen or the tighter region.
 4. Re-anchor on a more reliable nearby control, title, or label.
 5. Try a keyboard-first flow if the app supports it.
 6. Use `/exec` only if GUI control is blocked and shell-level intervention is genuinely cleaner.
 Do not skip from "uncertain click" straight to random retries.
 ## Concrete screenshot -> `image` -> action example
 Example loop:
 1. `GET /screen?screen=0` to capture the current app state
 2. if the UI is text-heavy, try `POST /ocr` first
 3. if OCR does not answer the real question, pass the screenshot to the OpenClaw `image` tool with a narrow prompt like:
   - "In this save dialog, which visible button is the primary action?"
   - "Is there a dismiss/close button in the top-right of this modal?"
 4. map the answer back to a Clickthrough target using the returned grid/region metadata
 5. click once with `POST /action`
 6. recapture the screen
 7. optionally use `POST /wait` or another `image`/OCR check to confirm the result
 The key rule is simple: screenshot first, interpret second, click third, verify fourth.
 Do not collapse those steps into fake certainty.
 ## App-specific playbooks (recommended)
 Build per-app routines for repetitive tasks instead of generic clicking.
 ### Launcher / search / start app playbook
 Use this when the goal is "open app X" or "bring up tool Y".
 1. check `GET /windows` first in case the app is already open
 2. if present, use `POST /windows/action` to focus or restore it
 3. if absent, prefer `POST /launch` when you know the executable path
 4. if launch path is unknown but the OS launcher/search UI is available, use a keyboard-first flow:
   - open launcher (`win`, `cmd+space`, or app-specific shortcut depending on host)
   - type exact app name
   - wait for stable results with `POST /wait` or recapture
   - verify the result text with OCR or the `image` tool
   - press Enter or click the exact result once
 5. verify the app window now exists or is focused
 Do not keep relaunching if the window already exists; that’s sloppy.
 ### Dialog confirmation playbook
 Use for modals like save/discard, delete confirmation, permission prompts, and installer dialogs.
 1. capture the dialog region with `POST /zoom`
 2. use OCR first for title/body/button labels
 3. if button hierarchy or emphasis matters, inspect the zoomed screenshot with the `image` tool
 4. identify the exact intended action (`Cancel`, `Save`, `Allow`, `Delete`, etc.)
 5. for destructive actions, require explicit user confirmation unless already requested
 6. click once and verify the dialog disappeared or changed state
 Good verification targets:
 - dialog title vanished
 - expected next window appeared
 - destructive side effect is visible and confirmed
 ### File picker playbook
 Use for open/save dialogs.
 1. verify the file picker window is focused
 2. OCR the visible breadcrumb/path area, filename field, and button row
 3. prefer keyboard-first entry when possible:
   - type or paste the target path/name into the focused field
   - use `tab` / `shift+tab` to move predictably between filename and action buttons
 4. if the target path is uncertain, use OCR plus the `image` tool to identify the active field and selected folder/file row
 5. verify the intended filename/path is visible before confirming
 6. activate `Open` / `Save` once and verify the picker closes
 If the picker stays open, stop and inspect why instead of hammering Enter like a maniac.
 ### Browser tab / window playbook
 Use for browser navigation, tab targeting, or web app recovery.
 1. use `GET /windows` to focus the correct browser window first
 2. prefer keyboard-first navigation:
   - `ctrl+l` / `cmd+l` to focus the address bar
   - `ctrl+tab` / `ctrl+shift+tab` for tab movement when order is known
   - `ctrl+w` only for explicitly requested close actions
 3. verify tab or page identity with OCR on the tab strip or page heading
 4. if multiple similar tabs are open, zoom into the tab strip and use the `image` tool to distinguish active vs inactive tabs
 5. after navigation, wait for visual stability or expected text before taking the next action
 Do not assume a page loaded just because the click landed. Verify it.
 ### Settings / preferences navigation playbook
 Use when the task involves toggles, dropdowns, sidebars, or nested settings panels.
 1. identify the current settings page with OCR on the heading/sidebar
 2. use OCR to find the specific section label before trying to toggle anything
 3. if the layout is dense, zoom into the relevant pane and use the `image` tool to distinguish labels from controls
 4. prefer small reversible actions: one toggle, one dropdown, one field edit at a time
 5. after each change, verify the control state changed visually or via visible text
 6. if a save/apply button exists, treat it as a separate confirmation step and verify completion
 Settings UIs love hiding side effects. Assume nothing.
 ### Spotify playbook
 - Focus app window before search/navigation.
 - Prefer keyboard-first flow for song start:
  1) `Ctrl+L` (search)
  2) type exact query
  3) Enter
  4) verify exact song+artist text
  5) click/double-click row
  6) verify now-playing bar
 - If now-playing does not match target track, stop and re-localize; do not keep clicking nearby rows.
--- a/tests/test_ocr_and_interact.py
+++ b/tests/test_ocr_and_interact.py
@@ -0,0 +1,114 @@
 import sys
 from PIL import Image
 from fastapi.testclient import TestClient
 from server import services
 from server.app import app
 from server.config import SETTINGS
 from server.models import ClickTextAction
 def _auth_headers() -> dict:
    token = SETTINGS.get("token", "")
    if not token:
        return {}
    return {"x-clickthrough-token": token}
 def test_extract_ocr_items_normalization(monkeypatch):
    class FakeOutput:
        DICT = "DICT"
    class FakeTesseract:
        Output = FakeOutput
        @staticmethod
        def image_to_data(_image, lang, config, output_type):
            assert lang == "eng"
            assert output_type == "DICT"
            return {
                "text": ["hello", " ", "world"],
                "conf": ["95.0", "-1", "62.5"],
                "left": [10, 12, 40],
                "top": [20, 25, 60],
                "width": [30, 10, 50],
                "height": [10, 10, 12],
            }
    monkeypatch.setitem(sys.modules, "pytesseract", FakeTesseract)
    items = services.extract_ocr_items(Image.new("RGB", (100, 100)), origin_x=100, origin_y=200, min_confidence=60, lang="eng", psm=None)
    assert len(items) == 2
    assert items[0]["text"] == "hello"
    assert items[0]["bbox"]["x"] == 110
    assert items[0]["center"]["y"] == 225
    assert items[1]["text"] == "world"
 def test_resolve_text_match_contains_exact_regex_and_nth():
    items = [
        {"text": "Save", "confidence": 70},
        {"text": "Save as", "confidence": 96},
        {"text": "SAVE", "confidence": 88},
    ]
    contains = services._resolve_text_match(ClickTextAction(text="save", match="contains", occurrence="first"), items)
    assert contains["text"] == "Save"
    best = services._resolve_text_match(ClickTextAction(text="save", match="contains", occurrence="best"), items)
    assert best["text"] == "Save as"
    exact_case = services._resolve_text_match(
        ClickTextAction(text="SAVE", match="exact", case_sensitive=True, occurrence="first"),
        items,
    )
    assert exact_case["text"] == "SAVE"
    regex_nth = services._resolve_text_match(ClickTextAction(text="^Save", match="regex", occurrence="nth", nth=2), items)
    assert regex_nth["text"] == "Save as"
 def test_interact_click_text_region_optional(monkeypatch):
    monkeypatch.setattr(services, "select_display", lambda screen: ({"screen": screen}, [], {"requested": screen, "selected": screen, "fallback": False}))
    monkeypatch.setattr(
        services,
        "capture_region_image",
        lambda screen, x, y, w, h: (Image.new("RGB", (20, 20)), {"x": x or 0, "y": y or 0, "width": w or 20, "height": h or 20}, {}, [], {}),
    )
    monkeypatch.setattr(
        services,
        "extract_ocr_items",
        lambda *args, **kwargs: [
            {
                "text": "Apply",
                "confidence": 93.0,
                "bbox": {"x": 10, "y": 20, "width": 20, "height": 10},
                "center": {"x": 20, "y": 25},
                "region_relative_bbox": {"x": 10, "y": 20, "width": 20, "height": 10},
            }
        ],
    )
    client = TestClient(app)
    response = client.post(
        "/interact",
        json={"screen": 0, "action": {"action": "click_text", "dry_run": True, "click_text": {"text": "Apply", "match": "contains"}}},
        headers=_auth_headers(),
    )
    assert response.status_code == 200
    body = response.json()["data"]
    assert body["resolved_target"]["x"] == 20
    assert body["click_text_match"]["matched"]["text"] == "Apply"
 def test_see_ocr_off_on_contract(monkeypatch):
    monkeypatch.setattr(
        "server.app.capture_region_image",
        lambda *args, **kwargs: (Image.new("RGB", (10, 10)), {"x": 0, "y": 0, "width": 10, "height": 10}, {"screen": 0}, [], {}),
    )
    monkeypatch.setattr("server.app.encode_image", lambda *args, **kwargs: "abc")
    monkeypatch.setattr("server.app.extract_ocr_items", lambda *args, **kwargs: [{"text": "x"}])
    client = TestClient(app)
    off = client.post("/see", json={"ocr": False, "with_grid": False}, headers=_auth_headers())
    assert off.status_code == 200
    assert "ocr" not in off.json()["data"]["meta"]
    on = client.post("/see", json={"ocr": True, "with_grid": False}, headers=_auth_headers())
    assert on.status_code == 200
    assert on.json()["data"]["meta"]["ocr"][0]["text"] == "x"
Author	SHA1	Message	Date
Space-Banane	22ca0097d1	Remove interact verify endpoint All checks were successful python-syntax / syntax-check (push) Successful in 31s Details	2026-05-04 15:59:43 +02:00
space	f05e0c56e6	Add AGENTS playbook and prioritize automation backlog All checks were successful python-syntax / syntax-check (push) Successful in 6s Details	2026-05-04 12:27:31 +02:00
space	211f38003e	Align interact verify test interval with schema minimum All checks were successful python-syntax / syntax-check (push) Successful in 10s Details	2026-05-03 21:04:39 +02:00
space	c1fc97e198	Fix auth headers in OCR/interact endpoint tests All checks were successful python-syntax / syntax-check (push) Successful in 1m7s Details	2026-05-03 21:01:29 +02:00
space	9e816e0417	Add pytesseract OCR, click_text interact action, and interact verify endpoint All checks were successful python-syntax / syntax-check (push) Successful in 6s Details	2026-05-03 20:57:34 +02:00
Paul Wähner	1c03cab457	refactor: simplify to see/interact/exec and split server modules All checks were successful python-syntax / syntax-check (push) Successful in 6s Details	2026-05-03 20:07:12 +02:00
Paul Wähner	aced5be25e	feat: migrate to v2-only API and unified response envelope All checks were successful python-syntax / syntax-check (push) Successful in 7s Details	2026-05-03 19:11:11 +02:00
Luna	2585bc3a7c	docs(skill): prefer vision for target localization All checks were successful python-syntax / syntax-check (push) Successful in 11s Details	2026-05-01 17:16:37 +02:00
Luna	66615c8a81	docs(skill): refine clickthrough control guidance	2026-05-01 17:15:19 +02:00