@gleanwork/mcp-server-tester 2.0.0-beta.0 → 2.0.0-beta.1

package/dist/reporters/mcpReporter.d.cts

@@ -165,6 +165,9 @@ interface HostEvent {
     server?: string;
     arguments?: Record<string, unknown>;
     output?: string;
+    /** Explicit tool-result error status, absent when not observed. */
+    isError?: boolean;
+    rawName?: string;
     id?: string;
 }
 

package/dist/reporters/mcpReporter.d.ts

@@ -165,6 +165,9 @@ interface HostEvent {
     server?: string;
     arguments?: Record<string, unknown>;
     output?: string;
+    /** Explicit tool-result error status, absent when not observed. */
+    isError?: boolean;
+    rawName?: string;
     id?: string;
 }
 

package/dist/types/index.d.cts

@@ -1,4 +1,4 @@
-export { A as AuthType, K as BuiltInRubric, N as BuiltInSanitizer, Q as CLIConfig, W as CLIOAuthClientConfig, X as CLIOAuthResult, Y as CLIOutputFormat, Z as CaseComparisonResult, _ as ClientCredentialsConfig, $ as CompareEvalRunsOptions, a0 as ComparisonOutcome, a1 as ContentBlock, a2 as CreateMCPClientOptions, a3 as CustomJudgeExecutor, a4 as CustomJudgeResult, a9 as EvalCase, aa as EvalCaseComparison, ab as EvalCaseComparisonOutcome, ac as EvalCaseRequest, l as EvalCaseResult, ae as EvalContext, o as EvalDataset, ag as EvalExpectBlock, ah as EvalExpectationResult, ak as EvalMode, ao as EvalRunComparisonLabels, ap as EvalRunComparisonResult, aq as EvalRunMetadata, ar as EvalRunnerOptions, r as EvalRunnerResult, ay as ExpectationBreakdown, az as ExpectationResultMap, aA as ExpectationType, aB as ExperimentMetric, aJ as FieldRemovalSanitizer, aY as HostType, aZ as HttpMCPConfig, a_ as IterationResult, G as Judge, F as JudgeConfig, a$ as JudgeExpectConfig, J as JudgeMatcherOptions, b0 as JudgeResult, b1 as JudgeValidatorConfig, L as LLMProvider, b2 as LLMToolCall, b4 as LoadDatasetOptions, b5 as MCPAuthConfig, b6 as MCPAuthFixtures, b7 as MCPClientCredentialsConfig, p as MCPConfig, b9 as MCPConformanceCheck, ba as MCPConformanceOptions, bb as MCPConformanceRaw, bc as MCPConformanceResult, bd as MCPConformanceResultData, be as MCPEvalData, bf as MCPEvalHistoricalSummary, bg as MCPEvalReporterConfig, bh as MCPEvalRunData, M as MCPFixtureApi, bi as MCPFixtureOptions, bj as MCPHostCapabilities, a as MCPHostConfig, b as MCPHostSimulationResult, bk as MCPHostSimulator, bl as MCPOAuthConfig, bm as MCPServerCapabilitiesData, bp as NormalizedToolResponse, O as OAuthSetupConfig, P as PatternValidatorOptions, bs as PlaywrightOAuthClientProviderConfig, bt as PredicateResult, bu as ProposeVariantsContext, bv as ProtectedResourceDiscoveryResult, bw as ProtectedResourceMetadata, bx as ProviderKind, by as RegexSanitizer, bz as ResultSource, R as RubricSpec, bC as SaveBaselineOptions, bD as SaveEvalRunComparisonOptions, bE as SaveServerComparisonOptions, bF as SchemaRegistry, S as SchemaValidatorOptions, bG as SerializedEvalDataset, bH as ServerComparisonOptions, bI as ServerComparisonResult, c as SizeValidatorOptions, d as SnapshotSanitizer, bJ as SnapshotSanitizers, bK as StdioMCPConfig, bN as StoredClientInfo, bQ as StoredEvalResultLoadOptions, bR as StoredEvalResultRef, bS as StoredEvalResultSaveOptions, bT as StoredEvalRunRef, bU as StoredOAuthState, bV as StoredServerMetadata, bW as StoredTokens, T as TextValidatorOptions, bY as TokenResult, g as ToolCallCountOptions, f as ToolCallExpectation, bZ as ToolMetadataOverride, b_ as ToolOverrideVariant, e as ToolPredicate, c0 as UsageMetrics, V as ValidationResult, c1 as VariantCandidateResult, c2 as VariantExperimentOptions, c3 as VariantExperimentReason, c4 as VariantExperimentResult, c5 as VariantExperimentRound, c6 as VariantImprovementProposal, c7 as VariantRecommendation } from '../index-aX_tCfTD.cjs';
+export { A as AuthType, K as BuiltInRubric, N as BuiltInSanitizer, Q as CLIConfig, W as CLIOAuthClientConfig, X as CLIOAuthResult, Y as CLIOutputFormat, Z as CaseComparisonResult, _ as ClientCredentialsConfig, $ as CompareEvalRunsOptions, a0 as ComparisonOutcome, a1 as ContentBlock, a2 as CreateMCPClientOptions, a3 as CustomJudgeExecutor, a4 as CustomJudgeResult, a9 as EvalCase, aa as EvalCaseComparison, ab as EvalCaseComparisonOutcome, ac as EvalCaseRequest, l as EvalCaseResult, ae as EvalContext, o as EvalDataset, ag as EvalExpectBlock, ah as EvalExpectationResult, ak as EvalMode, ao as EvalRunComparisonLabels, ap as EvalRunComparisonResult, aq as EvalRunMetadata, ar as EvalRunnerOptions, r as EvalRunnerResult, ay as ExpectationBreakdown, az as ExpectationResultMap, aA as ExpectationType, aB as ExperimentMetric, aJ as FieldRemovalSanitizer, aT as HostDiagnostics, a_ as HostType, a$ as HttpMCPConfig, b0 as IterationResult, G as Judge, F as JudgeConfig, b1 as JudgeExpectConfig, J as JudgeMatcherOptions, b2 as JudgeResult, b3 as JudgeValidatorConfig, L as LLMProvider, b4 as LLMToolCall, b6 as LoadDatasetOptions, b7 as MCPAuthConfig, b8 as MCPAuthFixtures, b9 as MCPClientCredentialsConfig, p as MCPConfig, bb as MCPConformanceCheck, bc as MCPConformanceOptions, bd as MCPConformanceRaw, be as MCPConformanceResult, bf as MCPConformanceResultData, bg as MCPEvalData, bh as MCPEvalHistoricalSummary, bi as MCPEvalReporterConfig, bj as MCPEvalRunData, M as MCPFixtureApi, bk as MCPFixtureOptions, bl as MCPHostCapabilities, a as MCPHostConfig, b as MCPHostSimulationResult, bm as MCPHostSimulator, bn as MCPOAuthConfig, bo as MCPServerCapabilitiesData, br as NormalizedToolResponse, O as OAuthSetupConfig, P as PatternValidatorOptions, bu as PlaywrightOAuthClientProviderConfig, bv as PredicateResult, bw as ProposeVariantsContext, bx as ProtectedResourceDiscoveryResult, by as ProtectedResourceMetadata, bz as ProviderKind, bA as RegexSanitizer, bB as ResultSource, R as RubricSpec, bE as SaveBaselineOptions, bF as SaveEvalRunComparisonOptions, bG as SaveServerComparisonOptions, bH as SchemaRegistry, S as SchemaValidatorOptions, bI as SerializedEvalDataset, bJ as ServerComparisonOptions, bK as ServerComparisonResult, c as SizeValidatorOptions, d as SnapshotSanitizer, bL as SnapshotSanitizers, bM as StdioMCPConfig, bP as StoredClientInfo, bS as StoredEvalResultLoadOptions, bT as StoredEvalResultRef, bU as StoredEvalResultSaveOptions, bV as StoredEvalRunRef, bW as StoredOAuthState, bX as StoredServerMetadata, bY as StoredTokens, T as TextValidatorOptions, b_ as TokenResult, g as ToolCallCountOptions, f as ToolCallExpectation, b$ as ToolMetadataOverride, c0 as ToolOverrideVariant, e as ToolPredicate, c2 as UsageMetrics, V as ValidationResult, c3 as VariantCandidateResult, c4 as VariantExperimentOptions, c5 as VariantExperimentReason, c6 as VariantExperimentResult, c7 as VariantExperimentRound, c8 as VariantImprovementProposal, c9 as VariantRecommendation } from '../index-CwFHKIoN.cjs';
 import 'zod';
 import '@playwright/test';
 import '@modelcontextprotocol/sdk/client/index.js';

package/dist/types/index.d.ts

@@ -1,4 +1,4 @@
-export { A as AuthType, K as BuiltInRubric, N as BuiltInSanitizer, Q as CLIConfig, W as CLIOAuthClientConfig, X as CLIOAuthResult, Y as CLIOutputFormat, Z as CaseComparisonResult, _ as ClientCredentialsConfig, $ as CompareEvalRunsOptions, a0 as ComparisonOutcome, a1 as ContentBlock, a2 as CreateMCPClientOptions, a3 as CustomJudgeExecutor, a4 as CustomJudgeResult, a9 as EvalCase, aa as EvalCaseComparison, ab as EvalCaseComparisonOutcome, ac as EvalCaseRequest, l as EvalCaseResult, ae as EvalContext, o as EvalDataset, ag as EvalExpectBlock, ah as EvalExpectationResult, ak as EvalMode, ao as EvalRunComparisonLabels, ap as EvalRunComparisonResult, aq as EvalRunMetadata, ar as EvalRunnerOptions, r as EvalRunnerResult, ay as ExpectationBreakdown, az as ExpectationResultMap, aA as ExpectationType, aB as ExperimentMetric, aJ as FieldRemovalSanitizer, aY as HostType, aZ as HttpMCPConfig, a_ as IterationResult, G as Judge, F as JudgeConfig, a$ as JudgeExpectConfig, J as JudgeMatcherOptions, b0 as JudgeResult, b1 as JudgeValidatorConfig, L as LLMProvider, b2 as LLMToolCall, b4 as LoadDatasetOptions, b5 as MCPAuthConfig, b6 as MCPAuthFixtures, b7 as MCPClientCredentialsConfig, p as MCPConfig, b9 as MCPConformanceCheck, ba as MCPConformanceOptions, bb as MCPConformanceRaw, bc as MCPConformanceResult, bd as MCPConformanceResultData, be as MCPEvalData, bf as MCPEvalHistoricalSummary, bg as MCPEvalReporterConfig, bh as MCPEvalRunData, M as MCPFixtureApi, bi as MCPFixtureOptions, bj as MCPHostCapabilities, a as MCPHostConfig, b as MCPHostSimulationResult, bk as MCPHostSimulator, bl as MCPOAuthConfig, bm as MCPServerCapabilitiesData, bp as NormalizedToolResponse, O as OAuthSetupConfig, P as PatternValidatorOptions, bs as PlaywrightOAuthClientProviderConfig, bt as PredicateResult, bu as ProposeVariantsContext, bv as ProtectedResourceDiscoveryResult, bw as ProtectedResourceMetadata, bx as ProviderKind, by as RegexSanitizer, bz as ResultSource, R as RubricSpec, bC as SaveBaselineOptions, bD as SaveEvalRunComparisonOptions, bE as SaveServerComparisonOptions, bF as SchemaRegistry, S as SchemaValidatorOptions, bG as SerializedEvalDataset, bH as ServerComparisonOptions, bI as ServerComparisonResult, c as SizeValidatorOptions, d as SnapshotSanitizer, bJ as SnapshotSanitizers, bK as StdioMCPConfig, bN as StoredClientInfo, bQ as StoredEvalResultLoadOptions, bR as StoredEvalResultRef, bS as StoredEvalResultSaveOptions, bT as StoredEvalRunRef, bU as StoredOAuthState, bV as StoredServerMetadata, bW as StoredTokens, T as TextValidatorOptions, bY as TokenResult, g as ToolCallCountOptions, f as ToolCallExpectation, bZ as ToolMetadataOverride, b_ as ToolOverrideVariant, e as ToolPredicate, c0 as UsageMetrics, V as ValidationResult, c1 as VariantCandidateResult, c2 as VariantExperimentOptions, c3 as VariantExperimentReason, c4 as VariantExperimentResult, c5 as VariantExperimentRound, c6 as VariantImprovementProposal, c7 as VariantRecommendation } from '../index-aX_tCfTD.js';
+export { A as AuthType, K as BuiltInRubric, N as BuiltInSanitizer, Q as CLIConfig, W as CLIOAuthClientConfig, X as CLIOAuthResult, Y as CLIOutputFormat, Z as CaseComparisonResult, _ as ClientCredentialsConfig, $ as CompareEvalRunsOptions, a0 as ComparisonOutcome, a1 as ContentBlock, a2 as CreateMCPClientOptions, a3 as CustomJudgeExecutor, a4 as CustomJudgeResult, a9 as EvalCase, aa as EvalCaseComparison, ab as EvalCaseComparisonOutcome, ac as EvalCaseRequest, l as EvalCaseResult, ae as EvalContext, o as EvalDataset, ag as EvalExpectBlock, ah as EvalExpectationResult, ak as EvalMode, ao as EvalRunComparisonLabels, ap as EvalRunComparisonResult, aq as EvalRunMetadata, ar as EvalRunnerOptions, r as EvalRunnerResult, ay as ExpectationBreakdown, az as ExpectationResultMap, aA as ExpectationType, aB as ExperimentMetric, aJ as FieldRemovalSanitizer, aT as HostDiagnostics, a_ as HostType, a$ as HttpMCPConfig, b0 as IterationResult, G as Judge, F as JudgeConfig, b1 as JudgeExpectConfig, J as JudgeMatcherOptions, b2 as JudgeResult, b3 as JudgeValidatorConfig, L as LLMProvider, b4 as LLMToolCall, b6 as LoadDatasetOptions, b7 as MCPAuthConfig, b8 as MCPAuthFixtures, b9 as MCPClientCredentialsConfig, p as MCPConfig, bb as MCPConformanceCheck, bc as MCPConformanceOptions, bd as MCPConformanceRaw, be as MCPConformanceResult, bf as MCPConformanceResultData, bg as MCPEvalData, bh as MCPEvalHistoricalSummary, bi as MCPEvalReporterConfig, bj as MCPEvalRunData, M as MCPFixtureApi, bk as MCPFixtureOptions, bl as MCPHostCapabilities, a as MCPHostConfig, b as MCPHostSimulationResult, bm as MCPHostSimulator, bn as MCPOAuthConfig, bo as MCPServerCapabilitiesData, br as NormalizedToolResponse, O as OAuthSetupConfig, P as PatternValidatorOptions, bu as PlaywrightOAuthClientProviderConfig, bv as PredicateResult, bw as ProposeVariantsContext, bx as ProtectedResourceDiscoveryResult, by as ProtectedResourceMetadata, bz as ProviderKind, bA as RegexSanitizer, bB as ResultSource, R as RubricSpec, bE as SaveBaselineOptions, bF as SaveEvalRunComparisonOptions, bG as SaveServerComparisonOptions, bH as SchemaRegistry, S as SchemaValidatorOptions, bI as SerializedEvalDataset, bJ as ServerComparisonOptions, bK as ServerComparisonResult, c as SizeValidatorOptions, d as SnapshotSanitizer, bL as SnapshotSanitizers, bM as StdioMCPConfig, bP as StoredClientInfo, bS as StoredEvalResultLoadOptions, bT as StoredEvalResultRef, bU as StoredEvalResultSaveOptions, bV as StoredEvalRunRef, bW as StoredOAuthState, bX as StoredServerMetadata, bY as StoredTokens, T as TextValidatorOptions, b_ as TokenResult, g as ToolCallCountOptions, f as ToolCallExpectation, b$ as ToolMetadataOverride, c0 as ToolOverrideVariant, e as ToolPredicate, c2 as UsageMetrics, V as ValidationResult, c3 as VariantCandidateResult, c4 as VariantExperimentOptions, c5 as VariantExperimentReason, c6 as VariantExperimentResult, c7 as VariantExperimentRound, c8 as VariantImprovementProposal, c9 as VariantRecommendation } from '../index-CwFHKIoN.js';
 import 'zod';
 import '@playwright/test';
 import '@modelcontextprotocol/sdk/client/index.js';

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@gleanwork/mcp-server-tester",
-  "version": "2.0.0-beta.0",
+  "version": "2.0.0-beta.1",
   "description": "Playwright-based testing and evaluation framework for MCP servers",
   "keywords": [
     "playwright",
@@ -28,6 +28,8 @@
       "import": "./dist/index.js",
       "require": "./dist/index.cjs"
     },
+    "./cowork-runtime": "./scripts/cowork_computer_use.py",
+    "./cowork-requirements": "./scripts/cowork-requirements.txt",
     "./types": {
       "types": "./dist/types/index.d.ts",
       "import": "./dist/types/index.js",
@@ -54,12 +56,14 @@
     "mcp-server-tester": "./dist/cli/index.js"
   },
   "files": [
-    "dist"
+    "dist",
+    "scripts/cowork_computer_use.py",
+    "scripts/cowork-requirements.txt"
   ],
   "scripts": {
     "build": "npm run build:ui && tsup && npm run build:copy-ui",
     "build:copy-ui": "cp -r src/reporters/ui-dist dist/reporters/",
-    "build:ui": "tsx src/reporters/build-ui.ts",
+    "build:ui": "node --import tsx src/reporters/build-ui.ts",
     "dev": "tsup --watch",
     "dev:ui": "tsx src/reporters/build-ui.ts --watch",
     "format": "prettier --write \"**/*.{ts,tsx,json,md}\"",

package/scripts/cowork-requirements.txt

@@ -0,0 +1,5 @@
+anthropic==1.6.0
+pyautogui==0.9.54
+mss==10.2.0
+Pillow==12.3.0
+pyobjc-framework-Quartz==12.2.2; sys_platform == "darwin"

package/scripts/cowork_computer_use.py

@@ -0,0 +1,409 @@
+#!/usr/bin/env python3
+"""Submit one Claude Cowork query through a bounded screenshot/action loop.
+
+The process stops immediately after the first submit key. MST owns native-session
+correlation, terminal validation, response extraction, and telemetry collection.
+"""
+
+from __future__ import annotations
+
+import argparse
+import asyncio
+import base64
+import io
+import json
+import math
+import os
+import re
+import subprocess
+import sys
+import time
+from typing import Any
+
+DISPLAY_WIDTH = 1280
+DISPLAY_HEIGHT = 800
+DEFAULT_MODEL = "claude-sonnet-4-6"
+DEFAULT_MAX_ACTIONS = 24
+
+
+def log(message: str) -> None:
+    print(f"[mst:cowork-cu] {message}", file=sys.stderr, flush=True)
+
+
+def screenshot() -> dict[str, Any]:
+    import mss
+    from PIL import Image
+
+    with mss.MSS() as capture:
+        monitor = capture.monitors[1]
+        raw = capture.grab(monitor)
+        image = Image.frombytes("RGB", raw.size, raw.bgra, "raw", "BGRX")
+        image = image.resize((DISPLAY_WIDTH, DISPLAY_HEIGHT), Image.Resampling.LANCZOS)
+        buffer = io.BytesIO()
+        image.save(buffer, format="PNG")
+        return {
+            "type": "image",
+            "source": {
+                "type": "base64",
+                "media_type": "image/png",
+                "data": base64.standard_b64encode(buffer.getvalue()).decode("ascii"),
+            },
+        }
+
+
+def screen_point(coordinate: list[int]) -> tuple[int, int]:
+    import pyautogui
+
+    width, height = pyautogui.size()  # Logical points, not Retina capture pixels.
+    return (
+        int(coordinate[0] * width / DISPLAY_WIDTH),
+        int(coordinate[1] * height / DISPLAY_HEIGHT),
+    )
+
+
+def type_query(text: str) -> None:
+    """Unicode keyboard events; line breaks must not submit the composer."""
+    import Quartz
+    import pyautogui
+
+    for line_index, line in enumerate(text.split("\n")):
+        if line_index:
+            pyautogui.hotkey("shift", "enter")
+        for start in range(0, len(line), 16):
+            chunk = line[start:start + 16]
+            units = len(chunk.encode("utf-16-le")) // 2
+            for down in (True, False):
+                event = Quartz.CGEventCreateKeyboardEvent(None, 0, down)
+                Quartz.CGEventKeyboardSetUnicodeString(event, units, chunk)
+                Quartz.CGEventPost(Quartz.kCGHIDEventTap, event)
+            time.sleep(0.01)
+
+
+def execute_action(action: dict[str, Any]) -> tuple[Any, bool]:
+    import pyautogui
+
+    name = action.get("action")
+    if name == "screenshot":
+        return screenshot(), False
+    if name in {
+        "left_click",
+        "right_click",
+        "middle_click",
+        "double_click",
+        "triple_click",
+    }:
+        x, y = screen_point(action["coordinate"])
+        button = {
+            "right_click": "right",
+            "middle_click": "middle",
+        }.get(name, "left")
+        clicks = {
+            "double_click": 2,
+            "triple_click": 3,
+        }.get(name, 1)
+        pyautogui.click(x, y, clicks=clicks, button=button)
+        return f"{name} completed", False
+    if name == "cursor_position":
+        x, y = pyautogui.position()
+        return {"x": int(x), "y": int(y)}, False
+    if name == "mouse_move":
+        x, y = screen_point(action["coordinate"])
+        pyautogui.moveTo(x, y)
+        return "mouse_move completed", False
+    if name == "left_click_drag":
+        start_x, start_y = screen_point(action["start_coordinate"])
+        end_x, end_y = screen_point(action["coordinate"])
+        pyautogui.moveTo(start_x, start_y)
+        pyautogui.dragTo(end_x, end_y, duration=0.4, button="left")
+        return "drag completed", False
+    if name == "type":
+        type_query(action["text"])
+        return "text entered", False
+    if name == "key":
+        text = str(action["text"])
+        aliases = {"return": "enter", "super": "command", "cmd": "command", "ctrl": "ctrl"}
+        keys = [aliases.get(part.strip().lower(), part.strip().lower()) for part in text.split("+")]
+        if len(keys) == 1:
+            pyautogui.press(keys[0])
+        else:
+            pyautogui.hotkey(*keys)
+        submitted = any(key in {"enter", "return"} for key in keys)
+        return f"key completed: {text}", submitted
+    if name == "scroll":
+        x, y = screen_point(action.get("coordinate", [DISPLAY_WIDTH // 2, DISPLAY_HEIGHT // 2]))
+        amount = int(action.get("scroll_amount", 3))
+        direction = action.get("scroll_direction", "down")
+        pyautogui.moveTo(x, y)
+        pyautogui.scroll(-amount if direction == "down" else amount)
+        return "scroll completed", False
+    if name == "wait":
+        time.sleep(min(float(action.get("duration", 1)), 5.0))
+        return "wait completed", False
+    raise RuntimeError(f"unsupported Computer Use action: {name}")
+
+
+TOKEN_FIELDS = (
+    "input_tokens", "output_tokens", "cache_creation_input_tokens", "cache_read_input_tokens",
+)
+
+
+class ComputerUseDriverError(RuntimeError):
+    def __init__(self, message: str, telemetry: dict[str, Any]):
+        super().__init__(message)
+        self.telemetry = telemetry
+
+
+class Telemetry:
+    """Only completed API responses and allowlisted scalar data; never request content."""
+
+    def __init__(self) -> None:
+        self.started = time.monotonic()
+        self.models: list[str] = []
+        self.responses = 0
+        self.usage: dict[str, int | float] = {}
+        self.coverage = {field: 0 for field in TOKEN_FIELDS}
+        self.actions = 0
+        self.attempted = 0
+        self.executed = 0
+        self.refused = 0
+
+    def observe(self, response: Any) -> None:
+        self.responses += 1
+        model = getattr(response, "model", None)
+        secrets = [value for key, value in os.environ.items()
+                   if value and re.search(r"token|key|secret|password|authorization", key, re.I)]
+        if (isinstance(model, str) and re.fullmatch(r"claude-[a-z0-9][a-z0-9.-]{0,99}", model)
+                and not any(secret in model for secret in secrets) and model not in self.models):
+            self.models.append(model)
+        usage = getattr(response, "usage", None)
+        for field in TOKEN_FIELDS:
+            value = getattr(usage, field, None)
+            if (type(value) in (int, float) and 0 <= value <= 9007199254740991
+                    and math.isfinite(value) and value == int(value)):
+                total = self.usage.get(field, 0) + value
+                if total <= 9007199254740991:
+                    self.usage[field] = total
+                    self.coverage[field] += 1
+
+    def snapshot(self, accounting: str) -> dict[str, Any]:
+        return {
+            "accounting": accounting,
+            "response_models": list(self.models),
+            "planner_response_count": self.responses,
+            "usage": dict(self.usage),
+            "usage_observation_counts": dict(self.coverage),
+            "duration_ms": max(0, (time.monotonic() - self.started) * 1000),
+            "action_count": self.actions,
+            "attempted_action_count": self.attempted,
+            "executed_action_count": self.executed,
+            "refused_action_count": self.refused,
+            # Anthropic Messages usage does not supply authoritative dollar cost.
+            "cost": {"status": "unavailable"},
+        }
+
+
+async def run(query: str, max_actions: int, mode: str) -> dict[str, Any]:
+    telemetry = Telemetry()
+    try:
+        result = await run_driver(query, max_actions, mode, telemetry)
+        result["telemetry"] = telemetry.snapshot("complete")
+        return result
+    except Exception as error:
+        raise ComputerUseDriverError(str(error), telemetry.snapshot("partial")) from error
+
+
+async def run_driver(query: str, max_actions: int, mode: str, telemetry: Telemetry) -> dict[str, Any]:
+    try:
+        import anthropic
+    except ImportError as error:
+        raise RuntimeError("Install the Computer Use dependencies: pip install anthropic pyautogui mss Pillow") from error
+
+    api_key = os.environ.get("ANTHROPIC_API_KEY")
+    if not api_key:
+        raise RuntimeError("ANTHROPIC_API_KEY is required for the Computer Use submission driver")
+
+    client = anthropic.Anthropic(api_key=api_key)
+    model = os.environ.get("MST_COWORK_CUA_MODEL", DEFAULT_MODEL)
+    log(f"starting driver with model={model}, max_actions={max_actions}")
+    subprocess.run(["open", "-a", "Claude"], check=False, capture_output=True)
+    log("requested Claude Desktop launch/focus")
+    time.sleep(float(os.environ.get("MST_COWORK_CUA_START_DELAY", "3")))
+
+    tools = [{
+        "type": "computer_20251124",
+        "name": "computer",
+        "display_width_px": DISPLAY_WIDTH,
+        "display_height_px": DISPLAY_HEIGHT,
+    }]
+    if mode == "hitl":
+        messages: list[dict[str, Any]] = [{
+            "role": "user",
+            "content": (
+                "Inspect the currently visible Claude Desktop Cowork task. If a human-in-the-loop "
+                "approval or choice prompt is blocking progress, choose the first visible option. "
+                "Do not submit a new query, do not change account settings, and do not approve a "
+                "prompt unless it is the visible task's first option. If no such prompt is visible, "
+                "stop without taking an action.\n\nTask to locate (do not type or submit this text):\n" + query
+            ),
+        }]
+        system = (
+            "You are a bounded HITL resolver. Use screenshots and Computer Use actions only. "
+            "Choose the first visible option when a task approval/choice prompt is present. "
+            "Never submit a query or change account settings. If no HITL prompt is visible, stop."
+        )
+    else:
+        tools.append({
+            "name": "fill_query",
+            "description": (
+                "Insert the unchanged original evaluation query into the focused "
+                "empty Cowork task composer. First locate and focus that composer using a screenshot. "
+                "This tool takes no text: the harness supplies the exact text. It may run only once. "
+                "After it succeeds, use computer key Enter to submit, never click Send."
+            ),
+            "input_schema": {"type": "object", "properties": {}, "additionalProperties": False},
+        })
+        messages = [{
+            "role": "user",
+            "content": (
+                "Control Claude Desktop. Open Cowork, create a fresh task, and focus its empty "
+                "prompt composer. Call fill_query with no arguments. It inserts the original "
+                "query for you. Do not reconstruct or type the query yourself. Then press "
+                "unmodified Enter once to submit. Do not wait for or read the answer."
+            ),
+        }]
+        system = (
+            "You are a bounded desktop submission operator. Use screenshots and Computer Use "
+            "actions to open/focus Claude, select Cowork, create a fresh task, and focus its composer. "
+            "Use fill_query, not computer type, to insert the query. After fill_query succeeds, "
+            "submit with unmodified Enter, never by clicking Send. Never submit twice. Stop after "
+            "submission. Do not approve permissions or change account settings."
+        )
+
+    # Ground the first action in a current screenshot, not an assumed layout.
+    messages[0]["content"] = [{"type": "text", "text": messages[0]["content"]}, screenshot()]
+    hitl_action_taken = False
+    actions_executed = 0
+    typed_query = False
+    for action_number in range(1, max_actions + 1):
+        log(f"requesting Computer Use plan {action_number}/{max_actions}")
+        response = client.beta.messages.create(
+            model=model,
+            max_tokens=1024,
+            system=system,
+            tools=tools,
+            messages=messages,
+            betas=["computer-use-2025-11-24"],
+        )
+        telemetry.observe(response)
+        messages.append({"role": "assistant", "content": response.content})
+        tool_results: list[dict[str, Any]] = []
+        for block in response.content:
+            if getattr(block, "type", None) != "tool_use":
+                continue
+            if actions_executed >= max_actions:
+                raise RuntimeError("Computer Use action budget exhausted; no further actions executed")
+            actions_executed += 1
+            telemetry.actions += 1
+            tool_name = getattr(block, "name", "computer")
+            action = block.input
+            action_name = action.get('action', 'unknown')
+            refusal = None
+            if mode == "hitl" and (tool_name != "computer" or action_name not in {"screenshot", "wait", "mouse_move", "cursor_position", "left_click", "scroll"}):
+                telemetry.refused += 1
+                raise RuntimeError("HITL cannot type, press keys, drag, or submit tasks")
+            if mode == "submit":
+                if tool_name == "fill_query":
+                    if typed_query:
+                        refusal = "The query is already entered. Do not fill again. Use unmodified Enter to submit."
+                    elif block.input != {}:
+                        refusal = "fill_query takes no arguments; the harness owns the original text."
+                    else:
+                        action = {"action": "type", "text": query}
+                        action_name = "type"
+                elif tool_name != "computer":
+                    refusal = "Unknown tool. Use computer for navigation and fill_query for text entry."
+                elif action_name == "type":
+                    refusal = "Do not reconstruct the query. Focus the empty composer, then call fill_query with no arguments."
+                elif action_name == "key" and any(k.strip().lower() in {"enter", "return"} for k in str(action.get("text", "")).split("+")):
+                    if not typed_query or str(action.get("text", "")).lower() not in {"enter", "return"}:
+                        refusal = "Only an unmodified Enter after fill_query succeeds may submit."
+                elif typed_query and action_name not in {"screenshot", "wait", "cursor_position"}:
+                    refusal = "The query is entered. Only screenshots or the single Enter submission are allowed."
+            if refusal:
+                telemetry.refused += 1
+                # A rejected proposal has no desktop side effects. Let the planner correct it
+                # within the same action budget; never retry a failed physical text entry.
+                tool_results.append({"type": "tool_result", "tool_use_id": block.id, "content": refusal, "is_error": True})
+                continue
+            log(f"executing action {actions_executed}: {action_name}")
+            if mode == "hitl" and action_name not in {
+                "screenshot",
+                "wait",
+                "mouse_move",
+                "cursor_position",
+            }:
+                hitl_action_taken = True
+            telemetry.attempted += 1
+            result, submitted = execute_action(action)
+            telemetry.executed += 1
+            if tool_name == "fill_query":
+                typed_query = True
+            if submitted and mode != "hitl":
+                log("submission boundary reached; stopping immediately")
+                return {
+                    "status": "submitted",
+                    "action_count": actions_executed,
+                    "model": model,
+                    "submission_action": {"action": "key", "text": "enter"},
+                }
+            if isinstance(result, dict) and result.get("type") == "image":
+                tool_content: Any = [result]
+            else:
+                tool_content = str(result)
+            tool_results.append({
+                "type": "tool_result",
+                "tool_use_id": block.id,
+                "content": tool_content,
+            })
+        if not tool_results:
+            if mode == "hitl":
+                log("no HITL action was needed")
+                return {"status": "hitl_checked", "action_count": actions_executed, "model": model}
+            raise RuntimeError("Computer Use planner stopped before submitting the Cowork query")
+        messages.append({"role": "user", "content": tool_results})
+
+    if mode == "hitl" and not hitl_action_taken:
+        log("no visible HITL prompt found within the bounded check")
+        return {
+            "status": "hitl_checked",
+            "action_count": actions_executed,
+            "prompt_found": False,
+            "model": model,
+        }
+    if mode == "hitl":
+        raise RuntimeError(
+            f"Computer Use HITL check exceeded {max_actions} actions after attempting a visible prompt"
+        )
+    raise RuntimeError(f"Computer Use submission exceeded {max_actions} actions without submitting")
+
+
+def main() -> int:
+    parser = argparse.ArgumentParser()
+    parser.add_argument("query")
+    parser.add_argument("--max-actions", type=int, default=DEFAULT_MAX_ACTIONS)
+    parser.add_argument("--mode", choices=["submit", "hitl"], default="submit")
+    args = parser.parse_args()
+    try:
+        print(json.dumps(asyncio.run(run(args.query, args.max_actions, args.mode))), flush=True)
+        return 0
+    except Exception as error:
+        log(f"driver failed: {error}")
+        result = {"status": "failed", "error": str(error)}
+        if isinstance(error, ComputerUseDriverError):
+            result["telemetry"] = error.telemetry
+        print(json.dumps(result), flush=True)
+        return 1
+
+
+if __name__ == "__main__":
+    sys.exit(main())