Scale computer-use 2.0 with open-source drivers, cross-OS sandboxes, and benchmarks for training, evaluation, and data generation.
Rust agentai-agentapplecomputer-usecomputer-use-agentcontainerizationcuadesktop-automationhacktoberfestlumemacosmanusoperatorswiftvirtualizationvirtualization-frameworkwindowswindows-sandbox
chore: remove internal scratchpad / handoff docs from repo root + libs (#1603)
Francesco Bonaccif83c43c7477c
diff --git a/WINDOWS_SESSION0_STATUS.md b/WINDOWS_SESSION0_STATUS.mddeleted file mode 100644index ff6872f84a..0000000000--- a/WINDOWS_SESSION0_STATUS.md+++ /dev/null@@ -1,158 +0,0 @@-# Windows Session-0 Stabilization — Status Report--## Good morning — TL;DR--While you slept I stabilized Windows cua-driver-rs end-to-end on the VM-you provisioned. **All 11 parity examples PASS** (was 6/11 before).-**Two real bugs fixed** (`doctor` segfault, `launch_app` hang), **one-focus-steal hardening** (UWP foreground-restore), **one quality filter**-(opaque system packages out of `list_apps`), **one startup banner**-(serve in Session 0 warns about GUI-tool limitations).--Branch `stab/windows-session0-fixes` is 7 commits ahead of-`origin/main`, all local — no PR opened per your ask. To turn into a-PR when ready:--```bash-git -C <repo-root> push -u origin stab/windows-session0-fixes-gh pr create --base main --head stab/windows-session0-fixes \- --title "fix(windows): Session-0 hardening + UWP focus-restore" \- --body-file WINDOWS_SESSION0_STATUS.md-```--**Still wants your eyes:** the UWP focus-restore needs visual-confirmation in Session 1+ (RDP in, run the snippet under "Visual-confirmation" below). I made it stronger than my first pass (now uses-the `keybd_event` workaround to lift Windows' SetForegroundWindow-restriction) but I can't visually verify from Mac side.-------**Branch:** `stab/windows-session0-fixes` (local-only, no PR yet per your ask)-**Base:** `origin/main` @ 69a9dbd5 (post-#1547)-**VM:** `<vm-name>` (<vm-host>) — Rust toolchain + VS Build Tools 2022 installed-**Dev loop:** edit on Mac → `scp` changed files to `~/cua/...` on VM → `cargo build --release -p cua-driver` (~20s incremental, ~5min cold) → run. Full reference: `libs/cua-driver-rs/DEV_LOOP_WINDOWS_VM.md`.--## TL;DR--**Two real bugs found + fixed**, both Session-0 (services / SSH-launched) edge cases:--1. **`doctor` crashed with `0xC0000005 ACCESS_VIOLATION`** — COM lifecycle bug in `ui_automation_available`: `IUIAutomation` dropped AFTER `CoUninitialize`, so `IUnknown::Release` ran against a torn-down apartment. Fix: flatten probe result before tearing down COM. ✅-2. **`launch_app` hung indefinitely in Session 0** — two paths: `resolve_aumid_by_name` walked `shell:AppsFolder` via the interactive shell broker (doesn't exist in Session 0 → hang). `launch_uwp` itself called `IApplicationActivationManager::ActivateApplication` which also needs interactive AppX runtime (same hang). Both now short-circuit on Session 0 with `current_session_id() == Some(0)` — `resolve_aumid_by_name` returns None (falls through to ShellExecuteEx PATH lookup, which works), `launch_uwp` fails fast with a descriptive error. ✅--**Plus**: foreground-restore after UWP activation (now with `keybd_event` workaround for Windows' foreground-lock restriction), three parity-example fixes (Session-0 awareness + #1545 contract), opaque-system-family UWP filter for list_apps, serve startup Session-0 banner, and `overlay_dump` example build fix.--## Bugs fixed (commits on `stab/windows-session0-fixes`)--| Commit | Subject |-|---|---|-| `a9d9c027` | fix(windows): doctor COM lifecycle + Session 0 hardening for launch paths |-| `9e91f683` | test(windows): parity examples Session-0 aware + #1545 contract |-| `7251e73b` | feat(serve): Session-0 warning banner on Windows daemon startup |-| `2eb3c4d3` | docs(parity): update list_apps + launch_app Windows status with live evidence |-| `ac70b94d` | feat(installed_apps): filter opaque-system-family UWP packages from list_apps |-| `6f580201` | fix(installed_apps): drop Package.IsFramework() probe — hangs in Session 0 |-| `d5b5821c` | fix(launch_uwp): claim 'last input' before SetForegroundWindow restore |--### Surprise bug found during this work (now fixed in `6f580201`)--`Package.IsFramework()` WinRT call **hangs in Session 0** (0 CPU, no progress, no error). I had added it as a "filter out framework packages from list_apps" optimization in `ac70b94d`, which silently broke list_apps for ~3 hours of debug-loop. Worth documenting: avoid WinRT property getters that may touch the AppX activation runtime when running in services context. The opaque-family-name filter still catches most framework noise.--## Validated on the VM (Session 0, SSH from Mac)--| Tool | Status | Notes |-|---|---|---|-| `cua-driver --version` | ✅ 0.2.2 | |-| `cua-driver list-tools` | ✅ 30 tools | matches Swift surface |-| `cua-driver doctor` | ✅ exit 0 | clean run, all probes report. Session-0 warning surfaced correctly. |-| `cua-driver call get_screen_size` | ✅ 1024x768 @ 1.0 | |-| `cua-driver call get_cursor_position` | ✅ 0,0 | |-| `cua-driver call check_permissions` | ✅ elevated, post_message, uia all true | |-| `cua-driver call get_config` | ✅ returns full config | |-| `cua-driver call list_windows` | ✅ empty `[]` (expected — no desktop) | |-| `cua-driver call list_apps` | ✅ 3.7s cold, returns ~150 apps (running + installed) | |-| `cua-driver call get_accessibility_tree` | ✅ returns process list | |-| `launch_app {"name":"notepad"}` | ✅ ShellExecuteEx, pid returned | SW_SHOWNOACTIVATE (no focus steal) |-| `launch_app {"path":"C:\\..\\calc.exe"}` | ✅ ShellExecuteEx, pid returned | |-| `launch_app {"aumid":"...Calculator!App"}` | ✅ fails fast with Session-0 message (no hang) | needs interactive session to actually launch |--### Parity tests-*(`cargo run --example *_parity` against running daemon)*--Pre-fix: 6 PASS / 5 FAIL — Session-0 expected-failures + stale post-#1545 assertions.-Post-fix: see `b8jvtpz9m.output` (in flight as of this writing).--## Visual confirmation needs RDP--**Background-launch focus invariant for UWP apps** is now stronger than the initial best-effort:--- Win32 path (ShellExecuteEx + SW_SHOWNOACTIVATE): ✅ proven by code — the API contract says SW_SHOWNOACTIVATE never activates.-- UWP path (`IApplicationActivationManager::ActivateApplication(AO_NONE)`): the AppX runtime brings the app to foreground (= Start-Menu-click semantics). Now does:- 1. Snapshot prior foreground via `GetForegroundWindow` before the call- 2. Inject a VK_NONAME (0xFC) `keybd_event` press+release immediately after to claim "owner of last input" — that lifts Windows' foreground-lock restriction- 3. Call `SetForegroundWindow(prior)` in a 3 × 50ms retry loop, logging each attempt at debug-- VK_NONAME (0xFC) doesn't map to any UI action, so no application sees a keystroke. The Session-0 short-circuit higher up bypasses this entire path so it only executes on interactive logons where focus actually matters.--To visually validate when you wake:--```powershell-# RDP into <vm-host> as <user>, then in a PowerShell on the desktop:-$exe = "C:\Users\<user>\AppData\Local\Programs\trycua\cua-driver-rs\bin\cua-driver.exe"-# Or use the dev build:-# $exe = "$env:USERPROFILE\cua\libs\cua-driver-rs\target\release\cua-driver.exe"--# Open something focusable first (e.g. Notepad++ via Start menu) and start typing into it.-# Then from a *different* PowerShell window, fire:-'{"name":"calculator"}' | & $exe call launch_app-# Calculator should launch into the taskbar/background; your Notepad++ focus-# should be preserved (no visible focus steal). If focus DOES move to-# Calculator, the SetForegroundWindow restriction blocked our restore —-# we have a fallback path (push to HWND_BOTTOM) on standby.-```--## Other improvements made--1. **`eprintln!` traces in list_apps / installed_apps replaced with `tracing::debug!`** — silent at default log level. Enable with `RUST_LOG=cua_driver=debug,installed_apps=debug,list_apps=debug` when diagnosing.-2. **Parity tests `list_apps_parity` / `list_windows_parity` / `launch_app_parity`** brought up to current contract (#1545's unified shape, Session-0 awareness, UWP fail-fast acceptance).-3. **`overlay_dump.rs` example** — fixed `PrintWindow` + `PRINT_WINDOW_FLAGS` imports (moved to `Win32::Storage::Xps` in windows-rs 0.58).--## Known issues NOT fixed (documented for next session)--1. **UWP display names still show family-name fallback for some packages** — system Components and packages with unresolved `ms-resource:` tokens AND no `Properties/DisplayName` in their AppxManifest.xml fall back to their FamilyName (e.g. `1527c705-839a-4832-9118-54d4Bd6a0c89_cw5n1h2txyewy`). #1547 closed the common case; the long tail is system packages users mostly don't care about. Possible follow-ups: (a) filter out packages whose family_name matches GUID pattern AND display name is empty/equal-to-family — they're usually unactivatable system components; (b) try `Package.PublisherDisplayName` as a third fallback.--2. **`SetForegroundWindow` restoration after UWP activation is best-effort** — Windows restricts foreground transfers and may silently ignore the call. Stronger options to consider: (a) `SetWindowPos(spawned_hwnd, HWND_BOTTOM, ..., SWP_NOACTIVATE)` to push the UWP window behind everything (more aggressive than macOS but guaranteed); (b) the `keybd_event` trick to give cua-driver "last input" status so SetForegroundWindow's restriction lifts (visible side effect of pressing Alt — usually OK).--3. **`overlay_dump` builds — but won't run usefully in Session 0** since the overlay needs a desktop. Build-fix only.--4. **Parity tests requiring foreground windows** — `list_windows`, `screenshot`, `get_window_state`, `click`, `type_text`, `hotkey`, `press_key`, `drag`, `scroll` etc. all need interactive desktop. They're tested by being called against the running daemon, but their assertions (e.g. "expected at least one visible window") are Session-0-aware where I touched them; the rest will naturally need Session 1+ to exercise meaningfully.--## How to take the local fixes forward--```bash-# View the local-only commits-git -C <repo-root> log --oneline origin/main..stab/windows-session0-fixes--# Bring them up onto a fresh branch for a PR-git -C <repo-root> checkout -b fix/windows-session0-git -C <repo-root> merge --ff-only stab/windows-session0-fixes--# Then `gh pr create` as usual-```--The local branch is rebased on current `origin/main` (post-#1547). No PR opened per your instruction.--## Other PRs of yours seen earlier in the session--- **#1546** — closed as superseded by #1547 (older diff of same domain)-- **#1547** — merged (UWP ms-resource: display name normalization)-- **#1523** — *parity(launch_app) macOS port of #1492* — has CR fixes pushed but **needs a port-onto-current-main**, not a simple rebase. Main moved significantly: `platform-macos/src/apps.rs` deleted and reorganized into `apps/mod.rs` + `apps/nsworkspace.rs`, `AppInfo` gained 3 new fields. Estimated ~30 min focused work to rebuild. Parked pending your call.--## Timeline--- Session started ~Sun May 18 00:00 (local)-- Toolchain install on VM: 23:08 → 23:34-- Dev loop online + first cargo build: 23:34 → 23:48 (~11min cold build)-- doctor + launch_app bugs identified + fixed: ~00:00 → ~02:00-- This report: ~02:35-- Resuming at next user-online checkpointdiff --git a/docs/scripts/HANDOFF-modal-s3-sync.md b/docs/scripts/HANDOFF-modal-s3-sync.mddeleted file mode 100644index b7ab10b12c..0000000000--- a/docs/scripts/HANDOFF-modal-s3-sync.md+++ /dev/null@@ -1,130 +0,0 @@-# Handoff: Deploy & Test Modal S3 Sync--## Context--PR #1430 (`feat/modal-s3-sync` branch in `trycua/cua`) adds a `sync_to_s3()` function to `docs/scripts/modal_app.py`. This function uploads generated documentation and code search databases from Modal volumes to S3, bridging the gap between Modal (which generates the DBs on a schedule) and a K8s-hosted MCP server (which serves them).--### Architecture--```-Modal (generates DBs daily) → sync_to_s3() → S3 → K8s CronJob (pulls from S3) → PVC → MCP Server-```--### What was built--1. **`sync_to_s3()` function** in `docs/scripts/modal_app.py`:- - Uploads SQLite + LanceDB files from two Modal volumes to `s3://trycua-docs-mcp-data/`- - Uses **Modal OIDC federation** (no static AWS keys) — Modal auto-injects `MODAL_IDENTITY_TOKEN`, which is exchanged via STS `AssumeRoleWithWebIdentity` for temporary credentials- - IAM role: `arn:aws:iam::296062593712:role/modal-docs-mcp-write-role` (already provisioned via Terraform in `trycua/cloud` repo)--2. **Integration into scheduled pipelines**:- - `scheduled_crawl()` (6 AM UTC) — calls `sync_to_s3()` after regenerating docs databases- - `scheduled_code_index()` (5 AM UTC) — calls `sync_to_s3()` after aggregating code databases--3. **S3 bucket** (`trycua-docs-mcp-data`) already exists with versioning, encryption, and cross-region replication. Currently has `docs_db/` data from a prior manual upload.--### What needs testing--The code has NOT been deployed yet. The previous environment could not reach PyPI (SSL handshake failures), so `modal` CLI was never installed.-------## Your Task--### 1. Install Modal CLI--```bash-pip install modal-```--### 2. Authenticate Modal--The Modal token should be available in `~/.modal/config.toml`. If not, ask the user to provide it or run:--```bash-modal token set --token-id <ask-user> --token-secret <ask-user>-```--### 3. Checkout the branch--```bash-cd /path/to/cua # the trycua/cua repo-git fetch origin-git checkout feat/modal-s3-sync-```--### 4. Deploy the Modal app--```bash-modal deploy docs/scripts/modal_app.py-```--This should succeed. The app name is `cua-docs-mcp` — it will register all functions including the new `sync_to_s3`.--### 5. Test `sync_to_s3` manually--```bash-modal run docs/scripts/modal_app.py::sync_to_s3-```--**Expected output:**-- "Syncing databases to s3://trycua-docs-mcp-data/ ..."-- Upload lines for docs SQLite, docs LanceDB directory, code SQLite, and code LanceDB directory-- "S3 sync complete: N files uploaded to s3://trycua-docs-mcp-data/"--**If OIDC auth fails** with an STS error about the role trust policy, check:-- The Modal app name must be `cua-docs-mcp` (defined at top of `modal_app.py`)-- The role trust policy restricts to `app_name:cua-docs-mcp` — if the app name changed, the OIDC claim won't match-- Verify the role exists: `aws iam get-role --role-name modal-docs-mcp-write-role`--**If volumes are empty** (no databases to upload), you'll need to generate them first:-```bash-# Generate docs databases (takes ~30 min)-modal run docs/scripts/modal_app.py --skip-code--# Generate code databases (takes ~1-2 hours)-modal run docs/scripts/modal_app.py --skip-docs-```--### 6. Verify S3 contents--```bash-aws s3 ls s3://trycua-docs-mcp-data/ --recursive --region us-west-2-```--**Expected structure:**-```-docs_db/docs.sqlite-docs_db/docs.lance/...-code_db/code_index.sqlite-code_db/code_index.lancedb/...-```--### 7. Report results--Let the user know:-- Whether `modal deploy` succeeded-- Whether `sync_to_s3` ran and how many files were uploaded-- The S3 listing output-- Any errors encountered-------## Key Files--| File | Purpose |-|------|---------|-| `docs/scripts/modal_app.py` | Modal app — crawling, indexing, S3 sync, MCP server |-| (cloud repo) `terraform/aws/docs-mcp-storage/main.tf` | S3 bucket, OIDC role, readonly IAM user |-| (cloud repo) `terraform/aws/oidc/modal-oidc.tf` | Modal OIDC provider |--## AWS Details--| Resource | Value |-|----------|-------|-| S3 Bucket | `trycua-docs-mcp-data` |-| Region | `us-west-2` |-| Write Role ARN | `arn:aws:iam::296062593712:role/modal-docs-mcp-write-role` |-| Readonly Creds Secret | `docs-mcp/readonly-credentials` (in AWS Secrets Manager) |-| Modal App Name | `cua-docs-mcp` |-| Modal Workspace ID | `ac-3LfmQEOVnLXl4ns0YBxNuA` |diff --git a/libs/cua-bench/plan.md b/libs/cua-bench/plan.mddeleted file mode 100644index b287106931..0000000000--- a/libs/cua-bench/plan.md+++ /dev/null@@ -1,531 +0,0 @@-# Plan: Tinker RL Training Integration with cua-bench--## Overview--This plan describes how to integrate [Tinker](https://tinker-docs.thinkingmachines.ai/) as an RL training backend for the OpenCUA model using `cua-bench` as the environment. The result is a closed training loop: `cua-bench` generates rollout trajectories and rewards → Tinker updates model weights via PPO → updated weights reload into vLLM → next rollout round.-------## 1. How Traces Are Collected Today--After running `cb run tasks --agent opencua`, each task execution produces a **HuggingFace Dataset** stored on disk via `cua_bench/tracing.py`.--### Trace event schema--| Field | Type | Description |-|---|---|---|-| `event_name` | `str` | One of `"reset"`, `"agent_step"`, `"solve"`, `"evaluate"` |-| `data_json` | `str` | JSON payload specific to event type |-| `data_images` | `Sequence[Image]` | Screenshots at that moment |-| `trajectory_id` | `str` | UUID identifying the full episode |-| `timestamp` | `str` | ISO-8601 UTC timestamp |--### Events recorded per task--| Event | When | Key `data_json` fields |-|---|---|---|-| `reset` | After env setup completes | `{task, task_index, elapsed}` |-| `agent_step` | After every agent action | `{step, agent, model, usage, output}` |-| `solve` | After oracle solution (non-eval runs) | `{completed: bool}` |-| `evaluate` | After evaluation function | `{result: <reward_value>}` |--The `agent_step` `output` field mirrors the raw message list from the ComputerAgent SDK, including typed actions (`click`, `type_text`, etc.) and the final `"DONE"` message.--### Trace storage paths--```-~/.local/share/cua-bench/runs/<run_id>/-├── task_0_trace/ # HF Dataset (arrow files)-│ ├── dataset_info.json-│ ├── state.json-│ └── data/-├── task_0_agent_logs/-│ └── trajectories/ # ComputerAgent internal trajectories-└── imgs/ # Optional PNG dumps (--save-pngs flag)-```--### Loading traces programmatically--```python-from datasets import load_from_disk--ds = load_from_disk("~/.local/share/cua-bench/runs/<run_id>/task_0_trace/")--# Extract reward from evaluate event-reward_row = next(r for r in ds if r["event_name"] == "evaluate")-reward = json.loads(reward_row["data_json"])["result"]--# Extract agent steps in order-steps = [r for r in ds if r["event_name"] == "agent_step"]-for step in steps:- data = json.loads(step["data_json"])- screenshot = step["data_images"][0] # PIL.Image-```-------## 2. Tinker API Overview--Tinker is a cloud-hosted RL/SL fine-tuning platform for LLMs. It abstracts distributed GPU training while exposing a simple Python API.--### Core client hierarchy--```-ServiceClient (entry point)-├── create_lora_training_client() → TrainingClient-│ ├── forward_backward(data, loss_fn) → Future-│ ├── optim_step(AdamParams) → Future-│ ├── save_state(name) → Future[str] (tinker:// path)-│ └── save_weights_and_get_sampling_client() → SamplingClient-├── create_sampling_client(checkpoint) → SamplingClient-│ ├── sample(prompt, params) → List[str]-│ └── compute_logprobs(prompt, target) → List[float]-└── create_rest_client() → RestClient- ├── list_checkpoints(run_id)- └── get_checkpoint_archive_url_from_tinker_path(path)-```--### Supported RL loss functions--| Loss | Use case |-|---|---|-| `"importance_sampling"` | Off-policy PPO without clipping |-| `"ppo"` | Clipped PPO (standard RLHF) |-| `"cispo"` | Clipped importance sampling coefficients |-| `"dro"` | Direct reward optimization with quadratic penalty |-| Custom | Via `forward_backward_custom()` with arbitrary differentiable loss |--### Checkpoint persistence--| Method | Saves | When to use |-|---|---|---|-| `save_weights_for_sampler(name)` | Weights only | Quick eval between epochs |-| `save_state(name)` | Weights + optimizer state | Full resume / multi-phase SL→RL |-| `save_weights_and_get_sampling_client()` | Weights → new SamplingClient | End of training step, for next rollout |--Checkpoint paths follow the scheme `tinker://<model_id>/<name>`. Default TTL is 7 days; use `RestClient` to extend.-------## 3. Proposed Architecture--```-┌─────────────────────────────────────────────────────────────┐-│ RL Training Loop │-│ │-│ ┌──────────────┐ rollouts ┌───────────────────────┐ │-│ │ cua-bench │ ─────────────▶│ Trace Collector │ │-│ │ (env) │ │ (HF Datasets) │ │-│ └──────┬───────┘ └──────────┬────────────┘ │-│ │ │ │-│ ┌──────▼───────┐ ┌──────────▼────────────┐ │-│ │ vLLM │ ◀─ weights ── │ Tinker │ │-│ │ (OpenCUA) │ │ TrainingClient │ │-│ └──────────────┘ │ (PPO / IS loss) │ │-│ └───────────────────────┘ │-└─────────────────────────────────────────────────────────────┘-```--### Component roles--| Component | Role |-|---|---|-| `cua-bench` | Rollout environment; generates trajectories + reward signals |-| `cua_bench/tracing.py` | Serializes trajectories to HF Datasets on disk |-| `Trace Collector` | Reads traces, extracts `(prompt, action, reward)` tuples |-| `Datum Builder` | Converts tuples to Tinker `Datum` objects (token sequences + masks) |-| `Tinker TrainingClient` | Runs PPO gradient updates on remote GPU cluster |-| `Tinker SamplingClient` | Not used for generation (vLLM does that), but used for `compute_logprobs` to get reference policy log-probs for IS ratio |-| `vLLM` | Serves the current policy for cua-bench rollouts |-| `Weight Sync` | Downloads updated LoRA weights from Tinker checkpoint → hot-reloads into vLLM |-------## 4. Implementation Plan--### Phase 1: Trace extraction pipeline--**New file**: `cua_bench/rl/trace_reader.py`--```python-"""-Read completed cua-bench run traces and convert them to RL experience tuples.-"""-import json-from pathlib import Path-from dataclasses import dataclass-from typing import List, Optional-from datasets import load_from_disk--@dataclass-class RLStep:- step: int- prompt_images: list # screenshots before action- action_text: str # raw action output from model- reward: float # 0.0 except at terminal step--@dataclass-class Episode:- trajectory_id: str- task_description: str- steps: List[RLStep]- terminal_reward: float # from "evaluate" event--def load_episode_from_trace(trace_dir: str) -> Episode:- ds = load_from_disk(trace_dir)- rows_by_event = {}- for row in ds:- rows_by_event.setdefault(row["event_name"], []).append(row)-- reward = float(json.loads(rows_by_event["evaluate"][0]["data_json"])["result"])- trajectory_id = ds[0]["trajectory_id"]-- steps = []- for row in rows_by_event.get("agent_step", []):- data = json.loads(row["data_json"])- steps.append(RLStep(- step=data["step"],- prompt_images=row["data_images"],- action_text=str(data["output"]),- reward=0.0,- ))-- # Assign terminal reward to last step only (sparse reward)- if steps:- steps[-1].reward = reward-- return Episode(trajectory_id, task_description="", steps=steps, terminal_reward=reward)--def load_run_episodes(run_dir: str) -> List[Episode]:- """Load all task traces from a cua-bench run directory."""- run_path = Path(run_dir)- episodes = []- for trace_dir in sorted(run_path.glob("task_*_trace")):- episodes.append(load_episode_from_trace(str(trace_dir)))- return episodes-```--### Phase 2: Datum construction--**New file**: `cua_bench/rl/datum_builder.py`--The Tinker `Datum` format requires tokenized inputs with explicit weight masks. For a vision-language model like OpenCUA (based on Qwen-VL), each step's datum looks like:--```-[system_tokens] [image_tokens] [user_instruction_tokens] [action_tokens]- ^^^^^^^^^^^^^^^^- loss computed here-```--Key decisions:-- **Input sequence**: System prompt + screenshot(s) + task description (this is the `model_input`)-- **Target sequence**: The action the agent took (this is the `loss_fn_inputs` target)-- **Weight mask**: 1.0 on action tokens, 0.0 on prompt tokens-- **Advantage weighting**: For PPO, multiply step reward by a per-step advantage estimate; apply as weight multiplier on action tokens--```python-import tinker.types as tt-from transformers import AutoTokenizer--def build_datum(step: RLStep, tokenizer, advantage: float) -> tt.Datum:- """Convert a single RL step to a Tinker Datum."""- # Tokenize action (target)- action_ids = tokenizer.encode(step.action_text, add_special_tokens=False)-- # Tokenize prompt context (images + instruction handled by rendering pipeline)- # ... (use Tinker's ImageChunk + EncodedTextChunk for multimodal)-- model_input = tt.ModelInput(- chunks=[- tt.EncodedTextChunk(token_ids=prompt_ids),- tt.ImageChunk(image=step.prompt_images[0]), # latest screenshot- tt.EncodedTextChunk(token_ids=action_ids),- ]- )-- # Shift targets: model predicts token[i+1] from token[i]- target_ids = action_ids # shifted internally by Tinker- weights = [max(0.0, advantage)] * len(action_ids) # zero-out negative advantages-- return tt.Datum(- model_input=model_input,- loss_fn_inputs=tt.LossFnInputs(- target_token_ids=target_ids,- weights=weights,- )- )--def build_batch(episode: Episode, tokenizer, gamma=0.99) -> list[tt.Datum]:- """Convert episode to list of Datums with discounted advantage estimates."""- # Compute discounted returns (simple Monte Carlo)- G = 0.0- returns = []- for step in reversed(episode.steps):- G = step.reward + gamma * G- returns.insert(0, G)-- baseline = sum(returns) / len(returns) if returns else 0.0-- datums = []- for step, G in zip(episode.steps, returns):- advantage = G - baseline- datums.append(build_datum(step, tokenizer, advantage))- return datums-```--### Phase 3: RL training loop--**New file**: `cua_bench/rl/train.py`--```python-"""-Main RL training loop using Tinker.--Usage:- python -m cua_bench.rl.train \- --base-model Qwen/Qwen3-8B \- --run-dir ~/.local/share/cua-bench/runs/<run_id> \- --epochs 10 \- --lr 1e-5-"""-import tinker-from tinker import ServiceClient-import tinker.types as tt--def train_epoch(training_client, episodes: list, tokenizer, args):- """Run one PPO update epoch over a batch of episodes."""- all_datums = []- for ep in episodes:- all_datums.extend(build_batch(ep, tokenizer, gamma=args.gamma))-- if not all_datums:- return-- # Submit forward-backward and optim step without awaiting between them- # (keeps operations in the same ~10s Tinker clock cycle)- fwdbwd_future = training_client.forward_backward_async(- data=all_datums,- loss_fn="ppo",- loss_fn_params=tt.PpoParams(- clip_low_threshold=0.8,- clip_high_threshold=1.2,- ),- )- optim_future = training_client.optim_step_async(- tt.AdamParams(- learning_rate=args.lr,- betas=(0.9, 0.999),- weight_decay=0.01,- grad_clip=1.0,- )- )-- # Collect results- fwdbwd_result = fwdbwd_future.result()- optim_future.result()- return fwdbwd_result--def main(args):- service_client = ServiceClient(api_key=args.tinker_api_key)-- training_client = service_client.create_lora_training_client(- base_model=args.base_model,- lora_config=tt.LoraConfig(rank=16, seed=42),- )-- tokenizer = training_client.get_sampling_client().tokenizer-- for epoch in range(args.epochs):- # 1. Run cua-bench rollouts with current policy- run_id = run_cuabench_rollouts(args)-- # 2. Load traces- run_dir = f"~/.local/share/cua-bench/runs/{run_id}"- episodes = load_run_episodes(run_dir)-- # 3. Log stats- avg_reward = sum(ep.terminal_reward for ep in episodes) / max(len(episodes), 1)- print(f"Epoch {epoch}: {len(episodes)} episodes, avg_reward={avg_reward:.3f}")-- # 4. Train- train_epoch(training_client, episodes, tokenizer, args)-- # 5. Save weights and get new sampling client- sampling_client = training_client.save_weights_and_get_sampling_client()-- # 6. Checkpoint full state every N epochs- if (epoch + 1) % args.checkpoint_every == 0:- ckpt_path = training_client.save_state(name=f"epoch-{epoch+1}").result()- print(f"Checkpoint saved: {ckpt_path}")-- # 7. Hot-reload updated weights into vLLM- sync_weights_to_vllm(training_client, args.vllm_url)--def run_cuabench_rollouts(args) -> str:- """Launch a cua-bench dataset run and return the run_id."""- import subprocess, re- result = subprocess.run([- "cb", "run", "dataset", args.tasks_path,- "--agent", "opencua",- "--model", args.model,- "--max-steps", str(args.max_steps),- ], capture_output=True, text=True)- run_id = re.search(r"run-[a-f0-9-]+", result.stdout).group(0)- # Wait for all sessions to complete (cb run watch or poll cb run info)- wait_for_run(run_id)- return run_id-```--### Phase 4: Weight sync to vLLM--After each Tinker training step the updated LoRA weights must reload into the running vLLM server so that subsequent cua-bench rollouts use the new policy.--**New file**: `cua_bench/rl/weight_sync.py`--```python-"""-Download LoRA checkpoint from Tinker and hot-reload into a running vLLM instance.-"""-import requests-import tarfile-import tempfile-from pathlib import Path--def sync_weights_to_vllm(training_client, vllm_base_url: str):- """- 1. Save weights to Tinker checkpoint- 2. Download tar archive- 3. POST to vLLM's /load_lora_adapter endpoint- """- # Save weights-only checkpoint (faster than save_state)- ckpt_path = training_client.save_weights_for_sampler(name="current").result()-- # Get signed download URL via RestClient- rest_client = training_client.service_client.create_rest_client()- url_response = rest_client.get_checkpoint_archive_url_from_tinker_path(ckpt_path)-- # Download tar archive- with tempfile.TemporaryDirectory() as tmpdir:- archive_path = Path(tmpdir) / "lora.tar"- r = requests.get(url_response.url, stream=True)- r.raise_for_status()- with open(archive_path, "wb") as f:- for chunk in r.iter_content(chunk_size=8192):- f.write(chunk)-- # Extract LoRA adapter files- with tarfile.open(archive_path) as tar:- tar.extractall(tmpdir)-- adapter_dir = Path(tmpdir) / "adapter"-- # Hot-reload via vLLM dynamic LoRA API- requests.post(- f"{vllm_base_url}/load_lora_adapter",- json={- "lora_name": "opencua-current",- "lora_path": str(adapter_dir),- }- ).raise_for_status()-```--### Phase 5: Resume training from checkpoint--```python-# Resume full training state (weights + optimizer momentum)-training_client = service_client.create_training_client_from_state_with_optimizer(- checkpoint_path="tinker://<model_id>/epoch-5"-)--# Resume weights only (fresh optimizer)-training_client = service_client.create_lora_training_client(- base_model=args.base_model,- lora_config=tt.LoraConfig(rank=16),- initial_checkpoint="tinker://<model_id>/epoch-5",-)-```-------## 5. File Structure--```-libs/cua-bench/-├── cua_bench/-│ └── rl/-│ ├── __init__.py-│ ├── trace_reader.py # Load HF Dataset traces → Episode objects-│ ├── datum_builder.py # Episode → Tinker Datum (tokenized, masked)-│ ├── train.py # Main RL training loop (entry point)-│ └── weight_sync.py # Download Tinker checkpoint → reload vLLM-├── plan.md # This file-└── pyproject.toml # Add tinker to [project.optional-dependencies]-```--Add to `pyproject.toml`:-```toml-[project.optional-dependencies]-rl = [- "tinker",- "transformers>=4.57.0,<5.0",-]-```-------## 6. End-to-End Training Script--```bash-# 1. Start vLLM with base model-vllm serve Qwen/Qwen3-8B \- --served-model-name opencua \- --enable-lora \- --host 0.0.0.0 --port 30000--# 2. Set env vars-export OPENCUA_BASE_URL=http://localhost:30000/v1-export OPENCUA_API_KEY=EMPTY-export TINKER_API_KEY=<your_tinker_key>--# 3. Run RL training loop-python -m cua_bench.rl.train \- --base-model Qwen/Qwen3-8B \- --tasks-path tasks/slack_env \- --model openai/opencua \- --max-steps 50 \- --epochs 20 \- --lr 1e-5 \- --gamma 0.99 \- --checkpoint-every 5 \- --vllm-url http://localhost:30000/v1-```-------## 7. Key Design Decisions--| Decision | Rationale |-|---|---|-| **Sparse reward** | `evaluate_task` returns a single terminal reward per episode; intermediate steps get reward 0.0 and the terminal step gets the full reward |-| **Monte Carlo returns** with simple baseline | Simple, unbiased, works well with sparse rewards; GAE would need a value network |-| **PPO loss in Tinker** | Standard choice for LLM RL; clip thresholds prevent oversized updates when off-policy rollouts diverge |-| **LoRA fine-tuning** | Tinker only supports LoRA currently; keeps checkpoint size small for fast sync to vLLM |-| **Weights-only sync to vLLM** | `save_weights_for_sampler` is faster than `save_state`; vLLM's dynamic LoRA API avoids server restart |-| **Async Tinker calls** | Submit `forward_backward_async` and `optim_step_async` without awaiting between them to keep both in the same ~10s Tinker clock cycle |-| **Screenshot as model input** | Each `agent_step` trace row has a screenshot; use the most recent as the visual context for that step's datum |-------## 8. Open Questions / Future Work--1. **Reference policy for IS ratio**: PPO needs log-probs from both current policy (from Tinker `compute_logprobs`) and sampling policy (from vLLM generation). Need to capture vLLM log-probs at rollout time and store them in trace metadata.--2. **Dense rewards**: Currently uses sparse evaluation reward. Could add partial rewards (e.g., intermediate task completion signals from `evaluate_task` called mid-episode) for faster learning.--3. **Multi-task batching**: Different task types may have very different reward scales. Per-task reward normalization or separate baselines may be needed.--4. **Rollout parallelism**: `cb run dataset` already parallelizes across tasks. Tinker's clock-cycle architecture means training throughput is bounded by how many rollout batches can be collected per ~10s cycle.--5. **Vision input**: OpenCUA is a vision-language model. Tinker supports Qwen3-VL via `ImageChunk`. The datum builder must encode screenshots as Tinker image chunks rather than base64 strings.--6. **SL warm-start before RL**: Use `save_state` after a supervised learning phase, then `create_training_client_from_state_with_optimizer` to continue into RL without losing optimizer momentum.diff --git a/libs/cua-driver-rs/DEV_LOOP_WINDOWS_VM.md b/libs/cua-driver-rs/DEV_LOOP_WINDOWS_VM.mddeleted file mode 100644index a71d5e7915..0000000000--- a/libs/cua-driver-rs/DEV_LOOP_WINDOWS_VM.md+++ /dev/null@@ -1,132 +0,0 @@-# Dev loop — cua-driver-rs on the Windows VM--Quick reference for iterating on Windows-specific cua-driver-rs code-directly on `<vm-name>` without round-tripping through GitHub-CD. Set up Sun 2026-05-18 by the autonomous session for the doctor-segfault / Session-0 hang investigations.--## Connection--- **SSH** (Session 0 — services context, no desktop):- ```bash- ssh -i ~/.ssh/<ssh-key> <user>@<vm-host>- ```-- **RDP** (Session 1+ with attached desktop, needed for visual / GUI tool validation):- ```bash- open <rdp-file-dir>/<vm-name>.rdp- ```- Credentials: `<user>` + password from your password manager.--## Toolchain on the VM (already installed)--- Git 2.54.0-- Rustup 1.29 + `stable-x86_64-pc-windows-msvc` (rustc 1.95.0)-- Cargo 1.95.0-- Visual Studio Build Tools 2022 (VCTools + Windows 11 SDK)-- Clone: `C:\Users\<user>\cua` (HEAD pinned to origin/main at clone time;- `git pull` to refresh)--## Iteration loop--1. **Edit locally on the Mac** (use Claude Code, your IDE, etc.).-2. **Push changed files to the VM** via `scp`:- ```bash- scp -i ~/.ssh/<ssh-key> \- <repo-root>/libs/cua-driver-rs/crates/platform-windows/src/<file>.rs \- <user>@<vm-host>:cua/libs/cua-driver-rs/crates/platform-windows/src/<file>.rs- ```-3. **Build on the VM** (~20 s incremental, ~5 min cold):- ```bash- ssh -i ~/.ssh/<ssh-key> <user>@<vm-host> \- "powershell -Command 'cd cua\\libs\\cua-driver-rs; cargo build --release -p cua-driver'"- ```-4. **Test on the VM** — use PowerShell `Start-Job` for clean timeouts;- pipe JSON args via stdin to bypass PowerShell's nested-quoting hell:- ```powershell- '{"name":"notepad"}' | & "$env:USERPROFILE\cua\libs\cua-driver-rs\target\release\cua-driver.exe" call launch_app- ```--## Gotchas the autonomous session hit--1. **PowerShell heredoc quoting** — `Start-Process -ArgumentList` with a JSON- arg containing `"` will silently strip the quotes. **Workaround:**- pipe the JSON via stdin to `cua-driver call <tool>` instead of- passing it as argv.--2. **`Start-Process -RedirectStandardError` buffering** — stderr from- the child process may not flush to the redirect file until the child- exits cleanly. If you `Stop-Process` mid-run you lose buffered- stderr. **Workaround:** use `Start-Job` with `Tee-Object` for live- capture, or write to a file from inside the script.--3. **PowerShell strings with `(`, `,`, em-dashes** — PowerShell parses- parentheses and commas inside double-quoted strings as method calls- in some contexts. Em-dashes get corrupted by `scp`. **Workaround:**- stick to plain ASCII single-quoted strings in any script you `scp`- over.--4. **`R` is a PowerShell alias for `Invoke-History`.** Don't name a- helper function `R`.--5. **Some parity examples hard-code `target/debug/cua-driver.exe`** —- if you've only run `cargo build --release` you need to also run- `cargo build` (debug) once, otherwise `list_tools_parity` and- `describe_parity` fail with "cannot find the path specified".--6. **`SetForegroundWindow` is restricted on Windows** — calling it- from a non-foreground process silently no-ops most of the time. The- foreground-restore in `launch_uwp` is best-effort. Plan visual- confirmation in Session 1+ for any focus-related work.--7. **Visual Studio Build Tools install is ~5 GB and takes ~10-15 min- over Azure VM network.** Don't reinstall — it's there.--## Running the parity suite--```powershell-$env:CUA_DRIVER_RS_TELEMETRY_ENABLED = "0"-$exe = "$env:USERPROFILE\cua\libs\cua-driver-rs\target\release\cua-driver.exe"--# 1. Start the daemon-Get-Process | Where-Object { $_.Name -eq 'cua-driver' } | Stop-Process -Force-Start-Process -FilePath $exe -ArgumentList 'serve' -NoNewWindow--# 2. Run all parity examples (skip overlay_dump — it's interactive-overlay-only)-cd $env:USERPROFILE\cua\libs\cua-driver-rs-$examples = @('list_tools_parity','describe_parity','get_screen_size_parity',- 'get_cursor_position_parity','config_parity',- 'agent_cursor_setters_parity','get_agent_cursor_state_parity',- 'set_agent_cursor_style_parity','list_windows_parity',- 'list_apps_parity','launch_app_parity')-foreach ($ex in $examples) {- cargo run --release --quiet --example $ex -p platform-windows- Write-Host "$ex exit=$LASTEXITCODE"-}--# 3. Cleanup-Get-Process | Where-Object { $_.Name -eq 'cua-driver' } | Stop-Process -Force-```--In Session 0 the autonomous session got 11/11 PASS as of 2026-05-18.--## Commands the autonomous session used most--```bash-# Quick file push-scp -i ~/.ssh/<ssh-key> <local> <user>@<vm-host>:cua/libs/cua-driver-rs/...--# Quick build + test with timeout (Mac side)-ssh -o ServerAliveInterval=20 -i ~/.ssh/<ssh-key> <user>@<vm-host> \- 'powershell -ExecutionPolicy Bypass -File some-script.ps1'--# Quick PSCMD via stdin pipe-echo '{"name":"notepad"}' | ssh -i ~/.ssh/<ssh-key> <user>@<vm-host> \- '"C:\\Users\\<user>\\cua\\libs\\cua-driver-rs\\target\\release\\cua-driver.exe" call launch_app'--# Kill all stale cua-driver procs-ssh -i ~/.ssh/<ssh-key> <user>@<vm-host> \- "powershell -Command 'Get-Process cua-driver -EA SilentlyContinue | Stop-Process -Force'"--# Check which sessions exist-ssh -i ~/.ssh/<ssh-key> <user>@<vm-host> "powershell -Command qwinsta"-```diff --git a/libs/python/mcp-server/QUICK_TEST_COMMANDS.sh b/libs/python/mcp-server/QUICK_TEST_COMMANDS.shdeleted file mode 100755index 3242c610b8..0000000000--- a/libs/python/mcp-server/QUICK_TEST_COMMANDS.sh+++ /dev/null@@ -1,63 +0,0 @@-#!/bin/bash-# Quick Test Commands for MCP Server Local Desktop Option-# Run these commands to test the implementation--set -e # Exit on error--echo "======================================================================"-echo "Testing MCP Server Local Desktop Option"-echo "======================================================================"-echo ""--# Change to repo root-cd "$(dirname "$0")/.."--# Test 1: Quick Logic Test (No setup required)-echo "Test 1: Quick Logic Test (No setup required)"-echo "----------------------------------------------------------------------"-python tests/quick_test_local_option.py-echo ""--# Test 2: Automated Tests (Requires pytest and packages)-echo "Test 2: Automated Tests (Requires pytest and packages installed)"-echo "----------------------------------------------------------------------"-if command -v pytest &> /dev/null; then- echo "Running pytest..."- pytest tests/test_mcp_server_local_option.py -v || echo "Note: Some tests may require full setup"-else- echo "⚠️ pytest not found. Install with: pip install pytest"-fi-echo ""--# Test 3: Existing MCP server tests-echo "Test 3: Existing MCP Server Tests"-echo "----------------------------------------------------------------------"-if command -v pytest &> /dev/null; then- echo "Running existing session management tests..."- pytest tests/test_mcp_server_session_management.py -v || echo "Note: Some tests may fail if dependencies are missing"-else- echo "⚠️ pytest not found. Install with: pip install pytest"-fi-echo ""--# Summary-echo "======================================================================"-echo "Test Summary"-echo "======================================================================"-echo "✅ Quick logic test completed"-echo ""-echo "Next steps for comprehensive testing:"-echo "1. Install dependencies:"-echo " pip install -e libs/python/core"-echo " pip install -e libs/python/computer"-echo " pip install -e libs/python/agent"-echo " pip install -e libs/python/mcp-server"-echo " pip install -e libs/python/computer-server"-echo ""-echo "2. For manual end-to-end testing, see:"-echo " tests/MANUAL_TEST_LOCAL_OPTION.md"-echo ""-echo "3. For detailed testing info, see:"-echo " tests/TESTING_SUMMARY.md"-echo ""-diff --git a/libs/python/mcp-server/quick_test_local_option.py b/libs/python/mcp-server/quick_test_local_option.pydeleted file mode 100755index 6e2caab229..0000000000--- a/libs/python/mcp-server/quick_test_local_option.py+++ /dev/null@@ -1,244 +0,0 @@-#!/usr/bin/env python3-"""-Quick test to verify the local desktop option logic without full setup.--This script tests the environment variable parsing and logic flow-without requiring VMs, computer-server, or MCP clients to be running.-"""--import os-import sys---def test_env_var_parsing():- """Test that environment variable is parsed correctly."""- print("Testing CUA_USE_HOST_COMPUTER_SERVER environment variable parsing...")- print("-" * 60)-- test_cases = [- # (env_value, expected_result, description)- ("true", True, "lowercase 'true'"),- ("True", True, "capitalized 'True'"),- ("TRUE", True, "uppercase 'TRUE'"),- ("1", True, "numeric '1'"),- ("yes", True, "lowercase 'yes'"),- ("Yes", True, "capitalized 'Yes'"),- ("false", False, "lowercase 'false'"),- ("False", False, "capitalized 'False'"),- ("FALSE", False, "uppercase 'FALSE'"),- ("0", False, "numeric '0'"),- ("no", False, "lowercase 'no'"),- ("", False, "empty string"),- ("random", False, "random value"),- (None, False, "not set (None)"),- ]-- passed = 0- failed = 0-- for env_value, expected, description in test_cases:- # Simulate the logic from session_manager.py line 59- if env_value is None:- actual = os.getenv("CUA_USE_HOST_COMPUTER_SERVER", "false").lower() in (- "true",- "1",- "yes",- )- else:- os.environ["CUA_USE_HOST_COMPUTER_SERVER"] = env_value- actual = os.getenv("CUA_USE_HOST_COMPUTER_SERVER", "false").lower() in (- "true",- "1",- "yes",- )-- status = "✓ PASS" if actual == expected else "✗ FAIL"- if actual == expected:- passed += 1- else:- failed += 1-- print(- f"{status} | Value: {env_value!r:15} | Expected: {expected!s:5} | Got: {actual!s:5} | {description}"- )-- # Clean up- os.environ.pop("CUA_USE_HOST_COMPUTER_SERVER", None)-- print("-" * 60)- print(f"Results: {passed} passed, {failed} failed")- return failed == 0---def test_session_manager_logic():- """Test the logic flow in session_manager.py without actual Computer creation."""- print("\nTesting session_manager.py logic flow...")- print("-" * 60)-- # Read the actual session_manager.py to verify the logic- import pathlib-- session_manager_path = (- pathlib.Path(__file__).parent.parent- / "libs"- / "python"- / "mcp-server"- / "mcp_server"- / "session_manager.py"- )-- if not session_manager_path.exists():- print(f"✗ FAIL | session_manager.py not found at {session_manager_path}")- return False-- content = session_manager_path.read_text()-- # Check for the key logic- checks = [- ('os.getenv("CUA_USE_HOST_COMPUTER_SERVER"', "Environment variable check present"),- ("use_host_computer_server=use_host", "use_host_computer_server parameter passed"),- ("Computer(", "Computer instantiation present"),- ]-- all_checks_passed = True- for check_str, description in checks:- if check_str in content:- print(f"✓ PASS | {description}")- else:- print(f"✗ FAIL | {description} - not found")- all_checks_passed = False-- print("-" * 60)- return all_checks_passed---def test_documentation_consistency():- """Verify documentation mentions the new feature."""- print("\nTesting documentation consistency...")- print("-" * 60)-- import pathlib-- docs_to_check = [- ("configuration.mdx", "CUA_USE_HOST_COMPUTER_SERVER"),- ("usage.mdx", "Targeting Your Local Desktop"),- ]-- docs_path = (- pathlib.Path(__file__).parent.parent- / "docs"- / "content"- / "docs"- / "libraries"- / "mcp-server"- )-- all_docs_ok = True- for doc_file, expected_content in docs_to_check:- doc_path = docs_path / doc_file- if not doc_path.exists():- print(f"✗ FAIL | {doc_file} not found")- all_docs_ok = False- continue-- content = doc_path.read_text()- if expected_content in content:- print(f"✓ PASS | {doc_file} contains '{expected_content}'")- else:- print(f"✗ FAIL | {doc_file} missing '{expected_content}'")- all_docs_ok = False-- print("-" * 60)- return all_docs_ok---def print_usage_examples():- """Print usage examples for both modes."""- print("\n" + "=" * 60)- print("USAGE EXAMPLES")- print("=" * 60)-- print("\n1. DEFAULT MODE (VM):")- print("-" * 60)- print(- """-{- "mcpServers": {- "cua-agent": {- "command": "/bin/bash",- "args": ["~/.cua/start_mcp_server.sh"],- "env": {- "CUA_MODEL_NAME": "anthropic/claude-sonnet-4-5-20250929"- }- }- }-}--Note: CUA_USE_HOST_COMPUTER_SERVER is not set, so VM mode is used (safe).-"""- )-- print("\n2. LOCAL DESKTOP MODE:")- print("-" * 60)- print(- """-Step 1: Start computer-server locally:- python -m computer_server--Step 2: Configure MCP client:-{- "mcpServers": {- "cua-agent": {- "command": "/bin/bash",- "args": ["~/.cua/start_mcp_server.sh"],- "env": {- "CUA_MODEL_NAME": "anthropic/claude-sonnet-4-5-20250929",- "CUA_USE_HOST_COMPUTER_SERVER": "true"- }- }- }-}--⚠️ WARNING: AI will have direct access to your desktop!-"""- )---def main():- """Run all quick tests."""- print("=" * 60)- print("QUICK TEST: MCP Server Local Desktop Option")- print("=" * 60)- print()-- results = []-- # Run tests- results.append(("Environment Variable Parsing", test_env_var_parsing()))- results.append(("Session Manager Logic", test_session_manager_logic()))- results.append(("Documentation Consistency", test_documentation_consistency()))-- # Print summary- print("\n" + "=" * 60)- print("SUMMARY")- print("=" * 60)- for test_name, passed in results:- status = "✓ PASSED" if passed else "✗ FAILED"- print(f"{status} | {test_name}")-- all_passed = all(result for _, result in results)-- if all_passed:- print("\n🎉 All quick tests passed!")- print_usage_examples()- print("\nNext steps:")- print("1. Run full automated tests: pytest tests/test_mcp_server_local_option.py")- print("2. Follow manual testing guide: tests/MANUAL_TEST_LOCAL_OPTION.md")- return 0- else:- print("\n❌ Some tests failed. Please review the output above.")- return 1---if __name__ == "__main__":- sys.exit(main())