orchestrator Files

Linux Orchestrator (Minimal)

A small Node.js HTTP service that orchestrates a desktop automation loop on Linux:

• Calls an OpenAI-compatible chat API (OpenAI/Azure/Anthropic via proxy, or LM Studio)
• Sends single-step actions to a desktop driver (if you have one) or executes local shell steps
• Stores tasks under data/tasks/<id> for later inspection

Env Vars

• ORCH_PORT: default 4101
• ORCH_MODEL: default qwen2.5-coder-1.5b (use your model id)
• OPENAI_API_BASE: default http://localhost:1234/v1 (LM Studio)
• OPENAI_API_KEY: optional key (LM Studio may not need one)
• DESKTOP_DRIVER_URL: optional driver endpoint (default http://127.0.0.1:39990/computer-use)
• MAX_STEPS: default 8
• MAX_RETRIES_PER_STEP: reflection retries on failure (default 2)
• MAX_RECURSION_DEPTH: limit for spawn_subtask (default 2)
• BASE_BACKOFF_MS: base milliseconds for retry backoff (default 500)
• PYTHON_BIN: python binary name/path used for tools that need Python (default "python3")
• MAX_CONTEXT_CHARS: character budget for context (default 6000)

Run npm start

Endpoints

• GET /health
• GET /tasks
• POST /tasks { goal }
• GET /tasks/:id
• GET /tasks/:id/messages
• GET /tasks/:id/context
• GET /tasks/:id/notes
• POST /tasks/:id/notes { note }
• GET /events/:id (SSE stream of task logs)

Available Tools

• screenshot, move_mouse, click_mouse, scroll, type_text, key_press, open_app
• run_shell, verify_result, spawn_subtask, spawn_template, done

Verification Tool

• verify_result supports:
• file_exists: { path }
• api_call: { url, method?, expected_status?, expectation? }
• test: { script, expectation? } → runs in bash
• vscode_test: { extension_path, test_type: compile|command, command? }

Notes

• On Linux, run_shell executes your command with /bin/bash -lc and captures stdout/stderr/exitCode.
• The desktop driver is optional; if unreachable, the retry loop will record the error and continue.
• Requires Node.js 18+ for global fetch.