ExecOrchestra

Heterogeneous Resource Routing and Orchestration

Coordinate tools, models, skills, and agents through executable plans.

Quick start  Models and services  Documentation

English · 简体中文

✨ Overview

ExecOrchestra is a heterogeneous resource orchestration framework for complex tasks. It combines tools, models, skills, and agents into executable workflows, allocating generative processing and direct execution as needed. Tools handle explicitly specified operations, while models and agents contribute where generation, judgment, or interaction is required. Clear execution records, intermediate artifacts, and delivered results help developers follow progress, diagnose issues, and extend their applications.

🧩 System framework

ExecOrchestra framework: task workflow, typed candidate construction across heterogeneous resources, and Direct, Localized, and Extended execution modes.

Organize heterogeneous resources around task requirements and determine the generative scope of each subtask.

Three execution modes

A plan gives the model as much control as the subtask calls for:

Execution mode How the work gets done
Direct The plan fixes the operations and their inputs. Tools carry them out without an intermediate model call.
Localized The plan gives a model a specific job, such as writing a short processing script. Tools do the reading, writing, or computation.
Extended A model or agent can inspect a tool’s response and choose its next action within the permitted tools and session limits.

🚀 Quick start

📋 Prerequisites

  • Operating system: Linux x86-64
  • Python: 3.11 or higher
  • Tools: Git, Conda, and a running Docker Engine accessible to your user
  • Model services: A compatible LLM API and Embedding model services (Qwen3-Embedding-4B by default).

📦 Installation

  1. Clone the repository

    git clone -b execorchestra --single-branch https://github.com/OpenBMB/ChatDev.git
    cd ChatDev
    cd execorchestra
    
  2. Set the environment and install dependencies

    conda create -n execorchestra python=3.11 -y
    conda activate execorchestra
    python -m pip install torch==2.11.0 --index-url https://download.pytorch.org/whl/cpu
    python -m pip install -r requirements.txt
    python -m scripts.prepare_release init
    test -f .env || cp .env.example .env
    
  3. Build the Docker tool environment

    docker build -f sgar_mvp/docker/Dockerfile.runtime -t sgar-runtime:rc1 .
    python -m sgar_mvp.scripts.seal_rc1_runtime
    
  4. Prepare the embedding model service

    For local deployment, install the embedding environment and download the pinned model and tokenizer.

    conda create -n execorchestra-embedding python=3.11 -y
    conda run -n execorchestra-embedding python -m pip install -r requirements-embedding.txt
    python -m scripts.download_embedding
    

    Downloads are stored under .cache/embedding-model and .cache/embedding-tokenizer.

Using an existing embedding service instead

Skip the embedding environment and model-weight download above. Download only the matching tokenizer:

python -m scripts.download_embedding --tokenizer-only

In sgar_mvp/config/embedding_release.json, set runtime.api_endpoint_url to the full embeddings URL and runtime.api_model_id to its served model name. Keep runtime.tokenizer_path as .cache/embedding-tokenizer. The endpoint must use the configured Qwen3-Embedding-4B revision and return normalized 2,560-dimensional vectors.

🔑 Configuration

  1. Fill in API endpoints and keys

    Edit ChatDev/execorchestra/.env. Replace the values below with your own; the file contains grouped comments for required and optional settings.

    LLM_BASE_URL="https://your-gateway.example/v1"
    LLM_API_KEY="your-model-api-key"
    SGAR_EMBEDDING_API_KEY="your-embedding-api-key"
    
    • LLM_BASE_URL / LLM_API_KEY: Your OpenAI-compatible model gateway and its access key.
    • SGAR_EMBEDDING_API_KEY: The embedding endpoint key. For local deployment, choose a private nonempty value; the service and client load the same .env.
  2. Load configuration into the framework terminal

    conda activate execorchestra
    set -a
    source .env
    set +a
    

    Repeat this step when opening a new framework terminal.

🧰 Runtime preparation

  1. (Optional) Start the local embedding service

    For local deployment, open a second terminal in ChatDev/execorchestra. On a dedicated GPU server, choose an available GPU and run:

    conda activate execorchestra-embedding
    set -a
    source .env
    set +a
    export CUDA_VISIBLE_DEVICES=0
    export SGAR_EMBEDDING_DEDICATED_GPU=1
    export SGAR_EMBEDDING_MODEL_PATH="$PWD/.cache/embedding-model"
    python -m scripts.embedding_server
    

    The default endpoint is http://127.0.0.1:8001/v1/embeddings. Keep it running during preparation and task execution; stop it with Ctrl+C afterward.

  2. Prepare resources in the framework terminal

    python -m scripts.prepare_release all --allow-unavailable
    

    This command checks models, control roles, tools, and agents, prepares the usable resource pool, builds its index, and verifies configuration in that order. It makes live API/tool requests and may incur charges; it stops at the first failed stage. Unavailable candidates are recorded in sgar_mvp/runtime_state/preparation_report.json. Omit --allow-unavailable when every configured candidate must be available.

⚡ Run a task

After preparation succeeds, write your request to a UTF-8 text file and place it with any attachments in a dedicated input directory. Replace the placeholders:

bash scripts/run.sh \
  --query-file "<task-file>" \
  --public-input-root "<input-directory>" \
  --run-dir "<new-run-directory>"
Parameter Meaning
--query-file File containing the task request and expected deliverable; it must be inside the declared input directory.
--public-input-root Directory the task may read. Repeat this option for multiple separate input roots.
--run-dir New or empty directory for this attempt’s logs and results.
--input NAME=PATH Optional named attachment inside a declared input root. May be repeated.
--network-policy declared Enable declared network access for web tasks after configuring search/browser services. Tool network access is disabled by default.

The terminal displays each stage, selected resources, intermediate outputs, and delivery status. The same readable trace is saved to <new-run-directory>/progress.log. Check run_manifest.json for completion status and delivery_manifest.json for the delivered file path (delivery_locator, relative to the run directory).

Optional: run an example (XBench case 157)
bash scripts/run_xbench_demo.sh

This XBench DeepSearch-2510 task asks for the name of a coauthor of an XBench paper. The example shows how ExecOrchestra searches the web, checks the paper's author list, and delivers an answer with source links. Example details

View output files and log details
In the run directory Use it to
progress.log Follow the same guided stages seen in the terminal, including intermediate output previews.
pipeline.log Inspect detailed requests, candidate scores, plan validation and runtime diagnostics. Use SGAR_CONSOLE=verbose to also show detailed diagnostic records live.
run_manifest.json Check the final run status and the recorded cause of a failure.
delivery_manifest.json, when present Locate the delivered output and verify its content hash.
experiment_report.md Read the run report. Benchmark answer grading is a separate step.

📚 Documentation

Start here What you will find
Install & run Detailed example verification and preparation troubleshooting.
Environment & services · Model configuration Local settings, model gateways, embeddings and tool dependencies.
Benchmark evaluation GAIA and Terminal-Bench entry points and scoring.
Resource catalog · Framework source Resource declarations, orchestration and execution logic.
Third-party notices All involved content from other sources in this project.

👥 Contributors

TL-Z
TL-Z
WilsonWukz
WilsonWukz
ffcosmos
ffcosmos
Wuuu-uu
Wuuu-uu
hendrick-wang
hendrick-wang
huatl98
huatl98
inwt2333
inwt2333
LLLLLLsk0231
LLLLLLsk0231
cuixitong
cuixitong
qianc62
qianc62

🤝 Acknowledgments

Shanghai Artificial Intelligence Laboratory
Description
ChatDev 2.0: Dev All through LLM-powered Multi-Agent Collaboration
Readme Apache-2.0 460 MiB
Languages
Python 69.7%
Vue 27.5%
JavaScript 2.2%
CSS 0.4%
Dockerfile 0.1%
Other 0.1%