Miroflow Agent Server

Hydra Overrides

Any extra arguments passed to ./scripts/run_miroflow_agent_server.sh ... are forwarded directly to api_server.py as Hydra overrides.

Use this to switch the server's default LLM config at launch time:

./scripts/run_miroflow_agent_server.sh llm=gpt-5
./scripts/run_miroflow_agent_server.sh llm=qwen-3
./scripts/run_miroflow_agent_server.sh llm=glm-5-agenthub

The currently exposed LLM choices are the config names under:

scripts/miroflow-agent-server/conf/llm/

At the time of writing, those include:

• claude-3-7
• claude-haiku-4.5
• gemini-2.5-flash
• gemini-3-flash-agenthub
• gemini-3-pro
• gemini-3-pro-agenthub
• glm-5-agenthub
• gpt-5
• gpt-5.2-agenthub
• qwen-3

Hydra override values use the config basename without the .yaml suffix.

Environment Variables

• Required:
  • OPENAI_API_KEY
• Optional:
  • OPENAI_BASE_URL
  • SERPER_API_KEY, SERPER_BASE_URL
  • JINA_API_KEY, JINA_BASE_URL
  • E2B_API_KEY
  • SUMMARY_LLM_API_KEY, SUMMARY_LLM_BASE_URL, SUMMARY_LLM_MODEL_NAME

If a run needs extra providers or tool integrations, check scripts/miroflow-agent-server/.env.example for the full expected environment.

Quick Reference

./scripts/run_miroflow_agent_server.sh
./scripts/stop_miroflow_agent_server.sh
./scripts/test_miroflow_agent_server.sh [image-file]
curl --noproxy '*' http://127.0.0.1:7210/health

Copied source lives in scripts/miroflow-agent-server/; logs go to logs/.

API Reference

The launched server exposes three client-facing endpoints:

• GET /health — lightweight health check for liveness/readiness
• POST /v1/api/execute — submit one request and receive streamed NDJSON output
• POST /v1/api/upload — upload one file into the sandbox bound to X-Session-Id

For detailed request/response shapes and examples in both curl and pure Python, see:

• API_USAGE.md

That reference also documents:

• required and optional headers
• session reuse via X-Session-Id
• NDJSON stream parsing for /v1/api/execute
• file upload flow for /v1/api/upload

Success Criteria

• curl --noproxy '*' http://127.0.0.1:7210/health returns HTTP 200 with {"status":"healthy"}
• A fresh server run writes a new log file under logs/
• The wrapper test can connect to port 7210

Common Failures

• Sandbox bind failure: in the Codex sandbox, opening a listening local HTTP port may fail even when the app is fine. Retry outside the sandbox before treating it as a server bug.
• Proxy interference: local curl checks should bypass proxies with --noproxy '*'.
• Missing dependencies: let preflight finish creating the skill-local .venv, or set MIROFLOW_PYTHON=/path/to/python to an environment that already has MiroFlow dependencies installed.
• Wrong process on port 7210: ./scripts/stop_miroflow_agent_server.sh only kills a listener if /proc/<pid>/cwd points at this skill's copied server directory and the command line matches api_server.py.

Verification Checklist

• ./scripts/run_miroflow_agent_server.sh has been started in a suitable environment
• ./scripts/stop_miroflow_agent_server.sh cleanly stops only the verified skill-owned listener on port 7210
• curl --noproxy '*' http://127.0.0.1:7210/health succeeds
• A fresh log file appears in logs/
• If sandbox startup failed, the same command was retried outside the sandbox

Constraints

• Treat scripts/miroflow-agent-server/ as a copied source snapshot; do not rewrite it unless the user explicitly asks to refresh the copy.
• If the upstream repo changes, re-copy needed files from ~/MiroThinker/apps/miroflow-agent instead of editing the snapshot in place.
• No automatic log rotation is configured; prune old files in logs/ manually if repeated runs make the directory grow too large.