Skip to content

mini-swe-agent.md

Latest commit

 

History

History
49 lines (32 loc) · 2.89 KB

mini-swe-agent.md

File metadata and controls

49 lines (32 loc) · 2.89 KB

Mini SWE Agent bridge

Just a show of the bridge in action. Not a best-effort or absolutely-correct implementation of the agent.

This example wraps mini-swe-agent behind ACP so Zed can run it as an external agent over stdio. It also includes a local Textual UI client connected via a duet launcher

Behavior

  • Prompts: text blocks are concatenated into a single task string. (Resource embedding is not used in this example.)
  • Streaming: only LM output is streamed via session/update → agent_message_chunk.
  • Tool calls: when the agent executes a shell command, the bridge sends:
    • tool_call with kind=execute, pending status, and a bash code block containing the command
    • tool_call_update upon completion, including output and a rawOutput object with output and returncode
  • Final result: on task submission (mini-swe-agent prints COMPLETE_TASK_AND_SUBMIT_FINAL_OUTPUT as the first line), a final agent_message_chunk with the submission content is sent.

Configuration

Environment variables control the model:

  • MINI_SWE_MODEL: model ID (e.g. openrouter/openai/gpt-4o-mini)
  • OPENROUTER_API_KEY for OpenRouter; or OPENAI_API_KEY / ANTHROPIC_API_KEY for native providers
  • Optional MINI_SWE_MODEL_KWARGS: JSON, e.g. { "api_base": "https://openrouter.ai/api/v1" } (auto-injected for OpenRouter if missing)

Agent behavior automatically maps the appropriate API key based on the chosen model and available environment variables.

If mini-swe-agent is not installed in the venv, the bridge attempts to import a vendored reference copy under reference/mini-swe-agent/src.

How to run

  • In Zed (editor integration): configure an agent server to launch examples/mini_swe_agent/agent.py and set the environment variables there. Use Zed’s “Open ACP Logs” to inspect tool_call/tool_call_update and message chunks.
  • In terminal (local TUI): run the duet launcher to start both the agent and the Textual client with the same environment and dedicated pipes:
python examples/mini_swe_agent/duet.py

The launcher loads .env from the repo root (using python-dotenv) so both processes share the same configuration.

TUI usage

  • Hotkeys: y → YOLO, c → Confirm, u → Human, Enter → Continue.
  • In Human mode, you’ll be prompted for a bash command; it will be executed and streamed back as a tool call.
  • Each executed command appears in the “TOOL CALLS” section with live status and output.

Files