Swe2 Runner — Project Management agent for Claude Code
Dispatches a bounded implementation task to an external SWE-2 agent (Cognition devin CLI) inside a named worktree, iterates with it until the task is done or blocked, and reports back the diff and v.
How to install Swe2 Runner
Installs to ~/.claude/agents/milwil-2-claude-swe2-dispatch-swe2-runner.md
mkdir -p ~/.claude/agents && curl -fsSL https://raw.githubusercontent.com/milwil-2/claude-swe2-dispatch/HEAD/agents/swe2-runner.md -o ~/.claude/agents/milwil-2-claude-swe2-dispatch-swe2-runner.md Restart Claude Code, or start a new session, for it to be picked up.
What Swe2 Runner does
name: swe2-runner description: Dispatches a bounded implementation task to an external SWE-2 agent (Cognition `devin` CLI) inside a named worktree, iterates with it until the task is done or blocked, and reports back the diff and verification result. Use when you want SWE-2 to do the editing while this session keeps ownership of integration, review, and git. Not a reviewer. model: inherit effort: high color: magenta tools: Read, Glob, Grep, Bash
You drive an external coding agent — SWE-
Alternatives in Project Management
- Acceptance Checker.Agent — Maps implementation and tests to task acceptance criteria, identifies remaining gaps, and provides a clear don 47 ★
- Task Checkbox Updates: — [ ] → [x] # Fully completed - [ ] → [~] # Partially done (add % if helpful) - [ ] → [ 45 ★
- Red Reviewer — Adversarial verification gate for the genai_incidents improvement plan 28 ★
Full documentation available on GitHub
View Source RepositoryRelated Agents
Spike
Bounded investigation that answers one specific technical unknown so a blocked task can resume. Returns a find
Agy Runner
Drives one delegated task through the agy MCP server and reports back, so the orchestrator is not blocked whil
Micro Worker
Minimal-footprint worker for ONE bounded, fully-specified task — a single file edit, a scoped change, a named
Engine Validation Runner
Use for focused verification while building — run the project's self-tests, or a named subset of them, and get
Task Runner
Owns the whole agent loop for ONE task. Main spawns it with a task and gets a single schema back; the runner r
Junior Worker
Use when a domain lead needs a bounded, well-specified task executed — one file, one feature, one wireframe, o
Related Skills
Cognition Claude Proxy
Use Devin CLI's model catalog (SWE-2, GLM-5.2, DeepSeek, 200+) with Claude Code as the backend
Iterate Revision
Revise one manuscript section against the reviewer team until DONE/BLOCKED/HALT. Writer ↔ reviewer loop driven
WindsurfAPI
Turn Windsurf / Devin Desktop's 100+ AI models (Claude, GPT, Gemini, DeepSeek, Kimi, GLM, SWE) into OpenAI-, A