corral-harness

Generated by scripts/gen-cli-docs.sh from corral-harness’s own -h output and its main.go doc comment. Do not hand-edit — run scripts/gen-cli-docs.sh and commit the result.

Usage

corral-harness — loops a headless coding agent (Claude Code, Gemini CLI, Codex, ...) as a swarm bee over MCP

Usage:
  corral-harness         claim one task, work it, complete it, exit
  corral-harness -h      print this help and exit

Env:
  CORRAL_BRAIN   brain URL (default http://localhost:9019)
  AGENT_NAME       swarm name (default Harness)
  AGENT_ROLE       role(s) to serve (default generalist): a single role, a
                 comma-separated list (e.g. "mutant-generator,test-writer,test-critic") to
                 claim any ready task in that set, or "any"/"*"/empty to
                 claim ANY ready task as a pure generalist
  AGENT_MODEL      the model driving this harness (e.g. gpt-5.1-codex); adds a
                 report_host step so findings attribute to it in model_comparison
  AGENT_BACKEND    the backend/vendor for AGENT_MODEL (e.g. openai, anthropic, gemini)
  AGENT_WORKSPACE  working directory for the harness (default .)
  HARNESS_CMD    command template; placeholders: {prompt} {mcp_config} {brain}
                 e.g. claude -p {prompt} --mcp-config {mcp_config} \
                      --allowedTools "mcp__corral__*,Read,Write,Edit,Bash" \
                      --permission-mode acceptEdits
  HARNESS_DESC   how to announce this harness (default derived from HARNESS_CMD)
  AGENT_ROUNDS     max tasks to run, 0 = forever (default 0)
  HARNESS_TIMEOUT_SECONDS  per-invocation kill deadline (default 900)
  HARNESS_IDLE_SECONDS     backoff when the queue is empty (default 30)
  AGENT_PROMPT_FILE optional file replacing the built-in bee prompt; the same
                 placeholders are substituted into it