agent-harness

# Agent Harness Router Use this skill when the request spans an LLM control harness and its concrete Codex or Claude runtime, or when the correct narrow workflow is unclear. Route first; then load only the selected leaf skill. ## Routing - Design, refactor, recover, or reconfigure the model/tool/state control plane, including provider, tool, module, or loop hot swap: use `agent-harness-engineering`. - Benchmark, replay, diagnose, compare, or release-gate complete harness behavior, including concurrent runtime generations and rollback: use `agent-harness-evaluation`. - Operate or diagnose Codex CLI: start with `codex-cli`, which routes exec and review automation, deferred completion, task supervision, plugin/MCP management, doctor/debug, rollout forensics, and app environments. - Operate or diagnose Claude Code: start with `claude-code`, which routes print automation, plugin/MCP management, doctor/debug, hooks/settings, agents, worktrees, and sessions. - Supply a password, passphrase, OTP, API token, device approval, or 1Password credential to a task without returning the value to the agent: use `credential-handoff`. - Prove what launchd, systemd timer, cron, or Windows Task Scheduler actually ran: use `scheduled-automation-runtime`. For mixed requests, separate the claim boundaries. Use the vendor leaf for exact commands, configuration, hooks, logs, or sessions; use harness engineering for runtime-control design; use harness evaluation for empirical ...