local-agent-delegationlisted

# Local Agent Delegation Use this skill to fan out one bounded prompt to multiple agent runtimes and collect their outputs for parent-agent review. ## Ground Rules 1. Keep the parent Codex thread responsible for scope, final judgment, commits, publish, and cleanup. 2. Default to read-only prompts. For edit tasks, create a fresh task worktree first and name the allowed files. 3. Do not pass secrets, tokens, private keys, `.env` contents, or broad personal paths into prompts. 4. Do not use local Hermes. Hermes delegation is only through the VPS SSH adapter. 5. Do not use `--yolo`, `--dangerously-skip-permissions`, or equivalent broad approval flags unless the user explicitly scoped that risk. 6. Do not send private project context to web agents. This skill is for local CLIs and Hermes VPS only. ## Adapters - `claude`: `claude -p ... --output-format json --permission-mode plan` - `mimo`: `mimo run ... --pure --format json --dir <cwd>`; no hard permission flag is assumed, so keep prompts read-only and rely on the helper's post-run mutation detector - `gemini`: `gemini -p ... --skip-trust --output-format json --approval-mode plan` with scoped `GEMINI_CLI_TRUST_WORKSPACE=true` for headless trusted-workspace compatibility - `hermes-vps`: `ssh` using explicit `HERMES_VPS_KEY`, `HERMES_VPS_HOST`, `HERMES_VPS_USER`, optional `HERMES_VPS_CWD`, and optional `HERMES_VPS_EXECUTABLE`, then remote `<hermes-executable> chat -q ... -Q --source tool` - `odysseus-api`: placeholder only; use