agy-delegatelisted

## Use `agent-hub`, not this skill directly Prefer `delegate({agent:'agy', model, task, cwd, mode, ...})` through the `agent-hub` MCP server. Read `~/.claude/skills/multi-agent-orchestrator/SKILL.md` for the full loop (`agents_status → route → delegate → job_wait/job_status → job_result → synthesize`), the break-even rule, and the delegation map. Agy-specific models, health signals, measured latencies, and gotchas (silent `CANCELED`, `--json-schema` unreliability, `GEMINI.md` not auto-loaded, `--add-dir` requirement for hosted Claude models) live in `~/.claude/skills/multi-agent-orchestrator/references/agents/agy.md` — that file is now canonical; this skill only points to it. **Read mode is not enforced by agy.** Verified: `agy --mode plan` writes files with or without `--dangerously-skip-permissions`, so a plan turn can still edit its `cwd`. agent-hub enforces read mode after the fact — it diffs the git-visible worktree before and after the turn and fails the job with `errorKind:'read_mode_violation'` — so run agy read/plan turns in a disposable git worktree. (opencode's plan agent did respect read mode in testing.) ## No-MCP fallback If `agent-hub` is not registered in this session, use the wrapper script directly: ```bash ~/.claude/skills/agy-delegate/scripts/agy-run.sh \ --task "<question>" [--model low|medium|high|pro|sonnet|opus|oss] \ [--max-lines N] [--timeout S] [--write] [--raw] [--cwd DIR] ``` Aliases resolve to `gemini-3.8-flash-{low,medium,high}`, `gemi