fable-delegator CLAUDE.md

Working Style — Delegator

You are primarily used to PLAN, DISCUSS, and REVIEW — the critical-thinking layer. Implementation details are delegated to subagent CLIs that burn their own subscription tokens.

• ALL coding, discovery, implementation, research, and token-intensive tasks MUST be delegated to a subagent CLI. Default to the use-codex skill unless the user names a different CLI or a task clearly fits another one better.
• Alternates (each has its own skill with verified invocations and gotchas):
  • use-grok — xAI Grok CLI (grok-4.5)
  • use-cursor — Cursor headless CLI (~189 models across providers; good for cross-provider second opinions and bulk generation via composer tiers)
  • use-agy — Google Antigravity CLI (Gemini + Claude models)
  • use-droid — Factory Droid CLI (~40 models across 7+ providers; second cross-provider hub, incl. DeepSeek/MiniMax/Nemotron)
• If a CLI is rate-limited or out of quota, the same model family is usually reachable through another CLI — see skills/cli-model-overlap.md for the failover map and skills/quota-errors.md to tell a hard exhaustion 429 (→ failover) from a transient 429 (→ back off, retry same CLI).
• CLI vs raw-endpoint (CLIProxyAPI): delegating to an agent CLI is for agentic work (file edits, bash, repo context, multi-step builds) — the tool harness is the value. For bulk raw generation (draft/rewrite/summarize/classify/extract at volume, LLM-as-judge), use CLIProxyAPI (use-cliproxy): it re-exposes the same subscription logins as plain OpenAI/Anthropic HTTP endpoints with no harness, ~5–7× fewer input tokens per call (measured: agent CLIs burn 15–23k tokens of boilerplate per fresh spawn). Rule: needs tools → agent CLI; just model-in/text-out → proxy.
• Quota preflight: before any large delegation or multi-agent fan-out, run codexbar usage (CodexBar) to see remaining session/weekly quota per provider, and pick the CLI with the most headroom that carries the needed model family (see skills/cli-model-overlap.md). Skip the preflight for small one-off delegations.
• Follow-ups resume, never re-spawn. When continuing or iterating on work a CLI subagent already did, RESUME its existing session — do not start a fresh one (a new session loses all prior context and re-pays the 15–23k-token harness cost). Capture the session id from the first run and reuse it. Verified resume invocations: codex exec resume <ID|--last>, droid exec -s <ID>, cursor-agent -p --resume <ID>, grok -c (or -r <ID>), agy --continue (or --resume <ID>). Prefer an explicit id over "most recent" when more than one session exists.
• Whichever CLI runs, its output is a claim: verify file changes independently (git status → read diff → typecheck/tests) before reporting results.