orchestration

Use when building with the Claude API and the task fits "plan big, execute small" - a cheap/fast executor model (Haiku/Sonnet) consults a stronger advisor model (Opus/Fable) mid-generation for strategic guidance. Covers the server-side advisor tool (advisor_20260301, beta), valid model pairings, multi-turn rules, cost controls and steering prompts. NOT a Claude Code command - it is a Messages API feature.

Use when the user asks how/what/where/why about how a system works (the current repo OR a documented dependency, e.g. "how does the retry loop work", "what does X command do", "where is Y configured", "why did this trace fail"). Delegates the doc/code research to a CLEAN background subagent with an auto-picked cheap model (haiku for lookups, sonnet for reasoning) so the heavy reading never runs in the main session's (possibly huge) context. NOT for editing code, building, or multi-step tasks.

Use BEFORE asking the user any question, choice, review or approval in chat when the project has a decision inbox (its URL is in CLAUDE.md or memory, or an artifact titled "Buzón de decisiones" exists) — post it there as an item and tell the user in one line that it is waiting in the inbox, instead of writing the question in chat. Also use when agents need answers that can wait (questions, multiple-choice decisions, reviews of images/videos/audio, blind A/B tests), when the user says "buzón de decisiones", "decision inbox", "déjame las preguntas en un sitio", "pendientes para mí", or hands over a list of pending decisions; to create an inbox for a project; and to read answers, write resolutions or reply to status requests.

Orchestration mode - run the current task with the "plan big, execute small" advisor/executor pattern inside this session. The session model (Opus 5) orchestrates, verifies and approves; bounded execution goes FIRST to Codex (subscription quota, ~zero marginal cost), then to cc-delegate once Codex runs out of quota, with Claude subagents sized by the Claude-usage mode; Fable 5 is consulted up front, as a one-shot advisor, only for decisions that qualify. Use when the user invokes /plan-big-execute-small, says "plan big execute small", "patron planner/executor", "advisor pattern", or asks to run a big task with cheap executors under big-model supervision. NOT for trivial tasks (one file, one step) - just do those directly.