| name | orchestrator |
| description | Chooses which model runs a delegated task (subagent, parallel fan-out, or Workflow stage) and how to invoke it. Use whenever spawning agents with the Agent tool, authoring or running a Workflow, delegating to a subagent, dispatching bulk/mechanical work, requesting a review, or otherwise deciding which model should do a piece of work. Trigger on "which model", "spin up agents", "fan out", "run a workflow", "delegate this", "have gpt/codex/fable/opus do X". |
Orchestrator — model routing
Pick the model per task, not per session. You are the orchestrator; the model
you assign to a subagent/stage is a first-class decision. Route by the table
below, then invoke via the mechanics section.
Roster (higher = better on every axis)
| Model | Cost | Intelligence | Taste | Invoke as |
|---|
| gpt-5.5 | 9 | 8 | 5 | codex plugin (see below) |
| fable-5 | 2 | 9 | 9 | model: "fable" |
| opus-4.8 | 4 | 7 | 8 | model: "opus" |
| sonnet-5 | 5 | 5 | 7 | model: "sonnet" |
Cost = what I actually pay (OpenAI's limits are generous, so gpt-5.5 is
effectively cheapest). Never use Haiku.
Route by task shape
| The task is… | Send it to |
|---|
| Bulk / mechanical / clear-spec impl, migrations, data | gpt-5.5 — cheap, token-efficient, smart enough |
| User-facing: UI, UX, copy, API design (needs taste ≥ 7) | fable-5 (best taste), else opus-4.8 |
| Hardest unsupervised reasoning (fuzzy spec, deep debug) | fable-5 (intel 9), or gpt-5.5 (intel 8) |
| Review of a plan or implementation | fable-5 or opus-4.8; add gpt-5.5 for an independent 2nd read |
| Anything else / catch-all | opus-4.8 |
Decision rules
- Defaults, not limits. You have standing permission to escalate: if a
cheaper model's output misses the bar, rerun or redo it on a smarter model
without asking. Judge the output, not the price tag — escalating costs less
than shipping mediocre work.
- When axes conflict for anything that ships: intelligence > taste > cost.
Cost is a tie-breaker only.
- Taste gate: never route user-facing work (UI/copy/API design) below taste 7
— that rules out gpt-5.5 for those.
- Reviews want a second, independent perspective. A fable/opus review plus a
gpt-5.5 pass catches more than either alone.
Invocation mechanics
Claude models (fable-5, opus-4.8, sonnet-5) — pass model on the call:
Agent tool: subagent_type + model: "fable" | "opus" | "sonnet".Workflow: agent(prompt, { model: "fable", ... }), or model on a phase.
gpt-5.5 → the codex plugin (do NOT hand-roll bash wrappers; use the plugin):
/codex:review [--base <ref>] — read-only quality assessment / branch review./codex:adversarial-review <focus> — skeptical design review (auth, tradeoffs,
reliability); append focus text to steer it./codex:rescue — subcontract active debugging, multi-file refactor, or an
implementation loop when a second pass is needed.- Async:
/codex:status, /codex:result, /codex:cancel (with --background).
- Inside a Workflow/subagent, delegate gpt-5.5 work via the
codex:codex-rescue agent type or the plugin's native slash commands /
codex-cli-runtime skills — never raw terminal wrappers.
- Fallback: if the codex CLI/plugin isn't installed in this project, substitute
a fable-5 (or opus-4.8) agent for the gpt-5.5 slot and note the swap.
Closed-loop QA (optional)
Keep the review gate on with /codex:setup --enable-review-gate — a stop hook
challenges Claude's output with Codex before finalizing, so broken code or weak
assumptions don't reach the main session unvetted.
Worked example — a 3-stage feature
- Plan the change → opus-4.8 (or fable-5) subagent.
- Implement the clear-spec pieces in parallel → gpt-5.5 via
/codex:rescue or codex:codex-rescue agents.
- Review the diff → fable-5 agent + a gpt-5.5
/codex:review pass;
reconcile both. Escalate any rejected fix to a smarter model rather than
patching blind.
Credit: the model-routing rubric — scoring delegatable work on cost,
intelligence, and taste, and routing each task to the cheapest model that clears
the bar — is adapted from Theo Browne (@t3dotgg).
