The smartest model runs the show. Cheaper models do the typing.
Claude Code lets every subagent run on a different model — and lets the session itself run on a different model than its subagents. This plugin exploits that with the architect pattern: your session runs on Fable 5, Anthropic's most capable model, acting as a full-time architect. It owns requirements, decomposition, specs, and verification — and routes every implementation task to the cheapest adequate lane:
| Lane | Producer | Invocation | Route here when |
|---|---|---|---|
| Routine | Sonnet | implementer agent (default) |
|
| The spec fully determines the outcome | |||
| Subtle | Opus | implementer with model="opus" |
|
| A Sonnet miss is expensive — concurrency, security, hard debugging | |||
| Cross-vendor | GPT-5.5 | codex-implementer agent |
|
| Correctness-critical, or you want a non-Anthropic second implementation | |||
| Judgment | Fable 5 | fable-advisor agent |
|
| Commitment boundaries — see below |
Tokens route by volume: the expensive model emits the fewest tokens (judgment and specs), cheap models emit the most (code). Implementation mechanics are ~90% of a session's tokens and Sonnet handles them at near-parity — so this runs roughly 60% cheaper than Fable-for-everything, at higher quality than Sonnet-for-everything.
The plugin ships the orchestration skill — the routing doctrine that teaches the session when to use each lane, the cost discipline that keeps the expensive model's own token volume minimal (emit judgment not volume, keep context lean, reason once then hand off), the five-part spec contract that makes context-free delegation safe, and the verification rules that keep cheap lanes honest.
claude plugin marketplace add DannyMac180/fable-advisor
claude plugin install fable-advisor
Then start your session as the architect:
/model fable
Lite mode — one file, 30 seconds. Don't want the full pattern? Copy agents/fable-advisor.md into
~/.claude/agents/
and keep your session on Sonnet. You get advisor consults at commitment boundaries without the orchestration layer (see "Advisor-only mode" below).Claude Code ≥ 2.1.170 with a subscription that includes Fable 5 (Pro, Max, Team, or Enterprise — all current consumer plans qualify).No Fable access(e.g. API-key billing)? Use/model opus
for the session and changemodel: fable
→model: opus
in the advisor file. Same pattern, model tiers shift down one.Codex lane (optional): thecodex-implementer
agent needs theOpenAI Codex CLIinstalled and authenticated (npm i -g @openai/codex
, thencodex login
). Without it the agent reportsSTATUS: unavailable
— the other lanes are unaffected.- Heads-up: if a pinned model isn't available on your account, Claude Code silently falls back to your session model — the pattern degrades quietly rather than erroring. If results feel unremarkable, check your plan. (This quiet fallback applies only to Claude model pins — the codex lane always fails loudly with a structured error.)
Model resolution order in Claude Code: CLAUDE_CODE_SUBAGENT_MODEL
env var → per-invocation model
parameter → agent frontmatter → session model.
With the session on Fable, just ask for work — the orchestration skill routes it:
Add rate limiting to our public API. Design it, delegate the
implementation, and verify the evidence before you call it done.
The architect writes the spec, picks the lane (rate limiting touches concurrency — likely implementer
with model="opus"
), reads the diff and verification evidence when the report comes back, and only then reports done.
To make the doctrine always-on, add one line to your project's CLAUDE.md
:
You are the architect running the most expensive model — minimize your
own token volume. Delegate all implementation through the orchestration
skill's routing table (never type code yourself), delegate broad codebase
exploration to cheap read-only agents, and verify evidence before
accepting any lane's report.
Even the architect gets a second opinion. The fable-advisor
agent is a read-only skeptic — consulted before architecture decisions, migrations, API designs, and whenever a problem has resisted two attempts. It reads your actual code and returns a verdict in under 300 words. It never implements. Running it from a Fable session still pays: it sees the code fresh, without your conversation's accumulated assumptions.
The inverse arrangement, for when you'd rather keep the session cheap: run the session on Sonnet and consult fable-advisor
only at commitment boundaries.
Migrate our checkout sessions from Postgres to Redis — plan it,
consult your advisor before committing, then implement.
A typical consult costs cents. To make it automatic, add to your project's CLAUDE.md
:
Before committing to any architecture decision, migration, or refactor
touching 3+ files, consult the fable-advisor agent and act on its verdict.
Is this Anthropic's "advisor tool"? No — that's a server-side API feature. These are plain Claude Code subagents plus a skill: readable, editable, no beta flags.
Does this work on claude.ai? No — subagent model routing is Claude Code only (CLI, desktop, VS Code, web).
Why not just run everything on Fable? You can. It's excellent. It's also ~3× the per-token cost of Sonnet, and most of a session's tokens are implementation mechanics that Sonnet handles at near-parity. Spend the premium where judgment lives.
Upgrading from v1? Nothing breaks: the fable-advisor
and implementer
agents are unchanged, and advisor-only mode works exactly as before. v2 adds the orchestration skill and the codex lane, and flips which pattern the docs lead with.
Why a GPT-5.5 lane in a Claude plugin? Vendor diversity. Models from one family share blind spots; an independent implementation from a different lineage catches what same-family review misses. The architect stays Claude — the lane is a producer, not a judge.
MIT