Senior Fable

The expensive model decides. The cheap models type.

Senior Fable is a Claude Code plugin that turns your top-tier model (Claude Fable 5) into a tech lead: it keeps decomposition, architecture, contested decisions and final synthesis — and delegates everything else to cheaper subagents. Mechanical work goes to a Sonnet fast-worker, long context-heavy investigations go to a read-only Opus deep-reasoner, and risky decisions get a second opinion from Codex (if installed).

The result: your most expensive tokens are spent only where top-tier judgment actually matters, while routine execution and codebase digging run on models that cost a fraction. Token savings without quality loss — and often with quality gains, because the messy exploration stays out of the main context window.

Why

Running a frontier model on boilerplate is like having your staff engineer format JSON. Two things are wasted:

• Tokens / usage limits. Every test file, rename and formatting pass burns top-tier budget.
• Context. Reading 40 files to find one bug pollutes the main conversation. After a compact, the reasoning that mattered may be gone.

How it works

| Work | Goes to | Why | |---|---|---| | Tests, boilerplate, formatting, mechanical edits with a clear spec | fast-worker (Sonnet) | Cheap and fast; no judgment needed | | Long digs: exploring a codebase slice, grinding logs, multi-file investigations | deep-reasoner (Opus, read-only) | Context isolation — the mess stays in the subagent | | Risky or contested decisions | codex:rescue (optional) | Independent review from a different model family | | Decomposition, architecture, reviewing results, final synthesis | the main model | This is what you pay top tier for |

Determinism is enforced, not requested:

• Least-privilege tools. deep-reasoner is denied Write/Edit at the harness level — it physically cannot modify files. fast-worker gets exactly six tools: no web, no spawning its own agents.
• Spec template. Every delegation ships as Goal / Files in-and-out of scope / Constraints / Definition of done / Report format. Subagents never see your conversation, so specs are self-contained.
• Acceptance gate. A result without the required report structure or without verification output is a failed delegation and is not integrated.
• Failure policy. Never retry the same spec verbatim; two failed rounds means the lead does it personally.
• One writer per file set. Parallel delegations must have disjoint file scopes — no edit races.
• YAGNI at decomposition. The cheapest delegation is the work that isn't needed: speculative parts are cut before routing, not built.

Install

/plugin marketplace add AndyShaman/senior-fable
/plugin install senior-fable@senior-fable

Restart the session, switch to Fable (/model fable), then activate:

/senior-fable

Give it substantial multi-step work and watch the panel: purple is the deep-reasoner digging, blue is the fast-worker typing.

Requirements & notes

• Model: built for sessions running on Claude Fable 5 (the strongest tier). On Opus/Sonnet sessions the skill deliberately does not apply — delegating "up" makes no sense.
• Codex is optional. If the OpenAI Codex plugin is installed, contested decisions get an independent second opinion; if not, that step is skipped automatically.
• Billing: on subscription plans this saves your top-tier usage limits; on API billing it saves money directly.
• Compaction-safe. The skill is deliberately small enough to be re-attached whole after a context compact, and instructs the model to re-invoke itself if delegation behavior fades.

What's inside

skills/senior-fable/SKILL.md   # the tech-lead playbook: matrix, spec template, gates
agents/deep-reasoner.md        # Opus investigator — read-only, returns distilled findings
agents/fast-worker.md          # Sonnet executor — verified mechanical work, shortest diff

Related work

• wshobson/agents — the canonical per-agent model-tier marketplace that inspired the routing idea.

Keywords

Claude Code plugin · multi-agent orchestration · subagents · token optimization · model routing · Claude Fable 5 · Opus · Sonnet · tech-lead pattern · orchestrator–worker · context isolation · AI pair programming · agent delegation

License

MIT © Andy Shaman