Fable5 Model Routing

Fable 5 Model Routing

Entry Gate

Before applying this skill:

1. Identify the executing harness. In Codex, do not apply this skill or use it as routing policy; continue with Codex's native capabilities. Explicit invocation does not override this exclusion.
2. In Claude Code or Cursor, inspect the user's actual goal, available context, relevant artifacts, stakes, ambiguity, and missing evidence before deciding whether the skill fits. Resolve discoverable facts rather than inferring a judgment problem from trigger words alone.
3. Apply the skill only when Fable 5 is available and the remaining problem has a consequential judgment core—such as a material trade-off, hidden premise, conflicting evidence, or defensible recommendation—for which Fable 5 has a concrete advantage over direct completion by the current agent.
4. A long task, a matching phrase, or an opportunity to use multiple agents is not enough by itself.

The fit decision is the lead's inspection, not another user confirmation. If the request already names the decision and grants its scope, resolve the eligible route and proceed. Ask only when a missing user-owned trade-off changes the recommendation; continue collecting independent facts meanwhile.

This gate controls the harness that leads the workflow. It does not prevent an eligible Claude Code or Cursor run from assigning bounded support work to a configured Codex or GPT-5.6 Sol lane after the judgment boundary is stable.

Purpose

After the entry gate passes, use Fable 5 as the lead for work whose value comes from judgment, ambiguity resolution, sustained context, or a defensible recommendation. Fable 5 can also perform difficult end-to-end work; do not delegate merely to keep its context empty. Delegate when concurrency, context isolation, fresh verification, tool specialization, or a measured cost/latency advantage improves the result.

Spend scarce Fable 5 capacity at the judgment frontier: the early framing, adaptive investigation, and interpretation needed to make later work genuinely bounded. Once the decision rule, research questions, coverage bar, and action specification are stable, route large follow-on research, implementation, and verification to the strongest suitable execution lane. If new evidence breaks an assumption or reopens an ambiguous choice, return that choice to Fable 5.

This skill assigns model roles. orchestrate-subagents owns work decomposition, packets, asynchronous coordination, and synthesis mechanics.

Quick Start

1. Pass the entry gate, then state the user-visible decision or outcome and its completion bar.
2. Identify the judgment core Fable 5 must own: the trade-off, risk, hypothesis, conflict, ambiguous implementation choice, and any investigation whose next move depends on interpreting what was just found.
3. Have Fable 5 investigate until it can write a stable decision or execution specification. Read references/lane-routing.md; delegate only work whose success can be judged from that packet without inventing a missing premise.
4. Give each lane a bounded question or action, evidence contract, stop condition, and escalation trigger. Large straightforward collection or web search may start earlier when Fable 5 has defined the coverage and source bar.
5. Use fresh-context verification for consequential long runs or when the lead may be anchored to its own approach. Routine work does not need a committee.
6. Ground every progress and completion claim in tool output, inspected sources, or a named artifact from the current run.
7. Before answering, use references/judgment-gate.md. Resolve conflicts and make the recommendation; do not concatenate worker summaries.

Fable 5 Owns

• The decision rule: what evidence would change the answer.
• Issue framing, hidden assumptions, stakeholder and time-horizon checks.
• Judgment-dependent discovery: source selection, interpretation, and high-ambiguity analysis where intermediate results change the next move.
• The specification that makes follow-on research or execution bounded, including escalation conditions for evidence that reopens the judgment.
• Cross-source conflict resolution and confidence calibration.
• The final recommendation, caveat, and condition that would reverse it.

Fable 5 may also own long-context reading or implementation when keeping the work together is more valuable than parallelism. Route by task shape, not by a blanket rule that collection is beneath the lead model.

Delegate When It Helps

• Independent evidence or implementation streams can run concurrently.
• A fresh-context reviewer can test the specification or challenge anchoring.
• A support lane has materially better repository, browser, data, or execution tools for a bounded task.
• Large structured results can be reduced without fresh semantic judgment at every step.
• Follow-on research, implementation, or action has a stable specification and does not require the worker to invent a decision rule or resolve a new value conflict.
• A lower-cost lane passes the same evidence and quality bar for routine work.

Every worker returns compact evidence: answer, sources or files inspected, decisive facts, caveats, confidence, and what remains unverified.

Long Runs

Use the harness's effort and runtime controls deliberately. Important work may benefit from higher effort; routine work may be better at medium or low effort. Do not default every lane to the maximum setting without an evaluation signal.

During long runs, give sparse outcome-based updates at real phase changes. Before claiming progress, point to the tool result or artifact that proves it. Never ask a model to reproduce, transcribe, or expose private reasoning; request evidence, assumptions, decisions, and concise rationale instead.

Output Behavior

Answer as the lead's judgment, not as a committee transcript. Put the decision or highest-impact finding first, then the evidence that moved it, the main caveat, and what would change the answer. Mention lanes only when their coverage or limitations affect trust.

Reference Files

Gotchas

• Do not use model prestige as a substitute for sources, tests, or direct inspection.
• Do not outsource the decision or average worker opinions.
• Do not delegate unresolved ambiguity disguised as a broad research request.
• Do not hide a material model or tool substitution; state it when it changes confidence, cost, latency, or reproducibility.