Three skills that turn Fable 5 into your coding orchestrator

Fable 5.1 or Opus 5.5 leads Foreman v0.6.4 lets the lead be Fable 5.1 or Opus 5.5. The lead plans the work and personally accepts the result.

Stay on Fable 5 Fable Safe Prompt rewrites a task in clear, neutral language so Fable 5's cybersecurity and biology safety checks do not misread ordinary work and quietly move you onto Opus 4.8 and leave you there.

Optional Codex workers If the Codex CLI is installed, Foreman can hand coding tickets to it. Without it, Foreman routes those tickets to other workers. A tiny test call says whether Codex or Grok can really do work, not just whether you are signed in. The verdicts are LIVE, SIGNED OUT, or BALANCE EXHAUSTED.

Route from one dated table One dated evidence table compares Claude Opus 5.5, Fable 5.1, OpenAI GPT-6 Astra, Sol, and Luna, and xAI Grok 4.7. One command on your machine shows which model gets each job, based on what is actually installed, signed in, and approved on your computer. Clear-cut jobs get a fixed default. Mechanical edits go to GPT-6 Luna, about a tenth of the cost of Claude Haiku. When two good options are close, the lead asks you once, recommends one, and remembers your answer for the rest of the session. An unattended run uses a sensible default and tells you what it chose.

Ask twice before expensive helpers GPT-6 Astra and Claude Fable are hired as helpers only after you say yes twice in the same session. If Fable 5.1 itself is leading your session, that's fine; the rule is about the helpers it hires.

Use Jev only with your own key Jev is a decision model, not a chatbot. With your own TypeSafe or OpenRouter key, the lead uses it for cheap, narrow sorting jobs: merging duplicate review findings, and checking whether cited evidence really supports a finding. That costs a tiny fraction of a normal AI call. Jev only reorders what the lead reviews. It never accepts work or makes security decisions. It starts in a trial mode, and the skill works the same without it.

Check the work before it is done The lead runs the real tests. An independent reviewer, always Claude Opus, checks behavior-changing work. When Claude built the change, a different AI family reviews it when Codex or Grok is available. The lead personally accepts the result.

Checked on real Claude Code runs A behavioral test suite runs real Claude Code leads through 15 scenarios: farming work out, review by a different AI family, following your instructions, recovering when a provider runs out, and asking before expensive models. It grades their actual tool calls against hidden answer keys. Every scenario passed on its most recent run. Those were single runs, not a statistical rate. In a blind test, an AI model answered 12 routing questions using only the skill's instructions: with the previous version it got 6 right, and with v0.6 it got all 12.