Why it mattersSplitting planning onto an expensive model and implementation onto cheap ones targets the part of an agent bill that grows fastest, and this pattern now ships as an installable skill.
fable-orchestrator, a routing skill for OpenAI's Codex harness that has Claude Fable 5.1 plan work while cheaper models implement it, reached 555 GitHub stars in the five days since developer codejunkie99 created it on 2 September. The counts come from the GitHub REST API, retrieved 7 September: 555 stars and 102 forks, under an MIT licence.
The skill splits an objective across three models from three vendors, each with a fixed job. Fable 5.1 plans and adjudicates, and by design never writes code or owns the workspace. GPT-5.6 Luna handles normal implementation. DeepSeek V4 Flash takes loops, repeated iteration and high-throughput work. When invoked, Fable returns a bounded task graph; Codex validates the graph, starts ready workers in parallel where useful, collects their evidence, verifies the result, and goes back to Fable only when the task needs another decision.
The boundaries are stated as plainly as the roles. The repository ships no proxy, no dashboard, no model catalog and no credential store: it consumes the OpenCode Go agent routes already configured in Codex Router. If neither allowed implementation route is callable, the workflow reports the blocker rather than substituting a model. The installable part is three files, the scripts are shell, and a dependency-light test checks the routing strings and scans the package for credential-shaped strings.
The same division of labour is surfacing in other repositories this week. codex-astra-luna-orchestrator, a skill by developer donvito that uses GPT-6 Astra to orchestrate and Luna to implement within OpenAI's own lineup, was created on 5 September and stood at 262 stars two days later, per the same API. fable-orchestrator is the cross-vendor version of the idea, with the larger traction number.
The economics explain the shape. Planning happens in few tokens and benefits most from the strongest model, while implementation burns the bulk of an agent's tokens and often does not, so pushing the implementation turns onto cheaper models attacks the part of the bill that grows fastest as agent use scales. Keeping the planner out of the workspace also gives the setup a clean audit line: every code change traces to a named implementation model, and every decision to the adjudicator. The cost of adopting it is operational: three vendors means three sets of keys, three rate-limit regimes and three failure modes inside one task graph, and this skill leaves all of that to the router by design. For a team already running Codex, the traction number says enough people want per-role model routing that it arrived as community scaffolding before it arrived as a product feature.
Source
Primary source: codejunkie99/fable-orchestrator on GitHub. Star and fork counts from the GitHub REST API, retrieved 2026-09-07.
Source: GitHub
This item was written by an AI system from the linked source. Reveneau is responsible for what it publishes.
Get AI News in your inbox
New developer tools, model and agent releases, and how teams are actually shipping with them. Short, and only when there is something worth reading.