realgarit/fable-baton — explained in plain English
Analysis updated 2026-05-18
Keep a premium AI model available longer in a session by routing routine edits and file reading to cheaper models.
Enforce a consistent delegation pattern across every coding session in every repository automatically.
Reduce how often a subscription plan's top-tier model quota runs out and forces a downgrade mid session.
| realgarit/fable-baton | hurapanda/chees | jenniferzhao0531/icml2026-guide-cn | |
|---|---|---|---|
| Stars | 21 | 21 | 21 |
| Language | HTML | HTML | HTML |
| Setup difficulty | easy | easy | easy |
| Complexity | 3/5 | 2/5 | 2/5 |
| Audience | developer | general | researcher |
Figures from each repo's GitHub metadata at analysis time.
Requires a Claude subscription or API access that includes Fable 5, otherwise it falls back to Opus.
fable-baton is a plugin for Claude Code, the command line coding assistant from Anthropic. It changes how work gets split up during a coding session. Instead of one AI model doing every step itself, from reading files to writing code to checking the result, the plugin turns the top model, called Fable 5, into a conductor that hands off routine work to cheaper models underneath it. The plugin sets up four roles. A scout, running on the smaller Haiku model, handles discovery work like reading files and logs and writing summaries. An executor, running on Sonnet, handles scoped implementation such as tests and routine edits. An architect, running on Opus, takes on complex implementation, deep debugging, and reviewing what the cheaper agents produced. A verifier, also on Haiku, checks that tests pass and that changes match the plan. Fable itself is meant to stay focused on judgment and planning rather than doing the hands on labor. The project uses three layers of hooks to keep this pattern in place through a session: one that loads the policy when a session starts, one that reminds the model of the delegation rules on every new prompt, and one that counts consecutive inline tool calls and nudges the model to delegate once a threshold is crossed. The author ran a small benchmark comparing sessions with the plugin on and off, and found that total dollar cost stayed about the same or sometimes rose slightly, but the top model's own share of the work and tokens dropped noticeably. The stated benefit is for people on a subscription plan where running out of access to the top model triggers an automatic downgrade mid session, since delegating cheaper work keeps that top model available longer for the judgment calls it is best suited for. Installation is done from inside a Claude Code session using plugin marketplace commands, followed by a setup step that configures the default model. It requires Claude Code with plugin support and access to Fable 5, falling back to Opus if that is unavailable. It is released under the MIT license.
A Claude Code plugin that turns the top AI model into a conductor, delegating routine coding work to cheaper Opus, Sonnet, and Haiku agents.
Mainly HTML. The stack also includes Claude Code, Node.js, Hooks.
Use freely for any purpose, including commercial use, as long as you keep the copyright notice.
Setup difficulty is rated easy, with roughly 5min to a first successful run.
Mainly developer.
This repo across BitVibe Labs
double-check against the repo, no cap.