---
source_hash: bd48e400de04280dee2884547487854b0d5ae316186f675ddecbbfac39e95c14
post: /posts/workflows/fable-made-me-start-using-codex
shape: doctrinal
---

# Context Pack: Fable made me start using Codex

Reference for the rewritten June 2026 account. Model names and assignments describe that reported setup, not current product guidance or a permanent ranking. Only the quoted excerpts are verbatim.

## Verbatim excerpts

> A model name at the top of a diagram doesn't make the rest coordinate itself.

> Before anything merges, a fresh-context audit checks the result without the history of building it.

> The models can change. Keep the discipline around the work.

## Routing contract

One Claude/Fable seat holds the brief, context, taste and judgement. Workers execute. Classify the task before choosing the model. Briefs and results are files, with plans, audits and Cortex links available for inspection. Local models handle memory and retrieval.

The exact retained YAML has five routes: heavy-code to codex, long-video to kimi, looping-build to minimax-m3, orchestration-and-judgement to claude, and memory-and-retrieval to local-models. Gates require a fresh-context audit before any merge and a human for publication.

A looping pass starts a clean worker, reads the disk plan, executes one slice, writes the result and exits. The next pass has no previous conversation. It must inspect actual work rather than accepting a saved claim. Stop when the plan is complete or a gate stops the work. Lower cost does not establish task capability.

## Two-day evidence and failures

The seat specified this page; a Claude Sonnet worker built the diagrams and components and returned a green build without the seat opening the editor. The table retains six reported roles: Fable routing, Sonnet contained builds, Opus 4.8 plans/synthesis, Codex heavy implementation, MiniMax M3 loops/audits and Kimi large-context reading. Local memory/retrieval models are also named separately.

Two cheaper implementation passes struggled before Codex finished the next one. A separate briefing bug prevented an audit verdict from being written and triggered about ten repeated audit runs. A Sonnet worker found the cause, added a cap and wrote tests that night. The cap is THREE RETRIES then human escalation, not three total attempts.

## Boundaries and teaching surfaces

The account planned to package routing and looping into a later ARI-OS update. It does not establish that the public installer implements the full private setup. The earlier pack's shipped/implemented assertion was unsupported and is removed.

Three diagrams remain under /posts/workflows/fable-made-me-start-using-codex/. ModelMatrixFidget reproduces four fixed historical matches; it does not rank current models or assert local no-upload video processing. TaskRouterFidget illustrates assignments, not live model execution or measured throughput. Component sources are under apps/web/components/blocks.

## References

- Body: content/posts/workflows/fable-made-me-start-using-codex.mdx.
- [ARI-OS overview](/documents/ari-os/overview).
- [Replaceable model](/documents/tools-and-plugins/the-model-is-the-replaceable-part).
- [Codex loop](/documents/discipline-and-process/the-codex-loop).
- [Outcomes](/documents/thoughts-and-scribbles/stop-prompting-start-defining-outcomes).