# Using Fable 5.1 as Architect + Opus 5.5 in Claude Code on a 100k+ LOC Rust repo — anyone switched their Architect to Opus 5.5 too? — retained record

> Retained reddit record from reddit:ClaudeAI.

Canonical: https://highsignal.app/data/records/87be7b32c27e026e

[← sources](https://highsignal.app/data)


source: [reddit](https://highsignal.app/data/reddit) (reddit:ClaudeAI)  published 2026-10-26 11:03 ingested 2026-10-27 02:36

[original source ↗](https://www.reddit.com/r/ClaudeAI/comments/1xftq1a/using_fable_51_as_architect_opus_55_in_claude/)

## retained text



```
I’m managing a \~100k+ line Rust codebase (GPU/wgpu desktop app, \~1,600 tests + pixel-exact golden image gates).

Right now I run a strict two-model local markdown pipeline:

1. **The Architect (Fable 5.1 in chat):** Maintains a \~130KB master spec, rules on audit findings against \~50 standing architectural laws, and cuts hyper-specific session handoffs (`§0` ground truth, `§Q` rulings, `§F` allowed file touch-list, `§W` ordered work, `§S` stop rules).
2. **The Implementer (Opus 5.5 in Claude Code):** Runs headless, pins every fix RED-first, runs 2 rounds of fresh-context subagent diff reviews (`r1`/`r2`), and outputs a structured `CCREPORT` back to the Architect. Everything stays uncommitted across \~7 sessions per block (currently 66 dirty files) until the final sitting, verified by blob hashes at open/close.

*Why I’m asking:*  
The workflow is rock solid, but because the Architect only sees the `HEAD` clone + the markdown `CCREPORT`s (not the live 66-file dirty tree), Fable occasionally makes small premise slips in the handoffs (e.g., referencing a helper removed in Session 1, or misplacing which module a predicate lives in), which Opus 5.5 then has to catch and log as a Contradiction/Deviation. Plus, reading tokens across 7 sessions are causing 1–2 context compactions per long session.

For anyone else running an *Architect → Claude Code Implementer* split on a 100k+ LOC codebase:

1. *Did you replace Fable 5.1 with Opus 5.5 for the Architect/Spec role?* Does Opus 5.5 hold long-horizon spec invariants and prose discipline as well as Fable, or does it try to micromanage the code too much in the handoff?
2. *For the reading-token / compaction problem on dirty trees:* Are you using Serena, Claude Code's native LSP, or a generated symbol/ctags map to keep reading tokens down across multi-session uncommitted diffs?

Curious to hear from people running large native repos.
```



document fetched 2026-10-27 00:23 · canonical https://reddit.com/r/ClaudeAI/comments/1xftq1a/using_fable_51_as_architect_opus_55_in_claude
