Fable Week: How I Used Anthropic's New Model to Overhaul Agent Vibes

A couple of weeks ago, Anthropic gave me early access to their new Fable model. Most people in that situation would spend the tokens on demos — a few impressive outputs to show off. I decided to do something different with the access.

I pointed it at the entire AgentVibes codebase. Every file. Every script. I asked it to find everything that was fragile, duplicated, or quietly broken — and to understand the codebase deeply enough to help me redesign the parts that had been accumulating debt since the first version.

That week of work shipped as v5.12.0 — “The Fable Week Overhaul.” Here’s what actually happened.

---

The Problem Fable Found Immediately

AgentVibes has to make a lot of decisions every time it speaks. Which voice to use. Which engine. Whether to play audio locally or route it to another machine over SSH. Background music. Volume. Mute state. That logic had originally lived in one place — and then, as the project grew to support Mac, Linux, Windows, remote audio, and a standalone voice server, the same logic got copied into four separate scripts written in four different styles.

Those copies drifted. A fix made in the Mac/Linux script was missed in the Windows script. A routing fix in the remote audio script didn’t make it to the voice server. Same family of bugs, showing up in different places, never fully exterminated because fixing one copy didn’t fix the others.

I knew this intellectually. But the codebase had grown large enough that the full picture — how exactly the four copies diverged, what each one handled differently, where the actual gaps were — lived in my head imperfectly.

Fable’s advantage here is context. It can consume the entire repo at once and hold it in working memory. I wasn’t describing the problem to it; it was finding the problem itself, then explaining it back to me more precisely than I’d been thinking about it.

---

The Workflow: Fable Analyzes, Claude Code Builds

The process was deliberate. I used Fable for what Fable is genuinely good at: deep analysis of an existing codebase. Feed it the whole thing, ask it to map the duplicated logic, understand the divergence points, and produce a clear specification for what a unified core should look like.

Then I handed that specification — and the full context Fable surfaced — to Claude Code to actually implement.

This isn’t two models competing. It’s two models doing what each does well. Fable spent the analysis phase. Claude Code spent the build phase. The combination is faster and more reliable than either alone — Fable’s whole-repo comprehension paired with Claude Code’s iterative, tool-assisted implementation.

By end of the week I had a real redesign, not just a patch.

---

What Got Built: One Brain for Everything

The result of the week is what we’re calling the shared core — a single module that makes all the voice routing decisions, with every other part of AgentVibes deferring to it.

One place to fix. One place to trust.

The immediate wins that came out of this:

Kokoro voices that were silent on Linux now work everywhere. This was a routing inconsistency between the Linux script and the voice server. With one shared core, both follow the same path. Fixed.

Your voice choices stick. Previously, the assistant would sometimes reset voice settings while repeating configuration back to you — overwriting your preferences with its own echoed values. That’s the kind of subtle drift that happens when multiple copies of the same logic have different assumptions about who’s authoritative. One core, one authority.

Volume, mute, and remote playback behave consistently across Mac, Linux, and Windows. Previously each platform had its own interpretation of what “muted” meant in edge cases. Now they all ask the same brain.

Safe by default. If the shared core isn’t available on a given machine for any reason, AgentVibes falls back to the old behavior — so voice never just stops working. We didn’t want the overhaul to create a new category of failure.

---

Previews Go Where You’re Listening

One behavior that was quietly wrong: voice and music previews always played on the local machine, even when you’d configured AgentVibes to route all audio to a remote receiver. If you’re SSH’d into a server and your speakers are on your laptop, previewing a voice meant hearing nothing.

That’s fixed in 5.12. If you have SSH remote configured, previews play on your receiver. If not, they play locally. This covers voice previews from Setup, Agent, and Settings screens, and music/track previews from the background music picker — Space to play, Space again to stop.

This was a small thing that was wrong everywhere, and fixing the shared core fixed it in every place at once.

---

A Simpler Voice Menu

The Voices tab had a redundant entry point — a separate listing that duplicated what was already in Setup. People would land there, get confused about where to actually make their choice, and end up in the wrong place.

Removed. Voice selection for any provider now lives in Setup exclusively. One place, clear navigation.

---

Reviewed Before Shipping

After the week of building, I ran three independent reviews over the changes — security, correctness, and regressions. Real reviews, not “does it run.” Every genuine issue that surfaced got fixed before release. The alpha went out first to let the community find anything the reviews missed. The stable release followed when nothing was found.

If you were on the alpha — thank you for the feedback.

---

What This Means for the Project Going Forward

The shared core is also groundwork. One of the new things it passes through is the full project path of whatever folder a message came from — not just the project name. That field is there for upcoming enhancements that depend on knowing exactly where work is happening, not just what it’s called.

No promises on timeline, but that data is now there.

---

Get the Update

If you already have Agent Vibes installed:

npx agentvibes update

Starting fresh:

npx agentvibes install

Source is at github.com/paulpreibisch/AgentVibes. Apache 2.0.

If you haven’t read the original post on what Agent Vibes actually does — how the remote receiver works, what BMAD party mode sounds like with 12 agents each speaking differently — it’s here: Agent Vibes: Finally, Your AI Agents Can Talk Back.

---