Blog

    November 1, 2026

    Print

    After the First Blaze

    After the first blaze repacking Muse where the work lives

    The idea in 30 seconds

    I wanted Meta's coding agent next to my files, not trapped in a terminal. So I wrapped the real harness instead of rebuilding it. Living in that wrapper for a few weeks showed me three frictions worth fixing: carrying context without re-explaining it, keeping long runs readable, and letting the tool sit where the host wants it to sit.

    Keep the center clean. Put the differences at the edges.

    The door I built

    A few months ago I asked a simple question:

    Can I use Muse the way I use Claude Code and Codex, right there in the editor?

    Meta had shipped a strong terminal agent. Muse is the harness. Muse Spark is the model family behind it. What was missing was the front door into VS Code and Cursor.

    There were three ways to build it.

    Rebuild the whole harness and call the model directly. Plug just the model into another chat. Or wrap what Meta already shipped and put a handle on it.

    I chose the third. A thin shell around muse exec: same sandbox, same sessions, same subagents. When Meta ships the next Muse-family model on that scaling ladder, the door rides the upgrade. No re-platforming.

    That was the Inception post: a Muse inside Muse. It was also a Trailcraft decision. Publish the tool, publish the method, leave something the next person can follow.

    If you want the build log, it's still on Hashnode: Inception: A Muse Inside Muse.

    This note is about what happened when I kept walking the trail I had just cut.

    What you learn by living in it

    Shipping is not finishing. Finishing is living in what you shipped and noticing where you still adjust your grip.

    Three things kept rubbing.

    1. I kept re-explaining the same files.

    The first version let me mention a file. The model then had to go read it. Sometimes it did. Sometimes it missed. Every miss was a round trip where I, the human, was the integration layer again. Telephone.

    You know that pattern if you read Finding / Founding BrainTrailz. The human carrying context between systems is integration debt.

    The fix was boring in the best way. Now when I point at something (a file, the open editors, the diagnostics in the Problems pane, the last terminal output), the extension packs it into the prompt before the model ever sees it. Compact in the sidebar as a chip. Expanded for the model as the content. No whisper chain.

    It sounds small. It removes an entire category of "you didn't read what I meant."

    2. Long runs got noisy.

    An agentic turn can touch fifty files. The first display showed every tool call the same way: open, chattery, hard to scan. Useful at five tools. Unreadable at fifty.

    The new default is quieter. A dense one-liner per tool, no chevron, no green badge on success. If there's nothing sensible to show, it collapses to a soft working… pulse. Failures still pop red and stay open.

    It's the same principle as a good trail log. If everything is highlighted, nothing is.

    3. The tool was sitting in the wrong doorway.

    My first placement put the chat on the right side. It looked clever in VS Code. Cursor rejected it outright: startup error, views dumped into Explorer. So I left clever on the floor and moved the view to the activity bar where both hosts agree it belongs. One-click pop-out into an editor tab when you want more room. A real shortcut now (⌘⇧R on Mac, Ctrl+Shift+R on Linux) instead of a menu you hunt for.

    Keep the center clean. Put the host differences at the edges.

    The pattern underneath

    If you've followed the other field notes, you'll recognize this.

    In Intent Is Not Enough the lesson was: intent tells a system what you want, not what good looks like, not where judgment belongs, not what evidence to leave behind. Context chips are a tiny version of that. The model doesn't need my intent to "look at this file." It needs the actual file. Evidence over explanation.

    In The Margin Stack the lesson was: the money isn't in the burger, it's in the combo: what happens when process work gets cheap and you keep the price while the cost collapses. The margin analog here is attention. A noisier tool log is expensive attention. A quiet one returns margin to the work.

    And from Finding / Founding BrainTrailz: one trail, N participants. The trail is the muse session: purpose, decisions, artifacts. The extension is just an adapter. When a new participant shows up, you don't redesign the center. You write another adapter at the edge.

    That's why this stays a thin shell. The harness is Meta's trail. My job is to make the handle fit the ground I'm actually standing on.

    What's in 0.5.1 if you want the list

    • Context chips: @ or +Attach for file contents, current file, open editors, Problems, terminal output. No more bare path round-trips.
    • Minimal tool rows: dense, static, silent runs collapse to a pulse, failures stay red.
    • Editor panel promoted: ⌘⇧R / Ctrl+Shift+R + sidebar pop-out button.
    • Startup status bar: Muse · session · mode from launch, not after you open chat.
    • Host fit: activity bar everywhere, folder chip only in multi-root, fewer duplicate commands.

    The full change log is on GitHub: codemarc/muse-code - CHANGELOG.

    Where this is going

    I shipped the door with itself the first time. Same story this round. The extension built the next layer of its own front door.

    That's not the point. The point is what the next person on the trail gets to skip.

    If you've been running muse in a terminal and wishing it lived next to your files the way Claude Code and Codex do, it's there now. Search Muse CLI Chat in the VS Code Marketplace or Open VSX (codemarc.muse-cli-chat), or grab the 0.5.1 VSIX on GitHub.

    And if you're reading this as a team lead rather than a builder: the shape matters more than the tool. Keep your center clean, put the host quirks at the edges, and publish the method so the next person doesn't have to cut the same trail twice.

    That's Trailcraft. Gear gets you there. Craft keeps the next person on the path.


    Marc J. Greenberg is Co-Founder and CTO of Beyond Limited Thinking and works with hospitality, proptech, and payments teams as a fractional CTO through Braintrailz, half-day strategy to 90-day engaged. He still ships. This extension is proof.

    Thin shell, real harness: one trail, N adapters