Copilot Brain: See What Copilot
Is Actually Doing in VS Code

Copilot in agent mode is a black box with a progress spinner. It picks a model without telling you which. It reads files, runs commands and calls other agents, and you see a tidy summary of maybe three of those steps. It spends tokens and premium requests, and the first time you find out how many is when the quota bar moves. When a turn takes four minutes, you have no way to know whether it's thinking hard or stuck in a loop.

Copilot keeps detailed records of all of it. It just doesn't show them to you.

Copilot Brain is a VS Code extension that puts those records in an Output channel, line by line, as the work happens.

▶ check whether the retention jobs config is wired up
  🛠 Auto → gpt-6-luna
    💭 The config lives in the solution scripts. Checking the deploy module first.
    🔧 view C:\Projects\CWE\solution\scripts\New-RetentionJobs.ps1
    🔧 Search for retentionconfigs|asyncoperations
    🔧 bash -lc "git log --oneline -5"
    🤖 Explore (gpt-5.3-codex) · 4 tools · 12,000 tok · 8.5s
11:31:11 ✓ gpt-6-luna | cli | in 64,000 (30,000 cached) out 1,234 (512 reasoning)
    ✔ session so far · 3 premium req · 2.09 AIC

That's one turn. The model Auto chose, its reasoning, every tool call as it completes, a subagent with its own model and token count, the tokens the request actually cost split into fresh and cached, and the running credit total for the session.

What Each Line Tells You

There's a status bar item too: requests and tokens so far for the window, with a per-model breakdown on hover. Click it to open the channel.

The Part That Took the Longest to Find

Copilot doesn't offer an API for any of this, and VS Code deliberately prevents one extension from reading another's Output channel. So the extension reads what Copilot writes to disk. The surprise was that there isn't one place — there are three, and which one is live depends on which backend is behind your chat.

• The Copilot Chat log and the ccreq: debug documents behind it. One line per request with the model and latency; the documents hold tokens, context limit and credit cost.
• The chat session files in VS Code's workspace storage, which VS Code appends to as the agent works. Tool steps and per-turn totals live here.
• The Copilot CLI event stream at ~/.copilot/session-state/<id>/events.jsonl.

That third one is the one to know about. Newer VS Code chat panes — the ones with Agent / Auto / Intelligence along the bottom — don't run on the Copilot Chat extension at all. They run on the agent host, which is Copilot CLI underneath, and they write to none of the first two sources. Point a log reader at the Copilot Chat log during one of those sessions and it looks almost idle: a couple of small background calls to a mini model and nothing else, while a full agent run is plainly happening in front of you.

It's also the richest of the three. It's the only source that carries readable thinking text, the auto-routing decision, and subagent runs with their own model and token counts. Copilot Brain reads all three, so it keeps working whichever backend your chat happens to be using.

Nothing Leaves Your Machine

Copilot Brain reads files Copilot has already written on your own disk and appends text to an Output channel. No account, no API key, no network call, nothing uploaded. It's about 400 lines of plain JavaScript with no dependencies and no build step.

Install

Download copilot-brain-*.vsix from the latest release, then:

code --install-extension copilot-brain-0.0.4.vsix

Or use the Extensions view → ... → Install from VSIX. Reload the window, use Copilot Chat, then open the Output panel and pick Copilot Brain from the dropdown.

Three settings control how much it prints: copilotBrain.showThinking, copilotBrain.showTools and copilotBrain.showBackground. All default to on. Turn off the last one to hide Copilot's own housekeeping requests — the small model calls that generate chat titles and todo lists — and keep only your turns.

The Rest of the Set

Copilot Brain is the fifth VS Code extension we've built for our own daily use. They're all free, MIT licensed, and installed from a VSIX rather than the marketplace.

Copilot HUD

Your Copilot quota as a status bar meter. Usage bars for premium requests, chat and completions, and a click-through dashboard with ring gauges, a burn-rate forecast that tells you whether you'll run out before the monthly reset, and warnings at 75% and 90%. It uses your existing VS Code GitHub sign-in, so there's nothing to configure. If Copilot Brain tells you what this request cost, Copilot HUD tells you how much of the month you have left.

VSCodeCoPilotHUD on GitHub

Copilot Auto Model

The smallest of the set, and the one that started this line of work: a single status bar item showing which model Auto actually used for your last chat turn. It deliberately ignores Copilot's background requests so the display stays on the model that answered you. No sign-in, no setup, about fifty lines of code. Copilot Brain grew out of asking what else was sitting unused in the same log file.

VSCodeAutoCopilotPlugin on GitHub

KnotDo for VS Code

Link a KnotDo list to a workspace and work your tasks without leaving the editor. Tick things off, set due dates and priority, and turn a TODO comment into a real task that jumps back to the line of code it came from. Shared lists can be scoped to a single project by title prefix.

VSCodeKnotDoPlugin on GitHub

NoteSmith for VS Code

Browse and edit your NoteSmith documents in the sidebar, as a normal editor tab. Useful for the notes that belong next to the code rather than in it — deploy runbooks, API keys' locations, the reasoning behind a decision you'll question in six months.

VSCodeNoteSmithPlugin on GitHub

Check out more free tools and apps at varo.industries/apps.