Build Mirror

Claude Code writes the code. Build Mirror tells you what it just did, in words you already know.

A panel sits beside your editor and watches the Claude Code session running in this folder. Every time Claude finishes a turn, a card appears explaining that turn: what changed, why it matters, how risky it is, and which files it touched. Anything unfamiliar in the explanation is underlined, and hovering it gives you the plain meaning.

It is for the person shipping real software with an agent who does not want to end up with a codebase they cannot read.

What a card tells you

• Summary. What this turn actually did, one or two sentences.
• Why it matters. What it means for your project, not for computers in general.
• Risk. Low, medium or high, with the few words that justify it. Read off the transcript by plain rules first, so the risk line cannot be talked down by a model having a good day.
• Files. Every file the turn wrote to, taken from the transcript itself rather than from the model, so it never names a file that was never touched.
• How it came together. The decisions inside the change, one point per line.
• Words. The terms in the explanation, each with what it means here.

You can also ask a follow-up question on any card, and the answer gets the same treatment.

Interrupting Claude halfway through is ordinary use, so it is handled as such. What you type while it is still working counts as part of the same turn rather than starting a new one, so the actions taken before you cut in stay on the card, and your correction is explained next to the request that started it.

The whole session, on request

Once there are two turns to join up, a button appears offering to explain the entire session as one piece of work: what it set out to do, where it actually got to, the highest risk it reached at any point, every file it changed and what each is for now, and a section for what is still unfinished. That last one is the reason to open it. Work started and abandoned, a step cut off when you interrupted, a check nobody ran, a decision handed back to you and never made.

It is a button rather than something that happens on its own, because it costs one request every time you press it. Reading a long session costs more than reading a turn, so this is the most expensive thing in the extension and the only one that waits to be asked.

Alerts, the moment they happen

Some actions are worth catching while the agent is still working, not after. Deleting things, publishing, deploying, installing packages, anything touching a key or a password. Build Mirror spots these with plain rules on your own machine, so they cost nothing, need no API key, and arrive instantly. Set buildMirror.alerts to important if you only want the hard-to-undo ones.

Getting set up

1. Install the extension. The Build Mirror icon appears in the activity bar.
2. Choose who writes the explanations in Settings, under Build Mirror:
   • OpenAI, Gemini, DeepSeek, Groq, OpenRouter and anything else that speaks the same API. Set buildMirror.baseUrl and buildMirror.model.
   • Ollama, a model on your own machine. Free and private, but a small model gives noticeably weaker explanations.
   • Anthropic, the best explanations and the highest cost.
3. Run Build Mirror: Set API Key from the command palette. Not needed for Ollama.
4. Open Claude Code in the same folder and work as usual.

buildMirror.instructions is worth a minute of your time. Tell it what you already know and the explanations land at your level, for example I write Python but have never touched TypeScript or I have shipped software for ten years, skip the analogies and be blunt.

Requirements

• Claude Code, with a session open in the same folder as your VS Code window.
• An API key for one of the providers above, or Ollama running locally.

What leaves your machine

Build Mirror reads Claude Code's own session transcript from ~/.claude/projects, and sends the parts of each finished turn that need explaining to whichever provider you chose. That includes your prompt, Claude's reply, and the code the turn wrote. Nothing else is collected, there is no telemetry, and there is no server of ours in the middle.

Your API key is kept in VS Code's secret storage, not in settings, so it never lands in a settings file you might commit or share. Build Mirror: Remove API Key deletes it.

If you would rather nothing left the machine at all, set the provider to Ollama. The alerts keep working either way, since they never needed a model.

What it costs

One request per finished turn, plus one per follow-up question you ask, plus one each time you ask for the whole session. A cheap model like gpt-5-mini is a few cents a day of ordinary work. Each extra mirrored session is explained separately, so two open mirrors cost about twice as much as one.

Several sessions at once

Running more than one Claude Code session in a folder is normal, and each one gets its own mirror. Build Mirror: Open a Mirror for Another Session opens one beside the first, up to four at a time. buildMirror.autoOpenNewSessions decides whether new sessions ask first, open straight away, or are ignored.

Known limits

• Explanations arrive after a turn finishes, not while it runs. The alerts are the live half.
• Build Mirror sees Claude's changes, not your own. If you edit a file while Claude is working on it, nothing here will tell you the two of you collided.
• A session summary covers the last 60 turns. Past that it says which turns it left out rather than pretending to cover them.
• Only the first folder of a multi-root workspace is watched.
• Nothing on disk says whether a Claude Code session is still open, so a very old transcript can occasionally be offered as though someone were still using it.

Licence

MIT.