What leaves your machine Diagram

The short answer

• Only two calls leave the machine
• The changed files (minus excluded ones) and the request go to DeepSeek
• Single claims with the code they cite go to Anthropic

Not shown here

• The replay scripts, which also call Anthropic but only on benchmark data
• Connected but out of scope: Hook entry, Router, Project tests, Ranker, Delivery, polywatch CLI

What it guarantees

• Excluded files never leave the machine: .env, keys, *secret*, *credential* and node_modules are dropped before the reviewer prompt is built

What leaves your machine

Which parts of my code are sent to which provider?

A code reviewer is a data flow. Anyone installing polywatch on a private codebase needs to know exactly what is sent where before the first turn is reviewed.

The long read

Everything inside the box runs locally. The turn log, the settings, the results and the ledger are plain files in the project's .polywatch folder.

The first crossing is the review. The worker builds a prompt from the changed files and the user's request. Files matching the exclusion list (.env files, keys, anything named secret or credential, node_modules) are dropped before the prompt is built, and each file is capped at 60,000 bytes. The prompt goes to DeepSeek by default; the settings can point it at Anthropic or at a local OpenAI-compatible server instead.

The second crossing is confirmation. For each claim it checks, the worker sends Anthropic the claim, the cited file or the windows of it around the names the claim mentions, and the request when it is under 20,000 characters.

Nothing else leaves. The ledger of outcomes, which is what calibrates the ranking, never does.

Terms used here

turn: One round of Claude Code work: the user's prompt and everything Claude does in answer, up to when it stops.
tier: polywatch's label for how hard a change is to get right, from its size and whether it touches concurrency, state or protocol logic.
claim: One specific, checkable statement a reviewer makes about a defect: which file, where, and what is wrong.
finding: A claim that survived confirmation (it was not refuted), with a score. Confirmed findings carry Opus's evidence.
confirmed: Opus checked the claim against the code and said it holds.
refuted: Opus checked the claim against the code and said it does not hold. Refuted claims are never shown.
ledger: An append-only file of every review and every outcome, used to calibrate the ranking.
outcome: The user's verdict on one finding after the fact: real bug or false alarm.