dsh-receipts

You have been using an AI coding assistant for months, yet you cannot clearly say what you actually built, how much it cost, or what it delivered. dsh-receipts turns the session logs sitting on your local machine into a structured usage-and-impact report. It reads DeepSeek Harness transcripts in pure local mode (Zstd-compressed logs supported), cross-references git history across your repositories, and produces Markdown daily / weekly / monthly reports plus a self-contained single-file HTML receipt that opens offline and can be shared with a manager or teammate without any setup on their end. The entire mining pipeline is zero-network, zero-config, and zero-model-call by design; the only optional LLM pass is a single under-20-KB polishing step, keeping cost negligible.

Reports can be sliced by period (7 to 365 days), filtered by repository substring, and optionally redacted to mask absolute paths. The plugin enforces strict privacy boundaries: it reads only your dshHome directory, writes only to your output folder, and ships with network-whitelist test assertions so that no data ever leaves the machine.

It is a natural fit for solo developers who need to justify AI tooling spend to stakeholders, engineering leads who want periodic retrospectives on how much real work the assistant actually unblocked, and any engineer who wants a reproducible, auditable paper trail of AI-assisted development. Topic classification is deterministic keyword clustering rather than semantic analysis, so results are stable, explainable, and easy to extend. The project is MIT-licensed, fully self-implemented, and ships with comprehensive test coverage including multi-frame Zstd regression tests and offline HTML self-containment verification.