Skip to content

Releases

Whetstone 0.1.0

Choose a tag to compare

@boffti boffti released this 28 Oct 20:19

The first public release. Whetstone is a private practice space that runs on your machine. Your own AI agent writes the questions for you; real tests grade them.

Practice

  • A question page with a lesson, the problem, a four-rung hint ladder, a Monaco editor and pytest grading. Hints open one test run at a time; the reference opens after a solve or a confirmed give-up.
  • Question types: build, predict, fix, review, tests (write the tests; Submit runs them against hidden buggy versions), design (a design note, then code for the guarantee it declares) and whiteboard system design. examples/ has a finished question of each type.
  • Predict then explain: guess which tests pass before a run and before Submit, then explain one hidden case in a sentence after a solve.
  • Spaced review: a solved idea comes back as a variant on a new surface after 3 days, 3 weeks, 3 months, 6 months and a year. Mastery is a clean solve plus one passed review.
  • Today: your next step, one alternative and the next review, then your focus, each skill with an edge that dulls with time. A Progress page has the calendar, honest numbers, milestones and records; the game layer rewards mastery, never volume, and can be hidden.
  • Built for a side pane next to your agent: in the Code view, the problem's sections open as popovers over the editor. It also works full screen and on a phone, in light and dark.
  • One tap after Submit says how a question felt (too easy, about right, too hard, felt artificial); your agent reads it when it writes the next one.

Your agent

  • Whetstone has no AI of its own. Your agent (Claude Code, Codex, Cursor or another) drives the bin/whetstone CLI on your own subscription, guided by AGENTS.md and eight skills: profile, curate, author, agent, tutor, pair, review and interview.
  • Prepare buttons queue requests that your agent picks up while it watches (whetstone watch, whetstone request claim|done|fail). Buttons show queued, preparing and Start; a stopped agent's claim can be cancelled or retried. Prepare all queues one module at a time, in path order, so the first is ready while the rest are built.
  • Prepare asks what a question is for: your own work (from your profile), a company you are preparing for (its own interview reports lead the brief), or a domain. Prepare again for… reframes a question you already have as a linked variant; the original keeps its progress.
  • Live pair: your agent watches your session and replies in the editor's margin, at the line that matters, as a quiet reviewer, a coach or an interviewer.
  • whetstone today starts a session with the reviews due, the question in progress and each path's next step; it can run from a Claude Code SessionStart hook.
  • whetstone-profile drafts your profile from what your agent already knows and, with consent, your GitHub activity, including what your AI co-writes.

System design

  • A Design page of the design questions companies ask, most asked first. Draw on an Excalidraw whiteboard; your agent reads the drawing as a Mermaid diagram and grades the design against a rubric, quoting it.

Catalog and paths

  • 2,771 paraphrased interview reports from 397 companies, organised into 271 modules, each with links to its sources. Questions record which reports they are grounded in.
  • 78 learning paths: systems to build, modelled on big-tech systems, across 11 domains. Filters, a sort, and a route map per path: point at a stop for its module, pick it to see every path that uses it.
  • Company logos (Simple Icons, CC0) and domain icons, served from your machine.

Privacy and security

  • Everything stays on your machine: your profile, questions and progress live in the git-ignored my/ folder. No account, no telemetry.
  • The server answers only your own machine (local Host and Origin checks against DNS rebinding and cross-site requests).
  • Test runs get an allowlisted environment (no ANTHROPIC_API_KEY), their own process group, and in Docker a memory cap, a process cap and a read-only content/. A failed hidden test shows only its message, never the expected values.
  • Dependencies are pinned by hash (uv.lock), images by digest, and CI actions by commit SHA; CI runs tests on Python 3.11 to 3.13, CodeQL, zizmor and Scorecard.

Quick start: see the README. Whetstone has no AI of its own: your agent (Claude Code, Codex, Cursor or another) does the writing, on your own subscription.