github amosblomqvist/learnharness v0.3.6
0.3.6

latest releases: v0.6.1, v0.6.0, v0.5.0...
22 days ago

ChatGPT sign-in

Connect your ChatGPT account in the desktop app and use available ChatGPT plan models without an OpenAI API key. API-key connections remain separate. Addresses #19, with related subscription support requested in #4, and the ChatGPT portion of #60. Google and Claude subscription sign-in are not included in this release.

  • Add and switch between saved accounts, identified by email.
  • Disconnect an account to remove its saved credentials from this device.
  • Control background use with one setting shared across ChatGPT accounts. It is enabled by default; background work counts toward your ChatGPT plan usage.
  • Credentials are stored locally as owner-only plaintext JSON in chatgpt-credentials.json, alongside auth.json. ChatGPT credential storage does not use OS Keychain or Electron safeStorage.

ChatGPT sign-in requires the desktop app's private server connection. It is not available in a browser-only session or the external dev-server launch mode. Model availability depends on your account.

Context usage and compaction

Addresses #38.

  • The context meter now uses the latest completed conversation response's measured Pi usage, including input, output, and cache tokens—not the compaction summarizer's smaller request.
  • After compaction, the meter shows ? / maximum until a new conversation response provides a measurement.
  • Retained-history selection now follows Pi 1.4's cutoff logic and includes its image-size allowance, while preserving LearnHarness's educational summaries and plan continuity.
  • Compactions that would discard no active history are skipped.

Retained-history sizing remains approximate. Encrypted reasoning is not assigned an invented token count or stripped, and compaction does not guarantee a fixed post-compaction request size.

More accurate cost estimates

Also addresses accounting findings from #38.

  • Cost estimates now use Pi AI's model catalog and pricing calculator, fixing cached-input accounting.
  • New usage records preserve one-hour cache-write detail and individual background-job calls for more accurate pricing.
  • Records with insufficient historical detail remain unpriced rather than showing misleading zero costs. Some older records may therefore show an unknown cost.

Dollar figures are estimates using the currently shipped catalog, not provider invoices. ChatGPT subscription calls remain plan usage, not API-dollar charges.

Clearer provider settings

Settings now opens to Providers, organized into Model API keys, Subscriptions, Dictation, and Search. ChatGPT disclosures and account controls live in dialogs instead of crowding the main settings and setup screens. API keys remain the primary setup route.

Other fixes

  • Fixed missing Google provider dependencies in packaged desktop builds, which could leave Gemini sessions stuck on Probing. This covers the gcp-metadata failure reported in #12 and #54, and the missing-dependency portion of #23. The separate saved-error visibility concern in #23 is not claimed as resolved.
  • Research no longer requires a search-provider API key. Without one, research can still use direct URLs and page fetching; configured search providers remain available.
  • Fixed slow model loading in development: Next.js internal IPC is no longer mistaken for the desktop ChatGPT bridge.
  • Added packaged-provider initialization checks to catch missing dependency payloads before release.

Don't miss a new learnharness release

NewReleases is sending notifications on new releases.