stackchain/timmy-talking-turd
1
0
Fork 0
Code Issues 41 Pull Requests 6 Actions Packages Projects Releases 15 Wiki Activity

15 Releases 15 Tags

RSS Feed
  • daily-2026-08-26 5dac04b1e6
    Compare

    Timmy review release 2026-08-26
    All checks were successful
    Quality gates / quality (push) Successful in 1m56s
    Details
    Pre-Release

    timmy released this 2026-08-26 12:22:14 +00:00 | 0 commits to main since this release

    Review release

    • Version: 2026-08-26-5dac04b1e609
    • Exact commit: 5dac04b1e6097563afac469ac399a301a7d8d86a
    • Completed ticket / PR: #11 / #66
    • Vision profile / model shown: selfhost / SmolVLM2-2.2B-Instruct (synthetic deterministic API fixture; no model weights included)

    Tangible change demonstrated

    The mobile photo flow now has separate Take photo and Choose from gallery controls. The real-app recording visibly cancels the camera picker, shows the recovery message, keeps gallery and manual logging available, selects a clearly synthetic Type-4 fixture from the gallery, and shows that analysis remains disabled until explicit consent. Touch indicators and on-screen explanations are burned into the demo.

    Verified gates

    • 67/67 unit and security tests
    • mobile UI green path
    • photo-first acceptance flow
    • camera/gallery recovery at 390×844 and 393×852
    • sleek Hermes-chat acceptance flow
    • dependency audit: 0 high vulnerabilities
    • syntax, diff hygiene, secret scan, and forbidden-artifact scan
    • actual Playwright browser recording from the running app
    • H.264/yuv420p 720×1280 media probe
    • complete MP4 decode
    • contact-sheet and full-resolution middle-frame visual QC

    Receipts

    • Source SHA-256: ececea0418fddeaf69c525f9d08b63ff51064df15ec31eedc6dff7427c2e068f
    • Demo-video SHA-256: e6cd5c865c4f5b570769885b39caa1036d1077f568e6dc0882ac3574cca0e099

    Safety / accuracy boundary

    Timmy may suggest only visible Bristol form, broad color, and image quality with bounded confidence and abstention. It does not diagnose disease or infer symptoms; user confirmation remains mandatory. Known limitation: open-weight VLM integration is exercised, but Bristol-form accuracy is not clinically validated. The demo uses synthetic data and deterministic responses, not a live model-quality benchmark.

    Downloads