Skip to content

Qualify signed-in Kinetic mailbox and model journeys after deployment #54

Description

@sarthakagrawal927

Current queue disposition — 11 September 2026

Real Google OAuth, mailbox search, semantic indexing/search and reload persistence passed. Account switching/token refresh, compatibility-host and broader sender/digest/ranking qualification remain follow-ups.

This remains an open repository backlog issue. Its outstanding acceptance criteria are not marked complete. The owner-approved minimum-use review has moved it out of the immediate project queue; the original requirements below and existing README references are preserved. To resume, remove queue: excluded and unarchive the project card.

Completed release work — 8 September

  • Rerun the isolated browser mailbox workflow: real account-scoped IndexedDB, synthetic mail/embeddings, distinct account cursors, colliding IDs and delayed prior-account responses.
  • Publish the checked account-isolation repair from 6a4eb83. Worker version 99493f23-22e2-431e-9ae3-78cd2e732063, deployment d6eae844-69a4-4df3-a0ac-6aaa2e5d7a9f, 100% traffic. Prior production was 274233b.
  • Correct current deployment runbooks to pnpm run deploy; bare pnpm deploy invoked pnpm's unrelated built-in command and failed before deployment.
  • Reproduce a live discovery-routing regression: GET auth session and mailbox routes returned discovery 404s; advertised OpenAPI returned asset 404. Three exported-Worker tests failed on original source and pass with 210e6ad. Product APIs now bypass public discovery; discovery endpoints are handled before static assets.
  • Full local quality and docs pass after routing repair: 103 tests, unchanged coverage/complexity gates, 45 Markdown files checked.
  • Exact-source CI 34223394319 passed. Published 210e6ad as Worker c8900fd9-0d92-4973-b2dd-7e1509427081, deployment 1cd0620c-d582-4262-a00c-070fda259190, 100% traffic. In a separate logged-out browser context: health 200, auth session 200, mailbox 401, agent catalog 200 and OpenAPI 200 (3.1.0). Canonical public terms also renders correctly at 390px.

Remaining qualification

  • Owner-assisted real Google OAuth completion and account switching. The canonical sign-in button reaches Google's account form; this does not prove callback completion. Owner sign-in is pending in the open Kinetic tab.
  • Read/search/sender/digest journey on permitted mailbox data.
  • Client ONNX ranking quality and model download behavior.
  • Verify compatibility-host and canonical OAuth callback completion.

Gmail remains read-only. No compose, draft, reply, archive or delete features. No private messages or unsubscribe requests were accessed/performed during this release check. Synthetic local cache/API tests do not qualify live mailbox or model quality. Keep shareability false until the required evidence exists.

Actual browser model evidence — 8 September

  • Run the actual MiniLM fp32 model on synthetic messages in an isolated Chrome browser. Model file 90,387,606 bytes; 384 finite normalized dimensions.
  • Reproduce application CSP blocking the current HF model redirect and pinned ONNX runtime. Source repair 9e4cca9 permits the observed model host, exact runtime CDN path, and cached blob module factory. The new pnpm verify:browser-model passes with zero CSP violations. Full local quality passes (104 tests); docs pass (46 files). Exact CI 34225523012 passed; release and live verification completed below.
  • Compare BGE small q8 on the same eight synthetic messages and ten queries: intended target first for 10/10, versus MiniLM missing three flight queries. BGE model file 34,014,426 bytes. This tiny corpus is diagnostic evidence, not broad ranking qualification. Receipt and interpretation are in docs/knowledge/learnings/browser-model-2026-09-08.md.
  • Publish 9e4cca9 at 100%: Worker c94fa462-d536-4080-a12a-f51976d25e90, deployment c8acbcec-3c3e-4e33-884e-a871aaf4d56e. Live app policy verified, health/session 200 and mail 401. Deployed Transformers bundle plus app MiniLM fp32 settings produced 384 finite normalized dimensions from synthetic text in 6195 ms with zero CSP violations. Model disposed and test tab closed. This does not qualify signed-in indexing or real mail.
  • Qualify broader retrieval examples before replacing MiniLM. The 9 September 16-message/18-query diagnostic scored MiniLM 16/18 and BGE 17/18; MiniLM remains current.
  • Pin existing MiniLM revision and persist matching model/config identity; exclude legacy/foreign/malformed/sparse vectors, preserve emails and stale vectors as pending, and validate before reindex writes. Actual IndexedDB sync/reindex/reload and 124 tests pass. Released 24142fe; receipts below.
  • Remove the implication that cosine similarity is a calibrated match probability. Released 24142fe uses a numeric Similarity label.

No model replacement or mailbox access occurred in this check. Shareability remains false.

Indexing cancellation repair — 8 September

  • Reproduce four completion-accounting regressions: pre-cancelled and mid-inference cancellation claimed a whole batch, cancellation during persistence miscounted completed work, and zero-sized batches claimed no pending mail.
  • Repair 451b699 counts successful writes only, checks cancellation before model loading and after inference, and leaves unprocessed mail pending. Six focused tests plus actual browser IndexedDB cancellation/retry pass. Full quality passes 110 tests with unchanged gates; 46 docs pass.
  • Published 451b699 after CI 34226441025 passed. Worker b42e2e27-b358-4ee7-a338-f9cc9d4cbc69, deployment b2eb7176-344c-46f6-9a62-3ac37e6281e0, 100% traffic. Fresh guest browser loads the new index-QyIj82F-.js bundle; app/health/session 200 and mail 401. This is separate from remaining ranking and owner-login qualification.

Activity

  1. sarthakagrawal927 commented on Sep 9, 2026

    @sarthakagrawal927
    MemberAuthor

    Bounded embedding-cache repair — 9 September

    Fresh isolated Chrome, actual models and IndexedDB: MiniLM ranked intended messages first for16/18 queries; BGE17/18 on16 synthetic messages including cancelled-flight, paid/unpaid invoice and promotional distractors. This diagnostic does not justify replacing MiniLM. A deliberately foreign-tagged same-width cached vector was nevertheless returned by actual search with cosine1.00000008; existing counts treated any vector as indexed.

    Design: retain the existing MiniLM fp32/mean/normalized model, pin its existing observed revision, and identify the full embedding contract (model/revision/dtype/pooling/preparation version). Cache validity requires matching identity and384 finite normalized dimensions. Untagged, foreign or malformed vectors remain stored with original mail, but are excluded from search and counted pending for ordinary account-scoped reindexing. New vectors carry identity; synchronization preserves identity with vectors. Never label legacy values as current without inference. Replace percent-match wording with an explicit similarity score. No model replacement, production data/schema operation or deployment.

    Scenarios: same-dimension foreign vectors, untagged legacy values, short/NaN vectors cannot rank or count indexed; valid current vectors survive refresh/reload; pending records reindex without losing email or another account's data; cancellation/retry remains accurate. Cold-cache model loading and actual existing model output remain separately verified. Broader ranking and real-mail/account gates stay open.

    • Source repair and focused regression checks.
    • Real browser IndexedDB preservation/reindex/reload proof, plus cold model load.
    • Full owning quality/docs and exact-source CI.

    Completed with runtime 24142fe, CI 34323190467 and the release receipt in the latest issue comment. Broader #54 account/mailbox gates remain open.

  2. sarthakagrawal927 commented on Sep 9, 2026

    @sarthakagrawal927
    MemberAuthor

    Released cache correctness and required dependency patches on 9 September.

    • Runtime 24142fe passed CI 34323190467 (124 tests, real IndexedDB verifier, full quality/build/docs); all six deploy gates passed without bypass.
    • pnpm run deploy published Worker 9ad4e473-50cb-4cfb-8735-895e4a909941, deployment 5df53469-9e8c-40fa-b4bb-59febc65be2d, exact full-SHA tag at 100% traffic. Known rollback b42e2e27-b358-4ee7-a338-f9cc9d4cbc69 restores 451b699; not exercised.
    • Ordinary https://mail.significanthobbies.com/app serves byte-identical /assets/index-Dw903OH4.js; SHA256 b618032cc0f5f6d27644941802440921f9eb8c8437b61fc7de2bc0ef4146eded. Real 390px landing→Google sign-in button has no overflow or page errors. Health/session 200 and anonymous mail 401; Google configuration present. No test CSP or response mocks in this hosted check.
    • Legacy vectors are retained, never silently deleted or labelled current; emails stay available. Semantic search excludes incompatible/malformed/sparse vectors until explicit normal indexing regenerates them. Current model identity survives sync/reload; invalid inference cannot acquire identity. Similarity is no longer presented as probability.
    • MiniLM retained: fresh 16-message/18-query diagnostic scored 16/18 versus BGE 17/18. Both miss a boarding-pass query against a promotional distractor. This is not broad ranking qualification. Pinned MiniLM cold-cache load/inference passes 6116ms, 384 finite normalized values, zero CSP violations.
    • Scoped dependency patch: Astro 7.2.8 plus HTML whitespace preservation, Sharp 0.35.4, js-yaml 4.3.2, svgo 4.1.0. Zero critical/high advisories; no new direct production dependency or provider configuration changes.

    Durable evidence: docs/knowledge/learnings/embedding-cache-2026-09-09.md, comparative corpus/receipt JSON, release JSON and inspected local/hosted mobile screenshots. The localhost telemetry CORS limitation is documented separately; hosted guest check had no page errors.

    Keep #54 open for real Google callback/account switching, permitted mailbox search/senders/digest and compatibility-host acceptance. No credentials, real mail or unsubscribe operation was used; no D1 migration or native/public installation claim. Browsers and servers closed, task-owned temporary files cleaned after retaining evidence. Documentation receipt commit does not trigger another deployment.

  3. added
    queue: excludedRetained outside the immediate-use project queue; remove label and unarchive to resume.
    on Sep 11, 2026
  4. sarthakagrawal927 commented on Sep 18, 2026

    @sarthakagrawal927
    MemberAuthor

    Closing as part of the owner-approved Fleet inactive-project reconciliation on 2026-09-18. The source and history remain available; if the project is reactivated, open a new scoped issue against the current product direction.

Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Metadata

Metadata

Assignees

No one assigned

    Labels

    queue: excludedRetained outside the immediate-use project queue; remove label and unarchive to resume.

    Type

    No type

    Projects

    No projects

      Milestone

      No milestone

      Relationships

      None yet

      Development

      No branches or pull requests

      Issue actions