Skip to content

WORKLOG.md

Latest commit

 

History

History
161 lines (136 loc) · 9.03 KB

WORKLOG.md

File metadata and controls

161 lines (136 loc) · 9.03 KB
Raw

work log: alpha.8 → alpha.10

Durable record of decisions, evidence, and next steps for the alpha.8–alpha.10 push. Newest entries last. Release mechanics are in RELEASING.md.

state at start (2026-09-26)

  • alpha.7 released (GitHub prerelease + crates.io, PyPI, npm, Maven). research/power-states-paper merged into main as PR #4 (f0135e8).
  • portable correctness / research-evidence red on main since alpha.6: paper/audit/archive-inventory.json was not regenerated after the power-state archives landed. Lint, clippy, and workspace tests were green.
  • Uncommitted alpha.8 work from the previous session: TUI ? Functions tab (10 hard-coded entries), TypeScript bigint range validation, docs home.

alpha.8

  • Decision: ship the existing work plus the CI fix as one small release; defer the full function browser to alpha.9.
  • TypeScript bug: BigUint64Array.from([-1n]) stores 2^64-1, so a negative ID was written as a different record ID. Now RangeError before any native call. Test: sdk/typescript/test/qenlo.test.ts (6/6 pass locally).
  • CI fix: regenerated the inventory (1654 → 1833 files; deterministic across two runs). bump_alpha.py now regenerates it on every version bump.
  • Local checks: cargo fmt --check, cargo clippy -p qenlo-browser --all-targets -D warnings, cargo test -p qenlo-browser, TS typecheck + tests, full research-evidence job steps (17 unit tests, alpha.5 gate reproduction, processed-data diff), check_docs, check_release_versions.
  • Docs home verified rendering in a browser at docs/#/.

candidate work (ranked)

  1. alpha.9 function browser: catalog every public Collection method, group by task, type-to-filter, example per entry, and a test that fails if a public method is missing (drift guard).
  2. alpha.10 README / landing hero: lead with what Qenlo is, the evidence map, and honest results (wins and failed bets), with links to raw data.
  3. Surgical perf/reliability items found during review (see below).

alpha.8 shipped

  • CI root cause #2: after regenerating the inventory, Linux still differed in 45 places. glob() follows filesystem order (NTFS sorted, ext4 hashed), so manifest_references reordered. Fixed by sorting on the name string (Windows Path ordering is case-insensitive, POSIX is not). portable correctness fully green on d1ed007, first time since alpha.5. Tagged alpha.8 there.

alpha.9

  • Function browser (delegated to Sonnet, reviewed): 33/33 public Collection methods, / filter, drift test that re-parses crates/qenlo/src/lib.rs. Review found the 80x24 layout hid the summary and example entirely (8-row shortcut panel) and truncated names; fixed with an adaptive one-line panel and a TestBackend render test at 80x24. Every example identifier was checked against the source (Filter = Predicate, derives Default, etc.).
  • Reliability scan (Sonnet, read-only) → four fixes, each with a test: closed collections served stale get_record/scan_records; FFI free ran close() outside catch_unwind and dropped its error; Python __del__ was silent (now ResourceWarning, stdlib convention); TypeScript leaked handles and the directory lock without close() (now FinalizationRegistry).
  • Deliberately not done: scan_records(filter) materializes every matching slot per page (O(matches) per page). Candidate for alpha.10 if measured.

alpha.9 shipped

  • Found while verifying alpha.8 assets: SHA256SUMS hashed itself mid-write, so sha256sum -c always failed one line. Fixed in sdk-release.yml before tagging alpha.9 (5300b6c). The alpha.8 asset digests all match GitHub's recorded digests; only the self-line is bogus.
  • Line endings: Python write_text on Windows turned two LF .rs files into CRLF in 212eb52; restored in 5300b6c. Use byte writes from now on.

alpha.10

  • Landing page: the three hero cards had rendered ~2px wide on the live site since alpha.8. .cards is container-type: size and sizes cards from 100cqh; with flex: 1 1 auto, the research section added below it took the height and 100cqh resolved to 0 (a min-height alone did not fix it; a definite flex basis did). Verified 426px desktop, 361px tablet, 347px mobile in a browser. The dot-matrix glyph set lacked 2/3/4/8 (Sonnet caught it), so 4.6 would have drawn as 0.0.
  • Hero numbers replaced: prototype 1M / 0.17 ms / 100% (unshipped PyTorch CUDA prototype that lost to FAISS in its own gate) → 4.6x, 235.7%, 1,833, each traced to a file. README restructured (install, examples, research table); wording fixed where the draft implied causation ("speed tracks power state") or overstated ("learned rule"). All 1,833 inventory entries are hashed, so "every file" is accurate.
  • Web/desktop function browser: GET /api/functions from the same catalog; checked in a browser (/ focuses search; "batch" → 3 of 33).
  • Deferred with reason: sqlite-vec comparison (roadmap #2). Needs a release build of the bench and matched filter semantics; too heavy for a modest overnight run. Next step: add an SqliteVec backend to scripts/oss_replay.py (~25 lines) and run the 100K x 384 cell.

GTM

  • X and LinkedIn were unavailable until the user signed in mid-session. Posted on X (voice skill + short style): CI glob-order story and alpha.8. Drafts and a posted log live in .claude/gtm-drafts.md (untracked).

alpha.10 re-tag

  • Post-publication proof (RELEASING.md) on alpha.8: PyPI wheel works on Windows; the npm package fails with "native library not found". Cause: the release job packed native/windows-x64/qenlo_ffi.dll, the loader looks for native/win32-x64/ (Node's process.platform). Every npm version back to at least alpha.3 has this layout; SDK CI only exercises Linux and local dev uses a win32-x64 folder, so nothing caught it. Moving the published DLL into win32-x64/ made the same package pass (search + 18-field report).
  • The first alpha.10 tag (806ca76) was pushed before this was found. Its release build was cancelled before any GitHub release or registry upload, the tag was deleted, and alpha.10 is re-tagged on the fixed commit. No published artifact ever carried the first tag.

final state (2026-09-26)

  • Released and published: alpha.8 (d1ed007), alpha.9 (5300b6c), alpha.10 (5c2400c); all on main, GitHub prereleases with 11 assets each, and on crates.io, PyPI, and npm (alpha dist-tag). Maven Central is still not actually published (job passes without credentials).
  • Verification per RELEASING.md "post-publication proof", Windows only: clean uv env from PyPI and clean npm project from the registry for alpha.8, alpha.9, alpha.10: create in-memory collection, add, filtered search, execution report, close. alpha.10 npm works without any workaround; alpha.8/9 npm need QENLO_LIBRARY_PATH on Windows (noted on their release pages). sha256sum -c SHA256SUMS passes for alpha.9 and alpha.10 downloads.
  • GTM (user signed in mid-session): five X posts and one LinkedIn post, voice skill + short style, links in .claude/gtm-drafts.md.

open items / next steps

  1. Done: the maintainer moved npm latest from 0.1.0-alpha.1 to 0.1.0-alpha.10. The publish workflow still tags only alpha, so latest will lag again on the next release unless it is moved at publish time.
  2. SDK CI tests TypeScript only on Linux, which is how the Windows layout bug survived; a Windows/macOS job that installs the packed tarball would catch the whole class.
  3. sqlite-vec comparison (roadmap #2), plan above.
  4. Linux/macOS post-publication proofs and a Rust registry consumer build were not run locally (CI builds and tests those platforms from source).
  5. Desktop app Functions tab inherits the web UI; not separately launched.

alpha.11 (2026-09-26)

  • Building qenlo-memory (github.com/a3ro-dev/qenlo-memory, a cross-agent memory on the Python SDK) showed that flush() never compacted. It returned early on durable_generation == generation, which every WAL commit satisfies, so one WAL file per write piled up and open replayed all of them. Fix: track the loaded or written snapshot generation (OpenedStore.snapshot_generation, Collection.snapshot_generation), and skip only when the snapshot is current. close() keeps the old durable-only check so it never writes a full snapshot. Rust regression test flush_compacts_committed_wal_files; checked through the Python FFI too (300 WAL files -> 0, reopen keeps 300 rows).
  • Python README: wheel list now matches what is published (no Linux aarch64 or Intel macOS), and the backend strings match the FFI's Debug output.
  • Deliberately unchanged: records have no payload and deleted ids are never reused. qenlo-memory keeps text in sqlite next to the collection, which is the intended pattern.
  • Published alpha.11 (crates.io, PyPI, npm alpha); proof on Windows from PyPI: in-memory add/search/report, and a durable collection's flush() leaves 0 WAL files. npm latest still pointed at alpha.10, so sdk-publish.yml now moves latest to each release after publishing (future releases).