[MeharPro]

Status: ready to send (after the sender's by-hand items in docs/adoption/hunts/superset-sqllab-issue.md, "Before sending").

What

A race hunt on Apache Superset's SQL Lab (master 730016b8, run natively, GenClass injected from outside). EditorAutoSync marks editors as saved at the time the PUT /tabstateview/<id> response arrives. An edit typed while that request is in flight therefore counts as saved, is never sent, and a reload brings back the old SQL. Explore reported it (RACE 2 of 6 at 600 ms, 3 of 6 at 900 ms, on exactly the loads whose server copy lost the edit) and its replay reproduced it. A plain Playwright script reproduces it 3 of 3 at 600, 900 and 1,500 ms (0 of 3 at 0 and 300 ms). Two smaller bugs of the same family the brief predicted were also confirmed and fixed, each as its own patch: Format SQL overwrites text typed while formatting, and the Save query modal saves twice on a double click.

Only files under docs/adoption/hunts/ are added or changed; nothing in packages/.

Files

• docs/adoption/hunts/superset-sqllab.md: the report (setup, scenarios and verdicts incl. negatives, the finding in causal order, root cause, fixes, tests, fixed variant, problems with GenClass, time).
• docs/adoption/hunts/superset-sqllab.patch: the autosave fix plus a Jest regression test (git format-patch; applies at 730016b8 and at the brief's f44d3992).
• docs/adoption/hunts/superset-sqllab-format.patch and superset-sqllab-savequery.patch: the two secondary fixes, each with its test.
• docs/adoption/hunts/superset-sqllab-issue.md: the issue (Superset's bug-report form), the pull request (Superset's template), the optional Format SQL issue and PR, the disclosure paragraph and the "before sending" checklist.
• docs/adoption/hunts/superset-sqllab/: harness (hunt.mjs, repro-autosave.mjs, repro-format-save.mjs, lib.mjs, summarise.mjs) and results/{unfixed,fixed}/ (tool texts and summaries, each file under 25 KB).
• docs/adoption/hunts/README.md: the row.

To send (a person, from a personal account, after the owner's go-ahead)

1. Re-search apache/superset issues and PRs for anything filed since 2026-11-03 ("tabstateview", "EditorAutoSync", "autosave" + "SQL Lab", "unsaved" + "sqllab").
2. Reproduce by hand: DevTools throttling at about 1 s latency. Type in SQL Lab, wait for the autosave PUT, type more while it is pending, wait 10 s, reload: the second edit is gone.
3. File issue A from superset-sqllab-issue.md through the bug-report form (fields map one to one).
4. Fork apache/superset, git am docs/adoption/hunts/superset-sqllab.patch on a branch from master, set yourself as author, keep the Generated-by: Claude Code trailer, read the diff, then run npm run test -- src/SqlLab/components/EditorAutoSync with and without the fix and pre-commit run on the staged files.
5. Open PR B (title fix(sqllab): keep editor changes made while the tab state is being saved) with "Fixes #". Superset keeps PRs open at least 24 h and needs CI green plus one approval. Answer review yourself.
6. Format SQL (section C of the drafts) only after A/B get a maintainer response. Save query: no draft, later if at all.

Measured (every number with its M and latencies)

	unfixed	fixed
autosave, plain Playwright, response held 0/300/600/900/1,500 ms, 3 trials each	lost 0/0/3/3/3 of 3	lost 0 of 3 at 0/600/900/1,500
Format SQL, plain, held 0/300/900/1,500 ms	typed text lost 0/0/3/3 of 3	0/0/0/0 of 3
Save query, two clicks 150 ms apart, held 0/300/900 ms	duplicate 0/3/3 of 3	0/0/0 of 3
explore autosave, slow_ms 300/600/900, 2 baselines + 6 seeds each	no race / RACE 2 of 6 / RACE 3 of 6 (replay reproduced both); oracle agrees per load	no race at all three
explore format	RACE 6/5/5 of 6	no race / no race / RACE 3 of 6 (false: formatted vs unformatted, text kept)
explore savequery	RACE 2 of 2 (false: harness tab names) / no verdict / no verdict	RACE 5/2/6 of 6 (false: harness tab names, client-id key, unexplained text)
Jest regression tests (3 patches)	each fails 3 of 3	each passes 3 of 3

False reports: on the fixed build, 0 of 72 explore loads had any bug by the harness's own oracle, but 4 of 9 fixed sessions still carried a RACE line. On the unfixed build, the verdict was right in 6 of 9 sessions.

Problems with GenClass itself (report section 7)

• When a bug needs latency, the delayed fifo baselines can carry it: at 900 ms the "failing run" explore named was the correct one.
• slow_ms explores "response before or after the user's next keystrokes", but the docs describe only reordering.
• Superset's Redux store (production build, no DevTools enhancer) was not discovered, and component state came under minified names.
• Stale-write findings on a bookkeeping counter (autosave). Format-SQL replays said "inconclusive" for a timing-dependent mangling.
• A PUT keyed by an un-normalised 11-character client id fed a false RACE line.
• No identical-in-flight witness for the Save query double submit (16 of 18 seeded loads sent two identical POSTs).
• Save query double submit: no verdict, because the second click exists only in the failing loads.

Checks

• npm run typecheck -w @genclass/runtime and npm run build -w @genclass/runtime pass: build run (the harness uses dist/); typecheck not run separately, since no runtime source changed
• unit tests pass: not run, since no GenClass code changed
• no packages/runtime/test/review-*.test.ts was edited
• no model-visible text changed
• no public-surface change
• a correct app is not made worse: no runtime change
• sim/ and demos/ not touched
• no dependency change (the local xlsx workaround stayed in the Superset clone, out of every patch)

Notes for the reviewer

• Read superset-sqllab.md sections 3 (the tables), 4 (the finding) and 7 (GenClass problems) first.
• Branch: the session was told to push to claude/beautiful-einstein-5r1xn0 rather than adopt/superset-sqllab.
• Upstream checks run in the session on the changed Superset files: pre-commit oxfmt, oxlint, custom rules, type-checking-frontend (after npm run plugins:build), all passed; every src/SqlLab Jest suite: 47 suites, 526 passed, 1 skipped.
• Not run: Superset's Playwright e2e suite (a Jest unit test covers each fix); the server-side PUT ordering the brief suspected (it needs a save slower than the 5 s debounce; noted, not fixed).
• Local deviations: xlsx repointed to an npm mirror because cdn.sheetjs.com is blocked by this container's proxy; zstd installed for the build.
• Session time: about 1 h 40 min.

---

Generated by Claude Code