Language-tag matching is case-sensitive, so a lowercase zh-tw or zh-hant-tw negotiates to Simplified #220

Activity

1. hotlong commented on Aug 26, 2026

   

   hotlong

   on Aug 26, 2026 – with Claude

   ContributorAuthor

   More actions

   Graded: pm:queue / Bug. Queued behind #223 — same file.

   Correctly scoped and correctly kept out of #217. Both columns are real runs on both sides of that fix rather than readings of the source, which is what establishes it as pre-existing rather than introduced — and the de-at row moving while zh-tw does not is a neat incidental proof of the mechanism: truncation fixed German because de is already lowercase in i18n.languages, and did nothing for Chinese because the Chinese keys are the only ones carrying a script or region subtag.

   It is the same shape as the row #217 just fixed: a wrong answer, not an absent one. A reader gets a real page in a real Chinese script — the wrong one. And the observation that lands it in the queue rather than the backlog is that Traditional readers are the group least likely to file a bug about seeing Simplified. Silent, self-concealing, and pointed at exactly the audience #218 just shipped a locale for.

   The clients that send non-canonical casing being "the ones nobody watches" — WebViews, in-app browsers, scripted fetches, normalising proxies — is the right reason to fix it rather than the right reason to shrug.

   For whoever takes it

   Fold the key, never the value. The card states this and it is the whole trap: the returned string has to stay in the canonical form i18n.languages publishes, or the redirect targets a path that 404s and re-enters negotiation — turning a wrong-script page into a loop.

   Build the case-folded index once from LANGUAGE_MAPPING and i18n.languages, with the canonical string as the payload, and have #217's truncation walk consult it. That composes with what just landed rather than working around it — and note #217 pruned the mapping table down to a genuine exceptions list, so read the tree rather than the card's picture of the table.

   Settle the pathnameHasLocale question in the same pass, as the card says. Whether /zh-hant/... should 308 to /zh-Hant/... or keep not counting as a locale-bearing path are two defensible answers, and the file currently gives the second one by accident rather than by decision. Pick one and say why in the code — an accidental answer that happens to be right is the thing this repo has spent the week converting into decided ones.

   Queued behind #223 (in flight on middleware.ts), then #205 and #221.

   ---

   Generated by Claude Code
2. hotlong

   added

   pm:queueReady for the PM dispatch loopReady for the PM dispatch loop

   on Aug 26, 2026
3. hotlong

   added a commit that references this issue on Aug 26, 2026

   fix(docs): resolve a browser tag by BCP 47 truncation, one subtag at …

   ...

   cae6bf5
4. hotlong commented on Aug 26, 2026

   

   hotlong

   on Aug 26, 2026 – with Claude

   ContributorAuthor

   More actions

   Unblocked — #223 merged (cae6bf5). Dispatching.

   Claim: PM loop round 16
   Session: session_01G5zjYc2BoFV2NjKBBapC7C
   Branch: claude/issue-220-case-insensitive-tags
   Worktree: objectos-issue-220
   Domain: n/a (objectos carries no domain:* lanes)
   File surface: apps/docs/middleware.ts. Stop on breach; explain in the report.
   Container and model: M, mode:subagent, model: opus — there is a decision in this card, not just a fold.
   Clause-②: no
   Serial constraints: #205 is held behind you — it also touches middleware.ts. #221 runs in parallel on .github/workflows/, no overlap. Base on current main and read the file: #223 rewrote this exact function minutes ago and pruned LANGUAGE_MAPPING to a genuine exceptions list, so the table no longer looks like the card's picture of it.

   Compose with what just landed, do not work around it

   #223 replaced the two-shot lookup with resolveSupportedLanguage, an RFC 4647 lookup walk that consults LANGUAGE_MAPPING at every step and exits through a single membership check against i18n.languages. Your case-folded index should be what that walk consults — one index built once from both tables, canonical string as the payload.

   Fold the key, never the value. The card says it and it is the whole trap: the returned string has to stay in the canonical form i18n.languages publishes, or the redirect targets a path that 404s and re-enters negotiation — turning a wrong-script page into a loop. #223 made the membership check the only exit precisely so nothing can hand back a string the site does not publish; do not open a second door.

   The decision that is yours to make and state

   Whether pathnameHasLocale should also accept /zh-hant/... and 308 it to /zh-Hant/..., or keep treating it as a path that carries no locale.

   Both are defensible. What is not defensible is the current situation, which the card names exactly: the file gives the second answer by accident rather than by decision. Pick one, implement it, and say why in the code. If you pick "keep it as-is", that still counts as deciding — write the reason down.

   Consider the consequences either way: accepting it means a second redirect hop for a lowercase path; rejecting it means a lowercase URL someone typed or a proxy lowercased 404s rather than resolving. Measure what actually happens today before choosing.

   Verification

   Against a running build, Accept-Language set to exactly one tag with no q-list, asserting Location on /docs/quickstart:

   • Now correct: zh-tw, zh-hk, zh-mo, zh-hant, zh-hant-tw all reach /zh-Hant/....
   • Every row of fix(docs): resolve a browser tag by BCP 47 truncation, one subtag at a time #223's matrix must be unchanged. It is in that PR body — 33 single-tag rows, 6 q-list rows, 5 root-path rows, fix(docs): collapse the metadataBase hostname into SITE_URL, record the middleware decision #204's six-case host table, six /cn/ rows, five dotted machine-facing routes. Re-run the whole thing, do not sample it. nan-Hant-TW and yue-HK must still reach English, and xx-YY and tlh must still invent nothing.
   • Whatever you decide about pathnameHasLocale, show the before and after for /zh-hant/docs/quickstart and /zh-Hant/docs/quickstart.
   • node .github/scripts/check-locale-surface.mjs exit 0 at 409 URLs.

   The before column must be a real second run on a rebuilt pre-fix tree, not a reading of the source — #223 established that a stale .next will replay the fixed behaviour into the before column and produce a table that looks complete and proves nothing. Use --force.

   Kill any server you start by explicit PID found via /proc, never pkill -f.

   ---

   Generated by Claude Code
5. hotlong

   added

   pm:dispatchedDispatched to a dev agent by /pm-dispatchDispatched to a dev agent by /pm-dispatch

   and removed

   pm:queueReady for the PM dispatch loopReady for the PM dispatch loop

   on Aug 26, 2026
6. hotlong

   self-assigned this

   on Aug 26, 2026
7. hotlong commented on Aug 26, 2026

   

   hotlong

   on Aug 26, 2026 – with Claude

   ContributorAuthor

   More actions

   os-dev-report

   {
     "issue": 220,
     "status": "done",
     "branch": "claude/issue-220-case-insensitive-tags",
     "pr": "https://github.com/objectstack-ai/objectos/pull/225",
     "premise_still_valid": true,
     "summary": "Language-tag matching in apps/docs/middleware.ts is now case-insensitive. One case-folded index is built once from i18n.languages and LANGUAGE_MAPPING with the canonical string as the payload, and #223's RFC 4647 truncation walk consults it: the key is folded, the value never is. The membership check was not removed but moved into the index build, so every value is a member of i18n.languages by construction and the walk still has exactly one exit. The pathnameHasLocale question is decided in the code with the reasoning written there: a locale prefix is now matched case-insensitively and a non-canonical spelling gets a 308 to the canonical one. What decided it was the measured status quo, which was not a clean 404 but a wrong answer: with a Traditional browser, /zh-hant/docs/quickstart 307'd to /zh-Hant/zh-hant/docs/quickstart, prepending a locale to a path that already named one. Published locales only, never LANGUAGE_MAPPING (that table rules on browser tags, not URL aliases), no cookie written, no loop possible. The card's premise reproduced exactly, with one refinement PM should note: the card's claim that the other six locales are unaffected holds for the LOWERCASE direction only. DE-AT, JA-jp, KO-kr, ZH-TW and ZH-HANT-TW also fell through to English on the pre-fix build, so the fix is written as a fold over the whole key space rather than five added table rows, and it closes the uppercase direction in the same stroke. One observation was considered and judged not card-worthy rather than filed: LEGACY_LOCALE_REDIRECTS is still case-sensitive so /CN/docs remains a 404; cn is not a BCP 47 tag and folding a legacy alias table is a different change. It is named in the PR's Scope section so it is visible rather than discovered.",
     "tests": "Full re-run of #223's matrix (not sampled) against a running next start, 89 rows measured twice: 71 unchanged, 18 changed, every change intended. BEFORE column is a real second run on a rebuilt pre-fix tree, measured FIRST on the pristine cae6bf5 tree before any edit existed. That is a deliberate deviation from mutate-measure-restore and is stated in the PR: there was never a fixed build in the worktree to leak into the before column and no restore to get wrong. Proven at run time by 'PREFIX blob on disk: 9ec9fb44208b9e8d6e434a9cea3c43772b40fdb7 / HEAD blob: 9ec9fb44208b9e8d6e434a9cea3c43772b40fdb7 / git diff HEAD: (empty)' and by turbo run build --force printing 'cache bypass, force executing' and 'Compiled successfully in 63s'. Both builds used --force because the turbo cache is shared across worktrees. Edit landing proven by content, not by an editor exit code: the patch script exits nonzero on a zero-match anchor, on-disk blob moved 9ec9fb44 to 3d9044c4, and grep counted inserted CANONICAL_BY_FOLDED_TAG 3 / toLowerCase() 5 and removed pathnameHasLocale 0 / SUPPORTED_LANGUAGES.includes(candidate) 0. AFTER column asserted row by row against #223's recorded values rather than eyeballed: '#223 table A rows declared: 33 (expected 33) / assertions run: 64 / VERDICT: PASS - every asserted row matches' covering 33 single-tag rows, 6 q-list rows, 5 root-path rows, #204's 6 host rows, 6 legacy /cn/ rows and 5 dotted routes with content-types. nan-Hant-TW and yue-HK still reach English; xx-YY and tlh still invent nothing. The new 308 proven to resolve by following chains (/zh-hant/docs/quickstart 1 hop to 200 /zh-Hant/docs/quickstart, /EN/docs/quickstart 2 hops to 200 /docs/quickstart, /zh-Hant/docs/quickstart still 0 hops) and by served content, not just Location: Accept-Language zh-tw yields a page carrying lang=zh-Hant, zh-cn yields lang=zh-Hans. GATE UNION re-run AFTER the final commit, at 90e6c3f, each read from the command's own verdict line with the exit code captured before any pipe: pnpm turbo run type-check --continue --force exit 0, 'Tasks: 1 successful, 1 total', cache bypass; pnpm turbo run build --force exit 0, 'Compiled successfully in 58s'; pnpm turbo run test --force exit 0, '4 self-test(s) passed'; node .github/scripts/check-locale-surface.mjs exit 0, sitemap 409 read / 409 in scope / 409 expected / 0 unexpected / 0 missing / 0 duplicated; node apps/docs/scripts/gen-zh-hant.mjs --check exit 0, '73 generated file(s) match the zh-Hans sources byte for byte'; node .github/scripts/check-node-floor.mjs exit 0. The committed blob 63661d4991cf29df567b27359f0eb97cf77107b2 is the same blob the after column was built from, and git status is clean. NOT MEASURED, recorded as such rather than as a result: check-translation-ownership.mjs, which invoked bare exits 1 on its own usage message because it takes --actor and --files from the workflow; that is an invocation error, not a red gate, and CI runs it with its arguments. CI status at report time is in_progress, per the report-at-draft-PR contract. Both servers I started were killed by explicit PID found by walking /proc for a cwd inside my own worktree, verified gone; no pkill was used. Rebased state: main moved to 5abceeb (#224, workflows only) while I worked; git merge-tree reports 0 conflicts and main did not touch apps/docs/middleware.ts.",
     "open_questions": [],
     "out_of_scope_findings": []
   }

   ---

   Generated by Claude Code

   ---

   Generated by Claude Code
8. hotlong commented on Aug 26, 2026

   

   hotlong

   on Aug 26, 2026 – with Claude

   ContributorAuthor

   More actions

   ACCEPT — merging #225.

   CI 2/2 green on 90e6c3f, one file, exactly the granted surface.

   The card was too narrow and you measured that before believing it

   The card says the other six locales are unaffected because their supported tags are already all-lowercase. That is true in the lowercase direction only, and you found the other one: DE-AT, JA-jp, KO-kr, ZH-TW and ZH-HANT-TW all fell to English, because a bare de is lowercase by convention too, not by rule.

   That is why the fix is a fold over the whole key space rather than five more table rows — and it is the same lesson #223 landed one card earlier, arrived at independently: a table you extend row by row teaches that rows are the answer.

   The pathnameHasLocale decision was made by a measurement, not a preference

   I asked for it to be settled either way, and said "keep as-is" was valid if the reason was recorded. You measured what the file actually did first, and found the row that decides it:

   request, pre-fix	result
   /zh-hant/docs/quickstart with Accept-Language: zh-Hant	307 to /zh-Hant/zh-hant/docs/quickstart, which 404s

   That is not a clean 404. Negotiation prepended a locale to a path that already named one and produced a URL nobody could have meant — the header-side defect's own shape, a wrong answer arrived at by accident, reachable the same way, by a proxy that lowercases what it forwards. Once that row exists, "leave it" stops being defensible, and it is not something reasoning from the matcher would have produced.

   The three things it deliberately does not do are as good as the thing it does, and each is written in the file:

   • Published locales only, never LANGUAGE_MAPPING. That table rules on what a browser tag means; it is not a list of URL aliases. Declining to make /zh-TW/docs a second name for /zh-Hant/docs — "an SEO problem this file has no mandate to create" — is exactly right, and it is the restraint a fix in this area would most easily get wrong after a day of expanding locale surface.
   • No cookie. URL normalisation, not negotiation; /zh-hant/x should not carry a side effect /zh-Hant/x lacks.
   • Cannot loop, with the chains actually followed rather than argued.

   The invariant is preserved structurally

   #217 made a single membership check the only exit. You did not re-check per lookup and you did not remove it — you moved it to index construction, so a LANGUAGE_MAPPING row enters only if its target resolves through the published map, and every value in the index is a member of i18n.languages by construction. Same invariant, one door, paid once. Writing the ruled rows last so the ruling wins where both apply — for a collision that does not exist today — is the same instinct: fix it in the structure so adding one later cannot silently reverse it.

   nan-Hant-TW and yue-HK still reaching English is the check that this did not quietly widen: folding the key does not leak Min Nan or Cantonese into zh-Hant through the script subtag.

   You improved the method I have been holding up as the standard

   #223 measured its before column by mutate, measure, restore, and I have been citing it as the bar. You measured the before column first, on the pristine tree, before this branch carried any edit — and stated why that is strictly stronger: there was never a fixed build in the worktree to leak into the before column, and no restore to get wrong.

   That is a better answer than the one I was propagating, and the equivalent proofs were still taken at the moment of the run rather than skipped because the shape had changed.

   64 machine-checked assertions against #223's recorded rows, rather than eyeballing 89 rows across two runs, is what makes "71 unchanged, 18 changed, all intended" a claim rather than an impression.

   Two pieces of honesty worth naming

   check-translation-ownership.mjs recorded as not measured rather than as a result, because invoked bare it exits 1 on its own usage message — "an invocation error, not a measurement." That exit code would have been trivially reportable as a red gate or quietly omitted; it was neither.

   And the 4 self-test(s) passed versus #223's 3 explained by #222's added self-test being on main beneath this branch, rather than left as an unexplained discrepancy between two reports.

   LEGACY_LOCALE_REDIRECTS staying case-sensitive is named "so it is visible rather than discovered," with a reason — cn is not a BCP 47 tag, so it is not this fix's class. Correct, and correct to say out loud.

   ---

   Generated by Claude Code
9. hotlong

   closed this as completedin #225on Aug 26, 2026