[qwen-code-review-bot]

Verified daily translation batch from scheduled run https://github.com/QwenLM/qwen-code-docs/actions/runs/39691454871.

Review and merge to publish: the main push triggers deploy-docs.yml, which rebuilds and deploys the site. Later scheduled runs pause while this PR is open, so merging it also unblocks the next one.

Files whose structure could not be verified were restored from HEAD and are not included here; the run log lists them.

[yiliang114]

Patrol-Run: qwen-docs-pr-review

Approving. Verified mirror fidelity and baseline arithmetic by computation rather than by reading for drift.

What I checked:

• EN mirror is byte-identical to upstream QwenLM/qwen-code@79242671 for all 10 changed files, after applying sync.ts's own two systematic transforms — its RELATIVE_LINK regex with fence tracking, plus dropping the page-top language-switch line. managed-agent-reference-deployment.md and workflow-intent.md match only after the language-switch strip, which is the expected shape.
• All 26 added (file, lang) hashes in last-sync.json recompute to the sha256 of the corresponding upstream content at 79242671. Real catch-up, not hash shuffling. commit moves d098ff23 -> 79242671.
• Nav gate — the load-bearing one, since this batch adds 8 new pages: en/users/_meta.ts:35 registers 'workflow-intent', so findUnlistedEntries(content/en) stays empty and the post-merge deploy-docs.yml check has nothing to trip on.
• computer-use.md is the largest rewrite, and all 8 languages land on EN's shape (82 lines, six H2, 12 fences, identical per-fence line counts). The only fence-content deltas are translated // comments and ASCII-diagram labels in de/fr/pt-BR/ru; commands, flags, env vars and identifiers are untouched. I read EN against zh and de end to end — no claim inverted, nothing dropped.
• The new workflow-intent.md is much shorter in ja (45) and ko (44) than EN (61), so I read both in full to rule out truncation. It is prose re-wrapping, not lost content — every EN paragraph is covered.
• Placeholder/mojibake scan across all 29 non-EN changed files: zero hits (no U+FFFD, no {{, no stray %s, no truncation markers). Fences balance everywhere.
• MDX brace risk: the only bare brace outside a code span in the changed set is ko/users/features/auto-mode.md:69 {#classify-all-shell-commands}, a heading-ID attribute, and that file was not touched in ko by this PR.

Non-blocking:

1. Seven copies of the new workflow-intent.md carry a spurious space in the SDK import inside the code fence — from ' @qwen-code/sdk' instead of from '@qwen-code/sdk': de:34, fr:36, ja:22, ko:21, pt-BR:33, ru:34, zh:29. EN is clean. No build impact (inside a fence, MDX-safe) but it does break copy-paste. main already ships 17 instances of the same pattern (14 in */developers/daemon-client-adapters/web-shell.md, 3 in {ru,pt-BR,ko}/developers/sdk-typescript.md), so this batch takes the total to 24. That points at the translation pipeline rather than at this PR, so I am not blocking on it — worth fixing once at the source instead of per batch.
2. managed-agent-reference-deployment.md still lags EN in de/fr/ko/pt-BR/ru (nine H2 and seven fences vs EN's twelve and ten). Those five were not touched here and keep their old 295649941c… hash, so they stay in the backlog and self-heal under the per-(file, lang) model. zh and ja were updated and do match EN.
3. de/users/features/computer-use.md:63 reads "js completes in einem Aufruf" — the English verb survived into a German sentence. Meaning is intact; cosmetic.
4. Several files in this batch end without a trailing newline. Pre-existing pattern, no build impact.
5. Anchor fragments still reuse EN slugs on localized pages — the known pipeline-wide issue; neither the build nor the tests validate fragments.

Not checked: I did not read de/fr/pt-BR/ru prose end to end, so a semantic inversion in a language I did not open would not have been caught — zh and de (full read), ja and ko on the new page, plus the shape and fence checks are what I can stand behind. I did not build the site or run the translator locally; the conclusion rests on the three PR checks plus byte-exact mirror comparison and hash recomputation.