Community pre-validation batch — 43 PRs checked; 35 merged (Oct 31)

[Antonio-Faure]

I spent some time pre-validating the open PR queue so the maintainer can merge in batches instead of re-running the same checks. This is a community snapshot, not a request for action — every PR below has the exact command and result in its review comment.

Maintainer checklist (Oct 31)

CLEAN PRs ready to merge — all checks green, no failures:

• fix(math_animator): keep the post-render layout report from triggering repairs #2246 — fix the post-render layout review retry in feat(math_animator): dry-run geometry check gates each render #2191 (22 checks pass).
• refactor(settings): remove the orphaned legacy model editor stack #2251 — remove 3,549 LOC of orphaned settings editors; full Next.js build passes.
• refactor(web): remove the dead watching marks panel and video client API #2252 — remove 535 LOC of dead watching/video client; full Next.js build passes.
• chore: remove dead Python modules and helpers #2253 — remove 456 LOC of dead Python helpers; focused tests pass.
• fix(web): keep chat scroll position when the reader scrolls up #2254 — fix [Bug]: 模型正在写字的时候，界面会自动往下滚着跟随；但这套跟随会盖掉我自己的滚动——我往上滚去看前面讲过的内容，它会把我拽回底部，一直拽，读到一半的内容反复被顶走。 #2072: keep the reader's scroll position while streaming; 12 checks pass, 3 pending, no failures.
• fix(docker): skip production frontend build in development #2002 — Docker development target no longer rebuilds the production frontend; 20 checks pass, 1 skipped, 16 focused tests.
• fix(notes): hand the agent a ready-made link to the page it wrote #2258 — notebook-page link fix for [Bug]:Agent创建notebook和markdown文件之后给我链接，我一点开链接，发现page not found #2097 (ready-made Markdown link); 22 checks pass, 47 notes tests.
• fix(documents): follow XLSX workbook sheet relationships #2247 — XLSX fallback follows workbook sheet relationships; 214 utils tests pass (no CI matrix yet).
• fix(storage): isolate temporary files for attachment writes #2248 — concurrent attachment writes use unique temp files; 50 storage/workspace tests pass (no CI matrix yet).
• fix(files): decode Windows-1254 text before falling back to Latin-1 #2136 — Windows-1254 before Latin-1; 158 utils tests pass, 2 xfailed ([Bug]: Windows-1254 Turkish text that is valid GBK is decoded as Chinese #2133 remains).

Issues already fixed on dev — ready to close:

• [Bug]: CJK emphasis persistence repair corrupts inline code containing dollar signs #2104 — CJK emphasis repair + inline $ (integration merge 7b307357d)
• [Bug]:Native Word equations lose their structure during DOCX extraction #2144 — DOCX OMML equations (fix(docx): preserve native OMML equation structure #2163)
• [Bug]:「资源」浮层向下溢出，窗口最大化时不可见 #2053 — resources popover viewport (fix(web): keep the composer resources popover inside the viewport #2127)
• [Bug]: UI strings containing ":" render empty or garbled (e.g. blank "Query:" label) #2177 — translation keys split on : (fix(i18n): stop splitting translation keys on ":" #2178)
• [Bug]: Parser installation guidance points to a CLI command unsupported in the official Docker image #2164 — parser installation guidance (fix(settings): provide supported parser installation guidance #2180)
• [Bug]: Remaining state-file reads use the locale encoding (GBK) instead of UTF-8, breaking migrations on Chinese Windows #2147 — UTF-8 state-file reads (fix: read state files with the writers' UTF-8 encoding #2149)
• [Bug]:PDF reference lists are silently dropped during indexing #2222 — PDF reference lists (fix(rag): index MinerU reference lists and skip page furniture #2223)
• [Bug]: Knowledge base routes return 500 instead of 404 for a missing knowledge base (v1.7.0 regression) #2184 — knowledge 404 instead of 500 (fix(knowledge): return 404 instead of 500 for a missing knowledge base #2185)
• [Bug]: Editing a persisted message silently reuses stale model and reply-language settings #2106 — edited message uses live model/language (fix(web): edited questions pick up the live model and reply language #2124)
• Add Portuguese (pt-BR) as a full interface language #2201 — Portuguese interface (feat(i18n): add Portuguese (pt) interface localization #2202)
• [Feature Request]: Turkish (tr) UI locale and reply language #2129 — Turkish interface/reply language (feat(i18n): add Turkish as an interface and reply language #2135)
• [Bug]: Memory sentence chunking misses unspaced CJK boundaries #2130 — unspaced CJK memory boundaries (fix: recognize unspaced CJK memory sentence boundaries #2131)
• [Feature Request]: Show a Little Tutor entry point next to the quoted source message #2118 — Little Tutor entry next to quoted message (feat(web): add Little Tutor message entry #2207)
• [Feature Request]: Render RTL content (Arabic, Hebrew, Persian) correctly in chat with dir=auto #2101 — RTL dir="auto" rendering (feat(web): render right-to-left text correctly with per-block dir=auto #2221)
• MarginNote search treats % and _ in a query as SQL LIKE wildcards #2214 — MarginNote LIKE wildcards (fix(capabilities): match LIKE wildcards literally in the MarginNote store (#2214) #2215)
• Navigation indexes poll archived workspaces and log a traceback for each poll #2196 — archived workspaces in navigation indexes (fix(workspace): skip archived workspaces in navigation indexes (#2196) #2197)
• [Bug]: Archived workspaces are still read by navigation indexes, logging a traceback on every poll when their folder is missing #2074 — same archived-workspace traceback (fix(workspace): skip archived workspaces in navigation indexes (#2196) #2197)
• Feature request: native xAI (Grok) TTS and STT providers #2186 — xAI TTS/STT (feat(providers): add xAI model coverage and Mistral embeddings/voice #2203)
• [Bug]: Agentic chat sends temperature to Azure gpt-5 / o-series deployments #2181 — Azure gpt-5/o-series temperature (fix(llm): leave temperature out for gpt-5 and o-series agentic calls #2187)
• [Bug]: Linked-folder sync writes a full index version per file with no retention — O(N²) disk growth and stalled API on large folders #2173 — linked-folder per-file version publish (fix: publish one LlamaIndex version per folder sync #2208; retention pruning still open)
• Legacy TutorBot migration keeps non-ASCII directory names as partner_id → partner page always shows "Partner not found" #2166 — legacy non-ASCII partner_id (fix(web): decode the partner id from the route #2167)
• [Feature Request]: Add LLMTR as an OpenAI-compatible LLM provider #2155 — LLMTR provider (feat(providers): add LLMTR as an OpenAI-compatible gateway #2156)
• [Bug]: Mastery Path re-poses a question the learner already answered correctly (fallback to the one source-verifiable item; focus packet can belong to another objective) #2073 — Mastery Path re-pose (fix: Mastery Path re-poses a question the learner already answered correctly (fallback to the one source-verifiable item #2137)
• [Bug]: Mastery Path concept objectives can never pass — every mastery_assess is refused as "teaching slice", so the tutor re-quizzes forever #2070 — Mastery concept objectives (fix: Mastery Path concept objectives can never pass — every mastery_assess is refused as "teaching slice", so the tuto #2103)
• [Bug]:kb create --docs-dir silently overwrites same-named documents and still reports success #2041 — same-named KB documents (fix(knowledge): keep same-named documents when creating a KB #2045)
• [Bug] Math Animator visual QA (VisualReviewAgent) is unreachable: enable_visual_review defaults to False and no caller passes it #2123 — Math Animator visual QA, opt-in via Expose opt-in visual review for generated animations #2142 (feat(math_animator): turn on visual review when a vision model is available #2195 still open for auto-enable)
• Docker image: rebuild on Node 22.23.4 (CVE-2026-111497) #2110 — Docker Node 22.23.4 pin (fix: pin Docker Node image to security release #2146)
• [Feature Request]: Offered chat completion reveal the answers, there should be an option to hide them. #2242 — hide the next-message hint (feat: let learners turn off the chat composer's next-message hint #2140)
• [Bug]: Figures from source-page presentation / generated images never appear where they are explained (stack at end of reply or invisible) #2036 — presented source page + duplicate image card (fix(mastery): show a presented source page at the call that presented it #2219 + fix(web): don't repeat a generated image as a card when the reply embeds it #2220)
• [Bug] chunk_with_boundary returns more than budget chunks whenever overlap_ratio > 0 #2111 — chunk budget with overlap (Integrate reviewed October PRs with correctness and compatibility fixes #2126 / e1c05b852)
• [Docs]: 说明 /api/auth/is_first_user 在认证关闭时的返回值 #2175 — is_first_user when auth is disabled (answered by design)
• [Bug]: Theme never follows the OS after the first visit, and there is no System/Auto option #2100 — theme follows the OS (fix(web): follow the OS theme by default and add a System option #2121)
• Image description model dropdown never lists any model: frontend filters on supports_vision, backend never sends it #2037 — image-description model dropdown (fix(settings): let a declared vision model drive image description #2039)
• [Bug]: Exact-substring quote check rejects faithful citations (line breaks/markup differ), so mastery_quiz fails repeatedly and the tutor re-streams the same text #2075 — faithful citations rejected by exact-substring check (fix: Exact-substring quote check rejects faithful citations (line breaks/markup differ), so mastery_quiz fails repeatedl #2090)
• [Bug]: share_mastery_history rewrites already-sent history (anchors the latest copy), breaking prompt-cache prefixes on every repeated state #2077 — share_mastery_history rewrote sent history (fix(mastery): keep shared history prefixes stable for prompt caching #2085)
• [Bug]: Agent loop bypasses OpenAICompatProvider cache_control, so Claude on OpenRouter never caches in chat/Mastery Path (0% cache reads) #2078 — agent loop bypassed cache_control (fix: Agent loop bypasses OpenAICompatProvider cache_control, so Claude on OpenRouter never caches in chat/Mastery Path ( #2081)
• Chat with PDF attachment fails on zhipu/GLM: text-only fallback retry never fires (error 1210 not matched) #2088 — GLM text-only retry never fired (fix(llm): match zhipu/GLM multimodal rejection so text-only retry fires #2089)
• [Bug]: Resend silently ignores the current model and explicit reply-language selection #2105 — Resend ignored current model/language (fix: use the current model and reply language when editing or resending #2122)
• [Feature Request]: App-wide display preferences - font size scale, line height, and font choice for chat and learning views #2152 — text size/line spacing (feat(web): add account-scoped learning text size and spacing #2176; font picker still open)
• [Bug]: Chat "Play aloud" does not play on iOS browsers (audio.play() called after TTS fetch, outside the user gesture) #2020 — iOS play-aloud (fix(chat): unlock iOS reply playback (#2020) #2021)
• [Bug]: Codex OAuth inference omits session-id and changes cache affinity when summaries update #1991 — Codex OAuth cache affinity (fix(llm): scope Codex prompt cache key to the conversation and send session-id #1992)
• [Bug]:精通之路单选题选项在作答卡片上被打乱后，回传给导师的「字母答案」与学习者实际点击的选项内容不一致 #1993 — shuffled mastery option letters (fix(mastery): show the picked option's text next to its letter in the grade seed #2000)
• KB reindex loses all chunking/embedding progress on interruption — request for checkpoint/resume in the embedding stage #1983 — KB reindex checkpoint/resume (fix(rag): resume embedding checkpoints (#1983) #2027)
• README: refresh UI screenshots and walkthroughs for v1.6.4/v1.6.5 #1262 — README screenshots refreshed (d02fb9850)
• Watching: align active transcript with video captions and center on mobile #1254 — watching transcript alignment (fix(watching): clamp transcript alignment to middle 50% and respect reduced motion #1351)
• Watching notes: preserve capture timestamps and drafts across async saves #1256 — watching note timestamps/retry (feat(watching): preserve note timestamps and add save retry UX #1350)

Batch 3 (older issues) — 11 fixed, 12 feature-implemented, 5 questions, 1 superseded; individual comments posted:

#1240, #1228, #1223, #1220, #1215, #745, #621, #473, #431, #429, #199, #1209, #1130, #1109, #992, #971, #916, #725, #684, #661, #613, #375, #304, #655, #612, #486, #440, #401, #1238.

• [Bug]: 模型正在写字的时候，界面会自动往下滚着跟随；但这套跟随会盖掉我自己的滚动——我往上滚去看前面讲过的内容，它会把我拽回底部，一直拽，读到一半的内容反复被顶走。 #2072 — chat auto-scroll hijacks the reader's position (fix(web): keep chat scroll position when the reader scrolls up #2254, pending merge)

Oct 31 merge batch — final status

By 11:28 UTC, 35 of the 43 PRs in this list are merged, including #2203, #2204, #2206 and #2216. The remaining 8:

PR	State
#2009	clean merge against current dev; ready
#2136	clean merge against current dev; ready
#2008	clean merge against current dev; ready
#2007	test-only conflict in tests/utils/test_document_images.py (new OMML tests from #2163); source merges cleanly
#2148	superseded by e1c05b852 (integrated via #2126); rebasing would add nothing
#2002	ready — 15 Docker tests passed; earlier blocker was a branch-behind artifact
#2139	blocked — throttle keys on the constant document count
#2172	blocked — missing import time

Other outcomes:

• fix: keep BibTeX quoted values that contain a braced quote #2158 merged after the fix: format BibTeX names with Jr parts and braced "and" #2157 test-append conflict was resolved.
• fix(docx): preserve native Word (OMML) equation structure in text extraction #2154 is superseded by fix(docx): preserve native OMML equation structure #2163 (merged); it now conflicts in the same files.
• feat: add Vietnamese localization #2174 (Vietnamese) merged; feat(i18n): add Turkish (tr) support and Turkish documentation #2231 (Turkish) is a separate locale PR that conflicts with feat(i18n): add Turkish as an interface and reply language #2135.
• Integration check on dev f756f371: full pytest 25 failed / 11,693 passed; all 25 reproduce on pre-batch 40c051b, so none is a regression from the batch. npm run test:node 1318 passed; typecheck clean.

Blocked

PR	Issue
#2139	Throttle keys on the constant total document count, not documents processed: 10,961 docs never checkpoint, 100/200 docs save every call. verify_sources() does not flush. Needs a processed-since-save counter and a regression test.
#2172	recovery.py calls time.time() without import time; test_journal_recovery.py gives 10 failed / 18 passed, all NameError. Needs the import before merge.

i18n merge order

The locale PRs overlap heavily, so they cannot be merged in arbitrary order:

• feat(i18n): add Turkish as an interface and reply language #2135 (Turkish) is currently CONFLICTING.
• feat(i18n): add Turkish as an interface and reply language #2135 ∩ feat: add Vietnamese localization #2174 = 35 shared files; feat(i18n): add Turkish as an interface and reply language #2135 ∩ feat(i18n): add Japanese as an interface language #2213 = 18; feat(i18n): add Japanese as an interface language #2213 ∩ feat: add Vietnamese localization #2174 = 18; feat(i18n): add Portuguese (pt) interface localization #2202 ∩ feat: add Vietnamese localization #2174 = 13.
• Suggested approach: land one new locale at a time and rebase the next; keep fix(i18n): complete the French catalog and follow it for partner replies #2204 (French completion, no new language registration) separate from the new-locale queue.
• feat(i18n): add Turkish as an interface and reply language #2135 should not be merged as-is.

Remaining queue (not pre-validated in this batch)

94 open PRs are not in the table above. The main groups:

• Large / deep review: feat(web): add account-scoped learning text size and spacing #2176 (learning text size), feat(web): add Little Tutor message entry #2207 (Little Tutor entry), feat(desktop): add lightweight Electron shell without bundled runtime #2183 (Electron shell), Feature/practice qti support #2044 (QTI, ~95k lines), feat(learner): record reading time and guardian practice view (#1988) #2017 (reading time), feat(learning): add Guided practice to rebuild a worked solution #1900 (guided practice).
• i18n cluster: feat(i18n): add Turkish as an interface and reply language #2135 (Turkish) is still CONFLICTING; feat(i18n): add complete Spanish (Spain) localization #728 (Spanish) is stale. feat(i18n): add Portuguese (pt) interface localization #2202 (Portuguese) and fix(i18n): complete the French catalog and follow it for partner replies #2204 (French) were reviewed earlier in their own threads.
• Conflicting / stale: Feature/practice qti support #2044, feat(learner): record reading time and guardian practice view (#1988) #2017, feat(sync): 端云同步路由（分域分权 / 幂等 / 冲突留档 / 大文件通道） #1999, feat(learning): add Guided practice to rebuild a worked solution #1900, fix(partners): keep engine-injected paths out of IM tool hints #1693, fix(mastery): keep learning evidence bound to objective identity #1530, feat(reading): converge immersive watching into reading (#1378) #1384, fix: require recorded learner evidence before awarding mastery #1279, feat(rag): show LlamaIndex source figures in chat #1268, feat(subagent): add ACP client backend (Part of #1028) #1272, and others older than 30 days. Most need an author rebase before they can be evaluated.
• Small stale fixes worth a fresh rebase: feat(book): complete bare LaTeX delimiters in math reading #1608, feat(partners): roam web sessions across browsers #1565, feat: bundle a MarginNote 4 native add-on for the MN4 bridge #1243, feat(launcher): add Windows startup batch script #1154, Add GitHub Actions workflow for Python package publishing #632.

Merge-order hotspots

• provider_registry.py: feat(providers): add CloudGPU as an OpenAI-compatible gateway #2138 and feat(providers): add LLMTR as an OpenAI-compatible gateway #2156 both insert after runcrate; feat(providers): add xAI model coverage and Mistral embeddings/voice #2203 also touches the registry. Merge one provider PR, then rebase the others.
• i18n registration files: feat: add Vietnamese localization #2174, feat(i18n): add Japanese as an interface language #2213, feat(i18n): add Turkish as an interface and reply language #2135, feat(i18n): add Portuguese (pt) interface localization #2202 and fix(i18n): complete the French catalog and follow it for partner replies #2204 share languages.ts, settings_spec.py, interface_settings.py and the locale catalogs; see the order above.
• Dockerfile: fix(docker): skip production frontend build in development #2002 and fix: pin Docker Node image to security release #2146 touch the same Node stages and tests/scripts/test_docker_compose.py; fix(docker): skip production frontend build in development #2002 needs clarification first.
• CI approval: many of these are first-time contributors, so the maintainer still has to approve the workflow run before checks appear. The test evidence above is independent of that gate.

How to help

Pick a PR that is not in the table above, run its focused test command, and leave a short review with the command and result. If it fails, comment with the exact failure — that is more useful than a generic “LGTM”. The maintainer merges in batches, so a few verified PRs per batch go a long way.

I did not touch the maintainer’s running instance or any PR branch.

[Antonio-Faure]

Two more from the #2289–#2294 batch:

• fix(tutor): retry a failed source attribution after a cooldown #2303 — a failed tutor-source attribution now retries after a 1 h cooldown instead of setting the terminal tutor_source_backfill_attempted flag ([Bug]: Little Tutor 来源消息回填失败也写入 tutor_source_backfill_attempted，旧线程永不重试且无手动绑定入口 #2292, automatic half). Deliberate detaches stay final.
• feat(chat): surface Little Tutor threads in the command palette #2304 — Little Tutor threads appear in the command palette with their parent conversation, so they are discoverable across conversations ([Bug]: Little Tutor / Dito 侧栏会话刷新即丢、无跨对话入口、无法开独立空白会话 #2291, discovery half).

Also independently verified two community PRs that are waiting on workflow approval:

• chore: sync web/package-lock.json and stop masking missing peers #2295: Node 22 npm ci --dry-run fails on dev (missing highlight.js/lowlight) and passes on the PR head (1080 packages). Lockfile is correct; CI still needs approval.
• fix(question): keep a quiz question whose Markdown carries a fenced code sample #2296: pytest -q tests/agents/question on the PR head — 57 passed.

[Antonio-Faure]

Nov 1 queue update (~10:00 UTC)

No merges since the Oct 31 11:28 UTC batch. The open queue is now 141 issues / 123 PRs; the increase is mostly the Windows-compatibility series (#2266–#2286) and five detailed bug reports (#2289–#2294).

Still clean from the Oct 31 batch: #2246, #2251, #2252, #2253, #2254. #2257 (@Zhi-Xiang-Guo) is also green and fixes #2255; my #2264 was closed as its duplicate.

New fixes opened this morning (CI running, no failures yet):

• fix(workspace): renumber turn-event ids when moving a conversation #2298 — turn_events.id collisions no longer block conversation moves ([Bug]: 跨工作区迁移会话被知识库依赖与 turn_events 主键冲突双重阻断（「移动到工作区」与「数据迁移」均失败） #2289).
• fix(workspace): keep an explicit knowledge-base detach detached #2299 — an explicit knowledge_bases: [] is no longer resurrected from old request snapshots ([Bug]: 依赖闭包从历史 request_snapshot 复活已解除的知识库引用，会话无法真正解绑知识库 #2290).
• fix(memory): trace chat turns in L1 and rebuild a missing L2 document #2300 — chat turns now emit an L1 trace, and a missing L2/chat.md is rebuilt instead of suppressed by stale meta ([Bug] Main chat surface never emits an L1 trace, so Memory L2/L3 can never be built from conversations #2294).
• fix(chat): persist open side-chat tabs across reloads #2301 — open Little Tutor / Dito tabs survive a reload ([Bug]: Little Tutor / Dito 侧栏会话刷新即丢、无跨对话入口、无法开独立空白会话 #2291, part 1).

Review notes posted:

• chore: sync web/package-lock.json and stop masking missing peers #2295 — right diagnosis, but CI has never run on the head; also a trailing space, an unexplained fsevents "dev": true, and Dockerfile/README still use --legacy-peer-deps.
• fix(i18n): resolve the remaining inline UI copy through the locale catalogs #2287 / test: read UTF-8 sources and fixtures as UTF-8 #2283 — red jobs are unrelated flaky tests (single-flight cleanup / sync-scheduler timing).
• fix(skills): keep GBK file names when importing skill and visualizer zips #2286 — stacked on fix(knowledge): read zip member names written without the UTF-8 flag #2279; merge fix(knowledge): read zip member names written without the UTF-8 flag #2279 first or it lands both commits.

Still untouched: #2292 (backfill marks a failed attribution as attempted, so it never retries; needs a retry/status design plus manual binding UI) and the remaining UX asks in #2291 (global side-chat entry, standalone side chats).

[Antonio-Faure]

One more feature PR from today's request:

• feat(web): global update notice with auto-check, install button and progress #2305 — global update notice with OpenCode-desktop-style behaviour: automatic check on launch (plus 6 h and tab focus), persistent card with release summary, admin “update and restart” button, live job progress through the restart, per-version dismiss, and a one-time “What’s new” dialog after the version changes. Docker/source installs get the exact command plus a copy button. It is dynamically imported, so route budgets stay green (/chat 1061/1100 KB, root shell 364/390 KB).

[Antonio-Faure]

Nov 2 update (~09:00 UTC)

No merges since the Oct 31 11:28 UTC batch; the queue is now 136 issues / 168 PRs. Most of the growth is new contributor PRs, not regressions.

Two new fixes from today's triage:

• fix(update): keep installed modules in step and record a failed restart #2357 — managed updates now pin the installed module distributions to the same target as core in one pip transaction, so a 1.7.x → 1.7.y restart cannot observe a mixed set ([Bug]: In-app updater upgrades only core, so a 1.6.x → 1.7.x update leaves all modules uninstalled and DeepTutor fails to restart #2330). It also records a failed job when the restarted launcher exits before readiness instead of leaving restarting forever. 1.6.x → 1.7.x still needs a 1.6.16 backport or the manual deeptutor[full,parse] upgrade, because the old worker predates the module split.
• fix(llm): make multi-turn GPT-5.6+ prompt caching actually reuse history #2358 — GPT-5.6+ Responses requests now send explicit cache breakpoints over the last four durable user messages. The implicit breakpoint only sits at the trailing message, which changes every turn; [Bug]:缓存命中率极低 #2297 measured 3.0k cached tokens out of 12.3k while the requests shared a 10.1k-token prefix. Gated to native OpenAI bindings, with a per-model fallback if a provider rejects the fields.

Also from the notification batch:

• fix(deps): install the Docling parser in the official image without a compiler #2328 (Docling install without gcc) is still CLEAN; [Bug]: Parser installation guidance points to a CLI command unsupported in the official Docker image #2164 is closed and the remaining failure is tracked in bug: DEEPTUTOR_EXTRAS=parse-docling fails in the official image because webrtcvad needs gcc #2327.
• chore: sync web/package-lock.json and stop masking missing peers #2295 was reduced to the lockfile only and re-verified with Node 22 (npm ci --dry-run, 1080 packages); it still needs a first-time-contributor workflow approval.
• [Bug]: Codex OAuth inference omits session-id and changes cache affinity when summaries update #1991 is already fixed by the merged fix(llm): scope Codex prompt cache key to the conversation and send session-id #1992 (session-id + stable cache key); it can be closed.

All 15 of my open PRs are CLEAN.

[Antonio-Faure]

Nov 2, 22:15 UTC — the queue is blocked on CI approvals, not on code

Open queue: 145 issues / 193 PRs. Before reviewing anything, there is one mechanical bottleneck worth clearing.

39 open PRs have all three PR workflows in action_required. GitHub is waiting for a maintainer to approve the run, so the PR page shows “no checks reported” and nothing can progress. That includes six of today’s bug-fix PRs:

• fix(learning): read lesson JSON from reasoning-model replies #2375 — lesson JSON from reasoning-model replies ([Bug]: Prepare Lesson fails with "invalid source-selection response" when reasoning_effort is high/max #2367)
• fix(llm): drop a null status from replayed Responses output items #2376 — null status on replayed Responses items ([Bug]: Responses tool continuation fails on reasoning status=null #2373)
• fix(sandbox): inherit TEMP/TMP/USERPROFILE so Windows temp resolution works #2385 — Windows TEMP/TMP/USERPROFILE in the sandbox ([Bug]: Video Lesson always fails on Windows — sandboxed child gets no TEMP/TMP, so esbuild reads C:\Windows\temp and aborts #2382)
• fix: add chat context summary prompt resources #2388 — missing context-summary prompt resources ([Bug]: Missing chat context summary prompt resources during history compaction #2370)
• fix(agents): announce ask_user pauses on the event stream #2389 — wait_for_input on the ask_user pause ([Bug]: ask_user pause never emits wait_for_input — event-stream clients hang on mastery_path / deep_question #2387)
• fix(mastery): grade concept objectives within covered and own scope (#2363, #2364) #2366 — Mastery Path concept objectives within covered scope ([Bug]: Mastery Path objective with incomplete objective-scoped packet can never be assessed; each answer re-runs preparation for minutes #2363, [Bug]: Mastery Path concept objective stuck in partial loop: tutor follow-ups don't ask for the judge's missing element, judge requires next objective's content #2364)

The rest of the blocked set, oldest first:

#1785 #1995 #2010 #2044 #2247 #2248 #2250 #2256 #2261 #2265 #2295 #2296 #2306 #2320 #2321 #2322 #2324 #2325 #2334 #2335 #2337 #2350 #2351 #2352 #2353 #2356 #2359 #2365 #2377 #2383 #2384 #2391 #2394

A few of these have been waiting since Oct 5 (#1785), Oct 19 (#1995) and Oct 21 (#2010). Approving a run does not endorse the PR; it only lets CI say whether it is reviewable.

13 older PRs have no workflow runs at all (mostly pre-dating the current workflows or path-filtered): #329 #432 #454 #459 #492 #525 #563 #620 #632 #637 #638 #1279 #1693.

Once the batch is approved, the queue becomes readable by CI state: 141 PRs already have real results, and the mergeable subset can be triaged from checks instead of guessed.

If the whole batch is too much, the small blocked fixes first

These are under 150 changed lines and at most five files, so a green CI run is quick to judge:

#2350 (7 lines) #2352 (11) #2256 (16) #2389 (33) #2295 (39) #2385 (41) #2306 (45) #2265 (46) #2010 (57) #2322 (58) #2351 (67) #2248 (88) #2296 (115) #2376 (126) #2324 (130) #2247 (142) #2334 (149)

Two size outliers in the blocked set: #2044 is 95k lines across 137 files and #2394 is 8.3k lines across 189 files. They may deserve their own review slot, or a request to split, before CI time is spent.

Three new contributions from today

• fix(rag): keep dotted identifiers as BM25 tokens #2396 — BM25 kept only 26 from 3.4.26, so a query for an exercise could not distinguish it from any other chunk containing 26 ([Bug]: Hybrid BM25 tokenizer drops one-character tokens, so dotted identifiers like "3.4.26" cannot be matched #2331). The new pattern keeps dotted identifiers and one-character tokens, applies to every BM25 path, and versions the persisted sidecar because the library does not store its token pattern. 16 focused tests pass; the wider RAG suite matches dev’s 13 environment failures.

• ci: cancel superseded runs for the same PR or branch #2397 — adds concurrency to the four PR workflows. In the six hours before this comment there were 47 pull_request runs; without cancellation, every fix-up commit leaves the previous full matrix running. No test is skipped; only superseded commits are cancelled.

• fix(workspace): keep the write probe inside the Windows path limit #2398 — the workspace write probe used a 55-character filename, so a deep custom root on Windows crossed MAX_PATH and the service reported “not writable” for a perfectly writable folder ([Feature Request]: Run the Python test suite on Windows CI (136 tests fail on Windows today) #2354, product bug 2). The probe is now 19 characters, and WinError 206 / ENAMETOOLONG gets its own actionable message. 28 workspace tests pass.

All three are open from a fork with existing merged PRs, so their checks run without approval. All three are already CLEAN (#2396 22 checks, #2397 24, #2398 22). Landing #2397 first makes the approval batch cheaper: a newer push cancels the superseded matrix instead of leaving it in the queue.

[Antonio-Faure]

Nov 6 — v1.7.2 is out; v1.7.3 candidates

dev == main == 60fb3bdfd (v1.7.2); PyPI, Docker and Desktop workflows are green. Open queue: 139 PRs / 164 issues (41 CLEAN, 32 UNSTABLE, 66 DIRTY).

Six PRs look merge-ready now (CLEAN, small, tests, no auth/migration):

• fix(runtime): keep parallel tool calls apart when the provider omits index (#2482) #2488 — parallel tool calls without a provider index ([Bug]: Runtime accumulator merges Gemini parallel tool calls when stream index is missing #2482).
• fix(knowledge): report live progress for LlamaIndex folder batches #2483 — LlamaIndex folder-batch progress.
• fix(reading): keep existing-material picker open for multiple selections #2479 — reading material picker stays open ([Bug] Existing-material picker collapses after each selection #2478).
• fix(book): match language cues as whole words #2477 — book language cues match whole words.
• fix(web): give sockets an absolute ws:/wss: URL #2424 — absolute ws:/wss: socket URLs ([Bug]:Web chat never connects — new WebSocket() receives a relative URL ("/ws?dt_workspace=…") #2404).
• fix(graphrag): keep document parsing off the API loop #2257 — GraphRAG parsing off the API loop ([Bug]: GraphRAG document parsing blocks the API event loop during knowledge-base indexing #2255).

Next review: #2490, #2487, #2486 (the zhouyao423 cluster).

Two notes:

• fix: preserve book chat scroll position while streaming #2489 is red only on the flaky cron test test_due_job_fires_and_one_shot_is_removed; a rerun is enough. Rebase it after fix(web): keep chat scroll position when the reader scrolls up #2254 to use the new isFollowing API.
• Fix reindex freeze on large knowledge bases (O(n^2) run-state serialization) #2139 still needs splitting: the O(n²) throttle is good, but it drops the load_run() takeover guard and raises the stall timeout to 4 h.

I rebased #2254 onto v1.7.2; CI is running.

The 33 issues already fixed in v1.7.1 are still all open — closing them would take the issue queue from 164 to ~131:

#2186 #2144 #2118 #2201 #2173 #2123 #2166 #1220 #2106 #2105 #2104 #1915 #1908 #2072 #1983 #2047 #2037 #2041 #612 #902 #613 #1130 #1223 #1228 #661 #486 #640 #199 #431 #473 #745 #621 #429

[Antonio-Faure]

Nov 8 — v1.7.3 candidate review

Queue: 159 PRs / 167 issues. dev == main == 60fb3bdfd (v1.7.2); no merges since the release.

Top patch candidates (CLEAN, green, reviewed today):

1. fix(embedding): restore indexed responses to input order #2512 — indexed embedding responses attach vectors in the wrong order (silent data-integrity bug).
2. fix(session): reject retained history that exceeds its plan #2511 — reject oversized retained history before sending.
3. fix(mcp): refresh persisted OAuth grants after restart #2492 — refresh persisted MCP OAuth grants after restart.
4. fix(question): retry tool summarizer without stream_options when provider rejects it #2517 — retry question summarization without rejected stream_options.
5. fix(web): keep chat scroll position when the reader scrolls up #2254 — chat/book scroll follow; rebased onto v1.7.2, 20 checks.
6. fix: pick the JSON fence when a code snippet fence comes first #2513 — choose the later JSON fence instead of an earlier snippet.
7. fix(mastery): stop tutoring-memory retries from growing every turn #2510 — bound the failed mastery-condensation backlog.
8. fix: render and resume book page chat questions #2490 — book-page ask_user render/resume.
9. fix: refuse mastery visual cards whose pixels were not inspected #2502 — reject unsupported visual mastery cards.
10. fix(parsing): report an unsettled MinerU tree as cleanup_failed #2485 — classify an unsettled MinerU tree as cleanup_failed.

Review findings posted today:

• fix(runtime): keep parallel tool calls apart when the provider omits index (#2482) #2488 — blocking ambiguity for ID-less interleaved fragments; needs a failure policy and a test. fix(runtime): separate indexless streamed tool calls #2493 is the competing approach.
• fix(knowledge): report live progress for LlamaIndex folder batches #2483, fix(book): match language cues as whole words #2477, fix(web): give sockets an absolute ws:/wss: URL #2424, fix(graphrag): keep document parsing off the API loop #2257, fix(llm): omit optional replay status for custom Responses gateways #2487 — reviewed, merge-ready from my side (targeted tests run where possible).
• fix(reading): keep existing-material picker open for multiple selections #2479 — needs an interaction test and an accessible selected state.
• fix: preserve book chat scroll position while streaming #2489 — the Python 3.11 red is the pre-existing cron flake only; rerun. No fix(web): keep chat scroll position when the reader scrolls up #2254 hook dependency.
• fix(mastery): preserve turn input evidence and completion outcomes #2486 — two contracts in one PR; needs focused completion-policy review or a split.
• Reading feat(reading): open a body image full-screen in the text and EPUB readers #2473–feat(reading): keep the marks in step across tabs #2476 — a cumulative feature stack, not a patch batch: merge feat(reading): zoom the PDF reader #2474 → feat(reading): a usable annotation workflow in the reader #2475 → feat(reading): keep the marks in step across tabs #2476; feat(reading): open a body image full-screen in the text and EPUB readers #2473 is independent; feat(reading): a usable annotation workflow in the reader #2475 is on hold for selection/accessibility.

Duplicate / competing fixes:

• fix(graphrag): run the synchronous parser in a worker thread #2352 vs fix(graphrag): keep document parsing off the API loop #2257 — same GraphRAG parser offload; prefer the older green fix(graphrag): keep document parsing off the API loop #2257.
• fix(runtime): separate indexless streamed tool calls #2493 vs fix(runtime): keep parallel tool calls apart when the provider omits index (#2482) #2488 — same [Bug]: Runtime accumulator merges Gemini parallel tool calls when stream index is missing #2482 defect; decide the ambiguity semantics before choosing one.
• The reading PRs are cumulative branches, not independent duplicates; do not count their lines as separate patches.

Hotfixes to consider before 1.7.3:

• fix(session): reject retained history that exceeds its plan #2511 + fix(mastery): stop tutoring-memory retries from growing every turn #2510 — oversized history and mastery backlog; integrate the shared memory.py edits.
• fix(embedding): restore indexed responses to input order #2512 — reindex affected OpenAI-compatible indexes after the fix; native Ollama /api/embed is not implicated.
• [Bug]CodeBuddy provider: /auth/start can permanently wedge the settings API + incompatibilities with WorkBuddy 5.9.3 desktop #2495 — CodeBuddy auth deadlock; no PR yet.
• fix(runtime): pin backend imports to the launcher installation #2506 — Uvicorn app-dir stale import after upgrade; approval-gated.

Issue cleanup: 4/33 closed (#2186, #2166, #2144, #1223), 29 pending. Six active authors were pinged again today; eleven dormant issues are candidates for bulk close.

[Antonio-Faure]

Follow-up on the P0 hotfix review:

• fix(mastery): stop tutoring-memory retries from growing every turn #2510 — request changes. The bound is not hard: with the oldest archived exchange above the plan, the captured memory prompt was 82,670 tokens against a 20,000-token plan. It needs a cap on the final projected summary request (including prior memory/instructions) and a test for the oversized-first-exchange case.
• fix(mcp): refresh persisted OAuth grants after restart #2492 — hold, security. The issuer is always stamped after a successful code exchange, including the resource-origin /register fallback. An offline HTTPX probe reproduced a resource-issued refresh token being POSTed to a different issuer after restart. Fix the provenance rule and add committed cross-origin tests.
• fix(session): reject retained history that exceeds its plan #2511 — merge as a scoped guard. It correctly rejects oversized answer history, but the check runs after _summarize(); it is not a provider-request ceiling.
• fix(embedding): restore indexed responses to input order #2512 — merge. Ordering/validation correct; native Ollama unaffected. Reindex indexes built through gateways that returned reordered data[].
• fix(runtime): pin backend imports to the launcher installation #2506 — hold on CI. The --app-dir fix works locally (stale 1.6.15 on base, correct on PR, 1/2 workers), but the three workflows are action_required with no checks.
• fix(api): let learners use the shared voice endpoints #2440 and fix: refuse mastery visual cards whose pixels were not inspected #2502 — merge-ready from the full-API and visual-workflow reviews.