Repository navigation
[DONE] 정리 4단계 PR 11 — MQTT 벡터가 시험 브로커를 거쳐 토픽과 역할을 가상 패널에 돌려 보고 check:mqtt-vectors 가 생긴다 - #79
Merged
Conversation
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01Bqrhsc29GLhQvHPem3UshZ
The virtual panel speaks MQTT since PR 9 and 10, and nothing replayed docs/mqtt-spec.md against it. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01Bqrhsc29GLhQvHPem3UshZ
…tor from a zero count The spec does not say what the first knob/N is measured against, so the vector no longer leans on it. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01Bqrhsc29GLhQvHPem3UshZ
…ng out of the suite Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01Bqrhsc29GLhQvHPem3UshZ
…anup-4-pr11-vmdpix Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01Bqrhsc29GLhQvHPem3UshZ
…he folder is on dev Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01Bqrhsc29GLhQvHPem3UshZ
patternflow-bot
marked this pull request as ready for review
October 15, 2026 14:55
This was referenced Oct 15, 2026
Merged
engmung
pushed a commit
to engmung/Patternflow
that referenced
this pull request
Oct 15, 2026
engmung
pushed a commit
to engmung/Patternflow
that referenced
this pull request
Oct 15, 2026
Patternflow-lab#11 Patternflow-lab#12 Patternflow-lab#13 Patternflow-lab#15 Patternflow-lab#17 Patternflow-lab#14 Patternflow-lab#16 Patternflow-lab#19 Patternflow-lab#20 Patternflow-lab#18 Patternflow-lab#21 Patternflow-lab#23 Patternflow-lab#24 Patternflow-lab#25 Patternflow-lab#26 Patternflow-lab#28 Patternflow-lab#29 Patternflow-lab#30 Patternflow-lab#27 Patternflow-lab#31 Patternflow-lab#32 Patternflow-lab#34 Patternflow-lab#35 Patternflow-lab#33 Patternflow-lab#36 Patternflow-lab#37 Patternflow-lab#38 Patternflow-lab#39 Patternflow-lab#40 Patternflow-lab#41 Patternflow-lab#42 Patternflow-lab#43 Patternflow-lab#46 Patternflow-lab#44 Patternflow-lab#47 Patternflow-lab#48 Patternflow-lab#49 Patternflow-lab#45 Patternflow-lab#50 Patternflow-lab#51 Patternflow-lab#52 Patternflow-lab#54 Patternflow-lab#57 Patternflow-lab#55 Patternflow-lab#53 Patternflow-lab#56 Patternflow-lab#59 Patternflow-lab#58 Patternflow-lab#60 Patternflow-lab#61 Patternflow-lab#62 Patternflow-lab#63 Patternflow-lab#64 Patternflow-lab#65 Patternflow-lab#66 Patternflow-lab#67 Patternflow-lab#68 Patternflow-lab#69 Patternflow-lab#70 Patternflow-lab#72 Patternflow-lab#73 Patternflow-lab#71 Patternflow-lab#75 Patternflow-lab#74 Patternflow-lab#77 Patternflow-lab#78 Patternflow-lab#76 Patternflow-lab#79 Patternflow-lab#80 Patternflow-lab#81 Patternflow-lab#82 Patternflow-lab#84 Patternflow-lab#83 Patternflow-lab#85 Patternflow-lab#86 Patternflow-lab#87 Patternflow-lab#88 Patternflow-lab#90 Patternflow-lab#91 Patternflow-lab#89 Patternflow-lab#92
patternflow-bot
added a commit
that referenced
this pull request
Oct 16, 2026
… 않는다 (#94) * Add horizontal desktop printed enclosure and assembly instructions. Switch username to mbchars. * tools: case-models builds mbchars' horizontal desktop case mbchars.py puts the remix's printed parts together from its STLs (no assembled file exists): the three front sections where the joint bars' 36 mm insert pitch puts their lug holes, the covers on the sixteen inserts, the USB retainer on its boss and the stands on the bottom edge, each step checked against the README (410.28 x 176.16 x 42 mm), the board's encoder grid, the 320 x 160 mm panel and boolean overlaps. assemble.mjs learns ledTurn (the panel lies the long way here) and knobLook (the official knob, printed white), and lib/looks.mjs pla_orange, the 3MF's own #F46B16. The other three models rebuild byte-identical; the new one is 380 KB. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01QGBYm64PnfneJq2ZDJ8Gs7 * web: /build's case switch and 3D preview get a fourth case, mbchars' horizontal desktop one The card is the remix README's own facts and photos (resized to 1200 px), the preview shows the device in the case, and the switch is four tabs, two by two on a phone. A case too wide for the view at the one scale (this one, 41 cm across on a desktop's tall canvas) is drawn smaller by as little as keeps it clear of the arrows; every case still fits at the one scale on a phone, and the three upright ones on a desktop. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01QGBYm64PnfneJq2ZDJ8Gs7 * docs: changelog for mbchars' horizontal desktop case and the fourth preview case Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01QGBYm64PnfneJq2ZDJ8Gs7 * web: ZuperZuche's build joins the map - the first print, from Moscow Found by accident, with the PCBs ordered ten minutes later; the 3D printer came a month on, bought for the case, and the case was the first thing it printed. Moscow was given, so the pin is on the city. The blurb stops at the printer. The maker's note about making the patterns with AI is left out on purpose, and there are no pronouns for the maker: the message does not say. Three photos: a pixel-art Mount Fuji beside a modular synth, a pixel-art frog pond seen from above, and a cat plush in front of the panel. groups.test.ts pins the number of countries on the real map; Russia is new, so 11 becomes 12. Co-Authored-By: Claude Sonnet 5.5 <noreply@anthropic.com> * web: the 3D preview names whose case it is, in the build map's label On the Build tab the name on the case's tab (Official, Besoiobiy, SimonePDA, mbchars) stands over the device, on the build map's dark highlight. It follows the pick at once, as the tab does, and reads Official for a case whose model did not load. The knob hint moves under it while it is up. Changelog also gets the line for the arrows. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> * web, docs: the first flash starts in download mode, /flash goes to the guide's Flash chapter, and the README is new from its first screen The Play guide's 01 Flash (and /ko) puts hold BOOT, tap RST before Flash Patternflow on every first flash, assumes the Wi-Fi step after "Installation complete!" usually does not come (RST, Flash Patternflow again, Connect to Wi-Fi), and ends on the panel's own patternflow-xxxx network at 192.168.4.1 for when that does not take; the device simulator draws the NETWORK screen's hotspot. The Build guide's flash card, the Pattern tab's notes and the breadboard page give the same order. The bare /flash, which the device's Update page and every release's FLASHING.md point at, redirects to /guide/play#flash; /flash/manifest.json and /flash/bin/* are untouched. flash.test.ts pins the card order and anchors. The README opens with a one-line definition, the hero, the photosensitivity warning, four doors (Build, Play, Make patterns, Add sound) and First light; the story, community, facts and editions follow, shorter. The docs that give the first flash (firmware/README.md, PATTERN_GUIDE.md, SUPPORT.md, docs/assembly/) say the same order, the indexes list the guide pages, the three case remixes and the bundles that exist, and stale wording (the custom slots, "variant", the dead /flash page) is gone. CONTRIBUTING.md gets a row for the interactive guide's copy, and the pull request template asks for the row and the link check. From #7 (caaf2c1). Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01VaFtEmtx2AkdH1wnbQpV3Z * docs: the README opens with the fixed lines and three ways in, and First light follows the build The opening is L1 + L2 + L3 from docs/manifesto.md (Paik moves to Where it began, in L5's own words), with Play one in your browser, Build one and Get one assembled under the badges. First light moves below the build, gains the steps it was missing (unplug before seating, Origin as the sign of success, Format storage on the first Basics install) and states the J4 rule as data never power on the DevKit's ports. Stale lines are corrected against the repo: sound and MQTT shipped in editions, MIDI arrived in 3.9.0, 3.10.5 fixed four requests, and the photosensitivity warning is the manifesto's full text. From #8 (bfc4b22). Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01VaFtEmtx2AkdH1wnbQpV3Z * docs: BUILD_GUIDE flashes in the Play guide's order, and the README's campaign is still open BUILD_GUIDE §8.1 now goes download mode first (hold BOOT, tap RST), Flash Patternflow, the Wi-Fi as a second go after RST, the DevKit seated with the power bank unplugged, then the panel's own patternflow-xxxx network (K2 shows its name before the hotspot is up), then the Basics pack: the Play guide's 01 Flash card for card. §7 no longer seats the DevKit before it has firmware, so it is Wiring, and 7.4 names the step that powers up. The README's First light points at §8 again, and Crowd Supply reads as it stands: past its goal, open until 15 October; the 200% figure stays in the dated timeline. Besoiobiy's case README cites the renumbered steps. From #9 (16eca1b). Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01VaFtEmtx2AkdH1wnbQpV3Z * ## What changed **New journal post** — `fine` (ko + en), published 2026-10-13. Not highlighted. Putting the blog off but not past a week, a hand-written paper diary after months, being selected for Anthropic's Claude startup program, reading and music slipping, AI making time feel more pressing, the interview shoot for the introduction video at the startup center studio, the funding ending in two days, the apartment renovation that still hasn't happened, the visually impaired artist enjoying the Patternflow at the Seoul Station teardown, and parts bought for a workshop. **Two photos**, resized to 2000 px on the long side: - the interview studio (faces already pixelated in the source), after the interview paragraph; also the cover, since the author prefers not to lead with a face - the Seoul Station exhibition before the teardown, after the teardown paragraph ## Notes - The draft pasted the Korean section twice in `_temp/blog.md`; the second copy was clean and is the one used, so the text matches the English paragraph for paragraph (13 each). - Verified against the dev server: both language versions return 200, both images serve, and the index hero is this post with the interview photo as its thumbnail. Not in the highlighted rows. - `dev` and `main` were identical before this commit, so the PR carries only the post. 🤖 Generated with [Claude Code](https://claude.com/claude-code) BODY journal: new post "fine" Post (ko+en), not highlighted: the paper diary, the Claude startup program, the interview shoot for the introduction video, the funding ending in two days, the apartment still unrenovated, and the visually impaired artist enjoying the Patternflow at the teardown. Two photos at 2000 px: the interview studio (faces already pixelated), also the cover, and the exhibition before the teardown. The draft pasted the Korean section twice; the clean second copy is used. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> * docs: the cleanup has a baseline From #11 * docs: cleanup step 1 is planned From #12 * docs: the v2 build guide moved to `docs/legacy/BUILD_GUIDE_v2.md` From #13 * docs: the doc map lists the Korean/English pairs From #15 * docs: the doc map reads the YAML under `.github/` From #17 * docs: `CHANGELOG.md` is split by release into `docs/releases/` From #14 * docs: the community files moved to `.github/` and the audio guide to From #16 * docs: cleanup step 2 is planned From #19 * docs: cleanup step 3 is planned From #20 * docs: the project's writing voice is written down From #18 * docs: cleanup step 4 is planned From #21 * docs: the docs index and the repository map catch up with step 1 of the From #23 * docs: cleanup step 5 is planned From #24 * web: `check:features` fails when /features and the feature folders From #25 * web: `check:links-absolute` fails on a GitHub link into the repository From #26 * hardware: `check:bom` and `check:bom-schematic` hold the BOM to the From #28 * web: `check:editions` fails when the shelf and the bundles disagree From #29 * web: `check:presets` fails when a web preset cannot reach a panel From #30 * web: `check:human-lock` fails a pull request that changes a word of the From #27 * web: `check:changelog` fails when `CHANGELOG.md` and `docs/releases/` From #31 * web: every pull request runs the drift checks and the text lock From #32 * web: `check:roadmap` fails when a /roadmap node and its issues disagree From #34 * docs: the contributor docs say what CI runs on every pull request From #35 * web: `check:voice` points at the habits of generated prose, line by line From #33 * docs: `AGENTS.md` carries the writing-voice rule and the step 2 check From #36 * docs: the doc map is regenerated at the close of step 2 of the cleanup From #37 * web: five of the site's twelve lint warnings are gone From #38 * web: the seven `<img>` lint warnings carry a named exception From #39 * docs: where the panel size lives, line by line From #40 * web: pattern Lab's July storage keys say who wrote them and when From #41 * web: the site's redirects and old routes say what links them From #42 * web: web CI fails on a lint warning From #43 * firmware: the MIDI page counts 24 detents a turn, as the encoder has From #46 * web: the community's old formats say what wrote them and when From #44 * firmware: the vendored HUB75 driver names the upstream commit it came From #47 * firmware: the vendored HUB75 driver takes upstream's fix for its config From #48 * web: a new TODO or FIXME turns the drift checks red From #49 * web: one file holds the values the firmware and the site both carry From #45 * docs: the any-resolution RFC has a draft From #50 * docs: the Linux player has a draft RFC From #51 * firmware: the firmware reads the values it shares with the site from From #52 * web: the site's lint and test tooling is on its latest releases inside From #54 * docs: `AGENTS.md` names `shared/`, the HUB75 driver's `UPSTREAM.md` and From #57 * docs: the two step 5 drafts are listed in `docs/README.md` From #55 * web: the site reads the panel size and the knob numbers from From #53 * web: drizzle is on its latest patch releases From #56 * docs: the repository maps catch up with step 3 From #59 * web: tailwind is on its latest 4.x release From #58 * web: react Three is on its latest releases inside the current majors From #60 * web: zustand is on its latest release inside the current major From #61 * web: mediabunny is on its latest release inside the current major From #62 * web: changes from #63 From #63 * web: posthog-js is on its latest release inside the current major From #64 * web: better-auth is on its latest release inside the current major From #65 * web: react is on its latest release inside the current major From #66 * web: next is on its latest release inside the current major From #67 * docs: the doc map is regenerated at the close of step 3 of the cleanup From #68 * web: the guide's simulated board takes input that no hand made From #69 * tools: a virtual panel for the contract vectors From #70 * tools: the virtual panel answers device state over REST From #72 * tools: the virtual panel answers the pattern routes From #73 * tools: contract vectors have a format and a replay From #71 * web: `check:virtual-panel` runs every preset on the virtual panel From #75 * tools: the virtual panel speaks MQTT From #74 * tools: the virtual panel answers `/api/mqtt` From #77 * tools: the virtual panel speaks OSC From #78 * docs: the REST contract has its first vectors From #76 * docs: the MQTT contract has its first vectors From #79 * tools: oSC has contract vectors From #80 * docs: the REST vectors reach the rest of the core From #81 * docs: the MQTT vectors reach the channels and the snapshot, and From #82 * tools: the virtual panel takes the audio WebSocket From #84 * tools: the virtual panel speaks RTP-MIDI From #83 * tools: the audio WebSocket has contract vectors From #85 * tools: the virtual panel sends MIDI back From #86 * tools: mIDI has contract vectors From #87 * tools: mIDI out and `/api/midi` have contract vectors From #88 * docs: the maps of the tree name the virtual panel and the contract From #90 * docs: the REST and MIDI specs say one thing where they said two From #91 * docs: `AGENTS.md` names the virtual panel, the contract vectors and From #89 * docs: the doc map is regenerated at the close of step 4 of the cleanup From #92 * docs: changelog for 81 workshop fork PRs #11 #12 #13 #15 #17 #14 #16 #19 #20 #18 #21 #23 #24 #25 #26 #28 #29 #30 #27 #31 #32 #34 #35 #33 #36 #37 #38 #39 #40 #41 #42 #43 #46 #44 #47 #48 #49 #45 #50 #51 #52 #54 #57 #55 #53 #56 #59 #58 #60 #61 #62 #63 #64 #65 #66 #67 #68 #69 #70 #72 #73 #71 #75 #74 #77 #78 #76 #79 #80 #81 #82 #84 #83 #85 #86 #87 #88 #90 #91 #89 #92 * 시작 — 정리 마지막 기록을 원래 레포 dev 위에 다시 짠다 Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01DpjXsYZbwT5PpwzpNCEEqq * docs: the doc map lists the six links the cleanup's closing entry adds The progress log's entries for steps 1 to 5 and the gate link contract-vectors.md, the doc-parity list, the panel-geometry record and the two RFC drafts. The cleanup RFC's status line links the log. Each of the six resolves. The log and the RFC come over with the cleanup's records, so until they land the map is six rows ahead of the tree. From #93 (2f094fe). Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01DpjXsYZbwT5PpwzpNCEEqq --------- Co-authored-by: Mykyta Bilous <belousnikita@gmail.com> Co-authored-by: Claude <noreply@anthropic.com> Co-authored-by: Lee Seunghun <122682380+engmung@users.noreply.github.com> Co-authored-by: pi-hub <pi-hub@localhost>
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
ops / https://claude.ai/code/session_01Bqrhsc29GLhQvHPem3UshZ
정리 4단계 계획서 표의 PR 11 이다. MQTT 계약의 토픽 표와 두 역할을 벡터 아홉 개로 적고, 벡터마다 시험 브로커를 하나씩 띄워 가상 패널에 돌려 보는 check:mqtt-vectors 를 더했다. 머지되면 mqtt-spec.md 의 문장이 바뀌거나 가상 패널이 그 문장과 어긋날 때 PR 마다 drift 검사가 빨개진다.
좋아지는 것: 없음
할 일
web/scripts/drift/_broker.ts: 연결, 구독, 발행, retain 만 하는 MQTT 3.1.1 브로커와mqttSenderweb/scripts/drift/mqtt-vectors.ts:replay("docs/mqtt-vectors", { mqtt })docs/mqtt-vectors/topics-*.json아홉 개와README.mddocs/rest-api-vectors/README.md를 링크한다ffb5d2a에서 Drift checks(38754668255), Docs point at files that exist(38754668222), Web CI(38754668331) 셋 다 success975433a그대로라 merge 할 것이 없다한 일
_broker.ts는 벡터마다 127.0.0.1 의 빈 포트에 브로커를 하나 띄운다. CONNECT 엔 CONNACK, SUBSCRIBE 엔 QoS 0 SUBACK 과 그 토픽에 남은 retain 메시지, PINGREQ 엔 PINGRESP, DISCONNECT 엔 끊기다. 토픽은 글자 그대로만 맞춘다(와일드카드 없음,mqtt-spec.md12행). PUBLISH 는 그 토픽을 구독한 연결마다 retain 비트를 끄고 보내고, retain 이면 브로커에 남긴다. 빈 retain 은 남은 것을 지운다.mqtt.ts의 framing 을 가져다 쓰지 않고 따로 짰다.Sender 의
open()은POST /api/mqtt에 host 와 port 만 보낸다. 패널은 Broadcast 에 역할 off 로 남아 아무 데도 걸지 않는다. 역할은 벡터가 첫 스텝에서 고르고(POST /api/mqtt?role=subscriber), 패널이 다음 프레임에 걸므로 그다음 스텝은wait다.send는{ topic, payload }를,"retain": true면 남기면서, 구독자에게 보낸다.heard는 패널이 보낸 PUBLISH 마다{ topic, payload, retain }이다.settle()은/api/mqtt의state가connecting이 아니고, setImmediate 세 번과 HTTP 왕복 한 번 동안 브로커를 오간 바이트 수가 그대로일 때 끝난다. 2초를 넘기면 던진다.TCP 에서 두 번 막혔다. 처음엔 연결 직후 메시지가 넷쯤 사라졌다. 브로커가 SUBACK 열한 개를 한꺼번에 쓰면 Nagle 규칙 때문에 첫 것 뒤가 패널 쪽 커널의 지연 ACK(리눅스 40 ms)를 기다렸고, 벡터가 보낸 PUBLISH 가 그 뒤에 섰다. 패널은 CONNACK 4바이트와 SUBACK 둘(10바이트)만 받고, 브로커가 그 뒤에 쓴 110바이트는 실제 시간 40 ms 뒤에야 닿았다. 브로커 소켓에
setNoDelay(true)를 켰다. 반대쪽도 같았다.mqtt.ts의 소켓은 Nagle 이 켜져 있어, 발행자가 잇달아 보내는 두 번째 메시지가 브로커 쪽 커널의 지연 ACK 를 기다렸다. 그 파일은 PR 9 행이라 고치지 않았다. 브로커가 패널의 PUBLISH 마다 PINGRESP(2바이트)를 돌려보내 ACK 가 바로 가게 했다. 패널은 PINGRESP 를 keepalive 기다림을 푸는 것 말고는 무시한다.벡터 아홉, 저마다 스펙 한 문장이다.
topics-sleep.json: 32행 「Publishing1to<prefix>/sleep…」.docs/contract-vectors.md의 MQTT 예에 역할과 연결 스텝을 더했다.topics-sleep-words.json: 26행, 낱말 열(1, 0, on, off, true, false, sleep, wake, toggle 두 번)마다 status 와sleep/state.topics-sleep-state.json: 27행. 구독자로 붙자0하나만, REST 로 재우자1, 발행자로 다시 붙자1. 셋 다 retain 없이.topics-knob.json: 39행.knob/2에 0, 5, 5, -2, 25 를 보내면 패턴이 받은 클릭 합이 0, 5, 5, -2, 25.topics-param.json: 23행.param/3을 750, 1000, 0 으로 잡고 빈 페이로드로 놓는다.topics-pattern.json: 24행. 이름Wave Saw,Origin, slugwave_saw.topics-publisher-ignores.json: 41행. 발행자에게 knob, param, pattern 을 보내도 패널은 그대로이고 아무것도 보내지 않는다.topics-either-role.json: 32행. 구독자와 발행자 둘 다message배너를 띄우고(GET /vp/mqtt), 발행자가sleep을 따른다.topics-publisher.json: 13행. 손이 돌린 K2 의 누적 클릭-3,-2와 고른 패턴이 나간다.증거
로컬, web/:
doc-map 에 이 PR 의 새 줄은 넷이고 다 닿는다. README 에서
docs/mqtt-spec.md,docs/contract-vectors.md,tools/virtual-panel/README.md,docs/rest-api-vectors/README.md로.Nagle 을 찾은 실행. 브로커와 패널 소켓에 로그를 잠깐 달았다(커밋하지 않음). 구독자로 붙인 뒤
sleep을 20 ms 마다 1, 0, 1, 0, 1 로 보냈다:GitHub, 이 PR:
ffb5d2a에서 「Drift checks / check」(38754668255,PF_DRIFT_STRICT=1, human-lock 포함), 「Docs point at files that exist / check」(38754668222), 「Web CI / build」(38754668331)가 success. dev merge 전5c458ae의 strict drift 로그에도passed mqtt-vectors가 있다(job 116487316083).실행법
가정
docs/mqtt-vectors/README.md,docs/mqtt-vectors/topics-*.json,_broker.ts,mqtt-vectors.ts)에만 더했다.CHANGELOG.md,docs/rfc-cleanup-progress.md, doc-map CSV,mqtt.ts,mqtt-http.ts,_vectors.ts,docs/contract-vectors.md는 안 고쳤다. 라벨은 앞 PR 들처럼 달지 않았다._broker.ts에 둔다. D10 은 스위트가 송신기를 자기_파일에 둔다고 하고, 표가 k11 에 준_파일이 이것 하나다.POST /api/mqtt?role=subscriber그대로다.wait마다mqttConnected: true를 본다. 붙지 못했을 때 벡터가 뒤의 엉뚱한 줄이 아니라 그 자리에서 실패한다.settle()이 실제 시간으로 40 ms 넘게 기다리는 것이다. 지연 ACK 는 운영체제마다 달라 고르지 않았다.mqtt.ts가setNoDelay(true)를 켜면 필요 없어진다(for another k).0을 보내고 한 번 기다린다. 처음 받은 수를 무엇과 비교하는지 스펙이 말하지 않아서다(D8). 0 이 기준이면 어느 읽기로도 움직임이 0 이다. 그 뒤wait는 한 프레임에 온 둘을 합치는 구현(펌웨어의 coalesce)에서도 같은 결과가 나오게 둔 것이다.pattern의 페이로드는"<string>"이다. 24행이 「display name or slug」 라 이름만 맞다고 할 수 없다. 가상 패널은 이름을 보낸다.message의 retain(broadcast: yes · channels: no),/api/mqtt. README 의 마지막 문단이 벡터 없는 곳을 적는다(Director 토픽, 다시 거는 간격, 로그인 거절).firmware/와web/src/를 건드리지 않아build.sh all,check:ci,check:panels는 §2.1(1) 대상이 아니다. 계획서대로 typecheck, 이 PR 의 check, 깨뜨린 실행을 적었다.Changelog
Docs
docs/mqtt-vectors/holds nine, for the topic table and the two roles indocs/mqtt-spec.md: the sleep words andsleep/state,param/Nand its empty payload,knob/Nas a difference,patternby name or slug, and what a Publisher ignores and sends.npm run check:mqtt-vectorsinweb/replays them against a fresh virtual panel each, through a test broker of its own (web/scripts/drift/_broker.ts), and the drift workflow runs it on every pull request.for another k
"retain": true, so a retained snapshot can sit on the broker before the panel subscribes, and the broker hands it over on SUBSCRIBE with the retain flag set. Eachoutentry carriesretain.open()leaves the panel on Broadcast with the role off, so a channel vector starts withPOST /api/mqtt?channel=ch1(orlive,5) and await.rest-vectors.ts(k6's, now on dev) passes no sender, so adocs/rest-api-vectors/mqtt-*.jsonwith anmqttstep throwsthis suite brings no sender for mqtt, and one without it reads/api/mqttwith no broker set (stateno broker configured). For/api/mqtton a connected panel,rest-vectors.tsneedsimport { mqtt } from "./_broker"andreplay("docs/rest-api-vectors", { mqtt }). A vector opens a channel only for the kinds its steps use, so the other REST vectors pay nothing. That file is in k6's row, not k12's.AGENTS.mdcan listnpm run check:mqtt-vectorsanddocs/mqtt-vectors/besidedocs/mqtt-spec.md.docs/README.mdnamesdocs/mqtt-vectors/, andweb/ARCHITECTURE.mdorweb/README.mdcan say_broker.tsis a test broker, not a client.tools/virtual-panel/transports/mqtt.tsdoes not callsetNoDelay(true)on its socket. Against a broker that delays its ACK, the panel's second publish in a row waits about 40 ms._broker.tsworks around it (가정, fifth item).for an issue
docs/mqtt-vectors/topics-param.json(step 09, an emptyparam/320 ms after step 07 wrote0).docs/mqtt-spec.mdline 23, "empty payload releases the hold", and line 38, "An empty payload releases it explicitly." In the firmwareflushCoalescedRemote()(features/mqtt/core_mqtt.hline 648) callsPatternflowBus::releaseAbsolute(), which keeps a channel written less thanABSOLUTE_RELEASE_GRACE_MS(250 ms) ago (src/core_bus.hlines 106 to 111).releaseAbsoluteNow()beside it has no grace.docs/mqtt-vectors/topics-knob.json(step 11, a count 27 past the last).docs/mqtt-spec.mdline 39, "The panel diffs it against the last value it received and injects the difference as a detent delta."applyRemoteKnob()cuts the pending delta to ±16 (KNOB_PENDING_FLUSH,features/mqtt/core_mqtt.hlines 120 and 599 to 603).Upstream
npm run typecheck,npm run lintandnpm run check:drift(mqtt-vectors among them) pass; firmware/ and web/src/ are untouched.🤖 Generated with Claude Code
https://claude.ai/code/session_01Bqrhsc29GLhQvHPem3UshZ
Generated by Claude Code