Skip to content

Latest commit

 

History

History
491 lines (428 loc) · 53.2 KB

DESIGN.md

File metadata and controls

491 lines (428 loc) · 53.2 KB

obs-rust.com — design document

Revision 10 — 2026-11-18. Living document; supersedes Revision 9 (same day, in git history).

This revision records #31 and #29 as shipped (PRs #25 and #26, §2; #29 went live while it was written), runs §12 end to end as a judge would bar the cold trials, repairs what stalled or was ambiguous there (§12; D2, D3, D5, D8; shoot) and answers every D line itself (§11, "Pre-judgement"). One phrase joins BAN under the label A3; the probe, want and every label are unchanged. One item gates (§10): #32, new. Measured on https://obs-rust.com, 2026-11-18, 13:36 to 14:15 UTC (site ac1d7aa, then 09bc0e4).

Revision 9's script printed 1 FAIL line at 13:36 UTC (ac1d7aa), twice over: S4 /websocket/#demo (#29). This revision's prints 1 at 13:54 UTC (09bc0e4), twice over: A3 /examples/, the phrase added here, since the lede of /examples/ promises what its own fifth example contradicts (#32). S4 is green, and no other line printed in any of the four runs.

1. Position

obs-studio-rust is real, it is OBS, and you can run it, script it and read all of it today.

Audience Job Succeeds when
Streamers Decide to try it On the first screen of / they see the real app; within three taps they know what does not work yet and hold a verified installer for their OS
Builders Call OBS from code They pick WebSocket or MCP from one table, find a request with its fields, and send it to their own OBS without leaving the page
AI agents Answer or act correctly Given only https://obs-rust.com/llms.txt, a cold agent answers an install, API, MCP or port-process question with a citable URL in three fetches or fewer

When goals conflict: truth of a claim, then agent-readable text, then visual polish.

2. Shipped (verified on the live site, not read from PR descriptions)

Item Evidence on obs-rust.com PR
R2 P0 #1 to #6, #8; /mcp/; typed tooling. R3 P0 #2, #3: real app, agent router libobs figure from conversion.json; tools/check.py and basedpyright strict in pages.yml; self-hosted fonts, no third-party request on any page; /nope is 404; corpus schema_version 2 with typed ws-*, mcp-tool, ledger-* chunks; 30 MCP tools from the server's schema; tools/models.py. Hero is the upstream screenshot /assets/main-window.webp (34 KB) with a receipt. ## Router heads llms.txt and /ai/: 12 rows, each printed size equal to the bytes served; /ai/manifest.json lists 106 documents and every tenth matches its SHA-256 #2 to #9
R4 P0 #5, #8, #17a, #2 (part), #15a: stable and beta, "What to expect", phone and focus #stable offers v1.0.0 (7 files) and #beta offers v1.0.1-beta.20261108 (9 files); 16 of 16 hashes equal the line in that release's SHA256SUMS, on / and in site-download.md, with JS off too. Six limits beside the download, each linking blob/<40 hex>/README.md#L<n>; the cited lines support their sentence (A4). At 320, 390 and 1440 px: no sideways scroll, no text under 12 px, no control under 24 px, no nav link off screen, a :focus-visible ring. All chunk urls are absolute and every ninth distinct one is 200 #11, #12, #13
R5 P0 5b (desktop), 12, 4, 2: stable first, app first, promises reworded, figure pinned, agent path closed Eyebrow "obs-studio-rust · stable v1.0.0"; the action reads "Download v1.0.0 ↓" at 1440 px. llms.txt: "ships stable and beta builds", "Releases: stable obs-studio-rust-v1.0.0; beta obs-studio-rust-v1.0.1-beta.20261108"; router row "Install and verify the download" (A5). The screenshot is 660 px wide at 1440x900 and 343 px at 390x844, ending 0.72 and 0.74 screens down; h1 is 72 px and 38 px. None of the six banned phrases on any page (A3): "Built to stay compatible"; "clients written for obs-websocket v5 (OBS 28 and later) are expected to work against it unchanged". / links blob/fc58e78c…/conversion.json, whose libobs.all.percent is 60.36, shown as 60.4% (A6). The --dangerously-skip-permissions block opens with a .caution paragraph. 0 of 1709 chunks hold <img, <p align or shields.io; "Linux" is in the first 1500 characters of /mcp/index.md; site-control.md says the MCP server runs "on the OBS machine itself (Linux)" and "A WebSocket client may run remotely" (A7) #15
R6 P1 #30, P0 5c: search that answers; the action on the phone's first screen On /ai/#q=install%20on%20windows the first of 25 hits is site-download.md#windows-x64-arm64; SetInputSettings and wait_for put their typed chunk first (A8). Six example queries; / focuses #q; the tallest of the first five cards is 150 px, each with a copy button; #meta is aria-live and reads "25 results for …" (S5). Merged 10:40 UTC during this revision; A8 and both S5 lines are green. At 390x844 the first screen holds the link "Download v1.0.0 ↓" (outside the header) and the whole monitor, 358 px wide, ending 0.73 screens down; B "/ first screen at 390 px" prints nothing (10:36 UTC, site 3ad32a9 and later) #18, #16
R7 P1 #26: one system on every page. R8 P1 #27: every page opens on its object Probed at 1440 and 390 px on the five pages, 11:36 UTC, site b498bf9 (merged 11:19 UTC): every text size is one of the 11 on the scale (38 distinct sizes before); h1 is 72 px and 40 px on all five; radii are 4 px, 6 px and 50% only; body text is 40 em wide at most (51 and 72 before); no card row ends in a lone card. site.css declares --t-12 to --t-96 and --r-ctl, --r-plate, --r-light; the HTML of the six pages holds no font-size or border-radius px literal. All 10 S1 lines are green and no A, B or S5 line regressed. D5 ("one hand") is the judge's to answer. Merged 11:48 UTC as 705f0da; probed 12:36 UTC. On /, /websocket/, /examples/, /ai/ and /mcp/ the data-object starts 0.23, 0.47, 0.48, 0.40 and 0.15 screens down at 1440 px (limit 0.6) and 0.38, 0.63, 0.63, 0.56 and 0.58 at 390 px (limit 0.85), 342 px or wider. Lengths in screens: 6.5 / 11.9, 11.4 / 25.5 (86.9 / 156.5 before), 6.2 / 8.0, 4.2 / 8.5, 7.8 / 13.7 (20.4 / 33.8 before). All 30 MCP tools are <details> and none on /mcp/ or / is open; with JS off /websocket/ shows 225 anchored elements for 207 entries and / 16 download links with 7 hashes. All 10 S2 lines are green #20, #21
R8 P1 #28: pictures and motion, but for the picture reopened as #31 Merged 12:22 UTC as 038dd2f; probed 12:36 UTC. og:image is /assets/social-card.png, 1200x630, 44,159 bytes (a 768 KB square before); assets/obs-rust-logo.png and obs-rust-bot-logo.png are deleted and 404. 100% of links, buttons and summaries ease on the five pages (0% to 25% before), by one transition rule in site.css, which holds two @keyframes (fill, tally) and one infinite; one element animates on /, none under reduced motion. The second picture was a mockup, replaced in the next row. The 8 S3 lines of Revision 8 are green #23
R9 P1 #31: a real second picture Merged 12:53 UTC as ac1d7aa; probed 13:36 UTC. / shows three pictures 300 px or wider: the hero and upstream's two status-bar captures, 338x22 px at 1440 px and 340x23 px at 390 px (560x34 native, never enlarged). /assets/gpu-usage-label-hw.png (5,716 bytes) and -sw.png (6,302 bytes) have the SHA-256 of docs/images/ at b29f7d7a…; their alts read the labels out ("GPU: 21% with a green dot", "CPU rendering with a red dot"); one receipt links blob/b29f7d7a…/docs/gpu-status.md#L33, which reads "On an RTX 3060 (NVML), and on the same machine forced onto llvmpipe", and A4 finds its words there. "mockup" is on no page, gpu-status-red.webp is 404, and / is 6.3 / 11.8 screens (6.5 / 11.9 before). S3 / pictures is green; D1 is "yes" (§11). Unchanged at 13:54 UTC, the receipt now at f90b7a43… #25
R9 P1 #29: a console that plays Merged 13:47 UTC as 09bc0e4; probed 13:54 UTC. /websocket/#demo fills #log with ops 0, 1, 2, 5, 6 and 7 from a transcript inlined as #replay-data; #st reads "replay · scripted" and the console says "scripted from the protocol of obs-studio-rust-v1.0.0, written by hand and checked against protocol.json", both in hazard yellow. console-1440.png shows four whole frames (Identified, Request, RequestResponse, Event), each with direction, op name and indented JSON, the console brought into view. The console still starts 0.47 and 0.63 screens down; the page is 11.5 / 25.6 screens (limit 15 / 30) and 40.7 KB gzipped (limit 60); with JS off 228 anchored elements show for 207 entries. S4, S2 /websocket/ and B /websocket/ with JS off are green; D8 is "yes" (§11) #26

3. Critique of the live site: one sentence untrue

Seen as a designer, in the six §11 pictures, in tall frames of / and in computed styles, 13:36 to 14:10 UTC. Every picture on / is a capture and the console plays. What remains:

  1. /examples/ promises what it contradicts, and opens on prose (#32). Its lede says "Each one runs unchanged against upstream OBS Studio and obs-studio-rust", yet example 05 is tagged "obs-studio-rust only", and nothing measures the rest. data-object is the whole first block, so S2 is green at 0.48 screens while the code starts 0.75 screens down. On / the lede says plugins, scenes and profiles "keep working", flatly (§7).
  2. Small, outside the judge's pictures (#33): strips wider than their plate, a nav on two rows at 390 px, raw markdown in search snippets.

4. Principles and constraints

  1. Receipts, visibly. Every figure is computed at build time and shows its source, commit and date, linked at a commit and never at a branch. No receipt, no number.
  2. The project's own measures only: conversion.json, the features ledger, the releases API, protocol.json, control-api.schema.json, the README, upstream docs/images/.
  3. Show the thing, say what does not work: the real application and release assets, and their limits beside the pitch, quoted from upstream files.
  4. Fail closed. If a source cannot be fetched or parsed the build fails; stale beats wrong.
  5. Text first; static, first-party, small. Every page an agent may need has a markdown twin. No framework or third-party origin; assets are self-hosted; JS only enhances.
  6. Craft is measured. A rule of §6 a script can read (a size, a radius, a position) is checked by a script; what only an eye can judge is a yes/no question in §11 D.

Constraints from the maintainer; a change that breaks one is not done:

  • Python runs only through uv run, strictly typed: pydantic models or frozen dataclasses for all data, no duck-typed dicts, basedpyright strict clean (the pages.yml step). A new input, such as the replay transcript, gets a model in tools/models.py.
  • One generator (tools/build.py, corpus.py, mcp.py, sitedocs.py), one gate (tools/check.py), the existing pages.yml. No new workflow file. Each item lands with its check in check.py, so §11 finds nothing _site/ did not show.

5. Information architecture

Nav, on one row (two at 390 px until #33): Download · WebSocket API · MCP · Examples · AI Corpus, GitHub. Five pages and the 404. Each page has one object, marked data-object, that leads it:

/             The program monitor (the real app). Then proof strip, download with its limits,
              what changed (with one more real capture), build with it
/websocket/   The console, docked beside the reference.   /websocket/#demo replays a session
/mcp/         The agent session panel.                    /mcp/index.md is its text twin
/examples/    The first runnable snippet.                 /404.html: the badge and the nav links
/ai/          The search box, then router and feeds.      /ai/#q=<query> runs a search into #hits
/security/    The disclosure status board, then pipeline, process, archives, published advisories.
              Aggregates only from security/archives.json; zips in security/archives/ (AES-256, the
              password never in this repo); published advisories from the GitHub API. Footer link.
              /ai/docs/*.md, corpus.json, manifest.json; /llms.txt, /llms-ctx.txt, /llms-full.txt
P1 depth:     /download/ /changes/ /progress/ /control/ /websocket/<Name>.md /mcp/tools/<name>.md

6. Visual system

The brand stays. Colour has fixed meaning; a colour used outside its meaning is a bug.

Token Value Means
--bg0/1/2, --plate, --edge #0b0c0d to #24282b, seams #33383c canvas, steel plates, 1 px seams
--rust #e8631a Rust code, primary action, the one accent per view
--cu #3fa58a verified or computed: receipts, links, connected state
--haz #f2b705 caution only: beta, unsigned, untested, destructive, recorded and not live; the focus ring
--red #ff4040 tally: on air, error. Never decoration
--c, --cpp #8a96a3, #5d6b7a legacy C and C++ in data graphics
--fg, --fg2, --dim #ece8e1, #c9c5bd, #8e949a text; --dim never under 13 px

Type scale, the only sizes on the site (px): text 12, 13, 15, 17, 19; display 22, 30, 40, 52, 72; 96 for the hero numeral and the 404 only. Big Shoulders Stencil Display 900 for h1 (72 at 1440 px, 40 at 390 px, on every page), h2 (52, 30) and numerals, six words at most; Barlow 400/600 for body (17, lede 19) and UI (15); JetBrains Mono for code (13), labels and receipts (12, 13). Sizes and radii are tokens in site.css (--t-*, --r-*); a page <style> may lay out its own object but names them by token, never as a px literal. Rhythm and shape: 1180 px max, 8 px scale, body text (15 px and up) 40 em wide at most. Every page starts the same: header, 72 px (24 px at 390 px), eyebrow, h1, lede, object; on a phone / puts the monitor before the lede, to keep it whole in the first screen. Sections are 96 px apart (64 px at 390 px). Radius is 4 px on controls and tags, 6 px on plates, 50% on lights. A card row never ends in one lone card. Tables and code scroll in their own box, never the page.

Signature moments, four and only four. The program monitor: bezel, tally and dock around the upstream screenshot, 640 px or wider on desktop, 330 px or wider and whole within the first screen on a phone. The receipt: one component, .receipt, patina-green mono under every computed number and limit, its link a 40-hex permalink. The headline figure: one per view, libobs lines ported; other denominators are labelled rows beneath it or leave the page. The tally light: lit red only for live state, the WebSocket console connection.

Imagery is real: captures of the running program from upstream docs/images/, copied or re-encoded at build, each with a receipt, and a 1200x630 social card made from the app screenshot. A mockup, render or infographic is not a picture here, whatever its label says. The badge is the only drawn mark. No stock art, no shipped file that no page references. Motion has a job; CSS only, all of it in site.css; off under prefers-reduced-motion. (a) Every link, button and summary eases colour, border and background over 120 to 180 ms, by one rule. (b) The conversion bar fills once, from zero to its figure, within 1.2 s. (c) The tally pulses while the console is connected, the one animation that repeats. (d) A copy control shows "copied" for 1.5 s. Nothing else moves.

7. Content and corpus

Voice: plain, specific, verifiable; no superlatives, no absolutes ("every", "nothing", "identical"). A promise or a limit ships in one of two forms: a sentence with a line receipt (a typed Limit in build.py whose pattern must still match the upstream file, rendered with a blob/<commit>/<file>#L<n> link), or a sentence reworded until it says only what the project measures ("clients written for obs-websocket v5 are expected to work against it unchanged", not "every client works").

Streamers: the first screen names the newest stable release; each download has its size, SHA-256, verify command and limits. Builders: one computed WebSocket-or-MCP table; examples name only what exists and promise only what is measured. Agents: the router heads llms.txt.

Corpus. Tiers by rule in build.py: core (README without badges, AGENTS, CONTRIBUTING, SECURITY, CODESTYLE, CONSTITUTION, ledger summary, site/*.md, MCP and WebSocket narrative) fills llms-ctx.txt; then reference, process, archive. Chunks split on headings, open with their breadcrumb, hold no markup, and are at most 2400 characters; typed chunks (ws-request, ws-event, ws-enum, mcp-tool, mcp-resource, ledger-entry, ledger-fix, later c-symbol) are one unit each. Fields: id, doc, group, tier, kind, breadcrumb, url (absolute, resolves), source, text, chars, tokens_est, sha256.

Search on /ai/ is lexical and client-side (embedding search stays rejected). Terms match whole words or word prefixes, never the inside of a word; one- and two-letter words and "on", "the", "to", "how" count only inside the whole phrase. Rank by term frequency, then boost a match in the title or breadcrumb, an exact id, and core over process and archive. A result is a card of at most 200 px: kind, title, breadcrumb, a two-line snippet with the match marked, "Open", and a copy button (URL and chunk id). The count is #meta, aria-live. Before a query, six example queries are links (#q=...). / focuses the box.

Feeds. llms.txt: 80 links and 25 KB at most, each [title](url): first sentence, naming the newest stable release and beta; llms-ctx.txt under 300 KB; manifest.json: bytes, SHA-256.

8. Budgets and accessibility

Checked by tools/check.py on _site/, and by §11 on the live site. Live values in brackets.

Budget Limit
Third-party origins requested by any page 0 [0]
HTML gzipped: / with CSS; /websocket/; /mcp/; /ai/ 40 KB [15.8]; 60 KB [40.7]; 20 KB [14.0]; 30 KB [9.0]
First search on /ai/, gzipped (P1) 150 KB [636 KB]
Fonts, all files; largest image on / 160 KB [157.7 KB]; 120 KB [34.2 KB, the hero]
llms.txt / llms-ctx.txt 80 links, 25 KB [55, 16.7 KB] / 300 KB [230.5 KB]
Social card 1200x630 PNG under 200 KB [1200x630, 44.2 KB]; unreferenced assets 0 [0]
Page length at 1440 / 390 px, in screens / 6.5 / 12 [6.3 / 11.8]; /websocket/ 15 / 30 [11.5 / 25.6]; /examples/ 7 / 10 [6.2 / 8.0]; /ai/ 5 / 10 [4.2 / 8.5]; /mcp/ 10 / 18 [7.8 / 13.7]

Accessibility is WCAG 2.2 AA; §11 B probes it at 320, 390 and 1440 px, and with JS off; check.py keeps heading order, alt, labels, the skip link and contrast. Long pages close detail and never drop it: an entry is a <details> whose summary carries the anchor.

9. WebSocket reference and console

Sticky index left, reference centre, console (#console) docked right and in the first screen, so "Try" never scrolls the reader away. Under 1200 px it is a bottom sheet that shows 15% of the screen or more (state, the last frame, Replay) and opens on "Try". The password is never stored. The log (#log) reads as a conversation: one row per frame with data-op, direction, op name, indented JSON and round-trip time. A typed form per request is depth (item 19).

Replay. /websocket/#demo, and a "Replay a session" button, play a checked-in transcript (Hello, Identify, Identified, a request, its response, one event), inlined at build, into the log within four seconds, with the console in view; no socket opens. It says, in --haz, where the frames came from: "recorded from obs-studio-rust-v" only if they were captured from that release, otherwise "scripted from the protocol of obs-studio-rust-v". #st reads "replay" and the tally stays unlit. The build fails if a frame names what protocol.json lacks or the tag is not the newest stable release.

10. Backlog

Numbers continue Revisions 3 to 5; a letter marks a split. An item is done when its check passes in tools/check.py on _site/ and in §11 on the live site.

P1 "stunning": what §11 A, S and D judge (1 FAIL line; P0 is empty)

One PR stands between the site and an LGTM. It adds its lines to check.py, keeps every green line green, and names a size or radius by its §6 token.

Order # Item Files §11 lines it turns green
1 32 /examples/ says only what is so, and opens on code (§3 item 1). (a) The lede is one sentence and drops "Each one runs unchanged against upstream OBS Studio and obs-studio-rust"; what is true goes beside each example, as "obs-studio-rust only" already does on 05. (b) data-object moves from #hello to its .tabset (tab row, install line, code), and the paragraph "Connect, authenticate…" moves below the code, so the code plate starts within 0.6 screens at 1440 px and 0.85 at 390 px (0.67 and 0.89 today). (c) On / the lede hedges as /websocket/ does: "are built to keep working". src/examples.html (lede, line 25; data-object, line 34); src/index.html (lede, line 154); tools/check.py (fails on "runs unchanged against" in any page, and on a [data-object] in examples.html that holds no <pre> or holds an <h2>) 1: A3 /examples/. D2, and D6 for examples-1440.png. S2 /examples/ (same limits, now measuring the code plate) must stay green at both sizes, with S1 /examples/ and B /examples/.

P1 and P2: depth (does not gate the LGTM)

# Item Acceptance check
19 Typed "Try" form from requestFields; a missing required field disables Send and says why. An empty required field blocks Send with a message; localStorage never holds the password.
33 Three loose ends (§3 item 2). The strips sit inside their plate; the nav is one row at 390 px; a search snippet shows text, not markdown. In src/index.html the wide rule offsets .shot img by the token that pads .shot (--s-1), and check.py fails when the two differ; at 390 px every nav link has the same top (Gate v2's probe, item 6); no snippet in #hits holds **, a backtick or two pipes (src/ai.html).
6 Gate v2: every §11 check, S included, in check.py on _site/ (the probe runs on local files); a post-deploy step in the existing job runs §11 A1. check.py exits non-zero for each of: a table wider than its page, 11 px text, a 14 px font size, a relative chunk url, a banned phrase, a limit whose pattern is gone; the deploy job fails if /llms.txt is not 200 text/plain or /nope is not 404.
4b Home in order: one headline figure; proof strip of four receipts without "Forks"; ledger preview leads with fixed bugs; a copy control per hash; the WebSocket-or-MCP table. One figure carries the hero numeral style; "N requests", "N events" and "N typed tools" equal protocol.json and mcp/schema.json.
2b, 1, 15 Corpus residue and search index: prose url ends in its heading's #anchor; no prose chunk under 300 characters; hall of fame out of core; c-symbol chunks; deterministic output; /ai/search.json; tools/eval_corpus.py. No prose url without #; one c-symbol chunk per Sphinx directive; two builds from the same inputs give a byte-identical corpus.json apart from the timestamp; the first search transfers under 150 KB gzipped; 90% of 30 or more checked-in questions have the expected chunk in the top 5.
11, 18, 13 Markdown twins, per-item pages, sitemap: index.md per page with rel="alternate"; /websocket/<Name>.md and /mcp/tools/<name>.md; JSON-LD on /; every document in sitemap.xml; server.json valid against the vendored schema. Every alternate link resolves; a .md exists for every name in protocol.json and every MCP tool; JSON-LD's version equals the newest stable tag; the sitemap lists as many URLs as the manifest has documents; server.json validates against tools/vendor/server.schema.json.
16 Examples as files examples/<lang>/*, highlighted at build, served raw, with one MCP agent session. Every request, event and tool name used exists in protocol.json or mcp/schema.json; Python examples pass uv run python -m py_compile.
20 to 25 (P2) /websocket/obs-websocket.d.ts; /libobs/ symbol index; topic bundles /ai/bundles/*.txt; MCP packages in server.json; contributor strip; /releases.xml and a launch-speed figure from JSON. tsc --noEmit accepts the .d.ts; every c-symbol id resolves; bundle sizes are in llms.txt; the package URL and hash equal a release asset and its SHA256SUMS line; the strip equals leaderboard.json.
9, 10, 14 /changes/, /control/, /progress/ from the ledger, the two schemas and conversion.json. Entry and figure counts equal their sources; each page reads complete with JS off; one receipt each.

11. The done bar

The judge works on the live site only; nothing is taken from a build log, a PR or this document's brackets. Verdict: LGTM when the script prints no FAIL line, all three trials in C pass, at most one of D1 to D4 fails and at most one of D5, D6, D8 and D9 fails; otherwise FAIL: and the failed lines. §12 is the procedure, with the command that extracts and runs the two blocks below. Budget 10 minutes: the script runs in 1 min 37 s to 2 min 30 s (killed, and failed, at 9 min); C about five, in parallel; D four. It needs bash, curl, jq, ffmpeg and the Chromium at ~/.hermes/tools/chromium-1208/chrome-linux64/chrome, and samples nothing at random: two runs against one deploy print the same lines.

<!doctype html><meta charset="utf-8"><style>html,body{margin:0;background:#000}iframe{display:block;border:0}</style><body><script>
const P=new URLSearchParams(location.search),w=+P.get('w'),h=+P.get('h'),H=P.get('hash'),f=document.createElement('iframe');
f.width=w;f.height=h;if(P.get('nojs')=='1')f.sandbox='allow-same-origin';f.src='page.html'+(H?'#'+H:'');document.body.append(f);
f.onload=()=>P.get('shot')||setTimeout(()=>{const d=f.contentDocument,W=f.contentWindow,q=s=>[...d.querySelectorAll(s)],box=e=>e.getBoundingClientRect(),cs=e=>W.getComputedStyle(e),
 vis=e=>box(e).width>0&&box(e).height>0,px=e=>parseFloat(cs(e).fontSize),bar=e=>e.closest('header,nav,footer'),uniq=a=>[...new Set(a)],txt=s=>(e=>e?e.innerText.trim():'')(d.querySelector(s)),text=q('body *').filter(e=>vis(e)&&[...e.childNodes].some(n=>n.nodeType==3&&n.textContent.trim())),img=q('img').filter(vis),
 hero=img.filter(e=>box(e).top<h).sort((a,b)=>box(b).width-box(a).width)[0],obj=q('[data-object]').filter(vis)[0],act=q('a,button,summary').filter(vis),cta=q('a[href*="#download"],a[href*="#stable"],a[href*="/releases/download/"]').filter(e=>vis(e)&&!bar(e)&&box(e).bottom<=h)[0],
 hit=q('#hits>*').filter(e=>vis(e)&&e.querySelector('a[href]'));
d.dispatchEvent(new W.KeyboardEvent('keydown',{key:'/',bubbles:true}));
const r={w:W.innerWidth,overflow_px:d.documentElement.scrollWidth-W.innerWidth,small_text:text.filter(e=>!e.children.length&&px(e)<12).length,small_controls:q('button,summary,input,select,header a,nav a').filter(e=>vis(e)&&(box(e).width<24||box(e).height<24)).length,
 nav_clipped:q('header a,nav a').filter(e=>vis(e)&&!e.closest('footer')&&(box(e).right>w+1||box(e).left<-1)).length,hero_w:hero?Math.round(box(hero).width):0,hero_bottom_vh:hero?+(box(hero).bottom/h).toFixed(2):9,cta:cta?cta.innerText.trim():'',
 dl_links:q('a[href*="/releases/download/"]').filter(vis).length,hashes:(d.body.innerText.match(/\b[0-9a-f]{64}\b/g)||[]).length,ids:q("[id]").filter(vis).length,line_receipts:q('a[href*="/blob/"][href*="#L"]').map(a=>({href:a.href,text:a.parentElement.innerText.replace(/\s+/g,' ')})),
 third_party:W.performance.getEntriesByType('resource').map(e=>e.name).filter(n=>!/^(https:\/\/obs-rust\.com\/|file:)/.test(n)),sizes:uniq(text.filter(e=>e.isConnected).map(e=>Math.round(px(e)*100)/100)).sort((a,b)=>a-b),h1:(e=>e?px(e):0)(q('h1').filter(vis)[0]),
 radii:uniq(q('body *').filter(vis).map(e=>cs(e).borderTopLeftRadius).filter(r=>r!='0px')).sort(),measure:Math.round(Math.max(0,...q('p,li').filter(e=>vis(e)&&px(e)>=15&&e.innerText.length>120).map(e=>box(e).width/px(e)))),
 orphans:q('body *').filter(e=>vis(e)&&cs(e).display=='grid'&&e.children.length>3).filter(e=>{const t=[...e.children].filter(vis).map(c=>Math.round(box(c).top)); return t.filter(x=>x==t[0]).length>2&&t.filter(x=>x==t[t.length-1]).length==1}).length,
 screens:+(d.documentElement.scrollHeight/h).toFixed(1),obj_top:obj?+(box(obj).top/h).toFixed(2):9,obj_w:obj?Math.round(box(obj).width):0,easing:act.length?+(act.filter(e=>cs(e).transitionDuration!='0s').length/act.length).toFixed(2):0,anim:q('body *').filter(e=>vis(e)&&cs(e).animationName!='none').length,
 pictures:img.filter(e=>box(e).width>=300).length,slash:d.activeElement?d.activeElement.id:'',hits:hit.length,hit_h:Math.round(Math.max(0,...hit.slice(0,5).map(e=>box(e).height))),hit_first:(a=>a?a.href:'')(d.querySelector('#hits a[href]')),
 hit_copy:hit.slice(0,5).filter(e=>e.querySelector('button')).length,tries:q('a[href*="#q="]').filter(vis).length,count:txt('#meta[aria-live]'),ops:uniq(q('#log [data-op]').map(e=>+e.dataset.op)).sort(),state:txt('#st'),recorded:(txt('#console').match(/(recorded|scripted)[^\n]*obs-studio-rust-v[\d.]+/i)||[''])[0]};
document.body.innerHTML='<pre>'+JSON.stringify(r).replace(/</g,'&lt;')+'</pre>'},H?6000:2500);
</script>
B=https://obs-rust.com; GH=https://github.com/obs-rust/obs-studio; RAW=https://raw.githubusercontent.com/obs-rust/obs-studio
C=~/.hermes/tools/chromium-1208/chrome-linux64/chrome; PAGES="/ /websocket/ /examples/ /ai/ /mcp/"; mkdir -p shots
fail(){ echo "FAIL $*"; }; get(){ curl -s -m 60 --retry 2 --compressed "$B$1"; }; gz(){ curl -s -m 60 -H 'Accept-Encoding: gzip' "$B$1" | wc -c; }
lim(){ [ "$2" -le "$3" ] || fail "$1 = $2 > $3"; }; code(){ curl -s -m 30 --retry 2 -L -o /dev/null -w '%{http_code}' "$1"; }; export -f code
alive(){ xargs -P 6 -I{} bash -c '[ "$(code "{}")" = 200 ] || echo "FAIL '"$1"': {}"'; }   # alive LABEL < urls: each is 200, six at a time
chrome(){ timeout 60 "$C" --headless=new --incognito --no-sandbox --disable-gpu --hide-scrollbars --allow-file-access-from-files --user-data-dir="$PWD/prof" $XF "$@" 2>/dev/null; }
page(){ get "$1" | sed "s|<head>|<head><base href=\"$B$1\">|;s/history\.\(replace\|push\)State/console.debug/g" > page.html; }   # the live page, framed from a local file
probe(){ k="p$(tr -c 'a-z0-9\n' _ <<<"$*").json"; [ -s "$k" ] && { cat "$k"; return; }; page "$1"   # probe PATH W H [nojs] [hash]: layout facts as JSON, cached
  for try in 1 2; do chrome --window-size=1700,1300 --virtual-time-budget=20000 --dump-dom "file://$PWD/probe.html?w=$2&h=$3&nojs=${4:-0}&hash=${5:-}" | grep -o '{"w":.*}' > "$k" && break; done; cat "$k"; }
want(){ r=$(probe $2); [ -n "$r" ] || { fail "$1: the probe returned nothing"; return; }   # want LABEL "PATH W H [nojs] [hash]" JQ-FAULTS
  jq -r --arg l "$1" "[$3]|if length>0 then \"FAIL \(\$l): \"+join(\"; \") else empty end" <<<"$r"; }
shoot(){ page "$2"; for try in 1 2 3; do chrome --window-size=$3,$(($4+87)) --virtual-time-budget=9000 --screenshot="shots/$1.png" "file://$PWD/probe.html?w=$3&h=$4&shot=1&hash=${5:-}" >/dev/null   # shoot NAME PATH W H [hash]
  [ "$(wc -c < "shots/$1.png")" -gt 30000 ] && return; done; fail "D no screenshot of $2 at $3 px"; }
# A1 the site is up, fresh, and serves each format with the right status and type
get /llms.txt > llms.txt; d=$(grep -o 'Generated [0-9-]*' llms.txt | cut -d' ' -f2); [ -n "$d" ] && [ $(( ($(date -u +%s) - $(date -u -d "$d" +%s)) / 86400 )) -le 2 ] || fail "A1 site generated '$d', over 2 days ago"
st(){ c=$1; ty=$2; shift 2; for p; do got=$(curl -s -m 60 -o /dev/null -w '%{http_code} %{content_type}' "$B$p")
  case "$got" in "$c $ty"*) ;; *) fail "A1 $p -> $got";; esac; done; }
st 200 text/html $PAGES; st 404 text/html /nope; st 200 text/plain /llms.txt /llms-ctx.txt /llms-full.txt; st 200 application/json /ai/corpus.json /ai/manifest.json /mcp/schema.json /websocket/protocol.json
st 200 text/markdown /mcp/index.md /ai/docs/site-download.md /ai/docs/site-control.md /ai/docs/site-examples.md
# A2 weight: a phone on a slow link, and an agent with one context window
lim "A2 / html+css gzipped" $(( $(gz /) + $(gz /assets/site.css) + $(gz /assets/fonts.css) )) 40000; lim "A2 /websocket/ html gzipped" "$(gz /websocket/)" 60000; lim "A2 llms.txt bytes" "$(wc -c < llms.txt)" 25000; lim "A2 llms-ctx.txt bytes" "$(get /llms-ctx.txt | wc -c)" 300000
# A3 no retired or absolute promise is back
BAN='rust has passed c|starting soon|nothing you rely on|keep the two identical|every client written|every v5 client works|each one runs unchanged'; for p in $PAGES /nope; do get $p | grep -oiE "$BAN" | sort -u | sed "s|^|FAIL A3 $p says: |"; done
# A4 limits: three or more on /, each with a commit-pinned line receipt whose lines say the same thing
probe / 1440 900 > home.json; [ "$(jq '.line_receipts|length' home.json 2>/dev/null || echo 0)" -ge 3 ] || fail "A4 / shows under 3 limits with a line receipt"
jq -r '.line_receipts[]?|[.href,.text]|@tsv' home.json | while IFS=$'\t' read -r href text; do
  [[ $href =~ /blob/([0-9a-f]{40})/([^#]+)#L([0-9]+)$ ]] || { fail "A4 receipt is not pinned to a commit and a line: $href"; continue; }
  win=$(curl -s -m 30 "$RAW/${BASH_REMATCH[1]}/${BASH_REMATCH[2]}" | sed -n "${BASH_REMATCH[3]},$(( BASH_REMATCH[3] + 3 ))p")
  n=0; for wd in $(grep -oE '[A-Za-z]{6,}' <<<"$text" | sort -u); do grep -qiF "$wd" <<<"$win" && n=$((n+1)); done
  [ "$n" -ge 2 ] || fail "A4 the cited lines do not support: ${text:0:80}"; done
# A5 downloads: the newest stable and any newer beta, every file with the SHA-256 from its own SHA256SUMS
get / > home.html; get /ai/docs/site-download.md > dl.md
REL=$(curl -s -m 60 ${GITHUB_TOKEN:+-H "Authorization: Bearer $GITHUB_TOKEN"} "https://api.github.com/repos/obs-rust/obs-studio/releases?per_page=100")
OURS='[.[]|select(.tag_name|startswith("obs-studio-rust-v"))]'; stable=$(jq -r "$OURS|map(select(.prerelease|not))[0].tag_name // empty" <<<"$REL" 2>/dev/null)
beta=$(jq -r "$OURS|(map(.prerelease)|index(false)) as \$i|.[:\$i][0].tag_name // empty" <<<"$REL" 2>/dev/null)
[ -n "$stable" ] || fail "A5 cannot read a stable release from the GitHub API (if rate limited, set GITHUB_TOKEN)"; for tag in $stable $beta; do curl -sL -m 60 "$GH/releases/download/$tag/SHA256SUMS" > "sums.$tag"
  for doc in home.html dl.md; do links=$(grep -o "releases/download/$tag/[^\")]*" $doc | sort -u)
    [ "$(grep -c . <<<"$links")" -ge 3 ] || fail "A5 $doc offers under 3 files of $tag"
    for l in $links; do h=$(awk -v f="${l##*/}" '$2==f||$2=="*"f{print $1}' "sums.$tag"); [ -n "$h" ] && grep -q "$h" $doc || fail "A5 $doc: ${l##*/} lacks the SHA-256 from SHA256SUMS"; done; done; done
[ -z "$stable" ] || grep -q "$stable" llms.txt || fail "A5 llms.txt does not name the newest stable release, $stable"
# A6 figures and names equal their sources
sha=$(grep -oE '[0-9a-f]{40}/conversion\.json' home.html | head -1); pct=$(curl -s -m 30 "$RAW/$sha" | jq -r '.libobs.all.percent|"\(.*10|round/10)|\(.*10|floor/10)"|gsub("\\.";"\\.")' 2>/dev/null)
[ -n "$sha" ] && [ -n "$pct" ] || fail "A6 / has no conversion.json receipt pinned to a commit"; [ -z "$pct" ] || grep -qE "(^|[^0-9.])($pct) ?%" home.html || fail "A6 / does not show the libobs percent of its receipt ($pct)"
get /websocket/protocol.json > p.json; get /mcp/schema.json > m.json; get /ai/docs/site-control.md > ctl.md
for n in "$(jq '.requests|length' p.json) requests" "$(jq '.events|length' p.json) events" "$(jq '.tools|length' m.json) typed tools"; do
  grep -q "$n" ctl.md && grep -q "$n" home.html || fail "A6 '$n' is not on / and in site-control.md"; done
names=$(get /ai/docs/site-examples.md | grep -oE "requestType\": ?\"[A-Za-z]+|call\('[A-Za-z]+|\.on\('[A-Za-z]+" | grep -oE '[A-Za-z]+$' | sort -u); [ "$(wc -w <<<"$names")" -ge 5 ] || fail "A6 site-examples.md names under 5 requests or events"
for n in $names; do jq -e --arg n "$n" '[.requests[].requestType,.events[].eventType]|index($n)' p.json >/dev/null || fail "A6 an example uses $n, which protocol.json lacks"; done
get /examples/ | grep -oE 'href="https://github\.com/[^"#]+' | cut -d'"' -f2 | sort -u | alive "A6 dead link on /examples/"
# A7 agent path: router sizes, manifest hashes, chunk urls, MCP platform
rt=$(sed -n '/^## Router/,/^## Feeds/p' llms.txt | grep -o '(https://[^)]*) ([0-9]* bytes)' | tr -d '()'); [ "$(grep -c bytes <<<"$rt")" -ge 8 ] || fail "A7 the router in llms.txt has under 8 rows"
while read -r u b _; do [ -z "$u" ] || [ "$(curl -s -m 60 --compressed "$u" | wc -c)" = "$b" ] || fail "A7 router size is wrong: $u"; done <<<"$rt"
get /ai/manifest.json > man.json; jq -r '.docs|to_entries[]|select(.key%10==0).value|"\(.sha256) \(.url)"' man.json | while read -r h u; do
  [ "$(curl -s -m 60 "$u" | sha256sum | cut -d' ' -f1)" = "$h" ] || fail "A7 manifest sha256 is wrong: $u"; done; get /ai/corpus.json > c.json
jq -r '([.chunks[]|select(.url|startswith("https://")|not)]|length) as $rel | ([.chunks[]|select(.text|test("<img |<p align|shields\\.io"))]|length) as $raw
  | (select($rel>0)|"\($rel) chunk urls are not absolute"), (select($raw>0)|"\($raw) chunks hold raw HTML or badge markup") | "FAIL A7 corpus: "+.' c.json
jq -r '.chunks[].url|split("#")[0]' c.json | sort -u | awk 'NR%9==1' | alive "A7 chunk url is dead"
for id in ws:request:SetInputSettings mcp:tool:wait_for; do jq -e --arg i "$id" '[.chunks[]|select(.id==$i)]|length==1' c.json >/dev/null || fail "A7 no chunk $id"; done
get /mcp/index.md | head -c 1500 | grep -qi linux || fail "A7 /mcp/index.md does not name its platform in the first 1500 characters"; grep -qi linux ctl.md || fail "A7 site-control.md does not name the MCP platform"
# A8 search on /ai/: a plain question and two exact names each put their own document first (the first hit, whatever it links)
while read -r q ok; do want "A8 search '$q'" "/ai/ 1440 900 0 q=$q" "(select(.hit_first|test(\"$ok\")|not)|\"first hit is \(.hit_first), want $ok\")"; done <<'Q'
install%20on%20windows site-download|obs-rust\\.com/#
SetInputSettings /websocket/#SetInputSettings
wait_for /mcp/#wait_for
Q
# B what a visitor sees, from headless Chromium: probe.html reads the fetched live page in a same-origin frame
ALL='(select(.overflow_px>0)|"scrolls sideways by \(.overflow_px) px"),(select(.small_text>0)|"\(.small_text) texts under 12 px"),(select(.small_controls>0)|"\(.small_controls) controls under 24 px"),(select(.nav_clipped>0)|"\(.nav_clipped) nav links off screen"),(select(.third_party|length>0)|"third-party request to \(.third_party[0])")'
for p in $PAGES /nope; do for s in "390 844" "1440 900"; do want "B $p at ${s% *} px" "$p $s" "$ALL"; done; done; want "B / at 320 px" "/ 320 640" "$ALL"
HERO='(select(.hero_w<MINW or .hero_bottom_vh>1)|"largest image on the first screen is \(.hero_w) px wide, ending at \(.hero_bottom_vh) screens"),(select(.cta=="" or (.cta|test("beta";"i")))|"first-screen download action reads: \(.cta)")'
want "B / first screen at 1440 px" "/ 1440 900" "${HERO/MINW/640}"; want "B / first screen at 390 px" "/ 390 844" "${HERO/MINW/330}"
want "B / with JS off" "/ 1440 900 1" '(select(.dl_links<6 or .hashes<6)|"\(.dl_links) download links and \(.hashes) SHA-256 on the page")'
want "B /websocket/ with JS off" "/websocket/ 1440 900 1" "(select(.ids<$(jq '(.requests|length)+(.events|length)' p.json))|\"only \(.ids) anchored entries are visible\")"; get /assets/site.css | grep -q ':focus-visible' || fail "B site.css has no :focus-visible ring"
# S craft, measured, one line per item and page: S1 one system (#26), S2 the object leads (#27), S3 pictures and motion (#28), S4 console (#29), S5 search (#30)
SCALE='[12,13,15,17,19,22,30,40,52,72,96]'; RAD='["4px","6px","50%"]'
S1="(select(.sizes-$SCALE|length>0)|\"type sizes off the scale: \(.sizes-$SCALE|map(tostring)|join(\" \"))\"),(select(.radii-$RAD|length>0)|\"radii off the system: \(.radii-$RAD|join(\" \"))\"),(select(.measure>40)|\"body text runs \(.measure) em wide\"),(select(.orphans>0)|\"\(.orphans) card grids end in a lone card\"),(select(.h1!=H1)|\"h1 is \(.h1) px, not H1\")"
S2='(select(.obj_top>TOP or .obj_w<MINW)|"its object ([data-object]) starts \(.obj_top) screens down, \(.obj_w) px wide"),(select(.screens>MAX)|"\(.screens) screens long, over MAX")'
s2(){ x=${S2//TOP/$1}; x=${x//MINW/$2}; echo "${x//MAX/$3}"; }; while read -r p m1 m2; do want "S1 $p at 1440 px" "$p 1440 900" "${S1//H1/72}"; want "S1 $p at 390 px" "$p 390 844" "${S1//H1/40}"; want "S2 $p at 1440 px" "$p 1440 900" "$(s2 0.6 560 $m1)"; want "S2 $p at 390 px" "$p 390 844" "$(s2 0.85 330 $m2)"
  want "S3 $p eases" "$p 1440 900" '(select(.easing<0.9)|"\(.easing*100|round)% of links, buttons and summaries have a transition, want 90%")'; done <<'T'
/ 6.5 12
/websocket/ 15 30
/examples/ 7 10
/ai/ 5 10
/mcp/ 10 18
T
want "S3 / pictures" "/ 1440 900" '(select(.pictures<2)|"\(.pictures) real pictures 300 px or wider, want 2")'
! grep -qiE 'mock-?up' home.html || fail "S3 / pictures: / shows a picture it calls a mockup; §6 wants captures of the running program"
og=$(grep -oE 'property="og:image" content="[^"]*' home.html | cut -d'"' -f4); curl -s -m 60 "$og" -o og.png; [ "$(od -An -tu1 -j16 -N8 og.png | awk '{print $1*16777216+$2*65536+$3*256+$4 "x" $5*16777216+$6*65536+$7*256+$8}')" = 1200x630 ] && [ "$(wc -c < og.png)" -le 200000 ] || fail "S3 the social card is not a 1200x630 PNG under 200 KB"
want "S3 / motion" "/ 1440 900" '(select(.anim<1 or .anim>6)|"\(.anim) animated elements, want 1 to 6")'   # 901 below: a second cache key
XF=--force-prefers-reduced-motion want "S3 / with reduced motion" "/ 1440 901" '(select(.anim>0)|"\(.anim) elements still animate")'
kf=$(get /assets/site.css | grep -o '@keyframes' | wc -l); inf=$(get /assets/site.css | grep -o 'infinite' | wc -l); pk=$(for p in $PAGES /nope; do get $p; done | grep -c '@keyframes')
[ "$kf" -le 3 ] && [ "$inf" -le 1 ] && [ "$pk" = 0 ] || fail "S3 site.css: $kf @keyframes (at most 3), $inf that repeat (at most 1), $pk pages with their own (0)"
want "S4 /websocket/#demo" "/websocket/ 1440 900 0 demo" '(select([0,2,6,7]-.ops|length>0)|"the replay shows ops \(.ops|map(tostring)|join(" ")), want 0 2 6 7"),(select(.recorded=="")|"the replay does not say recorded or scripted, and from which release"),(select(.state|test("^connected";"i"))|"the tally claims a live connection")'
want "S5 /ai/ idle" "/ai/ 1440 900" '(select(.slash!="q")|"/ does not focus the search box"),(select(.tries<4)|"\(.tries) example queries offered, want 4")'
want "S5 /ai/ results" "/ai/ 1440 900 0 q=install%20on%20windows" '(select(.hits<3 or .hit_h>200)|"\(.hits) hits, the tallest of the first five \(.hit_h) px, over 200"),(select(.hit_copy<5)|"\(.hit_copy) of the first five hits have a copy button"),(select(.count|test("[0-9]")|not)|"no aria-live result count")'
# D pictures for the judge, in shots/: the live page in a frame of the stated size, over 87 rows of window that §12 crops away
shoot home-1440 / 1440 900; shoot home-390 / 390 844; shoot examples-1440 /examples/ 1440 900; shoot mcp-390 /mcp/ 390 844; shoot console-1440 /websocket/ 1440 900 demo; shoot search-1440 /ai/ 1440 1400 q=install%20on%20windows

The script is Revision 9's plus one phrase in A3's BAN, "each one runs unchanged" (§3 item 1), and 87 px in shoot: this Chromium keeps 87 px of a window for itself, so until now every picture ended 87 px short of its frame. The probe frames a local copy of the live page with history.replaceState neutralised; the 40 em measure is body text only.

C. Cold trials. Each is one fresh sub-agent with a fetch tool, no repository and no prior context, three fetches per question; the judge checks the answer against the named source.

  1. Streamer, from /: the Windows x64 file of the newest stable release with its SHA-256, and one thing to expect. Pass: the hash equals the line in sums.<stable tag>; the limit is one of the receipted sentences in home.json.
  2. Builder, from /: which surface for "a button on another PC that switches scene" (WebSocket) and for "an agent on the OBS machine that waits until recording is active" (MCP), SetCurrentProgramScene's fields, wait_for's arguments, where the MCP server runs. Pass: both choices right, fields equal p.json and m.json, the platform is Linux.
  3. Agent, from /llms.txt: the newest stable Linux AppImage with its SHA-256; SetInputSettings's fields; an MCP client config and a three-call session; how to claim an issue; the percent of libobs ported and its commit. Pass: four of five, each citing a fetched URL that holds the fact.

D. Judgement. Each line is a yes/no question; a "no" fails it and the judge names the element at fault. D1 to D4 ask whether the site is honest, D5 to D9 whether it is stunning. For D5 to D9, open the named picture (the live page at the stated size, by step 4 of §12) and answer from it and the named criteria, not from the S lines, which already count.

  1. The real thing. On the first screen (home-1440.png), could / not be mistaken for a template: the real application readable, one headline figure with its receipt, one primary action? And is each picture on / (the <img> lines of step 4, the badge aside) a capture of the running program, by its alt and receipt, none a mockup, render or infographic (§6)?
  2. Of the ten sentences step 4 prints (each page's lede, and the first text under its first h2), is none a superlative, or a promise made flatly or absolutely ("each", "every", "never", "keep working") with neither a receipt beside it nor a hedge ("expected to")?
  3. Are limits as easy to find as the pitch: in step 4's outline of /, do three or more limits with a line receipt sit inside #download, after its links, before the next section?
  4. Does colour carry only the meanings of §6: hazard marks real cautions (beta, unsigned, the permission-skipping prompt, a replay); the tally is unlit unless the console is connected?
  5. One hand. home-1440.png beside examples-1440.png, and home-390.png beside mcp-390.png: does each pair read as two pages of one site: the same header, the same start (eyebrow, h1, lede) at the same left edge and weight, the same plates and buttons? (On a phone / shows its monitor before its lede, by §6; that is not a fault.)
  6. The object leads. In each of the six pictures, is the first thing the eye lands on the page's object (the app, the console with traffic, the search box with results, the agent session, a runnable snippet), readable, with nothing decorative larger than it?
  7. (D7 is retired: S3 measures motion.) The console plays. In console-1440.png (/websocket/#demo, nothing connected): could someone new to obs-websocket follow the exchange: three or more frames in view (the log scrolls in its box; S4 counts the ops), each with direction, op name and indented JSON; "recorded" or "scripted" with a release tag, in the caution colour; the tally unlit? (#demo eases the console into view, so the picture may start below the h1, a little higher or lower from run to run; not a fault.)
  8. Search answers. In search-1440.png ("install on windows", a 1400 px tall frame): does the first result read as the answer (the site's download or install text, not a CI note); are three or more results on screen, each with kind, title, breadcrumb and a marked snippet of two lines or fewer; and are the result count and a copy control visible?

Pre-judgement, Revision 10. The design lead's own answers, by step 4 of §12 as written, from the pictures of 13:54 UTC. The judge answers afresh and takes nothing from this table.

Line Answer Seen
D1 yes home-1440.png: the AppImage capture, 660 px wide, its docks, colour bars and mixer plainly OBS; "60.4% ported to Rust" over its receipt; one orange action. All three <img> are captures, by alt and receipt (§2)
D2 no /examples/: "Each one runs unchanged against upstream OBS Studio and obs-studio-rust", absolute and untrue of example 05. /: "keep working", flat. #32
D3 yes The outline reads #download, 16 download links, 6 limits with a line receipt, #changes; a tall frame shows "What to expect" directly under the stable files
D4 yes Hazard yellow marks the replay tag and its provenance line in console-1440.png, and lower on / only the limit labels and the permission-skipping caution. The monitor's tally is grey; nothing is lit red
D5 yes Both pairs share the header, the eyebrow, h1 weight and left edge, plates and buttons. The nav wraps to two rows on both phones alike (#33)
D6 no In examples-1440.png the eye lands on two headings; the code starts 0.75 screens down, under a paragraph, a tab row and an install line (#32). The other five open on their object: the app, the console with traffic, the agent session, the search box with results
D8 yes Four whole frames, each with direction, op name and indented JSON; "replay · scripted" and "scripted from the protocol of obs-studio-rust-v1.0.0" in yellow; nothing lit red
D9 yes First result "Download OBS Studio Rust › … › Windows (x64, arm64)"; five in view, each with kind, title, breadcrumb, a marked two-line snippet and "Copy link + id"; "25 results for …". Raw ** in snippets (#33)

Two are "no", both for /examples/ and both within the allowance; #32 is their row. Today the verdict is FAIL on the script line A3 /examples/ alone. #32 should leave no "no" and no line, to be seen on the live site.

12. Judge protocol

You are the judge of https://obs-rust.com. Decide from the live site alone, by §11 of DESIGN.md on main of obs-rust/obs-rust.github.io. Edit and fix nothing; take no figure from a PR, a build log or this document. Stop at the first step that fails. No step needs rm, a cd outside a subshell or $(mktemp -d), which a harness may refuse.

  1. Run the script where DESIGN.md is as on main: a checkout at origin/main, or any directory after curl -sfO https://raw.githubusercontent.com/obs-rust/obs-rust.github.io/main/DESIGN.md. Give the command a 10-minute limit (it takes two, over many a tool's default), or run it in the background and wait for its exit. T is the first unused of judge-1, judge-2, …: mkdir refuses a used one, whose cached probes are stale. It lies outside the session's scratch space, where an image reader may refuse a file over a few KB:
    D="$PWD/DESIGN.md"; T=~/.cache/obs-rust-judge/judge-1; export GITHUB_TOKEN=${GITHUB_TOKEN:-$(gh auth token 2>/dev/null)}
    mkdir -p ~/.cache/obs-rust-judge && mkdir "$T" && (cd "$T" && awk '/^```html probe/{f="probe.html";next} /^```bash judge/{f="judge.sh";next} /^```$/{f=""} f{print > f}' "$D" \
      && timeout 540 bash judge.sh > verdict.txt; echo "exit $?, $(grep -c '^FAIL' verdict.txt) FAIL"; cat verdict.txt; ls -l shots)
  2. Read it. Pass: exit 0, no line starting FAIL, six pictures in shots/. If not, run step 1 once more in the next unused directory. A line counts when its label, the text up to the first colon, prints in both runs; exit 124 or a missing picture in both counts as script did not finish. If anything counts, skip steps 3 and 4.
  3. Cold trials, §11 C. Start the three at once, each a fresh sub-agent given only the first paragraph of C and its own trial. If you cannot start sub-agents, run each yourself with curl under the same fetch limit, answering only from the bodies you fetched. Check each answer against its named source in $T. All must pass; a failure is C1, C2, C3.
  4. Judgement, §11 D. Make copies an image reader accepts (720 px wide, 15 to 45 KB), without the 87 rows of window under each frame, then list what / says of its pictures (D1), the outline of / (D3) and the ten sentences (D2):
    mkdir -p "$T/small"; for f in "$T"/shots/*.png; do ffmpeg -v error -y -i "$f" -vf "crop=iw:ih-87:0:0,scale='min(720,iw)':-1" -q:v 12 "$T/small/$(basename "$f" .png).jpg"; done
    ls -l "$T/small"; grep -iE '<img |(Picture|Screenshot|Captures?):' "$T/home.html" | cut -c1-420
    grep -oE '<section[^>]* id="[a-z]+"|/releases/download/|/README\.md#L[0-9]+' "$T/home.html" | sed -E 's/.*id="/section #/;s/"$//;s|/rel.*|download link|;s|/README.*|limit with a line receipt|' | uniq -c
    for p in / /websocket/ /examples/ /ai/ /mcp/; do curl -s -m 60 --compressed "https://obs-rust.com$p" | tr '\n' ' ' | sed 's/<p class="lede/\n&/; s/<h2/\n&/' | sed -n '2,3p' | sed -E 's/<\/p>.*//; s/<[^>]*>/ /g; s/  +/ /g' | cut -c1-300 | sed "s|^|$p: |"; done
    Open the six in $T/small, one read each. A refused read is retried once (an image budget may refuse one read and allow the next); one refused for its size is remade with -q:v 20. For small text, at most two crops at full size: ffmpeg -v error -y -i "$T/shots/console-1440.png" -vf crop=720:450:<x>:<y> "$T/small/crop-1.jpg". A picture that still cannot be read is a "no" for its lines. Then answer D1 to D4, D5, D6, D8 and D9, each yes or no, from the pictures and, for D1 to D3, those lists. A "no" names the element at fault. Pass: at most one "no" among D1 to D4 and at most one among D5, D6, D8 and D9.

Verdict. Reply with one line and nothing else: LGTM if steps 2, 3 and 4 all passed, otherwise FAIL: <reason>, the failed labels in script order joined by "; " (the first six, then +N more), each D label with its fault: FAIL: A3 /examples/; S4 /websocket/#demo or FAIL: D6 feed cards on /ai/ outweigh the search box; D8 the frames show no JSON. A "no" within step 4's allowance is not a failure and is not reported.

13. Non-goals

No framework, bundler, analytics, cookie banner or third-party embed. No server-side search, hosted MCP endpoint or embedding service. No hand-edited copy of upstream docs. No blog, light theme or new workflow file. No claim the build cannot check. No stock art, carousel, video or scroll-jacking.