Four-stroke engine cutaway, built by an AI agent

The video

Open the live player

What the player is. It is a repackaging of the agent's finished animation files (the same drawing, timing and engine code as the video), wrapped with a Play/Pause button and a time slider by the owner's assistant. It draws the animation live in your browser instead of playing a video. It plays by itself, unless your device asks for reduced motion, in which case it waits for you to press Play. Add #t12.5 to the address to open it paused at 12.5 seconds. The player runs scripts, makes no requests to other sites, and uses the fonts on your own device (Inter and JetBrains Mono if you have them installed), so its letter shapes differ slightly from the video.

How it was built

The four-stroke engine video (32 seconds, 1920 x 1080, 30 frames per second, silent, 960 frames) was built by an AI agent, Hermes Agent 0.23.0, running on a Windows workstation with an AMD Ryzen Threadripper PRO 7975WX (32 cores), 192 GB of DDR5 RAM and three NVIDIA GPUs (two GeForce RTX 4090 with 24 GB each and one RTX 5090 with 32 GB). The agent's model was GLM 5.3 Flash, run locally with the Project Maya runtime. It was a heavily compressed build (IQ2_S, 116 GB) spread over the three GPUs and system memory. Until about 10:50 on 9 November (rounds 1 to 3 and the first attempt at round 4) it was served as four parallel seats of 131,072 tokens each; from stop 4 on it was served as two seats of 524,288 tokens each, with the same model, files, quantization and hardware (see the paragraph on rounds 4 to 10). The agent's configuration asked for a very high thinking level ("xhigh"). The local router's log shows "high" for the 442 successful agent calls up to 10:32 on 9 November; that the router treats "xhigh" and "max" the same as "high" is from the owner's notes. The model server's own "high" setting is a thinking budget of 32,768 tokens, then the answer (server log).

Two kinds of work went to other models. NVIDIA's Nemotron 3.5 Bolt, a 74-billion-parameter model in a 2-bit build, ran on one of the two graphics cards in a Linux server with an AMD Ryzen 9 9950X3D2 (16 cores), 96 GB of DDR5 RAM (89 GB usable) and two AMD Radeon AI PRO R9700 cards with 32 GB each. Up to 10:33 on 9 November it handled 75 text calls inside four sub-agent sessions (reading documentation and reviewing frames) and 55 picture checks. Up to the same time GLM itself did 39 picture checks and wrote the 23 summaries that shrink the conversation to make room.

The counts in this paragraph and in the table below stop at 10:33 on 9 November, which covers rounds 1 to 3 and the first, unsuccessful attempt at round 4 (they were not recounted for rounds 4 to 10, apart from the figures in the next three paragraphs). Up to then the build took 15 hours 49 minutes of wall-clock time, of which about 8 hours was active work. The lead model made 415 calls and the agent ran about 530 tool calls, writing 584 thousand output tokens. The conversation was compacted 21 times, the router refused four requests as too long, and the agent hit its limit of 60 calls per turn once, after which the limit was raised to 150. The video itself is a HyperFrames 0.9.26 project: one hand-built SVG drawing, animated by a GSAP 3.14.2 timeline that serves only as a clock, and rendered through headless Chrome 152 and ffmpeg 6.1.1 to H.264. Up to that point there were six renders (three drafts, three final), each taking 25 to 38 seconds.

Rounds 4 to 10, part 1 (stops 3 to 8, rounds 4 to 8). The first request for round 4 (stop 3, delivered at 09:48) produced no edit in about 45 minutes. The agent log shows two single thoughts of about 32,900 and 30,800 output tokens (about 11 minutes each) and two context compactions of about 4 minutes each after the prompt arrived (a third compaction finished at 09:40, about eight minutes before it); the turn ended at 10:33 with the router's "chat too long" refusal. Stop 4 re-sent round 4 at 12:27 with the design already decided; its renders were written at 12:46 and the turn's report came at 13:02. Stop 5 (13:02) and stop 6 (13:09) followed at once; their renders were written at 13:04 and 13:22, and their turns ended at 13:09 and 13:24. Stop 7 (13:30) and stop 8 (13:55) followed the same way: their renders were written at 13:32 and 13:33 (stop 7) and at 14:11 and 14:12 (stop 8), and their turns ended at 13:38 and 14:14. The five turns used 67, 19, 21, 20 and 47 model calls (the api_calls figure on each turn's "Turn ended" line in the agent log; short follow-up turns of one to three calls that the log shows after them are not counted). Rounds 4 to 8 added ten renders (five drafts of 24.9 to 29.4 seconds and five finals of 37.6 to 38.2 seconds, by the render tool's own timing), sixteen in all up to stop 8. Between 10:27:08, when the last compaction before stop 4 finished, and 14:14 there was no compaction. The next one started at 14:14:58, a second after the stop 8 turn ended, and finished at 14:16:20 (the log: 403 messages cut to 64, a rough token estimate of about 290,500 down to about 33,200; the database has one row starting "[CONTEXT COMPACTION" after stop 4, written at 14:16:17). It started at 290,503 tokens. The agent's configuration asked for compaction at 60 percent of its window from about 12:15, but the log's own threshold lines do not settle which threshold applied (one quotes 85 percent of 491,520, which is 417,792 tokens; another shows an effective threshold of 368,640). No model-written summary was produced: the summary call failed with the logged message "Auxiliary chat stream stalled: no new output for 60.0s (77.3s elapsed, timed out)", and the log records a fallback ("deterministic_fallback", 341 items dropped), so the older messages were dropped without a summary. The stall was a 60-second no-output limit; a second large request (247,518 tokens in, 70.9 seconds, logged at 14:16:20) was running at the same moment. Afterwards that limit was raised to 300 seconds in the agent's configuration (auxiliary.compression.no_progress_timeout, about 14:59 on 9 November). From 12:27 to 14:14 the database holds 40 distinct assistant tool calls to vision_analyze, counted by tool-call id (stop 4: 15, stop 5: 3, stop 6: 6, stop 7: 4, stop 8: 12); it has 52 assistant rows with such calls, because 12 rows are exact replays (same ids and timestamps) written after the 14:16 compaction. The agent's configuration sends these picture checks to the Bolt model, with GLM as the fallback after a timeout; how many fell back was not counted. The model server changed between stop 3 and stop 4: GLM 5.3 Flash on the same machine, with the same files and quantization, went from four seats of 131,072 tokens of context each to two seats of 524,288 each (through two seats of 262,144 at about 10:55; the final layout from about 11:27). Only at about 12:15 was the agent's own context window raised from 122,880 to 491,520 tokens, and its configuration asked for compaction at 60 percent of it instead of 75 (the log does not settle which threshold applied at 14:14). The earlier rounds therefore ran on the old layout and stop 4 onward on the new one; the model, hardware and quantization did not change.

Rounds 4 to 10, part 2 (stops 9 to 11, rounds 9 and 10). Stop 9 (16:56:55) is the owner's own typed request, not a prompt written by the assistant. The agent worked on it in two turns and made no render. In the first turn it asked four clarifying questions with its form tool (at 16:59:35); the form went unanswered, the tool waited 625 seconds, and the turn ended at 17:10:13 after 6 model calls. At 17:16:17 the owner sent his marked screenshot, and the second turn ran from then until 17:54:46 and used 55 model calls (the agent log's api_calls figure; the router's log also shows 55 calls to the lead model in that span, although the brief written afterwards says 35). In it the agent asked a second set of questions, which the owner answered after 593 seconds (17:34:38), and a third set at 17:54:15, which was cancelled unanswered 31 seconds later, when the turn ended at 17:54:46 (reason "interrupted_by_user" in the log; who interrupted is not recorded). The owner's assistant then read his screenshot and wrote the decided brief (stop 10, 18:12:31, round 9). Its first turn ended at 18:25:05 after 46 calls (the brief had asked the agent to stop at about 40 tool calls); a short "continue" at 18:25:44 told the agent to back up the older renders and render; the draft was written at 18:26:35 (24.0 seconds by the render tool's own timing) and the final at 18:27:18 (36.8 seconds); and the second turn ended at 18:30:05 after 10 calls. Stop 11 (18:36:38, round 10) ran in one turn that ended at 18:50:26 after 27 calls; its draft was written at 18:47:04 (25.3 seconds) and its final at 18:47:48 (37.1 seconds), which is the video above. Together stops 9 to 11 used 144 model calls in five turns (6 + 55, 46 + 10, 27; one-to-three-call follow-up turns are not counted, as before). Rounds 9 and 10 added four renders, twenty in all. There was no context compaction after 14:16:20: the agent log has none, and the database has no row starting "[CONTEXT COMPACTION" after the one written at 14:16:17. The router's log puts the agent's context at 53,734 tokens at the first call of the second stop 9 turn and about 135,000 tokens (the router's real count for that call; its exact count was 139,987) at its last. The database holds 25 distinct vision_analyze tool calls from stop 9 on, counted by tool-call id (stop 9: 13, stop 10: 11, stop 11: 1), and no replays this time.

Rounds 4 to 10, part 3: the agent's thinking degraded during stop 9. The agent's database keeps the thinking behind each reply. Counting the share of "=" characters in it (thoughts of 300 characters or more, by clock hour), the share was 0.3 to 1.5 percent in every hour of the build before 15:00, and 11.7 percent in the 17:00 hour, when 53 percent of the thoughts were above 5 percent and 16 of 32 ended in a long run of "= = = =" (the longest run 510 characters); up to 31 percent of the characters of a single thought were "=". Inside the second stop 9 turn (17:16 to 17:54, 55 calls, context 53,734 to about 135,000 tokens, the router's real count for the last call; its exact count was 139,987) it climbed steadily: 0.0 to 1.8 percent until 17:24, 2.5 percent at 17:36, 3.5 at 17:42, 4.7 at 17:43, 6.9 at 17:44, 18.5 at 17:44, and 8 to 31 percent in every thought from then to 17:54. One thought was about 15,700 tokens and took about 5 minutes (the router's log: 15,697 tokens in 309.9 seconds, the call started at 17:37:15). The agent software sends its stored thinking back to the model on later calls by default (this is from the owner's assistant's check and was not re-checked for this page), so each degraded thought was part of what the model saw next; the model is a 2-bit class build (IQ2_S). A feedback loop in a heavily compressed model is therefore the likely cause, but it is not proven: the records show the rise, not the reason for it. The turn ended at 17:54:46 with the log reason "interrupted_by_user" (who interrupted is not recorded); the owner's assistant then wrote the decided brief (stop 10) with rules: short reasoning in plain sentences, stop at about 40 tool calls per turn, no picture-analysis loops for geometry, numbers instead. In the next three turns (stops 10 and 11), 8 of the 11 thoughts of 300 characters or more were between 0 and 2.2 percent; the other 3 were brief spikes of 12.1, 18.0 and 33.5 percent in confused moments of stop 10, so the 18:00 hour averaged 6.9 percent.

The agent's own review kept passing defects that were visible in the frames, so the decisive checks came from outside it. The owner watched the first full render and rejected it. His assistant, a Claude Sonnet 5.5 session, wrote the fix prompts, except stop 9, which is the owner's own typed message. Separate reviewer agents (not the build agent) then looked at full-resolution frames at fresh timestamps, without seeing the build agent's reports, at six checkpoints. They were Claude Sonnet 5.5 as well, the same model family as the assistant that wrote the prompts, so they are separate agents but not a different kind of model (the model is recorded for the first, the fifth and the sixth checkpoint and presumed the same for the others). At the first two they found problems the build agent had passed, such as an exhaust valve train drawn off its axis, and one item still open after round 3 (the callout leader dots). After round 4 three more reviewer agents checked full-resolution frames at new times and confirmed the leader markers, but found the dial needle still reading as a line through IN and EX and the last caption dim; these became stop 5. After round 6 three more checked 24 frames and 14 pairs of touching parts and found every pair separated, with the connecting rod over the crank disc as the weakest separation; that became stop 7. After round 8 three more separate reviewer agents (Claude Sonnet 5.5, not shown the build agent's reports) looked at frames at 32 new times. By measuring pixels against the specification's lift curve they found the lobe nose direction matching the specification and the roller drop and valve-head drop matching the lift curve to within about 1 px, the roller resting on the round base whenever a valve is shut, and the poppet heads reading as real poppet valves, sitting flush on the seat wedges when shut with a clean gap when open. Their only remarks were minor: the closed heads and wedges hang about 7 mm below the chamber roof because the specification's 8 mm head thickness puts them there (a real valve face sits flush; left as is), the stem's outline runs through the head, and the third reviewer also noted the crankcase and sump clipped at the bottom edge of the engine panel in the zoomed frames (about 23 to 26 seconds), probably the intended zoom. The first of those remarks was dealt with in round 9, which raised the closed heads 16 px so they sit flush. After round 9 three more separate reviewer agents (Claude Sonnet 5.5, not shown the build agent's reports) compared the new render with round 8 at 32 times. One of them found the tooth of casting gone at the intake port mouth at all ten times it was assigned (0.5, 2.4, 4.5, 6, 9, 11.5, 13.6, 16.5, 20 and 27.5 seconds) and at the exhaust port mouth at 6, 11.5, 13.6 and 27.5 seconds, the times its report names. In the chamber they found no gas notch at x 818 to 827, 880 to 922 or 972 to 990, the closed valve faces flush, and the electrode, spark and flame fine. The earlier fixes held: the dots on the heads after the 16 px raise, the cam timing, the poppet valves, the outlines and the tints. Their one reservation was that the head casting gradient was faint and other parts were still flat; that became stop 11. Minor remarks they left standing: the dark outline strokes on the closed head wedges read busy; the bottom outline of an open head takes a darker tint of the gas colour (as before); the spark glow is a hard-edged disc (as before); and the red-to-blue gas crossfade passes through purple for a few frames near t = 15.9 seconds (not changed). These checks passed because the reviewers measured positions, whereas earlier looks at the frames did not catch the lobe phase error. Ten fix rounds were applied after the owner's rejection: rounds 1 and 2 after stop 1, round 3 after stop 2, rounds 4 to 8 after stops 4 to 8, round 9 after stops 9 and 10 (stop 9 is the owner's own request, stop 10 the assistant's brief for it) and round 10 after stop 11.

The facts, in one table

Correction. Rounds 1 to 3 and the first attempt at round 4 ran with the model server set to 4 seats of 131,072 tokens each; stop 4 onward ran with 2 seats of 524,288 tokens each. The server's layout can change again, so a check of that server today may show neither. This page describes the layouts used while each stop was worked on, not the current one.

The counts come from the agent's own session records, the model router's log, the model servers' configuration and logs, and the render logs. Most of it was checked against a record; a few items come from the owner's notes and are marked as such.

The prompts, stop by stop

A stop is a point where the agent had finished, or had stopped, a piece of work and new instructions were sent. After stop 1 the agent worked its fix rounds 1 and 2, and after stop 2 round 3. Stop 3 asks for round 4, which was not applied; stop 4 repeats it with the design already decided, and it was applied. Stops 5 to 8 are rounds 5 to 8. Stop 9 is different: it is the owner's own typed request (port edges, chamber spaces, shading), after which the agent asked him questions and the assistant wrote the decided brief for round 9 (stop 10). Stop 11 is round 10, and round 10 is the video above. There are eleven stops in all, after the initial prompt (stop 0).

1. Stop 0

   Initial prompt

   One written brief and nothing else: build the 32-second video from scratch. It fixes the engine geometry and valve timing, the gas colours, every line of on-screen text, the technique (one hand-built SVG drawing driven by a single clock) and the exact commands to run, and it says the agent cannot ask questions. It also tells the agent to hand the reading of documentation digests and the visual review of its own screenshots to helper sub-agents (called Bolt in the prompt).

   Show the prompt (11,887 characters)

   Delivered 2026-11-08 18:43:44 (local time of the computer running the agent). Plain-text file

   TASK: Build final.mp4, a 32 s 1920x1080 30 fps silent video of a four-stroke petrol engine cross-section (cutaway) as a HyperFrames project: one hand-built inline SVG, GSAP used only as the clock. You cannot ask questions: take the simpler reading, log it in REVIEW.md.
   
   0. TOOLS
   Terminal calls time out at 180 s: run installs and renders in the background with a log and poll. Run 'npx hyperframes telemetry disable' once; never run publish, feedback, cloud render, auth login, events, usage, npx skills. Library (book hyperframes_en_2026-11; no slash commands): py -3.14 library.py search 'words' | read PATH [--part N] [--grep WORD]. Save your context with delegate_task (Bolt, 2 at once): Bolt 1 digests skills/hyperframes/SKILL, references/runtimes, references/format; Bolt 2 digests skills/hyperframes-cli/references/lint-validate-inspect, preview-render, references/types/explainer, registry/blocks/exploded-stack. Each digest: 400 words max (contract, commands, flags, gotchas), appended to NOTES.md. Read other pages only when an error names them. You write all code.
   
   1. SPEC
   npx hyperframes init four-stroke-2d --non-interactive --skip-skills. Root data-composition-id='engine' data-start='0' data-duration='32' data-width='1920' data-height='1080'; bg #0E1116; no audio. Write SPEC.md first (shot list; mark 'a chain turns the cam the same way as the crank' as standard mechanics, unsourced). theta = 45*t degrees (7.5 rpm): 4.0 s per 180-degree stroke, 32 s = two 720-degree cycles, EngineState pose(32) = pose(0), last frame 31.967 s.
   
   2. ENGINE (all numbers in assets/engine-state.js; invent none)
   th = theta mod 720, a = theta mod 360; th 0 = intake TDC, 360 = firing TDC. Strokes of 180 from th 0: INTAKE, COMPRESSION, POWER, EXHAUST.
   Crank r 40, rod l 136, bore 80 (mm), axis at origin. Pin height x = r*cos(a) + sqrt(l^2 - r^2*sin(a)^2): 176.00 TDC, 96.00 BDC, 129.98 at 90 (46.02 travelled, not half). Crank pin (r*sin a, r*cos a), clockwise on screen; wrist pin stays on the axis; pin-to-pin always 136. Crown 26 above pin; roof 212.
   Valves 22 either side, open downward, max lift 8, visible springs; INTAKE left, head 28; EXHAUST right, head 24; plug on axis, 10 wide.
   Timing (typical; engines vary) is given at 1.0 mm lift: intake 710 and 225, exhaust 495 and 10, spark 330. lift = 4*(1-cos(2*pi*u)), u = ((th-start) mod 720)/305, 0 if u>1; start 675 intake, 460 exhaust; peaks 8 at th 107.5, 612.5. So lift is 0 for intake th 260-675, exhaust 45-460; 1.0 (+-0.02) at the edges; 1.6 at th 0; 1.0-2.3 in the overlap. Open flags are windows: intake th>=710 or <=225; exhaust th>=495 or <=10; both only th>=710 or <=10 (20 degrees).
   Two overhead cams at th/2 (half speed, same sense), axes 22 either side, 298 up. ROLLER followers (radius 4) on the line through the cam axis. Pitch curve R(psi) = 10 + 4*(1+cos(2*pi*psi/152.5)) for |psi| <= 76.25 from the nose, else 10; roller centre 288 closed; drawn lobe = R minus 4. lift = R - 10 holds only for this follower. Nose angle clockwise from up = theta/2 + 126.25 (intake), theta/2 - 126.25 (exhaust): straight down at peak. Valve lift is read from the lobe at the contact; never animate a valve alone.
   Gas (state, never on metal): fresh #3FA7FF, compressed #FFD23F, flame #FF7A1A, burnt #E5383B. th 710-40: red to blue smoothstep; 40-225 blue; 225-330 mix(blue, yellow, k), k = (71.26-s)/64.42, s = travel below TDC; 330-390 chamber stays yellow plus a red burnt disc about the plug, radius b*40, b = smoothstep(330,390,th), height = live chamber (roof minus crown), clipped to the chamber, thin orange rim; 390-710 red. gasColor yellow until th 390, then red. spark = max(0, 1-|th-330|/12), white flash. Port dots: blue in at intake, red out at exhaust, only while that window is open; phase from a cumulative-lift table built at load; dot i at i/N; no randomness.
   Parts (cut at centre plane): piston with 3 rings, rod, crank with counterweight, flywheel R 70 behind, block (bore wall 40, jacket 44-54 in #4C6A73, outer 60), head, valves with springs and rollers, 2 cams, plug, ports bore 24, crankcase, sump to -100.
   
   3. ON-SCREEN COPY (exact; nothing else)
   Inter for words (title 700, stroke name 900, rest 400), JetBrains Mono 400 or 700 for numbers; bundled, no @font-face, no other weights. Text >= 30 px, contrast >= 4.5:1. HUD only in x 96-560 and x 1360-1824 (width 464, wrap OK), y 54-1026; engine only in x 580-1240.
   Left, top y: 'FOUR-STROKE ENGINE' 44 px + 'Single-cylinder petrol cutaway. Slow motion, 7.5 rpm.' (54); 'CRANK ANGLE' + floor(th) as '345°', mono 72 px (215); 'CAM ANGLE' + floor(th/2), mono 48 px (345); 'CYCLE 1 OF 2' / 'CYCLE 2 OF 2' (switch 16.0 s) and 'STROKE 1 OF 4' to '4 OF 4' (445); stroke name 'INTAKE' 'COMPRESSION' 'POWER' 'EXHAUST' 56 px (540); caption 32 px (625); footnote 'Valve and spark angles are typical values. Real engines vary. Geometry, flow and burn are simplified.' (870).
   Right: dial centre (1592,260) radius 150, every position and the pointer at th/2 degrees clockwise from the top: blue arc intake open (710 to 225), red arc exhaust open (495 to 730), bright overlap mark (710-730), tick 'SPARK' at 330, ticks '0°' '180°' '360°' '540°', labels 'IN' 'EX', caption 'ONE LAP = 720° CRANK = 360° CAM' (465); timing inset (550): 18-mark crank sprocket beside a 36-mark cam sprocket twice its size, chain line, marks at theta and theta/2, caption 'CRANK 2 : CAM 1'; callout pill (720-810); legend swatches 'Fresh charge' 'Compressed charge' 'Flame front' 'Burnt gas' 'Dial arcs: valve open' (825).
   Captions, one per 4 s from 0: 1 'Piston moves down. The intake valve is open and fresh air-fuel mix flows in.' 2 'The intake valve closes just after the bottom. The rising piston squeezes the charge.' 3 'Hot gas expands and pushes the piston down. Only this stroke delivers power.' 4 'The exhaust valve is open. The rising piston pushes burnt gas out.' 5 'The camshaft turns at half crank speed: 360° of cam for 720° of crank.' 6 'Each valve is pushed open by its cam lobe and closed by its spring.' 7 'One power stroke every two crank turns. The flywheel carries the other three.' 8 'Both valves are open together for only about 20°, at the end of exhaust and the start of intake.' Stroke name + caption fade out over the 0.25 s ending on each boundary and in over the 0.25 s starting on it; frame 0 and the last frame fully shown.
   Callouts (30 px, opaque #0E1116 pill, leader line to the part; fades lie INSIDE the window, 0.25 s each end): 4.4-6.6 and 20.4-22.6 'INTAKE CLOSES 45° AFTER BOTTOM' (intake valve); 6.8-9.0 and 22.8-25.0 'SPARK 30° BEFORE TOP' (plug); 10.4-12.6 and 26.4-28.6 'EXHAUST OPENS 45° BEFORE BOTTOM' (exhaust valve); 15.0-17.2 'OVERLAP: BOTH OPEN, ONLY 20°' (between valves); same text 30.0-32.0, fade-in only, held to the last frame.
   
   4. TECHNIQUE
   assets/engine-state.js: globalThis.EngineState in an IIFE; EngineState.at(t) is pure (no clock, random, earlier frames) and returns theta, th, stroke, strokeName, cycle, pinY, s, crankPin, rodTilt, intakeLift, exhaustLift, intakeOpen, exhaustOpen, camAngle, camNoseIntake, camNoseExhaust, spark, burn, flameR, chamberH, gasColor, zoom; every drawn thing and HUD number reads it.
   One SVG viewBox 0 0 1920 1080, 2 px/mm, crank axis (900,800): sx = 900+2x, sy = 800-2y. <g clip-path='url(#win)'> (static rect x 580-1240, y 70-1010) around <g id='engine' transform='translate(900 390) scale(z) translate(-900 -390)'>; 2 px #3A4350 border outside both. Cut faces: diagonal hatch (#C9D1D9 on #1B2027); steel #8E99A6, graphite #2B3038; gradients only, no filters. HUD = one class='clip' wrapper (data-start='0' data-duration='32'); each caption group a clip with its own data-start and data-duration, opacity on a child.
   GSAP: save https://cdn.jsdelivr.net/npm/gsap@3.14.2/dist/gsap.min.js once into assets/vendor/gsap.min.js and reference it locally (no CDN tag; no network when rendering). const clock = {t:0}; ONE paused tl; tl.to(clock, {t:32, duration:32, ease:'none', onUpdate: () => render(clock.t)}, 0); render(0) at load; compute EVERY opacity inside render(t), no other tweens; await document.fonts.ready; window.__timelines['engine'] = tl as the LAST statement; never tl.play(). If snapshots at 0, 2, 4 s show one piston height, onUpdate is not firing on seek: use hf-seek (root data-no-timeline; addEventListener('hf-seek', e => render(e.detail.time)); section 'The seek protocol' in runtimes) and drop GSAP. Frames seek out of order: carry nothing over.
   Spark flash and flame are the brightest things; all else calm. Engine-group zoom about (900,390), smoothstep keys at t mod 16: 6.4 s 1.0, 8.0 s 1.2, 9.4 s 1.2, 11.8 s 1.0.
   Never: audio, media files, network requests, Date.now, performance.now, Math.random, setTimeout, setInterval, own requestAnimationFrame, repeat:-1, CSS perspective, glow, bounce easing, faces, stock footage, compression ratio, horsepower, a brand, any rpm but 7.5; both valves open outside the overlap; sinusoidal piston; valve events exactly at TDC or BDC; a gas colour changing meaning.
   
   5. COMMANDS (only these; invent no flags)
   1) npx hyperframes doctor ; npx hyperframes browser ensure (offline: set HYPERFRAMES_BROWSER_PATH to Chrome). Log both in REVIEW.md.
   2) Write SPEC.md, the module and assets/engine-state.check.cjs (require('vm').runInThisContext(require('fs').readFileSync('assets/engine-state.js','utf8'))). Assert every number above, and: stroke 0,1,1,2,2,3,3,0 at t 3.999,4,7.999,8,11.999,12,15.999,16; lift 0 inside the zero windows and 1.0 (+-0.02) at th 710, 225 (intake), 495, 10 (exhaust); pin-to-pin 136 (1e-6) for t 0-32 step 0.01; lift peak 8.00 at t=2.3889 (intake) and 13.6111 (exhaust); both flags true exactly for th>=710 or <=10; lobe radius at the contact minus 10 = lift (+-0.01) at th 10, 107.5, 180 (intake), 505, 612.5, 700 (exhaust); spark 1 at t=22/3 and 70/3, 0 beyond +-0.267 s; gasColor #3FA7FF at t=3, #E5383B at t=11.5; t=32 equals t=0 (1e-9). node assets/engine-state.check.cjs must print PASS before you draw.
   3) Build index.html; npx hyperframes lint after each big edit.
   4) npx hyperframes check --json: zero errors, browserSkipped not true, zero console errors; film_silent is expected (accept in REVIEW.md); fix or justify other warnings.
   5) npx hyperframes snapshot --at 0,2,4,5,6.5,7.3333,7.4,8,11,13.61,15.9,16,31.9 (PNGs: .hyperframes/artifacts/snapshots/<id>/frame-NN-at-Xs.png). Send each PNG with its Expect line to a Bolt subagent (vision); record its answer in REVIEW.md; if nobody can view it write NOT VIEWED, never passed. Expect: t=0 piston TDC, both valves about 1.6 up (inside the overlap), 000°; t=2 pin 129.98, intake about 7.7, chamber blue; t=4 BDC 96.00, intake about 4.3; t=5 intake 1.0 (closing), charge pure blue, intake callout; t=6.5 blue-yellow (k about 0.64); t=7.3333 spark flash, piston 6.8 below TDC, valves shut, zoom 1.12; t=7.4 flash fading; t=8 TDC, flame half grown INSIDE the chamber, zoom 1.2; t=11 exhaust 1.0, piston about 71 below TDC, gas red; t=13.61 exhaust 8; t=15.9 overlap callout, 715°; t=16 pose of t=0, CYCLE 2 OF 2, caption 5; t=31.9 overlap callout. Also: wrist pin on axis, rod joins both pins, roller touches the nose at peak lift, no gas colour on metal, no clipped or overlapping text.
   6) snapshot --at 4,0,4 --no-end: the two 4 s PNGs (frame-00-at-4s.png and frame-02-at-4s.png in the printed folder) must be identical (cmp or sha256sum).
   7) snapshot --review; follow references/review; write REVIEW.md; at most 3 fix rounds.
   8) render --quality draft --output review.mp4 --workers 1; fix; render --quality delivery --output final.mp4 --workers 1.
   9) ffprobe -v error -show_entries stream=codec_type,width,height,r_frame_rate,duration -of default=nw=1 final.mp4: one video stream, 1920x1080, 30/1, 32.0 (+-0.05), no audio.
   REPORT: files (absolute paths), each command with exit code, decisions, what you could not verify; never claim a check you did not run.

2. Stop 1

   The owner rejects the first render

   He watched the first finished video and found it not acceptable: the piston was not attached to the crank (drawn far above the cylinder), the valve springs were floating or hidden, and a dial label collided with a caption. The prompt also says to look for everything else that is wrong, and lifts the limit of three fix rounds to six.

   Show the prompt (5,032 characters)

   Delivered 2026-11-09 06:12:08 (local time of the computer running the agent). Plain-text file

   NEW TASK from Kavan: he watched final.mp4 and it is NOT acceptable yet. His 3-round fix cap from before is lifted: you now have up to 6 fix rounds. Fix everything below, re-verify by looking at real frames, re-render, and report. Do not argue the old verdicts: your earlier review passed these defects, so trust the evidence below and your own crops, not the contact-sheet scans.
   
   WHAT HE SAW (and I confirmed on frames at 0.5, 2, 4, 5.5, 7.3, 9, 11, 13, 15.5, 18, 20, 22.5, 25, 27, 29, 31 s of final.mp4):
   
   1. THE PISTON IS NOT ATTACHED TO THE CRANK. The piston body (the grey block with 3 ring notches at its sides, about 150 px wide) is drawn FAR ABOVE the cylinder, floating over the cylinder head near the top of the engine area, and it moves the wrong way / by the wrong offset (tall stack at t=2, flat at t=0.5). Meanwhile the rod's small end and wristPin circle sit down in the bore with NOTHING attached: there is a black gap between the bottom of the blue/red gas block and the wrist pin. Fix: pistonG must follow the wrist pin exactly. Per SPEC.md the crown is flat at pinY+26 (the gas block's bottom edge), half-width 38, 3 rings at crown offsets -6/-13/-20, skirt down to pinY-34, wrist pin through the piston, rod joining wrist pin and crank pin in EVERY frame. I read the code and the likely root cause is this: the pistonG rects in index.html (around lines 74-84) are authored with the crown top at y=0 and the group 120 px tall (60 mm: crown to skirt), but the per-frame update (around line 511) sets transform translate(0 2*st.s), which is only the travel below TDC, so the piston is drawn near the very top of the viewBox. It should be translate(0 crownY) with crownY = 800 - 2*(st.pinY + 26), the same value you already compute two lines below for the gas block. Verify that, do not trust me blindly: also check the rest of the group (ring notches, skirt) lines up with the wrist pin once it is moved. Check numerically at t = 0, 2, 4, 5.5, 7.33, 9, 11, 13, 16, 20, 25, 31.7 s: piston crown y == gas block bottom y == wrist pin y minus 26 mm, in the same units, from the DOM (getBoundingClientRect or the SVG attributes), then confirm by eye on full-resolution crops. The solid grey piston must be visible between the gas block and the rod in every frame.
   
   2. THE VALVE SPRINGS ARE FLOATING OR HIDDEN. The right (exhaust) spring zigzag is drawn OUTSIDE the head casting to the right, floating in the air beside the head. The left (intake) spring is partly hidden behind the head block (z-order). Per SPEC: spring wells at x = valveX +/- 14, y 246-280 INSIDE the head extents (x +/-66, y 212-280), seat washer at y 248 (fixed, stem hole +/-4), retainer on the stem at y = 284 - lift, coil zigzag 5 turns between washer and retainer. Both springs must sit inside their wells, be drawn IN FRONT of the head casting, touch the washer at the top and the retainer at the bottom, and compress/extend with the valve. Also check each valve stem runs continuously from the valve head up through the washer to the roller/cam follower with no gaps, and that the cam towers/rocker bits do not cover the springs.
   
   3. The dial text collision: the 'SPARK' label next to '360' overlaps the caption 'ONE LAP = 720 deg CRANK = 360 deg CAM' (right side, around y 440-470 at 1920x1080). Separate them (move one, do not remove either).
   
   4. LOOK FOR EVERYTHING ELSE WRONG. Do a proper pass on the engine area at the 16 times above, using FULL-RESOLUTION CROPS of the head, bore, rod and crank (not 360 px contact sheets), one picture call at a time (vision_analyze runs on Bolt now, a few seconds per frame, but it is only trustworthy for text and colour: for geometry use the DOM numbers). Things I want checked: every part that should touch actually touches (plug into the roof, valve heads on their seats when closed, rod to both pins, crank pin to crank web), no part floating, no part hidden behind another, leader lines that cross the whole engine, gas colour only inside the chamber above the crown (never on metal), no clipped or overlapping text anywhere.
   
   HOW TO WORK:
   - Before touching anything, copy the current deliverable: final.mp4 -> final_v1.mp4 and review.mp4 -> review_v1.mp4, and copy index.html -> index_v1.html.
   - Fix in index.html, keep the spark flash, gas colours, timing and the callouts working as they are now (they are correct). Re-run lint and check, snapshot at the 16 times, then render review.mp4 (draft) and final.mp4 (delivery) and re-run the ffprobe gate (h264, 1 video stream, 1920x1080, 30 fps, 32.0 s +/-0.05, no audio).
   - Append a new section to REVIEW.md and REPORT.md: what was wrong, the numbers before and after (piston crown y vs wrist pin y per time), what you fixed per round, anything you found beyond the three listed, anything still imperfect.
   - You have up to 150 tool calls in this turn. Work in short steps, write progress notes to NOTES.md as you go (round number, what changed), and end with a final report in chat that names the files and lists each defect as fixed or not fixed.

3. Stop 2

   Three separate reviewer agents (not the build agent)

   Three reviewer agents looked at the new video at full resolution at 21 different times, without seeing the agent's own reports. All three flagged the same blocking defect: the whole exhaust valve train was drawn about 60 pixels too far to the right (30 mm in the drawing's own scale: the stem sat at +52 mm where the drawing's specification puts it at +22 mm), so the exhaust valve sat outside the combustion chamber. They also listed dial ticks and labels, callout leader lines, a clipped spark flash, a malformed exhaust port and smaller items. The agent's own report had said that nothing known was still imperfect. The prompt also asks the agent to make the animation viewable as a plain HTML page. The live player linked from this page was made separately; see the note by the player link.

   Show the prompt (6,769 characters)

   Delivered 2026-11-09 07:41:24 (local time of the computer running the agent). Plain-text file

   NEW TASK from Kavan, in two parts: PART 1 = fix round 3 (an independent check of your new final.mp4 found real defects), PART 2 = make the animation viewable as a plain HTML page. Do PART 1 first, then PART 2, then one final report. Your fix budget is still open (up to 6 rounds; this is round 3). You have up to 150 tool calls.
   
   WHY ROUND 3: three independent reviewers looked at final.mp4 at full resolution at 21 different times, without seeing your reports. All three agreed on the same blocking defect and several smaller ones. Their images are in folders A, B and C of the review images (files head_T.png, low_T.png, full_T.png): look at head_0.5.png and head_4.png first. The piston fix, the spring fix and the SPARK/ONE LAP separation are confirmed good. Do not undo them.
   
   PART 1 - DEFECTS, most important first
   
   1. BLOCKING: THE WHOLE EXHAUST VALVE TRAIN IS ~60 px (30 mm) TOO FAR RIGHT. SPEC says valve axes are at x = -22 mm and +22 mm about the bore axis, i.e. screen x = 856 (intake, correct) and 944 (exhaust). Your exhaust stem, spring, cam and roller are at x = 1004 (+52 mm). In round 2 you moved the cam train from 1044 to 1004 to meet the stem, which locked in the misplacement: the real error was the exhaust STEM position (it was at 1004 in v1 as well). Consequences you can see: the exhaust valve head (x 980-1028) sits OUTSIDE the chamber (chamber/gas spans x 820-980), over the right bore wall and water jacket, so it can never open into the cylinder; the exhaust callout leader and the exhaust port dots end at x 944-948 where there is no valve; the red port dots are drawn across the stem and spring. Fix: move the ENTIRE exhaust column to axis x = 944: valve head (24 mm wide), stem, barrel, retainer, washer, spring, roller, cam (cam axis x = valve axis per SPEC), the cam housing/tower block, and anything else keyed to 1004 or 1044. Mirror the intake side (856 = 900 - 44) exactly. The exhaust head must then sit inside the chamber x-range, over the gas, and the port dots must end at the seat where the valve is. Check the two cams do not collide (88 px apart now; max lobe radius is 14 mm = 28 px each).
   
   2. Dial: a stray slanted tick sits glued to the front of the label so it reads '\SPARK' (it is about 20-25 px outside the ring, not on it; the 330 deg spark tick should sit on the ring near x 1629, y 405). Put the spark tick on the ring at the right angle or remove the stray stroke. The short 90-degree ticks run through the '180 deg' and '540 deg' labels (a tick strikes through the '1'); move the labels outward. The red 'EX' label starts on the red arc; nudge it clear. The needle tip dot lands on 'IN' / 'EX' at some times (cosmetic: nudge labels if easy).
   
   3. Callout leader lines: each anchor must land on the part it labels (intake valve head at (-22, 218) mm, plug at (0, 230) mm, exhaust valve head at (22, 218) mm, between valves at (0, 218) mm per SPEC). Today the 'OVERLAP' leader anchors at the plug tip and the exhaust leader at an empty spot. The leader lines also run diagonally across the whole engine (head hatch, gas, piston crown, right water jacket) like a scratch: route them so they do not cross the gas or the piston (for example draw them behind the engine group, or run them above the head and then out to the right).
   
   4. The white SPARK flash at t=7.33 is clipped only to the bore width, so it paints over the piston crown, the top ring land and half of the wrist pin. Clip it to the chamber rect (crown to roof) so it never covers the piston.
   
   5. Exhaust port looks broken: the intake port is a smooth curved band but the exhaust port is a jagged black wedge with a notch cut through the right edge of the head and no opening into the chamber; they should be mirror images (SPEC bezier). The left under-deck block is hatched and the right is plain grey: make them mirror-symmetric or deliberate.
   
   6. Smaller things, your call if they are quick: the connecting rod and the crank web are the same grey (the rod merges into the web where they overlap: add a thin darker outline or slightly different shade to the rod); two grey rectangles above the head are offset ~45 px from the cam centres (check what they are; if they are stray, fix or remove); the HUD readouts truncate instead of round (329.85 shows 329): check the SPEC rule and follow it.
   
   7. Remove the probe hook 'window.__hfRender = render;' from index.html before the final render (the measuring tool does not need it).
   
   HOW: copy final.mp4 -> final_v2.mp4 and review.mp4 -> review_v2.mp4 and index.html -> workspace/v2/index_v2.html first. Use the measuring tool, it now has two new checks that FAIL today for defect 1:
     node measure_engine.mjs            (add --times and --shots DIR as before)
   New checks: valve axes mirrored about the bore axis, and both valve heads inside the chamber x-range. All checks must pass, at 12 or more times, before you render. Then look at full-resolution crops of the head at 0.5, 2, 5.5, 11, 13 and 25 s by eye (vision_analyze is for text and colour only). Then render review.mp4 and final.mp4, run the ffprobe gate (h264, 1 video stream, 1920x1080, 30/1, 32.0 s +/-0.05, no audio), append the round-3 numbers to REVIEW.md, REPORT.md and NOTES.md.
   
   PART 2 - VIEWABLE AS HTML (Kavan asked: 'is it viewable as html?', and wants YOU to build it). Today index.html cannot be opened by double click: its GSAP timeline is paused and only HyperFrames drives it, so a browser shows one still frame at t=0. Build player.html in four-stroke-2d (do NOT change index.html): ONE self-contained file made from the FINISHED index.html (copy after part 1 so the fixes are included): inline assets/vendor/gsap.min.js and assets/engine-state.js, same SVG/HUD markup and engine code, optional Google Fonts links for Inter and JetBrains Mono with the same fallbacks. Behaviour: the 1920x1080 stage scales to the window width keeping 16:9; a Play/Pause button; a scrub slider 0-32 s with a time readout; loops; drive it with requestAnimationFrame and tl.seek(t, false) (the false matters: the default suppresses the onUpdate that redraws). Autoplay unless prefers-reduced-motion is set (Kavan's Chrome reports reduce-motion, so make the Play button obvious). Optional deep link: open paused at a time from a hash like #t12.5. Test it in headless Chrome at three times (use the hash) and compare the screenshots against the same frames of final.mp4: they must match. Report the path and size, and add a line about player.html to REPORT.md.
   
   FINAL REPORT: per defect fixed or not fixed with evidence (measure tool output before/after), the ffprobe gate, the player.html test, files written. Be honest about anything still imperfect; your earlier report said 'nothing remains known-imperfect' and the independent check disagreed.

4. Stop 3

   A second set of reviewer agents

   Three fresh reviewer agents checked the round 3 video at 21 new times. They confirmed the earlier fixes and agreed on one item still not fixed: each callout leader line must end in a visible dot on the part it labels. Two cosmetic items were added (the dial needle drawn across the IN and EX labels, and whether the on-screen numbers truncate or round). This prompt asked for round 4. It was not applied: the agent made no edit in about 45 minutes of work (two single thoughts of about 32,000 tokens each, about 11 minutes apiece, and two context compactions of about 4 minutes each after the prompt arrived), and the turn ended on a router refusal at 10:33. Stop 4 sent it again.

   Show the prompt (3,923 characters)

   Delivered 2026-11-09 09:48:15 (local time of the computer running the agent). Plain-text file

   NEW TASK from Kavan: FIX ROUND 4 (small). Do this AFTER you finish the current job (player.html and its report); if you already finished it, start now. Three independent fresh reviewers checked your round-3 final.mp4 at 21 new times: they confirmed your round-3 fixes (exhaust valve train mirrored at x=944 and over the chamber, both ports mirror images, dial tick and labels, spark flash inside the chamber, under-deck steps symmetric, rod outline). Well done. ONE round-3 item is NOT fixed and all three reviewers agree:
   
   1. CALLOUT LEADER LINES (visible). Each callout leader must end with a VISIBLE dot ON the part it labels. Today (a) the intake-close leader (callout A/A2) and the exhaust-opens leader (C/C2) end on the valve STEM, about 35-40 px above the valve head, and you cannot see the dot because it is #8E99A6, the same grey as the steel stem/plug it sits on; the line just seems to vanish into the stem; (b) the spark leader (B/B2) dot sits on the plug body 34-48 px above the plug tip; (c) the between-valves leader (D1/D2) dot is 16 px above the roof; (d) the lines take odd long routes: the intake leader goes from the pill up to y=125, across to x=790 and down over the casting and port; the exhaust and overlap leaders make a V-bend at about (1000,308) that scratches back across the red exhaust particle trail and in front of the exhaust stem and spring.
   Fix: give every leader dot a clearly visible look against steel (for example a bright fill such as #FFD23F or white with a 2 px dark outline #0E1116, radius 6-7), draw the dots above all engine parts, and anchor them: A/A2 and C/C2 at the CENTRE OF THE VALVE HEAD (the anchor must follow the head as it opens downward, not stay at a fixed mm point above it), B/B2 at the plug tip (or the lowest visible part of the plug), D1/D2 at the roof between the valves on the bore axis. Route the lines so they do not run over the red/blue port particle trails, the springs or the gas; a simple elbow is fine. Keep the spec anchor points in mind (SPEC.md line 112) but visibility and landing on the labelled part win; note any deviation in REPORT.md.
   
   MEASURING TOOL (it was extended): node measure_engine.mjs now has THREE leader checks (dot on top, dot contrasts with the steel, dot lands on its part: valve heads <= 8 px, plug <= 40 px from the spark centre, between valves <= 14 px). It only looks at leaders whose callout is showing, so run it with times inside the callout windows: --times 5.5,7.33,8,11.5,16,21.5,24,27.5,31 (windows 4.4-6.6, 6.8-9, 10.4-12.6, 15-17.2, 20.4-22.6, 22.8-25, 26.4-28.6, 30-32). Right now the contrast and landing checks FAIL on your current index.html; all other checks pass and must keep passing. Add --shots DIR and look at head crops at 5.5, 7.33, 11.5, 16 by eye (vision_analyze is for text and colour only).
   
   2. Cosmetic, quick: the dial needle SHAFT runs through the letters of 'IN' (near t=1.9) and 'EX' (near t=11 and t=26). Draw the IN/EX labels above the needle (z-order) or nudge them so the shaft does not strike through them.
   
   3. The HUD truncates instead of rounding (329.85 shows 329, 22.5 shows 22). Check SPEC.md for the rule; if the spec says round, fix it; if it says floor or is silent, leave it and say which.
   
   THEN: re-run measure_engine.mjs at the 12+ times (ALL checks must pass, including the new leader ones), back up final.mp4 to final_v3.mp4 and review.mp4 to review_v3.mp4 BEFORE rendering, render review.mp4 and final.mp4, ffprobe gate (h264, 1 video stream, 1920x1080, 30/1, 32.0 s +/-0.05, no audio), append round-4 numbers to REVIEW.md/REPORT.md/NOTES.md, and REBUILD player.html from the finished index.html (keep the build in a small script so it can be regenerated) and repeat the 3-time screenshot comparison against the new final.mp4. Final report: per item fixed or not fixed with the tool output, the gate, the player test, files written; be honest about anything still imperfect.

5. Stop 4

   Round 4 again, with the design decided

   The assistant re-sent round 4 after the stalled attempt, now with the design already decided so that the agent only had to implement it, and told it to keep its reasoning short and work in small steps. The leader lines and dots had been the same grey as the steel, and their anchors were fixed points above the valve heads. The prompt specifies white lines with a dark halo, a route from the label up to a bus line above the engine and straight down the labelled part's own axis, dots on the plug tip, the valve heads and between the valves, drawn after every engine part. It also asks for the dial labels to be drawn above the needle and for the HUD rounding to be checked against the specification, and notes that the model server's seats and the agent's context window had been changed at 12:15 (the prompt says 12:15; the seat change itself was about 11:27, and only the agent's own context window and its compaction threshold changed at about 12:15). The agent applied it: the renders were written at 12:46 and its report came at 13:02. Three reviewer agents then confirmed the leader markers.

   Show the prompt (4,860 characters)

   Delivered 2026-11-09 12:27:24 (local time of the computer running the agent). Plain-text file

   continue. FIX ROUND 4, with the design already decided so you only have to IMPLEMENT it. Your last attempt at this ran 45 minutes and made no edit: two single thoughts of about 32,000 tokens each (11 minutes apiece) and three 4-minute chat compactions. Do not re-derive the design. If something is ambiguous, pick the simplest option, write one line about it in NOTES.md and move on. Keep your reasoning short. Work in small steps: one edit, run the measuring tool, look at two crops, next edit.
   
   CONTEXT: three independent reviewers checked your round-3 final.mp4. Everything from round 3 is confirmed except ONE item: the callout leader lines. In index.html (lines ~159-193) each leader group has fixed polyline segments from the pill at (1360,765) to a bend and a last segment whose end follows an anchor; the dots are circles #leaderDotA..D2 filled #8E99A6, the same grey as the steel, so you cannot see them, and the anchors are fixed spec points ABOVE the valve heads (at (856,364) and (944,364)) so they land on the stem. The bend at (1000,310) for B/C/D also sits in the exhaust port and scratches across the red particle trail.
   
   THE DESIGN (implement exactly this):
   1. Colour: every leader line and dot is NOT steel grey. Lines: white #F2F5F8, stroke-width 2, drawn over a dark under-stroke #0E1116 stroke-width 5 (a halo, so it reads on steel and on dark). Dots: fill #F2F5F8, stroke #0E1116 stroke-width 2, r 6. Do not use any gas colour (blue #3FA7FF, yellow #FFD23F, orange #FF7A1A, red #E5383B), they mean something in the legend.
   2. Route (all callouts): pill corner (1360,765) -> straight line up to a point on a horizontal 'bus' above the engine -> along the bus -> straight DOWN the labelled part's own axis to the dot. The final vertical segment must be exactly vertical every frame: set its x1 AND x2 to the zoomed anchor x (tx), so the camera zoom never skews it. Bus heights (screen y, before zoom; they are drawn in screen space and may stay unzoomed): A/A2 (intake) y=125; C/C2 (exhaust) y=140; B/B2 (plug) y=110; D1/D2 (between valves) y=95. Bus start x for the diagonal from the pill: A 1032, C 1040, B 1048, D 1056 (pick other values if lines would cross each other, but never route through the engine window x 580-1240, y 70-1010 except for the vertical drop).
   3. Anchors (the DOT position, per frame): A/A2 = centre of #valveHeadIntake (x = its centre x, y = its centre y, so it follows the head as it opens downward); C/C2 = centre of #valveHeadExhaust (same); B/B2 = the plug tip: x=900 (zoomed), y = roof y minus 8 px, i.e. the lowest visible part of the plug; D1/D2 = on the bore axis at the roof, y = roof y minus 4 px. Compute them with the same zoom transform you already use (zsx/zsy) from the same numbers that place the valve head, so they cannot drift.
   4. Z-order: the whole leaderLinesTop group (lines and dots) goes AFTER every engine part, the cams, springs, plug and the head casting in the SVG so nothing covers a dot. Keep the fade of each group with its callout.
   5. Delete the old fixed anchor table values that no longer apply; keep the second-cycle duplicates (A2..D2) in sync.
   
   CHECK (the measuring tool has three leader checks; they all FAIL today):
     node measure_engine.mjs --times 5.5,7.33,8,11.5,16,21.5,24,27.5,31 --shots workspace/meas4
   They must all pass together with every other check. Then look at head crops at 5.5, 11.5, 16, 24 (vision_analyze, one call at a time: it is for text and colour only): the white line must run down its own valve or plug and end in a clearly visible dot on that part.
   
   OTHER SMALL ITEMS:
    a. Dial needle: the shaft runs through the letters of 'IN' (near t=1.9) and 'EX' (near t=11 and t=26). Draw the IN and EX labels above the needle in z-order.
    b. HUD rounding: the HUD truncates (329.85 shows 329). Look at SPEC.md for the rule; if it says round, round; if it says floor or nothing, leave it and say which in REPORT.md.
   
   THEN: back up final.mp4 to final_v3.mp4 and review.mp4 to review_v3.mp4 BEFORE rendering; render review.mp4 and final.mp4; ffprobe gate (h264, 1 video stream, 1920x1080, 30/1, 32.0 s +/-0.05, no audio); append round 4 to REVIEW.md, REPORT.md, NOTES.md. player.html is an iframe of index.html, so it needs no rebuild; re-run the 3-time player screenshot comparison once. Final report: per item fixed or not fixed with the measuring-tool output, the gate, files written; be honest about anything still imperfect.
   
   NOTE ON YOUR SETUP (changed at 12:15): your context window is now 491,520 tokens (compaction at about 295K by your count instead of 92K), so you will be compacted far less often, and Maya now runs 2 seats instead of 4: you are the only user, so nothing queues. Long prompts are fine; reading a 100K prompt cold takes about 23 s. Do not spend time on any of that: it only explains why your calls will be quicker and your chat longer.

6. Stop 5

   Two small items from a third set of reviewers

   Three fresh reviewer agents had checked the round 4 video frame by frame, confirmed the leaders fixed and found no regressions, and left two items: the dial needle still read as a line through the IN and EX letters, and the last caption was dim on the final frame, although the specification says the last frame is fully shown. The prompt gives the exact fix for each (a dark halo on the two letters; holding the last caption instead of fading it out), says the HUD rounding is not a defect, and skips the player screenshot test. The agent applied both, and rendered at 13:04.

   Show the prompt (3,403 characters)

   Delivered 2026-11-09 13:02:18 (local time of the computer running the agent). Plain-text file

   continue. FIX ROUND 5, two small items, design already decided, just implement. Skip the player screenshot test (player.html is an iframe of index.html, I checked it myself); finish whatever you are doing in one step and start. Keep your reasoning short: one edit, run the measuring tool, look at crops, next edit.
   
   THREE fresh independent reviewers checked your round-4 final.mp4 frame by frame. Good news first: the callout leaders are CONFIRMED FIXED by all three (white line with dark halo, exactly vertical drop, clearly visible dot on the intake head, plug tip, exhaust head and between the valves, dot follows the heads as they open). Piston, springs, mirrored valve trains, ports, flash, dial ticks, SPARK / ONE LAP spacing: all hold, no regressions. The HUD truncates: SPEC.md line 80 says floor, so that stays and is not a defect (say so once in REPORT.md).
   
   TWO things remain:
   
   1. DIAL NEEDLE STILL READS AS A LINE THROUGH THE LABELS. In index.html (lines ~306-307) the IN text (x 1649, y 212) and EX text (x 1527, y 298) are already drawn after the needle group, but the needle still shows through the gaps between the letters and across them: at t=1.96 and 17.96 (IN), t=10.98 and 27.0 (EX). Fix: give both text elements a dark halo so the needle disappears behind the lettering. On each of the two text elements add these attributes: paint-order="stroke" stroke="#0E1116" stroke-width="10" stroke-linejoin="round". Keep their fill colours and keep them AFTER the pointerG group in the SVG. The page background is #0E1116, so a gap in the needle around the letters is correct and expected. Do not change anything else on the dial.
   
   2. LAST-FRAME CAPTION IS DIM. SPEC.md line 124-125: the last frame is fully shown (opacity 1 at t=32). Today caption 8 (cap8, 'Both valves are open together...') fades OUT in the last 0.25 s, so the held final frame (t=31.967) shows it at about 10 percent. Fix in the fade code (index.html lines ~510-511 and ~631): make the cap8 entry held, the same way callD2 is: fadeCaps entry for k=8 gets held: true, and the loop passes it: for (const f of fadeCaps) fade(f.el, f.win, f.held). That keeps the fade-IN at t=28 to 28.25 and removes the fade-out. Do not touch caps 1-7.
   
   CHECK:
     node measure_engine.mjs --times 5.5,7.33,8,11.5,16,21.5,24,27.5,31 --shots workspace/meas5
   All checks must still pass. Then BEFORE rendering back up final.mp4 to final_v4.mp4 and review.mp4 to review_v4.mp4 (do not overwrite the older backups). Render review.mp4 and final.mp4. Then extract and LOOK at these frames from the new final.mp4 (ffmpeg -ss T, crop the dial 330x260 at 1420,130 and scale 4x with flags=neighbor): t=1.96, 10.98, 17.96, 27.0 for the dial: the needle must not be visible through the IN / EX letters (a break in the needle around the letters is correct). And t=31.95 and t=28.1 for the left column: the caption text at 31.95 must be as bright as the one at t=28.5 or t=29, and at 28.1 it must still be fading in.
   
   THEN: ffprobe gate (h264, 1 video stream, 1920x1080, 30/1, 32.0 s +/-0.05, no audio); append round 5 to REVIEW.md, REPORT.md, NOTES.md. Final report: per item fixed or not fixed with the evidence, the gate, files written. One more thing to mention honestly in REPORT.md as accepted and not fixed: at about t=11.5-12.5 and 27.5-28.4 the exhaust leader's vertical touches the first red particle of the exhaust trail by about 3 px (cosmetic).

7. Stop 6

   Outlines and slight tints

   The owner had circled the engine head on a screenshot: the cam lobes, cam towers, bearing rings, lifters, valves and head casting were all the same steel grey and merged where they touch, the springs were dark on dark, and the spark plug was dark on near-black. In his words, lifters, valves and springs should be outlined, light things get a dark outline and dark things a light one, or the cams and lifters could get slightly different colours. The prompt asks for both: outlines by fill colour, a light edge under each spring, a dark edge on the bearing rings, and muted tans for the cam lobes and lifters and a lighter grey for the valves, with no blue, yellow, orange or red on metal because those colours mean gas in the legend. The agent applied it and rendered at 13:22.

   Show the prompt (4,779 characters)

   Delivered 2026-11-09 13:09:41 (local time of the computer running the agent). Plain-text file

   continue. FIX ROUND 6 (start it right after round 5 is rendered and reported; if round 5 is still running, finish it first). Design already decided, just implement. Keep your reasoning short: one edit, run the measuring tool, look at crops, next edit.
   
   THE PROBLEM (Kavan circled it on a screenshot of the engine head): the cam lobes, cam towers, bearing rings, rollers (lifters), valve barrels / stems / heads and the head casting are all the SAME steel grey, so where they touch they merge into one blob. The springs are dark on a dark band and dark on grey, the plug is dark on a near-black background. His words: lifters, valves and springs should be outlined; maybe everything should be outlined; dark things get a light outline, light things get a dark outline; or change the colours of the cams and lifters slightly so you can tell the parts apart. Do BOTH: outlines and slight tints.
   
   THE DESIGN (implement exactly this; index.html, engine group around lines 36-150):
   
   A. LIGHT parts (everything filled url(#steelGrad), i.e. head block, cam towers, crankcase pan, tubes and outer disc, crank disc and crank pin disc, piston body, rod bosses, valve heads / stems / barrels, cam lobes, rollers): dark outline stroke #0E1116, stroke-width 2, stroke-linejoin round. The rod polygon, rodSmall and the piston already have dark outlines on some parts: make all consistent. Easiest: one CSS rule block keyed on fill (for example [fill="url(#steelGrad)"] { stroke:#0E1116; stroke-width:2; stroke-linejoin:round }) plus the same for any element you give a new flat fill. Note CSS beats presentation attributes. Do NOT add an outline to the big black cavity rects (fill #0E1116) or to the hatch bands.
   
   B. DARK parts (fill #2B3038: both plug rects, valve retainers and washers, wrist pin, pin boss, the 6 piston ring-groove rects, the two cam bearing holes r12, the counterweight, the dimmed flywheel disc): light outline stroke #AEB8C2, stroke-width 1.5 (same CSS approach, key on fill="#2B3038").
   
   C. SPRINGS: the dark coil needs a light edge. Draw an underlay polyline for each spring (springIntakeUnder, springExhaustUnder) with stroke #C9D1D9, stroke-width 9, stroke-linejoin round, fill none, using exactly the same points as the coil, placed just BELOW the dark coil in z-order; set the dark coil to stroke-width 5. Both polylines must be updated every frame with the same points. The spring must still be in front of the head casting.
   
   D. BEARING RINGS (the two r20 stroke rings at (856,204) and (944,204)): add an underlay circle at the same centre, r20, fill none, stroke #0E1116, stroke-width 10, just below each ring, so the ring has a dark edge on both sides.
   
   E. SLIGHT TINTS (so the parts differ even before you look at the outlines). Change only the fills of these, nothing else:
      - cam lobes (camIntake, camExhaust): #B7A88E (a muted tan)
      - rollers (rollerIntake, rollerExhaust, the lifters): #D8CEBA (a lighter tan)
      - valve stems, barrels and heads (valveStem*, valveBarrel*, valveHead*): #C5CDD5 (a lighter cool grey than the casting)
      Everything else keeps the steel gradient. Do NOT use blue, yellow, orange or red on any metal: those hues mean gas in the legend. These tans are muted on purpose.
   
   F. Do not touch the callout leaders, the dial, the captions, the HUD, the zoom or the animation. Outline widths may scale with the 1.0 to 1.2 zoom; that is fine.
   
   CHECK: node measure_engine.mjs --times 5.5,7.33,8,11.5,16,21.5,24,27.5,31 --shots workspace/meas6 . All 17 checks must pass. If a check fails ONLY because it samples a colour you changed on purpose (for example a contrast or spring-visible test), adapt that check in measure_engine.mjs minimally, and say exactly what you changed and why in REPORT.md. Do not loosen any check that a real defect would trip.
   Then LOOK (ffmpeg -ss T from the new final.mp4; head crop = crop=330:260:760:150 then scale 4x; lower crop = crop=500:640:650:370 then scale 2x) at t=2.0, 5.5, 11.5, 24 (head) and t=2.0, 9.7 (lower). For each of these pairs, state in REPORT.md what visibly separates them: cam lobe | cam tower, roller | cam lobe, valve barrel | stem | head, retainer | spring, spring | head casting, plug | background, piston | cylinder wall, rod | crank disc. Every pair needs a visible edge or a visible tint step. If any pair still merges, fix it and look again.
   
   THEN: BEFORE rendering back up the round-5 renders: final.mp4 -> final_v5.mp4 and review.mp4 -> review_v5.mp4 (do not overwrite older backups). Render review.mp4 and final.mp4; ffprobe gate (h264, 1 video stream, 1920x1080, 30/1, 32.0 s +/-0.05, no audio); append round 6 to REVIEW.md, REPORT.md, NOTES.md. Final report: per item done or not done with the evidence, the gate, files written; be honest about anything still imperfect.

8. Stop 7

   The connecting rod gets its own tint

   Three fresh reviewer agents checked the round 6 video at 24 different times. They found all 14 pairs of touching parts now read as separate parts, no metal looking like a gas colour, and every earlier fix still holding. All three found one weak spot: the connecting rod, where it lies over the light crank disc, was nearly the same grey as the disc with only a thin dark line between them. The prompt gives the fix (one flat light grey on the rod, different from the valves, keeping the dark outline), asks for the measuring tool and four crops as checks, and asks for two things the reviewers noted to be listed as accepted and not fixed: the callout leader covering the roller, spring and stem for the 2.2 seconds of each callout (the approved design) and the small diagonal tick where the last coil of each spring meets the stem. The agent applied it and rendered at 13:33.

   Show the prompt (2,849 characters)

   Delivered 2026-11-09 13:30:51 (local time of the computer running the agent). Plain-text file

   continue. FIX ROUND 7, one small item, design already decided, just implement. Keep your reasoning short: one edit, run the measuring tool, look at two crops.
   
   Three fresh independent reviewers checked your round-6 final.mp4 at 24 different times. Result: ALL 14 part pairs now read as separate parts (cam lobe | tower, roller | lobe, barrel | stem | head, spring | retainer, spring | casting, plug | background, piston | wall, rings and pin | piston, crank pin | disc, counterweight | flywheel, valve head | seat), no metal looks like a gas colour, and every earlier fix still holds (leaders, dial halo, last-frame caption, piston, springs, symmetry). Well done.
   
   ONE weak spot all three found: the CONNECTING ROD where it lies over the light crank disc / web (visible at t=13, 18, 19.5, 21, 26, 28, 29 and others). Rod and disc are nearly the same steel grey and only a 1.5 px dark line separates them. Kavan's rule was: if parts are the same colour, tint one of them slightly. So:
   
   THE FIX (index.html, rod group around lines 66-72): tint the rod family a step lighter than the crank and casting steel, using ONE flat colour #B4BEC9 (cool light grey; the valves are #C5CDD5, keep that different) on: rodPoly, rodBig, rodSmall. Keep their existing dark outline (stroke #0E1116, stroke-width 2) on all three; rodBig and rodSmall currently get their outline from the CSS rule keyed on fill="url(#steelGrad)", so after you change their fill you must add them (by id) to that outline rule, or give them stroke attributes directly. Do not change the crank disc, crank pin disc, counterweight, piston or anything else. No blue, yellow, orange or red hue on the rod.
   
   CHECK: node measure_engine.mjs --times 5.5,7.33,8,11.5,16,21.5,24,27.5,31 --shots workspace/meas7 (all 17 must still pass). Then look (lower crop = crop=500:640:650:370 then scale 2x, from ffmpeg -ss T of the NEW final.mp4) at t=13, 19.5, 26 and 29: the rod must read as a lighter part than the crank disc where they overlap, with its outline still visible, and the rod must still look like one part with its big-end boss and small end.
   
   THEN: BEFORE rendering back up the round-6 renders: final.mp4 -> final_v6.mp4 and review.mp4 -> review_v6.mp4 (do not overwrite older backups). Render review.mp4 and final.mp4; ffprobe gate (h264, 1 video stream, 1920x1080, 30/1, 32.0 s +/-0.05, no audio); append round 7 to REVIEW.md, REPORT.md, NOTES.md. Final report: done or not done with evidence, the gate, files written. Two things the reviewers noted that you should list in REPORT.md as accepted and not fixed: (a) the callout leader runs down the valve axis and covers the roller, spring and stem for the 2.2 s of each callout (this is the approved design), (b) the last coil of each spring ends in a small diagonal tick on the stem axis at the retainer (by design, it seats on the stem).

9. Stop 8

   The valves were out of time and the wrong shape

   The owner looked at the outlined engine and said the valves looked out of time and maybe the wrong shape; the prompt says he was right on both and gives the design. Part A, timing: the two cam lobes were rotated by the cam angle only, while the specification gives each lobe's nose angle as the cam angle plus 126.25 degrees (intake) or minus 126.25 degrees (exhaust). The prompt asks for each lobe to be rotated by its nose angle, for two new measuring checks (rotation equal to the nose angle; roller touching the lobe within 3 px), and for a control run on a copy with the old rotation to show the checks fail. Part B, shape: the flat rectangular valve heads become poppet-valve heads (flat bottom, conical face, neck into the stem) with fixed seat wedges added to the head. The agent applied both and rendered at 14:12 (the round 8 video). Its own report says the lobe error was there from the first build and that neither its own checks nor the earlier reviewers caught it, and that the worst roller-to-lobe gap went from 15.5 px to 1.8 px. Three more reviewer agents then checked the result by measuring (see How it was built).

   Show the prompt (6,380 characters)

   Delivered 2026-11-09 13:55:30 (local time of the computer running the agent). Plain-text file

   continue. FIX ROUND 8: the valves, two parts. The owner looked at the outlined engine and said the valves are out of time and maybe the wrong shape. He is right on both. The design is already decided: just implement. Keep your reasoning short: one edit, run the checks, look at crops, next edit.
   
   PART A - TIMING (root cause found and confirmed by arithmetic, so do not hunt for another cause).
   index.html lines ~589-590 rotate BOTH cam lobe groups by st.camAngle only:
     $('camIntakeG').setAttribute('transform', 'rotate(' + st.camAngle ... ' 856 204)');
     $('camExhaustG').setAttribute('transform', 'rotate(' + st.camAngle ... ' 944 204)');
   But SPEC.md line 40 says the lobe NOSE angle (clockwise from up) is theta/2 + 126.25 for the intake and theta/2 - 126.25 for the exhaust, so that the nose points straight down at peak lift, and assets/engine-state.js already returns exactly those as st.camNoseIntake and st.camNoseExhaust. The drawn lobe path (camPath) has its nose at phi = 0, so the group must be rotated by the NOSE angle. The valves lift correctly from the spec curve, but the lobes are drawn 126.25 degrees out of phase, so the roller floats away from the nose while the valve is open (gap up to 15.5 px at peak lift) and the lobe overlaps the roller while the valve is shut (6.4 px). A probe with the lobe rotated by the nose angle gives a worst roller-to-lobe gap of 1.8 px (discretisation of the 2.5 degree lobe path): read cam_probe.py (it only prints, never writes) and run it with py -3.14 -I.
   THE FIX: rotate camIntakeG by st.camNoseIntake and camExhaustG by st.camNoseExhaust (same centres 856,204 and 944,204). Nothing else in Part A.
   CHECK A (add to measure_engine.mjs as two new checks, then run it): (1) at every sampled time the camIntakeG rotation equals camNoseIntake and camExhaustG equals camNoseExhaust, both within 0.05 degrees; (2) the roller touches the lobe: with the lobe path rotated by the transform, the distance from the roller centre to the nearest lobe-surface point minus the roller radius (8 px) is within +-3 px at every sampled time, for BOTH valves (use the maths in cam_probe.py, or measure it from the DOM with getPointAtLength / a point-in-polygon sampling, your choice). The existing 17 checks must also still pass. Control: temporarily revert the rotation in a COPY of index.html (never in place) with --file and confirm the new checks FAIL on that copy (they must fail on the old behaviour or they prove nothing).
   LOOK A: 4x head crops (crop=330:260:760:150 then scale 4x) at t=1.0, 2.0, 3.5, 5.0, 12.0, 13.6, 17.8: where a valve is open the lobe nose must visibly be pushing its roller down (nose pointing straight down at the lift peak: intake peak at t = 2.39 s and 18.39 s, exhaust peak at t = 13.61 s and 29.61 s); where a valve is shut the roller rests on the round base of the lobe.
   
   PART B - SHAPE. The valve heads are flat rectangular plates (T-shaped valves). Real poppet valves have a flat-bottomed mushroom head: a short vertical margin at the rim, a conical seat face going up and inward, then a neck into the stem. Implement it like this, keeping every existing id and every position the measuring tool and the leader dots use:
    B1. Keep the existing <rect id="valveHeadIntake"> and <rect id="valveHeadExhaust"> exactly as they are for geometry (the leader dots and the measuring tool read them) but make them invisible: fill="none" stroke="none" (do not delete them, do not change their per-frame y/height code).
    B2. Add one polygon per valve, drawn where the old rect was drawn in z-order: valveShapeIntake and valveShapeExhaust, fill #C5CDD5, with the same dark outline as the other tinted parts (add the fill to the existing CSS outline rule, or set stroke #0E1116, stroke-width 2, stroke-linejoin round on the polygon). Author each polygon at the CLOSED position and move it every frame by transform translate(0 dy) where dy = (current head rect y) - (closed head rect y), so it follows the head exactly. Points, with cx = the valve axis x (856 intake, 944 exhaust), y0 = the closed top of the head rect, W = the head rect width (56 intake, 48 exhaust), head height 16 px: (cx-W/2, y0+16), (cx+W/2, y0+16), (cx+W/2, y0+13), (cx+11, y0+2), (cx+7, y0), (cx-7, y0), (cx-11, y0+2), (cx-W/2, y0+13). The stem (14 px wide, half-width 7) joins the neck at y0, so stem and head read as one valve; the stem and barrel keep their tint.
    B3. The seat. A cone needs a matching seat or the closed valve leaves ugly triangular gaps. In the head casting, at each valve, add a seat wedge on each side of the valve: a steel-gradient polygon with the usual dark outline whose underside is exactly the head cone, i.e. left wedge = (cx-W/2, y0+13), (cx-11, y0+2), (cx-W/2, y0+2); right wedge = (cx+W/2, y0+13), (cx+11, y0+2), (cx+W/2, y0+2) (read the real pocket numbers in the headG group around lines 96-125: the pocket walls are at x = cx-W/2 and cx+W/2, y0 is the roof line; adapt if your numbers differ, the intent is: with the valve closed the head cone is in contact with the wedge undersides, no gap, and with the valve open a clear gap opens between cone and wedge so the gas path reads). Put the wedges after the dark pocket rects and before the valve shapes in z-order. They do not move.
    B4. Do not change the valve lift, the port paths, the gas dots, the leaders, the dial, the captions, the springs or the outlines elsewhere.
   CHECK B: the measuring tool's 17 checks (plus the two from part A) all pass. 4x head crops at t=7.0 and 9.0 (both valves closed): the cones seat against the wedges with no visible gap; at t=2.0 (intake open) and t=13.6 (exhaust open): the open valve shows the poppet shape with a gap between cone and wedge; the leader dot still sits on the head (t=5.5, 11.5, 21.5, 27.5).
   
   THEN: BEFORE rendering back up the round-7 renders: final.mp4 -> final_v7.mp4 and review.mp4 -> review_v7.mp4 (do not overwrite older backups). Render review.mp4 and final.mp4; ffprobe gate (h264, 1 video stream, 1920x1080, 30/1, 32.0 s +/-0.05, no audio); append round 8 to REVIEW.md, REPORT.md, NOTES.md. Final report: per part done or not done with the numbers (the roller gap before and after, the control run) and the gate, files written; honest about anything still imperfect. Say plainly in REPORT.md that the lobe rotation bug existed from the first build and that your earlier reviews did not catch it.

10. Stop 9

    The owner's own request: port edges, chamber spaces, shading

    This one is not written by the assistant: it is the owner's own typed message (typos and all). He asks for three things: the abrupt edge at the intake and exhaust ports, which he marked in red; the combustion chamber's open spaces next to the valves, which real engines do not have; and shading and gradients on certain parts to show depth and detail. He also tells the agent to ask him for any input it needs before it starts. A marked screenshot was attached to the chat (the text of this message does not show it). The agent asked in three rounds by a form tool. The first set of questions (16:59) timed out unanswered after 10 minutes; the third (17:54) was cancelled. He answered the second set at 17:34: for the ports, "Yes — continuous ports from face to pocket"; for the chamber, "There is three spaces in the chamber that needs to be filled in left of the intake port in between the 2 ports and right of the exhaust port."; for what to shade, "all of it"; and for how strong, "Noticeable — clearly sculpted, still the flat-colour style" (the last is one of the choices he was offered). The agent made no render in this stop. Its thinking degraded badly during it (see How it was built); the turn ended at 17:54:46 with the log reason "interrupted_by_user" (who interrupted is not recorded), and the assistant then wrote stop 10.

    Show the prompt (431 characters)

    Delivered 2026-11-09 16:56:55 (local time of the computer running the agent). Plain-text file

    this is what i want fixed the intake and exhaust ports have an obrupt edge marked in red that needs to be removed
    also the compression chamber has areas next to the valves (open space) that is not really there in real engines and needs to be fixed also i would like some shading and gradiants on certain parts to show some depth and detail. ask me for any input (before you start) you might need if you dont understand anything ask

11. Stop 10

    Round 9, with the owner's marks mapped to pixels

    The assistant could read the owner's marked screenshot (the agent could not) and mapped his five red marks to engine pixels: marks 1 and 2 are pointed peaks of casting just above the intake and exhaust port mouths at the head's side faces; marks 3, 4 and 5 are gas notches at the chamber roof, left of the intake valve, between the two valves around the plug, and right of the exhaust valve. It tells the agent not to ask the owner anything more and not to run picture-analysis loops for geometry (use the coordinates and pixel probes), to keep its reasoning short, to stop at about 40 tool calls, and why: the agent's thinking had filled with runs of "= = = =". Part 1, ports: each port's centre line started at the side face heading down at 42 degrees, so the 48 px butt end left a pointed tooth of casting; the fix makes the stroke leave the face perpendicular. Part 2, chamber: raise the closed valve heads and seat wedges 16 px so the shut face is flush with the chamber roof, make the roof solid between and beside the valves, and narrow the plug well from 32 to 24 px with a 6 px electrode. Part 3, shading: gradients per family of parts (about 8 defs), with outlines, round 6 hues and the no-gas-colours-on-metal rule kept. The agent applied all three (its report says 12 gradient definitions), and the renders were written at 18:26 and 18:27. Three reviewer agents then checked the result (see How it was built).

    Show the prompt (6,255 characters)

    Delivered 2026-11-09 18:12:31 (local time of the computer running the agent). Plain-text file

    continue. FIX ROUND 9, design already decided by the owner's assistant, who CAN read the owner's marked screenshot (you could not). Do NOT ask the owner any more questions: he already answered you (continuous ports, fill the three chamber spaces, shading on all parts and noticeable). Do NOT run vision loops for geometry: use the coordinates below and pixel probes. Keep your reasoning short and in plain sentences. About 30 minutes ago your thinking filled with runs of '= = = =' (up to 31 percent of its characters) because one turn ran 38 minutes and 35 calls and each thought fed the next; if you notice your thoughts turning into symbols or repeats, stop thinking and make the next edit. Work in small steps. After EACH of the 3 parts: run the measuring tool, look at one crop, write 5 lines of progress, then continue. If you reach about 40 tool calls, end the turn with a short status and wait for 'continue'.
    
    WHAT THE OWNER MARKED (his screenshot is an attachment to his message; the owner's assistant mapped the 5 red marks onto engine pixels, an overlay is at head_t6_marks.png, yellow rulers in engine px; do not rely on viewing it, use the numbers):
     Mark 1: from (765,251) curving down to (797,286): the pointed peak of casting just ABOVE the intake port mouth at the head's left face.
     Mark 2: from (1006,280) up to (1050,244): the same pointed peak above the EXHAUST port mouth at the right face.
     Marks 3, 4, 5: at the chamber roof, y 392 to 399: (3) x 818-827, the gas notch left of the intake valve head and its seat wedge; (4) x 880-922, the gas gap between the two valves around the plug; (5) x 972-990, the notch right of the exhaust valve head.
    
    PART 1, PORTS (ROOT CAUSE FOUND, tiny fix). In index.html around lines 449-457 the port centrelines are single quadratic beziers in mm: portIn = [[-66,252],[-46,234],[-22,226]] and portEx = [[22,226],[46,234],[66,252]]. The stroke (48 px wide, butt cap) therefore STARTS exactly at the head's side face (x=768 and x=1032, y=296) heading down-inward at 42 degrees, so the butt cap is a diagonal cut across the face and leaves a pointed triangular tooth of casting above (and below) each mouth. That tooth is what the owner marked. FIX: make each port leave the head perpendicular to its face: set the FIRST control point's y equal to the start point's y so the start tangent is horizontal: portIn = [[-66,252],[-46,252],[-22,226]] and portEx = [[22,226],[46,252],[66,252]]. The cap is then a vertical line exactly on the face, the mouth is a clean 48 px opening and the teeth disappear. The gas dots use the same P, so they follow the new curve; keep the intake/exhaust mirror symmetry. Check both mouths in a 5x crop at t=6 (valves shut): no triangular tooth above or below either mouth; the passage runs smoothly into the valve pocket.
    
    PART 2, CHAMBER (marks 3, 4, 5). Real engines have a flat chamber roof with the valve faces flush in it: the closed valve head must not hang into the chamber and the gas must not wrap around it. Today the closed head spans y 376-392 (its top at the roof y=376), with seat wedges around it, so gas shows beside and under it. DESIGN: raise the whole closed valve-head assembly by 16 px so the closed face is FLUSH with the roof: closed head y 360-376 (rect id valveHeadIntake/Exhaust y = mmY(220 - lift) = 360 + 2*lift; poppet polygons valveShapeIntake/Exhaust and the four seat wedges translated up by 16 px; the stems end at the head, shorten each stem by 16 px at its lower end; the head still moves DOWN by 2*lift px as it opens, and the roller, spring and cam geometry do not change). The roof between and beside the valves is then solid casting at y=376 from x=820 to x=980 (draw the casting down to y=376 everywhere in that span), so no gas notch remains at marks 3, 4 or 5. The plug well is an open 32 px slot around the 20 px plug (x 884-916): narrow the well to 24 px (x 888-912) so the plug is embedded in solid head, and let the plug tip end flush at the roof with a small electrode 6 px long below it (y 376-382, same dark tone with the light outline) so the spark still visibly fires at the tip (spark centre stays near (900,380)). Update everything that referenced the old head position: the leader dots' anchor (the dot sits on the head centre = new head rect centre), measure_engine.mjs checks that use the head rect (adapt only what the geometry change forces, say exactly what you changed), the gas clip, flame start. Then verify at t=6 (shut), t=2.4 (intake open) and t=13.6 (exhaust open) in 5x crops: no gas visible beside a shut valve, a clean gap when open, roof continuous.
    
    PART 3, SHADING. Add noticeable depth and detail with SVG gradients, keeping the flat-colour style, the round-6 hues, the outlines, and NO blue, yellow, orange or red hue on metal. Per family: head casting (vertical light to darker toward the bottom/roof), piston (crown highlight, darker skirt below the rings), bore walls/tubes (inner shadow toward the bore), crank web, crank disc and flywheel (radial turned-metal: lighter centre, darker rim), connecting rod (lengthwise gradient, lighter in the middle), valve stems (horizontal cylindrical gradient: dark edges, light centre), valve heads and seat wedges (diagonal light-to-dark following the cone), cam lobes and rollers (radial highlight), plug (subtle vertical). Extend the outline CSS rule (or add stroke attributes) to the new gradient fills so outlines stay. Use a small set of defs (about 8 gradients), not one per element. The shading must be visibly stronger than today at 1x on a 1920x1080 frame, yet text, leaders, dial and HUD stay untouched.
    
    CHECK: node measure_engine.mjs --times 5.5,7.33,8,11.5,16,21.5,24,27.5,31 --shots workspace/meas9 : all checks must pass (the cam-lobe checks from round 8 included). Do not loosen a check that a real defect would trip.
    THEN: BEFORE rendering back up the round-8 renders: final.mp4 -> final_v8.mp4 and review.mp4 -> review_v8.mp4 (do not overwrite older backups). Render review.mp4 and final.mp4; ffprobe gate (h264, 1 video stream, 1920x1080, 30/1, 32.0 s +/-0.05, no audio); append round 9 to REVIEW.md, REPORT.md, NOTES.md. Final report: per part done or not done with evidence, the gate, files written, honest about anything imperfect.

12. Stop 11

    Round 10: make the shading noticeable where it was faint

    Three reviewer agents had found the ports, the chamber and the dots on the heads good, and shading visible on the piston, crank disc, valve stems, plug and rollers. Their one miss: the large grey head-casting band had a gradient of only a few grey levels because it was spread over the whole head block, and the hatch bands, water jackets, crankcase tubes, outer crank disc and dark crank ring were still flat. The prompt lists seven changes (a gradient across exactly the visible band with named colours, a highlight line and a shadow line on its edges, cylindrical gradients on the crankcase tubes, a radial gradient on the outer crank disc, shading for the dark crank ring and counterweight, a left-to-right teal gradient on the water jackets, and the same light-to-dark on the cam towers). It asks for measured proof instead of a look: at least 35 grey levels between top and bottom of the casting band and at least 25 on a crankcase tube, the outer crank disc and a water jacket. The agent reports 94.3, 42.7 and 44.0, 36.1 and 26.2, and rendered at 18:47 (the video above).

    Show the prompt (4,296 characters)

    Delivered 2026-11-09 18:36:38 (local time of the computer running the agent). Plain-text file

    continue. FIX ROUND 10, one item: make the shading NOTICEABLE on the parts where it is still faint. Design decided, just implement. Keep your reasoning short and in plain sentences; if your thoughts turn into symbols or repeats, stop thinking and edit. Stop at about 40 tool calls and wait for 'continue'.
    
    Three independent reviewers checked your round-9 render. Verdict: the ports are fixed (no tooth at either mouth, continuous passages), the chamber is fixed (valve faces flush with the roof, no gas notch beside or between the valves, plug embedded with its electrode, spark and flame stay inside, dots still on the heads, cam timing and poppet valves hold, no colour shift), and shading is clearly visible on the piston, crank disc, valve stems, plug and rollers. ONE miss: the large grey head-casting band (engine x 770-1030, y 256-362) has a gradient of only a few grey levels (headGrad runs #A9B4BF to #7C8794 over the WHOLE 136 px block including the hatch bands, so the visible band gets a sliver of it), and the hatch bands, the teal water jackets, the crankcase tubes, the outer crank disc and the dark crank ring are still flat. The owner asked for noticeable shading on all parts, flat-colour style kept.
    
    DO THIS (index.html defs and fills; keep every outline, tint and hue rule; no blue, yellow, orange or red on metal):
     1. headGrad: change to gradientUnits="userSpaceOnUse" x1=0 y1=256 x2=0 y2=362 so the visible casting band spans the WHOLE range, with stops 0 #CBD5DF, 0.5 #98A3AE, 1 #5F6A77 (about a 35 percent lightness range). Make sure every element that uses headGrad still shows correctly (the band is a rect y 240-376; the hatch bands drawn over it keep their pattern).
     2. Add a 2 px highlight line along the top edge of the casting band (y 256-258, fill #E3EAF0, opacity 0.55) and a 2 px shadow line along its bottom edge (y 360-362, fill #3E4753, opacity 0.5), x 768-1032, drawn before the valve train and the plug so nothing is covered.
     3. Crankcase tubes (the two rects x 780-820 and 980-1020, y 376-694): horizontal cylindrical gradient, edges #7F8996, centre #BAC4CE (new def tubeGrad, userSpaceOnUse per tube or two defs).
     4. Outer crank disc (the r=160 circle) and the pan: radial turned-metal like crankGrad but centred up-left: centre #B5C0CB, mid #8E99A6, rim #66707C.
     5. Dark crank ring / flywheel (the #2B3038 dimmed disc and ring, and the counterweight): radial or linear from #3C444E to #1E232A so they read as shaded dark metal (keep their light outline).
     6. Water jackets (teal #4C6A73 bands): left-to-right gradient, #5F8190 to #3B555D (still teal, not a gas colour).
     7. Cam towers and bearing rings: the same vertical light-to-dark as the new headGrad (use headGrad with userSpaceOnUse is fine only if it looks right; otherwise a copy towerGrad y 204-240).
     Extend the outline CSS rule to any new fill so outlines stay.
    
    CHECK: (a) node measure_engine.mjs --times 5.5,7.33,8,11.5,16,21.5,24,27.5,31 --shots workspace/meas10 : all checks pass. (b) A NUMBER, not a vision look: from the live-page screenshot of the tool (or a new headless screenshot of the edited page at t=6), print the mean grey of the casting band at engine px (800, 262) and (800, 356): their difference must be at least 35 levels (0-255), and the same kind of top-to-bottom or left-to-right difference of at least 25 levels on a crankcase tube cross-section, on the outer crank disc (upper-left versus lower-right of its visible rim) and on a water jacket. Print the numbers in your report. (c) Then ONE look at a 4x head crop from that screenshot: shading visible at 1x, no banding, no coloured metal.
    THEN: BEFORE rendering back up the round-9 renders: final.mp4 -> final_v9.mp4 and review.mp4 -> review_v9.mp4 (do not overwrite older backups). Render review.mp4 and final.mp4; ffprobe gate (h264, 1 video stream, 1920x1080, 30/1, 32.0 s +/-0.05, no audio); append round 10 to REVIEW.md, REPORT.md, NOTES.md. Final report: done or not done with the numbers, the gate, files written, anything imperfect. Two reviewer remarks to list in REPORT.md as accepted, not fixed: the gas crossfade from red to blue passes through purple for a few frames around t=15.9 (a blend between the two phases; you did not change it), and the spark glow is a hard-edged pale disc (as before).

Messages that only restarted the agent after tool-limit or router-limit stops, tool coaching, and automatic notices are not shown. Two more messages in the stop 9 to 10 stretch are not shown either: the owner's message that attached his marked screenshot (a picture, with a one-line question whether the agent could see it) and a short "continue" that told the agent to finish and render round 9.

Notes

• Who wrote the prompts. The initial brief was submitted from the owner's desktop chat; who wrote it is not recorded. The assistant wrote the fix prompts, except one: stop 9 is the owner's own typed message. The other ten fix prompts were written by an AI assistant (a Claude Sonnet 5.5 session) on the owner's behalf: stops 1, 6 and 8 from the owner's feedback, stops 2, 3, 4, 5, 7 and 11 from the reviewers' findings (stop 4 re-sends the round-3 reviewers' open item with the design decided), and stop 10 from the owner's marked screenshot and his answers to the agent's questions. For stops 4 to 11 this rests on the assistant's own record: the agent's database keeps the text of those prompts but not who sent them. Stop 6 carries the owner's request for outlines and tints, stop 8 his remark that the valves looked out of time and maybe the wrong shape, and stop 9 his own words. The agent's own reports are not included.
• Verbatim. Each prompt is shown exactly as it was sent, including the owner's first name where it appears in them; everywhere else on this site he is called "the owner". Only the names of folders on the owner's computer were replaced by generic names (for example library.py, ffprobe, measure_engine.mjs), the path of the owner's attached screenshot and of an overlay image was removed from stop 10, and the file-attachment header that wrapped the initial brief was dropped. The initial prompt names the agent's helper sub-agents and tools (Bolt, vision_analyze, delegate_task) as the agent's own environment knew them.