A mech game prototype, pushed by GPT-6-Astra

agent
Codex 0.167.2
model
gpt-6-astra
harness
Codex
started
2026-11-15 03:13 UTC
took
83 min · 276 turns

user

+0:00

Your brief is in ~/context/brief-sortie-full.md. Read it in full and carry it out.

agent

+0:04

I’ll read the full brief and repository instructions, then carry out the work and verify the result.

  • exec_commandcat ~/context/brief-sortie-full.md (+1 more)

    command

    cat ~/context/brief-sortie-full.md
    
    pwd && rg --files -g 'AGENTS.md' -g 'package.json' -g 'README*' -g 'SKILL.md' -g 'pyproject.toml' -g 'Cargo.toml' -g 'go.mod' . ~/context

    result, 4 lines

    Script completed
    Wall time 0.1 seconds
    Output:
    {"chunk_id":"cf1286","wall_time_seconds":0.000006001,"exit_code":0,"original_token_count":967,"output":"You are running as GPT-6-Astra (gpt-6-astra) at effort high, in a session of your own. The owner, severian (they/them), asked to see what GPT-6-Astra would do with one of their lines of work, sortie, and to see one visual artifact you can make with what the project already has. The project's own session runs on a different model and is not to be touched; you are a separate, parallel look.\n\n## What sortie is\nThe owner's long-held dream: an original browser mech game (influences: Gundam 00's look, Lancer's frames and mounts, Front Mission's part swapping, Code Geass's knightmare traits for the light class). Mechs are built by Blender scripts. The look should be sleek and realistic, or a stylised sketch look, never toon robots; shape was always the real problem. On 11-08 the owner decided to restart with one mech from scratch, to find a faster, better way of working together; they haven't started yet.\n\nRead, in this order:\n1. `docs/handbook.md` in your worktree (entry point: each rule marked earned or guessed; guessed rules don't bind the restart)\n2. `~/context/reflect/sortie.html` (the 11-08 reflection; the box at the top is the restart plan)\n3. `~/context/project_sortie_mech_game.md` (history of the owner's reactions and decisions)\n4. Whatever else in the repo you judge useful (docs/heads.md, docs/sizes.md, docs/lore.md, the kit, the tools the handbook lists).\n\n## Your workspace\n- Your current folder is the git worktree `~/work/sortie` (branch `astra`, made for you from main c257287). Work only here; you may commit here.\n- `node_modules` in your worktree is a copy of the main checkout's: use it read-only (no installs).\n- Blender 5.2.3 is at `~/.local/bin/blender`; only with `-b`.\n- Output folder: `~/out/` (copied to the owner's review route afterwards).\n- Keep designs original: no Gundam names or lookalikes. The Phantom Brigade study notes are learning only; nothing from its data goes into the work.\n\n## Rules\n- Headless only: no windows on the owner's desktop. Never run anything that drives Windows programs (no tools/winvroid, no powershell.exe, cmd.exe or anything under /mnt/c). Browser rendering via headless Chromium only (Playwright is at `~/work/sortie/node_modules/playwright`; page-side code compiled by tsx can throw \"__name is not defined\", so keep page-side code as plain .js evaluated as text). Blender only with `-b`.\n- Don't install packages. Don't run the full test suite (a test run once ate 36 GB); targeted checks only.\n- Nothing runs on the C1 machine; no ssh. No network uploads, no publishing anywhere. Never push.\n- Edit files only in your worktree and your output folder. Never push.\n\n## Deliverable\nOne self-contained HTML page, `index.html` in your output folder, with its images (or a short clip) beside it. The owner finds wordy interfaces tiring: organise, don't narrate; short labelled sections, plain words, no em dashes. At the top: \"GPT-6-Astra (gpt-6-astra), effort high\" and the date.\n1. **Your read of this line of work:** what's working, what isn't, what you'd keep and drop, and what you would do differently in the restart. Give opinions plainly; disagree with the reflection page where you do.\n2. **What you'd do next** if you ran the restart's first session: a short, concrete sequence.\n3. **The one visual:** a single visual artifact made with what the project already has (its tools, kits, renderers, models, data). Your choice of what it is; pick the thing that best shows what you'd do with this work. Show it large, then say in a few lines how it was made (the commands, wall time) and what you'd change next.\nKeep it to one visual; don't build a gallery. Aim to be done in roughly 90 minutes of work.\n\nWhen done, end your turn with: the page path, a 5-line summary, and any commits on your branch. If something only the owner can decide blocks you, say so and stop rather than guessing.\n"}{"chunk_id":"1a960f","wall_time_seconds":0.000005501,"exit_code":0,"original_token_count":13,"output":"~/work/sortie\n./README.md\n./package.json\n"}
  • exec_commandcat docs/handbook.md (+1 more)

    command

    cat docs/handbook.md
    
    for p in /AGENTS.md [redacted: a lookup for an instructions file, reads as a home folder] ~/AGENTS.md ~/work/AGENTS.md ~/work/sortie/AGENTS.md; do if [ -f "$p" ]; then echo "$p"; cat "$p"; fi; done

    result, 4 lines

    Script completed
    Wall time 0.1 seconds
    Output:
    {"chunk_id":"561984","wall_time_seconds":0.00000509,"exit_code":0,"original_token_count":2794,"output":"# Handbook\n\nHow to work on sortie, for whoever picks it up cold. Read this first. Written 2026-11-08, after the\nretro. The owner has not reviewed it. Detail lives in the links.\n\n## Where things stand\n\n- **The game runs.** Flight, contracts, factions, salvage, thrusters, audio and tests are in good shape.\n  README.md has the controls, `CLAUDE.md` the code rules.\n- **The mech look is not done.** The owner is starting over with one mech from scratch, to find a faster\n  and better way to work on mechs. **They start it. Don't start mech work on your own.**\n- **The old makers are parked.** The common pattern and the houses' atelier are review models only, not\n  in the game. The game still flies the striker and the older frames.\n- **Why it went this way:** the retro, `~/citadel-render/workers/handoff/reflect/sortie.html` (outside the\n  repo, sources marked). The first-session plan is the box at its top. It's dense and the owner isn't\n  commenting on it, so don't wait for a reaction.\n- State of every doc, script and big file: [housekeeping.md](housekeeping.md).\n\n## The first session\n\nNothing from the old rounds carries over unless the owner brings it back. Work directly, no workers in\nbetween, so a comment turns into a new picture quickly. Log each step, so you can both see the speed.\n\n1. **Ask one question.** Think of a moment in any game or show when a mech made you feel \"that's it\".\n   What was it doing? Follow up from the answer. No boards or menus yet.\n   `tools/log.sh \"owner: ...\"`\n2. **Show 12 outlines.** Write a spec (format at the top of `tools/outlines-lib.mjs`) and show it.\n   Commit nothing yet: `review/out/` is gitignored.\n   `npm run -s outlines -- review/out/one-mech/outlines-01.mjs`\n   `tools/show.sh review/out/one-mech/outlines-01.png`\n3. **Redraw from what they say** until two or three feel alive. They cross out, you don't ask them to pick.\n4. **Block one out.** Copy `kit/blocks/dummy.py` to `kit/blocks/<name>.py`, set joints and boxes (API in\n   `kit/block.py`), read the contact sheet yourself, then show the clip.\n   `npm run -s kit -- block_<name>`\n   `npm run -s clip -- block_<name> --quick`\n   `tools/show.sh review/out/one-mech/clips/block_<name>.mp4`\n5. **Stop and compare notes.** What felt slow, what helped. The log is `review/out/one-mech/log.txt`.\n\n## Rules: earned or guessed\n\n**Earned:** the owner said it, or it visibly worked for them. **Guessed:** I invented or borrowed it and the\nowner never chose it. **A guessed rule does not bind the restart.** Ask, or test it, before relying on one.\nDocs still state many guessed rules as fact, so check here first.\n\n### The look\n\n| Rule | Status | Evidence |\n|---|---|---|\n| Realistic machines, no toon robots. Shape is the real issue, a stylised render is acceptable | Earned | \"I don't want toon robots\" |\n| Judge from every angle, not only front, side and back | Earned | \"you need to take into account other angles\" |\n| Backs must look finished | Earned | \"the backs are weak\", \"the back is bare\" |\n| More segmentation, depth, parts that stick out | Earned | \"more segmentation\", \"parts sticking out\" |\n| Faces need depth | Earned | \"flat at the front\" |\n| Feet, head and chest get the closest look | Earned | Raised again and again, feet in four messages |\n| Cockpit hatch, knees and elbows should look covered | Earned, weakly | \"are we deliberately not covering\", a question, I took it as a yes |\n| Heavier means more plates and intricacy, not only bulk | Earned | \"more armor plating, not just bulkier ones\" |\n| Keep real variety between makers | Earned | \"all four look distinct right now and I appreciate that\" |\n| Domed, rimmed, layered plates beat cones and tubes | Earned, once | Atelier round 1 \"bland\", the board \"great\" |\n| Mechs wear their faction's colours and bearings | Earned | \"would have to follow colors and bearings per faction\" |\n| The goal: functional and distinctive, yet cohesive | Earned | The owner's words |\n| Low bloom on glows, thruster look as it stands | Earned | \"too much bloom\", \"the thruster reel was sick\" |\n| Head findings in [heads.md](heads.md), section \"settled\" | Earned | Owner's reviews |\n| **The frame shows at the joints** | **Guessed** | Phantom Brigade study. Owner questioned exposed parts three times. Never asked |\n| One shared inner frame per size class | Guessed | Same study. Never chosen, and it's why the four atelier lines share limbs |\n| Joint covers (quilt, bellows, mail fauld) | Guessed | No verdict yet |\n| Large flat faces, one chamfer width | Guessed | Same study. Pulls against \"more segmentation\" |\n| Detail comes from decals, not geometry | Guessed | So maker plates have seams off, and no decals exist for them |\n| Main 70, accent 20, trim lines. Accent never bands the leg | Guessed | Owner: \"we don't decide about colors for now\" |\n| Two or three silhouette moves per maker | Guessed | Plausible, untested on its own |\n| A grade carries 1.5 to 2 times the plates | Guessed | The direction is earned, the number is mine |\n| Proportion targets (head .10, crotch .54, knee .39) | Guessed | Middle of GM and Jegan. The owner named Gundam 00 first, and 00 measures differently |\n| Reading distances of 2 km, 300 m, 50 m | Guessed | Never checked in play |\n| Standard mount pads at forearm, shoulder, pack, hip | Guessed | Fits the parts idea, never tested |\n| Four house lines and their names | Guessed | Owner liked all four. The framing and names are mine |\n| Current colours, including admin teal | Placeholder | Owner: \"keep what you already made\" for now |\n\n### How to work\n\n| Rule | Status | Evidence |\n|---|---|---|\n| Moving, in context, beats stills on grey | Earned, once | The thruster reel. Sample of one, keep testing |\n| Variety side by side beats refining one thing | Earned | Head rounds and the atelier board |\n| Research first, from games the owner knows | Earned | Thrusters. Proportions only helped once measured |\n| Polish rounds raise appeal | **Disproven** | Round 6: \"cleaner but is blander\" |\n| Two-view shells are enough | **Disproven** | Atelier round 1: flowerpots and tin. Learned once before, in the first days |\n| A critic finding no big issues means it's good | **Disproven** | \"No P1 is evident\", then \"bland? unappealing?\" |\n| Ratios and plate counts show progress | **Disproven** | \"I can't give you numbers I like\" |\n| The owner can pick from a menu | **Disproven** | \"I can't pick from the head set\", \"all A-D\" |\n| Code-built mechs can look great close up | Unproven | Fine at the chase camera so far. Could be the limit |\n\n### The code\n\nProven by shipping and tests, keep these: the sim stays pure, rendering only reads it, material roles are\nthe contract between kit and game, parts are rigid on the 22 joints, models come from `kit/` scripts\n(`CLAUDE.md`). Rebuilding everything after one change, with byte-identical checks, has been reliable.\n\n## Running things\n\n| To do | Run | Takes |\n|---|---|---|\n| Outline sheet | `npm run -s outlines -- <spec>` | about 1 s |\n| One block model | `npm run -s kit -- block_<name>` | about 3 s |\n| Quick clip: chase and turntable, 12 fps | `npm run -s clip -- <model> --quick` | about 35 s |\n| Full clip: chase, orbit, turn, bend, scale | `npm run -s clip -- <model>` | about 4 min |\n| Show the owner a file | `tools/show.sh <file>` | |\n| Log a step | `tools/log.sh \"note\"` | |\n| Dev server, game at `:5199` | `npm run dev` | |\n| Tests, types | `npm test`, `npm run typecheck` | under a minute when idle |\n| Rebuild one model | `npm run kit -- <name>` | seconds to minutes |\n| Rebuild a full lab | `npm run kit -- lab_pattern` | 6 to 8 min |\n| Still shots, dev server up | `node tools/shoot.mjs <dir> name=/?view=...` | |\n| Shots on C1 | `tools/c1-render.sh`, read its header first | |\n\nA salvage test has timed out when the machine was loaded. Check the load before calling it a failure, and\ndon't call it clean on reruns alone. Clips make their own server and stop it. Any model with the rig works, not lab lineups. Add `--look kit`\nor `--look field` to compare looks, since the owner never settled one and the default is `field`. Viewer and\nbench parameters are in the headers of `src/view/viewer.ts` and `src/view/thrusters.ts`.\n\n## Showing the owner options\n\n- **Rough and many first.** A sheet of outlines beats one polished model.\n- **One number per option,** so \"3 and 7 are dead\" is a complete answer.\n- **Moving, from the game camera, with a pilot for scale,** once there's anything in 3D.\n- **Before and after, side by side,** when something changed.\n- **Open it for them:** `tools/show.sh`. Images and clips go through `cw view`, and copies land in\n  `Downloads\\citadel-shots\\one-mech`. The terminal can't show images.\n- **Keep the words short.** One question at the end. No ratios. Longer write-ups go on a page.\n- You can't watch video. Judge from the clip's `.sheet.jpg`, and say so. The owner judges motion.\n\n## Reading the owner's feedback\n\nThey react precisely and rarely choose. Their strongest praise is short.\n\n| They say | It usually means | Do |\n|---|---|---|\n| \"X looks off\", \"weird\" | That part, trust it, even without a reason | Change it, or show that part three ways |\n| \"bland\", \"plastic\", \"toy\" | Surface, light or motion, more than outline | Try surface and context before reshaping |\n| \"not sure how to explain it\" | Real, and not a request for a question list | Show contrasts, ask what's closer |\n| \"I can't pick\", \"all of them\" | Wrong question, or they like the variety | Keep them all, or ask what's off instead |\n| \"go ahead\", \"you can proceed\" | Permission, not approval of the result | Keep showing |\n| \"sick\", \"that's done\" | The target feeling | Note what made it work |\n| Silence on something | Not a yes | Don't build rules on it |\n\nDon't turn one comment into a rule for every maker. For direction questions, interview first: one question\nat a time, play it back in their words, goals last.\n\n## Standing constraints\n\n- **The owner decides look, names and spending.** Open decisions: `led show 1005` and the \"Open\" sections in\n  [lore.md](lore.md) and [manufacturers.md](manufacturers.md). Don't repeat the list in every report.\n- **No paid services** without a fresh go, and put any cost in the first line. The retro lists tools, and\n  recommends nothing paid.\n- **Third-party art never gets committed or published.** References stay in `review/out/`. Phantom Brigade\n  material stays local, principles only. Use references to learn, never to copy a design.\n- **No headed browsers on the workstation.** They pop up on the owner's desktop. Headless here, or C1.\n- **C1:** `ssh severian@[redacted: a machine address on the home network]`, stay under `~/ci/sortie`, nice and ionice, cores 0 to 11, at most 3\n  browsers, ports 5800 to 5899, kill servers when done, short GPU runs, ask the lead before a burst.\n- **Design critic:** Codex only, `codex exec -m gpt-6-astra -c model_reasoning_effort=xhigh \"...\" < /dev/null`.\n  Always `< /dev/null`. Ask what the one idea is and what feels cheap. It never decides done.\n- **Prose in this repo** is terse: no em dashes, no semicolons. Commit and push freely, private repo.\n- **Keep [lore.md](lore.md), [manufacturers.md](manufacturers.md) and the [wiki](wiki.html) in step.**\n"}{"chunk_id":"f8f5c2","wall_time_seconds":0.000004481,"exit_code":0,"original_token_count":0,"output":""}
  • exec_commandcat ~/context/reflect/sortie.html

    command

    cat ~/context/reflect/sortie.html

    result, 7 lines

    Script completed
    Wall time 0.1 seconds
    Output:
    Warning: truncated output (original token count: 25281)
    Total output lines: 1
    
    {"chunk_id":"5d3069","wall_time_seconds":0.00000513,"exit_code":0,"original_token_count":49517,"output":"Warning: truncated output (original token count: 49517)\nTotal output lines: 428\n\n<!doctype html>\n<html lang=\"en\">\n<head>\n<meta charset=\"utf-8\">\n<meta name=\"viewport\" content=\"width=device-width, initial-scale=1, viewport-fit=cover\">\n<meta name=\"claude:viewport-fit-sides\" content=\"cover\">\n<title>Sortie Mech Reflection</title>\n<style>\n/* Layout: a mech design sheet. A title block on top, the answer next, then four labelled parts in one reading column. Comparison rows stack at phone width. Colours come from the game's own palettes: admin petrol teal, houses wine, a cool steel neutral. */\n@font-face{font-family:\"Chakra Petch\";font-weight:600;font-style:normal;font-display:swap;src:url([redacted: data-URI prefix][redacted: base64 data]) format(\"woff2\")}\n:root{\n  --bg:#eef1f3; --panel:#e1e6ea; --ink:#171b20; --muted:#4e5863; --line:#c5ccd4; --link:#14605d; --teal:#1b5c5a; --wine:#781c33;\n  --src:#d6ebe6; --src-ink:#14504d; --ex:#f1e6c9; --ex-ink:#6d5208; --gue:#efdbe0; --gue-ink:#6e1c31;\n  --display:\"Chakra Petch\",\"Segoe UI Variable Display\",\"Segoe UI\",system-ui,sans-serif;\n  --body:\"Segoe UI Variable Text\",\"Segoe UI\",system-ui,-apple-system,\"Helvetica Neue\",Arial,sans-serif;\n  --mono:ui-monospace,\"Cascadia Mono\",Consolas,Menlo,monospace;\n}\n@media (prefers-color-scheme:dark){:root:not([data-theme=\"light\"]){\n  --bg:#11151a; --panel:#1a2027; --ink:#e6eaee; --muted:#a0abb6; --line:#2e3640; --link:#6fcfc4; --teal:#6fcfc4; --wine:#e593a6;\n  --src:#173430; --src-ink:#8fdccf; --ex:#352c12; --ex-ink:#e6cd85; --gue:#3a1d26; --gue-ink:#efb0bf; color-scheme:dark}}\n:root[data-theme=\"dark\"]{\n  --bg:#11151a; --panel:#1a2027; --ink:#e6eaee; --muted:#a0abb6; --line:#2e3640; --link:#6fcfc4; --teal:#6fcfc4; --wine:#e593a6;\n  --src:#173430; --src-ink:#8fdccf; --ex:#352c12; --ex-ink:#e6cd85; --gue:#3a1d26; --gue-ink:#efb0bf; color-scheme:dark}\n*{box-sizing:border-box}\nhtml{-webkit-text-size-adjust:100%}\nbody{margin:0;background:var(--bg);color:var(--ink);font:400 1rem/1.55 var(--body)}\n.wrap{max-width:64rem;grid-template-columns:minmax(0,1fr);margin:0 auto;padding-block:2rem 4rem;padding-left:max(16px,env(safe-area-inset-left,0px));padding-right:max(16px,env(safe-area-inset-right,0px));display:grid;gap:2.5rem}\na{color:var(--link);text-underline-offset:2px}\na:focus-visible{outline:2px solid var(--teal);outline-offset:2px}\nimg{max-width:100%;height:auto;display:block}\nh1,h2,h3{font-family:var(--display);font-weight:600;line-height:1.2;text-wrap:balance;margin:0}\nh1{font-size:clamp(1.8rem,4.5vw,2.6rem)}\nh2{font-size:1.5rem}\nh3{font-size:1.08rem;color:var(--teal)}\np{margin:0}\n.prose{max-width:44rem;display:grid;gap:.7rem}\nul,ol{margin:0;padding-left:1.2rem;display:grid;gap:.4rem}\n.part{display:grid;grid-template-columns:minmax(0,1fr);gap:1.4rem;border-top:2px solid var(--ink);padding-top:1rem}\n.part>header{display:grid;gap:.2rem}\n.tag{font:500 .74rem/1.3 var(--mono);letter-spacing:.09em;text-transform:uppercase;color:var(--muted)}\n.sub{display:grid;grid-template-columns:minmax(0,1fr);gap:.75rem}\n\n/* title block */\n.block{display:grid;grid-template-columns:minmax(0,1fr);gap:1rem}\n.fields{display:grid;grid-template-columns:repeat(auto-fit,minmax(11rem,1fr));border:1px solid var(--line)}\n.fields div{padding:.5rem .7rem;border-right:1px solid var(--line);border-bottom:1px solid var(--line);margin:0 -1px -1px 0;min-width:0}\n.fields b{display:block;font:500 .7rem/1.3 var(--mono);letter-spacing:.08em;text-transform:uppercase;color:var(--muted)}\n.fields span{font-size:.92rem}\n\n/* answer */\n.answer{background:var(--panel);padding:1.1rem 1.2rem;display:grid;gap:.6rem}\n.answer h2{font-size:1.2rem}\n.start{border:2px solid var(--teal);padding:1.1rem 1.2rem;display:grid;grid-template-columns:minmax(0,1fr);gap:.6rem}\n.start h2{font-size:1.45rem}\n.start p,.start ol{max-width:50rem}\n.answer p{max-width:50rem}\n\n/* rows: label, body, consequence */\n.rows{display:grid;grid-template-columns:minmax(0,1fr);border-top:1px solid var(--line)}\n.row{display:grid;grid-template-columns:12rem minmax(0,1fr);gap:.3rem 1.2rem;padding:.8rem 0;border-bottom:1px solid var(--line)}\n.row>h3,.row>.lab{font-family:var(--display);font-weight:600;font-size:1rem;line-height:1.3;color:var(--ink);margin:0}\n.row .means{grid-column:2;color:var(--ink);font-size:.97rem;background:var(--panel);padding:.45rem .65rem}\n.row .means b{font:500 .7rem/1.3 var(--mono);letter-spacing:.08em;text-transform:uppercase;color:var(--teal);display:block}\n.row>div{min-width:0;display:grid;grid-template-columns:minmax(0,1fr);gap:.35rem}\n.row>.means{display:block}\n.row ul{font-size:.97rem}\n\n/* chips */\n.c{display:inline-block;font:500 .66rem/1.5 var(--mono);letter-spacing:.04em;padding:0 .35rem;vertical-align:.08em;white-space:nowrap}\n.c.s{background:var(--src);color:var(--src-ink)}\n.c.e{background:var(--ex);color:var(--ex-ink)}\n.c.g{background:var(--gue);color:var(--gue-ink)}\n\n/* feedback groups */\n.four{display:grid;grid-template-columns:repeat(auto-fit,minmax(13rem,1fr));gap:1.2rem 1.4rem}\n.four>div{display:grid;gap:.4rem;align-content:start;min-width:0}\n.four ul{font-size:.95rem;padding-left:1rem}\n.q{font-style:italic}\n\n/* figures */\nfigure{margin:0;display:grid;grid-template-columns:minmax(0,1fr);gap:.35rem}\nfigcaption{font-size:.86rem;color:var(--muted);max-width:50rem}\n.two{display:grid;grid-template-columns:repeat(auto-fit,minmax(min(18rem,100%),1fr));gap:1.2rem;align-items:start}\n.two>*{min-width:0}\n.figs{display:grid;grid-template-columns:minmax(0,1fr) minmax(0,1.7fr);gap:1.2rem;align-items:start}\n\n/* tables */\n.scroll{overflow-x:auto;border-top:1px solid var(--line)}\ntable{border-collapse:collapse;width:100%;font-size:.92rem}\nth{font:500 .7rem/1.3 var(--mono);letter-spacing:.07em;text-transform:uppercase;color:var(--muted);text-align:left;padding:.5rem .6rem .35rem 0;border-bottom:1px solid var(--line);vertical-align:bottom}\ntd{padding:.5rem .7rem .5rem 0;border-bottom:1px solid var(--line);vertical-align:top}\ntbody th{font:600 .92rem/1.35 var(--body);letter-spacing:0;text-transform:none;color:var(--ink);padding-top:.5rem;vertical-align:top}\n.band td{font:500 .72rem/1.3 var(--mono);letter-spacing:.08em;text-transform:uppercase;color:var(--wine);padding-top:1rem}\n.cost{font-variant-numeric:tabular-nums;font-weight:600;width:6rem}\n.tools{min-width:52rem}\n.week{min-width:40rem}\n.compare{min-width:34rem}\n.v-try{color:var(--teal);font-weight:600}\n.v-no{color:var(--muted)}\n.note{font-size:.88rem;color:var(--muted);max-width:50rem}\n@media (max-width:40rem){\n  .figs…15281 tokens truncated… labels. <a href=\"https://www.gunpla101.com/version-ka-101\">Hobby guide</a> <span class=\"c e\">excerpt</span></li>\n    </ul></div>\n      <div class=\"means\"><b>For sortie</b>The kit uses weighted normals nowhere. Maker plates have seams off (<code>kit/makers/shell.py:107</code>) because the rule said detail comes from decals, and the makers have no decals, only a service band and stencil blocks. Bare plates are my top suspect for \"plastic\" and \"bland\" <span class=\"c g\">guess</span>. This is the cheapest thing to test.</div></div>\n\n    <div class=\"row\"><h3>Joints and armour</h3><div><ul>\n      <li>Harness moves on overlapping lames with leather straps and sliding rivets, and mail covers the gaps. <a href=\"https://www.medievalcollectibles.com/glossary/terminology/armor-terminology/\">Glossary</a>, Met catalogue notes <span class=\"c e\">excerpt</span></li>\n      <li>One armourer notes that many visible rivets connect to nothing. <a href=\"https://www.ageofarmour.com/education/gorget.php\">Age of Armour</a> <span class=\"c e\">excerpt</span></li>\n      <li>The Geass rig was kept simple on purpose, and the hand rig was the hardest, since a robot hand can't close like a fist. <a href=\"https://cgworld.jp/feature/201903-codegeass1.html\">CGWORLD</a> <span class=\"c s\">source</span></li>\n      <li>Designs change once things move. Shinkawa revised details after seeing in-game movement, and Titanfall's designs were changed for the animators. <a href=\"https://blog.playstation.com/?p=207343\">PlayStation Blog</a>, <a href=\"https://www.cgchannel.com/2014/04/qa-titanfall-lead-artist-joel-emslie/\">CG Channel</a> <span class=\"c s\">source</span></li>\n      <li>Gunpla kits let plates slide and rotate with the joint so a pose still looks closed. I found no good source today. <span class=\"c g\">guess</span></li>\n    </ul></div>\n      <div class=\"means\"><b>For sortie</b>I can't find a range-of-motion sheet in any round. Add reach, crouch and boost poses before any detail. That would catch the shoulder bar problem early, and a mechanism only has to look believable.</div></div>\n\n    <div class=\"row\"><h3>Scale</h3><div><ul>\n      <li>Pacific Rim: <span class=\"q\">to communicate huge scale, you want to move things slowly</span>, with <span class=\"q\">little hits, and hard stops with mechanical recoil</span> in the small parts. <a href=\"https://artofvfx.com/?p=4715\">Art of VFX</a> <span class=\"c s\">source</span></li>\n      <li>An action figure beside the maquette made the Ogre read as 20 feet tall. <a href=\"https://www.cgchannel.com/2014/04/qa-titanfall-lead-artist-joel-emslie/\">CG Channel</a> <span class=\"c s\">source</span></li>\n      <li>The Origin sized ships with human models: <span class=\"q\">If the window is this big, then how many meters high is the bridge?</span> <a href=\"https://www.gundam-the-origin.net/en/special/interview09.html\">The Origin</a> <span class=\"c s\">source</span></li>\n    </ul></div>\n      <div class=\"means\"><b>For sortie</b>We already have pilots, a hangar and a catapult. Put a pilot in every hero shot, add hatches and rungs at human size, and let small parts snap while big masses drift.</div></div>\n\n    <div class=\"row\"><h3>Anime and real time</h3><div><ul>\n      <li>Lancelot siN: flat fills, line weight painted in large textures, limbs stretched to fit each layout. Some other cuts had about 10 staff retouching frames in Photoshop. <a href=\"https://cgworld.jp/feature/201903-codegeass1.html\">CGWORLD 1</a>, <a href=\"https://cgworld.jp/feature/201903-codegeass3.html\">3</a> <span class=\"c s\">source</span></li>\n      <li>A game can't cheat per shot, so the model must hold from every angle. Our chase camera mostly sees the back. <span class=\"c g\">guess</span></li>\n    </ul></div>\n      <div class=\"means\"><b>For sortie</b>Anime references flatter a front three-quarter drawing. Judge from the chase camera first. That may be why backs keep coming up.</div></div>\n  </div>\n\n  <div class=\"sub\">\n    <h3>Your four influences</h3>\n    <div class=\"scroll\"><table class=\"compare\">\n      <thead><tr><th></th><th>What I could source</th><th>My reading <span class=\"c g\">guess</span></th></tr></thead>\n      <tbody>\n        <tr><th>Gundam 00</th><td>Designs by Takayuki Yanase, Ebikawa and others (<a href=\"https://mechabay.com/gundam-00\">Mechabay</a>). Our measure: waist a tenth of height, legs near 60 percent, limbs of three or four overlapping plates over a dark inner frame. No designer interview found on why.</td><td>Long straight lines, sharp wedges and thin plates. The frame does show here, as a dark slim core. Ours reads as grey scaffolding.</td></tr>\n        <tr><th>Code Geass</th><td>On the Re;surrection film: Akira Yasuda's original designs, Eiji Nakata's main frames, Kenji Teraoka's cockpits (<a href=\"https://cgworld.jp/feature/201903-codegeass1.html\">CGWORLD</a>). Our measure in docs/light-frames.md.</td><td>The character is in the stance: a hunch, long arms, skating on wheels. Our sheets stand upright, which removes it.</td></tr>\n        <tr><th>Front Mission</th><td>Modular: body, two arms, legs (<a href=\"https://www.mobygames.com/game/13788/front-mission/?s=date\">MobyGames</a>). Shinkawa drew one new wanzer for Left Alive. No credits for the classic wanzers found.</td><td>Military vehicles first. Mismatched parts must still look right, so it's the best model for your parts system.</td></tr>\n        <tr><th>Lancer</th><td>A baseline maker plus four big makers (<a href=\"https://lancer.wiki.gg/wiki/The_Big_Four\">wiki</a>). No source on its art rules.</td><td>Useful for maker structure, as docs/lore.md already says. It teaches little about 3D form.</td></tr>\n      </tbody>\n    </table></div>\n  </div>\n\n  <div class=\"two\">\n    <div class=\"sub\">\n      <h3>Where code-built falls down</h3>\n      <ul>\n        <li><b>Differences you can't see.</b> Kate Compton's \"ten thousand bowls of oatmeal\": unique on paper, the same to the eye. <a href=\"https://gamedeveloper.com/design/practical-procedural-generation-for-everyone-\">Game Developer</a> <span class=\"c s\">source</span> Our heavies carry up to 73 plates, and Codex still says medium and heavy look alike at 2 km on two of the four lines.</li>\n        <li><b>Seeing.</b> Vision models struggle with Blender tasks that are <span class=\"q\">relatively easy for human Blender users</span>. <a href=\"https://arxiv.org/abs/2504.01786v1\">BlenderGym</a> <span class=\"c e\">excerpt</span> A similar system's stated limit is missing subtle spatial artifacts. <a href=\"https://api.emergentmind.com/papers/2508.08228\">LL3M</a> <span class=\"c e\">excerpt</span></li>\n        <li><b>The vocabulary shows through.</b> Two-view shells look like slabs, lofts look like tubes. Each kit has a look. <span class=\"c g\">guess, from our rounds</span></li>\n        <li><b>No hand.</b> An artist pushes a form until it's right. I change a number and wait minutes.</li>\n        <li><b>Thin geometry breaks.</b> Coplanar slivers on the light shins, a pinhole on the knee wings.</li>\n      </ul>\n    </div>\n    <div class=\"sub\">\n      <h3>Where it wins, so keep it</h3>\n      <ul>\n        <li>The joint contract, mount pads, material roles and livery swaps.</li>\n        <li>Rebuilding every model after one frame change, with byte-identical proof.</li>\n        <li>Effects and shaders, as the reel showed.</li>\n        <li>Variation you then curate. Compton kept the Spore planet seeds that looked good and threw out the rest: <span class=\"q\">You don't need to 'do it right', just do what works.</span> <a href=\"https://gamedeveloper.com/design/practical-procedural-generation-for-everyone-\">Game Developer</a> <span class=\"c s\">source</span></li>\n      </ul>\n    </div>\n  </div>\n</section>\n\n<section class=\"part\" id=\"tools\">\n  <header><div class=\"tag\">Part 3</div><h2>Tools</h2></header>\n  <div class=\"prose\"><p><b>You need to buy nothing.</b> The first four rows need no install and are what I'd use. The well-known Blender hard-surface add-ons I could check are built for a person with a mouse, so they barely help a scripted pipeline.</p></div>\n  <div class=\"scroll\"><table class=\"tools\">\n    <thead><tr><th>Cost</th><th>Tool</th><th>What it does</th><th>Fit here</th><th>Licence, signs of life</th><th>Verdict</th></tr></thead>\n    <tbody>\n      <tr class=\"band\"><td colspan=\"6\">Free, already here</td></tr>\n      <tr><td class=\"cost\">$0</td><th><a href=\"https://docs.blender.org/manual/en/5.1/modeling/modifiers/normals/weighted_normal.html\">Weighted Normal modifier</a> with Bevel</th><td>Keeps flat faces flat and puts the highlight on the bevel</td><td>Scriptable headless. Unused in the kit today.</td><td>GPL, in Blender. Manual current at 5.x.</td><td class=\"v-try\">Try first</td></tr>\n      <tr><td class=\"cost\">$0</td><th><code>lib.subd</code> with creases</th><td>Domes and swelling plates with crisp edges</td><td>Ours, already built. Your approved heads came from this line of work.</td><td>In the repo</td><td class=\"v-try\">Default for hero plates</td></tr>\n      <tr><td class=\"cost\">$0</td><th><a href=\"https://docs.blender.org/manual/en/latest/modeling/geometry_nodes/volume/index.html\">SDF grid nodes</a></th><td>Mesh to SDF, boolean, offset and fillet, then back to mesh: forms that blend into each other</td><td>Listed for 5.2 LTS, we run 5.2.3. Graph built from Python. Thin plates may need fine grids <span class=\"c g\">guess</span>.</td><td>GPL, in Blender. I read the node list only.</td><td class=\"v-try\">One-day test on a shoulder</td></tr>\n      <tr><td class=\"cost\">$0</td><th>SVG silhouette sheets</th><td>Dozens of rough outlines in minutes</td><td>No install. The kit already takes 2D outlines.</td><td>Ours to write</td><td class=\"v-try\">Use for the sketch step</td></tr>\n      <tr class=\"band\"><td colspan=\"6\">Free, to install</td></tr>\n      <tr><td class=\"cost\">$0</td><th><a href=\"https://github.com/gumyr/build123d\">build123d</a></th><td>Python CAD with true fillets and chamfers</td><td>Scriptable, but outside Blender. The page lists STL and STEP export. Complex fillets may fail <span class=\"c g\">guess</span>.</td><td>Apache 2.0. 3.6k stars, 3,454 commits.</td><td>Only if bevels fall short</td></tr>\n      <tr><td class=\"cost\">$0</td><th><a href=\"https://github.com/ahujasid/blender-mcp\">BlenderMCP</a></th><td>Lets an agent drive a live Blender and see viewport shots</td><td>Could shorten the loop. Adds a live session, and telemetry to switch off.</td><td>MIT. About 12k stars per one listing.</td><td class=\"v-no\">Skip for now</td></tr>\n      <tr><td class=\"cost\">$0</td><th><a href=\"https://extensions.blender.org/add-ons/hardflow/\">Hardflow</a></th><td>Draw-to-cut booleans, snapping, decals, kitbash</td><td>Mouse-driven. No scripting described.</td><td>GPLv3+. 1.22.0, Oct 2026 note, 7,504 downloads.</td><td class=\"v-no\">Only if you want to model by hand</td></tr>\n      <tr><td class=\"cost\">$0</td><th><a href=\"https://www.renderjuice.com/blender/extensions/bool-tool\">Bool Tool</a></th><td>Boolean shortcuts</td><td>The kit already calls booleans.</td><td>GPL. About 1M downloads.</td><td class=\"v-no\">Skip</td></tr>\n      <tr><td class=\"cost\">$0</td><th><a href=\"https://github.com/hlorus/CAD_Sketcher\">CAD Sketcher</a></th><td>Constraint-based 2D sketches</td><td>Interactive. Calls itself experimental.</td><td>GPL-3.0. Needs Blender 5.0+.</td><td class=\"v-no\">Skip</td></tr>\n      <tr><td class=\"cost\">$0</td><th><a href=\"https://github.com/microsoft/TRELLIS.2\">TRELLIS.2</a></th><td>One image to a textured 3D mesh</td><td>Needs an NVIDIA GPU with 24 GB, and C1 has a Radeon. A fused mesh won't ride our joints <span class=\"c g\">guess</span>.</td><td>MIT.</td><td class=\"v-no\">Can't run here</td></tr>\n      <tr><td class=\"cost\">$0 if it runs</td><th>Local image model on C1</th><td>Rough concept sheets for you to react to, private, nothing ships</td><td>Untested on C1, and the GPU is shared. I didn't research models today.</td><td>Depends on the model</td><td>Your call, question 5 below</td></tr>\n      <tr class=\"band\"><td colspan=\"6\">Paid, none recommended</td></tr>\n      <tr><td class=\"cost\">$38</td><th><a href=\"https://superhivemarket.com/products/hard-ops--boxcutter-ultimate-bundle\">Hard Ops / Boxcutter</a></th><td>The standard hard-surface toolset</td><td>Menus, pies and clicks. No scripting mentioned.</td><td>GPL. Blender 3.5 to 5.2.</td><td class=\"v-no\">Only for a human modeller</td></tr>\n      <tr><td class=\"cost\">$44.99</td><th><a href=\"https://superhivemarket.com/products/meshmachine\">MESHmachine</a></th><td>Chamfer and fillet tools, normal detail plugs</td><td>Page doesn't say how it's driven.</td><td>GPL. Blender 4.2 to 5.2.</td><td class=\"v-no\">No</td></tr>\n      <tr><td class=\"cost\">From $54.99</td><th><a href=\"https://superhivemarket.com/products/decalmachine\">DECALmachine</a></th><td>Mesh decals in Blender</td><td>The game already places decals at runtime.</td><td>GPL. Blender 4.3 to 5.1.</td><td class=\"v-no\">No</td></tr>\n      <tr><td class=\"cost\">$20 to about $70</td><th><a href=\"https://blendermarket.com/products/fluent/ratings?page=5\">Fluent</a></th><td>Interactive boolean modelling</td><td>Same problem as Hard Ops <span class=\"c g\">guess</span>. Listings conflict on price.</td><td>Not checked</td><td class=\"v-no\">No</td></tr>\n      <tr><td class=\"cost\">$29, or $69 commercial</td><th><a href=\"https://superhivemarket.com/products/robot-designer\">Robot Designer</a></th><td>Geometry Nodes robot generator, 800+ sliders, 25 presets</td><td>Someone else's look, with a real oatmeal risk.</td><td>Terms mix GPL with resale limits. Published about 7 months ago, Blender 4.5 to 5.2.</td><td class=\"v-no\">No</td></tr>\n      <tr><td class=\"cost\">Not priced</td><th>A concept artist</th><td>One real four-view sheet to model against</td><td>It's the step every source has and we lack.</td><td></td><td>Only if the free week fails, and only on your go</td></tr>\n    </tbody>\n  </table></div>\n</section>\n\n<section class=\"part\" id=\"forward\">\n  <header><div class=\"tag\">Part 4</div><h2>When you start this again</h2></header>\n\n  <div class=\"two\">\n    <div class=\"sub\">\n      <h3>What changes</h3>\n      <ul>\n        <li><b>Design and building become separate steps.</b> Rough and many first, then model one.</li>\n        <li><b>One hero at a time, judged moving.</b> Clips are the review. Stills are for diagnosis.</li>\n        <li><b>You veto, you don't choose.</b> I bring variety and ask \"what's off?\"</li>\n        <li><b>Surface gets equal time with outline.</b> Weighted normals, sculpted plates, designed seams, real decals, one settled look.</li>\n        <li><b>Rules come last.</b> <code>docs/visual-language.md</code> goes back to draft until something lands.</li>\n        <li><b>The critic changes job.</b> I ask Codex \"what's the one idea here, and what feels cheap?\" It never decides done. Your friends see clips too, since they gave useful reads early on.</li>\n      </ul>\n    </div>\n    <div class=\"sub\">\n      <h3>What I'd ask you, one at a time</h3>\n      <ol>\n        <li>Think of a moment in any game or show when a mech made you feel \"that's it\". What was it doing?</li>\n        <li>In your ideal first hour of sortie, where's your mech on screen when it matters most?</li>\n        <li>You've asked about the inner frame three times. Do you want to see it at all, and where?</li>\n        <li>Is Gundam 00 still the target for the plain line mech, or did the knight board move that?</li>\n        <li>Are rough private AI sketches fine for this, and do you already have an image tool you like?</li>\n      </ol>\n      <p class=\"note\">No boards or menus during this. I'd play your answers back in your words and keep one short settled / open note. The clean-or-worn question gets answered by the reel, so I wouldn't ask it.</p>\n    </div>\n  </div>\n\n  <div class=\"sub\">\n    <h3>The first week</h3>\n    <div class=\"scroll\"><table class=\"week\">\n      <thead><tr><th>Day</th><th>What happens</th><th>Your part</th><th>Done when</th></tr></thead>\n      <tbody>\n        <tr><th>1. Interview</th><td>The questions above. No building.</td><td>About 30 minutes</td><td>A short settled / open note</td></tr>\n        <tr><th>2. Reel</th><td>No modelling. A 10-second clip of each existing mech (striker, pattern, all four knight lines): turn, boost, stop, plus one shot next to a pilot. Same clip in clean and worn looks. Makers need a small hook-up to the thruster bench.</td><td>Watch, say what's off, send it to friends if you like</td><td>The look is settled, and we know which mech already has some pull</td></tr>\n        <tr><th>3. Sketches</th><td>30 or so rough silhouettes in SVG, pushed hard, none averaged from references. Rough AI sheets too if you said yes.</td><td>Cross out what's dead</td><td>Three survivors, no scoring</td></tr>\n        <tr><th>4. Surface test</th><td>One chest and shoulder from a survivor, built with weighted normals, sculpted plates, the SDF fillet, drawn seams and decals. Shown at the chase camera and close up, moving.</td><td>Say which looks expensive and which looks plastic</td><td>One method chosen, the rest dropped</td></tr>\n        <tr><th>5. Hero</th><td>One mech at one grade, with a reach, crouch and boost check before detail. Shown only as a reel.</td><td>React as you did to the thrusters</td><td>You say \"that's done\" about some of it, or name what's off</td></tr>\n      </tbody>\n    </table></div>\n    <p class=\"note\">You've since chosen to start over with one mech from scratch, so the plan at the top replaces this week. The test for the week was one part of one mech reaching the thruster reel's feeling. Grades, other makers and rules wait until then. The open shared shoulder bar question waits on your answer to question 3.</p>\n  </div>\n\n  <div class=\"two\">\n    <div class=\"sub\">\n      <h3>What I'd stop doing</h3>\n      <ul>\n        <li>Numbered polish rounds on one model</li>\n        <li>Building twelve models before one is loved</li>\n        <li>Turning one comment into a rule for every maker</li>\n        <li>Putting ratios and plate counts in your reports</li>\n        <li>Grey sheets of small tiles as the verdict</li>\n        <li>Asking you to pick, or carrying a rolling list of naming decisions in every report</li>\n        <li>Filling spare usage with systems while the look is open</li>\n        <li>Long terminal reports: one page, clip first, one question at the end</li>\n      </ul>\n    </div>\n    <div class=\"sub\">\n      <h3>Risks I can't remove</h3>\n      <ul>\n        <li><b>Code-built may cap out.</b> My guess is that it holds at the chase camera and falls short in hangar close-ups. If the week's hero doesn't land, I'll say so plainly. The next step would cost money: one commissioned four-view sheet, not priced, on your go only.</li>\n        <li><b>Sketches can promise more than the kit can build.</b> So day 4 comes before day 5, and only proven methods go on the hero.</li>\n        <li><b>I can't watch video.</b> I judge frames. You and your friends are the only judges of motion.</li>\n      </ul>\n    </div>\n  </div>\n\n  <p class=\"note\"><b>What I couldn't read.</b> Paradox forum (HBS diaries), Polycount, the Met and Age of Armour blocked direct fetches, so those are search excerpts. Kate Compton's original zine returned 403, so I used Game Developer's summary. CGWORLD and Tamashii Web were machine translated. I found no designer's own account of Gundam 00's shape choices, no source on Armored Core 6's or Lancer's art rules, and nothing usable on Gunpla inner frames. Prices are from listing pages today and can change.</p>\n</section>\n\n</div>\n</body>\n</html>\n"}