Subagent: Two-handed longsword clips

<teammate-message teammate_id="team-lead" summary="Body and sword changed: refit clips"> The character changed under your clips, so please refit them to the new body and sword. Your current keys look good in the sheet (real guard, lunge) but on the new body the build reports the left upper arm 7.8 cm inside the body at attack 0.38 to 0.47 s and the right upper arm up to 7.4 cm inside from 0.62 to 1.25 s.

What changed (all picked up by env WHOLE_DESIGN=wren, which you must now set on every build):

  • Body proportions (tools/whole/wren.py SHAPE): head 1.06, eye 1.08, neck 0.93, arms 0.95, hands 0.92, torso 0.97, leg_len 1.05. If you cache a shape (your shape.pkl), rebuild it from CAST["wren"]["shape"] after importing tools/whole/build.py with WHOLE_DESIGN=wren.
  • The sword is now wren.longsword: blade 0.93 m, grip 0.25 m, the right fist's middle 5.2 cm under the guard, the grip running 19.8 cm on behind it toward the pommel. Same hand frame as outfit.sword (blade along the hand's -z, fist offset (0.045, -0.018, 0)). Left fist 9.5 to 10 cm down the grip is right.
  • Costume: no long skirt any more. A cuirass ending at the natural waist, a leather waist, tassets on the thighs, a pauldron on the LEFT shoulder standing about 4 cm off it with four lames reaching about 11 cm down the arm, a small spaulder on the right, vambraces, and a high ponytail falling behind the head to mid-back (keep the blade and the fists out of it in the wind-up: the tail hangs 5 to 8 cm behind the skull).

Fast builds: WREN_QUICK=1 WHOLE_DESIGN=wren WHOLE_STANCE=longsword LOWMARCH_SITE=whole-clips OMP_NUM_THREADS=1 OPENBLAS_NUM_THREADS=1 python3 tools/whole/build.py arm_f_slender wren builds her with under layers, hair and sword only (no armour) in about 10 to 15 s plus your solve. Use that to iterate, then one full build (drop WREN_QUICK) to check the pauldron and tassets in the pictures. build.py no longer runs the slow plate-through-garment check by default (it cost 9 minutes with your per-sample keys); the arm-through-body lines still print.

Also needed: fewer keys in the exported clips if you can do it without losing the arcs (for example solve every sample but export every second one), since each key is checked and exported. Target: both clips report no arm through the body on the new body. Send your report when that holds; I'm using your file as it stands in the meantime. </teammate-message>

<teammate-message teammate_id="team-lead" summary="Clips adopted: report now"> Your clips2h.py as of 19:56 builds clean on the final character (34 keys, no arm through the body) and I've adopted it for the final stills and clip. Please stop editing that file now, so the committed version matches what was rendered, and send your report: what the guard and attack are, the check lines from your last build, what still looks wrong in the pictures, and wall time per step. </teammate-message>
  • Bashstat -c '%y' tools/whole/clips2h.py && md5sum tools/whole/clips2h.py && date

    command

    stat -c '%y' tools/whole/clips2h.py && md5sum tools/whole/clips2h.py && date

    result, 3 lines

    2026-11-10 20:15:33.293139517 +0800
    ebac3749c6c9d7153a904c53c7a3deae  tools/whole/clips2h.py
    Tue Nov 10 08:18:45 PM PST 2026
  • SendMessage{"to":"team-lead","summary":"Report; file on disk differs from your 19:56 render","message":"Report on clips2h.py. The file on disk is NOT the version you rendered: it was last written at 20:15 (+0800, md5 ebac3749) and exports 36 keys, while your 19:56 copy had 34. I have stopped editing. I also never did the refit to the new body: your refit message arrived after I'd finished, and your later message told me to stop. All my numbers below are from the OLD body (no WHOLE_DESIGN), so they don't describe the new character.\n\nWhat changed after the 34-key version (from memory; I don't know which landed before 19:56, and there's no git copy to diff against):\n- Follow-through keys 0.68 / 0.90 / 1.05 retuned so the blade trails down and back and stays 6.5 cm or more from the legs.\n- 0.20 key: blade tilted back (0.38, 0.82, 0.42).\n- Lunge hips shift 0.17 -> 0.13 and lead step 0.30 -> 0.26, so both feet stay planted.\n- Feet placed at rest-pose height and flat (BATTLE_IDLE's own feet float about 1 cm, lead toe up).\n- STEADY: lead and lift ease to a stop at every key, so the foot no longer overshoots and slides.\n- Samples 1.1 and 1.2 added, which is where 34 -> 36 came from.\n- Docstrings and line wrapping.\nIf you want the committed file to match your render, commit the copy you rendered from. Or rebuild from the current file with WHOLE_DESIGN=wren: I haven't run it on the new body.\n\nGuard: middle guard. Left foot forward, rear foot turned out 35°, hips bladed 22° and sunk 6 cm. Fists before the navel, elbows about 110°, point at the enemy's face. 2 s breathing loop (clips.sink), blade dipping a beat behind the body.\n\nAttack:\n- 0.20: blade rising on the right.\n- 0.32 to 0.44: wound over the right shoulder, weight back, shoulders turned 42° (0.44 is held).\n- 0.50: blade over the head, lead foot lifted 6 cm.\n- 0.56: cut on a 26 cm lunge, hips and shoulders turning the right side through (+10/+22°), lean 14°.\n- 0.62: blade down.\n- 0.68: overshoot low left, shoulders +34°, blade trailing down and back.\n- 0.90 to 1.05: settle.\n- 1.15 to 1.5: step back up into the guard.\n\nHow it's solved: every sample is solved from Hermite-eased key numbers. Feet by leg IK, clavicle swing toward the reach, right grip, left fist 9.5 cm down the actual sword axis, elbow directions retried until the arm clears checks.Body, head turned back level to the enemy. The right fist sits in the chest's frame: in model space the upper arms went through the bust whenever the shoulders turned.\n\nCheck lines from my last build (old body, 36 keys):\n wren: 36 keyframes checked; no arm through the body, 25 with an arm resting on it\nAll 25 \"resting\" hits are 0.3 to 1.4 cm outside the skin; the tightest is the right upper arm at 0.3 cm (attack 0.90 / 1.05).\n\nMargins from my own checks on the old body:\n- Fist spacing 9.5 cm in every sample, left fist at most 0.2 cm off the axis.\n- Blade and fists at least 13.9 cm from the posed head and hair (tightest at 0.44).\n- Assumed pauldron zone (left shoulder skin pushed out 5.5 cm): at least 14.5 cm.\n- Blade at least 6.5 cm from the legs, tip never below 0.14 m.\n- Rear foot fixed exactly; rear-hip extension at most 15°; elbows 50 to 118°.\n- None of this is checked against your new pauldron (about 4 cm off, lames 11 cm down the arm) or the high ponytail.\n\nWhat still looks wrong in the pictures:\n1. The follow-through (0.68 to 1.05) doesn't read. Her left side faces away from the camera, so the blade sits behind her body: \"hands at the belt holding a hilt\". Losing the skirt helps; even with the dress hidden, only part of the blade showed behind her thigh.\n2. In a still, the 0.56 cut reads more like a thrust (arms out, blade forward and down). The diagonal is only in the motion from 0.50 to 0.62.\n3. At 0.20 the rising blade crosses her face on screen, although it's about 20 cm in front of it.\n4. Her arms rest on the chest in most keys; on the new body a cuirass will meet the arm plates. The bust was the limiting shape.\n5. The guard is compact (hands close to the chest). In the default placement her point overlaps the soldier's halberd head.\n6. Checks run only at the solved samples. Hair is spring-simulated and the export blends linearly between keys.\n\nWall time:\n- Solve: about 500 s for the first version; 7 to 20 s after the speed-up (FastRig, which checks itself against the slow path and falls back on any mismatch, plus a 4-round grip).\n- Build: 39 to 163 s per run on a loaded machine (last one 94 s wall, 62 s user), 9 builds.\n- Shoot: about 6 s a frame alone; 8 frames in parallel 20 to 40 s, about 12 rounds.\n- Most of the session went into key tuning with headless scripts.\n\nNot done from your refit message, for whoever picks it up:\n- the refit to the new body and sword;\n- exporting every second sample;\n- the new pauldron and ponytail checks.\n\nChanges I'd suggest elsewhere (not made):\n- shoot.py: a relative WHOLE_SITE fails silently because the scratch copy's symlinks break. Wrapping it in os.path.abspath fixes it.\n\nFiles:\n- ~/Projects/lowmarch-fable52/tools/whole/clips2h.py\n- ~/Projects/lowmarch-fable52/review/out/whole-clips/shots/sheet.png\n- ~/Projects/lowmarch-fable52/review/out/whole-clips/shots/sheet_no_dress.png\n- scratch tools under review/out/whole-clips/ (git-ignored)","type":"message","recipient":"team-lead","recipient_kind":"name","content":"Report on clips2h.py. The file on disk is NOT the…"}

    input

    {
      "to": "team-lead",
      "summary": "Report; file on disk differs from your 19:56 render",
      "message": "Report on clips2h.py. The file on disk is NOT the version you rendered: it was last written at 20:15 (+0800, md5 ebac3749) and exports 36 keys, while your 19:56 copy had 34. I have stopped editing. I also never did the refit to the new body: your refit message arrived after I'd finished, and your later message told me to stop. All my numbers below are from the OLD body (no WHOLE_DESIGN), so they don't describe the new character.\n\nWhat changed after the 34-key version (from memory; I don't know which landed before 19:56, and there's no git copy to diff against):\n- Follow-through keys 0.68 / 0.90 / 1.05 retuned so the blade trails down and back and stays 6.5 cm or more from the legs.\n- 0.20 key: blade tilted back (0.38, 0.82, 0.42).\n- Lunge hips shift 0.17 -> 0.13 and lead step 0.30 -> 0.26, so both feet stay planted.\n- Feet placed at rest-pose height and flat (BATTLE_IDLE's own feet float about 1 cm, lead toe up).\n- STEADY: lead and lift ease to a stop at every key, so the foot no longer overshoots and slides.\n- Samples 1.1 and 1.2 added, which is where 34 -> 36 came from.\n- Docstrings and line wrapping.\nIf you want the committed file to match your render, commit the copy you rendered from. Or rebuild from the current file with WHOLE_DESIGN=wren: I haven't run it on the new body.\n\nGuard: middle guard. Left foot forward, rear foot turned out 35°, hips bladed 22° and sunk 6 cm. Fists before the navel, elbows about 110°, point at the enemy's face. 2 s breathing loop (clips.sink), blade dipping a beat behind the body.\n\nAttack:\n- 0.20: blade rising on the right.\n- 0.32 to 0.44: wound over the right shoulder, weight back, shoulders turned 42° (0.44 is held).\n- 0.50: blade over the head, lead foot lifted 6 cm.\n- 0.56: cut on a 26 cm lunge, hips and shoulders turning the right side through (+10/+22°), lean 14°.\n- 0.62: blade down.\n- 0.68: overshoot low left, shoulders +34°, blade trailing down and back.\n- 0.90 to 1.05: settle.\n- 1.15 to 1.5: step back up into the guard.\n\nHow it's solved: every sample is solved from Hermite-eased key numbers. Feet by leg IK, clavicle swing toward the reach, right grip, left fist 9.5 cm down the actual sword axis, elbow directions retried until the arm clears checks.Body, head turned back level to the enemy. The right fist sits in the chest's frame: in model space the upper arms went through the bust whenever the shoulders turned.\n\nCheck lines from my last build (old body, 36 keys):\n  wren: 36 keyframes checked; no arm through the body, 25 with an arm resting on it\nAll 25 \"resting\" hits are 0.3 to 1.4 cm outside the skin; the tightest is the right upper arm at 0.3 cm (attack 0.90 / 1.05).\n\nMargins from my own checks on the old body:\n- Fist spacing 9.5 cm in every sample, left fist at most 0.2 cm off the axis.\n- Blade and fists at least 13.9 cm from the posed head and hair (tightest at 0.44).\n- Assumed pauldron zone (left shoulder skin pushed out 5.5 cm): at least 14.5 cm.\n- Blade at least 6.5 cm from the legs, tip never below 0.14 m.\n- Rear foot fixed exactly; rear-hip extension at most 15°; elbows 50 to 118°.\n- None of this is checked against your new pauldron (about 4 cm off, lames 11 cm down the arm) or the high ponytail.\n\nWhat still looks wrong in the pictures:\n1. The follow-through (0.68 to 1.05) doesn't read. Her left side faces away from the camera, so the blade sits behind her body: \"hands at the belt holding a hilt\". Losing the skirt helps; even with the dress hidden, only part of the blade showed behind her thigh.\n2. In a still, the 0.56 cut reads more like a thrust (arms out, blade forward and down). The diagonal is only in the motion from 0.50 to 0.62.\n3. At 0.20 the rising blade crosses her face on screen, although it's about 20 cm in front of it.\n4. Her arms rest on the chest in most keys; on the new body a cuirass will meet the arm plates. The bust was the limiting shape.\n5. The guard is compact (hands close to the chest). In the default placement her point overlaps the soldier's halberd head.\n6. Checks run only at the solved samples. Hair is spring-simulated and the export blends linearly between keys.\n\nWall time:\n- Solve: about 500 s for the first version; 7 to 20 s after the speed-up (FastRig, which checks itself against the slow path and falls back on any mismatch, plus a 4-round grip).\n- Build: 39 to 163 s per run on a loaded machine (last one 94 s wall, 62 s user), 9 builds.\n- Shoot: about 6 s a frame alone; 8 frames in parallel 20 to 40 s, about 12 rounds.\n- Most of the session went into key tuning with headless scripts.\n\nNot done from your refit message, for whoever picks it up:\n- the refit to the new body and sword;\n- exporting every second sample;\n- the new pauldron and ponytail checks.\n\nChanges I'd suggest elsewhere (not made):\n- shoot.py: a relative WHOLE_SITE fails silently because the scratch copy's symlinks break. Wrapping it in os.path.abspath fixes it.\n\nFiles:\n- ~/Projects/lowmarch-fable52/tools/whole/clips2h.py\n- ~/Projects/lowmarch-fable52/review/out/whole-clips/shots/sheet.png\n- ~/Projects/lowmarch-fable52/review/out/whole-clips/shots/sheet_no_dress.png\n- scratch tools under review/out/whole-clips/ (git-ignored)",
      "type": "message",
      "recipient": "team-lead",
      "recipient_kind": "name",
      "content": "Report on clips2h.py. The file on disk is NOT the…"
    }

    result, 1 line

    {"success":true,"message":"Message sent to team-lead's inbox","msg_id":"14a03df8-aff3-4462-a32a-ed5ecdfc3144","routing":{"sender":"clips2h","senderColor":"blue","target":"@team-lead","targetColor":"red","summary":"Report; file on disk differs from your 19:56 render","content":"Report on clips2h.py. The file on disk is NOT the…"}}

I've sent the report to the lead. I'm no longer editing clips2h.py, but the copy on disk isn't the one they rendered and adopted. It was last written at 20:15 (+0800) and exports 36 keys; their 19:56 version had 34. Since 34 keys, I retuned the follow-through keys, fixed the feet so they stay planted at rest height, added samples at 1.1 and 1.2 s, and tidied the docstrings. I told them to commit the copy they rendered from, or rebuild from the current file.

I didn't refit the clips to the new body and sword. That request arrived after I'd finished, and their next message told me to stop. So every check number I sent comes from the old body. On the old body, the last build reported no arm through the body for either clip, with the closest arm 0.3 cm outside the skin. None of it has been checked against the new pauldron or the high ponytail.

What still looks wrong:

  • Follow-through: her left side faces away from the battle camera, so the blade is mostly hidden behind her body.
  • Cut at 0.56 s: reads more like a thrust in a still frame.
  • 0.20 s: the rising blade crosses her face on screen.
  • Arms: they rest on the chest in most keys.

Not done, and listed in the report for whoever picks it up: the refit to the new body, exporting every second sample, and the pauldron and ponytail checks. I also suggested shoot.py take an absolute site path, since a relative WHOLE_SITE fails without saying why.

Files are in ~/Projects/lowmarch-fable52:

  • tools/whole/clips2h.py
  • review/out/whole-clips/shots/sheet.png
  • review/out/whole-clips/shots/sheet_no_dress.png