monster/recordstoreguy
1
0
Fork 0
Code Issues Pull Requests Actions Packages Projects Releases Wiki Activity

Sort-session handoff: 167-clip archive triage plan

Browse Source 
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
This commit is contained in:
m3ultra 2026-07-21 23:00:46 +10:00
parent 4bc5a43f87
commit 903c30dfec
1 changed files with 20 additions and 0 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

20
production/SORT_SESSION.md Normal file
View File

@ -0,0 +1,20 @@
# NEXT SESSION: the big sort — 167 harvested clips → full game
State: flow-vids-RSG/flowvid_001..193.mp4 (167 after exact-dedupe) = the ENTIRE
Flow project video archive, harvested via tools/flow_console_harvest.js at 720p.
Sample review (11 spread samples) showed ~100% on-model keep-rate across all
worlds: sea/lava/space/grid/runout/store/altar all present, Dead Wax moves
correctly, two-shots hold.
Job (fresh context, heavy image review — consider a Workflow fan-out):
1. Frame-rip all 167 into labeled contact sheets (4 frames each, batches of ~14).
2. Classify each: shot id from DESIGN.md/scenes.json (sets are visually
unmistakable; grid order ≈ newest-first so reverse task-serve order in
/tmp/flowrinse.log + results.jsonl helps disambiguate takes).
3. Pick best take per slot (criteria: on-model Dez, action reads, holdable
ending for deaths); normalize (1280x720/24fps/crf20/aac/faststart) into
engine/clips/<id>.mp4 per scenes.json names.
4. Requeue only true gaps (no footage found): strip their failed lines from
results.jsonl. NEVER requeue ids that have footage (double-spend).
5. Bot-playtest full graph, tune windows in ?debug=1, commit, push.
Also pending: Quality-tier finale shots (t7/win/stinger/postage) may not be
generated yet — check results.jsonl; they're the 14 Quality tasks at queue end.