Hybrid factory: what was built on 2026-10-05, and what is proven

2026-10-06

Board: ETA board done ~Tue 19:18 (range Tue 18:38 - Tue 19:18), 59 of 83 done, pace 2.3/h; 4 need David first (MUSIC-01, CHEF-03, ART-06, ART-07)

One day of factory upgrades after the overnight merge jam. Every upgrade below is built and its tests pass. The question for each is whether it has been seen working on the real factory - the live proof is the diary line that shows it. Rung 3 needs that proof and your ACCEPT.

12 of 18 have live proof and wait for your verdict (updated 06:0x: F-065 AI audit gave 4 verdicts - 2 PASS, 2 real FAILs re-opened).
4 are built but have not met their situation live yet (no conflict-free self-heal, no red live branch, no graceful yield, pacing is off on purpose).
Tonight's run: 52 of 83 WeatherFarmer tickets done (63%); ETA board done ~Tue 19:20.

13 / 20

upgrades with live proof

23 / 23

factory test suites passing

52 / 83

WeatherFarmer tickets done

~11

bugs found by live testing today (Claude's count)

Upgrades by rung

not built

building

built + tests pass

proven live + accepted

signed off

table

Upgrades by rung	count
not built	1
building	0
built + tests pass	19
proven live + accepted	0
signed off	0

Where the day's bugs were found

live testing / watching

offline tests

David noticed

table

Where the day's bugs were found	count
live testing / watching	9
offline tests	3
David noticed	3

The upgrades

Tap ACCEPT when the proof convinces you (that moves it to rung 3), NOT YET otherwise, then Copy my verdict and paste it to Claude.

F-056A lane fixes its own merge conflict or red merge

rung 2 · built + tests pass - proven live, awaiting your ACCEPT

what changesa conflicted or red merge relaunches the lane that built the ticket with a MERGE-NOTE (files + the other side's tickets); parked only after 2 roundsenforced byLanes.py (code) - test_lanes S15-S17fires whenMerge-Lane exits 3 (conflict) or 4 (red)what a lane / Qwen seeslane-notes/<ticket>/MERGE-NOTE.mdhow we know it workslane_conflict_relaunch in the diary, then lane_merged for the same ticket

Live proof

15:10:21 [o1] merge of BUG-04 CONFLICTED - relaunching the lane to resolve it (round 1)

15:30 lane_merged o1 BUG-04 - ticket done

23:32:20 [s1] merge of ICON-01 went RED - relaunching the lane to resolve it (round 1)

10 relaunches today, 0 needed Claude

BLDG-01 conflicted 4 times in one day - lanes keep editing the same game files; F-062 (batch merges) would cut this.

F-059The GPU makes art without Qwen

rung 2 · built + tests pass - proven live, awaiting your ACCEPT

what changesLanes.py runs drain_run.py when art waits; drain.lock / merge.lock take turns; never two drains; a refused drain retries in 10 minenforced byLanes.py drain_tick + drain_run.py + Merge-Lane wait_for_drain (code) - S21, S24fires whenart requests pending and no Qwen jobwhat a lane / Qwen seesnew sprites in android/assets, committedhow we know it worksdrain rows by=lanes in the diary; the pending count falls

Live proof

16:34 drain by=lanes - 25 generated

17:24 drain by=lanes - 25 generated

18:4x a second drain refused: 'another drain (pid 41708) is running'

11 drains today, ~150 images, all committed

Found and fixed live: the drain starved behind merges for an hour (18:40), and a refused drain was read as stuck (22:30). The 10-min retry is live since 00:17 - not yet seen.

F-060The Qwen switch: one decision, holds veto it

rung 2 · built + tests pass - proven live, awaiting your ACCEPT

what changesQwen runs on spare capacity or when Claude is out/silent; any hold (art, dirty folder, a lane waiting, gaming, David) keeps it offenforced byLanes.py qwen_may_run + own holds (code) - S18-S22, test_qwen_autofires whenevery minutewhat a lane / Qwen seesQwen starts, or stops at its job endhow we know it worksevery flip is a qwen_switch row with its reason

Live proof

16:28:25 switch off - hold: assets (35 art requests waiting)

16:33:55 claude_usage_back - a hand-set outage handed back by the real probe in 5 s

17:45:46 switch ON - 17 tickets waiting with no free Claude lane

17:55:48 switch off - hold: waiting-on-qwen

23:30 Qwen started after your game closed (check_mem had refused it during the game)

Unproven: a REAL Claude usage limit (only a hand-set one); the silence trigger is the backstop. Fixed tonight: its 'silence' clock froze behind holds (22:30 false 213 min) - live since 00:17, check owed.

F-066Qwen yields gracefully (your idea)

rung 2 · built + tests pass - proven live, awaiting your ACCEPT

what changesa hold mid-job raises the wrap-up flag with the reason: Qwen commits, writes its form, stops in minutesenforced byLanes.py request_wrapup + the wrapup hook (code) - S25fires whena hold appears while a Qwen job runswhat a lane / Qwen sees'SUPERVISOR WRAP-UP ... A HOLD NEEDS YOU OUT OF THE WAY - <reason>' in every commandhow we know it worksqwen_wrapup_requested, then job_end within ~10 min

Live proof

09:11:31 hold waiting-on-qwen (two Claude merges waited) -> qwen_wrapup_requested

09:25:49 Qwen job_end exit_reason self_stop, 24 min, 20 min before its 09:46 deadline; form committed, tree clean

Qwen's last words: 'Stopping now; a HOLD needs me out of the way.'

14 minutes from request to stop - Qwen finished its current step first, by design.

F-064Finish-time estimate (ETA) in every health line and report

rung 2 · built + tests pass - proven live, awaiting your ACCEPT

what changesETA.py: board done/left weighted by est, 6-h/24-h pace, ETA + range, the tickets that need Davidenforced bya script (no AI) - test_eta.pyfires whenevery HEALTH line; the top of every reportwhat a lane / Qwen seesnothinghow we know it worksback-test: as of 18:00 it said Tue 20:24 - Wed 12:34; the range should hold the real finish

Live proof

00:58 HEALTH ... ETA board done ~Tue 23:21 (range Tue 20:54 - Tue 23:21), 43 of 83 done, pace 1.9/h; 7 need David first

02:40 HEALTH ... ETA ~Tue 20:47, 47 of 83 done, pace 2.0/h

Early in a run the 24-h pace includes idle hours, so the first estimates are pessimistic (as of 12:00 it said Friday); the range covers it.

F-065AI audit of every merged ticket (rung 3 for app tickets)

rung 2 · built + tests pass - proven live, awaiting your ACCEPT

what changesan auditor lane (Opus) runs a merged ticket's T0/T1 checks on the live build, writes AUDIT.md; PASS marks tested-AI, FAIL re-opens, 2 FAILs blockenforced byLanes.py start_audit + Lane-Runner --audit-job + Merge-Lane --audit (code) - S26, merge G/Hfires whena ticket merges done after auditSincewhat a lane / Qwen seesthe auditor: an AUDIT JOB prompt; a builder: AUDIT.md findings when its ticket is re-openedhow we know it workslane_audited rows with a verdict; board rows gain tested-AI

Live proof

03:24:42 BLDG-03 FAIL - the generated shrine has no feather totem; re-opened, lane s1 fixing it

04:09:56 DROP-02 PASS -> tested-AI

04:44:31 ZEN-03 PASS -> tested-AI

05:05:43 ICON-01 FAIL - marked done before its icons were judged; re-opened

4 verdicts, 2 real defects caught that the builder lanes had marked done

The first verdict was misread as 'NO' (a wrapped line); the parser is now strict (AUDIT: PASS|FAIL), BLDG-03 corrected. Network outages killed 2 audits; fixed (NET rule).

F-068Audit the backlog, easiest first

rung 2 · built + tests pass - not proven live yet

what changeswhen no new merge waits, the auditor takes the oldest-merged backlog ticket with the fewest device checksenforced byLanes.py audit_target (code) - test_lanes S29fires whenthe auditor is idle and auditBacklog is onwhat a lane / Qwen seesthe auditor: the same AUDIT JOB prompthow we know it workslane_audit_job on a ticket merged before auditSince

Live proof: none yet

Not met live yet - new merges keep the auditor busy (by design they go first).

NETA network outage is never a lane death

rung 2 · built + tests pass - not proven live yet

what changes'can't reach the API' / ENOTFOUND / ECONNRESET... -> the lane waits 2 min and retries the same ticket; no death counted, no hand-over to Qwenenforced byLanes.py NET rule (code) - test_lanes S30fires whena lane run ends with a network errorwhat a lane / Qwen seesnothinghow we know it workslane_requeued with network=true, the same lane continues

Live proof

found live: 05:29-05:48 two outages counted as deaths, handed RAIN-01 to Qwen in quiet hours and gave up CHEF-02's audit (repaired by hand)

fix live 05:55 (commit 5765771)

Waiting for the next outage to show the new line.

F-067This page: Factory Reports

rung 2 · built + tests pass - proven live, awaiting your ACCEPT

what changesfactory_report.py: JSON -> this page, published by one commandenforced bya script; CLAUDE.md: rung-3 reports go herefires whenClaude publishes a reportwhat a lane / Qwen seesnothinghow we know it worksyou are reading it

Live proof

https://prettycooldudes.com/factory-reports

Your verdict comes back by copy-paste, not automatically.

F-055A dirty main folder no longer jams merges

rung 2 · built + tests pass - not proven live yet

what changesMerge-Lane commits only the factory's own files (asset queue, generated art, ATTRIBUTION) then retries; anything else is never touchedenforced byMerge-Lane self_heal_dirty (code) - test_merge_lane E/Ffires whena fast-forward is refused by uncommitted factory fileswhat a lane / Qwen seesnothinghow we know it worksmerge_self_heal row, then the merge lands

Live proof: none yet

Not needed live yet: the drain now commits everything itself, so the situation has not come back.

F-057A red live branch is one alarm

rung 2 · built + tests pass - not proven live yet

what changesMerge-Lane emits live_red once per failing set per hour; the watcher wakes Claudeenforced byMerge-Lane + Watch-Problems (code)fires whenthe live branch fails tests on its ownwhat a lane / Qwen seesnothinghow we know it worksa live_red row

Live proof: none yet

Live has not gone red since it was built.

WATCHClaude watches: problems at once, a look after 7.5 quiet minutes

rung 2 · built + tests pass - proven live, awaiting your ACCEPT

what changesWatch-Problems.py under the Monitor tool; HEALTH shows each lane live / SILENTenforced bythe script + CLAUDE.md on-duty rulefires whena problem event; 7.5 min of a quiet diarywhat a lane / Qwen seesnothinghow we know it worksClaude answered every problem line today

Live proof

caught: BUG-04, TALK-01, BLDG-01 conflicts; the red live branch 10:31; the Qwen merge freeze 14:16; the dead merge lock 16:57; the drain starving 18:40

Rested on Claude re-arming it every 30 min; a Claude Code restart (20:08) stopped it until the next wake.

LOCKSDead merge locks and duplicate Lanes.py fix themselves

rung 2 · built + tests pass - proven live, awaiting your ACCEPT

what changesa merge.lock whose holder is dead is taken over at once; a second Lanes.py refuses to startenforced byMerge-Lane pid_alive, Lanes.py lanes.pid (code)fires whena restart killed a merge / an accidental second startwhat a lane / Qwen seesnothinghow we know it works'merge.lock held by N, which is not running - taking it over'; 'already running (pid N)'

Live proof

19:0x second Lanes.py start: 'already running for this project (pid 41264)' - exit 3

The lock takeover was coded after the 16:58 jam - not yet seen live.

WIPQwen's unfinished work is saved on a stop

rung 2 · built + tests pass - proven live, awaiting your ACCEPT

what changesthe supervisor stashes Qwen's uncommitted work to wip/qwen-<time> before exitingenforced byFactory-Supervisor.ps1 (code)fires whenthe supervisor is told to stopwhat a lane / Qwen seesnothinghow we know it worksqwen_wip_saved row; the main folder clean

Live proof

19:35:13 qwen_wip_saved - branch wip/qwen-20261005-1935

Nothing hands that branch to the next builder automatically (Claude left s2 a note for DECOR-01).

DIARYLanes record their tokens again

rung 2 · built + tests pass - proven live, awaiting your ACCEPT

what changesLane-Runner's event() crashed on every run end since F-049enforced bycode + test_diary_eventfires whenevery lane run endwhat a lane / Qwen seesnothinghow we know it workslane_end rows with tokens

Live proof

15:48:53 lane_end p1 plan tokens cost_usd 3.28

BURSTOpus bursts before each usage reset, and gaming mode

rung 2 · built + tests pass - proven live, awaiting your ACCEPT

what changesburst.py: a detached scheduler (last 30 min of each 5-h window) + a gaming watcherenforced byburst.py in its own tabs (survives a Claude restart)fires when00:50, 05:50, ... ; the game closed 10 minwhat a lane / Qwen seeslanes switch modelhow we know it worksburst and gaming_mode rows

Live proof

20:20:04 burst off - lanes back to normal

21:32:39 gaming_mode off - Minecraft Dungeons closed

Brittle by design: the window times assume back-to-back windows. The gaming watcher only knows one game - Rocket League ran with gaming mode off (check_mem still protected it). F-063.

F-061Budget pacing

rung 2 · built + tests pass - not proven live yet

what changesnear the window limit only one builder starts new workenforced byLanes.paced + Usage-Pace.pyfires whenafter calibration, at 85%what a lane / Qwen seesother lanes waithow we know it worksbudget_pace row

Live proof: none yet

Off on purpose - brittle (cost is recorded only at run end; your chats are not counted).

F-062Merges stop re-running the suite for art and config commits

rung 2 · built + tests pass - proven live, awaiting your ACCEPT

what changeswhen the live branch moves during a merge check only by art / the asset queue / .factory config, that move is folded into the tested merge and it lands at onceenforced byMerge-Lane fold_factory_move (code) - test_merge_lane I/Jfires whenlive moves mid-checkwhat a lane / Qwen seesnothing - tickets merge soonerhow we know it workslane_merge_folded rows; 'moved during the check' refusals drop from 11 a night to ~0; done->merged median under 18 min

Live proof

measured last night: 39 merges, 11 re-ran the full suite (6-8 min each) because a drain or a lanes.json switch committed mid-check; median wait 18 min, p90 58

06:25:24 lane_merge_folded s4 ICON-01 - live moved by factory files only, folded in, no second suite run

Part 2 (batch merges) waits for a re-measure after one night with part 1.

F-063Gaming mode for any game, by itself

rung 2 · built + tests pass - not proven live yet

what changesa watcher sees any game start (Windows' own game list, Steam, Xbox games) -> Qwen held, no art drain, models unloaded; 10 min after the game closes it lifts. Holds expire, so a forgotten 'on' heals.enforced bygaming.py watch (own tab) + Lanes / drain_run / supervisor drain gates (code) - test_gaming A-E, test_lanes S31fires whena recognised game startswhat a lane / Qwen seesQwen: the wrap-up line, then no new jobhow we know it worksdiary gaming_mode on auto=true game=<exe>, off ~10 min after; no drain row in between

Live proof

watcher running since 06:1x (tab gaming-watch)

Live proof is your next game. A launcher-only open (League client) counts as gaming.

F-058Qwen in its own folder

rung 0 · not built

what changesparked

Live proof: none yet

The real fix for merges waiting up to 45 min behind a Qwen job.

Stability matrix

state change	mechanism	offline test	live proof	rung
Claude out of usage	F-040 mark + F-060 switch	S19	16:33:55 hand-set outage handed back	2 (real limit unproven)
Claude silent 45 min	F-060 silence trigger	S20	22:30 fired - falsely (clock bug, fixed 9f17bc7)	2
art waiting	assets hold + drain	S21, S24, S25	16:28 hold + drain	2
lane blocked on Qwen	waiting-on-qwen hold, F-066 wrap-up	S25	09:11 wrap-up requested -> 09:25 Qwen self-stopped, tree clean	2
dirty main folder	dirty-main hold, F-055 self-heal, WIP save	E/F	17:38 hold kept Qwen off; 19:35 WIP saved	2
gaming	gaming hold + check_mem	S18	19:50 on, 21:32 off; 22:30 check_mem refused during Rocket League	2
merge conflict	F-056	S15, S16	15:10 BUG-04 relaunched and landed	2
red merge	F-056	S17	23:32 ICON-01 relaunched	2
red live branch	F-057 live_red	-	none yet	2
dead merge lock	pid_alive takeover	-	none yet (fixed by hand 16:58)	2
two Lanes.py	lanes.pid	-	19:0x second start refused	2
drain busy / refused	drain.lock, exit 3/4 = retry	S24	18:4x second drain refused	2
night (quiet hours)	quiet-hours hold 23:00-09:00	S27	00:35:28 hold appeared	2
a hold while a lane is busy (the frozen clock)	track_alive every tick	S28	no false silence since 02:02	2
network / API unreachable	NET rule: wait 2 min, retry, no death	S30	05:29-05:48 outages counted as deaths (old code); fix live 05:55 - new line not seen yet	2
a merged ticket is wrong	F-065 audit re-opens it	S26, merge G/H	03:24 BLDG-03 and 05:05 ICON-01 re-opened	2
Opus burst / usage reset	burst.py scheduler	-	00:50 and 05:50 bursts on, 01:21 off	2

Your verdict: ACCEPT what convinces you.
Owed live proofs: F-066 Qwen complying with a wrap-up (Qwen is off until 09:00, quiet hours), the 10-minute drain retry, F-068's first backlog audit, the NET line on the next outage.
5 tickets need you first: ICON-01, MUSIC-01, CHEF-03, ART-06, ART-07.

Hybrid factory: what was built on 2026-10-05, and what is proven

Upgrades by rung

Where the day's bugs were found

The upgrades

F-056A lane fixes its own merge conflict or red merge

F-059The GPU makes art without Qwen

F-060The Qwen switch: one decision, holds veto it

F-066Qwen yields gracefully (your idea)

F-064Finish-time estimate (ETA) in every health line and report

F-065AI audit of every merged ticket (rung 3 for app tickets)

F-068Audit the backlog, easiest first

NETA network outage is never a lane death

F-067This page: Factory Reports

F-055A dirty main folder no longer jams merges

F-057A red live branch is one alarm

WATCHClaude watches: problems at once, a look after 7.5 quiet minutes

LOCKSDead merge locks and duplicate Lanes.py fix themselves

WIPQwen's unfinished work is saved on a stop

DIARYLanes record their tokens again

BURSTOpus bursts before each usage reset, and gaming mode

F-061Budget pacing

F-062Merges stop re-running the suite for art and config commits

F-063Gaming mode for any game, by itself

F-058Qwen in its own folder

Stability matrix

Next