lampOS builders' queue #705

Continued in #1722 (19 Oct 2026 10:4x): #705 reached GitHub's limit of 2,500 comments. Post the queue, the hourly notes and builder-to-builder notes on #1722.

Read the newest comments, not the first page (16 Oct 13:41): this issue has hundreds of comments, and the API's first page is the oldest 100. Read new ones with gh api 'repos/autonomous-ai/lampOS/issues/705/comments?since=YYYY-MM-DDTHH:MM:SSZ&per_page=100' --paginate, giving the time of your last read. Eighth missed three tasks this way.

The order in which each lampOS builder takes its tasks. Kept by the first builder (the one with lamp-4ace); only it edits this body. Decided by the owner and the first builder on 2026-10-14 (fifth-P3), replacing the list in lampOS/docs/work-split.md, which keeps the rules, the loops, "What the lamp showed" and the history of finished tasks.

A builder's next task: the first issue in its own list below that is open and not status:taken. Claim it as lampOS/docs/joining.md says ("Tasks as GitHub issues"): add status:taken, comment taken by <name> · session <id> · <lamp or none>, then read the comments again and release it if the first "taken" comment is not yours. Questions and answers are comments on the task's issue. Only the first builder opens issues and makes or moves builder:<name> labels (#991, fifth-P8, decided 16 Oct). A problem found or a proposal is one comment, on the related issue or PR, or on this Queue issue headed <name>-P<n> or Found by <name> when none fits. The first builder opens an issue or says no, and no answer means not taken up. A builder takes only an issue labelled with its own name. A finished task is handed back as one pull request from <name>/<task> into lamp-voice, not a draft, with its ARM64 run green (joining.md, "One pull request per task", fifth-P6, decided 18:33); since 2026-10-15 14:50 the reviewer merges it if it is off unless set (lampOS/builders/reviewer.md), and the first builder merges the rest after a minute on lamp-4ace. A merge into lamp-voice does not close the issue: close it by hand.

Owner's order of 2026-10-14, 14:00: parts that stand on their own first, the harder ones later. And at 17:22: "reimagine the entire experience. make lamp the most fun creature to interact with": the design is lampOS/docs/creature.md; its pieces (#710-#714) come first. Stopping and interruption are tested by the first builder in the evenings.

Disk: every builder, every machine (the owner, 16 Oct 23:05)

• One build folder per checkout. One CARGO_TARGET_DIR for all your builds in a checkout, and one podman target volume for your ARM64 builds. No new target-* folder or volume per topic or per PR.
• Never share one between checkouts building at the same time, and never use another agent's. Cargo judges a workspace crate by its path and file times, so it can run the other checkout's stale test binary (10 Oct: "0 passed; 32 filtered out") or swap in the other build's crate (16 Oct, the reviewer's gate on lampOS: a skill's own short sound, sounds/NAME.wav, played by the voice (#938) #1144: E0460/E0463).
• Delete each run's output when you finish with it: source exports, OUT* folders, recordings, scratch clones, extra worktrees. Keep only what is installed on a lamp and its rollback.
• Past 20 GB, cargo clean the build folder. For a podman volume, remove and recreate it, then podman machine ssh <vm> -- sudo fstrim -av: the VM's disk image gives no space back without it.
• Below 20 GB free on the machine (df -h), clean before you build. The office iMac ran out at 22:55 on 16 Oct and an install on lamp-4ace failed half-way.

Ship V2 to homes by Black Friday, 27 November 2026 (the owner, 2026-10-15 16:13)

"we want to ship by black friday or latest by christmas to user's home." Six weeks. Everything below is for that date; anything not on this list waits until after it.

Must have: setup for a home (#806, third), signed updates (#737, second, with the owner's key), the voice never dying (#784, #762, fifth), speed (#799 second, #782 fourth), knowing who it is talking to and when not to answer (#800, #753, #792), privacy (#795, third), head touch (#711), the hello and faces (#759, #775), delegation and its news (#751, #765, #798), the backup voice (#778, sixth), giving way to a hand (#780, #790, fifth), and learning from real people (#803, sixth), and the first five minutes: the lamp walks its new person through where to sit, their name and face, voice, touch, privacy and giving way (#841, seventh).

Every component of the brain first (the owner, 2026-10-15 about 07:00)

"in the mvp, did you build ALL the components of the brain? and now have builders perfect/optimize each of those components?" Not yet. These parts of lampOS/docs/brain-build.md are missing, or only watching, on lamp-4ace, and each now has a builder. When every row works with the others, each builder owns a set of components to perfect.

The V2 MVP first: wide, then deep (the owner, 2026-10-15 00:37)

"prioritize the tasks for builders with ruthless priority. so we have an end to end experience on lamp first. then we'll optimize each component and edge cases later … wide first. deep later … have all the pieces in place make sure the whole works. then after that, have builders optimize each component."

So until every piece below is on lamp-4ace and works with the others, a builder works only on its MVP tasks. Everything else waits under "After the MVP", even if already taken: pause it, push what you have, and say so on its issue.

ARM64 runs, since 2026-10-15 15:02 (af7aed833): the run starts when your pull request into lamp-voice is opened (draft or not) and at each push to it, and a newer push cancels the older run. Pushes to <name>/* and lamp-<name> start none; rebase onto lamp-voice so your branch carries the new workflow. Every push to every branch, with four runners a run, had queued eleven runs behind eight, lamp-voice's own among them. Runs on lamp-<name> branches were cancelled at 15:02.

Never end a turn without a wake-up (the owner, 2026-10-15 14:21: builders 2, 3 and 4 were found idle). A session that ends its turn sits idle until someone types to it, and waiting for an ARM64 run or a review is not a reason to stop. Before you end any turn, schedule your own check in 10 to 15 minutes, with /loop or a cron in your session. That check re-reads this Queue and your issues, then carries on: the next step of your task, or your next task if it waits. If your list has nothing you can do, say so on this issue at once, in one line, and the first builder fills it within the hour.

Never idle, until V2 is done (the owner, 2026-10-15 about 06:48: "not just finishing the mvp, but keep building until v2 is done!"). If every MVP task in your list is waiting (on a lamp, a person, the owner's key or a review), take your next task under "After the MVP" and keep going. Say on the waiting issue what it waits on. When an MVP task becomes possible again, it comes first.

How a lamp is set up (phone, desktop) is not decided by the owner and is not in the MVP yet.

The first builder plays the whole lamp end to end every evening (a person's day: arrive, talk, ask, hand over work, get news, a reminder, touch, leave, night) and files each fault as an issue to a builder with room.

Today (16 Oct): 80%

Physical skills (the owner, 16 Oct 07:20): lampOS/docs/skills.md. Built-in behaviours become skills on lamp-skills; first five by 19 Oct: #938 (seventh), #939 (second), #940 (third), #941 (ten).

The owner, 16 Oct 05:5x: "let's move faster. let's aim to get to 80% today." Targets by part on #705 (first builder, 05:56). Accepted-on-lamp PRs are merged by the first builder the same hour.

reviewer (new, 2026-10-15 14:50; a Codex agent the owner starts)

Every ready pull request into lamp-voice, oldest first: review, run the gate, and merge the ones that are off unless set and pass the list in lampOS/builders/reviewer.md; leave the rest to the first builder with a comment. No pull request waits more than 30 minutes for its first review.

tests (renamed from pipeline on 2026-10-15; a Codex agent the owner starts; slow and flaky tests, the workflow being the devops team's)

0. next (17 Oct 21:4x) — Lamp V2 as its own device in the Autonomous app: online over MQTT, a device page, a lamp-v2 type #1396, the backend part in ecm-sds (office Mac's access): what lamp-v2 needs, after your Slow tests outside the realtime group: 118 tests of 1 s or more take 269 s one at a time (after #823) #1128 build.
1. now (17 Oct 16:2x) — peer reviews: lampOS: the self-test's head move reads each start once the joint holds still (#1329) #1340 (Ten's fix for your lampOS: the self-test's head move starts the next move before the head has settled, so its return can miss NEAR (passes by 0.1 count) #1329 hazard, the head move reads its start once the joint holds still), then the seventh's lampOS: the supervisor tests' inert child builds its worker only once connected, so its 250 ms initial-authority clock does not run during the connect #1330 after-bind PR when it opens. Then item 1.
2. now (17 Oct 06:5x) — three lamp-voice lib tests fail in the full one-thread run under load and pass alone (the reviewer, lampOS: with LAMP_TASK_NEWS_AT_PAUSE=on, tasks that finished together are told in one turn, at most three (#1161's P3) #1223 06:43, 46. The voice's timed process tests fail on a loaded host: make them deterministic #739's family): hand_controls::a_tap_on_a_quiet_lamps_head_plays_its_hm_as_a_turn_of_its_own (listening_not_ready), service_mute::active_physical_mute_cancels_owned_reply_then_reaps_without_completion and service_mute::repeated_verified_mute_cycles_keep_one_logger_and_never_open_audio_while_waiting ("privacy worker did not stop within 300 ms"): move each onto the test clock or a wait on the event, so it never depends on load
3. The 123 real-clock tests run on a test clock: 977 s of waiting gone, every test kept #823 — The 123 real-clock tests run on a test clock: 977 s of waiting gone, every test kept
4. A flaky lamp-policy test turns every ARM64 run red (program.rs:363) #783 — the sweep: the other program tests that kill a part and read its lines (moved from the eighth)
5. The ARM64 check in 10 minutes, with the same tests #788 — closed for the pipeline builder: the workflow and runners are the devops team's

nine (new, 2026-10-15 16:43; no lamp)

0. now (17 Oct 21:4x) — Lamp V2 as its own device in the Autonomous app: online over MQTT, a device page, a lamp-v2 type #1396, the lamp side with Ten: V1's MQTT contract written down (topics, the cmd: data answer, what "connected" needs); the private repos went to the office Mac's builders.
1. now (17 Oct 16:2x) — lampOS: the voice tells the model the name typed at setup and says hello by it (#806) #1323 merged at 16:14 (the typed name). Next is item 2 (No internet at home: the lamp says so once, keeps what works locally, and recovers by itself #830). For Sunday: B8's diff of lamp-4ace's voice unit against deploy/product/lamp-voice.service at the integration commit (Setup for a home: V2 ships with V1's setup flow (Wi-Fi, service, name), and a 5 s reset #806, 15:50).
2. One install for a new lamp: every part, its unit and its settings, from a signed build #811 — the rest of the install (lampOS: one install for a new lamp (#811) #815 merged 2026-10-15 22:19, with the second's fixes): setup's and the keeper's units at first boot, and Harness as a systemd user service with lingering on (the first builder's note on One install for a new lamp: every part, its unit and its settings, from a signed build #811, 20:19)
3. No internet at home: the lamp says so once, keeps what works locally, and recovers by itself #830 — No internet at home: lampOS: no internet at home: the voice starts offline, the clock is known, lamp/online (#830) #884's start offline passed on lamp-4ace (06:27, 2 min 22 s cut, no restart, answered 2 s after); the "listening with no link" line never printed; the reviewer's findings answered in 62f48cbdb

ten (new, 2026-10-15 16:46; no lamp)

0. now (17 Oct 21:4x) — Lamp V2 as its own device in the Autonomous app: online over MQTT, a device page, a lamp-v2 type #1396, the lamp part: the design (MQTT presence, the app's info nudge, the device API), before code.
1. now (17 Oct 16:2x) — lampOS: the self-test's air is not_fitted where no sensor was found since boot #1324 (air not_fitted: the reviewer's 16:08 P2, a missing test, and three P3s), lampOS: lamp-health show, #1327's P3s (first line = first row, zones and load for the owner) #1337 (lamp-health show: the reviewer's 16:08 P2), lampOS: switches.md's product column as #1332 sets it (docs only) #1335 once lampOS: the product voice runs lamp-4ace's switches of 17 October, bounded reference recovery and the call's wait #1332 merges, lampOS: the self-test's head move reads each start once the joint holds still (#1329) #1340's peer review by Tests.
2. today — In an English conversation, a turn heard in another language is answered in that language #926 (lampOS: in a conversation, a turn heard in another script or as a bare number is asked again (#926) #930): cut-ins heard as nonsense held, 0 wrong answers of 10 on lamp-4ace at 06:35; the "Sorry?" did not follow and a three-word fragment got past: 6eae0ebd5 next on the lamp. The self-test unit (lampOS: lamp-4ace's self-test unit, on demand with CAP_SYSLOG (#812) #929) merged 06:53 after its report was kept on lamp-4ace
3. Physical skills: the lamp skill creator, and Stretch break and Stuffy air written with it #941 — physical skills: the lamp skill creator, Stretch break and Stuffy air written with it (by 19 Oct)
4. Every lamp checks itself: a two-minute self-test of each part, at the factory and at first boot #812 — Every lamp checks itself: a two-minute self-test of each part, at the factory and at first boot (ship, 27 Nov)

second

0. now (17 Oct 16:2x) — lampOS: records no longer refuses every run on a lamp: a process ending mid-scan, the current link, and the product's records.json #1328 (records: the reviewer's re-check of 2dfa86830, then merge; its lamp check is tonight's integration install), lampOS: judge a sound while an answer waits by its words (#1123 step 2) #1277 to merge, then its "Stop." P3 as a follow-up.
1. lampOS: the Policy holds the lamp's actions, with LAMP_POLICY_HOLDS_ACTIONS (#800) #908 — the Policy holds lamp actions: 1ab26f986 passes holds.sh on lamp-4ace with Policy 8a609d9d3 (06:51: the unnamed light remark held, named ones run in 0 to 2 ms); the reviewer's P2 of 06:40 (request boundaries, bus_out.rs:432), then the first builder merges
2. lampOS: a story asked of the lamp is told by the lamp, with LAMP_TASKS_ASKED (#854) #903 — the story told by the lamp: accepted on lamp-4ace 03:37 (stories 4 of 4, named Claude Code 2 of 2); the reviewer's merge; the first builder plays handovers.sh again at 08:00
3. Physical skills: the say service, the Policy as the gate, Reminders and Task news as skills #939 — physical skills: the voice's say service, the Policy as the gate, Reminders and Task news as skills (by 19 Oct)
4. Jev 1.14: replay the Policy's logged decisions through 1.13 and 1.14, and switch if it is better #871 — Jev 1.14 pictures (lampOS: lamp-policy-replay compares decision models on the logged decisions, and lamp-policy --images asks again with the eyes' frame, both in shadow #873, --images, with picture_person): on lamp-4ace, watching; labelled scenes with the owner at the seat 2026-10-16 11:00, then lamp-policy-replay --pictures
5. 48. The Policy decides when the lamp speaks first: progress news, then hello #752 — 48. The Policy decides when the lamp speaks first: progress news acting since 20:07; the hello (Proposal second-P5: the hello on arrival asks the Policy (the second half of #752) #856, lampOS: a due hello asks the policy on voice/arrival with LAMP_HELLO_ASKS_POLICY=on, and lamp-policy --hello answers, acting with --act hello #883) watching on lamp-4ace
6. 45. Updates: a lamp keeps its lampOS current by itself #737 — 45. Updates: the publish: job, ready for the owner's key

(#770 moved to the seventh at 14:00: untaken since 10:40, and the seventh now owns the tasks part's side of delegation.)

#736 is done (5 of 5, first builder, d128c458a); #740 stays open for the other shape if you still want it. Later: #723 (the early end), #694 (the bare name; #750 not installed)

first (the first builder's own; never idle, the owner 16 Oct 10:30)

0. Sunday 18 Oct (now, 09:5x): the owner arrives about 15:00. Before then: lampOS: a line's voice print is asked for once 1.8 s of it is heard, not at 3 s, so scene 3's short lines are printed before their answers (#829) #1445's scene-3 trial on lamp-4ace (running), the clock read for Seventh's capture-worker question, lampOS: lamp-eyes' 10 s line says why each glance was detected and where a look's time goes (--look-on-change) #1452's eyes for 20 min with nobody there, lampOS: lamp-name rests its models while the room is quiet (--quiet-below, off unless set) #1440's ungated name calls (10:07), the build/quality report each hour. 13:30: lamp-voice's head on lamp-4ace (every part but the body, NO_BODY=1), then the battery, ten cut-ins, self-test, health, and lampOS: the voice tells lamp-tasks the news heard whole, so --close-told can close the harness the lamp started #1425's told check. From 15:00 with the owner: records aside, the arm folded and the new body started (Fifth's 4a page), button hold, head pads, the "Stop." recording (mic2 at 75 and 55, then mic1), setup from nothing with the Autonomous app (El's checklist after), testing together (lampOS: Expression's sway between sentences turns the base about 52 counts instead of rolling the head #1067's 4b story turn), then the Monday summary. Items below dated 16–17 Oct are done or folded into this.
1. Sunday, changed (17 Oct 21:1x): the owner is in only from about 15:00. Monday's build is cut at about 11:00 and soaks on lamp-4ace; setup from nothing waits for the owner on lamp-4ace (weekend: nobody at another lamp; lampOS builders' queue #705 21:1x, withdrawn 21:2x); lamp-4ace's hands-on steps and reboots wait for the owner.
2. now (17 Oct 20:3x) — the Monday summary is checked against its sources (Jarvis, 20:01) and corrected. The air sensor's cause is corrected (20:2x, robots/lamp/hardware: lamp-4ace's air sensor read again after a start by hand, not shown to be the bus speed #1379): at 400 kHz it reads 20 of 20; it came back after a start by hand. lampOS: the product sets I2C bus 0 to 100 kHz at each start, the air sensor's limit, with lamp-air limit-clock #1368 stays as the sensor's limit, and lampOS: lamp-air reads the SEN63C through a 400 kHz bus's filler, starts it when its name fails, and takes CO2 0x7FFF as not known #1366's start is the likely fix. The lamp checks of lampOS: the product sets I2C bus 0 to 100 kHz at each start, the air sensor's limit, with lamp-air limit-clock #1368 (works) and lampOS: the base strip checks a fifth of the MPR121's configuration each sample, so it keeps its 5 ms pace at 100 kHz #1374 (not possible: no base in lamp-4ace's pins) are done; their merges wait for a person's go. The second-voice A/B on lamp-4ace runs 20:3x–21:2x, then the battery, then lampOS: a voice waiting on the privacy switch or a mute says its microphones are off, so lamp-health reads it muted, not quiet #1376's lamp check. Tasks out tonight: lampOS: the voice ends at midnight when the lamp's daily log jobs run (lamp-4ace, 17 Oct 00:00:06) #1178 (eighth), lampOS: a home lamp keeps the person's time zone: memory and Expression follow /etc/localtime, setup writes zone.env again (#820) #1380 (third, done), session drops (twelve), 39. Seeing someone speak: whether the mouth of a face turned to the lamp is moving, so the lamp can tell it is being talked over #720 (fifth), lampOS: a home lamp keeps the person's time zone: memory and Expression follow /etc/localtime, setup writes zone.env again (#820) #1380's review (el).
3. now (17 Oct 16:5x) — the 18:00 freeze: the integration build of every part at one lamp-voice commit, installed on lamp-4ace with its checks (battery, ten cut-ins, self-test, the body), posted here for Tests' full ARM64 run and El's test channel. LAMP_SPOKEN_TO decided by the 16:31 A/B first, and the speaker-volume test for "Lamp, stop." ("Stop" said while the lamp talks stops it at once, every time (1 of 7 short stop words on 14 Oct) #828). lampOS: Expression forgets the oldest of 32 retired voice sessions instead of stopping after 32 voice restarts #1349 (Expression's session fence) mine, in review.
4. tonight — the midnight check at 00:10 (lampOS: the voice ends at midnight when the lamp's daily log jobs run (lamp-4ace, 17 Oct 00:00:06) #1178), the overnight soak; 07:00 the morning report, 07:30 the go/no-go to the owner.
5. Sunday 10:00–14:00 with the owner — the agenda on the review page: setup on lamp-4ace (the hotspot scripts back at 09:00, the owner's yes); also lampOS: a home lamp keeps the person's time zone: memory and Expression follow /etc/localtime, setup writes zone.env again (#820) #1380's zone check, lampOS: lamp-air reads the SEN63C through a 400 kHz bus's filler, starts it when its name fails, and takes CO2 0x7FFF as not known #1366's idle-start test and five edit-history revisions for the owner. Earlier plan: setup on lamp-06c1 if it is in the office (El, by 18:00), otherwise on lamp-4ace with the radio scripts put back (Setup for a home: V2 ships with V1's setup flow (Wi-Fi, service, name), and a 5 s reset #806 16:13, 16:16); the hands-on list; testing together; the Monday summary.
6. now — lamp-4ace: voice stack39 since 01:45 (raw mic2 stop words, lampOS: with LAMP_STOP_WORDS=heard-always a stop word said over the lamp's voice stops it (#828, on #970) #1047 73add3ae6: 8 of 10 cut-ins stop it, against 7); body back on the evening's 4396837, at rest and refusing to stand: at 01:58 an unchecked elbow goal register (37639) was clamped and written as goal 4095 (lampOS builders' queue #705, 02:20). No body on any lamp until Fifth's guard lands; a person looks at the elbow Monday. Midnight crash lampOS: the voice ends at midnight when the lamp's daily log jobs run (lamp-4ace, 17 Oct 00:00:06) #1178: check and unmute after every midnight. Next: "Lamp, stop." still never stops it
7. Every ready voice PR reviewed the hour it is ready, and merged on a clean review plus green ARM64 (16 Oct night: lampOS: the first request after a hand or the head stopped the lamp tells the model it was stopped (#829, on #1030 and #1107) #1148, lampOS: a skill's own short sound, sounds/NAME.wav, played by the voice (#938) #1144, lampOS: with LAMP_ASKER_FIRST=on speech well under the asker's takes no turn while its answer is awaited (#829) #1092, lampOS: a skill tool's look is the voice's own, not the model's to remember #1146, lampOS: with LAMP_WAIT_KEEPS_ANSWER=on a short sound does not cut off an answer still to come (#1123) #1130 sent back, lampOS: a look on the second voice waits 5 s, then says it couldn't get a good look (#1044) #1154 merged); lampOS: a mute left by the restart guard is known and a head tap wakes it, as sleep (LAMP_RESTORED_MUTE, #960) #997 re-checked tonight
8. Face records: only the owner's own record kept on lamp-4ace since 18:22 (four others set aside); then lampOS: with LAMP_NAMES=checked a name told is kept, only a name the person said, and a face keeps one name #1057 (LAMP_NAMES=checked) on lamp-4ace
9. Busy-room calls: lampOS: with busy-near each verdict line and the name note's line give the room microphone's rise #1083 (journal lines only now) and lampOS: with LAMP_SPOKEN_TO=day-and-night the name needs no face in a busy room only at night, and by day counts alone only heard surely (#829) #1158 (night and day, draft); count false calls and the owner's held calls before and after
10. lampOS: with LAMP_WAIT_KEEPS_ANSWER=on a short sound does not cut off an answer still to come (#1123) #1130 on lamp-4ace once its P1 and P2s are fixed, then count_1123.py again against the day's 1,377 / 867
11. Desk tutor: its own look (lampOS: a skill tool's look is the voice's own, not the model's to remember #1146's P2s) and its light reached (third), then a real page test with the owner on Monday
12. Heat: lampOS: let the eyes look twice a second instead of four while one face sits still #1114 is merged behind --steady-rate; switch it on for lamp-4ace's eyes and log the temperatures beside it
13. Always: lamp minutes for every pull request, the battery each hour, the local review page (~/lamp-review/lamp-review.html)

third

0. now (17 Oct 16:2x) — lampOS: the light steps warmer and cooler along white, warm and orange, and the lamp is told its light can change #1316 to merge, then the white follow-up as its own PR; lampOS: a stop that also ends the hotspot script is a clean stop, so the setup unit doesn't read failed (#1336 review) #1341 (a stop while the hotspot starts); the head-touch sheet for Sunday by 20:00 (33. The base's touch strip and the head's two pads: read, slide, and new meanings #711).
1. today — While it speaks, the lamp's head and body barely move: it reads as a speaker, not a creature #933 (stills 06:21: the head and body barely move while it speaks) for 80%; lampOS: the head pads, end to end (#711) #849 on lamp-4ace since 06:05 for the owner's hand at 11:00 (the reviewer's P2 on the tap window); lampOS: Schedule's requested stop is not a storage failure (#869) #928
2. 33. The base's touch strip and the head's two pads: read, slide, and new meanings #711 — the head pads, end to end (the owner, 2026-10-15 evening: the head carries every touch action; the base strip waits): tap stops it or makes it listen, a hand leans it in, three seconds sleeps it, a tap wakes it (docs/creature.md, "Touch")
3. Physical skills: arbitration at the body and the light, and move files from skills #940 — physical skills: arbitration at the body and the light, move files from skills (by 19 Oct)
4. lampOS: the privacy switch closes both microphones and the ring shows it, behind LAMP_PRIVACY_CONTROLS=on #795 — the privacy switch (pull request; the first builder's lamp minute with a hand on the slide)
5. Setup for a home: V2 ships with V1's setup flow (Wi-Fi, service, name), and a 5 s reset #806 — setup's follow-up for the reviewer's four findings on lampOS: lamp-setup, V1's setup flow and the reset button for a home lamp, off unless set #817
6. Proposal third-P1: Expression reads the system's time zone, so a lamp outside Vietnam keeps its days across daylight saving #820 — Expression, Memory and the policy read the system's time zone
7. 8. A shared log, settings, and reports from the parts that give none #699 — pull request lampOS: a shared log topic and lamp-settings, both off unless started #808, the shared log and settings

After the MVP: the base strip as a slider (#797 merged as a tool run by hand)

After the MVP: #699 — 8. A shared log, settings, and reports from the parts that give none

Done: #695 — 6 (on lamp-4ace since 16:53) · #696 — 11 (on lamp-voice; off on lamp-4ace until the voice uses it)

fourth

0. next (17 Oct 21:4x) — Lamp V2 as its own device in the Autonomous app: online over MQTT, a device page, a lamp-v2 type #1396, the app part in ecm-sds-mobile v3 (office Mac's access): a lamp-v2 entry, then a V2 device page on Ten's API, after lampOS: the product sets LAMP_SPOKEN_TO=on, so a busy room after the lamp's news is not deaf, with the five follow-ups test (#828) #1378 merges.
1. now (17 Oct 16:2x) — lampOS: a stop word said over the lamp's speech clears LAMP_FOLLOWS_ON's clock as a hand or head stop does (#1278's optional P3) #1338 (green ARM64, then merge); the two-people sheet for Sunday by 20:00.
2. lampOS: a bare call ends nothing, behind --call-wait-ms, with lamp-name's first sign on name/rising (part of #782) #891 — the bare call (on lamp-4ace since 00:40): the reviewer's findings, then merged; lampOS: a cut keeps its name on lamp-4ace's own path (the #809 follow-up, part of #782) #853 with it
3. Two people at the lamp: it answers the one who spoke to it, faces them, and leaves their talk alone #829 — step 2, the lamp speaks when spoken to: at the restored speaker level (2026-10-16 02:16) the scene's first unnamed question is answered 11 of 12 (0 of 12 at 01:04 was the louder lamp not hearing it); with the two tests asked at 02:00 (no lamp action from talk not for the lamp)
4. "Stop" said while the lamp talks stops it at once, every time (1 of 7 short stop words on 14 Oct) #828 — stop words: step 2 (a stop word heard while paused ends the answer), and the request's start kept from the earliest speech-like block when the onset is found late ("Lamp, stop. What is two plus two?" heard as "vestido", 02:00)
5. A pause at a comma ends the person's turn (smart-turn-fast, 400 ms at p=0.99) #782 — pull request lampOS: keep the call across a cut at a comma, behind LAMP_CUT_KEEPS_NAME, and let --endpoint-commit-ms choose how long a confident verdict waits #809 (comma pauses), then the first builder's 400 vs 550 minute

Done: #749 — 47 (pull request #773, merged 12:36). Next, the faults the first builder's end-to-end runs find. (#739 moved to the eighth.)

Done: #698 — 29 (pull request #743, merged 02:51; on lamp-4ace since 02:47) · #697 — 28 (on lamp-4ace since 2026-10-14 16:54) · #713 — 35 (pull request #727, merged 23:01) · #707 — 30 (pull request #721, merged; on lamp-4ace since 20:21) · #700 — 17 (pull request #722, merged)

sixth (new, 2026-10-15)

0. now (17 Oct 20:4x) — the Gemini session ending under a person: server_go_away, transport, peer_closed and stalled_response on lamp-52e6's sessions (what each cost, then the fix for the worst, for Monday's build; Twelve has barrier_timeout; lampOS builders' queue #705 20:4x).
1. Proposal sixth-P2: the second voice looks when it is asked to (#754 open limit 5) #1044 — the second voice looks when it is asked to (proposal sixth-P2).
2. Proposal sixth-P1: the second voice's first words under 1.5 s (it takes 3.0–3.6 s on lamp-52e6) #779 — the second route's first words under 1.5 s (proposal sixth-P1).
3. now (17 Oct 16:2x) — lamp-52e6: release e4f1d32b against head d4401a2fa, alternating blocks of five, 20 each ("Stop" said while the lamp talks stops it at once, every time (1 of 7 short stop words on 14 Oct) #828; lampOS builders' queue #705 16:04). Then lampOS: a stop word said over the lamp's speech clears LAMP_FOLLOWS_ON's clock as a hand or head stop does (#1278's optional P3) #1338's lamp minute.
4. today — lampOS: the startup hold's replay to Gemini fits its input queue and its audio age bound #923's P2 (replay bounded by the real queue); the speech-to-text first-word time for the fourth's stop phrases ("Stop" said while the lamp talks stops it at once, every time (1 of 7 short stop words on 14 Oct) #828), through the relay on lamp-4ace
5. lampOS: the second voice takes its final words with Finalize, a third of a second sooner (#779) #878 — the reviewer's P2 of 2026-10-15 23:57: the second voice's speech-to-text ends a line at an earlier speech_final and drops its tail ("set a timer" without "for twenty minutes"), in both arrival orders; the fourth's stop words ("Stop" said while the lamp talks stops it at once, every time (1 of 7 short stop words on 14 Oct) #828) will use the same speech-to-text
6. Gemini's link goes local-unavailable 44 ms after its first Ready when the startup hold replays a held line #892 — Gemini's link goes local-unavailable 44 ms after its first Ready when the startup hold (lampOS: at the voice's start a request waits up to 3 s for Gemini before the second route takes it (#779) #885, merged) replays a held line
7. Proposal sixth-P1: the second voice's first words under 1.5 s (it takes 3.0–3.6 s on lamp-52e6) #779 — the second route's latency: target p50 ≤ 2.5 s, p95 ≤ 4 s; direct provider key or the relay team: the owner decides 2026-10-16 11:00
8. "Lamp, that was weird": a person marks a bad moment for review #803 — "Lamp, that was weird": a person marks a bad moment for review

seventh (new, 2026-10-15)

0. now (17 Oct 16:2x) — lampOS: the supervisor tests' inert child builds its worker only once connected, so its 250 ms initial-authority clock does not run during the connect #1330's after-bind window as its own PR, then the room-mic design for "Lamp, stop." over the lamp's speech, on "Stop" said while the lamp talks stops it at once, every time (1 of 7 short stop words on 14 Oct) #828 before code (lampOS builders' queue #705 16:07).
1. now — Physical skills: lamp-skills, the skill format, and Welcome back as the first skill #938: lamp-skills, the skill format, Welcome back as the first skill (first on lamp-4ace 17 Oct)
2. lampOS: a full look when asked what the lamp sees, behind --full-look and LAMP_LOOK_FULL (#836) #842 — the full look on lamp-4ace (eyes since 03:43, with lampOS: let the eyes give a slow camera 2 s to start and 250 ms to stop, and log how long each open took (#911) #921 since 04:53): the first full-size still was lost between the eyes and the voice (03:57, 1 of 2), and the head wait hits its 2 s cap every time because a standing shoulder rests 22–34 counts short of its goal: settle on movement only (lampOS: a full look when asked what the lamp sees, behind --full-look and LAMP_LOOK_FULL (#836) #842 comments 03:24, 03:59)
3. lampOS: --zone system in Expression, Memory and the policy (#820) #844 — read the Memory part of the third's --zone system (9df579365, a355e3ef9, b3c3ca47e): the journal's day at local midnight; the first builder's lamp check follows after 08:00
4. What a person teaches and corrects is kept and obeyed in every later session #813 — kept rules when the lamp speaks first: lampOS: type a kept rule about the greeting or unasked news into memory/rules (#813) #904 merged 04:03; the second's Policy reader (--rules) and the end-to-end check with him
5. The first five minutes: the lamp walks its new person through where to sit, their name and face, their voice, touch, privacy and giving way #841 — onboarding (lampOS: the first five minutes: the lamp walks a new person through where to sit, their name and face, and preferences (#841) #858): the first builder walks it with the owner on lamp-4ace, 2026-10-16 11:00; fixes from that walk
6. "What do you see?": take a full-size picture after turning to look, and say when it can't make the thing out #836 — the full look's issue: closes with lampOS: a full look when asked what the lamp sees, behind --full-look and LAMP_LOOK_FULL (#836) #842

Done: #733 — 41 (pull request #769, merged 14:28: the hand-over played on lamp-4ace at 13:43 with the switch off, Codex checked by hand at 13:53; the switch on waits for the owner) · #764 — 56 (pull requests #771 and #775, merged; #775 at 14:28, the two-minute rule)

eighth (new, 2026-10-15; lamp-75c9)

0. now (17 Oct 16:2x) — tonight's readings after midnight (lampOS: the voice ends at midnight when the lamp's daily log jobs run (lamp-4ace, 17 Oct 00:00:06) #1178): the stall file at 00:05 and lampOS: a voice mic waited for at start is one plain line in the journal, the session's events and the voice's health until it starts again (#1178) #1326's mic line on lamp-cb94 (Twelve, 16:03). Then item 2 (eighth-P1: a lock kept on each lamp, so builders on any machine take turns at it #833).
1. now (17 Oct 07:2x) — tonight's midnight on lamp-4ace with lampOS: LAMP_STALLS=on records each stall of the voice's loop and the ring worker with what the lamp was doing (#1178) #1179 (a8444a116, merged 07:16) and LAMP_STALLS=on: write the reading of the stall file for 00:05 (which worker held, how long, what the lamp's counters said), a short doc section or a --summary reader, so the first builder posts it on lampOS: the voice ends at midnight when the lamp's daily log jobs run (lamp-4ace, 17 Oct 00:00:06) #1178 at 00:05
2. eighth-P1: a lock kept on each lamp, so builders on any machine take turns at it #833 — the lamp lock (pull request lampOS: a lock kept on each lamp as a 10-minute lease, so builders on any machine take turns at it (eighth-P1) #834): its ARM64 run is red; fix it and get it reviewed, so builders on any machine take turns on a lamp
3. A flaky lamp-policy test turns every ARM64 run red (program.rs:363) #783 — qualification: the tests builder took the flaky-test sweep (lampOS: drain complete test journals before killing health and memory #839 merged); keep only what is yours there

Done: #704 (the lamp weighed and measured, 2.686 kg) · #739 (closed 17 Oct 07:0x, #900) · #1179 (the stall recorder, merged 07:16)

fifth

0. now (17 Oct 16:2x) — lampOS: Expression's sway between sentences turns the base about 52 counts instead of rolling the head #1067 on lamp-52e6 (step 4b, the old sway_roll catalogue starting; post it on lampOS: Expression's sway between sentences turns the base about 52 counts instead of rolling the head #1067), then Sunday's body sheet with the owner (4a, 4b; lampOS: a standing lamp let go with its elbow fallen lowers its shoulder when taken up and stands again when next asked #1115's step 4).
1. today (16 Oct) — the canceller takes only 3–19 dB off the speaker tone (lampOS: the self-test's speaker tone in lamp-voice (#812) #895's selftest/speaker, falling within one voice run): find why (reference timing, mic2's DSP, the two USB clocks) and lift it toward 30 dB; it is the root of the lost stop words ("Stop" said while the lamp talks stops it at once, every time (1 of 7 short stop words on 14 Oct) #828)
2. lamp-4ace: the ring stops for good after one late SPI write, and the camera won't reopen for minutes after a close #911 — the ring stops for good after one late SPI write (SD stalls, The voice dies when the lamp's SD/MMC host re-tunes (a 0.3 s I/O stall) #784: 17:10, 23:21, 03:26; the light can't be asked for until the voice restarts), and the camera won't reopen for 2–18 min after any close (eyes or voice restart)
3. lampOS: let the body say on body/hand whether a hand is pushing it now (#711) #848 — body/hand for the head pads (33. The base's touch strip and the head's two pads: read, slide, and new meanings #711): the reviewer's P2 of 2026-10-15 21:50 (the hand's age after the preceding publication)
4. lampOS: give way to a hand and slump as V1 by default (--give-way off for the body as before) (#703) #857 — give-way and limp on by default: the first builder's minute on lamp-4ace with the owner pushing, 2026-10-16 11:00, then merged
5. 54. The voice comes up after a power cut without restarting #762 — 54. The voice comes up after a power cut without restarting: the owner's power cut on 2026-10-16 morning checks lampOS: a late privacy observation no longer ends the voice's start-up wait (#762) #835 on lamp-4ace
6. lampOS: let a joint go limp under a hand as V1 did, behind --limp on (#703) #780 — the follow-up asked there: a slump from a hand rests seconds, not two minutes

After the MVP: #720 (seeing someone speak): the fourth is adding Face.speaking to the eyes for #829 (2026-10-15 21:24); review its eyes part. #763 (a false face): pull request #766 merged 14:15, on lamp-4ace since 13:48; closed after tonight's empty office. (#704 moved to the eighth.)

Done: #715 — 37 (pull request #724, merged 01:00; on lamp-4ace with --look-at-name on since 00:57) · #710 — 32 (pull request #718, merged; on lamp-4ace since 18:41) · #719 — 38 (pull request #726, merged 23:00; its body on lamp-4ace)

twelve (no lamp)

0. now (17 Oct 16:2x) — lamp-cb94: the after-midnight reading counts from 09:02:20 on 668378ff1 (lampOS: a refused send names its socket, and a held-up speaker ends the capture worker unless reference recovery is bounded (#1178, lamp-cb94) #1298 with bounded), posted on lampOS: the voice ends at midnight when the lamp's daily log jobs run (lamp-4ace, 17 Oct 00:00:06) #1178.
1. now — lampOS: coordinator reminder test misses AckConfirmed after one ACK request #935 as lampOS: the reminder tests on real Stores run on the stepped scenario clock, for #935 #1009, ready: answer the reviewer until it merges
2. The provider worker's twenty_second_answer_through_a_five_second_reply_buffer_follows_capacity (ReadTimeout, then ConnectionReset at its unwrap; red in the reviewer's lampOS: voice: an answer's first words ask for the V1 moves too (#933, #940) #985 gate, 11:42): deterministic from its cause, one pull request
3. Then ask the first builder on lampOS builders' queue #705, with what you found in the other red voice tests of the day

el (lamp-06c1, with a person in the office)

0. now (17 Oct 21:4x) — Lamp V2 as its own device in the Autonomous app: online over MQTT, a device page, a lamp-v2 type #1396: review Nine's write-up of V1's MQTT contract against lamp-06c1's V1 behaviour (after the lampOS: with serve --close-told, lamp-tasks closes the harness it started once its answer is told and a read just before finds it still idle on that turn (#1390) #1395 peer review).
1. now (17 Oct 16:2x) — tonight: the test-signed channel from the integration commit's green ARM64 run (Setup for a home: V2 ships with V1's setup flow (Wi-Fi, service, name), and a 5 s reset #806, 15:5x). Now: is lamp-06c1 in the office on Sunday, and B7 on it, read-only (Setup for a home: V2 ships with V1's setup flow (Wi-Fi, service, name), and a 5 s reset #806, 16:13).
2. now — lamp-06c1 onto V2 (hold until 13:30), then set up the way a home is: One install for a new lamp: every part, its unit and its settings, from a signed build #811's install, then lamp-setup with a phone (Setup for a home: V2 ships with V1's setup flow (Wi-Fi, service, name), and a 5 s reset #806); every step that needed an engineer written down
3. Physical skills and V1 moves on lamp-06c1 with a person: Welcome back, Stretch break, the moves (El) #995 — Welcome back, Stretch break and the V1 moves on lamp-06c1 with a person
4. Lamp time for Jarvis's Physical skills: the Dance skill ("Lamp, dance"), with V1's dances (jarvis) #999 (lampOS: Dance, a skill the model calls with a mood and a length (#999) #1012) after the hold

jarvis (lamp-06c1, shared with el)

0. now (17 Oct 16:2x) — Sunday's sheets (Setup for a home: V2 ships with V1's setup flow (Wi-Fi, service, name), and a 5 s reset #806): fold in 16:13 and 16:16 (the radio scripts, drop-ins aside, the Policy key's fallback) and the integration commit once it's named on lampOS builders' queue #705.
1. now — Physical skills: the Dance skill ("Lamp, dance"), with V1's dances (jarvis) #999 Dance (lampOS: Dance, a skill the model calls with a mood and a length (#999) #1012): its lamp run after El's hold ends at 13:30
2. Physical skills: Task news and Stuffy air (from #939 and #941) #1021 — Task news as a skill
3. Physical skills: Task news and Stuffy air (from #939 and #941) #1021 — Stuffy air as a skill