Repository navigation
EPIC: night loop 2026-10-19 -- research, plan, tri loop, coverage, report #11063
Description
Owner, 2026-10-19 night (translated): research the weak points and the competitors, make a decomposed plan, implement all of it, end with a report and three collaboration options for the next loop; run autonomously all night; one continuous improvement loop that never breaks earlier work; self-critical, find anomalies and repair itself; new tri commands and skills; every task a GitHub issue; coverage to 100%; auto-fix on and merge when green.
Goal
The analog-emulation and lab work (epics #10486 and #11040) is weakness-checked, benchmarked against competitors, decomposed into issues, implemented, and left with a loop that can run again without repeating or undoing anything.
Plan
Phase 1, research: #11057 lab and network competitors; #11058 analog weaknesses against measured evidence.
Phase 2, speed and experience in tri: #11059 tri loop plan; #11060 tri learn; #11045 seal and t27b speed; #11020 tri remote; #11054 proactivity rule.
Phase 3, product: #11041 and #11042 lab decisions (t27c, t27b); #11043 lab job kind; #11044 installer, README and TRI CLAW; the open analog items of #10486 (#10942, #10881, #10882, #10911, M5 second half, M6 release).
Phase 4, coverage: #11061 tri coverage, gaps closed or filed.
Phase 5, report: #11062 night report and three options.
Loop contract (what keeps a new cron cycle from breaking earlier work)
- State lives in the ledger comment of this issue, never in a session. A tick reads it first.
- A lane listed as running is not started again; a row older than 45 minutes is rechecked, not trusted.
- Every PR action checks the PR state first and is idempotent; no force-push, branch updates go through update-branch or a merge commit.
- A blocked item gets a request with a default and a due time (CLAUDE.md "Proactivity: request, do not ask"), never a parked question.
- Each tick ends with Requests out, the anomalies found, and what changed in skills or tri.
- Merge only when green; a red is inherited only if it is a subset of master's failures.
Definition of done
All children closed or consciously deferred with an issue, coverage fraction printed with numerator and denominator, report and three options published, ledger row final.
Activity
gHashTag commentedon Oct 19, 2026
Loop ledger
Night loop of 2026-10-19. One row per tick, newest at the bottom. A tick re-reads this comment first, edits only its own row and the status table, and never starts a lane this ledger lists as running.
tri loop planprinted the six expected actions on a test snapshot, repeated output identical, NOOP with the last hash;tri seals driftread 0 moved on the clean ledger in 25.69 s, namedmoved: gen_hash_verilogfor a planted verilog drift andmoved: spec_hashfor a 1-byte edit,--fixresealed exactly that 1 spec; t27c rebuilt after the loop-plan lane's dead-function deletion (176.10 s wall under load 40; 47.27 s unloaded earlier); owner connected the AX7203 and the FPGA HDMI capture: board regression issue #11074 filed and its lane startedfpga-lint(1/60 modules fail read_verilog or hierarchy) andfpga-conformance, traced to master by the merge-rules session: specs/fpga/systolic.t27 (merged in #7719, tracked as #10780) makes the emitter write an illegal$signed(...)assignment target and yosys stops with a syntax error; it was red before the gen-verilog merge too, and the workflows trigger only on bootstrap/** and cli/**, so a PR shows them only when it touches those paths: inherited (my first reading, that #8844 caused it, was wrong)tri worktrees --reap-planlists as mine are the WIDGETS session's lanes (commit subjects 'widgets: ...', one committed 11 s earlier): hands offblocked; tri learn (#11060) lane started in the free CLI-build slot; self-critique filed as #11081 (tri worktrees printedminefor another session's lanes)scan(the guard against developer home paths): the document named the bench Mac's home directory in its toolchain line; filed #11083, fixed it in PR #11084 (one line, auto-merge on) and made the landing script refuse such a path before a push; checked the other files of tonight on master: no other occurrencescanis not a required check, so auto-merge does not wait for it: every non-inherited red of a merged PR has to be read, not only the required ones; load average 65 on 8 cores, memory free 30 %, disk 7.8 GBtri coverage <glob> --t27c target/release/t27c: analog 75/75, lab 55/55, board 5/5, remote 5/5, install 5/5 = 145/145 (the earlier unmeasured columns came from a scratch tree without target/). Master turned red oncoverageandspec-guardsat #11094: one seal (ml_host_io_model) missed by the emitter PR; PR #11105 sets that one hash (evidence in its body); root cause filed as #11103: on a Mac the two specs using host_open do not compile under Zig (a bare literal to variadic open), soseal --saverefuses; the one-line emitter fix needs the owner's word on FROZEN_HASH. A bug in my owntri learn: a skill file with a fence line like "``` numbers" made three Traps headings; fix filed as #11110 (PR in progress), the lessons PR #11108 is held with auto-merge off until it lands. X chat: his answer at 4:56 PM UTC says the GPU machine is not ready; acknowledged at 5:13 PM UTC, no plan for ittri learnbug abovecoveragecheck was red on the same stale ml_host_io_model seal because its base predates #11105 (the master runs after the seal fix were cancelled by newer pushes, so no green master run seen yet; recheck next tick); #11108 (lessons) has no red check, 6 pending. The charge-table lane runs two containers (ff, fs) next to the kick scan: six cores in all, the stated limitcoverageandspec-guardsread success again after the seal fix and the fence fix (the earlier cancelled runs were superseded by newer pushes). The charge-table lane finished all its remote decks (ff, fs, sf, tt, ss, inverter references) and is writing the spec; the kick scan (late aperture) is at 43 min on the remote host; removed the merged scratch trees of the lessons and fence lanes; night report draft filled except the two lane resultscoverageandspec-guardsread success on the newest finished runs. Machine inside limits (load 5, disk 6.3 GB, VM volume 5.0 GB), so a second analog lane started: #10881 (weak-inversion read, ngspice on the remote host with 2 threads, the VACASK lane keeps its 4)t27b test --check(the JIT passes, the interpreter faults on an 8-byte read of bytes never written), cause nine[3]usizescratch arrays declared without a value, now zero; PR #11136 (14 tests, --check exits 0, 1824 runtime asserts), the peer session told; my landing script now runst27b test --checkon every spec (a class of defect my L2 landing missed). The weak-inversion lane (#10881) is runningt27b test --check; a peer found itt27b test --checkmismatches 0 on all ten, seals re-saved, types CLEAN): the log-domain reader meets the 1 percent bar at the 576 midpoints (0.080 and 0.094 %), the 1000 held-out points (0.398 and 0.527 %) and the cell centres (0.467 and 0.705 %); the linear reader stays at 0.5 to 0.73; inverter delays within 0.025 % of ngspice backward Euler at 500 fF and within 0.0033 % at 10 pF. All three analog lanes of tonight are done; no lane running; trees of merged lanes removed (disk 5.1 GB)t27b test --checkmismatches 0, seals re-saved): the fitted converter's 8-conversion trace (time, the two plate voltages, the code) is written by the RAW writer and ngspice-47 reads it back with largest difference 0 on all 32 doubles of the binary file; a flipped data byte is reported by point and variable, a wrong point count is refused (by the spec's reader and by ngspice: "bad rawfile"); the ASCII file differs by 1 ulp on 4 of 32 values (16-digit format), stated in the PR. Disk 4.4 GB after the landing, zig cache cleared; a lane for S5 of #10942 (the late decisions of the converter kick: the bit path and decision 10 against 9) started, at most 2 containers of 2 threads on the remote hostpr-dashboard("unexpected end of JSON input" in the dashboard script, an infrastructure error unrelated to the change). No lane running; disk 5.6 GB, load 4.3. Nothing left in the analog list that fits the machine now (the board build needs 8 GB free; the external team's checks need their input)accept_sar.t27, each of the 13 SAR blocks measured from our own models and given PASS, FAIL or NOT MODELLED, never a pass without a number) filed and a third lane started; a written notice with a stop switch went to the contact in the chat (2:40 AM UTC) saying what is being done; the memory note on the private limits updated. Three lanes run: M3d board #11208, M3c-4 tables #11211, M5c checks #11212make installinstalls only the CLI (EARN=on, the lab node, is refused). A lane (agent a35c67b42f0659275) builds the app locally from that branch, launches it and finds how to drive its panes (docs/t27os/run-local.md, Refs #11037); conducting throughtriitself started (tri lanes,tri night,tri land --planwork;tri loop planwaits for a state file). #11214 (M3c-4) and #11215 (M5c) merged, #11211 and #11212 closed. M3d landing hit one conflict in mos_psp103_data.t27 (the board branch and #11214 both appended tests), resolved by keeping both tests; the landing resumed from the tests; dupe_scan then found the board top's copy of hex_at (a UART-capture parser of the SAR top): imported instead (use + pub), the emitted Verilog loses only that unused function (41 lines, none added), M3d landed as PR #11220 (13 specs tested clean, vacuous 0, seals re-saved, types CLEAN, auto-merge on). Load hit 85 (limit 40) during the landing plus the app build: the lane was told to use at most 2 parallel jobs and to wait for load under 20open -astarts all five (ZCode needs about 7 s cold; my first probe polled too early); editors.t27 knows 9 names and only 2 of those five; the Claude app's own Open-in menu knows only VS Code, Cursor, Windsurf, Zed (Cursor opened a file). (c) Window: matte black glass in a window, solid #000000 in full screen (the opaque backdrop fill of 10-19 hid the glass); editors open INSIDE the window as a pane; issue #11223, Sonnet lane started. Both lanes are stacked on claude/t27os-keys-v3 and push a branch only (no PR against master). Disk fell to 1.4 GB free at 10:55: freed my merged trees, the zig cache and four scratch dirs (2.3 GB), then the t27o session freed 2 GB at my request (4.1 GB); VM volume 6.0 GB used, load 10tri landhas not27b --checkstep and verifies seals only for changed specs, not for specs that import them (each let a defect through tonight). Both T27 OS lanes still running (13 and 12 minutes, below the 45-minute recheck); their branches are not on origin yet. Disk 5.0 GB free, VM volume 6.0 GB used, load 13.5: no emergency, no new lanet27b --checkexit 0, 0 mismatches; types ratchet CLEAN; dupe_scan 0 failing; no partner names, ASCII, no home paths. PR #11226 opened against the stack base claude/t27os-keys-v3 (PR #10977) and MERGED by the repo's merger 3 seconds later with 18 checks pending (a PR into a non-master base is merged at once, so all gates must run beforegh pr create: land_stack.sh does). PR #10977 now carries the new commits and re-runs its checks (30 pending). DeepSeek, Moonshot and MiniMax have documented Anthropic-protocol URLs but need PROVIDER_COUNT 9 -> 12, which touches glass_window.t27 and pty_server.t27 (lane #11223's files): left for a follow-up after #11223. Wiring left (window and server read the pane's error row, per-pane bad mask and failover count, key number to the respawn; the start-up loopkpre < 2must run to PROFILE_COUNT) is also for after #11223. The #11223 lane (glass, editors inside) is still running at 25 minsame_bytesin editors.t27): the lane gave it a different body (accumulate the XOR difference; a mutation of that line fails 2 of 21 tests) instead of blessing. Gates on the merged tree: editors 21, pty_logic 17, queen 9 tests, vacuous 0;t27b --checkclean on nine specs; types CLEAN; dupe 0 failing. Content:backdrop_alpha(full)(0.0 in a window, 1.0 in full screen), BLEND_CLEAR wipe, TINT_FULL 1.0; editor table 9 -> 27 (Antigravity, Antigravity IDE, ZCode, the JetBrains family, and others), Xcode through xed, the failure reason shown in the OPEN IN strip, strip 3 chips of 88 pt;editor_command/term_pick/editor_spawn_payload,wants_editorandlaunch_commandin pty_logic,open_target(is_god, term_found, ext_found)in queen. HOT RELOAD verified: the window binary built from the merged tree replaced the running app's executable at 11:23:30; the same pid 12779 re-checked in 20 s, the PTY server (pid 13403) and its panes untouched; a screenshot shows the translucent glass window (the wallpaper shows through the 0.30 tint). Full screen #000000 not verified live (the owner's green button). The lane resumed on PART C (the window call site: a T27 OS chip in the OPEN IN strip, the editor pane view, the Queen's A_EDIT with a once-per-space guard); the keys wiring (failover reads the pane error row, PROVIDER_COUNT 12) follows after it, same filesgit merge-treeshowed ONE conflicting file, docs/reports/t27b_expectations.json (two hunks: max_not_pass 133 vs 129; the t27os rows vs master's install.t27 row). Resolved in a scratch tree: max_not_pass 132 (master's 129 plus the three blocked t27os rows pty_client, pty_server, pty_sys; the not-pass count minus the cap stays 3 as on both sides), rows in path order (install, pty_client, pty_server, pty_sys); jq valid, paths sorted, types CLEAN, no markers; merge commit cbfac4e pushed to feat/t27os-petal-menu (a merge, no force push), so #10783 re-runs its checks with auto-merge still on. PR #10977 (base = the petal branch, labelno-auto-mergeset by someone) is left alone: no auto-merge enabled on it, the label is a hold; the stack reaches master through #10977 only after #10783 merges. Part C lane (aef393e9) still running (about 15 min)no-auto-mergeon #10977: after #10783 merges, retarget #10977 to master (default: I retarget it and re-run its gates in the next tick)security -istdin) and adds several keys per provider in the KEYS tab. From the owner's screenshot: the terminal area and KEYS panel are opaque black in windowed mode, the panel shows the terminal through it, the foot rows overlap: lane claude/t27os-glass-all (glass everywhere windowed, black everywhere full screen, panel narrows the terminal, foot rows)Status table