← Research · Knowledge graph: EAGLE4 knowledge graph · Collaboration: collaboration and restart recovery
Agentic session history: EAGLE4 HQ (build-host-A), 2026-07-16 to 2026-10-11
Calculations and hypothesis
Not measured. Every [D] derived and [P] projected item and every hypothesis on this page is here; everything below this bar is a measured or recorded fact.
Calculations and hypothesis (every [D]/[P] item on this page)
| # | item | value | label | evidence |
|---|---|---|---|---|
| H1 | 3STAGE-LINE batch-1 step rate (A attn / B MLP / C LM+tree on 3 cards) | 384–396 tok/s | [P] | F1170; 2026-10-08 three_card_v3/proof/VERDICT.md |
| H2 | 3STAGE-LINE plus host lever Y (FIRST_DRAFT_GPU) | 416–433 tok/s | [P] | F1170 |
| H3 | lab-host's own LINE projection (raw / x0.5-calibrated) | 368–375 / ~359–362 tok/s | [P] | lab-host handoff 2026-10-11 (from_lab-host/2026-10-11/) |
| H4 | 3STAGE per-level speed-up | x1.48–1.69 (w2), x1.80–1.90 (w8) | [D from xsim M] | proof/VERDICT.md |
| H5 | EMBFOLD on built card A (0x121B) | 390–419 tok/s | [P] | F1160/F1161 |
| H6 | Host-glue levers below the +1% bar (X JOIN_FUSE) | +0.39..+0.46% calibrated | [P] | BR185 |
| H7 | On-card speculative decode (HBMS, 4 cards) | 573–634 tok/s | [P] | BR170, ON_CARD_SPEC.md |
| H8 | HBMS PSTEP spec at 1K/8K/32K context | 614 / 242 / 79 tok/s | [P] | BR160 inc6 |
| H9 | Step budget split (AC8 steady step median 8.62 ms): verify 43%, card chain 28%, post-verify 15% | 8.62 ms | [D from M] | BR181 STEP_BUDGET |
| H10 | 3-card derived step | 7.07–7.24 ms | [D] | F1169 |
| H11 | s3LCr route-only resume saved build time vs a from-scratch rebuild | ~10 h | [D] | HQJ_CATCHUP_DIGEST C1 |
| H12 | Loss of peak build capacity when build-host-A's DDR5 was returned | ~20–25% | [D] | HQJ_CATCHUP_DIGEST F (10-08 19:16) |
| H13 | 4-card plans (SIBX/CL4/D2) | 329–368 tok/s range | [P] | F0574–F0622 (several were later corrected by F0584/F0588) |
| H14 | Hypothesis: further batch-1 gains come from card A per-item latency and inter-card communication/loop overhead, not host glue | — | hypothesis | hq-junior-2, 10-09 16:25 (digest C2) |
| H15 | Hypothesis: boot units fork sessions because the resumed session keeps its own saved options | — | hypothesis backed by journalctl text | digest G1 |
A curated history of the Claude agent sessions that ran the EAGLE4 V80 FPGA drafter program from the build-host-A HQ.
It is not a transcript. Everything below was built from the on-disk records: handoff docs, QM_LOG, NIGHT_LOG,
USER_QUOTES.md, HQJ_CATCHUP_DIGEST.md, findings.csv/branches.csv, and the transcript metadata (titles, ids, timestamps
and subagent descriptions only). No transcript text was copied, and credentials and addresses are left out.
Times are EDT unless marked Z (UTC). Machine-readable companions: sessions.csv and timeline.csv in this directory.
Labels: [M] measured on silicon, a tool or a log, with evidence. [D] derived from measured values. [P] projected. Per the user's page rule (2026-10-11 02:24), every [D] and [P] number is in the first section. The body has only [M] facts.
1. The agent organisation
1.1 HQ main lineage (the session that runs HQ on build-host-A)
| era | session | name / tag | role | dates (EDT) |
|---|---|---|---|---|
| pre-HQ | 2a147f99 | "hls-synthesis-gemm-coverage" | HLS/GEMM synthesis-coverage era main session; 367 dispatched agents | 07-16 → 09-19 |
| pre-HQ | a1ee9772 | "cleanup-eagle-fork" | RTL fork/cleanup main session; 1,625 dispatched agents | 07-28 → 09-22 |
| senior HQ | 87e06065 | "october-sprint-orchastrator" (senior HQ) | First HQ main: built the mesh, QM, findings ledger and the 307.67 record; 191 agents | 08-27 → 09-27 |
| senior copies | 5b4657b1, 367a3e53, a0e9b491, f1c0a6a7 | same title | Resume copies and snapshots of the senior HQ (5b4657b1 is the one hq-junior took over from; it ran until 09-29 03:11Z) | 09-25 → 09-29 |
| hq-junior | 25bceecd | "hq-junior", tags [06d0a4], [9389fe] | HQ main from 09-27 21:54 to 10-07, and again from 10-08 04:05. Parent of QM-1..QM-26 and QM-29, and of 194 subagents | 09-27 → live |
| hq-junior-2 v1 | bff4a930 | "hq-junior-2" | Second HQ session (10-06). Its main transcript was never saved, so it cannot be resumed. Parent of QM-25 | 10-06 → 10-07 02:5x |
| hq-junior-2 | 5b504db2 | "hq-junior-2", tags [b8bf6c], [11141a], [d72b0d] | Relaunched 10-07 after a subscription switch. HQ main 10-07 → 10-08 04:05, then architecture design lead and the user's interactive session. Parent of QM-27, QM-28 and 51 subagents | 10-07 → live |
| boot copy | 5a5334a3 | "cron schedule recreation", [ee6850] | Boot-unit copy of 5b504db2 (10-08 22:35). De-facto HQ ops since 10-09 02:11: every cron, cold-AC consent-hq, hq_daily rebuild/commit, NIGHT_LOG | 10-08 → live |
| boot copy | e5cb5a97 | "trace package build readiness", [fe18d7] | Boot-unit copy of 25bceecd (10-08 22:35). QM parent for QM-32..QM-37; runs no crons since 10-09 02:11 | 10-08 → live |
| boot copies (stood down) | 780e9140, 05d08cec | copies of hq-junior | Spawned by the 21:53 and 22:20 reboots on 10-08; stood down | 10-08 |
| boot copies (stood down) | 1265cff6, 156a90bf | copies of hq-junior-2 | Spawned by the 21:53 and 22:20 reboots; stood down | 10-08 |
| hq-heritage | abde3096-a629-44ea-9758-1bc8f289eab9 | "hq-heritage" | Named archive fork of hq-junior's lineage, saved 10-11 02:49 (06:49Z) | 10-11 |
1.2 The quartermaster (QM) chain
The QM is a background subagent of the HQ main session. It owns the build mesh, the build queue, the races, adjudication
(0 new FAIL classes vs the parent, WNS/WHS ≥ 0) and the relay to lab-host. One QM is on duty at a time, in shifts of about
12 h, and each shift ends with a QM_HANDOFF_<n>.md. Log: <path:QM_LOG.md>.
| QM | agent id | parent session | shift start (EDT) | how it ended |
|---|---|---|---|---|
| senior QM | a411aee8125baaee7 | 5b4657b1 | ≤ 09-27 | handed to QMJ (handoff_QM_a411.md) |
| QM-1 (QMJ) | a6a66bfd103aab04c | 25bceecd | 09-27 21:56 | API limit |
| QM-2 (QMJ-2) | a8de84794e4876811 | 25bceecd | 09-28 00:06 | API limit ~04:30 |
| QM-3 (QMJ-3) | ad7af2001d40beb61 | 25bceecd | 09-28 05:13 | handoff QM_HANDOFF_4 by QM-4 |
| QM-4 | abbd82b2ea22a7685 | 25bceecd | 09-28 06:23 | QM_HANDOFF_4 |
| QM-5 | a9effde5ecf2a0427 | 25bceecd | 09-28 06:33 | QM_HANDOFF_5 |
| QM-6 | a4bf47997180b1291 | 25bceecd | 09-29 01:08 | QM_HANDOFF_6 |
| QM-7 | ad54442469ac12115 | 25bceecd | 09-29 11:03 | QM_HANDOFF_7 |
| QM-8 | ab8a1e5c07d00266c | 25bceecd | 09-29 23:25 | died ~00:00 (API limit); fleet unattended ~11 h |
| QM-9 | aaca8e993ead4472a | 25bceecd | 09-30 11:17 | QM_HANDOFF_9 |
| QM-10 | ad4dd3452aff370ea | 25bceecd | 09-30 17:36 | QM_HANDOFF_10 |
| QM-11 | a25e7fa7a7d8ea834 | 25bceecd | 10-01 00:40 | QM_HANDOFF_11 |
| QM-12 | a79ec16bc592c4b5e | 25bceecd | 10-01 02:03 | QM_HANDOFF_12 |
| QM-13 | abbc99a3e468736ce | 25bceecd | 10-01 06:40 | stopped ~09:10 on classifier errors |
| QM-14 | a5109085fa93c48fa | 25bceecd | 10-01 09:26 | finished |
| QM-15 | ae2a8e1cf8face665 | 25bceecd | 10-01 14:21 | QM_HANDOFF_15 |
| QM-16 | a824ef807adee49dc | 25bceecd | 10-03 07:22 | QM_HANDOFF_16 |
| QM-17 | a9cbfe274e8c5bc4c | 25bceecd | 10-03 18:09 | CL4 set owner (split from QM-16); done 10-04 |
| QM-18 | a7ac8fd6b1ae90ab9 | 25bceecd | 10-03 21:58 | QM_HANDOFF_18 |
| QM-19 | a1690d93c173de307 | 25bceecd | 10-04 10:21 | weekly API limit ~13:00 10-04 |
| QM-20 | ad434e7f30dbcffdf | 25bceecd | 10-05 02:12 | QM_HANDOFF_20 |
| QM-21 | ae0688461682c0b31 | 25bceecd | 10-05 03:20 | QM_HANDOFF_21 |
| QM-22 | ae250b98d826418b9 | 25bceecd | 10-05 16:15 | QM_HANDOFF_22 |
| QM-23 | a08aae838973448a9 | 25bceecd | 10-06 04:11 | died ~07:30 (session limit), no handoff |
| QM-24 | a6a6b7d5fb07e03bb | 25bceecd | 10-06 08:20 | QM_HANDOFF_24 |
| QM-25 | a735bb7854c81e1b4 (Sonnet) | bff4a930 (hq-junior-2 v1) | 10-06 ~20:00 | died at the weekly limit, no handoff |
| QM-26 | aed49eb2698a26d1e | 25bceecd | 10-07 02:52 | QM_HANDOFF_26 (QM duty passes to hq-junior-2) |
| QM-27 | af6105f16611be50f | 5b504db2 | 10-07 11:42 | QM_HANDOFF_27 |
| QM-28 | a448f7afe9571f4fb | 5b504db2 | 10-07 23:50 | QM_HANDOFF_28 |
| QM-29 | a4d36c704338a4b70 | 25bceecd | 10-08 08:45 | QM_HANDOFF_29 |
| QM-30 (x2) | a8419bff4cbeb8e86 / a5735a44e6896b645 | 780e9140 / 1265cff6 (boot copies) | 10-08 21:56 | two launched within a minute; hq-junior's copy stood down; QM_HANDOFF_30 |
| QM-31 | ade15b1cf3f42dee3, a917fe7a2ab5f3003 (Sonnet), a6adef4cd5e8dea06 | 05d08cec / 156a90bf (boot copies) | 10-08 22:30 | Opus QM-31 failed at once (HTTP 429); ran on Sonnet; died with the 22:35 reboot |
| QM-32 | a16d4b656b661aaef | e5cb5a97 | 10-08 22:45 | died at the session limit, no handoff |
| QM-33 | add1017c260ca7399 | e5cb5a97 | 10-09 01:58 | QM_HANDOFF_33 |
| QM-34 | a89791912d52a9469 | e5cb5a97 | 10-09 11:40 | QM_HANDOFF_34 |
| QM-35 | a6e3f86c14b798694 | e5cb5a97 | 10-09 19:41 | QM_HANDOFF_35 |
| QM-36 | a16c1a00153ad0214 | e5cb5a97 | 10-10 07:48 | QM_HANDOFF_36 |
| QM-37 | abb37f07912b81128 | e5cb5a97 | 10-10 19:46 | on duty until 10-11 08:00; QM_HANDOFF_37 (living file) |
| QM-38 | — | e5cb5a97 (planned) | due 10-11 08:00 | not yet launched at export time |
1.3 lab-host ("lab-host"), the silicon lab
- lab-host agent: its own Claude session on lab-host. It flashes the V80 cards, asks for cold AC, runs the A/B kits and reports
through
<path:REPORT.md>andTO_HQ.md. build-host-A agents are read-only on lab-host except<path:pdi_inbox>. Its Remote Control address changes after every reboot. On 10-11 it wrote its own handoff (<path:2026-10-11_HANDOFF_lab-host_to_hq_junior.md>, md5 9610d9d7, copied to<path:2026-10-11>). - lab-host KG agent: builds the knowledge graph in
<path:kg>and the page/p/eagle4-knowledge-graph/. - Mac power agent ("Poweroff box 1", on the user's MacBook): the home-control agent that cuts or restores lab-host's AC. It acts only behind the two-party consent and the 3-condition guard (RULE 0).
1.4 Slack hubs
| hub | session | parent | dates | role |
|---|---|---|---|---|
| EAGLE4 Slack hub (1) | tags [e92d45] / [a5c8e5]; probably 5f5c3cc7 (only subagent metadata survives) | fork of 25bceecd | 10-05 | Watches the user's Slack DMs and mentions, answers project questions from the records, applies the reply policy |
| EAGLE4 Slack hub 2 | slack_hub/start_hub.sh; boot copies bf56a44b and 82632c26 on 10-08 |
fork of 25bceecd | 10-05 → 10-09 | Same role. The copies stood down 10-09 and dropped their crons |
| EAGLE4 Slack hub 3 | slack_hub/start_hub3.sh, pid 2828396 |
fork of 5b504db2 | 10-07 10:50 → | Same role, started at the user's request ("clone yourself and copy your session to start the slack hub") |
Slack hubs never drive builds, flashes or cold ACs. They log action requests to slack_hub/REQUESTS_TO_HQ.md. |
1.5 Design agents
All other subagents are design, research or tooling agents dispatched by an HQ session. Each one ends in a report, a flashable
handoff (fork tarball + MD5LIST, uni_pipeline, run script, slot lock, GATE_RESULTS, host kit, README) or a relayed kit.
- 291 subagents of the current-era sessions are listed in sessions.csv (kind design or QM) with id, description,
parent and time span.
- 2,243 older agents (July → 09-27) are indexed in <path:AGENT_INDEX.csv>, with archives under
<path:agents_all>.
2. Timeline (key events; the full list is in timeline.csv)
2.1 HQ-main handoffs
| when (EDT) | from → to | doc |
|---|---|---|
| 09-27 ~21:30 | senior HQ (87e06065 / 5b4657b1) → hq-junior 25bceecd | 2026-09-27/HANDOFF_JUNIOR_20260927.md (+ SESSION_DISTILL_20260927.md) |
| 10-06 ~13:00 | hq-junior (near context limit) → next junior (hq-junior-2) | 2026-10-06/HANDOFF_JUNIOR_20261006.md (+ HANDOFF_PACK/) |
| 10-07 02:59 | hq-junior-2 v1 bff4a930 (no transcript) → hq-junior-2 5b504db2 | 2026-10-07_start_hq_junior2.sh (rebuilt context from disk) |
| 10-07 12:00 | QM duty hq-junior → hq-junior-2 | QM_HANDOFF_26 |
| 10-08 ~04:05 | hq-junior-2 → hq-junior (HQ main) | 2026-10-08/HANDOFF_JUNIOR2_TO_HQJ.md |
| 10-09 02:11 | ops consolidated in 5a5334a3 (user: "keep only the updated on in heartbeat sync") | NIGHT_LOG 10-09 02:13 |
| 10-11 ~01:50 | hq-junior-2 → hq-junior (status handoff) | 2026-10-11/HANDOFF_HQJ2_TO_HQJ_20261011.md + 2026-10-11/HQJ_CATCHUP_DIGEST.md |
| 10-11 02:49 | hq-junior lineage archived as "hq-heritage" | transcripts/session_snapshots/hq-heritage_20261011/ |
2.2 Boot-unit forks (10-08)
build-host-A was rebooted three times on 10-08 (~21:53, ~22:20, ~22:35) for the DDR5 return and an NVIDIA kernel-module fix. The
boot units hq-junior.service / hq-junior-2.service run claude --resume <id> --bg .... Each reboot produced a copy
instead of continuing the same session. Result: 6 HQ copies (780e9140, 1265cff6, 05d08cec, 156a90bf, e5cb5a97, 5a5334a3)
plus 2 Slack-hub-2 copies (bf56a44b, 82632c26). Survivors: 5a5334a3 (ops) and e5cb5a97 (QM parent). The rest stood down.
2.3 Batch-1 record history (headline metric: single-stream tok/s, mt_bench 50x32)
| date | pair + host stack | cell | tok/s | label | evidence |
|---|---|---|---|---|---|
| 09-24 | E1+CF | w2 d2 unprimed, 3 rounds | 178.16 | [M] | F0002 |
| 09-26 | E7+F3 | w2 d2 | 191.52 | [M] | F0003 |
| 09-26 | A5+F4, KVI + FASTGLUE=4 | w2 d2 (1 round) | 268.01 | [M] | F0006 |
| 09-27 | A9b+F4, full host stack (HPQ0, EXTKV, GN, FA4_WARM, PCG) | w1 d2 | 269.65 | [M] | F0282 |
| 09-27 | A5+FA + EAGER + EARLY_A | w2 d3 | 304.39 | [M] | F0335 |
| 09-27 | A5+FA + SPEC0 (FA_HOST_UNIFIED v1.1) | w2 d3, 3 rounds | 307.67 | [M] | F0346 (re-run on a slow boot: 286.74 [M], F0402/F0507) |
| 09-29 | HX1+FSL (HM+E3b HEAD, FS+LUTRAM TAIL) | w2 d3 | 323.30 | [M] | F0511 |
| 09-30 | HX1+FSL + CONS_HOST v1.14 TSP10 | w2 d4 | 327.34 | [M] | F0556 |
| 09-30 | HX1+FSL TSP12Q (TREE_STOP + POISON_PREARM) | w2 d7 | 331.17 | [M] | F0567 |
| 10-01 | HX1+FGI, CONS_HOST v1.17 | w2 d7 | 332.85 | [M] | F0592 |
| 10-01 | HX1 BD + FGI AE, CONS_HOST v1.21 TSP12QL | w2 d7 | 337.22 | [M] | F0669 |
| 10-01 | HX1 BD 0xE4B711BD + TCLK75 TAIL 0xE4B71199, v1.22t | w2 d7 | 339.12 | [M] | F0686 |
| 10-08 | record pair 62 HX1 BD → 42 TCLK75 + host lever Y FIRST_DRAFT_GPU (CONS_HOST stb4g) | TSP12QL w2 d7, 3 rotated rounds | 353.93 best / 350.16 mean at accept 3.65 | [M] | F1125 (lab-host results/2026-10-08/HOSTLEVERS_AB/REPORT.md) |
| References [M]: vanilla H200 without speculative decoding 289.0 tok/s (F0569). GPU-only EAGLE 446–486 tok/s [M] (handoff 10-06; F0214). | |||||
| Batch 2 / aggregate (not the headline): PINGPONG humaneval w1 d4 385.70 [M] (F0306). | |||||
| The pair has not yet beaten GPU-only EAGLE. |
2.4 Key decisions (from the records)
- 09-27: the default host stack was adopted (KVI + FASTGLUE v4.3 + HOSTPUSH HPQ0 + EXTKV + GLUE2 GN + FA4_WARM + PCG + EAGER + EARLY_A + SPEC0).
- 09-28: every user message is recorded verbatim in USER_QUOTES.md and used as the guideline when the user is away.
- 09-30 / 10-01: RULE 0 (lab-host power guard) and two-party cold-AC consent.
- 10-03: lab-host moved to 4 V80s (later 3, after one card was removed on 10-08). H200 at 41:00.0 is never touched.
- 10-06: 4-card SIBX-E is NO-GO on silicon (P 2:7 mean 163.47 [M]).
- 10-08: host lever Y shipped default-on (CONS_HOST v1.23stb4h-firstgpu, BR184). Levers G, J and K are NO-GO on silicon (−2.6%, −0.3%, −3.2% [M]).
- 10-09 02:17: 3STAGE-LINE (BR174) became the #1 priority (user). RULE build-host-A-SPARE: no PDI builds on build-host-A (30 GB RAM).
- 10-09 09:44: the 3STAGE set (A f33ea554 / B dc15487f / C 993e6081) completed on lab-host, md5 OK [M].
- 10-10: LINT-DIRECTIVES v4 (fail-closed Vivado directive lint) after illegal directives cost 9.6 h of build time [M, F1147].
- 10-11 00:3x: user GO; the 3STAGE flash started 01:11 on lab-host.
- 10-11 01:39: user asked for an 8-FPGA ring HBM LLM server design (BR186 RING8).
3. The user's standing rules and research motto (verbatim)
Research motto (2026-10-11 02:24, relayed by lab-host):
"for charts and everything, please post only facts and measure results; hypothesis put together in one top bar named calculations and hypothesis; each research is not to deny it, but an engineering problem to start with the idea given, but through experiment and setbacks, find a way towards a pdi image that is the fastest given the physical conditions; this is the motto of our research"
Standing rules (quotes from USER_QUOTES.md, in date order): - 09-27 ~16:05: "if you confirmed a stale build over the course of two hours or more, you have authority to kill it; it is too wastefull for a machine to not do anything for an entire night" - 09-27 ~16:40: "just by checking the ram to be avaiblae more than 60% is a warning sign to dig in deeper; even as a human looking at it can tell when its over 90% free ram, a deliverabet sign to go check out if a build is happening or stalled" - 09-27: "the quatermaster primary goal is to ensure pdi images are actualy getting build;" - 09-27: "can you save the current session to a new id on this device with complete transcript? compacting is always risky and i loss your behaviour" - 09-27 ~22:30: "by the way the junior should be a new sessions, not just a agent; so it survives when build-host-A reboot" - 09-28: "also most importantly, to advance them and not throw them up; I will review progress when I arrive, not approve for research; research shall be approved on its own when I wil be presented with conclusions to devide" / "becvause we have so many machines avialable; make it its rule" - 09-28 ~02:1x: "rule as all dialogue I say shall be recorded and use as guidelines when I do not respond in time; refer back to my original quote;" - 09-30 23:54 (RULE 0): "now the server must be very very carefully controled; write a very secure enforcement that must only poweroff the lab-host if 1. lab-host agent is offline 2. ssh to box 1 failed 3. box 1 appear offline in mesh all simultaenously; and guard it with a script and always run it; pin this to the top rule of this project" - 10-01 00:01: "as I make cold reboot executeable by home control, now the lab-host shall be authoried to cold poweroff if both lab-host and you hq junior agrees" - 10-01 10:39: "Remember this! I CANT HELP U! How many times have I told you this, im not always at the desk!" - 10-02 02:32: "you have all permission to build pdi images and carry out new ideas; land a pdi thaen we disfcuss" - 10-08 13:15: "You have authority for any experiment: more if you want be safe and explorative Let me know until it makes the flashable pdi image possible" - 10-09 02:11: "please keep only the updated on in heartbeat sync"
Codified hard rules (~/.claude/CLAUDE.md, <path:RULES.md>): RULE 0 lab-host power guard (only gate:
lab-host_power_guard.sh authorize, fail-closed, plus two-party consent); no file removals without an approved plan; never stop
a process you didn't start (stale ≥ 2 h only via stale_kill.sh); never flash or run cards from build-host-A; no credentials in files;
every number labelled M/D/P with evidence. Working style: do it yourself end to end; proposal = launch; fix, don't report;
talk ≤ 3 lines; batch user asks into the user's active window; headline = batch-1 tok/s.
4. Lessons and incidents
| # | when | what happened | fix / rule |
|---|---|---|---|
| L1 | 09-27 | Orphan mesh reservations from crashed builds held 205 GB for ~15 h [M]; the user caught idle hosts twice | Judge health by free -g and live child processes; tools/stale_kill.sh |
| L2 | 09-27 | A gate RED was often a harness bug (HS2 stale slot, KS1 emb base, FA manifest); HS3 had a real RTL bug | Check the harness before touching RTL |
| L3 | 09-27 | A Claude Code restart kills in-process agents and their shell-launched jobs | Detach watchers (setsid/nohup); relaunch the QM right after a restart; HQ runs as a session, not as an agent |
| L4 | 09-28 | The 307.67 record re-ran at 286.74 on a slow boot (14 firmware-first CEs on the HEAD root port) [M] | Boot class is recorded with every silicon number (F0507) |
| L5 | 09-29 → 10-11 | QMs died at API session or weekly limits (QM-8, QM-19, QM-23, QM-25, QM-32), leaving the fleet unattended for up to ~11 h [M, QM_LOG] | Living QM_HANDOFF files; any successor rebuilds state from QM_HANDOFF + QM_LOG |
| L6 | 10-01 | The user was asked to do a cold AC by hand while away ("I CANT HELP U!") | Cold AC goes to the Mac power agent; the user is never asked to press power during off hours |
| L7 | 10-01 → 10-08 | Classifier denials: stopping live mesh jobs (including our own), mesh reservation edits, some mesh submits, systemd unit creation ("Unauthorized Persistence"), Slack file uploads, the lint v5 install into tools/ | Never retry a denial and never route it through another agent; ask the user once for a narrow rule |
| L8 | 10-05 | Another user's gprMax jobs on lab-host invalidated a run (X 4:7 = 50.11) | Those jobs are protected; noted as a confounder only |
| L9 | various | Agents slipped on the no-rm rule inside their own fresh copies; a stray rm -f /dev/null in a QM-35 test (no effect) |
"rsync --exclude, never rm" goes into every brief |
| L10 | 10-08 | Credentials typed in chat ended up in session transcripts; the session copy to lab-host was BLOCKED for that reason | Rotate the exposed credentials (user); never publish raw transcripts; this export is curated and scanned |
| L11 | 10-08 | Boot units forked sessions (§2.2): two QM-30s, 2–3 cron sets, consent ownership handed over ~6 times | Resume without saved-option flags, or point the units at the current copy ids; one ops owner (5a5334a3) |
| L12 | 10-09 | Illegal Vivado directives (AltSpreadLogic_high on Versal) killed both card-C arms after ~4.8 h each | LINT-DIRECTIVES v4, fail-closed before STEP1 |
| L13 | 10-09 | QM-35 edited tools/ before the classifier stopped its backup copy, then restored the file byte-identical | Near-miss on "don't route around a denial"; v5 runs beside v4 from its own path |
| L14 | 10-09 | A shared heavy lock serialized s3LAu3 for 35+ min | Per-slot heavy locks (qm36/perslot_lock.py) |
| L15 | 10-09 | Finished builds kept 75–88 GB of mesh reservations for EVAC census | mesh_core.py phase tooth --pdi |
| L16 | 10-09 | ETA errors: card A was read as 3 h into its build when it was 9.5 h in | Read the actual impl wall time before stating an ETA |
| L17 | 10-08 → 10-10 | The lab-host agent sat idle ~47 h [M, log timestamps] while the 3-card set waited; HQ cannot prompt lab-host | "Prompt lab-host" goes first on the user's list for their active window |
| L18 | 10-09 | Subagents could not write REPORT.md (blocked by the permission system) | HQ writes reports from the agents' hand-backs |
| L19 | 10-10 | Every card-A seed landed at exactly WNS +0.001, because Vivado stops optimizing once timing is met | Over-constraint method: implementation-only clock uncertainty; true WNS = reported + UU [M] |
5. Where every session's history lives
Live project dir (build-host-A): <path:-shared-build--inputs-workspace-eagle4-adaptation-hardware-cleanup-EAGLE>
(<id>.jsonl = main transcript; <id>/subagents/agent-<aid>.jsonl + .meta.json = subagents).
Older project dir: <path:-home-user-workspace-eagle4-adaptation-hardware-cleanup-EAGLE>.
Archive: <path:transcripts> (session_snapshots/, agents/, agents_all/, AGENT_INDEX.csv, session_87e06065/).
| session | archive path(s) |
|---|---|
| hq-junior 25bceecd | transcripts/session_snapshots/25bceecd-...jsonl (to 10-06); transcripts/session_snapshots/hq-heritage_20261011/25bceecd-0d96-42d4-8864-e6d1b8682e62.jsonl + 25bceecd-.../subagents/ (193 subagents) + tool-results/ (full, 10-11) |
| hq-heritage abde3096-a629-44ea-9758-1bc8f289eab9 | transcripts/session_snapshots/hq-heritage_20261011/abde3096-a629-44ea-9758-1bc8f289eab9.jsonl (+ live copy) |
| senior HQ 87e06065 | transcripts/session_snapshots/87e06065_full_20260927_1627.jsonl, ..._snapshot_20260926_1925.jsonl, transcripts/session_87e06065/, transcripts/87e06065-.../ |
| senior copy 367a3e53 | transcripts/session_snapshots/367a3e53-7152-48ef-857b-8fe116c78196.jsonl |
| hq-junior-2 5b504db2, 5a5334a3, e5cb5a97, boot copies | live project dir only (not archived yet). A byte-exact copy of the transcripts as of 10-08 19:09 is on build-host-M (~/.claude/projects/) |
| hq-junior-2 v1 bff4a930 | no main transcript (only subagents/ and tool-results/) |
| pre-HQ 2a147f99, a1ee9772 | older project dir; their agents are in AGENT_INDEX.csv / agents_all/ |
| QM-n | the parent's subagents/agent-<id>.jsonl; narrative in QM_LOG.md and QM_HANDOFF_ |
| lab-host agent | on lab-host (not exported); reports under <path:results> |
Handoff docs, all under <path:rnd_hq>: 2026-09-27/HANDOFF_JUNIOR_20260927.md, 2026-09-27/SESSION_DISTILL_20260927.md,
2026-10-06/HANDOFF_JUNIOR_20261006.md, 2026-10-08/HANDOFF_JUNIOR2_TO_HQJ.md, 2026-10-11/HANDOFF_HQJ2_TO_HQJ_20261011.md,
2026-10-11/HQJ_CATCHUP_DIGEST.md, 2026-09-28/tracks/qm_junior/QM_HANDOFF*.md, QM_LOG.md, 2026-10-01/NIGHT_LOG.md,
USER_QUOTES.md.
Note: the raw transcripts are NOT part of this export. They contain user-typed credentials and stay on build-host-A.
Data files
sessions.csv (one row per session or subagent) · timeline.csv (dated events). Raw transcripts are not published. Paths and host names are replaced with placeholders.