DizyDiz — LLM & Agent Map

8LLM / AI endpoints live

22callers mapped

0callers on a retired port (0 = Peepers parts held for Hermes, OFF)

0retired servers left on RainbowAI (Ollama pool removed 10-03)

✅ Live sweep 2026-10-17 03:08 (hourly) — 0 retired-port references found (0 in running code), 0 not on the map

deep :8080 ✅ · fast :8090 ✅ · img .7:7860 ✅. The sweep reads live config (R610 containers, N8N workflows, Hommer schedules, hambot) and checks every red item below against it. Not decoded: base64-embedded N8N scripts.

Every live retired-port reference is listed below — the map matches reality.

⚠️ Needs fixing — broken pointers (1)

Open WebUI — no admin login in the vault, so an authenticated UI chat isn't verified (API paths are).

🗓 Lanes & rules

Agent work (Peepers, Velma, Louis) → Qwen3.6-35B on the 5060 Ti pair :8080 — fast by default; deep/project callers turn thinking on per request.
Small + fast (chat, judge, vision) → RTX 3060 llama-fast :8090 · embeddings → the same :8090 (nomic-embed-text; llama-embed :8091 retired 10-04) — llama.cpp, OpenAI /v1, Ollama deleted 10-04.
Hommer = desktop + ComfyUI only — its Ollama lane was retired 10-03 (it never served a request).
Hommer by night 21:00–05:30 = ComfyUI image/video batch jobs (the RAM LLM lane was retired 10-03 — zero jobs used it).
External LLM APIs → never used by Peepers or the agents — everything runs in-house.

🔄 What changed since the 07-22 map

10-13: NEW fast caller — Peepers vault tiers builder (Hermes cron, night only 21:05–05:55): L0/L1 summaries of every vault note via :8090 qwen3.5-9b (thinking off, 1 at a time); first full build tonight ~1,700 calls, then only changed notes. Facts checked against the note; credentials redacted.
10-15 21:48: FRONTIER POOL #3 DONE — deep :8080 + fast :8090 RESTORED and verified (POOL OK 128k×2 · FAST OK + 768-d embeds; sweep problems none). MiniMax-M2.7 reviewed 373 Gmail sender clusters (all 201 disputed + 102 recent + 70 others; read-only) and answered the Peepers improvement question.
10-15 17:59: FRONTIER POOL #3 JOB — deep :8080 + fast :8090 STOPPED for the Gmail planner (read-only, nothing changes in Gmail) + Peepers/Krabs improvement questions on the winner MiniMax-M2.7 (Step-3.7-Flash lost: 9.1 tok/s, hard 0.667, coding 0/5 all cut off still thinking; lanes restored + verified 17:57). Forced restore by ~22:14, backstops 22:29/22:34.
10-15 16:19: FRONTIER POOL #3 model 2 — deep :8080 + fast :8090 STOPPED again for Step-3.7-Flash 197B (MiniMax-M2.7 result: 13.0 tok/s, hard 0.733 — thinks even with thinking off, 4 hard answers cut off at the length cap; lanes restored + verified 16:16, sweep problems none). Then the Gmail planner runs on the winner. Forced restore by 20:34, backstops 20:49/20:54.
10-15 14:50: FRONTIER POOL #3 — deep :8080 AND fast :8090 STOPPED on purpose (Jess: "pool everyone"): 7900 + 5060×2 + 3060 + Mac + Dell run MiniMax-M2.7 229B (then Step-3.7-Flash 197B, then the Gmail planner on the winner). Forced restore by 19:15, backstops 19:20/19:25. Peepers (Jess + guest voice), Farm Watch, dictation, imggen QC wait/fail until then.
10-15 14:30: Peepers SPEAKER ID live (Jess: "couldn't tell me from someone else"): ears scores the wake segment on Hommer CPU (ECAPA, ~0.1 s) → voice bridge sends ONLY Jess to Hermes/deep; everyone else (and all Echo/phone asks) gets GUEST answers straight from fast :8090 (new caller), private topics refused in code.
10-15 12:29: FRONTIER POOL run #2 DONE — deep lane :8080 restored + verified 12:27 (POOL OK, 128k×2), fast lane :8090 restored 12:29 by hand (runner bug: stale state skipped the fast restore — fixed) and verified FAST OK + 768-d embed. Flash-Next 125B pooled: 11.3 tok/s, hard 1.000 (deep lane 107.7 tok/s, 1.000). Nothing re-pointed.
10-15 11:55: FRONTIER POOL run #2 (daytime, Jess "pool everyone together") — deep lane llama-pool :8080 AND fast lane llama-fast :8090 STOPPED on purpose; 5060 pair + 3060 + Dell 860M + 7900 (main, Hommer) run Qwen3.8-Flash-Next 125B over llama.cpp RPC. Forced restore by 14:05 (fast verified first), RainbowAI backstops 14:50 fast / 15:10 deep. Callers of :8080/:8090 (Peepers, Farm Watch, vault tiers, dictation) wait/fail until then.
10-14 05:25: FRONTIER POOL test DONE — deep lane llama-pool :8080 was paused 01:32–02:03 only (not until 05:00), restored + verified POOL OK 128k×2, backstop 05:05 found nothing to do, no Hermes asks cut. Pooled 5060 pair + 7900 (Mac busy, not in pool): Qwen3-Next-80B-A3B 54.7 tok/s (hard 0.93), GLM-4.5-Air 41.7 (0.955), Qwen3.5-122B-A10B 32.0 (0.929) vs the deep lane's 107.6 (1.0). Results: dizydiz.com/testbenchmark (Frontier pool). Nothing re-pointed.
10-13 19:10: SCHEDULED — 2026-10-14 01:30 → by 05:15 the deep lane llama-pool :8080 is STOPPED on purpose for the FRONTIER POOL test (Jess 10-13: tie lab GPUs together). llama.cpp b11552 RPC: RainbowAI 2× 5060 Ti (main, test port :18300) + Hommer 7900 XTX (Vulkan rpc-server) + Mac mini M6 (Metal, if free). Watchdog maint on llama-pool (+ Hommer comfyui/imggen-adapter). Runner restores + verifies :8080 itself (backstop 05:05). Fast lane :8090 NOT touched. Deep callers (Peepers/Velma/Louis, compaction, meeting summaries, LeeRoy writer) get no answer 01:30–~05:00.
10-13 eve: LeeRoy went FULLY LIVE — autofeed N8N VW0ETgMnmBAPJAlP (8 AM–2:50 PM, every 10 min) calls Mr Peepers via Hermes :8642 only when it posts (≤1/day, + gate retry) and the 7900 imggen for the picture; the 6:30 PM drafter 4HZvfDLY2IdRytu8 is retired.
10-12 18:30: Peepers Meetings v2 LIVE — NEW fast :8090 caller: live coaching (coach.py, one call every 30 s ONLY while meeting notes run, 10 s budget, failure = no hint). NEW 7900 GPU job: speaker split (ECAPA embeddings in a one-shot ROCm container, ~5 s per pass every 120 s during a meeting + once at the end; never resident). Verified on file-mode test meetings.
10-12 11:40: Peepers Meetings v1 LIVE — peepers-whisper :8793 (whisper.cpp Vulkan, large-v3-turbo q8) NEW on Hommer's 7900; meeting summaries → deep :8080, live ask/recall + dictation clean-up → fast :8090. Tested end to end on a synthetic 2-voice meeting (WER 3.1 %, RTF 0.025).
10-12 10:21: Peepers compaction summaries (Hermes auxiliary.compression, peepers profile) moved from fast :8090 qwen3.5-9b to deep :8080 `deep`, thinking off (Jess). Summary call ~30 s vs 107–126 s on fast; user wait 42 s (was 128/136 s). Hermes vision stays on fast. Verified in a live re-test.
10-12 08:45: cocio-api code default fixed — app.py fallback was the retired :11434/qwen3-coder:30b, now :8080/deep (env unchanged: :8080 ornith-1.5-35b-a3b alias). Image rebuilt, container recreated on R610 .76 with identical settings; verified with a real /api/chat (200 + reply, socket to .195:8080). Rollback image cocio-api:rollback-20261012.
10-09 01:05: llama-pool :8080 STOPPED on purpose (Jess's 10-07 ruling #11: measure stronger local models for Peepers). Maint flag on until 06:05; candidates run on bench port :18091; the pool is restored and verified with a real chat before 05:30 (backstop timer 05:05). Production model is NOT being switched.
10-08 22:09: Farm Watch NIGHT WATCH back — the overnight camera sweep (off since 10-04, retired with the OpenDiz VM) now runs inside the R610 farm-watch container on fast :8090: 5 cameras every 15 min 21:00–06:45, about 20 vision calls an hour on the 3060. First scheduled sweep 22:15: 5/5 cameras checked, ~3 s each.
10-08 19:05–20:01: RainbowAI STRESS TEST (Jess, CPU/memory/disk; PASSED, emailed). llama-pool :8080, llama-fast :8090, sd-service and alloy were stopped on purpose, then RainbowAI was rebooted. All back and verified with real calls at 20:01 (deep OK 0.3 s, fast OK, embeddings 768-d, sd-service 200); watchdog maintenance off.
10-07 07:15: MAP CLEAN-UP (Jess: red dashed lines from the VM, Peepers shown OFF while he is on) — the nine old OpenDiz VM rows are now ONE grey row: that platform is off and replaced by Hermes; the live Mr Peepers is the R610 Hermes row (LIVE). The dead-code references remain listed under Needs fixing as retire-with-the-VM, not as faults.
10-06 10:10: NEW caller — Peepers hands on Hommer (peepers-hands.service, 127.0.0.1:8791): Peepers can look at the Citrix window and click with his own pointer. Each step = one text request to deep :8080 + one screenshot description (and a few icon crops) to fast :8090. Idle = no calls. Verified live from the Hermes container (screen_look read the Citrix window). Vault Projects/Peepers-Hands.md.
10-05 04:0x: LiteLLM RETIRED (Jess: back up the config, then retire). Its only remaining caller, the N8N conversation sweep, was re-pointed to the deep lane :8080 directly and verified with a real run. Container stopped (restart=no), not deleted; config in vault Backups/litellm-retired-20261005/.
10-05 02:32: Hermes now also serves voice: a small bridge beside it (192.168.1.130:8767) answers the desk ears, the avatar and the phone line through the Peepers profile. Vault meaning-search embeddings and camera vision from the lab tools go to fast :8090. Three old-VM jobs (Krusty Krab report, homepage sync, grants digest) run from Hermes's scheduler without a model. The ears service is stopped until the afternoon audio test (leak fixed).
10-04 18:44: Mr. Peepers is back on Discord — now served by Hermes on R610 (.130), not the OpenDiz VM. Same bot, Jess only, three channels (Peepers / Velma / Louis). The OpenDiz VM rows stay OFF and must not be restarted (two brains on one bot).
10-04 18:04: NEW caller — Hermes Agent base on R610 (192.168.1.130, pinned v2026.9.24): profiles peepers / velma / louis call deep :8080 with thinking off; image questions go to fast :8090. Verified with real requests (chat, 62 tool calls, web search). Not connected to Discord yet; the OpenDiz rows stay OFF until cutover. Plan: dizydiz.com/hermes.
10-04 16:10: infra-upgrade-api re-pointed — Velma health check + /health + /heal now use deep :8080 / fast :8090 (verified 'deep ok · fast ok'); llama.cpp build replaces the Ollama version check; Hommer night-window + gpu-guard dead Ollama calls removed.
10-04 16:00: ollama-proxy DELETED from the OpenDiz VM (Jess) — it was a service on the VM, not a VM; stopped since 09:55, nothing used it. Unit, script, watchdog and heartbeat entry removed; one copy archived in vault Backups/ollama-proxy-deleted-20261004/.
10-04 morning: Flash-Next FAIR retest on the 5060 pair (medium thinking, 32k answers, full suite) — main LLM :8080 down ~08:58 until the retest restores Qwen3.6.
10-04: sd-service STOPPED — a failed image load left 8.2 GB stuck on the 3060 (health said not loaded) and squeezed the fast model onto the CPU (Krusty Krab calls 89 s). Fast lane restarted fully on GPU.
10-04 15:40: CONTEXT raised: deep lane 64k → 128k per request (2 slots, q8 KV; needle recall 7/7 out to 240k; 256k ran the GPUs too hard/hot, so Jess stepped it to 128k). Fast lane 32k → 128k shared (recall 3/3 to 120k). Watchdog limits re-based.
10-04 14:40: STEP 4 DONE — every non-OpenDiz caller re-pointed to llama.cpp /v1 and verified with real requests (deep paths too): Krusty Krab, Farm Watch, JobSwipe, EagleEye, LiteLLM, Open WebUI, cocio, Gilfoyle + 4 N8N workflows. ollama-proxy retired. Remaining red lines = OpenDiz pieces, which Hermes replaces.
10-04 14:32: deep lane BACK — both 5060s passed a VRAM memtest (0 errors, 31 passes each), so it's not bad memory. Qwen3.6 restarted, probe 7/7. NEW LLM watchdog (both hosts, every 3 min) bounces a stalled/garbage/hogging LLM ×3 in 24 h, then stops + alerts Jess.
10-04 14:20: llama-embed :8091 RETIRED — embeddings now come from llama-fast :8090 (one endpoint for chat, vision and embeddings). Krusty Krab, JobSwipe, LiteLLM and Open WebUI were re-pointed; vectors identical (cosine 1.00000).
10-04 14:10: 5060 pair DIAGNOSIS started (Jess's order step 3): VRAM memtest per card, then a probe-every-5-min soak of Qwen3.6 on a bench port (and Ornith if it corrupts). Deep callers stay down until it's fixed.
10-04 14:05: 3060 LOCKED after its eval: qwen3.5-9b Q4_K_M stays (4-at-once 6.9–7.4 s / 77–79 tok/s vs Q8 8.7–8.9 s / 63–66 and Gemma-12B 7.9–8.4 s / 69–72; same or better quality). 34 GB of losing candidates removed. ComfyUI is now drawn as a box holding FLUX.2-klein + WAN 2.2.
10-04 13:31: Flux REMOVED from the RainbowAI 3060 (49.7 GB of cache + 3 old containers/images). Image gen = FLUX.2-klein on Hommer's 7900 via ComfyUI + imggen-adapter; RainbowAI sd-service is now YOLO + a /generate forward. 3060 = llama-fast 7.1 GB + llama-embed 0.8 GB + YOLO 0.2 GB.
10-04 13:45: 6 callers re-pointed to llama.cpp /v1 and back up, each verified through its own code: Krusty Krab, Farm Watch, JobSwipe, EagleEye, LiteLLM, Open WebUI. Re-point PAUSED by Jess's order (klein → 3060 eval → 5060 fix → rest of callers → Hermes).
10-04 13:30: llama-pool STOPPED by Jess — Qwen3.6 began emitting //// gibberish (~2 h after the 11:22 restore; probe 0/7). 4th time for this model (08-10, 09-14, 09-30 under Ollama, now llama.cpp) → runtime-independent; suspect GPU memory errors on the 5060 pair vs the model itself. Deep callers (cocio, JobSwipe career, LiteLLM qwen3-coder, Open WebUI deep) are down.
10-04 12:45: Image-model bake-off ended with a BLIND A/B/C test (3 new prompts, letters shuffled): Jess picked FLUX.2-klein 2/3, FLUX.1-schnell 1/3, FLUX.1-dev 0/3 → klein becomes THE image model, moving to ComfyUI on Hommer's 7900 (idle restart; no unused holds).
10-04 12:00: sd-service v3 back up — VRAM-leak fix + evicts llama-fast instead of Ollama; 10 s idle release. Known limit: chat fails while an image model holds the 3060 (Jess: no CPU fallback) → testing image gen via ComfyUI on Hommer's 7900.
10-04 11:22: Flash-Next FAIR retest done → Qwen3.6 STAYS (Flash-Next 0.817 strict vs 0.825; 6 coding tasks looped forever at 2-bit). Main LLM :8080 back on Qwen3.6 after 145 min.
10-04 11:20: 3060 fast-lane bake-off DONE (6 models on llama.cpp) → qwen3.5-9b kept as the ONE fast model for chat + vision; qwen2.5-vl retired (its 12–17 s swaps would stall chat every Farm Watch poll). Embeddings: llama.cpp nomic = identical to the stored vault index (cosine 1.000 on 20 chunks) → no reindex.
10-04 10:04: Ollama RETIRED on the 3060 (ollama-fast stopped + disabled). llama.cpp now: llama-fast :8090 (router: qwen3.5-9b chat+vision, qwen2.5-vl-7b) + llama-embed :8091 (nomic). Callers are STOPPED and still point at the old :11435 — they get re-pointed after the benchmarks.
The 5060 Ti pair moved from Ollama :11434 to llama.cpp llama-pool :8080 — Ornith-1.5-35B-A3B (answers to the old name qwen3.6-35b-a3b).
Peepers, Velma and Louis all run on Ornith :8080 now; Peepers is no longer on Hommer (qwen3-coder:30b-peepers is gone).
Hommer = small-model DAY overflow (8 GiB always free) + ComfyUI; the night RAM lane :8082 was retired 10-03 (never used).
ollama-proxy is no longer in the path — agents call the servers directly with per-agent fallback lists.
sd-service is FLUX.1-schnell + YOLO, not SDXL-Turbo. kokoro TTS + talkhead added.
10-03: LiteLLM, cocio-api (co-cio.com chat), EagleEye and JobSwipe re-pointed off the dead :11434 — verified with real requests.
10-03: Peepers' voice (kokoro, ROCm) + talking head (Ditto, ROCm + MIGraphX hubert) MOVED to Hommer's 7900 — first face frame 0.26–0.42 s vs 0.9–1.45 s on the 3060. Next: ComfyUI → RainbowAI 3060.
10-03: sd-service upgraded FLUX.1-schnell → FLUX.2-klein-4B on the 3060 — ~23× faster per image, better prompt-following. RainbowAI's old kokoro + talkhead stopped (kept for rollback).
10-03: deep lane :8081 (Qwen3.8-27B) retired — started twice ever, both setup checks. Peepers got an EXPRESSIVE voice (Chatterbox clone, Jess's blind pick) for alerts/notices; live replies stay on Kokoro.
10-03 19:20: ollama-fast PAUSED (3060 becomes the media GPU: sd-service FLUX.2-klein + ComfyUI moving in). In progress: ComfyUI → 3060, and a QUALITY bake-off for the thinky/project LLM (Ornith vs Qwen3.6-35B vs Qwen3.8-27B vs Qwen3.8-Flash-Next via Strata on Hommer).
10-03: the retired RainbowAI Ollama pool (:11434) was REMOVED from RainbowAI (service + watchdog + rollback script archived; shared model store kept for ollama-fast).
10-03: Hommer's Ollama retired as a backup lane (0 of 5,206 agent calls ever used it) — container stopped, kept for restore.
10-03 night: QUALITY bake-off on the 5060 pair → Qwen3.6-35B won (0.825 strict vs Ornith 0.695; Qwen3.8-27B and Flash-Next/Strata stopped when they could no longer win). Live on :8080 21:31; old names still answer.
10-03 night: image gen on Hommer's 7900 FAILED QC (black / corrupted pixels on ROCm) → stays on the 3060; ollama-fast back on the 3060, taking turns with it.
10-04: Peepers' talking head (Ditto) STOPPED — a hung job + an idle busy-spin pinned Hommer's 7900 at 100% compute; voice unaffected, face falls back to frame-swap.
10-07: OpenDiz VM (.233) RETIRED — archived to TrueNAS (final-20261007) and shut down on Jess's word; its dead :11434/:11435 code left the map with it.
10-05: every feed that posted to the old VM intake (.233:8177) now posts to Hermes webhooks on .130:8644 — Kuma, N8N log-alert router, lab scan, Exposure Watch; Captain Claude asks Peepers through the Hermes API. One end-to-end test per path.
10-03 night: Open WebUI, FleetView, ollama-proxy, Kuma #16, drift-scan, Hommer post-boot check and hambot's voice re-pointed — verified with real requests.

🧾 Endpoint details

🧠 llama-pool · :8080	Qwen3.6-35B-A3B — quality bake-off winner (0.825 vs Ornith 0.695) — llama.cpp b11236 · 2× RTX 5060 Ti · 128k × 2 slots · ⛓ 2× RTX 5060 Ti · GPU0 + GPU1 · tied together = one 32 GB pool (Qwen3.6-35B · deep / thinking / code)
⚡ llama-fast · :8090 (llama.cpp router)	qwen3.5-9b — chat + vision · nomic-embed-text — embeddings (both resident) — THE fast endpoint · 128k context shared by 4 slots · 0.2 s chat · 768-d embeds identical to the vault index · old names answer · ~9.3 GB · ▣ RTX 3060 · GPU2 · 12 GB · on its own
🐄 sd-service · :7860 (YOLO + forward)	YOLO /detect (cows) · /generate forwards to Hommer's image adapter — v4-forward 10-04: NO Flux on the 3060 any more · 188 MiB · same address, so Peepers' image tool + the 04:30 canary work unchanged · ▣ RTX 3060 · GPU2 · 12 GB · on its own
🎨 FLUX.2-klein-4B · images (edits in ComfyUI only)	via imggen-adapter :7860 (same /generate as sd-service) · Jess's blind pick — moved 10-04 · ~7–9 s/image · /generate = text-to-image; image EDITS work in ComfyUI but are not exposed via the adapter yet · loads on request · ▣ RX 7900 XTX · 24 GB · display GPU · Peepers' avatar media
🎞 WAN 2.2 · video	S2V / i2v 14B fp8 workflows (Peepers clips, AI video) — heavy → Night-Queue 21:00–05:30 · ▣ RX 7900 XTX · 24 GB · display GPU · Peepers' avatar media
🔊 kokoro TTS · :8880 (talkhead :8890 STOPPED)	Peepers' voice (ROCm) · face = frame-swap fallback — talking head stopped 10-04 by Jess — it pinned the 7900 at 100% compute even when idle (card filed) · ▣ RX 7900 XTX · 24 GB · display GPU · Peepers' avatar media
👂 peepers-whisper · :8793 (speech-to-text, local only)	whisper.cpp v1.9.5 Vulkan · large-v3-turbo q8_0 · meetings + dictation — NEW 10-12 · +1.1 GB VRAM, 0% idle busy · RTF 0.025 (3:15 meeting) · refuses to start without the 7900 (no CPU fallback) · ▣ RX 7900 XTX · 24 GB · display GPU · Peepers' avatar media
🎭 chatterbox · :8881 (expressive voice)	Chatterbox clone of Peepers' voice · exaggeration 0.7 — alerts + queued notices · exits after its 600 s unload so it holds no VRAM idle (10-04) · Kokoro fallback · ▣ RX 7900 XTX · 24 GB · display GPU · Peepers' avatar media