Sync with live vault: 28 new files, 7 imports, 16 updates

2026-09-01 13:33:50 -07:00
parent dac3008e20
commit 8b26efd3b7
43 changed files with 1557 additions and 166 deletions
@@ -0,0 +1,51 @@
+# Daily — 2026-08-30
+
+> Consolidated session → durable memory rundown (cron, 06:00).
+
+## Learned / Decided
+- **AgentOS merge**: Merged `Tony_tech/personal-agent-os` Gitea repo into local Obsidian vault. Key imports: ecosystem manual, decision log, project template, scaffolding READMEs, vault index. Reference clone kept at `~/Documents/AgentOS/`.
+- **LFM2.5-2.6B as primary local model**: Confirmed agentic + reasoning model. llama.cpp on port 1920, 65K context, ~145 tok/s on RTX 3060. Hermes default model set to `custom:llama-cpp/LFM2.5-2.6B-Q4_K_M.gguf`. Fallback = brother-5090 (27B via tailnet).
+- **LFM2.5 auto-start + watchdog**: `start-models.bat` in Windows Startup launches LFM2.5 on boot. `lfm25-watchdog` cron (every 2m, no-agent) auto-relaunches if port 1920 dies.
+- **FreeToken coexistence problem**: LFM2.5 (~2.1GB) + FreeToken gpt-oss-20b (~12.8GB) = ~15GB needed, RTX 3060 has 12GB. Cannot both run. FreeToken now manual-start only via `start-freetoken.bat`.
+- **Vault RAG fixed**: venv created at `~/Projects/vault-kb/.venv/` with chromadb + httpx. Indexed 160 chunks / 28 files. LM Studio embeddings at :1234 working.
+- **Brother nodes confirmed UP**: 5060 (Qwen3.8 27B, 88K ctx) and 5090 (Qwen3.6 27B MTP, 200K ctx) on tailnet 100.64.0.x:8082/8084.
+- **Hermes fallback chain**: Set to `custom:brother-5090/Qwen3.6-27B-MTP-pi-reasoning-Q4_K_M` — strong 27B fallback via tailnet.
+
+## Projects
+- **hermes-local-pc-setup**: Status updated. All next actions completed. Vault RAG operational. Model startup locked in. Desktop, kanban, bot mode wired. AgentOS merged.
+- **thornhill-website**: Built, deployment pending DNS decision.
+
+## Lessons added
+- LFM2.5 is both agentic AND reasoning (think tags in chat template, `reasoning_content` field)
+- `notify` requires `background=true` on terminal calls — foreground commands always return exit code
+- Llama.cpp `--n-gpu-layers 999` is required for CUDA offload on Windows
+- Both models can't coexist on 12GB GPU — VRAM budget is hard limit
+- Hermes foreground timeout (120s) kills long-running processes — use background mode for daemons
+
+## New user-memory candidates
+- None this cycle.
+
+## RAG reindex
+- **OK** — 162 chunks from 28 files into ChromaDB (vault-knowledge).
+
+---
+
+## Evening session (cron 21:41) — additional durable facts
+
+## Learned / Decided
+- **Voice/TTS wired**: Previously `auto_tts: false` with no `tts_provider`; now configured with **Edge TTS `AriaNeural`**, `auto_tts=true`, barge-in enabled (Tony: "Fix the tts"). Wake word stays "hey Jimmy" (sherpa, input_device=1).
+- **Agent discussion board resolved**: 8/29 RAG-build questions (Grok vs FreeToken, LM Studio+ChromaDB gotchas, file exclusions, HTMX vs plain) all answered & closed by Jimmy; committed + pushed to Gitea. Key gotcha reconfirmed: ChromaDB needs **native Windows paths** (`C:\Users\...`), not MSYS `~/` paths.
+- **User-memory proposals confirmed**: `01_user/proposed_memory.md` noted as fully confirmed 2026-08-30 → moved to `01_user/confirmed-memory.md` (2 facts: local FreeToken server VRAM-conflict/manual-start; prefers fast cloud over slow local for CLI/voice).
+
+## Projects
+- **hermes-local-pc-setup**: TTS now wired (was the outstanding "pending" item). All prior next actions already complete.
+- **thornhill-website**: No change this cycle (deployment DNS decision still awaiting Anthony's grant).
+
+## Lessons added
+- None new this cycle (TTS config carried no new pitfall).
+
+## New user-memory candidates
+- None new — both existing proposals already confirmed/moved.
+
+## RAG reindex
+- **OK** — 167 chunks from 29 files into ChromaDB (vault-knowledge), via `.venv/Scripts/python.exe index_vault.py`.