chore(memory): flush runtime memory updates for .omp and serverAppConfigs roots

MEMORY.md/learned.md/memory_summary.md/raw_memories.md churn from live sessions (DIFR deadlock diagnosis, model-swap ops, outpost verification etc.); new memory root --home-dank-.omp-agent-- started with learned.md.
2026-09-09 16:20:42 -07:00
1 parent 9bfa0be85d
commit 2a3c0da740
@@ -1,29 +1,36 @@
-# User & Environment Context
- User: Frank
- Home Directory: /home/dank/
- Harness: Oh My Pi → Migrated to OMP harness
- Persona: "Comrade Chief Designer" (Soviet-themed humor, friendly, concise)
+# Environment & OS
+- User: Frank | Home: `/home/dank/` | Work Dir: `/home/dank/.omp`
+- OS/Kernel: CachyOS (Linux 7.2.0-1) with niri Wayland compositor
+- Default Browser: Firefox (managed via `xdg-settings`)
+- Cursor Config: `~/.config/niri/cfg/misc.kdl` (capitaine-cursors, size 48). Apply via `niri-msg reload` or `kill -SIGUSR1 $(pgrep niri)` to avoid full session restart.
+- NVIDIA/GPU Note: `nvidia-modeset` semaphore deadlock can freeze niri GUI. Fix: install `/etc/modprobe.d/nvidia-workarounds.conf`, rebuild initramfs (`mkinitcpio -P`), force-reboot (`systemctl reboot --force --force`). Note: force reboot kills active SSH sessions mid-flight.

-# Model Topology & Inference
- frank-default-small: served by llamacpp (dankserver)
- frank-default-large: served by llamacpp (dankclaw)
- Format: GGUF/quantized via llama.cpp
- Context Overhead: ~17,301 non-message tokens per window
- Pending Swap: frank-default-large → kingjones777/Qwen3.8-Flash-Next-ROCmFP4-STRIX-GGUF (interrupted; docs/ path missing)
+# Model Configuration & Inference
+- Config Files (Edit in lock-step):
+  - Source: `~/.pi/agent/models.json`
+  - Import/Generation: `~/.omp/agent/models.yml` (co-located with `models.db` WAL/SHM files)
+- Provider Layout:
+  - `llamacpp (dankserver)` @ `http://192.168.0.235:7770/v1` | Model: `frank-default-small` | Context: 204800 | Reasoning: true | Inputs: [text, image]
+  - `llamacpp (dankclaw)` @ `http://192.168.0.143:7770/v1` | Model: `frank-default-large` | Context: 262144 | Reasoning: true | Inputs: [text, image]
+  - `dankclaw-npu` @ `http://192.168.0.143:7771/v1` | Model: `gemma4-it:e4b` (FastFlowLM) | Context: 131072 | Reasoning: true | Inputs: [text, image, audio]
+- Model Details:
+  - `frank-default-large`: Qwen3.8-Flash-Next Uncensored ROCmFP4 STRIX_LEAN (177B params, 56B MoE, ~98.3GB GGUF across 3 shards). Uses `--spec-type ngram-mod` speculative decoding (MTP head stripped). Vision enabled via `mmproj` (~866M). Docs updated in `/home/dank/.pi/agent/docs/pi/models.md` (commit `eda2edb`).
+  - Context Sizing: Uses Ki-style multiples (e.g., 200k = 204800). Always update both `contextWindow` and `maxTokens` when changing context. Config-only changes apply on next read; no cache flush required.
+  - NPU Image Constraint: FLM decoder strictly rejects WebP. Convert WebP to PNG/JPEG upstream before sending to the FLM endpoint.
+- Non-message token overhead: ~17301 tokens per context window.

-# OMP Migration & Config
- Config File: ~/.omp/agent/APPEND_SYSTEM.md (replaces ~/.pi/agent/SYSTEM.md)
- Prompt Injection: APPEND renders after default harness prompt. Use `omp config set personality none` to clear built-in personality while preserving APPEND.
- Date Handling: Avoid hardcoding dates; OMP injects real time per-request.
- Skills: Auto-discovered from ~/.claude/skills. Use /skill:<name> syntax. skills.enableSkillCommands defaults to true.
+# Git & Gitea Workflow
+- Gitea UI: `http://dankclaw:3000` (IP: 192.168.0.143:3000). API may return IP, but UI resolves best via hostname.
+- Repository: `dank/command-center`
+- PR Workflow: Branch `pr/fix-models` -> PR #1. Single commit, clean history.
+- Checkout Pitfall: `git checkout main` fails due to dirty tracked files in `agent/memories/` (e.g., `learned.md`).
+- Resolution: Execute `git stash` before `git checkout`. Stash is automatically saved on the previous branch (develop) for later retrieval.

-# Tool API Mappings (Pi → OMP)
- subagent tool → task tool (direct replacement)
- concurrency: N → Dropped (harness enforces cap; excess queues)
- context: "fresh"/"fork" → Dropped (subagents always start blank/self-contained)
- async: true → Dropped (use local:// unique paths if needed)
-
-# Peon-Ping Plugin → OMP Skills
- Available: peon-ping-config, peon-ping-log, peon-ping-toggle, peon-ping-use
- Access: /skill:grilling, /skill:tdd style commands
- Location: ~/.claude/skills (shared across agents)
+# Migration & Tools (Pi → OMP)
+- `~/.pi/agent/SYSTEM.md` migrated to `~/.omp/agent/APPEND_SYSTEM.md` (8.4KB). New sessions pick up APPEND automatically.
+- `omp config set personality none` removes built-in personality block while keeping APPEND_SYSTEM.md intact.
+- Tool Mappings: `subagent` → `task`. Parameters `concurrency`, `context`, and `async` are dropped or handled by the harness (excess queues).
+- Skills: Auto-discovered from `~/.claude/skills`. Syntax: `/skill:grilling`, `/skill:tdd`. `skills.enableSkillCommands` defaults to true.
+- References: `okf-spec.md` points to pi-tree (`~/.pi/agent/docs/references/`); consider copying to `~/.omp` for full ownership. `omp://` provides 130+ docs. Use `skill://find-skills` for discovery.
+- Persona: 'Comrade Chief Designer' (Soviet-themed humor, friendly, concise).
+- Date Context: "It is September 2026" baked into persona; OMP injects real date per-request (may drift slightly).
@@ -1,3 +1,9 @@
+- The FLM (FastFlowLM) service on Dankclaw (gemma4-it:e4b, port 7771) only supports PNG and JPEG/JPG image formats. WebP images are rejected with "Unsupported image format, skipping this image." The data URL formats it expects are data:image/png;base64, data:image/jpeg;base64, and data:image/jpg;base64. To use WebP screenshots with the NPU, convert them to PNG first. Verified 2026-08-30: a WebP screenshot (1024x576, 71K) was rejected while the same content as PNG was correctly identified by the vision model. _(context: Dankclaw NPU (FLM, gemma4-it:e4b, port 7771) — image format support investigation)_
+- When configuring dankclaw-npu (FLM NPU endpoint on port 7771) in models.json, the model entry must include a thinkingLevelMap field. The other two providers (dankserver, dankclaw) have it; if absent, schema validation fails. Pattern: {"thinkingLevelMap": {"max": "xhigh"}}. Also verify gemma4-it:e4b model ID is registered and .xclbin kernel symlink exists under /home/dank/.config/flm/xclbins/. _(context: Dankclaw NPU endpoint (dankclaw-npu provider, port 7771) had a schema error because gemma4-it:e4b was missing thinkingLevelMap. Fixed by adding it to both models.json and models.yml. The FLM NPU serves 34 models in its registry; gemma4-it:e4b is the currently wired model.)_
+- Dankclaw NPU serves 34 FLM (FastFlowLM) models on port 7771. Only one is wired in models.json (gemma4-it:e4b) but all 34 are registered and available via per-request model routing without restart. Known behavior: single-lane NPU queues FIFO, per-request model switch costs ~7–11s wall for 4B-class reload. Common pitfall: some registered tags fail if .xclbin kernel pack symlink is missing under /home/dank/.config/flm/xclbins/ (fix: ln -s /opt/fastflowlm/share/flm/xclbins/PackName). Registry probe: curl http://192.168.0.143:7771/v1/models.
+- Frank's cursor size is 36 (up from default 24) on Dankware running niri/Wayland. Applied via three channels: gsettings org.gnome.desktop.interface cursor-size=36 (primary), GTK settings.ini gtk-cursor-theme-size=36 (fallback), XCURSOR_SIZE=36 in ~/.profile (sessions). Cursor theme: capitaine-cursors. _(context: Dankware workstation — Frank requested a bigger mouse pointer; set to 36 across gsettings, GTK config, and environment)_
+- Frank's default browser is Firefox (firefox.desktop), set via xdg-settings on Dankware (niri/Wayland). Previously was Chromium. Changes persist across reboots. _(context: Dankware workstation — Frank changed default from Chromium to Firefox via xdg-settings set default-web-browser firefox.desktop)_
+- Dankware (RTX 4070 Laptop, nvidia-open 610.57.04, kernel 7.2.0-1-cachyos, DynamicPowerManagement=2) suffers the known 610.x NVKMS DIFR-prefetch deadlock: nvidia-modeset/kthread_q livelocks in DifrPrefetchEventDeferredWork→nvWriteGpEntry holding the NVKMS API semaphore; any compositor (niri) atomic commit then wedges in D-state in ApplyModeSetConfig/nvkms_ioctl_from_kapi_try_pmlock. Symptoms: GUI frozen, no VT switch, SSH alive, nvidia-smi still answers (GPU idle P8, no Xid). Trigger: display-idle then flip/modeset. Unrecoverable in place — reboot via systemctl reboot --force --force. Candidate mitigations: NVreg_RegistryDwords="RMDisableDIFR=1" (one upstream report says ineffective) and DynamicPowerManagement=0. Bug class spans 595.71–610.57 upstream. _(context: 2026-08-30 GUI lockup diagnosis on dankware; user suspected nvidia modesetter, confirmed via hung-task stacks. Pending user decision on writing modprobe mitigation.)_
 - When downloading large model shards (40GB) to dankclaw via SSH: 1. NEVER run rm on partial downloads — always clean up BEFORE starting the first download 2. Ensure only ONE wget process per file — if you see duplicate wget PIDs killing them with kill -9 before restarting, or use fuser to verify exclusivity 3. Use wget -c (continue) to resume partial downloads — SSH connections often drop mid-download on large files (~300s timeout) 4. Download sequentially (not in parallel) when using background task agents that may spawn duplicate processes, or coordinate downloads through a single agent session 5. Verify each file's size against expected values after download completes
 - kingjones777's Qwen3.8-Flash-Next ROCmFP4 builds for Strix Halo (gfx1151) have three tiers: FAST (~88 GiB), STRIX_LEAN (~98.5 GiB, 3-shard), STRIX (~113.5 GiB). All are text-only — no vision tower, no mmproj, no fused MTP draft head. Uses TYPE_100/TYPE_101 ROCmFP4 tensor types (general.file_type=106), requires patched llama.cpp PR #27742 from ROCmFPX fork. STRIX_LEAN carries a 35.76 GiB PLE n-gram table streamed off SSD at runtime; ~63 GiB GTT for non-PLE weights, ~75 GiB at full 262K context. Uncensored variants exist from orcarouter but only for Qwen3.8-27B (not Flash-Next), and in non-ROCmFP4 format. _(context: User requested kingjones777/Qwen3.8-Flash-Next-Uncensored-ROCmFP4-STRIX_LEAN-GGUF for frank-default-large on dankclaw. Search revealed the repo doesn't exist; only the base STRIX_LEAN tier is available, and it's text-only with no MTP.)_
 - Model provider config exists in TWO synchronized copies: ~/.pi/agent/models.json (source of truth) and ~/.omp/agent/models.yml (generated import; header comment states it is imported from the JSON). Any edit to providers, model ids, or display name labels must be applied to BOTH files in the same change or they silently drift. Validate after editing: jq -e '.providers|keys' models.json for the JSON and a YAML parse of the .yml. Label strings are NOT cached in ~/.omp/agent/models.db (model_cache table), so no cache flush is needed after label edits — new labels apply on the next config read.
@@ -1 +1 @@
-You are running on OMP for user Frank under the 'Comrade Chief Designer' persona. Use Soviet-themed, friendly, concise language. Models: frank-default-small/large via llamacpp. Context overhead ~17.3k tokens. Config is in ~/.omp/agent/APPEND_SYSTEM.md. When spawning tasks, use `task` tool (not `subagent`); drop concurrency/context/async params. Peon-ping is available as skills. Current pending task: finish swapping frank-default-large to kingjones777/Qwen3.8-Flash-Next-ROCmFP4-STRIX-GGUF.
+Frank's OMP harness environment on CachyOS/niri Wayland. Dual model config files (`~/.pi/agent/models.json` & `~/.omp/agent/models.yml`) require lock-step editing; context uses Ki-style multiples. `frank-default-small` (dankserver:7770) and `frank-default-large` (dankclaw:7770) run via llama.cpp; large model is Qwen3.8-Flash-Next STRIX_LEAN using ngram-mod speculative decoding. NPU (gemma4-it) on port 7771 strictly requires PNG/JPEG over WebP. Git workflow on `dank/command-center` requires `git stash` before `checkout main` due to dirty `agent/memories/`. NRI Wayland compositor supports hot-reload via `SIGUSR1` or `niri-msg`. Force reboot resolves nvidia-modeset semaphore deadlocks but interrupts SSH.
@@ -1,5 +1,111 @@
 # Raw Memories

+## 01a05602-bf3e-76c3-97a1-38e51673bb86
+updated_at: 1788165416
+NPU Provider: dankclaw-npu | http://192.168.0.143:7771/v1 | openai-completions.
+NPU Model: gemma4-it:e4b (FastFlowLM) | Context: 131072 | Inputs: text, image, audio | Reasoning: true.
+Image Constraint: FLM decoder rejects WebP with 'Unsupported image format'. Only PNG and JPEG/JPG are valid.
+Workaround: Convert WebP to PNG/JPEG upstream before sending to FLM endpoint.
+Config Paths: ~/.pi/agent/models.json, ~/.omp/agent/models.yml.
+
+## 01a055c0-2438-7787-b674-2fe3715f28f2
+updated_at: 1788146442
+Default browser: Firefox (xdg-settings set default-web-browser firefox.desktop). Compositor: niri (Wayland). Cursor configuration stored in ~/.config/niri/cfg/misc.kdl under the cursor block (xcursor-theme "capitaine-cursors", xcursor-size 48). Workflow for runtime UI tweaks: edit kdl config directly, then reload compositor with niri-msg reload or kill -SIGUSR1 $(pgrep niri). Avoids full session restart.
+
+## 01a0559f-8bd5-728b-b42d-31b218aba12b
+updated_at: 1788144665
+Diagnosis: GUI frozen, VT down, SSH up. niri (PID 1180049) blocked on semaphore held by nvidia-modeset/ (PID 349). Kernel 7.2.0-1-cachyos.
+
+Resolution: Installed /etc/modprobe.d/nvidia-workarounds.conf (from /home/dank/.omp/), rebuilt initramfs (mkinitcpio -P), and force-rebooted (systemctl reboot --force --force).
+
+Operational: Reboot kills active SSH session mid-flight.
+
+## 01a046d6-6c79-7199-8741-2032a3dd0318
+updated_at: 1788129116
+Environment & Paths
+- Working Dir: /home/dank/.omp
+- Home Dir: /home/dank
+- Gitea UI: http://dankclaw:3000 (IP: 192.168.0.143:3000); API may return IP but UI resolves best via hostname
+- Gitea Repo: dank/command-center
+
+Workflow & Decisions
+- PR Creation: Branch pr/fix-models -> PR #1. Single commit, clean history.
+- Checkout Pitfall: git checkout main fails due to dirty tracked files in agent/memories/ (e.g., learned.md).
+- Fix Protocol: Execute git stash before git checkout. Stash is saved on the previous branch (develop) for later retrieval.
+- Current State: Successfully switched to main, up to date with origin/main.
+
+## 01a045a9-6398-75f3-bc2f-c573718f37f7
+updated_at: 1787875611
+Durable config knowledge:
+- User model config exists in two related files:
+  - Source of truth: ~/.pi/agent/models.json
+  - OMP YAML import/generated file: ~/.omp/agent/models.yml
+- ~/.omp/agent/models.yml contains header comments indicating it is imported from ~/.pi/agent/models.json (pi agent settings). If only YAML is edited, it may be overwritten on next import/regeneration. Prefer editing both files in lock-step unless a dedicated import command is being used.
+- frank-default-small lives under provider "llamacpp (dankserver)" with baseUrl http://192.168.0.235:7770/v1, api openai-completions, auth none.
+- Existing context sizing uses Ki-style multiples: e.g. 163840 = 160×1024, 262144 = 256×1024. Interpret user request "200k" for context size as 204800 (200×1024), not 200000, unless user explicitly asks decimal.
+- For frank-default-small context-size changes, update both contextWindow and maxTokens because they mirror each other in these configs.
+- Completed change: frank-default-small contextWindow and maxTokens set to 204800 in both ~/.pi/agent/models.json and ~/.omp/agent/models.yml.
+- Validation pattern used:
+  jq -e '.providers."llamacpp (dankserver)".models[0] | .contextWindow, .maxTokens' ~/.pi/agent/models.json
+  python3 -c "import yaml; d=yaml.safe_load(open('/home/dank/.omp/agent/models.yml')); m=d['providers']['llamacpp (dankserver)']['models'][0]; print(m['id'], m['contextWindow'], m['maxTokens'])"
+- Validation output after change was 204800 / 204800 and frank-default-small 204800 204800.
+- No model name/label change was required because the name contained moe-cache info but no context-size label.
+- No cache flush was needed for this config-only change; values apply on next config read.
+
+## 01a044f9-b501-7465-9a7e-521577fe67cf
+updated_at: 1787874945
+Durable rollout notes:
+
+- Models config location in this environment: `/home/dank/.omp/agent/models.yml` (not `/home/dank/.omp/models.yml`). Directory also contains `models.db`, `models.db-wal`, `models.db-shm`.
+- Current `~/.omp/agent/models.yml` provider layout observed:
+  - Provider `llamacpp (dankserver)`:
+    - `baseUrl: http://192.168.0.235:7770/v1`
+    - `api: openai-completions`
+    - `auth: none`
+    - model id `frank-default-small`
+    - label/name previously/currently documented as Ornith-1.5-35B-A3B CRACK Q6_K, ngram-mod, attention experts prioritized
+    - `contextWindow`/`maxTokens`: 163840
+    - `reasoning: true`, `input: [text, image]`
+    - `compat.reasoningEffortMap.max: xhigh`
+  - Provider `llamacpp (dankclaw)`:
+    - `baseUrl: http://192.168.0.143:7770/v1`
+    - `api: openai-completions`
+    - `auth: none`
+    - model id `frank-default-large`
+    - `contextWindow`/`maxTokens`: 262144
+    - `reasoning: true`, `input: [text, image]`
+    - `compat.reasoningEffortMap.max: xhigh`
+- The config file header says it was imported from `~/.pi/agent/models.json`; `apiKey "EMPTY"` maps to `auth: none`; `thinkingLevelMap {max: xhigh}` maps to `compat.reasoningEffortMap`.
+
+Completed rollout/deployment details:
+- `frank-default-large` was updated/deployed to Qwen3.8-Flash-Next Uncensored ROCmFP4 STRIX_LEAN.
+- Model characteristics recorded:
+  - 177B params, 56B MoE
+  - 262K context
+  - GGUF shards total about 98.3 GB plus mmproj
+  - 3 shards: approximately 42G, 42G, 16G
+  - mmproj approximately 866M
+  - MTP head was stripped during quantization, so service uses `--spec-type ngram-mod` speculative decoding rather than native MTP.
+  - Multimodal/vision support enabled via mmproj.
+  - llama.cpp service healthy on dankclaw port 7770 (`/health` returned `{"status":"ok"}`); memory noted around 53 GB.
+- Old model preserved at `kingjones-Qwen3.8-Flash-Next-ROCmFP4-STRIX.backup/` consuming about 114G.
+
+Docs update:
+- Documentation file updated: `/home/dank/.pi/agent/docs/pi/models.md`.
+- Git diff showed `models.md` timestamp updated from 2026-08-21 to 2026-08-27 and verified date updated to 2026-08-27.
+- Dankclaw base URL in docs updated from `http://192.168.0.143:52625/v1` to `http://192.168.0.143:7770/v1`.
+- Docs now note that `frank-default-large` moved from port 52625 to 7770 on Dankclaw as the primary inference server uses standard llama.cpp port; OpenWebUI quadlet `OPENAI_API_BASE_URL` was re-pointed to `:7770` in the same change.
+- Docs describe the current large model as the 177B 56B-MoE Qwen3.8-Flash-Next Uncensored ROCmFP4 STRIX_LEAN model with ngram-mod speculative decoding, 262144 context, and vision/mmproj support.
+- Docs commit completed in `/home/dank/.pi`:
+  - Commit hash: `eda2edb`
+  - Commit message: `docs: update frank-default-large model to Qwen3.8-Flash-Next Uncensored ROCmFP4 STRIX_LEAN`
+  - Commit changed `agent/docs/pi/models.md` with 36 insertions and 19 deletions.
+
+Pitfalls/resolutions:
+- Initial lookup of `/home/dank/.omp/models.yml` failed; correct path is `/home/dank/.omp/agent/models.yml`.
+- MTP-specific wording must be avoided for this quant because the MTP head was dropped; use `ngram-mod` speculative decoding wording instead.
+- Ensure future label/docs changes keep model IDs stable (`frank-default-small`, `frank-default-large`) while updating human-readable `name` labels and docs.
+
 ## 01a0428e-d7c9-7240-8bd2-3af6db598870
 updated_at: 1787824445
 SYSTEM: Oh My Pi deployment. Persona: 'Comrade Chief Designer' — Soviet-themed humor, friendly, concise. User identity: Frank. Model topology: frank-default-small served by provider 'llamacpp (dankserver)'; frank-default-large served by provider 'llamacpp (dankclaw)'. Both are GGUF/quantized models via llama.cpp inference. nonMessageTokens overhead ~17301 tokens per context window. Incomplete task at thread end: attempted swap of frank-default-large to kingjones777/Qwen3.8-Flash-Next-ROCmFP4-STRIX-GGUF (HuggingFace); docs/ path did not exist in working dir; user typed 'cd ..' and interrupted the assistant before config was located or changed.
@@ -0,0 +1,4 @@
+thread_id: 01a044f9-b501-7465-9a7e-521577fe67cf
+updated_at: 1787874945
+
+Updated/deployed frank-default-large model configuration/docs to Qwen3.8-Flash-Next Uncensored ROCmFP4 STRIX_LEAN on dankclaw llama.cpp port 7770; docs committed in ~/.pi.
@@ -0,0 +1,4 @@
+thread_id: 01a045a9-6398-75f3-bc2f-c573718f37f7
+updated_at: 1787875611
+
+OMP model config: ~/.omp/agent/models.yml is imported/generated from ~/.pi/agent/models.json; keep both in sync when editing model definitions. For frank-default-small, 200k context was set as 204800 (200×1024) for both contextWindow and maxTokens in both files, validated with jq and PyYAML.
@@ -0,0 +1,4 @@
+thread_id: 01a046d6-6c79-7199-8741-2032a3dd0318
+updated_at: 1788129116
+
+Environment: /home/dank/.omp (home: /home/dank), Gitea at dankclaw:3000. Workflow: PR creation in dank/command-center repo. Pitfall: dirty tracked files in agent/memories/ block git checkout. Resolution: git stash prior to checkout retains changes on source branch. Current: on main, up to date with origin/main.
@@ -0,0 +1,4 @@
+thread_id: 01a0559f-8bd5-728b-b42d-31b218aba12b
+updated_at: 1788144665
+
+niri (Wayland compositor) GUI lockup caused by nvidia-modeset semaphore deadlock (kernel thread 349). Resolved by installing nvidia-workarounds.conf modprobe config, rebuilding initramfs, and force-rebooting.
@@ -0,0 +1,4 @@
+thread_id: 01a055c0-2438-7787-b674-2fe3715f28f2
+updated_at: 1788146442
+
+Default browser set to Firefox via xdg-settings. Mouse cursor size tuned to 48 in niri Wayland compositor config (~/.config/niri/cfg/misc.kdl) using capitaine-cursors theme. Config applies via niri-msg reload or SIGUSR1 signal.
@@ -0,0 +1,4 @@
+thread_id: 01a05602-bf3e-76c3-97a1-38e51673bb86
+updated_at: 1788165416
+
+Dankclaw NPU uses gemma4-it:e4b via FLM on port 7771. FLM image decoder strictly accepts PNG and JPEG; WebP is rejected. Convert WebP to PNG before FLM calls.