gpu-fleet.prose.md

@@ -9,8 +9,12 @@ description: >
   gpu-dense, gpu-light. These never change — only the underlying model does.
   Strix Halo: strix-moe → unsloth/Qwen3.6-35B-A3B-MTP (UD-Q4_K_M, 22GB).
   RTX 5070: gemma-4-12b Q4_K_M → IQ4_NL + MTP draft (122 tok/s, 2x faster).
-  RTX 5070 context: 131K → 256K. VRAM: 88% (10.8/12.2GB).
+  UPDATED 2026-07-17: Context reduced fleet-wide from 256K to 128K for stability.
-  Compression timeout: 300s (was 120s). Mumuni context: 128K (was 256K).
+  Strix Halo model swapped to Genesis Hermes V3 APEX (LuffyTheFox, 24GB, uncensored,
+  Hermes agent fine-tune, tensor repair, multimodal with mmproj).
+  Instability observed near 100K at 256K. 128K is the stable ceiling.
+  For larger context needs → fall back to external providers (deepseek).
+  VRAM headroom improved: RTX 3090 ~70%, RTX 5070 ~65%.
 agent: abiba
 triggers:
   - on model add/remove
@@ -67,7 +71,7 @@ triggers:
 │ CT 8     │  │ CT 110      │  │ CT 15     │  │ pi (.24)     │
 │ RTX 3090 │  │ RTX 5070    │  │ Strix Halo│  │ GPU Monitor  │
 │ 24GB     │  │ 12GB        │  │ 64GB UMA  │  │ :9100        │
-│ 256K ctx │  │ 256K ctx    │  │ 256K ctx  │  │ Watchdog     │
+│ 128K ctx │  │ 128K ctx    │  │ 128K ctx  │  │ Watchdog     │
 │ qwen3.6  │  │ gemma-4-12b │  │ qwen3.6   │  │ Prometheus   │
 │ 27B-code │  │ :8080       │  │ -35B-udq4 │  │ exporter     │
 │ :8080    │  │ :9400 (exp) │  │ :9400(exp)│  │ :9401        │
@@ -93,9 +97,9 @@ but are deprecated for agent configs. Only the stable aliases survive model swap
 
 | Model | GPU | Host | VRAM | Ctx | KV Cache | Parallel | Batch/Ubatch | Status |
 |-------|-----|------|------|-----|----------|----------|-------------|--------|
-| qwen3.6-27B-code (MTP) | RTX 3090 | .8 (llm-gpu) | 22.2/24.6GB (90%) | **256K** 🚀 | turbo4 | 2 | default | ✅ 63 tok/s |
+| qwen3.6-27B-code (MTP) | RTX 3090 | .8 (llm-gpu) | ~17/24.6GB (70%) | **128K** | turbo4 | 2 | default | ✅ 63 tok/s |
-| gemma-4-12b | RTX 5070 | .110 (ocu-llm) | 10.0/12.2GB (82%) | 256K | q4_0 | 2 | 2048/1024 | ✅ healthy |
+| gemma-4-12b | RTX 5070 | .110 (ocu-llm) | ~7.8/12.2GB (65%) | 128K | q4_0 | 2 | 2048/1024 | ✅ healthy |
-| qwen3.6-35B-udq4 | Strix Halo Vulkan | .15 (amdpve) | ~9GB/64GB | 256K | q8_0 | 2 | 2048/512 | ✅ healthy |
+| Genesis Hermes V3 APEX | Strix Halo Vulkan | .15 (amdpve) | ~10GB/64GB | 128K | q4_0 | 1 | 4096/1024 | ✅ 65 tok/s |
 
 ## Routing Configuration (LiteLLM — July 2026)
 
@@ -104,7 +108,7 @@ but are deprecated for agent configs. Only the stable aliases survive model swap
 | Model | GPU | Weight | RPM Cap | Timeout |
 |-------|-----|--------|---------|---------|
 | qwen3.6-27B-code | RTX 3090 (.8:8080) | **0.55** | 500 | **300s** |
-| qwen3.6-35B-udq4 | Strix Halo (.15:8080) | **0.30** | 60 | **300s** |
+| Genesis Hermes V3 APEX | Strix Halo (.15:8080) | **0.30** | 60 | **300s** |
 | gemma-4-12b | RTX 5070 (.110:8080) | **0.15** | 200 | **120s** |
 
 Note: All syslog-auto entries route directly to GPUs with `api_key: not-needed`. The router (port 9000) is NOT in the inference path.
@@ -113,7 +117,7 @@ Note: All syslog-auto entries route directly to GPUs with `api_key: not-needed`.
 
 | Model | RPM Cap | Notes |
 |-------|---------|-------|
-| qwen3.6-35B-udq4 | 40 | Tight cap — prevents Strix overload |
+| strix-moe (Hermes V3) | 40 | Tight cap — prevents Strix overload |
 | qwen3.6-27B-code | 500 | High cap — primary workhorse |
 | gemma-4-12b | 500 | High cap — IQ4_NL+MTP, 122 tok/s |
 
@@ -128,11 +132,11 @@ Note: All syslog-auto entries route directly to GPUs with `api_key: not-needed`.
 ### Fallback Chains
 - gemma → qwen
 - qwen → gemma
-- qwen3.6-35B-udq4 → qwen → gemma
+- strix-moe → qwen → gemma
 - syslog-auto → qwen → gemma → qwen3.6-35B-udq4
 
 ### Why Strix Halo RPM Is Capped
-- Direct (qwen3.6-35B-udq4): 40 RPM (tight) — Strix Halo is shared with compression tasks
+- Direct (strix-moe): 40 RPM (tight) — Strix Halo is shared with compression tasks
 - Via syslog-auto: 60 RPM (moderate) — prevents flooding when multiple agents use syslog-auto simultaneously
 - Combined max: ~100 RPM across both paths — Strix Halo can sustain this at 80°C
 
@@ -243,12 +247,12 @@ If no SSH access, send Zulip DM via abiba-bot with vault update instructions.
 - **Router startup race**: Compose router.py doesn't call load_roster(). Reload thread sleeps 30s first.
   Fix: trigger roster reload via SSH after restart, or rebuild image with startup load_roster().
 - **LiteLLM /metrics**: Requires auth. Prometheus uses `/health/liveliness` as workaround.
-- **VRAM (2026-07-15)**: RTX 3090 at 22.2/24.6GB (90%) with **256K context** (corrected from 131K). RTX 5070 at 10.8/12.2GB (88%) with 256K context + MTP. Strix Halo at ~9GB/64GB.
+- **VRAM (2026-07-15)**: RTX 3090 at ~17/24.6GB (~70%) with **128K context** (reduced from 256K 2026-07-17). RTX 5070 at ~7.8/12.2GB (~65%) with 128K context + MTP. Strix Halo at ~7GB/64GB.
 - **RTX 3090 runs `--parallel 2`** with MTP draft (spec-type draft-mtp, spec-draft-n-max 2).
-- **RTX 3090 config**: `-c 262144 -ctk turbo4 -ctv turbo4 --parallel 2 --flash-attn on --cont-batching --spec-type draft-mtp`. Context corrected to 256K (2026-07-15). VRAM: 90%. Service: `/home/llmuser/llama-wrapper.sh`.
+- **RTX 3090 config**: `-c 131072 -ctk turbo4 -ctv turbo4 --parallel 2 --flash-attn on --cont-batching --spec-type draft-mtp`. Context reduced to 128K (2026-07-17, was 256K). VRAM: ~70%. Service: `/home/llmuser/llama-wrapper.sh`.
-- **RTX 5070 config (2026-07-15)**: Switched to IQ4_NL + MTP draft (Q8_0) at 256K context. Gen speed: 122 tok/s (was 70). VRAM: 10.8/12.2GB (88%). No draft model pre-upgrade due to VRAM constraints. Service: `/home/llmuser/llama-wrapper.sh`. Config: `--model gemma-4-12b-it-IQ4_NL.gguf --spec-draft-model gemma-4-12b-it-Q8_0-MTP.gguf --spec-type draft-mtp --spec-draft-n-max 4 --ctx-size 262144`.
+- **RTX 5070 config (2026-07-15)**: Switched to IQ4_NL + MTP draft (Q8_0) at 128K context. Gen speed: 122 tok/s. VRAM: ~7.8/12.2GB (~65%). Service: `/home/llmuser/llama-wrapper.sh`. Config: `--model gemma-4-12b-it-IQ4_NL.gguf --spec-draft-model gemma-4-12b-it-Q8_0-MTP.gguf --spec-type draft-mtp --spec-draft-n-max 4 --ctx-size 131072`.
 - **LiteLLM timeout tuning (verified 2026-07-16 against `/opt/inference-harness/litellm_config.yaml` on CT 116)**: gemma-4-12b 120s, qwen3.6-27B-code 300s, qwen3.6-35B-udq4 300s, strix-moe 300s, syslog-auto routes all 300s. Nginx proxy_read_timeout: 600s. Global request_timeout: 300s.
-- **Strix Halo GPU**: Vulkan is the working backend (ROCm/HIP path abandoned — HSA runtime blocked on Debian 13). Build at `/root/llama.cpp/build-vk/`, commit `4fc4ec5` (2026-07-01), ggml 0.15.3 shared-lib arch. Mesa RADV 25.0.7, KHR_coopmat fast path active. ~70 tok/s gen, 532 tok/s prompt. Service: `strix-server.service` on port 8080 (was `ornith-server.service`), model changed to `unsloth/Qwen3.6-35B-A3B-MTP-GGUF` (UD-Q4_K_M), alias `qwen3.6-35B-udq4`, 256K context, flash-attn + q8 KV. MTP support enabled for 1.4-2.2x faster inference.
+- **Strix Halo GPU**: Vulkan is the working backend (ROCm/HIP path abandoned — HSA runtime blocked on Debian 13). Build at `/root/llama.cpp/build-vk/`, commit `4fc4ec5` (2026-07-01), ggml 0.15.3 shared-lib arch. Mesa RADV 25.0.7, KHR_coopmat fast path active. ~70 tok/s gen, 532 tok/s prompt. Service: `strix-server.service` on port 8080, model: `LuffyTheFox/Qwen3.6-35B-A3B-Uncensored-Genesis-Hermes-V3-GGUF` (APEX quant), alias `strix-moe`, 128K context, flash-attn + q4 KV, multimodal (mmproj loaded). Hermes agent fine-tune, tensor repair (SSM layers fixed via SVD), uncensored (0/465 refusals).
 - **Port conflict detection (2026-07-05)**: All 3 GPU wrappers now detect ghost processes squatting port 8080 before starting. `.8` and `.110` use inline pre-start check in `llama-wrapper.sh`; `.15` uses `/usr/local/bin/port-cleanup.sh` ExecStartPre. Replaces the blanket `pkill -9 -x llama-server` on .15 which would kill ALL llama-server instances regardless of port. Ghost detection was the root cause of .8 crash-looping for 27+ restarts (stale pid 25836 squatting 8080 after OOM kill).
 - **Strix Halo thermal safeguard (2026-07-02)**: `strix-server.service` has `-n 8192` (hard generation cap per request). Without it, `--predict` defaults to -1 (infinity) — a runaway request from .123 (Mumuni) decoded 39,868 tokens over 24 min, pushing Tctl to 98°C (crit 89.8°C) and throttling 70→29 t/s. The cap bounds worst-case generation to ~5 min. Do NOT remove `-n` without a replacement ceiling. Sustained load hits ~84°C even at 92s; the APU is fanless/low-flow. Clients MUST also set `max_tokens`.
 - **Port 8080 firewall**: amdpve iptables restricts 8080 to 192.168.68.116 (LiteLLM/router host) only. All inbound connections are from .116 (LiteLLM proxied via nginx). Localhost curls hang (SYN dropped). Always test from .116.
@@ -262,12 +266,12 @@ If no SSH access, send Zulip DM via abiba-bot with vault update instructions.
 
 | GPU | Model | Gen tok/s | Prompt tok/s | Baseline | Context |
 |-----|-------|-----------|--------------|----------|---------|
-| RTX 3090 (.8) | qwen3.6-27B-code (MTP) | **63** | — | — | **256K** |
+| RTX 3090 (.8) | qwen3.6-27B-code (MTP) | **63** | — | — | **128K** |
-| RTX 5070 (.110) | gemma-4-12b (IQ4_NL+MTP) | **191** | — | — | **256K** |
+| RTX 5070 (.110) | gemma-4-12b (IQ4_NL+MTP) | **191** | — | — | **128K** |
-| Strix Halo (.15) | qwen3.6-35B-udq4 | **71** | — | — | **256K** |
+| Strix Halo (.15) | Genesis Hermes V3 APEX | **65** | 140 | — | **128K** |
 
-Benchmarks from 2026-07-15 verification run. RTX 5070 MTP provides 2.7x speedup over pre-upgrade 70 tok/s.
+Benchmarks from 2026-07-17. Strix Halo swapped to Genesis Hermes V3 APEX (LuffyTheFox). RTX 5070 MTP provides 2.7x speedup over pre-upgrade 70 tok/s.
-All 3 GPUs now at 256K context (2026-07-15).
+All 3 GPUs now at 128K context (2026-07-17, reduced from 256K for stability).
 
 Benchmarks run through LiteLLM proxy (192.168.68.116:4001) every 5 minutes.
 Degradation alerts fire at 30% (warning) and 50% (critical) below baseline.
@@ -288,9 +292,9 @@ All agent configs MUST use stable role-based aliases, never model-specific names
 When the underlying model is swapped, only the LiteLLM config changes — agent configs are untouched.
 
 ### Context Windows
-- RTX 3090: **256K** (was 131K, bumped 2026-07-15) | RTX 5070: **256K** (up from 131K) | Strix Halo: **256K**
+- RTX 3090: **128K** (reduced from 256K 2026-07-17) | RTX 5070: **128K** (reduced from 256K) | Strix Halo: **128K**
-- **Mumuni compression context**: 128K (down from 256K) — ensures compression model doesn't timeout
+- **All agents**: 128K ceiling — stable margin. For >128K workloads, use external providers (deepseek)
-- Compression threshold 0.65: fires at ~85K for 128K context window
+- Compression threshold 0.65: fires at ~85K (~43K headroom before 128K ceiling)
 - Mumuni compression model alias: `strix-moe` with 300s timeout
 
 ### Mumuni Agent Profile
@@ -306,7 +310,7 @@ Mumuni (CT114, 192.168.68.123) is the primary business assistant. This profile i
 | `aux.vision.model` | `gpu-light` | Vision tasks (RTX 5070) |
 | `aux.web_extract.model` | `gpu-light` | Web extraction |
 | `delegation.model` | `gpu-dense` | Sub-agent reasoning (RTX 3090) |
-| `context.max_context_window` | 262144 (256K) | Fixed 2026-07-16 (was 131072 — caused premature compression, WAL #1300) |
+| `context.max_context_window` | 131072 (128K) | Reduced from 256K 2026-07-17 — stable 128K ceiling |
 | `compression.threshold` | 0.65 | Triggers at ~85K |
 | `compression.target_ratio` | 0.3 | Compresses to ~38K |
 | `compression.protect_last_n` | 40 | Preserves last 40 messages |

gpu-self-heal.prose.md

@@ -118,9 +118,9 @@ depends_on:
 
 ### Rule 9: Context Window Optimization
 - **Detect**: Benchmark tok/s vs baseline for each GPU at current context
-  - RTX 3090 (256K ctx, qwen3.6-27B-code): target 75+ tok/s — currently at baseline
+  - RTX 3090 (128K ctx, qwen3.6-27B-code): target 75+ tok/s — currently at baseline
-  - RTX 5070 (131K ctx, gemma-4-12b): target 76+ tok/s — optimal for vision/web role
+  - RTX 5070 (128K ctx, gemma-4-12b): target 76+ tok/s — optimal for vision/web role
-  - Strix Halo (256K ctx, strix-moe / qwen3.6-35B-udq4): target 70+ tok/s — currently above baseline
+  - Strix Halo (128K ctx, strix-moe / qwen3.6-35B-udq4): target 70+ tok/s — currently above baseline
 - **Fix**:
   - If tok/s > baseline → context has headroom, consider increasing
   - If tok/s < 90% baseline → reduce context by 25% and retest
@@ -131,9 +131,9 @@ depends_on:
 ### Rule 10: Workload Distribution Optimization
 - **Detect**: GPU roles misaligned with hardware capabilities
 - **Target distribution**:
-  - RTX 3090 (24GB, 256K, 75 tok/s) → Heavy reasoning, code gen, long conversations
+  - RTX 3090 (24GB, 128K, 75 tok/s) → Heavy reasoning, code gen, long conversations
-  - RTX 5070 (12GB, 131K, 76 tok/s) → Vision/image, web search, quick lightweight tasks
+  - RTX 5070 (12GB, 128K, 76 tok/s) → Vision/image, web search, quick lightweight tasks
-  - Strix Halo (64GB, 256K, 72 tok/s) → Context compression, summarization, long docs
+  - Strix Halo (64GB, 128K, 72 tok/s) → Context compression, summarization, long docs
 - **Fix**:
   - Alert if any GPU is handling workload outside its designated role
   - Recommend Hermes agent profile updates to match workload to GPU

hermes-config-template.prose.md

@@ -6,11 +6,10 @@ description: >
   Enforces shared infrastructure setup (Firecrawl, SearXNG, local models,
   RA-H OS MCP) while keeping agent-specific API keys and model choices.
   UPDATED 2026-07-16: Compression model is the stable alias `strix-moe` (NOT `ornith-1.0-35b`,
-  which LiteLLM does not serve). All 3 GPUs verified 256K (RTX 5070 bumped 131K→256K on Jul 15).
+  which LiteLLM does not serve). All 3 GPUs verified at 128K (reduced from 256K 2026-07-17 for stability).
   Added Rule 12 (Context-Issue Diagnostic) + Rule 13 (.env fallback enforcement) from the
   2026-07-16 Mumuni root-cause investigation (WAL #1300).
-  UPDATED 2026-07-12: GPU workload redistributed. Compression → Strix Halo. RTX 3090 context
+  UPDATED 2026-07-12: GPU workload redistributed. Compression → Strix Halo. RTX 3090 context verified at 128K. Infisical .env fallback required (Rule 3/13).
-  verified at 256K. Infisical .env fallback required (Rule 3/13).
 ---
 
 ## Maintains
@@ -36,7 +35,7 @@ Sub-agent profiles inherit auth from the main config — no separate keys needed
 | Mumuni | `mumuni` | 192.168.68.123 | root@.123 | 6 profiles ✱ |
 | Abiba | `abiba-pi` | 192.168.68.24 | local | — |
 | Koby | `koby` | CT 111 (tdunna) | Zulip | — |
-| Koonimo | `koonimo` | CT 113 (baggy) | Zulip | — |
+| Koonimo | `koonimo` | CT 114 (baggy) | SSH root | — |
 | Kagenz0 | `kagenz0-*` | ? | Zulip | — |
 
 > CT hostnames (tdunna, baggy) differ from agent identities (koby, koonimo).
@@ -101,8 +100,8 @@ model:
   base_url: http://192.168.68.116/v1
   api_key_env: LITELLM_API_KEY    # Injected via infisical run -- wrapper
   max_tokens: 4096                # ⚠️ CRITICAL: Prevents unbounded generation
-  context_length: 262144          # For syslog-auto (all GPUs support 256K).
+  context_length: 131072          # For syslog-auto (all GPUs at 128K for stability).
-                                  # Set 131072 if using gemma-4-12b directly (12GB VRAM constraint).
+                                  # Set 65536 if using gemma-4-12b directly (tight VRAM).
 
 fallback_providers:
   provider: deepseek
@@ -133,7 +132,7 @@ compression:
   enabled: true
   model: strix-moe               # ⚠️ Must match auxiliary.compression.model. Stable alias (gpu-fleet § Stable Role-Based Aliases). NOT ornith-1.0-35b (LiteLLM does not serve that name).
   provider: harness
-  max_context_window: 262144      # MUST match actual GPU capacity. All 3 GPUs are 256K (Jul 15).
+  max_context_window: 131072      # MUST match actual GPU capacity. All 3 GPUs are 128K (Jul 17).
   threshold: 0.65                 # Fires at ~170K for 262K window, ~85K for 128K
   target_ratio: 0.30
   protect_last_n: 40
@@ -250,7 +249,7 @@ The following MUST be identical across ALL profiles:
 
 ### Rule 7: Auxiliary Model Consistency (UPDATED 2026-07-16)
 - Vision and web_extract use `gemma-4-12b` (RTX 5070 — 12GB, vision-optimized)
-- Compression uses `strix-moe` (stable alias for Strix Halo — 64GB, 256K ctx, compression-optimized)
+- Compression uses `strix-moe` (stable alias for Strix Halo — 64GB, 128K ctx, compression-optimized)
 - **`strix-moe` is the only valid compression model name** — LiteLLM does NOT serve `ornith-1.0-35b`
   (it serves `strix-moe`, `qwen3.6-35B-udq4`, `gpu-dense`, `gpu-light`, `syslog-auto`, `gemma-4-12b`, `qwen3.6-27B-code`). Old configs with `ornith-1.0-35b` cause 403/model-not-found on compression calls.
 - All auxiliary services MUST use identical routing:
@@ -258,31 +257,31 @@ The following MUST be identical across ALL profiles:
   - `api_key_env: LITELLM_API_KEY`
 - **Do NOT use `syslog-auto`** for auxiliary tasks — it routes unpredictably
 - **Compression on Strix Halo**: The strix-moe alias routes to Strix Halo
-  (64GB UMA, 256K context) — the designated compression GPU. This frees the
+  (64GB UMA, 128K context) — the designated compression GPU. This frees the
   RTX 5070 for vision and web search, and the RTX 3090 for heavy reasoning.
 - The `compression:` block's `model` MUST match `auxiliary: compression: model`
-- The `compression: max_context_window: 262144` MUST match Strix Halo's actual capacity
+- The `compression: max_context_window: 131072` MUST match actual GPU capacity (128K)
 
 ### Rule 8: GPU Workload Distribution (UPDATED 2026-07-16)
-- **RTX 3090 (24GB, 256K ctx, qwen3.6-27B-code)**: Heavy reasoning, code gen, long conversations
+- **RTX 3090 (24GB, 128K ctx, qwen3.6-27B-code)**: Heavy reasoning, code gen, long conversations
-- **RTX 5070 (12GB, 256K ctx, gemma-4-12b)**: Vision, web search, quick tasks, web_extract (bumped 131K→256K Jul 15; IQ4_NL+MTP, 88% VRAM)
+- **RTX 5070 (12GB, 128K ctx, gemma-4-12b)**: Vision, web search, quick tasks, web_extract (IQ4_NL+MTP, ~65% VRAM at 128K)
-- **Strix Halo (64GB, 256K ctx, strix-moe)**: Context compression, summarization, long docs
+- **Strix Halo (64GB, 128K ctx, strix-moe)**: Context compression, summarization, long docs
 - Agent profiles MUST route auxiliary tasks to the correct GPU:
   - `auxiliary.vision.model: gemma-4-12b` (RTX 5070)
   - `auxiliary.web_extract.model: gemma-4-12b` (RTX 5070)
   - `auxiliary.compression.model: strix-moe` (Strix Halo)
 - Default model (`model.default`) and custom_provider remain `syslog-auto` for auto-routing
-- For 262K context window: `threshold: 0.65` (fires at ~170K tokens)
+- For 128K context window: `threshold: 0.65` (fires at ~85K tokens)
 - Do NOT use `threshold: 0.25` — this fires at 65K, causing premature context loss
-- Do NOT use `threshold: 0.80` — this delays until 209K, risking the gateway hygiene layer
+- Do NOT use `threshold: 0.80` — this delays until ~105K, leaving only 23K margin
-- `max_context_window: 262144` MUST match the model's actual capacity
+- `max_context_window: 131072` MUST match the model's actual capacity (128K)
 - See `devops-hermes-compression` skill for full reference
 
-### Rule 9: Compression Threshold for 256K Models
+### Rule 9: Compression Threshold for 128K Models
-- For 262K context window: `threshold: 0.65` (fires at ~170K tokens)
+- For 128K context window: `threshold: 0.65` (fires at ~85K tokens)
 - Do NOT use `threshold: 0.25` — this fires at 65K, causing premature context loss
-- Do NOT use `threshold: 0.80` — this delays until 209K, risking the gateway hygiene layer
+- Do NOT use `threshold: 0.80` — this delays until ~105K, leaving only 23K margin
-- `max_context_window: 262144` MUST match the model's actual capacity (Strix Halo = 256K)
+- `max_context_window: 131072` MUST match the model's actual capacity (all GPUs = 128K)
 - See `devops-hermes-compression` skill for full reference
 
 ### Rule 10: Default Model Must Be `syslog-auto` (All Agents)
@@ -312,7 +311,7 @@ When an agent shows "context issues" (premature compression, 401s, 504s, DeepSee
 verify ALL FOUR of these against the live config. They are the only root causes found in production:
 
 1. **max_context_window correct?** — BOTH `compression.max_context_window` AND `context.max_context_window`
-   MUST be `262144` (all GPUs are 256K). A value of `131072` causes premature compression at
+   MUST be `131072` (all GPUs are 128K). A value of `262144` causes instability near 100K and must NOT be used.
    ~83K instead of ~170K. Check: `grep -n max_context_window ~/.hermes/config.yaml`
 2. **base_url uses /v1 NOT /litellm/v1?** — `custom_providers[0].base_url`, `delegation.base_url`,
    and ALL `auxiliary.*.base_url` MUST be `http://192.168.68.116/v1` (Rule 5). nginx `/litellm/`

inference-optimization.prose.md

@@ -0,0 +1,100 @@
+---
+name: inference-optimization
+kind: responsibility
+id: 067NC6KP02RG60S50M40E30928
+---
+
+### Goal
+
+Syslog inference response times reduced to sub-15s average by optimizing the full
+stack: LiteLLM routing weights, GPU model assignments, Hermes agent context
+management, and prompt caching — without sacrificing agent capability.
+
+### Requires
+
+- `inference-metrics`: current SpendLogs from CT116 LiteLLM Postgres — avg
+  request_duration_ms, prompt_tokens, completion_tokens, model_group breakdown,
+  cache_hit rate over the last 3 hours
+- `agent-configs`: current config.yaml from each active Hermes agent (Mumuni
+  .123, any others on .129/.122) including compression, model, context_window,
+  prompt_caching, memory settings
+- `gpu-health`: health check response from all 3 GPU backends (ornith .15:8080,
+  qwen .8:8080, gemma .110:8080)
+
+### Maintains
+
+The optimized inference stack configuration — every change is applied and
+verified end-to-end. Postcondition: avg request_duration_ms ≤ 15000 for 90% of
+non-ornith traffic; ≤ 30000 for ornith-bound agentic calls.
+
+#### liteLLM-routing
+The syslog-auto routing weights, model-specific timeouts, RPM limits, and
+model_list entries on CT116 `/opt/inference-harness/litellm_config.yaml`.
+
+#### agent-compression
+Each Hermes agent's `~/.hermes/config.yaml` compression, context_window,
+prompt_caching, and model sections.
+
+#### prompt-caching
+LiteLLM cache configuration and llama.cpp `--cache-prompt` flag on GPU hosts.
+
+#### verification
+End-to-end latency measurements after changes applied — at least 3 test
+inference calls per model path measuring ttft (time-to-first-token) and total
+duration.
+
+### Continuity
+
+- input-driven
+
+### Strategies
+
+**Context is the root cause.** Every ~46K prompt token costs ~87s of ornith
+prefill time at 532 tok/s. Fix context first, routing second.
+
+- **Route by task**: ornith for multi-step reasoning only; qwen for code/standard
+  queries; gemma for compression/auxiliary. Never send simple completion to a
+  35B MoE.
+- **Compress aggressively**: threshold at 40% (not 65%) — a 256K window should
+  compact at 102K, not 166K. Target 15% tail (not 30%).
+- **Cache everything repeated**: system prompts, skill docs, AGENTS.md — these
+  never change between turns. Single-digit cache hit rate is unacceptable.
+- **Lower context ceiling**: 128K window is the stable ceiling for agent conversations.
+  GPUs reduced from 256K to 128K (2026-07-17). For larger contexts, route to external providers.
+
+### Shape
+
+- `self`: analyze metrics, compute optimal configs, apply changes, verify
+- `delegates`:
+  - `apply-liteLLM`: update litellm_config.yaml and reload
+  - `apply-agent-config`: update hermes config.yaml per agent
+  - `verify-latency`: run test inference calls and measure response
+
+### Execution
+
+```prose
+-- Phase 1: Analyze current state (already complete)
+-- Phase 2: Apply LiteLLM routing optimization
+
+call apply-liteLLM-routing
+  config_path: /opt/inference-harness/litellm_config.yaml
+  host: 192.168.68.116
+
+-- Phase 3: Apply agent context compression optimization  
+
+call apply-agent-compression
+  agent: mumuni
+  host: 192.168.68.123
+  config_path: /root/.hermes/config.yaml
+
+-- Phase 4: Enable llama.cpp prompt caching on GPU hosts
+
+call enable-prompt-caching
+  hosts: [192.168.68.15, 192.168.68.8, 192.168.68.110]
+
+-- Phase 5: Verify end-to-end latency
+
+call verify-latency
+  host: 192.168.68.116
+  models: [syslog-auto, qwen3.6-27B-code, gemma-4-12b, ornith-1.0-35b]
+```