Merge pull request 'fix: LiteLLM contracts — align with 2026-07-08 architecture changes' (#8) from fix/litellm-contracts-jul2026 into master

litellm-api-keys.prose.md

@@ -8,6 +8,10 @@ description: >
   Ensures agents never use the master key directly. Rotation is event-driven,
   not calendar-driven — rotate only on compromise, personnel change, or
   periodic security hygiene (quarterly/annually).
+  
+  Current key inventory and agent list: see gpu-fleet.prose.md § Agent Keys.
+  Source of truth for LiteLLM config: /opt/inference-harness/litellm_config.yaml
+  on CT 116. Last verified: 2026-07-09.
 ---
 
 ## Parameters
@@ -38,7 +42,8 @@ description: >
    - Generate new key with key_alias: "{agent_name}" (e.g., "tanko" — bare name, no date)
    - Set metadata: { "agent": "{agent_name}", "purpose": "agent-inference" }
    - Duration is null (permanent) — inherited from litellm default_key_generate_params
-   - Set models: ["syslog-auto", "qwen3.6-35B-A3B", "qwen3.6-27B-code", "gemma-4-12b"]
+   - Set models: ["syslog-auto", "qwen3.6-27B-code", "gemma-4-12b", "ornith-1.0-35b"]
+   - Note: qwen3.6-35B-A3B removed from fleet (was never deployed on any GPU)
    - Return the new key
 5. **If action == "rotate"**:
    - Generate new key with same alias (LiteLLM replaces the old key)

litellm-health.prose.md

@@ -2,34 +2,41 @@
 kind: function
 name: litellm-health
 description: >
-  Verifies the LiteLLM inference stack health. Current architecture (2026-06-30):
-  nginx:80 → LiteLLM:4000 → Router:9000(internal) → GPU(llama-server).
-  Router runs as internal-only backend behind LiteLLM. GPU monitoring via
-  Prometheus/Grafana and fleet dashboard.
+  Verifies the LiteLLM inference stack health. Current architecture (2026-07-09):
+  nginx:80 → LiteLLM:4000 → GPU(llama-server) via direct proxy.
+  Router (harness-router :9000) is DEPRECATED — container still runs but
+  is not in the request path. GPU monitoring via Prometheus/Grafana and
+  fleet dashboard (gpu-monitor :9100).
   Designed as a reusable contract for any Syslog agent.
+  
+  Source of truth: gpu-fleet.prose.md
 ---
 
-## Architecture (v3.2.0 — Layered: nginx → LiteLLM → Router → GPU)
+## Architecture (v4.0.0 — Direct: nginx → LiteLLM → GPU)
 
 ```
-Request → nginx:80 → LiteLLM:4000 → Router:9000(internal) → GPU(llama-server)
-                        │                    │
-                  Key validation      Model routing
-                  Fallback chains     Slot booking
-                  Complexity router   Circuit breakers
-                  Budget tracking     Redis-backed
-                        │                    │
-                  Prometheus ← metrics  ←───┘
+Request → nginx:80 → LiteLLM:4000 → GPU(llama-server, parallel 2)
+                        │
+                  Key validation
+                  Fallback chains
+                  Budget tracking
+                        │
+                  Prometheus ← metrics
                         │
                     Grafana :3001
+
+  harness-router :9000 — DEPRECATED, container still runs but
+  NOT in request path. nginx routes /v1 → LiteLLM directly.
+  Router slot booking + circuit breakers replaced by
+  LiteLLM native fallbacks + timeouts.
 ```
 
-**What changed (v3.1.0 → v3.2.0)**:
-- Router is BACK — runs internally on :9000 behind LiteLLM
-- nginx routes /v1/ and /admin/ → LiteLLM (not router directly)
-- Router :9000 is 127.0.0.1-only, not publicly accessible
-- GPU fleet dashboard: http://192.168.68.24:9100
-- Grafana: http://192.168.68.116:3001 (GPU dashboards via Prometheus)
+**What changed (v3.2.0 → v4.0.0 — 2026-07-08)**:
+- Router REMOVED from request path — LiteLLM proxies directly to GPU
+- All GPUs at parallel 2 (was parallel 1)
+- NVIDIA context reduced 256K→128K to free VRAM
+- LiteLLM timeouts tuned: gemma 25→120s, qwen 40→90s
+- nginx proxy_read_timeout: 600s, LiteLLM request_timeout: 300s
 
 ## Parameters
 
@@ -55,27 +62,29 @@ Request → nginx:80 → LiteLLM:4000 → Router:9000(internal) → GPU(llama-se
 
 ## GPU Fleet Topology
 
-| Host | IP | Hardware | Models Served | Engine |
-|------|-----|----------|---------------|--------|
-| llm-gpu | 192.168.68.8 | NVIDIA RTX 3090 (24 GB) | qwen3.6-27B-code | llama-server Docker |
-| ocu-llm | 192.168.68.110 | NVIDIA RTX 5070 (12 GB) | gemma-4-12b | llama-server Docker |
-| amdpve | 192.168.68.15 | AMD Strix Halo (CPU) | ornith-1.0-35b (35B) | llama-server bare-metal |
+| Host | IP | Hardware | Models Served | Engine | Context | Parallel |
+|------|-----|----------|---------------|--------|---------|----------|
+| llm-gpu | 192.168.68.8 | NVIDIA RTX 3090 (24 GB) | qwen3.6-27B-code | llama-server systemd | 128K | 2 |
+| ocu-llm | 192.168.68.110 | NVIDIA RTX 5070 (12 GB) | gemma-4-12b | llama-server systemd | 128K | 2 |
+| amdpve | 192.168.68.15 | AMD Strix Halo 64GB UMA | ornith-1.0-35b | llama-server systemd (Vulkan) | 256K | 2 |
 
 ## Model Fallback Chains (LiteLLM)
 
 | Primary | Timeout | Fallback | Timeout |
 |---------|---------|----------|---------|
-| qwen3.6-27B-code | 45s | gemma-4-12b | 30s |
-| gemma-4-12b | 30s | qwen3.6-27B-code | 45s |
+| qwen3.6-27B-code | 90s | gemma-4-12b | 120s |
+| gemma-4-12b | 120s | qwen3.6-27B-code | 90s |
 | ornith-1.0-35b | 120s | qwen → gemma | — |
-| syslog-auto (balanced) | — | qwen → gemma | — |
+| syslog-auto (balanced) | 90s | qwen → gemma | — |
+
+> Global: request_timeout=300s, nginx proxy_read_timeout=600s
 
 ## Containers on CT 116
 
 | Container | Image | Port | Health Check |
 |-----------|-------|------|-------------|
 | harness-litellm | berriai/litellm:1.90.0-rc.1 | :4000→:4001 | /health/liveliness |
-| harness-router | inference-harness-router | :9000 (127.0.0.1) | /health |
+| harness-router | inference-harness-router | :9000 (127.0.0.1) | /health (DEPRECATED — not in path) |
 | harness-nginx | nginx:alpine | :80 | HTTP 200 on /health |
 | harness-postgres | postgres:16-alpine | :5432 | pg_isready |
 | harness-redis | redis:7-alpine | :6379 | PING |

litellm-self-heal.prose.md

@@ -6,6 +6,9 @@ note: >
   Auto-remediation code was removed from the pi Zulip extension (retired 2026-07-04).
   This contract is now manual-only — triggers require explicit user request.
   Consider reimplementing as a standalone cron job or prose contract.
+  
+  Source of truth for GPU topology: gpu-fleet.prose.md
+  Last verified: 2026-07-09
 description: >
   Standing responsibility that monitors LiteLLM health and proactively
   fixes common issues. Reports every action via Zulip DM and RA-H OS
@@ -73,11 +76,11 @@ message is sent to the responsible agent with full context.
 
 ## GPU Fleet
 
-| Host | IP | Hardware | Models | Engine |
-|------|-----|----------|--------|--------|
-| amdpve | 192.168.68.15 | Strix Halo 64GB UMA | ornith-1.0-35b (16t, 262K ctx) | llama-server :8080 |
-| llm-gpu | 192.168.68.8 | RTX 3090 24GB | qwen3.6-27B-code | llama-server :8080 |
-| ocu-llm | 192.168.68.110 | RTX 5070 12GB | gemma-4-12b | llama-server :8080 |
+| Host | IP | Hardware | Models | Engine | Context |
+|------|-----|----------|--------|--------|---------|
+| amdpve | 192.168.68.15 | Strix Halo 64GB UMA | ornith-1.0-35b | llama-server :8080 (Vulkan) | 256K |
+| llm-gpu | 192.168.68.8 | RTX 3090 24GB | qwen3.6-27B-code | llama-server :8080 (systemd) | 128K |
+| ocu-llm | 192.168.68.110 | RTX 5070 12GB | gemma-4-12b | llama-server :8080 (systemd) | 128K |
 
 ## Remediation Rules
 
@@ -105,12 +108,10 @@ Detect → test prompt via LiteLLM returns non-200
 Fix → restart llama-server on GPU host → verify
 Escalate → after 2 failed restarts
 
-### Rule 7: Router Returns 503 (All GPUs Saturated)
-Detect → router logs show QUEUE_TIMEOUT with GPUs idle
-Root cause → roster not loaded at startup (empty TIER_MODELS)
-Fix → POST /admin/roster/reload (if endpoint available) or restart router
-Permanent fix → add load_roster() to router startup code
-Escalate → if reload doesn't resolve within 2 min
+### Rule 7: Router Returns 503 (All GPUs Saturated) — DEPRECATED
+Router is no longer in the request path (2026-07-08). LiteLLM proxies
+directly to GPU. This rule is retained for reference but is inactive.
+If 503 errors occur, check LiteLLM timeouts and GPU health directly.
 
 ### Rule 8: Agent Keys Invalid (401)
 Detect → agents report auth errors