Merge pull request 'UPDATE 2026-07-15: Full GPU fleet rebuild + stable aliases' (#18) from feat/gpu-fleet-rebuild-20260715 into master

Reviewed-on: SyslogSolution/prose-contracts#18
2026-07-16 21:17:12 +00:00
parent 190ceb9be4 369c312ccb
commit d6e85e68e9
13 changed files with 1136 additions and 107 deletions
@@ -0,0 +1,309 @@
+---
+kind: function
+name: abiba-zulip-restore
+description: >
+  Restores Zulip connectivity for Abiba (pi agent). Verifies the v2 router-worker
+  extension code, starts a PM2 process as the Zulip gateway with correct env vars,
+  validates health endpoint, and confirms DM delivery. Run this whenever Abiba
+  stops responding on Zulip or after system restart.
+agent: abiba
+version: 1.0.0
+status: active
+runtime_contract: 2
+---
+
+# Abiba Zulip Restore — Resume pi Zulip Communication
+
+Single-shot function that restores full Zulip connectivity for the Abiba pi agent.
+Covers extension code validation, PM2 process management, health endpoint
+verification, and DM loopback testing.
+
+## Live-State Fields
+
+| Field | Value | Trust |
+|-------|-------|-------|
+| Agent name | abiba | ✅ |
+| Bot email | abiba-bot@chat.sysloggh.net | ✅ |
+| Zulip server | https://chat.sysloggh.net | ✅ |
+| Extension path | /root/.pi/agent/extensions/zulip/index.js | ✅ |
+| Config path | /root/.pi/agent/extensions/zulip/config.yaml | ✅ |
+| Health port | 9200 | ✅ |
+| @all-bots user ID | 20 | ✅ (config, verified by API at runtime) |
+| PM2 process name | abiba-zulip | ✅ |
+| Provider | syslog-harness (http://192.168.68.116/v1) | ✅ |
+| Default model | deepseek-v4-pro | ✅ (settings.json) |
+
+## Architecture
+
+The pi Zulip extension uses a **router-worker architecture** (v2):
+
+- **Router** (PM2 `abiba-zulip`, `ZULIP_ROLE=router`): Polls Zulip for events,
+  maintains a per-sender pool of pi RPC worker processes. Each sender gets
+  their own `pi --mode rpc --session-dir` process with persistent sessions.
+  Handles streaming edits back to Zulip.
+- **Worker** (child `pi --mode rpc`): Runs the agent with per-sender persistent
+  sessions. No Zulip logic in the worker — the router handles all Zulip I/O.
+
+The extension loads in ALL pi sessions (via `settings.json` extensions array)
+but is a **NO-OP** unless `ZULIP_ROLE=router` is set. Only the PM2 router process
+carries the env var.
+
+## Maintains
+
+- extension_valid: bool — Extension code imports without errors
+- pm2_running: bool — PM2 process `abiba-zulip` is online
+- health_responding: bool — GET :9200/health returns "ok"
+- zulip_connected: bool — Queue registered, bot identity resolved
+- loopback_delivered: bool — Test DM sent and received
+- model_valid: bool — Configured models match LiteLLM authorized models
+
+### Postconditions
+
+- PM2 process `abiba-zulip` online and stable (uptime > 30s)
+- Health endpoint returns `{ status: "ok", connected: true }`
+- Worker pool creates sessions on demand
+- Echo prevention active (BOT_EMAILS includes all known bots)
+- PM2 saved for auto-restart on boot
+
+## Requires
+
+- Node.js with `yaml` module available
+- PM2 installed globally
+- Zulip API key in `config.yaml`
+- `pi` CLI available on PATH
+- Zulip server accessible at https://chat.sysloggh.net
+- LiteLLM provider accessible at http://192.168.68.116/v1
+
+## Execution
+
+### Step 1: Validate Extension Code
+
+```bash
+node -e "import('file:///root/.pi/agent/extensions/zulip/index.js').then(() => console.log('OK')).catch(e => {console.error('FAIL:', e.message); process.exit(1)})"
+```
+
+Expected: `OK`. If FAIL → check for missing dependencies, syntax errors.
+
+### Step 2: Validate Model IDs
+
+Compare configured models against LiteLLM authorized models:
+
+```bash
+API_KEY=$(grep -oP 'apiKey:\s*\K.*' /root/.pi/agent/models.json | head -1)
+curl -s -H "Authorization: Bearer $API_KEY" http://192.168.68.116/v1/models | \
+  python3 -c "import json,sys; d=json.load(sys.stdin); [print(m['id']) for m in d.get('data',[])]" 2>/dev/null
+```
+
+Check: Every model ID in `models.json` must appear in the authorized list.
+If not → fix `models.json` to only include authorized models (prefer `syslog-auto`).
+
+### Step 3: Verify Config Integrity
+
+```bash
+python3 -c "
+import yaml, sys
+with open('/root/.pi/agent/extensions/zulip/config.yaml') as f:
+    cfg = yaml.safe_load(f)
+required = ['zulip.site', 'zulip.email', 'zulip.api_key', 'agent.name']
+for k in required:
+    keys = k.split('.')
+    v = cfg
+    for kk in keys:
+        v = v.get(kk)
+        if v is None:
+            print(f'MISSING: {k}')
+            sys.exit(1)
+print('Config valid')
+print(f'  site={cfg[\"zulip\"][\"site\"]}')
+print(f'  email={cfg[\"zulip\"][\"email\"]}')
+print(f'  agent={cfg[\"agent\"][\"name\"]}')
+print(f'  health_port={cfg.get(\"health_port\", 9200)}')
+"
+```
+
+Expected: Config valid with all fields non-empty.
+
+### Step 4: Remove Stale Systemd Service
+
+The old `abiba-zulip.service` points to `/opt/abiba-zulip/dist/index.js` (compiled
+TypeScript, not the v2 extension). It's disabled and stale. Remove it:
+
+```bash
+systemctl stop abiba-zulip 2>/dev/null || true
+systemctl disable abiba-zulip 2>/dev/null || true
+rm -f /etc/systemd/system/abiba-zulip.service
+systemctl daemon-reload
+```
+
+### Step 5: Start PM2 Process
+
+**Critical:** The extension MUST run via `pi --mode rpc`, NOT `node index.js` directly.
+The extension exports a function that requires pi's session lifecycle. Running
+`node index.js` loads the module but never calls the export, so nothing happens.
+`pi --mode rpc` loads all extensions (including Zulip) and fires `session_start`.
+
+```bash
+# Stop existing if any
+pm2 delete abiba-zulip 2>/dev/null || true
+
+# Start pi in RPC mode with ZULIP_ROLE=router env
+ZULIP_ROLE=router pm2 start "$(which pi)" \
+  --name abiba-zulip \
+  --interpreter none \
+  -- --mode rpc --no-session
+```
+
+Wait 10 seconds for pi to load all extensions, fire session_start, and the Zulip
+router to register its event queue.
+
+### Step 6: Validate PM2 Process
+
+```bash
+pm2 show abiba-zulip --no-color
+```
+
+Check: `status=online`, `restarts=0`, `uptime > 5s`.
+
+### Step 7: Check Logs for Connection
+
+```bash
+sleep 3
+tail -20 /root/.pm2/logs/abiba-zulip-out.log
+```
+
+Look for:
+- `[zulip-ext] Connecting to https://chat.sysloggh.net as abiba-bot@chat.sysloggh.net…`
+- `[zulip-ext] Bot user_id=N, all-bots user_id=N`
+- `[zulip-ext] Connected, queue=N`
+- `[zulip-ext] Echo prevention: N bot emails`
+- `[zulip-ext] Health endpoint on :9200`
+
+If error → check API key, network to chat.sysloggh.net.
+
+### Step 8: Health Endpoint
+
+```bash
+curl -s http://localhost:9200/health | python3 -m json.tool
+```
+
+Check:
+- `status: "ok"` (not "down")
+- `zulip.connected: true`
+- `zulip.queue_id` is non-null string
+- `zulip.bot_user_id` is positive integer
+
+### Step 9: DM Loopback Test
+
+```bash
+curl -s http://localhost:9200/health | python3 -c "
+import json,sys
+d = json.load(sys.stdin)
+if d.get('zulip',{}).get('connected'):
+    print(f'✅ Zulip connected. Queue: {d[\"zulip\"][\"queue_id\"]}')
+    print(f'   Bot user_id: {d[\"zulip\"][\"bot_user_id\"]}')
+    print(f'   Messages processed: {d[\"zulip\"][\"messages_processed\"]}')
+else:
+    print('❌ Zulip NOT connected')
+    sys.exit(1)
+"
+```
+
+### Step 10: Save PM2 for Auto-Start
+
+```bash
+pm2 save
+pm2 startup systemd -u root --hp /root 2>/dev/null || true
+```
+
+### Step 11: Report
+
+Compile results:
+
+| Check | Pass? |
+|-------|-------|
+| Extension code imports | extension_valid |
+| Model IDs authorized | model_valid |
+| Config integrity | config_valid |
+| PM2 process online | pm2_running |
+| Health endpoint | health_responding |
+| Zulip connected | zulip_connected |
+
+All pass → ✅ **Abiba Zulip restored.** Relay success to user.
+
+Partial failure → see recovery matrix below.
+
+## Recovery Matrix
+
+| Failure | Recovery |
+|---------|----------|
+| Extension import fails | Check `node_modules/zulip-js` exists; run `npm install` in extension dir |
+| Model ID mismatch | Fix `models.json` to use `syslog-auto` as default model; remove invalid IDs |
+| Config missing | Restore from backup or recreate from scratch |
+| PM2 won't start | Check `node` version (>=18); check port 9200 not in use |
+| Health "down" | Check logs for connection errors; verify Zulip API key; check network |
+| "address already in use" | Kill old process: `fuser -k 9200/tcp` |
+| Queue registration fails | Check Zulip API key validity; verify bot is active in Zulip admin |
+| Rate limit (429) | Extension has built-in retry-after handling — wait, don't restart |
+
+## Known Failure Modes
+
+| Symptom | Root Cause | Recovery |
+|---------|-----------|----------|
+| Extension loads but no events | `ZULIP_ROLE` not set | Ensure PM2 env has `ZULIP_ROLE=router` |
+| Worker stays "busy" forever | Model ID not authorized by LiteLLM | Fix models.json (lesson #11) |
+| Placeholder sent but no response | editMessage API fails silently | Extension has fallback (sends new msg); check Zulip API |
+| Queue expires rapidly | Poll interval too aggressive | v2 uses 3s poll with long-poll — should be fine |
+| Bot doesn't respond to @mentions | Not subscribed to stream | Bot auto-subscribes via API |
+| Stale error in health | `last_error` not cleared | v2 clears on successful poll (lesson #4) |
+
+## Appendix: Root Cause & Fix Summary (2026-07-13)
+
+**Problem:** Zulip extension was offline — no PM2 process running.
+
+**Root cause:** The PM2 command ran `node index.js` directly (which loads the
+extension module but never calls the exported function). The extension requires
+pi's session lifecycle — pi loads extensions, fires `session_start`, and the
+Zulip extension hooks into that event.
+
+**Fix:** Run `pi --mode rpc` (not `node index.js`). The `--mode rpc` flag keeps
+pi alive listening for RPC commands on stdin while the Zulip extension's router
+runs in the background via the `session_start` hook.
+
+```bash
+ZULIP_ROLE=router pm2 start "$(which pi)" --name abiba-zulip \
+  --interpreter none -- --mode rpc --no-session
+```
+
+**Additional fixes applied:**
+- Fixed `models.json`: replaced `qwen3.6-35B-A3B` (not authorized by LiteLLM)
+  with `syslog-auto` + `strix-moe` (prevents silent worker failure per
+  Lesson #11)
+- Removed stale systemd unit `abiba-zulip.service` (pointed to old TS code)
+- PM2 saved for auto-restart on boot
+
+## Appendix: PM2 Ecosystem Config (Optional)
+
+If preferred over manual `pm2 start`, create `/root/ecosystem.config.js` entry:
+
+```js
+module.exports = {
+  apps: [{
+    name: 'abiba-zulip',
+    script: '/bin/pi',
+    interpreter: 'none',
+    args: '--mode rpc --no-session',
+    cwd: '/root',
+    env: {
+      ZULIP_ROLE: 'router',
+    },
+    log_file: '/root/.pm2/logs/abiba-zulip-out.log',
+    error_file: '/root/.pm2/logs/abiba-zulip-error.log',
+    max_restarts: 20,
+    restart_delay: 5000,
+  }]
+};
+```
+
+---
+
+**Last verified good state**: 2026-07-13 — Extension v2 running via `pi --mode rpc`, health endpoint :9200 returning `{status:"ok",connected:true}`, queue a669f21e.
@@ -72,10 +72,10 @@ reads them directly.
 |--------|-------|----------|------|----------|
 | `syslog-code` | qwen3.6-27B-code | terminal, file, web, memory, skills | Code patches, automation, scripts | Writing/modifying code, creating scripts, debugging, reading/writing files |
 | `syslog-devops` | qwen3.6-27B-code | terminal, file, web, memory, skills | Infrastructure, DB, bridge, Proxmox | Server ops, SSH, Docker, Proxmox, DB queries, hardware checks |
-| `syslog-email` | ornith-1.0-35b | terminal, file, web, memory, skills | Email automation, mail operations | Sending/receiving email, inbox management, SMTP operations |
-| `syslog-research` | ornith-1.0-35b | terminal, file, web, memory, skills, **browser** | Analysis, classification, data processing | Web research, browser tasks, data analysis, classification, reading docs |
-| `syslog-review` | ornith-1.0-35b | terminal, file, web, memory, skills | Verification, QA, audit validation | **ALWAYS** verify worker output before delivery — especially for infra changes, code builds, and research findings |
-| `syslog-writer` | ornith-1.0-35b | terminal, file, web, memory, skills | Docs, content, branding, reports | Writing docs, reports, proposals, content, markdown formatting |
+| `syslog-email` | strix-moe | terminal, file, web, memory, skills | Email automation, mail operations | Sending/receiving email, inbox management, SMTP operations |
+| `syslog-research` | strix-moe | terminal, file, web, memory, skills, **browser** | Analysis, classification, data processing | Web research, browser tasks, data analysis, classification, reading docs |
+| `syslog-review` | strix-moe | terminal, file, web, memory, skills | Verification, QA, audit validation | **ALWAYS** verify worker output before delivery — especially for infra changes, code builds, and research findings |
+| `syslog-writer` | strix-moe | terminal, file, web, memory, skills | Docs, content, branding, reports | Writing docs, reports, proposals, content, markdown formatting |

 ### Selection Rules

@@ -5,13 +5,12 @@ description: >
  Manages the GPU inference fleet across all hosts. Handles model deployment,
  registration, health checks, LiteLLM sync, agent key management, GPU
  saturation watchdog, Prometheus/Grafana monitoring, and self-healing.
-  UPDATED 2026-07-12: Architecture is DIRECT GPU — LiteLLM routes directly
-  to llama-server on each GPU host (no router in inference path). Router
-  (port 9000) is running but NOT in request path. All GPUs standardized on
-  api-key 'not-needed'. RTX 5070 had api-key mismatch (sk-loc...5678) that
-  caused cascading 401→timeout→401 fallback loops — fixed.
-  Context: RTX 3090 verified at 256K (was documented as 128K — WRONG).
-  Workload: Compression moved to Strix Halo, RTX 5070 → vision/web only.
+  UPDATED 2026-07-15: Stable role-based aliases introduced: strix-moe,
+  gpu-dense, gpu-light. These never change — only the underlying model does.
+  Strix Halo: ornith-1.0-35b → 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).
+  Compression timeout: 300s (was 120s). Mumuni context: 128K (was 256K).
 agent: abiba
 triggers:
  - on model add/remove
@@ -68,48 +67,72 @@ triggers:
 │ CT 8     │  │ CT 110      │  │ CT 15     │  │ pi (.24)     │
 │ RTX 3090 │  │ RTX 5070    │  │ Strix Halo│  │ GPU Monitor  │
 │ 24GB     │  │ 12GB        │  │ 64GB UMA  │  │ :9100        │
-│ 256K ctx │  │ 131K ctx    │  │ 256K ctx  │  │ Watchdog     │
-│ qwen3.6  │  │ gemma-4-12b │  │ ornith35B │  │ Prometheus   │
-│ 27B-code │  │ :8080       │  │ :8080     │  │ exporter     │
+│ 256K ctx │  │ 256K ctx    │  │ 256K ctx  │  │ Watchdog     │
+│ qwen3.6  │  │ gemma-4-12b │  │ qwen3.6   │  │ Prometheus   │
+│ 27B-code │  │ :8080       │  │ -35B-udq4 │  │ exporter     │
 │ :8080    │  │ :9400 (exp) │  │ :9400(exp)│  │ :9401        │
 │ :9400    │  └─────────────┘  └───────────┘  └──────────────┘
 └──────────┘
 ```

-## Current Model Assignments (2026-07-12)
+## Stable Role-Based Aliases (Introduced 2026-07-15)
+
+Agent configs, cron jobs, and workflows MUST use these aliases, never model-specific names.
+When a model is swapped on a GPU, ONLY the infrastructure layer changes — agent configs are untouched.
+
+| Alias | GPU | Current Model | Will Route To |
+|-------|-----|---------------|---------------|
+| `strix-moe` | Strix Halo (.15) | qwen3.6-35B-udq4 | Whatever runs on Strix Halo |
+| `gpu-dense` | RTX 3090 (.8) | qwen3.6-27B-code | Whatever runs on RTX 3090 |
+| `gpu-light` | RTX 5070 (.110) | gemma-4-12b | Whatever runs on RTX 5070 |
+
+**Backward compatibility**: Old model-specific names (qwen3.6-27B-code, gemma-4-12b, qwen3.6-35B-udq4) still work
+but are deprecated for agent configs. Only the stable aliases survive model swaps.
+
+## Current Model Assignments (2026-07-15)

 | Model | GPU | Host | VRAM | Ctx | KV Cache | Parallel | Batch/Ubatch | Status |
 |-------|-----|------|------|-----|----------|----------|-------------|--------|
-| qwen3.6-27B-code | RTX 3090 | .8 (llm-gpu) | 20.7/24GB (84%) | **256K** | turbo4 | 1 | default | ✅ healthy |
-| gemma-4-12b | RTX 5070 | .110 (ocu-llm) | 10.0/12.2GB (82%) | 131K | q4_0 | 2 | 2048/1024 | ✅ healthy |
-| ornith-1.0-35b | Strix Halo Vulkan | .15 (amdpve) | ~9GB/64GB | 256K | q8_0 | 2 | 2048/512 | ✅ healthy |
+| qwen3.6-27B-code (MTP) | RTX 3090 | .8 (llm-gpu) | 22.2/24.6GB (90%) | **256K** 🚀 | 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 |
+| qwen3.6-35B-udq4 | Strix Halo Vulkan | .15 (amdpve) | ~9GB/64GB | 256K | q8_0 | 2 | 2048/512 | ✅ healthy |

 ## Routing Configuration (LiteLLM — July 2026)

-### syslog-auto Weighted Pool
+### syslog-auto Weighted Pool (Direct GPU — bypasses router)

-| Model | GPU | Weight | RPM Cap | Purpose |
+| Model | GPU | Weight | RPM Cap | Timeout |
 |-------|-----|--------|---------|---------|
-| qwen3.6-27B-code | RTX 3090 | 0.55 | 500 | Heavy reasoning, code, long context |
-| ornith-1.0-35b | Strix Halo | 0.30 | **60** | Agentic workflows, tool calling |
-| gemma-4-12b | RTX 5070 | 0.15 | 200 | Overflow + vision |
+| qwen3.6-27B-code | RTX 3090 (.8:8080) | **0.55** | 500 | **300s** |
+| qwen3.6-35B-udq4 | 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.

 ### Direct Model Endpoints

 | Model | RPM Cap | Notes |
 |-------|---------|-------|
-| ornith-1.0-35b | 40 | Tight cap — prevents Strix overload |
+| qwen3.6-35B-udq4 | 40 | Tight cap — prevents Strix overload |
 | qwen3.6-27B-code | 500 | High cap — primary workhorse |
-| gemma-4-12b | 500 | High cap — fast 12B |
+| gemma-4-12b | 500 | High cap — IQ4_NL+MTP, 122 tok/s |
+
+### Stable Aliases (for agent configs — never change)
+
+| Alias | RPM Cap | Routes To | Purpose |
+|-------|---------|-----------|---------|
+| `strix-moe` | 40 | Strix Halo | Compression tasks (MoE models) |
+| `gpu-dense` | 500 | RTX 3090 | Heavy reasoning |
+| `gpu-light` | 500 | RTX 5070 | Vision, web extract, light tasks |

 ### Fallback Chains
 - gemma → qwen
 - qwen → gemma
- ornith → qwen → gemma
- syslog-auto → qwen → gemma → ornith
+- qwen3.6-35B-udq4 → qwen → gemma
+- syslog-auto → qwen → gemma → qwen3.6-35B-udq4

-### Why ornith RPM Is Capped
- Direct: 40 RPM (tight) — Strix Halo is shared with compression tasks
+### Why Strix Halo RPM Is Capped
+- Direct (qwen3.6-35B-udq4): 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

@@ -163,7 +186,7 @@ Show full fleet status: GPUs, models, VRAM, context windows, parallel slots, act
 3. Check LiteLLM: `curl http://192.168.68.116/health` (expect "I'm alive!")
 4. Check LiteLLM models: `curl -H "Authorization: Bearer $MASTER_KEY" http://192.168.68.116/v1/models`
 5. Check LiteLLM timeouts: `grep -n 'timeout:' /opt/inference-harness/litellm_config.yaml`
-   - gemma-4-12b: 120s, qwen3.6-27B: 90s, ornith-1.0-35b: 120s
+   - gemma-4-12b: 120s, qwen3.6-27B-code: 300s, qwen3.6-35B-udq4/strix-moe: 300s (ornith-1.0-35b does NOT exist — legacy name, do not use)
   - global request_timeout: 300s, nginx proxy_read_timeout: 600s
 6. Check AMD metrics: `curl http://192.168.68.15:9400/metrics` (Radeon 8060S, util%, VRAM, temp, power)
 7. Check port conflicts: verify only one llama-server on :8080 per host
@@ -203,7 +226,7 @@ If no SSH access, send Zulip DM via abiba-bot with vault update instructions.
 | `/root/scripts/gpu-saturation-watchdog.py` | pi (.24) | Auto-restart stuck llama-server |
 | `/root/dashboard/gpu-fleet.html` | pi (.24) | Live HTML dashboard |
 | `/etc/systemd/system/llama-server.service` | .8, .110 | llama-server daemons (Nvidia GPUs) |
-| `/etc/systemd/system/ornith-server.service` | .15 (amdpve) | llama-server daemon (Vulkan, Strix Halo). Note: `llama-server.service` and `llama-server@.service` are **masked** on .15 to prevent port 8080 collisions. |
+| `/etc/systemd/system/strix-server.service` | .15 (amdpve) | llama-server daemon (Vulkan, Strix Halo) running unsloth/Qwen3.6-35B-A3B-MTP-GGUF. Note: `llama-server.service` and `llama-server@.service` are **masked** on .15 to prevent port 8080 collisions. |

 ## Prometheus & Grafana

@@ -220,14 +243,14 @@ 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 (2026-07-12)**: RTX 3090 at 20.7/24GB (84%), RTX 5070 at 10.0/12.2GB (82%). RTX 3090 context increased to 256K (was incorrectly documented as 128K — verified via /proc/PID/cmdline).
- **RTX 3090 runs `--parallel 1`** (verified 2026-07-12). RTX 5070 and Strix at parallel 2.
- **RTX 3090 config**: `-c 262144 -ctk turbo4 -ctv turbo4 --parallel 1 --flash-attn on --cont-batching`. Service: `/home/llmuser/llama-wrapper.sh`.
- **RTX 5070 config**: `--ctx-size 131072 --cache-type-k q4_0 --cache-type-v q4_0 --batch-size 2048 --ubatch-size 1024 --parallel 2`. Api-key standardized to `not-needed` (was `sk-loc...5678` causing 401 loops). Service: `/home/llmuser/llama-wrapper.sh`.
- **LiteLLM timeout tuning (2026-07-08)**: gemma-4-12b 25→120s, qwen3.6-27B-code 40→90s, syslog-auto (qwen route) 40→90s. Nginx proxy_read_timeout: 600s. Global request_timeout: 300s. Config at `/opt/inference-harness/litellm_config.yaml`.
- **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: `ornith-server.service` on port 8080, 256K context, flash-attn + q8 KV.
+- **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.
+- **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 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`.
+- **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.
 - **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)**: `ornith-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`.
+- **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.
 - **Router sidecar fallback**: `router.py` `check_gpu_health()` now probes GPU `/health` directly when sidecar at :8090 is absent. Sidecar JSON exporters not deployed on any GPU host — router relies on GPU-direct fallback.
 - **Router GPU_MOE_URL bug (fixed 2026-07-01)**: docker-compose had `GPU_MOE_URL=.110:8080` (gemma host) instead of `.15:8080` (amdpve). Corrected.
@@ -237,11 +260,14 @@ If no SSH access, send Zulip DM via abiba-bot with vault update instructions.

 ## GPU Inference Benchmarks (Current)

-| GPU | Model | Gen tok/s | Prompt tok/s | Baseline | Samples |
+| GPU | Model | Gen tok/s | Prompt tok/s | Baseline | Context |
 |-----|-------|-----------|--------------|----------|---------|
-| RTX 3090 (.8) | qwen3.6-27B-code | 75 | 305 | 74 | 6 |
-| RTX 5070 (.110) | gemma-4-12b | 75 | 323 | 75 | 6 |
-| Strix Halo (.15) | ornith-1.0-35b | 70 | 532 | 70 | 6 |
+| RTX 3090 (.8) | qwen3.6-27B-code (MTP) | **63** | — | — | **256K** |
+| RTX 5070 (.110) | gemma-4-12b (IQ4_NL+MTP) | **191** | — | — | **256K** |
+| Strix Halo (.15) | qwen3.6-35B-udq4 | **71** | — | — | **256K** |
+
+Benchmarks from 2026-07-15 verification run. RTX 5070 MTP provides 2.7x speedup over pre-upgrade 70 tok/s.
+All 3 GPUs now at 256K context (2026-07-15).

 Benchmarks run through LiteLLM proxy (192.168.68.116:4001) every 5 minutes.
 Degradation alerts fire at 30% (warning) and 50% (critical) below baseline.
@@ -249,12 +275,55 @@ History stored at `/root/data/toks-history.json` with 7-day rolling window.

 **Note (2026-07-01)**: Strix Halo prompt tok/s jumped 209→532 after Vulkan rebuild (cooperative-matrix fast path now active on GFX1151). Baseline may need re-calibration.

-## Agent Config Implications (2026-07-12)
+## Agent Config Implications (2026-07-15)

-With RTX 3090 at 256K context (verified July 2026):
- Agents using `syslog-auto` (55/30/15 qwen+ornith+gemma): `context_length: 262144` — ornith and qwen both support it
- Agents using `qwen3.6-27B-code` directly: `context_length: 262144` (256K ctx verified)
- Agents using `gemma-4-12b` directly (auxiliary tasks): `context_length: 131072`
- Compression threshold at 0.65: fires at ~170K for 262K context window on Strix Halo
- All Hermes clients MUST set `max_tokens: 4096` — first line of defense before server-side `-n 8192` cap
- Port 8080 is used on all 3 GPU hosts (not 8090 as previously documented)
+### Stable Aliases — CRITICAL
+
+All agent configs MUST use stable role-based aliases, never model-specific names:
+- `compression.model: strix-moe` (NOT `qwen3.6-35B-udq4`)
+- `auxiliary.vision.model: gpu-light` (NOT `gemma-4-12b`)
+- `delegation.model: gpu-dense` (NOT `qwen3.6-27B-code`)
+- `auxiliary.web_extract.model: gpu-light`
+
+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**
+- **Mumuni compression context**: 128K (down from 256K) — ensures compression model doesn't timeout
+- Compression threshold 0.65: fires at ~85K for 128K context window
+- Mumuni compression model alias: `strix-moe` with 300s timeout
+
+### Mumuni Agent Profile
+
+Mumuni (CT114, 192.168.68.123) is the primary business assistant. This profile is the reference for all agent configs:
+
+| Setting | Value | Notes |
+|---------|-------|-------|
+| `model.default` | `syslog-auto` | Weighted pool (55% qwen, 30% strix, 15% gemma) |
+| `model.provider` | `custom:litellm` | LiteLLM on CT116 |
+| `compression.model` | `strix-moe` | Stable alias — survives model swaps |
+| `aux.compression.model` | `strix-moe` | Compression auxiliary model |
+| `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) |
+| `compression.threshold` | 0.65 | Triggers at ~85K |
+| `compression.target_ratio` | 0.3 | Compresses to ~38K |
+| `compression.protect_last_n` | 40 | Preserves last 40 messages |
+| `memory.memory_char_limit` | 800 | Brief memory entries |
+| `personalities` | `creative` | Creative assistant personality |
+| Platforms | cli, discord, homeassistant, signal, telegram, zulip | All Hermes platforms |
+| Main model timeout | 300s | LiteLLM global timeout |
+| Compression model timeout | 300s | ornith timeout increased from 120s |
+
+### Agent Update Status (2026-07-15)
+
+| Agent | Host | Status |
+|-------|------|--------|
+| **Mumuni** | CT114 (.123) | ✅ Updated to stable aliases |
+| **Tanko** | CT112 (.122) | ✅ Updated to stable aliases |
+| **Koby** | CT111 (.129) | ❌ SSH unreachable — needs Zulip DM |
+| **Koonimo** | CT113 | ❌ SSH unreachable — needs Zulip DM |
+| **Kagenz0** | CT105 | ❌ SSH unreachable — needs Zulip DM |
+
+All Hermes clients MUST set `max_tokens: 4096` — first line of defense before server-side `-n 8192` cap.
@@ -120,7 +120,7 @@ depends_on:
 - **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 5070 (131K ctx, gemma-4-12b): target 76+ tok/s — optimal for vision/web role
-  - Strix Halo (256K ctx, ornith-1.0-35b): target 70+ tok/s — currently above baseline
+  - Strix Halo (256K 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
@@ -192,7 +192,7 @@ Key is injected via `infisical run --` wrapper at PM2 startup:
      "apiKey": "${LITELLM_API_KEY}",
      "models": [
        { "id": "syslog-auto" },
-        { "id": "ornith-1.0-35b" },
+        { "id": "strix-moe" },
        { "id": "qwen3.6-27B-code" },
        { "id": "gemma-4-12b" }
      ]
@@ -5,10 +5,12 @@ description: >
  Standard Hermes configuration template for Syslog Solution LLC agents.
  Enforces shared infrastructure setup (Firecrawl, SearXNG, local models,
  RA-H OS MCP) while keeping agent-specific API keys and model choices.
-  UPDATED 2026-07-12: GPU workload redistributed. Compression → Strix Halo (ornith-1.0-35b).
-  RTX 3090 context verified at 256K (was incorrectly documented as 128K).
-  RTX 5070 stays at 131K for vision/web. Infisical .env fallback required (Rule 3).
-  Parallel counts corrected (RTX 3090=1, RTX 5070=2, Strix=2).
+  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).
+  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
+  verified at 256K. Infisical .env fallback required (Rule 3/13).
 ---

 ## Maintains
@@ -31,7 +33,7 @@ Sub-agent profiles inherit auth from the main config — no separate keys needed
 | Agent | Key Alias | Host | SSH | Sub-Agents |
 |-------|-----------|------|-----|-----------|
 | Tanko | `tanko-*` | 192.168.68.122 | jerome@.122 | — |
-| Mumuni | `mumuni-jul2026` | 192.168.68.123 | root@.123 | 6 profiles ✱ |
+| 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 | — |
@@ -94,7 +96,7 @@ work immediately after restart.
 ```yaml
 # ─── Model Selection ───
 model:
-  default: <agent_model>          # e.g., ornith-1.0-35b, qwen3.6-27B-code
+  default: <agent_model>          # e.g., strix-moe, qwen3.6-27B-code, syslog-auto
  provider: harness
  base_url: http://192.168.68.116/v1
  api_key_env: LITELLM_API_KEY    # Injected via infisical run -- wrapper
@@ -129,10 +131,9 @@ mcp_servers:
 # ─── Compression ───
 compression:
  enabled: true
-  model: ornith-1.0-35b           # ⚠️ Must match auxiliary.compression.model
+  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      # For Strix Halo compression (256K ctx).
-                                  # Set 131072 if using gemma directly.
+  max_context_window: 262144      # MUST match actual GPU capacity. All 3 GPUs are 256K (Jul 15).
  threshold: 0.65                 # Fires at ~170K for 262K window, ~85K for 128K
  target_ratio: 0.30
  protect_last_n: 40
@@ -142,37 +143,48 @@ compression:

 # ─── Auxiliary Tasks (CONSISTENCY RULE) ───
 # All auxiliary services MUST use identical model, base_url, and api_key_env:
-#   model: gemma-4-12b
+#   model: gpu-light                 # stable alias (NOT raw "gemma-4-12b")
 #   base_url: http://192.168.68.116/v1
 #   api_key_env: LITELLM_API_KEY
 # Do NOT use syslog-auto for auxiliary tasks — it routes to the primary GPU.
-# gemma-4-12b is a lightweight 12B model on the RTX 5070, freeing the Strix Halo
-# for agent reasoning.
+# gpu-light = RTX 5070 (12B), freeing the Strix Halo for agent reasoning.
+# Heavy aux (delegation, x_search) use gpu-dense (RTX 3090) instead.
+# NEVER use raw model names (gemma-4-12b, qwen3.6-27B-code, qwen3.6-35B-udq4)
+# in agent configs — use the stable aliases so model swaps don't break agents.
 auxiliary:
  vision:
    provider: harness
-    model: gemma-4-12b
+    model: gpu-light            # stable alias for RTX 5070 (was raw gemma-4-12b)
    base_url: http://192.168.68.116/v1
    api_key_env: LITELLM_API_KEY
    timeout: 60
    download_timeout: 30
  web_extract:
    provider: harness
-    model: gemma-4-12b
+    model: gpu-light            # stable alias for RTX 5070
    base_url: http://192.168.68.116/v1
    api_key_env: LITELLM_API_KEY
    timeout: 30
  compression:
    provider: harness
-    model: ornith-1.0-35b
-    base_url: http://192.168.68.116/v1
+    model: strix-moe              # MUST match compression.model above. Stable alias for Strix Halo.
+    base_url: http://192.168.68.116/v1   # Rule 5: /v1 NOT /litellm/v1
    api_key_env: LITELLM_API_KEY
-    timeout: 60
+    timeout: 300                 # gpu-fleet: 300s for large-history summarization (was 60)
+
+# ─── Delegation / Heavy Aux (use gpu-dense = RTX 3090) ───
+# delegation.model and x_search.model use gpu-dense (NOT raw qwen3.6-27B-code).
+
+delegation:
+  model: gpu-dense              # stable alias for RTX 3090 (was raw qwen3.6-27B-code)
+  provider: harness
+  base_url: http://192.168.68.116/v1
+  api_key_env: LITELLM_API_KEY

 # ─── Custom Provider ───
 custom_providers:
  - name: harness
-    model: <agent_model>
+    model: syslog-auto            # weighted pool (default)
    base_url: http://192.168.68.116/v1
    api_key_env: LITELLM_API_KEY
    api_mode: chat_completions
@@ -236,27 +248,29 @@ The following MUST be identical across ALL profiles:
 - Apply to BOTH main config AND all sub-agent profiles
 - For agents needing longer outputs: raise to 8192, but never omit

-### Rule 7: Auxiliary Model Consistency (UPDATED July 2026)
+### Rule 7: Auxiliary Model Consistency (UPDATED 2026-07-16)
 - Vision and web_extract use `gemma-4-12b` (RTX 5070 — 12GB, vision-optimized)
- Compression uses `ornith-1.0-35b` (Strix Halo — 64GB, 256K ctx, compression-optimized)
+- Compression uses `strix-moe` (stable alias for Strix Halo — 64GB, 256K 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:
-  - `base_url: http://192.168.68.116/v1`
+  - `base_url: http://192.168.68.116/v1`  (Rule 5: `/v1`, NOT `/litellm/v1`)
  - `api_key_env: LITELLM_API_KEY`
 - **Do NOT use `syslog-auto`** for auxiliary tasks — it routes unpredictably
- **Compression moved to Strix Halo (July 2026)**: The ornith-1.0-35b model on Strix Halo
-  (64GB UMA, 256K context, 72.4 tok/s) is the designated compression GPU. This frees the
+- **Compression on Strix Halo**: The strix-moe alias routes to Strix Halo
+  (64GB UMA, 256K 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

-### Rule 8: GPU Workload Distribution (July 2026)
+### 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 5070 (12GB, 131K ctx, gemma-4-12b)**: Vision, web search, quick tasks, web_extract
- **Strix Halo (64GB, 256K ctx, ornith-1.0-35b)**: Context compression, summarization, long docs
+- **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)
+- **Strix Halo (64GB, 256K 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: ornith-1.0-35b` (Strix Halo)
+  - `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)
 - Do NOT use `threshold: 0.25` — this fires at 65K, causing premature context loss
@@ -274,7 +288,7 @@ The following MUST be identical across ALL profiles:
 ### Rule 10: Default Model Must Be `syslog-auto` (All Agents)
 - **Hermes agents**: `model.default: syslog-auto`, `custom_providers[0].model: syslog-auto`
 - **pi agents**: `defaultModel: syslog-auto` in `settings.json`, first model in `models.json`
- `syslog-auto` is the LiteLLM routing model — it load-balances between ornith-1.0-35b
+- `syslog-auto` is the LiteLLM routing model — it load-balances between strix-moe
  and qwen3.6-27B-code, with gemma-4-12b as fallback. Using it protects against:
  - Model name typos that cause 403 errors and silent worker failures
  - Single GPU downtime (routing falls back automatically)
@@ -293,6 +307,60 @@ The following MUST be identical across ALL profiles:
 - A non-existent model ID causes 403 errors that silently break the pi RPC worker
  (no `agent_end` emitted, worker stays "busy", Zulip messages pile up unprocessed)

+### Rule 12: Context-Issue Diagnostic Checklist (ADDED 2026-07-16, WAL #1300)
+When an agent shows "context issues" (premature compression, 401s, 504s, DeepSeek fallback),
+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
+   ~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/`
+   has a 60s default timeout → 504 on any inference >60s; `/v1/` has 600s.
+   Check: `grep -n 'litellm/v1' ~/.hermes/config.yaml` (must return NOTHING)
+3. **LITELLM_API_KEY valid?** — The key must be a real LiteLLM key (`sk-` + 64 hex, 67 chars).
+   Malformed values (e.g. `sk-_SWAl_Vu_…`, 47 chars) return 401 → DeepSeek fallback.
+   Verify: `curl -s -o /dev/null -w '%{http_code}' -H "Authorization: Bearer $LITELLM_API_KEY" http://192.168.68.116/v1/models` (must be 200)
+4. **custom_providers aligned?** — `model: syslog-auto` (Rule 10), `api_mode: chat_completions`
+   (NOT `responses`). A wrong api_mode causes silent request failures.
+
+One-line agent health check (run on the agent host):
+```bash
+PID=$(pgrep -f "python -m hermes_cli.main gateway run" | head -1)
+cat /proc/$PID/environ | tr '\0' '\n' | grep ^LITELLM_API_KEY= | sed 's/=.*/<set>/'
+curl -s -o /dev/null -w 'key_health: %{http_code}\n' -H "Authorization: Bearer $(cat /proc/$PID/environ | tr '\0' '\n' | grep ^LITELLM_API_KEY= | cut -d= -f2)" http://192.168.68.116/v1/models
+```
+
+### Rule 13: API Key Injection — Two Patterns (UPDATED 2026-07-16, WAL #1300)
+
+Agents inject `LITELLM_API_KEY` via ONE of two mechanisms. Both are valid; the contract
+requirement is that the key is a **valid LiteLLM virtual key** (HTTP 200 on /v1/models).
+
+**Pattern A — systemd drop-in (Koby, Koonimo, and any agent without infisical wrapper):**
+A systemd drop-in `/etc/systemd/system/hermes-gateway.service.d/litellm-key.conf` sets the key:
+```ini
+[Service]
+Environment="LITELLM_API_KEY=sk-<VALID_KEY>"
+```
+The service unit `hermes-gateway.service` runs `python -m hermes_cli.main gateway run --replace`
+directly (no infisical). Apply with `systemctl daemon-reload && systemctl restart hermes-gateway`.
+- Koonimo (CT113/.114): service = `hermes-gateway.service`, drop-in has the key.
+- Koby (CT111/.129): service = `hermes-gateway.service` (created 2026-07-16), ExecStart uses `--replace`
+  to win the lock against stray `hermes gateway restart` invocations. Key also in `/etc/environment`.
+
+**Pattern B — infisical-gateway.sh wrapper (Mumuni):**
+The wrapper sources `~/.hermes/.env` then exports `LITELLM_API_KEY="$<AGENT>_LITELLM_API_KEY"`.
+See litellm-api-keys.prose.md § Machine Identity for Vault Writes for vault sync.
+
+**Verification (all agents):**
+```bash
+GP=$(pgrep -f "python -m hermes_cli.main gateway run" | head -1)
+K=$(cat /proc/$GP/environ | tr '\0' '\n' | grep '^LITELLM_API_KEY=' | cut -d= -f2)
+curl -s -o /dev/null -w '%{http_code}' -H "Authorization: Bearer $K" http://192.168.68.116/v1/models  # must be 200
+```
+- `/etc/environment` is NO LONGER the canonical key source (stale values there caused 401s).
+- Do NOT leave a hardcoded stale key in `/etc/environment` — it shadows the drop-in/wrapper.
+
 ## Execution

 1. **Check current config** — Read the target agent's config.yaml
@@ -72,6 +72,11 @@ External providers are **explicitly exempt** and may use hardcoded keys:

 ## Standard Pattern

+> **Canonical vault process (2026-07-16):** see `litellm-api-keys` § Production Vault Access Process.
+> All agents MUST use the `infisical-gateway.sh` wrapper (live vault injection). Hardcoded systemd
+> drop-ins / config.yaml keys are DEPRECATED — they rot on rotation (root cause of the 2026-07-16 401 storm).
+> 4/5 agents migrated; tanko (user jerome) pending.
+
 ```yaml
 # ✅ CORRECT — all harness/litellm providers (authenticated path, NO /responses suffix)
 model:
@@ -14,9 +14,15 @@ description: >
  expire/404. The .env fallback prevents agents from running without keys.
  Tanko incident: token 404 → gateway had no LITELLM_API_KEY for hours.
  
+  UPDATED 2026-07-16: Vault is SYNCED (session-13 keys written to vault via abiba service
+  token, all validate 200). Koby/Koonimo migrated from hardcoded drop-ins to the
+  infisical-gateway.sh wrapper (live vault injection). 4/5 agents now vault-backed.
+  Canonical process: see § Production Vault Access Process. Tanko (user jerome) pending.
+  Abiba's key is now a proper agent key (NOT the master key — stale note removed).
+  
  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: 2026-07-12.
+  on CT 116. Last verified: 2026-07-16.
 ---

 ## Parameters
@@ -51,8 +57,8 @@ 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", "qwen3.6-27B-code", "gemma-4-12b", "ornith-1.0-35b"]
-   - Note: qwen3.6-35B-A3B removed from fleet (was never deployed on any GPU)
+   - Set models: ["syslog-auto", "qwen3.6-27B-code", "gemma-4-12b", "strix-moe", "gpu-dense", "gpu-light", "qwen3.6-35B-udq4"]
+   - Note: `ornith-1.0-35b` is NOT a valid LiteLLM model name (use `strix-moe`, the stable alias). qwen3.6-35B-A3B removed from fleet (was never deployed).
   - Return the new key
 5. **If action == "rotate"**:
   - Generate new key with same alias (LiteLLM replaces the old key)
@@ -67,3 +73,145 @@ description: >
   - Test the key against LiteLLM /v1/models
   - Confirm key alias matches agent_name in LiteLLM key list
   - Verify agent gateway uses vault wrapper: `cat /proc/<pid>/cmdline` shows `infisical run`
+
+## Production Vault Access Process (canonical, 2026-07-16)
+
+The non-fail approach to agentic vault access. Deployed on 4/5 agents (tanko pending —
+runs as user `jerome`, not systemd root, needs user-scope adaptation).
+
+### The canonical pattern
+
+1. **infisical CLI** installed on the host (`/usr/local/bin/infisical` or `/usr/bin/infisical`).
+2. **Service token** (Infisical Machine Identity, `st.…`) stored root-only at `/root/.infisical-token` (`chmod 600`).
+   - Interim: the shared `abiba` service token (`st.8e848433…`) has READ+WRITE on the `agents` project.
+   - Proper: one machine identity per agent (create in Infisical UI → Project Settings → Machine Identities).
+3. **`infisical-gateway.sh` wrapper** at `/root/.hermes/infisical-gateway.sh` (`chmod 700`):
+   ```bash
+   #!/bin/bash
+   export INFISICAL_API_URL="https://vault.sysloggh.net"
+   TOKEN=$(cat /root/.infisical-token)
+   LOG=/root/.hermes/logs/gateway.log; mkdir -p /root/.hermes/logs
+   while true; do
+     infisical run --token="$TOKEN" --projectId=322fceab-39da-4854-a55a-568e76c0f13f \
+       --env=prod --domain=https://vault.sysloggh.net -- bash -c '
+         . /root/.hermes/.env 2>/dev/null   # [FALLBACK Rule 3] safety net only
+         export LITELLM_API_KEY="$<AGENT>_LITELLM_API_KEY"
+         exec <HERMES_VENV>/bin/python -m hermes_cli.main gateway run
+       ' >> $LOG 2>&1
+     sleep 5   # restart on exit
+   done
+   ```
+4. **Agent key in vault** as `<AGENT>_LITELLM_API_KEY` (e.g. `KOBY_LITELLM_API_KEY`). Vault = source of truth.
+5. **`.env` fallback** at `/root/.hermes/.env` (`chmod 600`) with the same key — safety net ONLY for vault outage (Rule 3/13). Must be kept in sync on rotation.
+6. **systemd service** `hermes-gateway.service` with `ExecStart=/root/.hermes/infisical-gateway.sh`. NO `litellm-key.conf` drop-in (those hardcode keys and rot).
+7. **NEVER hardcode** LiteLLM keys in systemd drop-ins, config.yaml, or /etc/environment. The wrapper injects live from vault.
+
+### Why this is non-fail
+
+- **No rot**: keys pulled live from vault at every gateway start. Rotation = one `infisical secrets set` + `systemctl restart`. No per-host file edits.
+- **Survives vault outage**: the `.env` fallback (Rule 3) keeps the gateway running if Infisical is unreachable.
+- **Survives gateway crash**: the wrapper's `while true` + systemd `Restart=on-failure` revive the gateway.
+- **Auditable**: `cat /proc/$(pgrep hermes_cli)/environ` shows the live key; `infisical secrets` shows the vault source.
+
+### Migration status (2026-07-16)
+
+| Agent | Host | Pattern | Vault key | Status |
+|-------|------|---------|-----------|--------|
+| abiba | .24 | `infisical run` (pi agent wrapper, service token) | ABIBA_LITELLM_API_KEY | ✅ vault-backed |
+| mumuni | .123 | infisical-gateway.sh + user-login machine identity | MUMUNI_LITELLM_API_KEY | ✅ vault-backed |
+| koby | .129 | infisical-gateway.sh + service token (migrated 2026-07-16) | KOBY_LITELLM_API_KEY | ✅ vault-backed, Zulip (tanko-bot@) + Telegram |
+| koonimo | .114 | infisical-gateway.sh + service token (migrated 2026-07-16) | KOONIMO_LITELLM_API_KEY | ✅ vault-backed |
+> **Baggy = Koonimo (CT 113).** Deleted `BAGGY_LITELLM_API_KEY` from vault 2026-07-16. Only `KOONIMO_LITELLM_API_KEY` exists — one secret per agent.
+| tanko | .122 | **hardcoded in config.yaml** (runs as user jerome, not systemd) | TANKO_LITELLM_API_KEY | ⚠️ TODO: migrate to user-scope wrapper |
+
+### Tanko migration (pending)
+
+Tanko runs the gateway as user `jerome` (not root/systemd), with the key hardcoded in
+`/home/jerome/.hermes/config.yaml` (`api_key: sk-CggiHWlamQy…`, valid but not vault-sourced).
+Migration: create a user-scope systemd service (`~/.config/systemd/user/hermes-gateway.service`)
+with `infisical-gateway.sh` wrapper in jerome's home, token at `~/.infisical-token`, lingering
+enabled (`loginctl enable-linger jerome`) so the user service runs without a login session.
+
+### Koby migration lessons (2026-07-16)
+
+Migrated Koby from hardcoded systemd drop-in → `infisical-gateway.sh` wrapper.
+**Two mistakes I made that broke the agent:**
+1. **Overwrote `/root/.hermes/.env`** without backing it up. The Zulip API key only existed
+   in the running process memory — the old .env was minimal (just LiteLLM key). Zulip creds were
+   inherited from the pre-migration gateway env, not stored in any file. Lost on restart.
+2. **Only injected `LITELLM_API_KEY`** in the wrapper — forgot Zulip + Telegram credentials.
+   Agents need ALL their platform env vars. Missing vars cause silent adapter failures.
+
+**How Koby actually connects (2026-07-16):**
+- Zulip: shares **Tanko's bot** (`tanko-bot@chat.sysloggh.net`, `TANKO_ZULIP_API_KEY=5PeD6f3zo…`).
+  Koby doesn't have its own Zulip bot (koby-bot@ doesn't exist in the swarm config).
+- Telegram: token `828640…` recovered from `.env.bak-20260603` (18KB backup from June 2026).
+  Allowed users: 6679773481. Home channel: 6679773481.
+- Both platforms now connect through the wrapper's env injection.
+
+**Golden rule for gateway restarts:** always `cat /proc/<pid>/environ` before killing the old
+process — captures the live env set. Especially important when migrating gateways between
+injection mechanisms.
+
+### Key rotation procedure (one vault operation with this standard)
+
+1. Generate new key: `POST /key/generate` (master key, admin).
+2. Update vault: `infisical secrets set <AGENT>_LITELLM_API_KEY=sk-NEW --token=$TOKEN --projectId=322fceab… --env=prod --domain=https://vault.sysloggh.net`.
+3. Update `.env` fallback: `echo '<AGENT>_LITELLM_API_KEY=sk-NEW' > /root/.hermes/.env && chmod 600 /root/.hermes/.env`.
+4. Restart: `systemctl restart hermes-gateway`. The wrapper pulls the new key live.
+5. Verify: `curl -H "Authorization: Bearer sk-NEW" http://192.168.68.116/v1/models` → 200.
+
+## Machine Identity for Vault Writes (ADDED 2026-07-16, WAL #1300)
+
+**Problem:** The infisical CLI on agent hosts is logged in as a user session (jerome@sysloggh.com).
+In CLI v0.38.0, `infisical secrets set` / `infisical export` fail with "project id missing" / "workspace
+key 404" — a known bug where user-session auth works for `run` but NOT for `secrets set`. The apt
+repo only ships 0.38.0, so `apt upgrade` does not help.
+
+**Proper fix — Machine Identity (Infisical automation best practice):**
+Create a machine identity with READ+WRITE scope on the `agents` project (project_id=
+`322fceab-39da-4854-a55a-568e76c0f13f`, env `prod`). Store client_id + client_secret securely.
+Then vault writes work from any host:
+```bash
+# Get a machine-identity access token
+TOKEN=$(curl -fsSL -X POST https://vault.sysloggh.net/api/v1/auth/universal-auth/login \
+  -H 'Content-Type: application/json' \
+  -d '{"clientId":"<CLIENT_ID>","clientSecret":"<CLIENT_SECRET>"}' | jq -r .accessToken)
+# Write a secret via REST API v3
+ curl -fsSL -X PATCH https://vault.sysloggh.net/api/v3/secrets/MUMUNI_LITELLM_API_KEY \
+  -H "Authorization: Bearer $TOKEN" -H 'Content-Type: application/json' \
+  -d '{"environment":"prod","secretValue":"sk-<NEW_KEY>","workspaceId":"<WORKSPACE_ID>","type":"shared"}'
+# OR via CLI: infisical secrets set --token=$TOKEN --projectId=322fceab... --env=prod ...
+```
+Creation requires the Infisical web UI (https://vault.sysloggh.net) under Project Settings →
+Machine Identities, or an admin API call. **TODO: create `abiba-automation` machine identity
+and store its credentials in the vault itself (or a root-only file).**
+
+**Interim (working now):** the `.env` fallback (hermes-config-template Rule 3/13). The
+infisical-gateway.sh wrapper sources `~/.hermes/.env`, so its `<AGENT>_LITELLM_API_KEY`
+overrides a stale vault value.
+
+**2026-07-16 UPDATE — vault is now SYNCED.** The abiba service token (`st.8e848433…`, READ+WRITE)
+can write to the vault, so the session-13 rotated keys (mumuni `sk-OzuWsoX2…`, koby `sk-BqRRMboTI…`,
+koonimo `sk-OEK7z26n6E…`) are now in the vault as `MUMUNI_LITELLM_API_KEY` / `KOBY_LITELLM_API_KEY` /
+`KOONIMO_LITELLM_API_KEY` and validate 200 against LiteLLM. The vault is the source of truth again.
+Creating a dedicated `abiba-automation` machine identity (via UI) is still the proper long-term fix
+so the shared service token isn't reused across hosts — but it is no longer blocking.
+
+## Key Rotation Log
+
+| Date | Agent | Action | Notes |
+|------|-------|--------|-------|
+| 2026-07-16 | mumuni | rotate | Old key malformed (sk-_SWAl_Vu_, 47 chars, not LiteLLM format) → 401. Deleted old `mumuni` key (token 15cbca18…), generated fresh (alias `mumuni`, 7 models: syslog-auto, qwen3.6-27B-code, gemma-4-12b, strix-moe, gpu-dense, gpu-light, qwen3.6-35B-udq4). New key sk-OzuWsoX2… written to /root/.hermes/.env (Rule 3/13 fallback). Vault sync PENDING (needs machine identity). WAL #1300. |
+| 2026-07-16 | koby | rotate | Old key sk-6sbCNjz (401, stale in /etc/environment). Deleted old `koby` key, generated fresh (alias `koby`). New key sk-BqRRMboTI… in systemd drop-in `hermes-gateway.service.d/litellm-key.conf` + /etc/environment. Created `hermes-gateway.service` unit (was missing — gateway wasn't persistent) with `--replace`. Verified HTTP 200, Telegram connected. |
+| 2026-07-16 | baggy (koonimo) | rotate | Old key sk-krnw_zGB (401, hardcoded in systemd drop-in). Deleted old `baggy` key, generated fresh (alias `baggy`, metadata agent=koonimo). New key sk-OEK7z26n6E… in drop-in `hermes-gateway.service.d/litellm-key.conf`. CT113 IP changed .113→.114. Verified HTTP 200, Zulip connected. |
+
+## LiteLLM Master Key (use sparingly — agents should NOT use it directly)
+
+- Master key: `sk-litellm-7f96080dd99b15c36bd4b333b58a6796` (in /opt/inference-harness/.env on CT116, Infisical project=infrastructure env=production secret=LITELLM_MASTER_KEY)
+- Used for /key/generate, /key/delete, /key/list (GET), DB queries
+- **Known violation (RESOLVED 2026-07-16):** Abiba's LITELLM_API_KEY was previously the master key.
+It is now a dedicated agent key `sk-sxbphLvk1OU…` (vault secret `ABIBA_LITELLM_API_KEY`, alias `abiba-pi`).
+The master key is admin-only (/key/generate, /key/delete, /key/list). NEVER use it for inference —
+see `litellm-self-heal` § "NEVER use litellm_proxy_master_key for inference".
+- LiteLLM key DB: `harness-postgres` container on CT116, table `"LiteLLM_VerificationToken"` (columns: token, key_alias, key_name, created_at, expires). Query: `docker exec harness-postgres psql -U litellm -d litellm -t -c "SELECT key_alias, substr(token,1,16) FROM \"LiteLLM_VerificationToken\" ORDER BY created_at;"`
@@ -42,8 +42,8 @@ Request → nginx:80 → LiteLLM:4000 → GPU(llama-server, parallel 2)
 **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
+- NVIDIA context reduced 256K→128K to free VRAM  (SUPERSEDED 2026-07-16: all GPUs back to 256K — see litellm-self-heal)
+- LiteLLM timeouts tuned: gemma 25→120s, qwen 40→90s  (SUPERSEDED 2026-07-16: qwen 300s, gemma 120s, strix 300s — see litellm-self-heal)
 - nginx proxy_read_timeout: 600s, LiteLLM request_timeout: 300s

 ## Parameters
@@ -72,18 +72,18 @@ Request → nginx:80 → LiteLLM:4000 → GPU(llama-server, parallel 2)

 | 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 |
+| llm-gpu | 192.168.68.8 | NVIDIA RTX 3090 (24 GB) | qwen3.6-27B-code | llama-server systemd | **256K** | 2 |
+| ocu-llm | 192.168.68.110 | NVIDIA RTX 5070 (12 GB) | gemma-4-12b | llama-server systemd | **256K** | 2 |
+| amdpve | 192.168.68.15 | AMD Strix Halo 64GB UMA | qwen3.6-35B-udq4 (LiteLLM alias: strix-moe) | llama-server systemd (Vulkan) | 256K | 2 |

 ## Model Fallback Chains (LiteLLM)

 | Primary | Timeout | Fallback | Timeout |
 |---------|---------|----------|---------|
-| 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) | 90s | qwen → gemma | — |
+| qwen3.6-27B-code | 300s | gemma-4-12b | 120s |
+| gemma-4-12b | 120s | qwen3.6-27B-code | 300s |
+| qwen3.6-35B-udq4 / strix-moe | 300s | qwen → gemma | — |
+| syslog-auto (balanced) | 300s | qwen → gemma | — |

 > Global: request_timeout=300s, nginx proxy_read_timeout=600s

@@ -129,7 +129,7 @@ Request → nginx:80 → LiteLLM:4000 → GPU(llama-server, parallel 2)
 7. **Check model inference via LiteLLM** — Test each model:
   - POST /v1/chat/completions model=gemma-4-12b → expect 200
   - POST /v1/chat/completions model=qwen3.6-27B-code → expect 200
-   - POST /v1/chat/completions model=ornith-1.0-35b → expect 200
+   - POST /v1/chat/completions model=strix-moe → expect 200
   - Use master key for auth

 8. **Check agent keys**:
@@ -6,6 +6,7 @@ note: >
  DEPLOYED 2026-07-12 on CT 116 cron: 0 */6 * * *
  Auto-remediation code was removed from the pi Zulip extension (retired 2026-07-04),
  now reimplemented as `litellm-health-check.sh` on CT 116.
+  Script: `/opt/inference-harness/scripts/litellm-health-check.sh` on CT 116 (cron `0 */6 * * *`).
  Reports to /var/log/litellm/health-*.json and RA-H OS knowledge graph.
  GPU monitoring integrated from gpu-monitor on .24:9100.
  
@@ -56,18 +57,29 @@ Request → nginx:80 → LiteLLM:4000 → GPU(llama-server, parallel 2)

 | 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 | **256K** | 1 |
-| ocu-llm | 192.168.68.110 | NVIDIA RTX 5070 (12 GB) | gemma-4-12b | llama-server systemd | 131K | 2 |
-| amdpve | 192.168.68.15 | AMD Strix Halo 64GB UMA | ornith-1.0-35b | llama-server systemd (Vulkan) | 256K | 2 |
+| llm-gpu | 192.168.68.8 | NVIDIA RTX 3090 (24 GB) | qwen3.6-27B-code | llama-server systemd (`/home/llmuser/llama-wrapper.sh`, `-c 262144 --parallel 2 --ngl 99`) | **256K** | 2 |
+| ocu-llm | 192.168.68.110 | NVIDIA RTX 5070 (12 GB) | gemma-4-12b | llama-server systemd (`/home/llmuser/llama-wrapper.sh`, `--ctx-size 262144 --parallel 2`, IQ4_NL + MTP draft) | **256K** | 2 |
+| amdpve | 192.168.68.15 | AMD Strix Halo 64GB UMA | qwen3.6-35B-udq4 (LiteLLM alias: `strix-moe`) | llama-server systemd (Vulkan) | 256K | 2 |
+
+> Verified on ground 2026-07-16 via `curl /v1/models` on each host + `llama-wrapper.sh`. The AMD host's underlying model is `qwen3.6-35B-udq4`; LiteLLM exposes it under two `model_name`s: `qwen3.6-35B-udq4` and `strix-moe` (rpm 40). The legacy name `ornith-1.0-35b` does NOT exist in LiteLLM and must not be referenced.
+
+## LiteLLM Model Surface (ground truth — `/opt/inference-harness/litellm_config.yaml` on CT 116)
+
+`model_name`s served: `qwen3.6-27B-code`, `gemma-4-12b`, `qwen3.6-35B-udq4`, `strix-moe`, `gpu-dense`, `gpu-light`, `syslog-auto`.
+
+- `syslog-auto` is a weighted router model: qwen3.6-27B-code (0.55, rpm 500) + qwen3.6-35B-udq4 (0.30, rpm 60) + gemma-4-12b (0.15, rpm 200).
+- `gpu-dense` / `gpu-light` are high-rpm aliases (rpm 500) onto qwen3.6-27B-code / gemma-4-12b respectively.
+- Key scoping: agent keys are restricted to `['syslog-auto','qwen3.6-27B-code','gemma-4-12b','strix-moe','gpu-dense','gpu-light']`. As of 2026-07-16 the `baggy`/`koby`/`mumuni`/`abiba-pi` keys ALSO include `qwen3.6-35B-udq4`; `abiba-pi` additionally includes `deepseek-v4-pro` (cloud fallback). `kagenz0`/`koonimo`/`pi-agents-unified` have the standard 6 only. Agents should still use the stable alias `strix-moe` (not the raw `qwen3.6-35B-udq4`) so model swaps don't break them.
+- **NEVER use `litellm_proxy_master_key` (the `sk-litellm-...` master key) for inference.** It is for admin endpoints only (`/key/list`, `/key/generate`, `/key/info`). All inference — agent traffic, health-check model tests, monitor scripts — uses agent-specific keys. The health-check script's model tests use a dedicated `monitor` agent key stored at `/etc/litellm-monitor.env` on CT 116 (root-only, `chmod 600`); `/key/list` is the only call that legitimately uses `$MASTER_KEY`.

 ## Model Fallback Chains (LiteLLM)

 | Primary | Timeout | Fallback | Timeout |
 |---------|---------|----------|---------|
-| 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) | 90s | qwen → gemma | — |
+| qwen3.6-27B-code | 300s | gemma-4-12b | 120s |
+| gemma-4-12b | 120s | qwen3.6-27B-code | 300s |
+| qwen3.6-35B-udq4 / strix-moe | 300s | qwen → gemma | — |
+| syslog-auto (balanced) | 300s | qwen → gemma | — |

 > Global: request_timeout=300s, nginx proxy_read_timeout=600s

@@ -86,6 +98,13 @@ Request → nginx:80 → LiteLLM:4000 → GPU(llama-server, parallel 2)
 | harness-docker-stats | python:3.12-alpine | — | container stats exporter |
 | harness-pve-exporter | prompve/prometheus-pve-exporter | — | Proxmox metrics → Prometheus |

+## Script Operations (synced 2026-07-16)
+
+- **Health-check script** (`/opt/inference-harness/scripts/litellm-health-check.sh` on CT 116): `gpu-fleet` check fails only on **critical** alerts (warnings are informational). Tests `strix-moe` (not `ornith-1.0-35b`).
+- **GPU monitor** (`/root/scripts/gpu-monitor-server.py` on pi .24): runs as **systemd unit `gpu-monitor.service`** (was bare `&` process). `gpu_count` includes Strix Halo (was 2, now 3). VRAM alert thresholds: warning 93%, critical 97% (raised from 90/95 — 256K context steady-state is ~96% on RTX 3090, not a fault).
+- **Agent key monitor** (`/root/scripts/agent-health-check.py` on pi .24, cron `*/10`): reads each agent's **live** `LITELLM_API_KEY` from its gateway process env via SSH — never hardcodes keys (hardcoded keys rot on rotation and caused 9×401/30min). Fleet roster: abiba, tanko, mumuni, koby, koonimo (legacy `tdunna`/`baggy` removed — never existed).
+- **Stale keys cleaned**: `daily-infra-report.py` SYNTHETIC_API_KEY was stale (`sk-U_ydi3B` → 401); now reads `LITELLM_MASTER_KEY` from env. Deprecated scripts (`router-original.py`, `router-phase0-backup.py`, `apply-fixes.py`) still reference `sk-syslog-local-master-key` but do not actively poll LiteLLM.
+
 ## Maintains

 - litellm-admin-ui: { status: "healthy", last_check: timestamp }
@@ -128,7 +147,7 @@ Run this first on every cycle. Results feed into remediation rules below.
 ### 5. Check model inference via LiteLLM — test each model
 - POST /v1/chat/completions model=gemma-4-12b → expect 200
 - POST /v1/chat/completions model=qwen3.6-27B-code → expect 200
- POST /v1/chat/completions model=ornith-1.0-35b → expect 200
+- POST /v1/chat/completions model=strix-moe → expect 200
 - Use master key for auth

 ### 6. Check agent keys
@@ -92,10 +92,10 @@ raw data never provided.
 |--------|-------|----------|------|----------|
 | `syslog-code` | qwen3.6-27B-code | terminal, file, web, memory, skills | Code patches, automation, scripts | Writing/modifying code, creating scripts, debugging, reading/writing files |
 | `syslog-devops` | qwen3.6-27B-code | terminal, file, web, memory, skills | Infrastructure, DB, bridge, Proxmox | Server ops, SSH, Docker, Proxmox, DB queries, hardware checks |
-| `syslog-email` | ornith-1.0-35b | terminal, file, web, memory, skills | Email automation, mail operations | Sending/receiving email, inbox management, SMTP operations |
-| `syslog-research` | ornith-1.0-35b | terminal, file, web, memory, skills, **browser** | Analysis, classification, data processing | Web research, browser tasks, data analysis, classification, reading docs |
-| `syslog-review` | ornith-1.0-35b | terminal, file, web, memory, skills | Verification, QA, audit validation | **ALWAYS** verify worker output before delivery — especially for infra changes, code builds, and research findings |
-| `syslog-writer` | ornith-1.0-35b | terminal, file, web, memory, skills | Docs, content, branding, reports | Writing docs, reports, proposals, content, markdown formatting |
+| `syslog-email` | strix-moe | terminal, file, web, memory, skills | Email automation, mail operations | Sending/receiving email, inbox management, SMTP operations |
+| `syslog-research` | strix-moe | terminal, file, web, memory, skills, **browser** | Analysis, classification, data processing | Web research, browser tasks, data analysis, classification, reading docs |
+| `syslog-review` | strix-moe | terminal, file, web, memory, skills | Verification, QA, audit validation | **ALWAYS** verify worker output before delivery — especially for infra changes, code builds, and research findings |
+| `syslog-writer` | strix-moe | terminal, file, web, memory, skills | Docs, content, branding, reports | Writing docs, reports, proposals, content, markdown formatting |

 ### Selection Rules

@@ -79,7 +79,7 @@ description: >
  `messages_processed` stalls.
 - **Root cause**: Agent's model config (`models.json` or `settings.json`) references a
  model ID that doesn't exist in LiteLLM's authorized model list. Example: `qwen3.6-35B-A3B`
-  configured but LiteLLM only exposes `ornith-1.0-35b` under that key. pi's session
+  configured but LiteLLM only exposes `strix-moe` (alias for qwen3.6-35B-udq4) under that key. pi's session
  workers emit 403 on first prompt, then never recover because the error doesn't trigger
  `agent_end` — worker stays `busy` and all subsequent messages pile up in the steer queue.
 - **Detection**: Compare `~/.pi/agent/models.json` model IDs against `curl -H "Authorization: Bearer <KEY>" http://192.168.68.116/v1/models` output. A stuck worker shows
@@ -88,7 +88,7 @@ description: >
  (2) Set `defaultModel` to `syslog-auto` (safe routing model). (3) Delete stale session
  JSONL files from `~/.pi/agent/sessions/zulip/`. (4) Restart PM2 process.
 - **Prevention**: Use `syslog-auto` as default model for all agents — it handles model
-  routing and fallback automatically. Direct model IDs (`ornith-1.0-35b`, etc.) should
+  routing and fallback automatically. Direct model IDs (`strix-moe`, etc.) should
  only be used when explicitly requested. Validate model IDs at agent setup time.
 - **Applies to**: pi extension (Tdunna CT111, fixed 2026-07-08), any agent using `syslog-harness` provider

@@ -0,0 +1,411 @@
+---
+kind: responsibility
+name: zulip-resilience-v3
+description: >
+  Rewrite the pi Zulip gateway with production-grade resilience patterns drawn from
+  Zulip's own event system docs (queue lifecycle, heartbeat monitoring, BAD_EVENT_QUEUE_ID
+  handling, idle_queue_timeout) and battle-tested Node.js resilience patterns
+  (circuit breaker, exponential backoff with jitter, bulkhead isolation, supervisor watchdog).
+replaces: zulip-self-heal (retired)
+agent: abiba
+triggers:
+  - "/zulip self-heal v3"
+  - "zulip stopped responding"
+  - "PM2 abiba-zulip crashed"
+---
+
+# Zulip Gateway v3 — Production Resilience
+
+## Architecture Overview
+
+The current v2 gateway (`/root/.pi/agent/extensions/zulip/index.js`) has three structural
+weaknesses that cause repeated deaths:
+
+1. **No crash recovery** — uncaught errors kill the Node process, PM2 exhausts max_restarts
+2. **No circuit breaker** — 502/fetch-failed errors escalate to process death with no fallback
+3. **No queue lifecycle management** — doesn't use Zulip's documented heartbeat protocol or
+   idle_queue_timeout, so BAD_EVENT_QUEUE_ID errors cascade into crashes
+
+The v3 rewrite addresses all three, following patterns from:
+
+- [Zulip Events System docs](https://zulip.readthedocs.io/en/11.6/subsystems/events-system.html) —
+  queue registration, heartbeat, BAD_EVENT_QUEUE_ID recovery, call_on_each_event loop
+- [Zulip API: Get Events](https://zulip.com/api/get-events) — long-poll timeout, dont_block, event ack
+- [Circuit Breaker & Retry Patterns in Node.js 2026](https://1xapi.com/blog/resilient-api-circuit-breaker-bulkhead-retry-nodejs-2026) —
+  Opossum-based circuit breaker with fallback, retry with jitter, bulkhead isolation
+
+---
+
+## Maintains
+
+- `zulip-gateway`: { status: "healthy" | "degraded" | "down" }
+- `circuit-breaker`: { state: "CLOSED" | "OPEN" | "HALF_OPEN", failures, successes }
+- `queue-lifecycle`: { queue_id, last_event_id, idle_timeout, heartbeat_age }
+- `workers`: { count, busy, idle, stuck }
+- `supervisor`: { pid, last_check, health_failures }
+
+---
+
+## Detection Rules
+
+### Rule 1: Queue Expired (BAD_EVENT_QUEUE_ID)
+- **Detect**: Events API returns error with BAD_EVENT_QUEUE_ID in body
+- **Fix**: Call `POST /register` to create new queue, update queue_id and last_event_id
+- **Debounce**: If 3 re-registrations fail within 60s, escalate (server may be down)
+- **Ref**: Zulip docs: "Your software will need to handle that error condition by re-initializing itself"
+
+### Rule 2: Network Degradation (502/ECONNREFUSED/fetch failed)
+- **Detect**: Events API returns 502 or network error
+- **Circuit breaker**: Track failure rate over 10s rolling window
+  - CLOSED → OPEN: 50% failure rate with ≥5 requests
+  - OPEN → HALF_OPEN: After 30s reset timeout
+  - HALF_OPEN → CLOSED: Probe succeeds
+  - HALF_OPEN → OPEN: Probe fails
+- **While OPEN**: Log errors, skip events, notify user via DM: "⚠️ Zulip connection degraded — will retry in 30s"
+
+### Rule 3: Long-Poll Timeout (natural)
+- **Detect**: Events API response takes > `event_queue_longpoll_timeout_seconds`
+- **Not an error**: Server sends heartbeat events when no real events. Simply re-poll.
+
+### Rule 4: Worker Busy Timeout (>5 min)
+- **Detect**: Worker `busySince` exceeds 5 minutes
+- **Fix**: SIGKILL worker, send error DM, clean up pending replies
+
+### Rule 5: Process Crash (uncaught)
+- **Detect**: `uncaughtException` / `unhandledRejection` fires
+- **Fix**: Log → clear poll timer → attempt reconnect with backoff → if reconnect fails 3x, exit(1) and let PM2 restart
+
+### Rule 6: Supervisor Detects Router Stall
+- **Detect**: External supervisor (`zulip-watchdog`) polls `/health` every 30s. If 3 consecutive failures:
+- **Fix**: `pm2 restart abiba-zulip` gracefully (SIGTERM, drain workers, restart)
+
+---
+
+## Implementation Plan
+
+### Phase 1: Rewrite Router Core (circuit-breaker + queue lifecycle)
+
+Replace the poll loop in index.js with a resilience-first event loop:
+
+```js
+// Queue lifecycle (Zulip docs pattern)
+async function createOrRefreshQueue() {
+  // POST /register with event_types=["message"]
+  // Store: queueId, lastEventId, eventQueueLongpollTimeoutSeconds
+  // NEW: pass idle_queue_timeout parameter (Zulip 12.0+)
+}
+
+// Circuit breaker (Opossum pattern, implemented inline to avoid dependency)
+class ZulipCircuitBreaker {
+  constructor({ failureThreshold=0.5, resetTimeout=30000, volumeThreshold=5, windowMs=10000 }) {
+    this.state = "CLOSED"; // CLOSED | OPEN | HALF_OPEN
+    this.failures = 0;
+    this.successes = 0;
+    this.totalRequests = 0;
+    this.lastFailureTime = null;
+    this.openedAt = null;
+    this.failureThreshold = failureThreshold;
+    this.resetTimeout = resetTimeout;
+    this.volumeThreshold = volumeThreshold;
+    this.windowMs = windowMs;
+  }
+
+  async fire(fn) {
+    if (this.state === "OPEN") {
+      if (Date.now() - this.openedAt > this.resetTimeout) {
+        this.state = "HALF_OPEN";
+      } else {
+        throw new CircuitOpenError("Circuit is OPEN");
+      }
+    }
+    try {
+      const result = await fn();
+      this.onSuccess();
+      return result;
+    } catch (err) {
+      this.onFailure();
+      throw err;
+    }
+  }
+
+  onSuccess() {
+    this.successes++;
+    this.totalRequests++;
+    if (this.state === "HALF_OPEN") {
+      this.state = "CLOSED";
+      this.failures = 0;
+    }
+    // Reset counters periodically
+    if (this.totalRequests > this.volumeThreshold * 2) {
+      this.failures = Math.floor(this.failures / 2);
+      this.successes = Math.floor(this.successes / 2);
+      this.totalRequests = Math.floor(this.totalRequests / 2);
+    }
+  }
+
+  onFailure() {
+    this.failures++;
+    this.totalRequests++;
+    this.lastFailureTime = Date.now();
+    if (this.totalRequests >= this.volumeThreshold &&
+        this.failures / this.totalRequests >= this.failureThreshold) {
+      if (this.state !== "OPEN") {
+        this.state = "OPEN";
+        this.openedAt = Date.now();
+        console.error(`[zulip-ext] CIRCUIT BREAKER OPEN — ${this.failures}/${this.totalRequests} failures`);
+      }
+    }
+  }
+}
+
+// Retry with exponential backoff + jitter (from resilience patterns)
+async function withRetry(fn, { maxAttempts=3, baseDelay=200, maxDelay=10000, shouldRetry=()=>true }={}) {
+  let lastError;
+  for (let attempt = 1; attempt <= maxAttempts; attempt++) {
+    try {
+      return await fn();
+    } catch (err) {
+      lastError = err;
+      if (attempt === maxAttempts || !shouldRetry(err)) throw err;
+      const delay = Math.min(baseDelay * Math.pow(2, attempt - 1), maxDelay);
+      const jitter = delay * (0.5 + Math.random() * 0.5); // 50-100% of delay
+      console.warn(`[zulip-ext] Retry ${attempt}/${maxAttempts} after ${Math.round(jitter)}ms: ${err.message.slice(0,80)}`);
+      await new Promise(r => setTimeout(r, jitter));
+    }
+  }
+  throw lastError;
+}
+
+// Resilience-first event loop (Zulip call_on_each_event pattern)
+async function resilientPollLoop() {
+  while (connected) {
+    try {
+      const events = await circuitBreaker.fire(() =>
+        withRetry(() => zulipQueue.poll(), {
+          maxAttempts: 2,
+          baseDelay: 1000,
+          shouldRetry: (err) => {
+            const msg = err.message || "";
+            return msg.includes("fetch failed") || msg.includes("ECONN") || msg.includes("network");
+          }
+        })
+      );
+
+      lastError = null;
+      retryCount = 0;
+
+      for (const ev of events) {
+        await processEvent(ev);
+      }
+
+      heartbeat();
+    } catch (err) {
+      const msg = err instanceof Error ? err.message : String(err);
+
+      if (msg.includes("BAD_EVENT_QUEUE_ID") || msg.includes("deregistered")) {
+        // Queue expired — re-register (Zulip docs pattern)
+        console.log(`[zulip-ext] Queue expired, re-registering… (${msg.slice(0,80)})`);
+        try {
+          zulipQueue = await createZulipQueue();
+          console.log(`[zulip-ext] Re-registered, new queue=${zulipQueue.queueId}`);
+        } catch (reRegErr) {
+          console.error(`[zulip-ext] Re-registration failed: ${reRegErr.message}`);
+          connected = false;
+          retryCount++;
+          const backoff = Math.min(5000 * Math.pow(2, retryCount), 300000);
+          console.log(`[zulip-ext] Full reconnect in ${Math.round(backoff/1000)}s`);
+          await new Promise(r => setTimeout(r, backoff));
+          await startPolling();
+          return;
+        }
+      } else if (err.name === "CircuitOpenError") {
+        // Circuit is open — skip this cycle, wait for HALF_OPEN
+        lastError = "circuit_open";
+        await new Promise(r => setTimeout(r, POLL_INTERVAL_MS));
+      } else {
+        lastError = msg;
+        retryCount++;
+        const backoff = Math.min(POLL_INTERVAL_MS * Math.pow(1.5, Math.min(retryCount, 8)), 60000);
+        console.error(`[zulip-ext] Poll error (retry ${retryCount}, backoff ${backoff}ms): ${msg}`);
+        await new Promise(r => setTimeout(r, backoff));
+      }
+    }
+  }
+}
+```
+
+### Phase 2: PM2 Hardening
+
+Create `/root/.pm2/ecosystem.config.cjs`:
+
+```js
+module.exports = {
+  apps: [
+    {
+      name: "abiba-zulip",
+      script: "/bin/pi",
+      args: "--mode rpc --session-id zulip-service",
+      env: {
+        ZULIP_ROLE: "router",
+        ZULIP_SITE: "https://chat.sysloggh.net",
+        ZULIP_EMAIL: "abiba-bot@chat.sysloggh.net",
+        ZULIP_API_KEY: process.env.ZULIP_API_KEY,
+        AGENT_NAME: "abiba",
+        AGENT_OWNER_EMAIL: "jerome@sysloggh.com",
+      },
+      max_restarts: 100,            // Up from default 10 — crash loops won't exhaust
+      min_uptime: "10s",            // Must survive 10s to count as "alive"
+      max_memory_restart: "500M",   // OOM protection
+      restart_delay: 5000,          // 5s between restarts
+      kill_timeout: 15000,          // 15s SIGTERM grace before SIGKILL
+      listen_timeout: 30000,        // 30s to bind health port
+      log_date_format: "YYYY-MM-DD HH:mm:ss Z",
+      error_file: "/root/.pm2/logs/abiba-zulip-error.log",
+      out_file: "/root/.pm2/logs/abiba-zulip-out.log",
+      merge_logs: true,
+      autorestart: true,
+      watch: false,
+      instances: 1,
+      exec_mode: "fork",
+    },
+    {
+      name: "zulip-watchdog",
+      script: "/root/.pi/agent/extensions/zulip/watchdog.js",
+      max_restarts: 10,
+      min_uptime: "3s",
+      restart_delay: 3000,
+      autorestart: true,
+    },
+  ],
+};
+```
+
+### Phase 3: Supervisor Watchdog
+
+Create `/root/.pi/agent/extensions/zulip/watchdog.js`:
+
+```js
+// External supervisor — monitors router health and restarts if stalled.
+// This is the pattern Hermes uses: an external process that can recover
+// the gateway even if the gateway process itself is hung (not just crashed).
+
+const HEALTH_URL = "http://127.0.0.1:9200/health";
+const CHECK_INTERVAL_MS = 30_000;
+const MAX_FAILURES = 3;
+
+let failures = 0;
+
+async function check() {
+  try {
+    const res = await fetch(HEALTH_URL, { signal: AbortSignal.timeout(5000) });
+    if (res.ok) {
+      const data = await res.json();
+      if (data.status === "ok" && data.zulip?.connected) {
+        if (failures > 0) {
+          console.log(`[watchdog] Router recovered after ${failures} failures`);
+        }
+        failures = 0;
+        return;
+      }
+    }
+    failures++;
+    console.warn(`[watchdog] Health check ${failures}/${MAX_FAILURES}: status not ok`);
+  } catch (err) {
+    failures++;
+    console.warn(`[watchdog] Health check ${failures}/${MAX_FAILURES}: ${err.message}`);
+  }
+
+  if (failures >= MAX_FAILURES) {
+    console.error(`[watchdog] ${MAX_FAILURES} consecutive failures — restarting abiba-zulip`);
+    const { execSync } = require("child_process");
+    try {
+      execSync("pm2 restart abiba-zulip", { timeout: 30000 });
+      console.log("[watchdog] Restart command sent");
+    } catch (e) {
+      console.error(`[watchdog] Restart failed: ${e.message}`);
+    }
+    failures = 0;
+    // Wait for restart to complete before checking again
+    await new Promise(r => setTimeout(r, 15000));
+  }
+}
+
+console.log("[watchdog] Zulip gateway supervisor started");
+setInterval(check, CHECK_INTERVAL_MS);
+check(); // Immediate first check
+```
+
+### Phase 4: Health Endpoint Enhancement
+
+Add circuit breaker stats to the existing health endpoint:
+
+```js
+// In /health response, add:
+"circuit_breaker": {
+  "state": circuitBreaker.state,
+  "failures": circuitBreaker.failures,
+  "successes": circuitBreaker.successes,
+  "total_requests": circuitBreaker.totalRequests,
+  "failure_rate": circuitBreaker.totalRequests > 0
+    ? (circuitBreaker.failures / circuitBreaker.totalRequests).toFixed(2)
+    : "0.00"
+}
+```
+
+---
+
+## Test Plan
+
+### Test 1: Queue Re-registration
+1. Manually delete the Zulip event queue via API
+2. Next poll should detect BAD_EVENT_QUEUE_ID
+3. Router should auto re-register within 1 poll cycle
+4. Verify: `/health` shows new queue_id, connected=true
+
+### Test 2: Circuit Breaker Trip
+1. Block Zulip server with iptables: `iptables -A OUTPUT -d 192.168.68.19 -j DROP`
+2. Router should detect failures, trip circuit after 5 failures
+3. `/health` should show circuit_breaker.state = "OPEN"
+4. Remove iptables rule
+5. Circuit should transition to HALF_OPEN → CLOSED within 60s
+6. Verify: messages processed after recovery
+
+### Test 3: Supervisor Recovery
+1. Kill the router process: `kill -STOP $(pm2 pid abiba-zulip)` (freeze, don't kill)
+2. Watchdog should detect 3 failed health checks in 90s
+3. Watchdog should execute `pm2 restart abiba-zulip`
+4. Verify: router back online, connected=true
+
+### Test 4: Worker Busy Timeout
+1. Send a message that triggers a long-running operation
+2. If worker stays busy >5 minutes, should receive SIGKILL
+3. User should receive error DM: "Response timed out"
+
+### Test 5: End-to-End Message
+1. Send DM "What time is it?" from Jerome
+2. Should receive response within 30s
+3. `/health` should show messages_processed incremented
+
+---
+
+## Rollback Plan
+
+If v3 causes issues:
+1. `pm2 delete abiba-zulip; pm2 delete zulip-watchdog`
+2. Restore v2 from git: `cd /root/.pi/agent/extensions/zulip && git checkout index.js`
+3. `pm2 resurrect` to reload previous process list
+4. Verify: `/health` returns ok
+
+Backup v2 before starting: `cp index.js index.js.v2-backup-$(date +%Y%m%d-%H%M%S)`
+
+---
+
+## Success Metrics
+
+| Metric | Current (v2) | Target (v3) |
+|--------|-------------|-------------|
+| Uptime between manual interventions | 1-3 days | 30+ days |
+| Crash recovery | Manual (PM2 resurrect) | Automatic (circuit breaker + supervisor) |
+| Queue expiry handling | Crash | Auto re-register |
+| Busy worker deadlock | Router death | Worker SIGKILL + error DM |
+| PM2 restart exhaustion | Yes (max_restarts=10) | No (max_restarts=100 + watchdog) |