Merge pull request 'fix(audit): stop requiring the retired gpu-light alias; derive model fields; sweep retired names' (#80) from fix/retired-alias-sweep-20260912 into master

.gitea/workflows/pr-pipeline.yaml

@@ -43,17 +43,21 @@ jobs:
           echo "=== Prose Contract Frontmatter Validation ==="
           FAILED=0
           for f in $(find . -name "*.prose.md" -not -path "./.git/*" -not -path "./runs/*"); do
+            # NOTE: use herestrings, not `echo "$FM" | grep ...`. Under the runner's
+            # `-e -o pipefail`, `grep -q` exits on first match and can SIGPIPE the
+            # producer, making the pipeline report non-zero and raising a false
+            # "Missing name/description" whose file set varies run to run.
             FM=$(sed -n '/^---$/,/^---$/p' "$f" | sed '1d;$d')
             [ -z "$FM" ] && { echo "  ❌ $f: No YAML frontmatter"; FAILED=$((FAILED+1)); continue; }
 
-            KIND=$(echo "$FM" | grep '^kind:' | awk '{print $2}')
+            KIND=$(grep '^kind:' <<< "$FM" | awk '{print $2}')
             case "$KIND" in
               function|responsibility|gateway|pattern|test|template|architecture|enforcement) echo "  ✅ $f: kind=$KIND" ;;
               *) echo "  ❌ $f: Invalid kind='$KIND'"; FAILED=$((FAILED+1)) ;;
             esac
 
-            echo "$FM" | grep -q '^name:' || { echo "  ❌ $f: Missing name"; FAILED=$((FAILED+1)); }
-            echo "$FM" | grep -q '^description:' || { echo "  ❌ $f: Missing description"; FAILED=$((FAILED+1)); }
+            grep -q '^name:' <<< "$FM" || { echo "  ❌ $f: Missing name"; FAILED=$((FAILED+1)); }
+            grep -q '^description:' <<< "$FM" || { echo "  ❌ $f: Missing description"; FAILED=$((FAILED+1)); }
           done
           [ $FAILED -gt 0 ] && { echo "❌ FRONTMATTER FAILED ($FAILED error(s))"; exit 1; }
           echo "✅ Frontmatter validation passed"

audit-hermes-config.py

@@ -36,6 +36,59 @@ def warn(rule, message):
     WARNINGS.append(f"[{rule}] {message}")
 
 
+# Derivation rule: a model name is any scalar under a mapping key named `model` or
+# `model_name`, at any depth. The top-level `model:` SECTION is the exception where the model
+# name lives under `default`/`model`/`model_name` inside that section, so it is descended
+# specially. The only other exception is key `models` (litellm key-generation params carry a
+# list of model names). EXTEND THE ALLOWLIST for a new exception; do NOT add another field by
+# hand.
+MODEL_KEYS = ("model", "model_name")
+MODEL_SECTION_KEYS = ("default", "model", "model_name")
+MODEL_LIST_KEYS = ("models",)
+
+
+def _iter_model_values(node, path=""):
+    """Yield (path, value) for every model-name-bearing scalar in a config."""
+    if isinstance(node, dict):
+        for key, value in node.items():
+            child = f"{path}.{key}" if path else key
+            if key in MODEL_KEYS:
+                if isinstance(value, dict):
+                    for subkey in MODEL_SECTION_KEYS:
+                        subvalue = value.get(subkey)
+                        if isinstance(subvalue, str):
+                            yield (f"{child}.{subkey}", subvalue)
+                    for subkey, subvalue in value.items():
+                        if isinstance(subvalue, (dict, list)):
+                            yield from _iter_model_values(subvalue, f"{child}.{subkey}")
+                elif isinstance(value, list):
+                    yield from _iter_model_values(value, child)
+                else:
+                    yield (child, value)
+            elif key in MODEL_LIST_KEYS:
+                yield from _iter_model_list(value, child)
+            elif isinstance(value, (dict, list)):
+                yield from _iter_model_values(value, child)
+    elif isinstance(node, list):
+        for i, item in enumerate(node):
+            yield from _iter_model_values(item, f"{path}[{i}]")
+
+
+def _iter_model_list(node, path):
+    """Yield scalars under an allowlisted `models` key (list of names or list of dicts)."""
+    if isinstance(node, list):
+        for i, item in enumerate(node):
+            yield from _iter_model_list(item, f"{path}[{i}]")
+    elif isinstance(node, dict):
+        for key, value in node.items():
+            if key in MODEL_KEYS and isinstance(value, str):
+                yield (f"{path}.{key}", value)
+            elif isinstance(value, (dict, list)):
+                yield from _iter_model_list(value, f"{path}.{key}")
+    else:
+        yield (path, node)
+
+
 def audit(path):
     with open(path) as f:
         cfg = yaml.safe_load(f)
@@ -89,15 +142,16 @@ def audit(path):
     )
 
     # --- Rule 8: GPU Workload Distribution ---
+    # gpu-light (and gemma-4-12b) were retired 2026-09-12; the RTX 5070 stable alias is gpu-vision.
     check(
-        aux.get("vision", {}).get("model") == "gpu-light",
+        aux.get("vision", {}).get("model") == "gpu-vision",
         "Rule 8",
-        f"auxiliary.vision.model must be gpu-light (got {aux.get('vision', {}).get('model')!r}) — RTX 5070 stable alias",
+        f"auxiliary.vision.model must be gpu-vision (got {aux.get('vision', {}).get('model')!r}) — RTX 5070 stable alias",
     )
     check(
-        aux.get("web_extract", {}).get("model") == "gpu-light",
+        aux.get("web_extract", {}).get("model") == "gpu-vision",
         "Rule 8",
-        f"auxiliary.web_extract.model must be gpu-light (got {aux.get('web_extract', {}).get('model')!r}) — RTX 5070 stable alias",
+        f"auxiliary.web_extract.model must be gpu-vision (got {aux.get('web_extract', {}).get('model')!r}) — RTX 5070 stable alias",
     )
 
     # --- Rule 9: Compression Threshold ---
@@ -178,21 +232,35 @@ def audit(path):
         f"custom_providers[0].base_url must end with /v1 (got {cp.get('base_url')!r})",
     )
 
-    # --- No raw model names (Rule 7/8 spirit) ---
-    raw_names = {"gemma-4-12b", "qwen3.6-27B-code", "qwen3.6-35B-udq4", "ornith-1.0-35b"}
-    for section_path, section_dict in [
-        ("model", model), ("compression", comp),
-        ("auxiliary.vision", aux.get("vision", {})),
-        ("auxiliary.web_extract", aux.get("web_extract", {})),
-        ("auxiliary.compression", aux.get("compression", {})),
-        ("delegation", deleg),
-    ]:
-        m = section_dict.get("model", "")
-        if m in raw_names:
+    # --- Retired/raw model names (Rule 7/8 spirit) ---
+    # The audit's job is to catch configs that are BROKEN, not to enforce a style preference.
+    # NON-RESOLVING names (removed 2026-09-12, verified 400/403 via live LiteLLM) must hard-FAIL:
+    #   gpu-light -> gpu-vision ; gemma-4-12b -> gpu-vision
+    #   crew-auto -> syslog-auto (its 64K cap is retired; no cap in force) ; ornith-1.0-35b -> strix-moe
+    # RESOLVING names (verified 200) are discouraged but working, so they only WARN:
+    #   qwen3.6-27B-code -> gpu-dense ; qwen3.6-35B-udq4 -> strix-moe
+    # Failing a working alias would reject valid configs - the exact defect this change fixes.
+    non_resolving = {
+        "gpu-light": "gpu-vision",
+        "gemma-4-12b": "gpu-vision",
+        "crew-auto": "syslog-auto (its 64K cap is retired; no cap in force)",
+        "ornith-1.0-35b": "strix-moe",
+    }
+    raw_but_live = {
+        "qwen3.6-27B-code": "gpu-dense",
+        "qwen3.6-35B-udq4": "strix-moe",
+    }
+    for field_path, value in _iter_model_values(cfg):
+        if value in non_resolving:
+            check(
+                False,
+                "Rule 7/8",
+                f"{field_path} = {value!r} is retired and no longer resolves (2026-09-12) — use {non_resolving[value]}",
+            )
+        elif value in raw_but_live:
             warn(
                 "Rule 7/8",
-                f"{section_path}.model = {m!r} — raw model name, use stable alias instead "
-                f"(gpu-light, gpu-dense, strix-moe, syslog-auto)",
+                f"{field_path} = {value!r} is a raw-but-live model name — prefer the stable alias {raw_but_live[value]}",
             )
 
     # --- Report ---

gpu-fleet.prose.md

@@ -216,7 +216,7 @@ If no SSH access, send Zulip DM via abiba-bot with vault update instructions.
 - **Alert migration**: All alerts now go to `#agent-hub` topics (`alerts-gpu`, `alerts-pm2`, `alerts-infra`) instead of DMs. Cross-agent visibility enabled.
 - **tok/s benchmarks**: Measured every 5 min via LiteLLM proxy. Baselines tracked with 30%/50% degradation thresholds.
 - **NetBird 502**: Tanko routes through NetBird for litellm.sysloggh.net. Use direct IP if NetBird down.
-- **Alias-retirement follow-up (2026-09-12)**: the agent-facing templates (`hermes-config-template.prose.md`, `hermes-agent-baseline.prose.md`), `litellm-api-keys.prose.md`, and the executable `audit-hermes-config.py` still reference the retired `gpu-light`/`gemma-4-12b` names, and `audit-hermes-config.py` Rule 8 currently fails a config whose vision/web_extract model is `gpu-vision`. These are tracked separately and are intentionally NOT updated in this change.
+- **Alias-retirement sweep (2026-09-12)**: `gemma-4-12b`, `gpu-light` and `crew-auto` are retired and replaced by `gpu-vision` / no cap respectively. The agent-facing templates (`hermes-config-template.prose.md`, `hermes-agent-baseline.prose.md`), `litellm-api-keys.prose.md`, `gpu-self-heal.prose.md`, `hermes-key-enforcement.prose.md`, `inference-optimization.prose.md`, `litellm-client-timeouts.prose.md` and the executable `audit-hermes-config.py` were all updated to the live canonical alias in the same change. **koby's config on .129 still names `gpu-light` (and `gemma-4-E4B`); .129 is report-only, so that is recorded for its owner and NOT edited here.**
 
 ## GPU Inference Benchmarks (Current)
 
@@ -239,7 +239,7 @@ History stored at `/root/data/toks-history.json` with 7-day rolling window.
 ### Stable Aliases — CRITICAL
 
 All agent configs MUST use stable role-based aliases, never model-specific names:
-- `compression.model: strix-moe`
+- `compression.model: syslog-auto`
 - `auxiliary.vision.model: gpu-vision`
 - `delegation.model: gpu-dense`
 - `auxiliary.web_extract.model: gpu-vision`
@@ -249,25 +249,25 @@ When the underlying model is swapped, only the LiteLLM config changes — agent
 ### Context Windows
 - RTX 3090: **128K** (reduced from 256K 2026-07-17) | RTX 5070: **128K** (reduced from 256K) | Strix Halo: **128K**
 - **All agents**: 128K ceiling — stable margin. For >128K workloads, use external providers (deepseek)
-- Compression threshold 0.60: fires at ~77K (~51K headroom before 128K ceiling)
+- Compression threshold 0.65 (audit Rule 9): fires at ~85K (~43K headroom before 128K ceiling)
 - **Pi agents (Abiba)**: `compaction.reserveTokens: 52739` (≈60% of 128K)
-- Mumuni compression model alias: `strix-moe`
+- Mumuni compression model alias: `syslog-auto`
 
 ### Mumuni Agent Profile
 
-Mumuni (kagentz CT105, 192.168.68.14 — migrated from CT100 2026-08-29) is the primary business assistant. This profile is the reference for all agent configs:
+Mumuni (kagentz CT105, 192.168.68.14 — migrated from CT100 2026-08-29) is the primary business assistant. This profile is the reference for all agent configs. The compression values below are the current required values per template Rules 7/9 and `audit-hermes-config.py`; whether Mumuni's LIVE config currently complies is a separate operational question.
 
 | Setting | Value | Notes |
 |---------|-------|-------|
 | `model.default` | `syslog-auto` | Balanced default (pool router) |
 | `model.provider` | `custom:litellm` | LiteLLM on CT116 |
-| `compression.model` | `strix-moe` | Stable alias — survives model swaps |
-| `aux.compression.model` | `strix-moe` | Compression auxiliary model |
+| `compression.model` | `syslog-auto` | Rule 7: auto-routing, prevents Strix Halo overload |
+| `aux.compression.model` | `syslog-auto` | Must match `compression.model` (Rule 7) |
 | `aux.vision.model` | `gpu-vision` | Vision tasks (RTX 5070) |
 | `aux.web_extract.model` | `gpu-vision` | Web extraction |
 | `delegation.model` | `gpu-dense` | Sub-agent reasoning (RTX 3090) |
 | `context.max_context_window` | 131072 (128K) | Reduced from 256K 2026-07-17 — stable 128K ceiling |
-| `compression.threshold` | 0.60 | Triggers at ~77K (~60% of 128K) — optimized for 128K context |
+| `compression.threshold` | 0.65 | Rule 9: triggers at ~85K for a 128K window |
 | `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 |

gpu-monitor.prose.md

@@ -27,8 +27,8 @@ agent: abiba
 ┌──────┐ ┌──────┐ ┌────────┐
 │.8:8080│ │.110  │ │.116:80 │
 │RTX3090│ │:8080 │ │nginx   │
-│gemma  │ │RTX5070│ │router  │
-└──────┘ │qwen27B│ │LiteLLM │
+│qwen   │ │RTX5070│ │router  │
+└──────┘ │vision │ │LiteLLM │
          └──────┘ │dashboard│
                   └────────┘
 ```

gpu-self-heal.prose.md

@@ -12,7 +12,7 @@ description: >
   Router (port 9000) DECOMMISSIONED 2026-09-11; references replaced with direct GPU routing.
   Benchmark baselines refreshed to live values.
   Prometheus exporters removed — not deployed; fall back to direct sidecar probes.
-  Stable role-based aliases (strix-moe, gpu-dense, gpu-light) from gpu-fleet.
+  Stable role-based aliases (strix-moe, gpu-dense, gpu-vision) from gpu-fleet.
 agent: abiba
 depends_on:
   - gpu-monitor.prose.md (live data source on .24:9100)
@@ -52,8 +52,8 @@ depends_on:
 
 Key notes:
 - All models use direct GPU routing via LiteLLM (`api_key: not-needed`). Router (port 9000) was decommissioned 2026-09-11 and is NOT in the inference path.
-- Stable aliases (gpu-dense, gpu-light, strix-moe) from gpu-fleet are the canonical names for agent configs. Model-specific names still work but are deprecated.
-- RTX 5070 tok/s is ~145 for Qwen3.5-9B — gpu-light is the fastest endpoint. Route vision/web/light work there first. NOTE: Qwen3.5-9B is multimodal (image+text), NOT text-only like gemma-4-12b was.
+- Stable aliases (gpu-dense, gpu-vision, strix-moe) from gpu-fleet are the canonical names for agent configs. Model-specific names still work but are deprecated. The retired names `gpu-light` and `gemma-4-12b` were superseded by `gpu-vision` on 2026-09-12 and no longer resolve (400 `Invalid model name`).
+- The RTX 5070 is the fastest endpoint per token — `gpu-vision` is its canonical alias. Route vision/web/light work there first. The RTX 5070 model is multimodal (image+text).
 - Strix Halo is 62.9 tok/s (89% of 70.5 baseline) — below optimal but stable. Check for competing workloads.
 - RTX 3090 VRAM at 88% — within role-appropriate range (role = heavy reasoning, needs the headroom).
 - RTX 5070 VRAM at 83% — role-appropriate for vision/web (smaller batch sizes).
@@ -64,7 +64,7 @@ Key notes:
 - **Detect**: Any GPU temp >85°C sustained for 2+ consecutive polls
 - **Fix**: 
   1. Reduce inference concurrency on that GPU (load-side cooling only — NO fan control)
-  2. Redirect new requests to cooler GPUs via LiteLLM fallback chains (gemma → qwen, qwen → gemma)
+  2. Redirect new requests to cooler GPUs via LiteLLM fallback chains (gpu-vision → gpu-dense, gpu-dense → gpu-vision)
   3. If all GPUs hot, alert about cooling infrastructure
 - **Verify**: Temp drops below 80°C within 5 minutes
 - **Escalate after**: 3 verification failures → Zulip alert
@@ -157,13 +157,14 @@ Key notes:
 ### Rule 10: Workload Distribution Optimization (updated 2026-07-18)
 - **Detect**: GPU roles misaligned with hardware capabilities
 - **Target distribution**:
-  - RTX 3090 (gpu-dense, 24GB, 74.9 tok/s) → Heavy reasoning, code gen, long conversations (slowest per-token but largest context capacity). Weight: 0.55 (LiteLLM).
-  - RTX 5070 (gpu-light, 12GB, ~145 tok/s) → Vision (image+text), web search, lightweight tasks. Weight: 0.15 (LiteLLM).
-  - Strix Halo (strix-moe, 64GB, 62.9 tok/s) → Context compression, summarization, long docs (MoE model). Weight: 0.30 (LiteLLM).
+  - RTX 3090 (gpu-dense, 24GB, 74.9 tok/s) → Heavy reasoning, code gen, long conversations (slowest per-token but largest context capacity).
+  - RTX 5070 (gpu-vision, 12GB, ~145 tok/s) → Vision (image+text), web search, lightweight tasks.
+  - Strix Halo (strix-moe, 64GB, 62.9 tok/s) → Context compression, summarization, long docs (MoE model).
 - **Note**: RTX 5070 is the fastest endpoint per token. Route high-volume, low-complexity work there first.
+- **Weights are not restated here** — the live `syslog-auto` pool weights and rpm caps live in CT 116 `/opt/inference-harness/litellm_config.yaml`, the single source of truth.
 - **Fix**:
   - Alert if any GPU is handling workload outside its designated role
-  - Recommend agent alias updates to match workload to GPU role (use stable aliases: gpu-dense, gpu-light, strix-moe)
+  - Recommend agent alias updates to match workload to GPU role (use stable aliases: gpu-dense, gpu-vision, strix-moe)
   - Track per-GPU request distribution via LiteLLM spend logs
 - **Verify**: Each GPU's request pattern matches its designated role within 24h
 - **Escalate**: If role mismatch persists >48h → agent alias audit needed
@@ -311,6 +312,6 @@ Pushed to `SyslogSolution/health-logs/gpu/{run_id}.json` — versioned, searchab
   If monitor response > 1MB, log a warning and skip the cycle rather than crashing.
 
 ### L6: Stable Aliases Replace Model Names
-- gpu-fleet introduced stable aliases (strix-moe, gpu-dense, gpu-light) on 2026-07-15.
+- gpu-fleet introduced stable aliases (strix-moe, gpu-dense, gpu-light) on 2026-07-15; `gpu-light` was superseded by `gpu-vision` on 2026-09-12.
 - Self-heal must use aliases for reporting and alerting, not model-specific names.
 - **Rule**: All alert messages and KG nodes use the stable alias as the GPU identifier.

hermes-agent-baseline.prose.md

@@ -80,7 +80,7 @@ custom_providers:
 auxiliary:
   vision:
     provider: harness
-    model: gemma-4-12b              # or syslog-auto
+    model: gpu-vision              # RTX 5070 stable alias (Rule 8; do not use syslog-auto for aux)
     base_url: http://192.168.68.116/litellm/v1
     api_key_env: LITELLM_API_KEY
     api_key: <value from: infisical secrets get LITELLM_API_KEY --project=agents --env=production>   # ← MANDATORY workaround
@@ -95,7 +95,7 @@ auxiliary:
     threshold: 0.65
     target_ratio: 0.3
     provider: harness
-    model: syslog-auto               # or gemma-4-12b
+    model: syslog-auto               # Rule 7: compression must be syslog-auto
     base_url: http://192.168.68.116/litellm/v1
     api_key_env: LITELLM_API_KEY
     api_key: <value from: infisical secrets get LITELLM_API_KEY --project=agents --env=production>   # ← MANDATORY workaround
@@ -185,7 +185,9 @@ Abiba (CT100) runs pi via PM2 with the Zulip extension.
 Config files: `~/.pi/agent/models.json`, `~/.pi/agent/settings.json`.
 
 **models.json** — Must only list models authorized for the agent's LiteLLM key.
-Key is injected via `infisical run --` wrapper at PM2 startup:
+`/v1/models` is key-scoped and the live registry is CT 116
+`/opt/inference-harness/litellm_config.yaml`; treat the list below as a snapshot and re-read
+the registry before applying. Key is injected via `infisical run --` wrapper at PM2 startup:
 ```json
 {
   "providers": {
@@ -197,9 +199,8 @@ Key is injected via `infisical run --` wrapper at PM2 startup:
         { "id": "syslog-auto" },
         { "id": "strix-moe" },
         { "id": "gpu-dense" },
-        { "id": "gpu-light" },
-        { "id": "qwen3.6-27B-code" },
-        { "id": "gemma-4-12b" }
+        { "id": "gpu-vision" },
+        { "id": "qwen3.6-27B-code" }
       ]
     }
   }

hermes-config-template.prose.md

@@ -102,7 +102,7 @@ model:
                                   # and falls back to 256K when /v1/models lacks a context
                                   # field (llama-server does). Without this override, agents
                                   # silently run syslog-auto at 256K (verified 2026-08-09).
-                                  # Set 65536 if using gemma-4-12b directly (tight VRAM).
+                                  # Set 65536 if pinning a single model directly (tight VRAM).
 
 fallback_providers:
   provider: deepseek
@@ -143,25 +143,25 @@ compression:
 
 # ─── Auxiliary Tasks (CONSISTENCY RULE) ───
 # All auxiliary services MUST use identical model, base_url, and api_key_env:
-#   model: gpu-light                 # stable alias (NOT raw "gemma-4-12b")
+#   model: gpu-vision                # stable alias (NOT a raw model name)
 #   base_url: http://192.168.68.116/litellm/v1   # Rule 5 (2026-08-09): canonical authenticated; /v1 also OK
 #   api_key_env: LITELLM_API_KEY
 # Do NOT use syslog-auto for auxiliary tasks — it routes to the primary GPU.
-# gpu-light = RTX 5070 (12B), freeing the Strix Halo for agent reasoning.
+# gpu-vision = 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.
+# NEVER use raw model names (e.g. qwen3.6-27B-code, qwen3.6-35B-udq4; gemma-4-12b is retired
+# and no longer resolves) in agent configs — use the stable aliases so model swaps don't break agents.
 auxiliary:
   vision:
     provider: harness
-    model: gpu-light            # stable alias for RTX 5070 (was raw gemma-4-12b)
+    model: gpu-vision           # stable alias for RTX 5070
     base_url: http://192.168.68.116/litellm/v1   # Rule 5 (2026-08-09): canonical authenticated; /v1 also OK
     api_key_env: LITELLM_API_KEY
     timeout: 60
     download_timeout: 30
   web_extract:
     provider: harness
-    model: gpu-light            # stable alias for RTX 5070
+    model: gpu-vision           # stable alias for RTX 5070
     base_url: http://192.168.68.116/litellm/v1   # Rule 5 (2026-08-09): canonical authenticated; /v1 also OK
     api_key_env: LITELLM_API_KEY
     timeout: 30
@@ -252,10 +252,11 @@ The following MUST be identical across ALL profiles:
 - For agents needing longer outputs: raise to 8192, but never omit
 
 ### Rule 7: Auxiliary Model Consistency (UPDATED 2026-07-16)
-- Vision and web_extract use `gemma-4-12b` (RTX 5070 — 12GB, vision-optimized)
-- Compression uses `strix-moe` (stable alias for Strix Halo — 64GB, 128K ctx, compression-optimized)
-- **`strix-moe` is the only valid compression model name** — LiteLLM does NOT serve `ornith-1.0-35b`
-  (it serves `strix-moe`, `qwen3.6-35B-udq4`, `gpu-dense`, `gpu-light`, `syslog-auto`, `gemma-4-12b`, `qwen3.6-27B-code`). Old configs with `ornith-1.0-35b` cause 403/model-not-found on compression calls.
+- Vision and web_extract use `gpu-vision` (RTX 5070 — 12GB, vision-optimized)
+- Compression uses `syslog-auto` — the Strix Halo weighted pool (64GB, 128K ctx, compression-optimized); do NOT pin `compression.model` to `strix-moe` (audit Rule 7 rejects it)
+- **`ornith-1.0-35b` is NOT a valid compression model name** — LiteLLM does not serve it
+  (do not restate the served model list here — CT 116 `/opt/inference-harness/litellm_config.yaml`
+  is the single source of truth for models, aliases, weights and fallbacks). Old configs with `ornith-1.0-35b` cause 403/model-not-found on compression calls.
 - **OPERATIONAL DECISION (2026-07-23): Use `syslog-auto` for compression across all agents.**
   The `syslog-auto` alias routes to the Strix Halo, but uses the weighted pool instead of pinning
   to `strix-moe` directly. This prevents sustained Strix Halo thermal load because the pool can
@@ -264,7 +265,7 @@ The following MUST be identical across ALL profiles:
 - All auxiliary services MUST use identical routing:
   - `base_url: http://192.168.68.116/litellm/v1`  (Rule 5, 2026-08-09: canonical authenticated; `/v1` also OK)
   - `api_key_env: LITELLM_API_KEY`
-- **Do NOT use `syslog-auto`** for auxiliary tasks — it routes unpredictably
+- **Do NOT use `syslog-auto` for `vision`/`web_extract`** — it routes unpredictably; compression is the deliberate exception (see the OPERATIONAL DECISION above)
 - **Compression on Strix Halo**: The strix-moe alias routes to Strix Halo
   (64GB UMA, 128K context) — the designated compression GPU. This frees the
   RTX 5070 for vision and web search, and the RTX 3090 for heavy reasoning.
@@ -273,11 +274,11 @@ The following MUST be identical across ALL profiles:
 
 ### Rule 8: GPU Workload Distribution (UPDATED 2026-07-16)
 - **RTX 3090 (24GB, 128K ctx, qwen3.6-27B-code)**: Heavy reasoning, code gen, long conversations
-- **RTX 5070 (12GB, 128K ctx, gemma-4-12b)**: Vision, web search, quick tasks, web_extract (IQ4_NL+MTP, ~65% VRAM at 128K)
+- **RTX 5070 (12GB, 128K ctx, gpu-vision)**: Vision, web search, quick tasks, web_extract (IQ4_NL+MTP, ~65% VRAM at 128K)
 - **Strix Halo (64GB, 128K ctx, syslog-auto)**: 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.vision.model: gpu-vision` (RTX 5070)
+  - `auxiliary.web_extract.model: gpu-vision` (RTX 5070)
   - `auxiliary.compression.model: syslog-auto` (Strix Halo)
 - Default model (`model.default`) and custom_provider remain `syslog-auto` for auto-routing
 - For 128K context window: `threshold: 0.65` (fires at ~85K tokens)
@@ -297,8 +298,8 @@ The following MUST be identical across ALL profiles:
 - **Koby Exception**: Per captain ruling 2026-08-11, Koby is a DeepSeek-primary external agent; its primary model remains `deepseek-v4-flash` (via api.deepseek.com to preserve DeepSeek-specific reasoning, while other sections follow Rule 10.
 - **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 strix-moe
-  and qwen3.6-27B-code, with gemma-4-12b as fallback. Using it protects against:
+- `syslog-auto` is the LiteLLM routing model — it load-balances across the live pool
+  (see CT 116 `/opt/inference-harness/litellm_config.yaml` for the current members and weights). Using it protects against:
   - Model name typos that cause 403 errors and silent worker failures
   - Single GPU downtime (routing falls back automatically)
   - Key/model authorization mismatches

hermes-key-enforcement.prose.md

@@ -169,16 +169,20 @@ The agent picks up the new key via `infisical run --` at gateway startup.
 
 **Keys are permanent and use bare agent name aliases.**
 
-- **Duration**: `null` — keys never expire. This is enforced by `default_key_generate_params` in `litellm_config.yaml`.
+- **Duration**: `null` — keys never expire. NOT enforced today: CT 116 `litellm_config.yaml` has no `default_key_generate_params` block, and a key generated with no explicit models comes back with an empty models list. OPEN policy question: should agent keys expire by default? (captain security-policy decision, raised separately.)
 - **Alias convention**: bare agent name only (e.g., `tanko`, `mumuni`, `koby`, `koonimo`). No dates, no versions. The alias IS the identity.
 - **Rotation triggers**: compromise, personnel departure, or quarterly security hygiene. NOT calendar-driven.
 - **Max budget**: $100 per key (config default).
 
 ```yaml
-# In litellm_config.yaml — ensures all future keys inherit these defaults:
+# NOT currently set in the authority; recommended value. CT 116 litellm_config.yaml has no
+# default_key_generate_params block today, and a key generated with no explicit models comes back
+# with an EMPTY models list. `models` is a literal key-generation parameter, so this is a value to
+# ADD — re-read the live registry at CT 116 /opt/inference-harness/litellm_config.yaml and
+# re-verify before applying.
 litellm_settings:
   default_key_generate_params:
-    models: ["syslog-auto", "qwen3.6-27B-code", "gemma-4-12b"]
+    models: ["syslog-auto", "qwen3.6-27B-code", "gpu-vision"]
     duration: null        # ← permanent
     max_budget: 100
     metadata:
@@ -286,13 +290,13 @@ auxiliary:
     api_key: sk-<agent-key-from-vault>           # ← workaround (get via: infisical secrets get LITELLM_API_KEY --project=agents --env=production --plain)
     api_key_env: LITELLM_API_KEY
     base_url: http://192.168.68.116/litellm/v1
-    model: gemma-4-12b
+    model: gpu-vision
     provider: harness
   compression:
     api_key: sk-<agent-key-from-vault>           # ← workaround (same as above)
     api_key_env: LITELLM_API_KEY
     base_url: http://192.168.68.116/litellm/v1
-    model: gemma-4-12b
+    model: syslog-auto
     provider: harness
 ```
 

inference-optimization.prose.md

@@ -23,7 +23,7 @@ management, and prompt caching — without sacrificing agent capability.
   .123, any others on .129/.122) including compression, model, context_window,
   prompt_caching, memory settings
 - `gpu-health`: health check response from all 3 GPU backends (strix-moe .15:8080,
-  qwen .8:8080, gemma .110:8080)
+  gpu-dense .8:8080, gpu-vision .110:8080)
 
 ### Maintains
 
@@ -56,8 +56,8 @@ duration.
 **Context is the root cause.** Every ~46K prompt token costs ~87s of
 prefill time at 532 tok/s. Fix context first, routing second.
 
-- **Route by task**: qwen for code/standard queries; gemma for
-  compression/auxiliary; strix-moe for compression tasks.
+- **Route by task**: gpu-dense for code/standard queries; gpu-vision for
+  vision/web-auxiliary; syslog-auto for compression.
 - **Compress aggressively**: threshold at 40% (not 65%) — a 128K window should
   compact at 51K, not 85K. Target 15% tail (not 30%).
 - **Cache everything repeated**: system prompts, skill docs, AGENTS.md — these
@@ -96,8 +96,10 @@ call enable-prompt-caching
   hosts: [192.168.68.15, 192.168.68.8, 192.168.68.110]
 
 -- Phase 5: Verify end-to-end latency
+-- `models` is a literal verification parameter (a snapshot only): the authoritative registry is
+-- CT 116 /opt/inference-harness/litellm_config.yaml; re-read it before use.
 
 call verify-latency
   host: 192.168.68.116
-  models: [syslog-auto, qwen3.6-27B-code, gemma-4-12b]
+  models: [syslog-auto, qwen3.6-27B-code, gpu-vision]
 ```

infrastructure-control.prose.md

@@ -222,7 +222,7 @@ description: >
 
 **Prometheus targets**:
 - 192.168.68.8:9400 (RTX 3090 — qwen)
-- 192.168.68.110:9400 (RTX 5070 — gemma)
+- 192.168.68.110:9400 (RTX 5070 — gpu-vision)
 - 192.168.68.15:9400 (Strix Halo — qwen3.6-35B-udq4)
 - harness-litellm:4000 (LiteLLM health)
 

litellm-api-keys.prose.md

@@ -67,8 +67,10 @@ description: >
 4. **If action == "create"**:
    - 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", "strix-moe", "gpu-dense", "gpu-light", "qwen3.6-35B-udq4"]
+   - Duration is whatever the caller passes; NO default enforcement exists today (CT 116 `litellm_config.yaml` has no `default_key_generate_params` block, and a key with no explicit models returns an empty models list). Agent keys are permanent by policy, not by that block. OPEN policy question: should agent keys expire by default? (captain security-policy decision, raised separately.)
+   - Set models: read the live key-scoped set rather than hardcoding one — `/v1/models` is key-scoped,
+     and the authoritative registry is CT 116 `/opt/inference-harness/litellm_config.yaml`. Do not add
+     retired names (`gemma-4-12b`, `gpu-light`, `crew-auto` — all retired 2026-09-12).
    - 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"**:

litellm-client-timeouts.prose.md

@@ -28,7 +28,7 @@ description: >
 | syslog-auto | 28.8s | 25.4s | 68 calls took 30-120s; tail to ~300s under load |
 | qwen3.6-27B-code | 23.0s | — | same backend class as syslog-auto |
 | strix-moe | 7.5s | — | Strix Halo, healthy |
-| gemma-4-12b | 2.6s | — | RTX 5070, healthy |
+| gemma-4-12b (retired 2026-09-12; RTX 5070 now `gpu-vision`) | 2.6s | — | RTX 5070, healthy |
 
 Sep 6 incident timeline: failures 04:00-07:00 EDT (0% GPU util = wedged
 backend), full recovery 07:00-08:00 with ZERO client failures once requests
@@ -53,8 +53,7 @@ proxy queuing.
 
 ### 2. Auxiliary tasks — keep template timeouts, one correction
 
-- vision: 60s (keep), web_extract: 30s (keep) — gemma-4-12b averages 2.6s;
-  these are fine.
+- vision: 60s (keep), web_extract: 30s (keep) — the 2.6s average was measured on `gemma-4-12b` (retired 2026-09-12); the live RTX 5070 alias is `gpu-vision`.
 - compression: 300s (keep — this was already raised from 60 per gpu-fleet).
 - **gpu-dense delegation/x_search: set timeout >= 120s.** The RTX 3090
   (qwen3.6-27B-code backend, 23.0s avg) is the same speed class as

tests/test_audit_hermes_config_alias.py

@@ -0,0 +1,191 @@
+"""Regression tests for the 2026-09-12 retired-alias sweep in audit-hermes-config.py.
+
+WHY THIS FILE EXISTS: the executable audit pushed agent configs toward a DEAD alias.
+Rule 8 required `auxiliary.vision.model == "gpu-light"` and
+`auxiliary.web_extract.model == "gpu-light"`, but `gpu-light` (and its raw predecessor
+`gemma-4-12b`) were retired on 2026-09-12 and now return 400 `Invalid model name`; the
+live RTX 5070 alias is `gpu-vision`. A config that adopted the correct canonical alias
+therefore FAILED our own audit, so the audit was actively enforcing a broken config.
+
+These tests execute the real CLI (`python3 audit-hermes-config.py <config>`) and assert
+observable behaviour — exit code and the emitted rule message — for the live alias and
+for both retired names. No network, vault, or SSH access is required.
+"""
+from __future__ import annotations
+
+import pathlib
+import subprocess
+import sys
+
+ROOT = pathlib.Path(__file__).resolve().parent.parent
+AUDIT = ROOT / "audit-hermes-config.py"
+
+BASE = """
+model:
+  api_key: ""
+  api_key_env: LITELLM_API_KEY
+  base_url: http://192.168.68.116/v1
+  max_tokens: 4096
+  default: syslog-auto
+  provider: harness
+fallback_providers:
+  provider: deepseek
+  model: deepseek-v4-flash
+  api_key_env: DEEPSEEK_API_KEY
+compression:
+  model: syslog-auto
+  provider: harness
+  threshold: 0.65
+  max_context_window: 131072
+auxiliary:
+  vision:
+    model: {alias}
+    provider: harness
+  web_extract:
+    model: {alias}
+    provider: harness
+  compression:
+    model: syslog-auto
+    provider: harness
+delegation:
+  provider: harness
+custom_providers:
+  - name: harness
+    key_env: LITELLM_API_KEY
+    base_url: http://192.168.68.116/v1
+"""
+
+
+def _run_config(tmp_path, name, text):
+    cfg = tmp_path / name
+    cfg.write_text(text)
+    proc = subprocess.run(
+        [sys.executable, str(AUDIT), str(cfg)],
+        capture_output=True, text=True,
+    )
+    return proc.returncode, proc.stdout
+
+
+def _run(tmp_path, alias):
+    return _run_config(tmp_path, f"{alias}.yaml", BASE.format(alias=alias))
+
+
+def test_live_canonical_alias_passes(tmp_path):
+    """The RTX 5070 alias that actually resolves must satisfy Rule 8."""
+    code, out = _run(tmp_path, "gpu-vision")
+    assert code == 0, out
+    assert "RESULT: PASS" in out
+
+
+def test_retired_gpu_light_is_rejected(tmp_path):
+    """A config pinned to the retired alias must fail, not pass."""
+    code, out = _run(tmp_path, "gpu-light")
+    assert code == 1, out
+    assert "auxiliary.vision.model must be gpu-vision" in out
+    assert "RESULT: FAIL" in out
+
+
+def test_retired_gemma_is_rejected(tmp_path):
+    """The retired raw model name must fail Rule 8 as well."""
+    code, out = _run(tmp_path, "gemma-4-12b")
+    assert code == 1, out
+    assert "auxiliary.vision.model must be gpu-vision" in out
+    assert "RESULT: FAIL" in out
+
+
+def test_corrected_compression_example_passes(tmp_path):
+    """The corrected workaround (vision=gpu-vision, compression=syslog-auto) must PASS."""
+    code, out = _run(tmp_path, "gpu-vision")
+    assert code == 0, out
+    assert "[Rule 7] compression.model must be syslog-auto (got 'syslog-auto')" in out
+    assert "[Rule 7] auxiliary.compression.model must be syslog-auto (got 'syslog-auto')" in out
+    assert "RESULT: PASS" in out
+
+
+def test_retired_alias_in_delegation_is_rejected(tmp_path):
+    """delegation.model has no dedicated value rule, so a retired name there used to PASS."""
+    code, out = _run_config(
+        tmp_path,
+        "delegation-gpu-light.yaml",
+        BASE.format(alias="gpu-vision").replace(
+            "delegation:\n  provider: harness",
+            "delegation:\n  provider: harness\n  model: gpu-light",
+        ),
+    )
+    assert code == 1, out
+    assert "delegation.model = 'gpu-light' is retired" in out
+    assert "RESULT: FAIL" in out
+
+
+def test_retired_alias_in_custom_providers_is_rejected(tmp_path):
+    """custom_providers[*].model is model-bearing; a retired name there must fail."""
+    code, out = _run_config(
+        tmp_path,
+        "custom-provider-gpu-light.yaml",
+        BASE.format(alias="gpu-vision").replace(
+            "  - name: harness\n    key_env: LITELLM_API_KEY",
+            "  - name: harness\n    model: gpu-light\n    key_env: LITELLM_API_KEY",
+        ),
+    )
+    assert code == 1, out
+    assert "custom_providers[0].model = 'gpu-light'" in out
+    assert "RESULT: FAIL" in out
+
+
+def test_raw_but_live_alias_warns_but_passes(tmp_path):
+    """Raw-but-live names resolve (200), so they warn only; failing them rejects valid configs."""
+    code, out = _run_config(
+        tmp_path,
+        "raw-qwen.yaml",
+        BASE.format(alias="gpu-vision").replace(
+            "delegation:\n  provider: harness",
+            "delegation:\n  provider: harness\n  model: qwen3.6-27B-code",
+        ),
+    )
+    assert code == 0, out
+    assert "delegation.model = 'qwen3.6-27B-code' is a raw-but-live model name" in out
+    assert "prefer the stable alias gpu-dense" in out
+    assert "RESULT: PASS" in out
+
+
+def test_retired_alias_in_fallback_providers_is_rejected(tmp_path):
+    """fallback_providers.model is model-bearing; a retired name there must fail."""
+    code, out = _run_config(
+        tmp_path,
+        "fallback-gpu-light.yaml",
+        BASE.format(alias="gpu-vision").replace("  model: deepseek-v4-flash", "  model: gpu-light"),
+    )
+    assert code == 1, out
+    assert "fallback_providers.model = 'gpu-light'" in out
+    assert "RESULT: FAIL" in out
+
+
+def test_retired_alias_in_x_search_is_rejected(tmp_path):
+    """x_search.model was previously not enumerated; the derivation must catch it."""
+    code, out = _run_config(
+        tmp_path,
+        "x-search-gpu-light.yaml",
+        BASE.format(alias="gpu-vision").replace(
+            "delegation:\n  provider: harness",
+            "delegation:\n  provider: harness\nx_search:\n  model: gpu-light",
+        ),
+    )
+    assert code == 1, out
+    assert "x_search.model = 'gpu-light'" in out
+    assert "RESULT: FAIL" in out
+
+
+def test_retired_alias_in_nested_auxiliary_block_is_rejected(tmp_path):
+    """A nested auxiliary sub-block outside the named three must still be derived."""
+    code, out = _run_config(
+        tmp_path,
+        "nested-aux-gpu-light.yaml",
+        BASE.format(alias="gpu-vision").replace(
+            "  compression:\n    model: syslog-auto\n    provider: harness\ndelegation:",
+            "  compression:\n    model: syslog-auto\n    provider: harness\n"
+            "  tasks:\n    summarize:\n      model: gpu-light\ndelegation:",
+        ),
+    )
+    assert code == 1, out
+    assert "auxiliary.tasks.summarize.model = 'gpu-light'" in out
+    assert "RESULT: FAIL" in out