@iowarp/clio-coder 0.4.1 → 0.4.3

package/src/domains/providers/models/local-models/clio-coder-local-coding-targets.yaml

@@ -9,8 +9,11 @@
 # https://huggingface.co/Jackrong/Qwopus3.6-35B-A3B-Coder-MTP-GGUF
 # https://huggingface.co/nvidia/Nemotron-Cascade-2-30B-A3B
 # https://huggingface.co/unsloth/NVIDIA-Nemotron-3-Nano-Omni-30B-A3B-Reasoning-GGUF
+# https://huggingface.co/nvidia/NVIDIA-Nemotron-3.5-Lightning-30B-A3B-BF16
 # https://huggingface.co/mradermacher/Qwen3.5-35B-A3B-Claude-4.6-Opus-Reasoning-Distilled-i1-GGUF
 # https://huggingface.co/deepreinforce-ai/Ornith-1.0-35B-GGUF
+# https://huggingface.co/meta-models/Muse-Glimmer-30B
+# https://huggingface.co/bottlecapai/ThinkingCap-Qwen3.6-27B
 #
 # This knowledge base is intentionally narrow. It describes only the local
 # models curated for Clio's local coding workflows and leaves cloud GPT models
@@ -26,8 +29,7 @@
 # not be established from git history or the release-test reports under
 # docs/release-notes/ read `unknown` rather than a plausible guess. A
 # `measuredUnder` block is provenance for a reader; nothing in the engine
-# consumes it (`extractLocalModelQuirks` narrows only kvCache, sampling, and
-# thinking).
+# consumes it (`extractLocalModelQuirks` narrows only sampling and thinking).
 #
 # `llamaCpp.parallel` is the recommended `--parallel` value for starting the
 # server, and families that carry one also carry a `parallelSlots` note. It is
@@ -73,13 +75,12 @@
       thinking:
         temperature: 1
         topP: 1
-        maxTokens: 32768
       instruct:
         temperature: 1
         topP: 1
-        maxTokens: 8192
     runtimePreference:
-      llamaCpp: "Use the OpenAI-compatible chat-completions surface with chat_template_kwargs.reasoning_effort set to low, medium, or high."
+      llamaCpp: "Use the OpenAI-compatible chat-completions surface with chat_template_kwargs.reasoning_effort set to low, medium, or high. The analysis channel arrives as reasoning_content."
+      lmstudio: "Reasoning effort travels as the top-level reasoning_effort field. The analysis channel arrives as the OpenAI-compatible `reasoning` field, on the streamed delta and on the final message, with reasoning_content absent (measured 2026-09-02 on dynamo, gpt-oss-20b and gpt-oss-120b); Clio reads both spellings into thinking content."
       openaiCompat: "Preferred for Harmony reasoning-effort control when the local gateway exposes chat_template_kwargs."
     thinking:
       mechanism: effort-levels
@@ -91,7 +92,9 @@
         GPT-OSS uses OpenAI Harmony formatting and supports low, medium, and
         high reasoning effort. Clio coerces off/minimal to low, captures
         non-final Harmony channels as ThinkingContent, and strips Harmony
-        control tokens from visible output.
+        control tokens from visible output. The analysis channel reaches Clio
+        as reasoning_content on llama.cpp and as reasoning on LM Studio; both
+        are captured the same way.
 
 - family: agenticqwen-30b-a3b-i1
   matchPatterns:
@@ -130,13 +133,11 @@
         temperature: 0.5
         topP: 0.9
         topK: 20
-        maxTokens: 16384
         reasoningBudget: 4096
       instruct:
         temperature: 0.5
         topP: 0.9
         topK: 20
-        maxTokens: 4096
     gpuTiers:
       "32gb": "Default 32 GB local llama.cpp orchestrator profile; benchmarked with roughly 1 GiB VRAM headroom on a 32 GiB class card."
     runtimePreference:
@@ -170,6 +171,10 @@
     - nemotron-3-nano-omni-30b-a3b-reasoning
     - nemotron-3-nano-omni
     - nemotron-nano-omni
+    # Both router spellings. The live mini router advertises the id with `moe`
+    # in it; issue #263 and the round-5 handoffs use the shorter one.
+    - nemotron3-30b-moe-omni
+    - nemotron3-30b-omni
   capabilities:
     chat: true
     tools: true
@@ -185,18 +190,45 @@
     contextWindow: 1048576
     maxTokens: 131072
   quirks:
+    measuredUnder:
+      hardware: "mini fleet node, AMD Radeon AI PRO R9700 32 GB, Vulkan backend, 4 request slots"
+      runtime: llamacpp
+      build: "worker llama-server from /home/akougkas/src/llama.cpp-b10687/build-vulkan/bin, behind a router reporting build_info b1-c841aee (the string the ornith-1.5 entry below records for this same host on this date)"
+      model: "Nemotron-3-Nano-Omni-30B-A3B-Reasoning UD-Q4_K_M with Nemotron-Omni-mmproj-F16 (arch nemotron_h_moe)"
+      llamaCpp: >-
+        --ctx-size 1032192 --parallel 4 --kv-unified --cache-type-k f16
+        --cache-type-v f16 --batch-size 16384 --ubatch-size 2048 --flash-attn
+        true --n-gpu-layers 99 --jinja --reasoning on --temperature 0.6 --top-k
+        40 --top-p 0.95 --repeat-penalty 1.05
+      date: "2026-09-02"
+      source: "issue #263 thinking sweep: one non-streaming POST /v1/chat/completions per variant on a one-line arithmetic prompt, max_tokens 4096"
+      note: |
+        The on-off classification is measured, and it replaces a budget-tokens
+        dial that never reached this runtime. With no thinking field the model
+        spent 54 completion tokens with 140 characters of reasoning_content;
+        with chat_template_kwargs.enable_thinking false it answered in 4 tokens
+        with none; with it true it spent 42 tokens and 119 characters.
+        reasoning_effort moved nothing: none, low, and high produced 43, 48, and
+        39 tokens and a reasoning trace every time. The budget-tokens mechanism
+        this entry used to declare is downgraded to informational for
+        openai-completions on llama.cpp with a qwen chat template
+        (acceptsBudgetTokensField in model-runtime-capabilities.ts), so no
+        budget ever reached the server and the model reasoned at every level.
+        The router agrees with the measurement, advertising reasoning:on_off
+        with default_reasoning:on.
+
+        The measuring server is not the recommended profile below: it ran
+        ctx-size 1032192 with f16 KV against the 819200 and q8_0 recommended
+        here, and loaded only the vision mmproj, so the audio capability stays
+        the card's claim rather than this run's.
     sampling:
       thinking:
         temperature: 0.6
         topP: 0.95
         topK: 20
-        maxTokens: 20480
-        reasoningBudget: 16384
-        gracePeriod: 1024
       instruct:
         temperature: 0.2
         topK: 1
-        maxTokens: 1024
     gpuTiers:
       "24gb": "Use a 4-bit quant with reduced load context for evaluation. Prefer the 32GB tier for sustained Clio writes."
       "32gb": "Primary strong local model for multimodal main-agent work. Keep openai-compat available for tool-call fallback."
@@ -213,17 +245,131 @@
       parallel: 4
       parallelSlots: "Start the server with --parallel 4, and add --kv-unified unless each of the four slots is meant to get 204800 tokens rather than the whole 819200. Clio admits four concurrent requests on this endpoint once /props reports total_slots 4."
     thinking:
-      mechanism: budget-tokens
-      budgetByLevel:
-        low: 1024
-        medium: 4096
-        high: 16384
+      mechanism: on-off
+      chatTemplateKwargs:
+        static:
+          # Card recipe for the reasoning mode (scratchpad/catalog/models/nvidia-Nemotron-3-Nano-Omni-30B-A3B-Reasoning-FP8.md:7): reasoning_budget=16384.
+          reasoning_budget: 16384
+        lmstudio: unsupported
       guidance: |
-        Reasoning is template-driven; thinking blocks emit through the qwen
-        chat template. Preserve thinking across coding-task turns when the
-        chain matters; otherwise let the budget cap output growth.
+        Reasoning is template-driven and the switch is
+        chat_template_kwargs.enable_thinking; intermediate levels coerce to on.
+        The checkpoint reasons when nothing is sent, so off has to be carried on
+        the wire rather than omitted. reasoning_effort is inert on llama.cpp;
+        LM Studio reads the same on-off switch from that field instead.
     serving: "Use qwen3_coder tool parsing when served through OpenAI-compatible HTTP. LM Studio uses the same OpenAI-compatible tool surface."
 
+# NVIDIA's 3.5 Lightning generation, not a requant of the Nano Omni family
+# above: a different HF repo (nvidia/NVIDIA-Nemotron-3.5-Lightning-30B-A3B),
+# text-only where Nano Omni carries vision and audio, and a Mamba-2 plus MoE
+# plus attention hybrid rather than the omni checkpoint. They share the vendor,
+# the nemotron_h_moe arch string, and the 30B-A3B parameter count, which is not
+# the same as sharing weights.
+- family: nemotron-3.5-lightning-30b-a3b
+  matchPatterns:
+    # No bare `nemotron-3.5-lightning`: the 30b spelling is the shortest one
+    # that cannot swallow a future 8B or 70B Lightning id, which would inherit
+    # this checkpoint's sampler, its 1M window and a dial measured elsewhere.
+    - nvidia-nemotron-3.5-lightning-30b-a3b
+    - nemotron-3.5-lightning-30b-a3b
+    - nemotron-3.5-lightning-30b
+    - nemo3.5-30b-moe
+  capabilities:
+    chat: true
+    tools: true
+    toolCallFormat: qwen
+    reasoning: true
+    thinkingFormat: qwen-chat-template
+    structuredOutputs: json-schema
+    vision: false
+    audio: false
+    embeddings: false
+    rerank: false
+    fim: false
+    contextWindow: 1048576
+    maxTokens: 65536
+  quirks:
+    measuredUnder:
+      hardware: "mini fleet node, AMD Radeon AI PRO R9700 32 GB, Vulkan backend, 1 request slot"
+      runtime: llamacpp
+      build: "worker llama-server from /home/akougkas/src/llama.cpp-b10687/build-vulkan/bin, behind a router reporting build_info b1-c841aee"
+      model: "NVIDIA-Nemotron-3.5-Lightning-30B-A3B Q4_K_M (arch nemotron_h_moe)"
+      llamaCpp: >-
+        --ctx-size 1048576 --parallel 1 --kv-unified --cache-type-k f16
+        --cache-type-v f16 --batch-size 16384 --ubatch-size 2048 --flash-attn
+        true --fit off --n-gpu-layers 99 --jinja --reasoning on
+        --chat-template-kwargs {"enable_thinking": true} --temperature 1.0
+        --top-k 40 --top-p 0.95 --repeat-penalty 1.05
+      date: "2026-09-02"
+      source: "issue #263 thinking sweep: one non-streaming POST /v1/chat/completions per variant on a one-line arithmetic prompt, max_tokens 4096"
+      note: |
+        The mechanism is measured on this runtime and the card agrees with it.
+        With no thinking field the model spent 97 completion tokens with 300
+        characters of reasoning_content; with
+        chat_template_kwargs.enable_thinking false it answered in 3 tokens with
+        none; with it true it spent 120 tokens and 353 characters.
+        reasoning_effort is inert: none, low, and high produced 127, 114, and
+        117 tokens with a reasoning trace every time. The NVIDIA card states the
+        same control, "Reasoning Mode: Configurable on/off via chat template
+        (enable_thinking=True/False)", and the router advertises reasoning:on_off
+        with default_reasoning:on.
+
+        The same divergence is already recorded in code for this checkpoint:
+        model-runtime-capabilities.ts notes a 2026-08-11 fleet measurement on
+        nvidia-nemotron-3.5-lightning-30b-a3b where LM Studio suppressed
+        reasoning for reasoning_effort "none" and ignored enable_thinking false
+        while llama.cpp did the reverse. That id is a matchPattern here.
+
+        Sampler provenance: temperature 1.0 and top_p 0.95 are the card's
+        "Recommended Sampling"; top_k 40 and repeat_penalty 1.05 are the router
+        preset's launch arguments and have no card value. The card publishes no
+        separate thinking and instruct profiles, so both carry the same numbers.
+        maxTokens follows nemotron-cascade-2, the other 1M-context Nemotron in
+        this catalog; the card caps output only by the context window.
+    sampling:
+      thinking:
+        temperature: 1.0
+        topP: 0.95
+        topK: 40
+        repeatPenalty: 1.05
+      instruct:
+        temperature: 1.0
+        topP: 0.95
+        topK: 40
+        repeatPenalty: 1.05
+    gpuTiers:
+      "32gb": "Text-only agent target on a 32 GB card. Q4_K_M weights plus f16 KV reach the full 1048576-token window at parallel 1 on the measured server."
+    runtimePreference:
+      llamaCpp: "OpenAI-compatible chat completions with --jinja and --reasoning on. The router profile pins chat_template_kwargs.enable_thinking true, so a request that wants silence must send false rather than omitting the field."
+      openaiCompat: "Use when a gateway fronts the same llama.cpp chat-completions surface with qwen tool parsing."
+      lmstudio: "LM Studio reads the on-off switch from reasoning_effort, not from chat_template_kwargs; see the 2026-08-11 note above."
+    llamaCpp:
+      ctxSize: 1048576
+      cacheTypeK: f16
+      cacheTypeV: f16
+      flashAttn: true
+      nGpuLayers: 99
+      parallel: 1
+      parallelSlots: "Start the server with --parallel 1 so one agent gets the whole 1048576-token context, and add --kv-unified if --parallel is raised or the context is divided across slots. Clio discovers total_slots 1 and admits one request at a time on this endpoint, so a dispatch raised while the orchestrator streams is refused rather than queued."
+      batchSize: 16384
+      ubatchSize: 2048
+      chatTemplateKwargs:
+        enable_thinking: true
+    thinking:
+      mechanism: on-off
+      chatTemplateKwargs:
+        static:
+          # Card: "For coding agents, add extra_body={"chat_template_kwargs": {"force_nonempty_content": True}}" (scratchpad/catalog/models/nvidia-NVIDIA-Nemotron-3.5-Lightning-30B-A3B-BF16.md:21).
+          force_nonempty_content: true
+        lmstudio: unsupported
+      guidance: |
+        Nemotron 3.5 Lightning's chat template exposes thinking as on or off
+        through enable_thinking; intermediate levels coerce to on. It reasons by
+        default, so off has to be carried on the wire. On LM Studio the same
+        switch is read from reasoning_effort instead, which Clio sends for
+        on-off models on that runtime.
+    serving: "NVIDIA-Nemotron-3.5-Lightning-30B-A3B, MoE hybrid of Mamba-2, MoE and attention layers with 3B active parameters. The card recommends nemotron_v3 reasoning parsing and qwen3_coder tool parsing; the mini router serves the Q4_K_M gguf with XML tool calls through llama.cpp's qwen3_coder parser. Native context is 1M."
+
 - family: qwen3.6-35b-a3b
   matchPatterns:
     - qwen3.6-35b-a3b
@@ -245,14 +391,14 @@
     rerank: false
     fim: false
     contextWindow: 262144
-    maxTokens: 65536
+    # Vendor maximum for complex problems. Evidence: scratchpad/catalog/models/Qwen-Qwen3.6-35B-A3B.md.
+    maxTokens: 81920
   quirks:
     sampling:
       thinking:
         temperature: 0.6
         topP: 0.95
         topK: 20
-        maxTokens: 16384
         reasoningBudget: 4096
       instruct:
         temperature: 0.7
@@ -318,22 +464,17 @@
         family of behavior as qwopus3.6-35b-a3b-coder. The recommended
         llama.cpp serving profile below is the model card's, not that
         measurement's, and no argv from the measuring host survives.
-    kvCache:
-      kQuant: q8_0
-      vQuant: q8_0
     sampling:
       thinking:
         temperature: 0.6
         topP: 0.95
         topK: 20
-        maxTokens: 16384
         reasoningBudget: 4096
       instruct:
         temperature: 0.6
         topP: 0.95
         topK: 20
         repeatPenalty: 1.05
-        maxTokens: 4096
     gpuTiers:
       "32gb": "32 GB local profile with Q4_K_M, ctx=262144, q8 KV, flash attention, and parallel=1."
     runtimePreference:
@@ -358,6 +499,76 @@
         contains the final answer.
     serving: "Ornith-1.0-35B is an agentic coding model; the card recommends 262K context, qwen3 reasoning parsing, qwen3_coder/qwen3_xml tool parsing, and sampler temperature=0.6, top_p=0.95, top_k=20."
 
+- family: ornith-1.5
+  matchPatterns:
+    - ornith-1.5-35b-a3b
+    - ornith1.5-35b-moe
+    - ornith-1.5-35b
+    - ornith1.5-35b
+    - ornith-1.5
+    - ornith1.5
+  capabilities:
+    chat: true
+    tools: true
+    toolCallFormat: qwen
+    reasoning: true
+    thinkingFormat: qwen-chat-template
+    structuredOutputs: json-schema
+    vision: false
+    audio: false
+    embeddings: false
+    rerank: false
+    fim: false
+    contextWindow: 262144
+    maxTokens: 65536
+  quirks:
+    measuredUnder:
+      hardware: "mini fleet node, AMD GPU on the Vulkan backend, 4 request slots"
+      runtime: llamacpp
+      build: b1-c841aee
+      model: "Ornith-1.5-35B-A3B Q4_K_M (arch qwen35moe)"
+      llamaCpp: >-
+        --ctx-size 786432 --parallel 4 --no-kv-unified --cache-type-k q8_0
+        --cache-type-v q8_0 --batch-size 16384 --ubatch-size 2048 --jinja
+        --reasoning on
+      date: "2026-09-02"
+      source: "prompt-infrastructure round 2 (v0.4.2 development): live probe of the on-off switch, and a debugger dispatch receipt"
+      note: |
+        The 1.5 release is a Qwen3.5-MoE derivative and its chat template
+        honors enable_thinking, which the 1.0 measurement above did not find.
+        Measured on this date: with chat_template_kwargs.enable_thinking false
+        the model answered a one-line arithmetic prompt in 4 completion tokens
+        with no reasoning_content; with it true the same prompt spent 79
+        tokens, 158 characters of them reasoning. Under the inherited
+        always-on classification a debugger worker dispatched with thinking
+        off spent 1,061 of its 1,518 output tokens on a reasoning trace, and
+        the router itself advertises the model as reasoning:on_off. The
+        generic `ornith` pattern still routes 1.0 checkpoints to the entry
+        above; these longer patterns win for 1.5 ids.
+    sampling:
+      thinking:
+        temperature: 0.6
+        topP: 0.95
+        topK: 20
+        reasoningBudget: 4096
+      instruct:
+        temperature: 0.6
+        topP: 0.95
+        topK: 20
+        repeatPenalty: 1.05
+    thinking:
+      mechanism: on-off
+      guidance: |
+        Ornith 1.5's Qwen3.5 chat template exposes thinking as on or off
+        through enable_thinking; intermediate levels coerce to on. LM Studio
+        2026-09-03 exposed a split vocabulary on this exact Q4_K_M: native
+        discovery advertised {off,on}, while its OpenAI-compatible request
+        validator rejected literal `on` and accepted only effort-level values.
+        `low` ran only after a warning and fallback to on, so Clio now omits an
+        active override for this default-on model and sends `reasoning_effort:
+        none` only when disabling it.
+    serving: "Ornith-1.5-35B-A3B follows the 1.0 card: 262K context, qwen3 reasoning parsing, qwen3_coder tool parsing, sampler temperature=0.6, top_p=0.95, top_k=20."
+
 - family: qwen3.6-27b
   matchPatterns:
     - qwen3.6-27b
@@ -376,7 +587,8 @@
     rerank: false
     fim: false
     contextWindow: 262144
-    maxTokens: 32768
+    # Vendor maximum for complex problems. Evidence: scratchpad/catalog/models/Qwen-Qwen3.6-27B.md.
+    maxTokens: 81920
   quirks:
     measuredUnder:
       hardware: unknown
@@ -400,7 +612,6 @@
         minP: 0.0
         presencePenalty: 0.0
         repetitionPenalty: 1.0
-        maxTokens: 32768
       instruct:
         temperature: 0.7
         topP: 0.80
@@ -408,7 +619,6 @@
         minP: 0.0
         presencePenalty: 1.5
         repetitionPenalty: 1.0
-        maxTokens: 32768
     gpuTiers:
       "32gb": "Fits at f16 KV with parallel=1 and ctx=262144 (Q4_K_XL weights ~19.5 GB plus ~32 GB KV)."
     runtimePreference:
@@ -444,6 +654,7 @@
     rerank: false
     fim: false
     contextWindow: 262144
+    # Vendor maximum for final responses. Evidence: scratchpad/catalog/models/Qwen-Qwen3.8-27B.md.
     maxTokens: 131072
   quirks:
     measuredUnder:
@@ -468,9 +679,6 @@
         control were established earlier, on 2026-08-18, from a llama.cpp host
         and an LM Studio host whose build strings were not recorded; treat those
         two findings as unknown-build.
-    kvCache:
-      kQuant: q8_0
-      vQuant: q8_0
     sampling:
       thinking:
         temperature: 1.0
@@ -479,7 +687,6 @@
         minP: 0.0
         presencePenalty: 0.0
         repetitionPenalty: 1.0
-        maxTokens: 32768
       instruct:
         temperature: 0.7
         topP: 0.80
@@ -487,14 +694,13 @@
         minP: 0.0
         presencePenalty: 1.5
         repetitionPenalty: 1.0
-        maxTokens: 32768
     gpuTiers:
       "24gb-rocm": "24 GB ROCm class: IQ4_NL uses about 17.3 GB and UD-Q4_K_XL uses about 19.5 GB. Serve ctx=131072 with q8_0/q8_0 KV, parallel=2, and draft-mtp speculative decoding."
       "32gb-cuda": "32 GB CUDA class: Q6_K uses about 23.8 GB. Serve ctx=131072 with parallel=4 because the full 262144-token context is VRAM-bound."
       "128gb-unified-vulkan": "128 GB unified-memory Vulkan class: IQ4_NL uses about 17.3 GB and fits ctx=262144 with parallel=4. Vulkan backends show much slower prefill than ROCm/CUDA on this family; prefer ROCm where available."
     runtimePreference:
       llamaCpp: "Primary verified path. OpenAI-compatible chat completions with --jinja, --reasoning on, --reasoning-effort medium as server defaults; reasoning_content and tool_calls both confirmed to parse correctly. A request can override to off with chat_template_kwargs.enable_thinking:false, or to a different active level with reasoning_effort in {low, medium, xhigh} (never minimal/high/max on the wire; the stock template raises a fatal jinja exception on those and Clio's landed capability resolver already clamps to this set)."
-      lmstudioOpenaiCompat: "Clio sends reasoning_effort over LM Studio's OpenAI-compatible HTTP port and omits chat_template_kwargs because LM Studio does not use that field for this model. With no override, LM Studio defaults this model to thinking off, unlike raw llama.cpp's template default of thinking on/xhigh. reasoning_effort:'none' explicitly disables it."
+      lmstudioOpenaiCompat: "Clio sends reasoning_effort over LM Studio's OpenAI-compatible HTTP port and omits chat_template_kwargs because LM Studio does not use that field for this model. The upstream template defaults to thinking on/xhigh; an LM Studio deployment may configure another default. reasoning_effort:'none' explicitly disables it. Through LiteLLM, this control also requires allowed_openai_params:['reasoning_effort']; Clio selects the LM Studio spelling only when every deployment of that alias explicitly declares model_info.runtime: lm-studio."
       lmstudioNative: "Clio's LM Studio runtime uses the OpenAI-compatible HTTP chat adapter with native REST model discovery and residency management."
     llamaCpp:
       ctxSize: 131072
@@ -516,7 +722,6 @@
       effortByLevel:
         low: low
         medium: medium
-        high: xhigh
         xhigh: xhigh
       guidance: |
         Official Qwen3.8-27B chat template (chat_template.jinja) validates
@@ -524,19 +729,32 @@
         fatal jinja exception (hard 500, not a graceful fallback) on anything
         else, including minimal/high/max. The 27B card gives no /think or
         /no_think soft switch for this model; unlike Qwen3, that idiom has no
-        effect here. Off must go through chat_template_kwargs.enable_thinking:
-        false, never through reasoning_effort. preserve_thinking should stay
+        effect here. The template defaults to xhigh with thinking enabled.
+        The picker offers off/low/medium/xhigh; released high and max choices
+        normalize to xhigh rather than reaching the strict template unchanged.
+        Template-level off uses chat_template_kwargs.enable_thinking:false.
+        preserve_thinking should stay
         on for an agentic harness (the template's own recommendation for
         agent scenarios; improves KV cache reuse). llama.cpp's own
         reasoning_effort:"none" also disables reasoning at the server layer
         and is a separate mechanism from the template's enable_thinking flag;
-        both were verified independently on llama.cpp and on LM Studio's HTTP
-        port.
+        the none sentinel is consumed before strict template effort validation.
+        LM Studio's HTTP port also accepts none. A 2026-09-05 same-route
+        LiteLLM probe showed none and the template flag alone still reasoned
+        because the generic OpenAI route filtered the effort parameter. Adding
+        allowed_openai_params:["reasoning_effort"] produced zero reasoning
+        twice; an allowed low control still reasoned. Clio forwards only an
+        effort resolved from the model/runtime contract, and never infers the
+        upstream control runtime from a host name, port, or model alias.
 
         Thinking off is verified for this family, and no floor is imposed on
-        it. JetBrains report that Qwen3.8 without reasoning gets stuck in a
+        it. The official card and template were rechecked at revision
+        1d4bf0f2ff6012fd82039f2fa52739d0dd7c60c0:
+        https://huggingface.co/Qwen/Qwen3.8-27B/blob/1d4bf0f2ff6012fd82039f2fa52739d0dd7c60c0/chat_template.jinja
+        JetBrains report that Qwen3.8 without reasoning gets stuck in a
         loop repeating the same tool call indefinitely under Junie. Clio's
-        shipped default is orchestrator.thinkingLevel: off, which on llama.cpp
+        shipped default is chat.thinkingLevel: off, an explicit Clio preference
+        rather than the vendor's template default. On llama.cpp it
         sends chat_template_kwargs.enable_thinking:false and on LM Studio's
         OpenAI-compatible port sends reasoning_effort:"none". Clio 0.3.8 ran
         that configuration on both runtimes (llama.cpp build b226-2115b73d8 on
@@ -544,8 +762,12 @@
         2026-08-29 and 2026-08-30: an interactive build-and-resume session, a
         three-worker fleet step, and a ten-task eval suite totalling 116 model
         calls, none of which repeated a tool call into a loop. The catalog
-        therefore imposes no thinking floor and coerces no level for this
-        family. If a future build or quantization regresses into the loop
+        therefore imposes no thinking floor for this family. A provisional 2026-09-03 LM Studio A/B on the heavily quantized
+        `qwen3.8-27b-gsq-rco` alias corroborated the switch: four thinking-off
+        scenarios reported zero reasoning tokens and no repeated-tool loop,
+        versus visible reasoning at the active level. The host had LM Link
+        enabled, so this records behavior only—not hardware throughput or a
+        controlled quality claim. If a future build or quantization regresses into the loop
         JetBrains describe, the guards that catch it are Clio's own, not a
         catalog setting: the tool-prose-loop detector
         (src/interactive/tool-prose-loop.ts), which is armed on the
@@ -590,9 +812,6 @@
         exact NVFP4 build it was seen on. The KV-budget arithmetic in the
         gpuTiers and the serving note is derived from the official Google
         Gemma 4 31B card.
-    kvCache:
-      kQuant: q8_0
-      vQuant: q8_0
     sampling:
       thinking:
         temperature: 1.0
@@ -601,13 +820,11 @@
         minP: 0.0
         presencePenalty: 0.0
         repetitionPenalty: 1.0
-        maxTokens: 32768
       instruct:
         temperature: 0.7  # 0.7 keeps tool-call JSON parseable; gemma card's 1.0 is for chat
         topP: 0.95
         topK: 64
         minP: 0.0
-        maxTokens: 32768
     chatTemplate: "gemma-channel"
     leakageNote: |
       LM Studio's gemma4 chat template emits chain-of-thought as plain text
@@ -643,6 +860,11 @@
     - gemma-4-31b-it-qat-ud-q4_k_xl-mtp
     - gemma-4-31b-it-qat-ud
     - gemma-4-31b-it-qat
+    # The mini router's short id for this exact build; its alias is
+    # Gemma-4-31B-it-qat-UD-Q4_K_XL-MTP-262K, the first pattern above. A bare
+    # `gemma4-31b` would also swallow a future `gemma4-31b-nvfp4` id that
+    # belongs to the turbo family above, so the full router id is the pattern.
+    - gemma4-31b-dense
   capabilities:
     chat: true
     tools: true
@@ -658,9 +880,35 @@
     contextWindow: 262144
     maxTokens: 32768
   quirks:
-    kvCache:
-      kQuant: q8_0
-      vQuant: q8_0
+    measuredUnder:
+      hardware: "mini fleet node, AMD Radeon AI PRO R9700 32 GB, ROCm backend, 1 request slot"
+      runtime: llamacpp
+      build: "worker llama-server from /home/akougkas/src/llama.cpp-b10687/build-rocm/bin, behind a router reporting build_info b1-c841aee"
+      model: "Gemma-4-31B-it-qat UD-Q4_K_XL with mtp-gemma-4-31B-it draft and mmproj-F16 (arch gemma4)"
+      llamaCpp: >-
+        --ctx-size 139264 --parallel 1 --kv-unified --cache-type-k f16
+        --cache-type-v f16 --batch-size 1024 --ubatch-size 256 --flash-attn true
+        --fit off --n-gpu-layers 99 --spec-type draft-mtp --spec-draft-n-max 4
+        --n-gpu-layers-draft 99 --jinja --reasoning off --temperature 0.8
+        --top-k 64 --top-p 0.95 --repeat-penalty 1.05
+      date: "2026-09-02"
+      source: "issue #263 thinking sweep: one non-streaming POST /v1/chat/completions per variant on a one-line arithmetic prompt, max_tokens 4096"
+      note: |
+        The on-off mechanism this entry already claimed is now measured on
+        llama.cpp rather than inherited from the gemma-4 template. With no
+        thinking field, and with chat_template_kwargs.enable_thinking false, the
+        model answered in 3 completion tokens with no reasoning_content; with it
+        true the same prompt spent 68 tokens, 128 characters of them reasoning.
+        reasoning_effort is inert: none, low, and high all stayed at 3 tokens
+        with no trace, so this family is on-off and not effort-levels on this
+        runtime. The router serves it with --reasoning off, which is why the
+        no-field baseline is silent.
+
+        The measuring server is not the 32gb profile recommended below. It ran
+        ROCm at ctx-size 139264 with f16 KV, batch 1024 and ubatch 256, where the
+        block below recommends Vulkan at 262144 with q8_0 KV. The serving note
+        and gpuTiers keep the 262k recommendation; treat the argv above as the
+        configuration the token counts came from.
     sampling:
       thinking:
         temperature: 1.0
@@ -669,14 +917,12 @@
         minP: 0.0
         presencePenalty: 0.0
         repetitionPenalty: 1.0
-        maxTokens: 32768
       instruct:
         temperature: 0.7  # 0.7 keeps tool-call JSON parseable; router profiles may serve this qat build at 0.8
         topP: 0.95
         topK: 64
         minP: 0.0
         repeatPenalty: 1.05
-        maxTokens: 32768
     chatTemplate: "gemma-channel"
     gpuTiers:
       "32gb": "32 GB Vulkan llama.cpp profile with UD-Q4_K_XL weights, ctx=262144, q8 KV, flash attention, parallel=1, F16 gemma4v mmproj (vision), and MTP draft (spec-type draft-mtp, n_max 4) both active."
@@ -726,9 +972,6 @@
     contextWindow: 122880
     maxTokens: 32768
   quirks:
-    kvCache:
-      kQuant: q8_0
-      vQuant: q8_0
     sampling:
       thinking:
         temperature: 1.0
@@ -737,13 +980,11 @@
         minP: 0.0
         presencePenalty: 0.0
         repetitionPenalty: 1.0
-        maxTokens: 32768
       instruct:
         temperature: 0.7  # 0.7 keeps tool-call JSON parseable; gemma card's 1.0 is for chat
         topP: 0.95
         topK: 64
         minP: 0.0
-        maxTokens: 32768
     chatTemplate: "gemma-channel"
     measuredUnder:
       hardware: unknown
@@ -779,10 +1020,26 @@
     serving: "Distilled from Claude 4.6 Opus reasoning traces onto Google Gemma 4 31B; sampler matches the gemopus card values."
 
 - family: qwopus3.6-27b-v1-preview
+  # A bare `qwopus3.6` pattern used to sit here and swallowed every Qwopus3.6
+  # router id that was not spelled 27b. The mini router's `qwopus3.6-35b-moe`
+  # landed on this 27B distill and inherited its budget-tokens dial, which emits
+  # nothing on llama.cpp, so that model ran with a four-level dial that never
+  # reached the wire (issue #263). The matcher in knowledge-base.ts ranks by
+  # pattern length and ignores file order, so the only durable guard against the
+  # next `qwopus3.6-*` id landing here is not to declare the short pattern.
+  #
+  # The cost is real and is chosen deliberately, here and at the other four
+  # families that keep only their full spellings. An unlisted `qwopus3.6-*` id
+  # now matches nothing, and a model with no family gets whatever
+  # `inferThinkingMechanism` reads off the live probe: `none` with the dial
+  # pinned to off and every thinking field stripped when the server reports no
+  # reasoning, `on-off` when it reports some. Both are honest about knowing
+  # nothing. Inheriting a sibling's measured dial is not, and it is the failure
+  # mode that costs a session its reasoning silently. Adding the id is a
+  # one-line catalog edit; noticing that a dial reaches nothing takes a sweep.
   matchPatterns:
     - qwopus3.6-27b-v1-preview
     - qwopus3.6-27b
-    - qwopus3.6
   capabilities:
     chat: true
     tools: true
@@ -818,7 +1075,6 @@
         minP: 0.0
         presencePenalty: 0.0
         repetitionPenalty: 1.0
-        maxTokens: 32768
       instruct:
         temperature: 0.7
         topP: 0.80
@@ -826,7 +1082,6 @@
         minP: 0.0
         presencePenalty: 1.5
         repetitionPenalty: 1.0
-        maxTokens: 32768
     gpuTiers:
       "32gb": "Fits at f16 KV with parallel=1 and ctx=262144 thanks to the Q4_K_M weights (~16.5 GB)."
     runtimePreference:
@@ -850,21 +1105,31 @@
     - qwopus3.6-35b-a3b-coder-mtp
     - qwopus3.6-35b-a3b-coder
     - qwopus3.6-35b-a3b
+    # The mini router's short id for this same gguf. At 17 characters it outbids
+    # the 13-character `qwopus3.6-27b` of the preview family, whichever order
+    # the two entries load in.
+    - qwopus3.6-35b-moe
   capabilities:
     chat: true
     tools: true
     toolCallFormat: qwen
-    # Reasoning class: effort-levels, measured against an LM Studio host on
-    # 2026-08-08 over its OpenAI-compatible surface. The card
-    # frames the Coder-MTP line as thinking-off execution, but the wire
-    # disagrees: on "what is 17+25" this model spends 98 of 103 completion
-    # tokens on reasoning, and a wiki planning dispatch spent 89,501. It is an
-    # always-reasoning model unless told otherwise.
+    # Reasoning class: on-off, and which field carries the switch depends on the
+    # runtime. The card frames the Coder-MTP line as thinking-off execution, but
+    # the wire disagrees on both hosts measured: the model reasons unless it is
+    # told not to, or stays silent unless it is told to, and never ignores the
+    # control that runtime honors.
+    #
+    # llama.cpp, 2026-09-02: chat_template_kwargs.enable_thinking is the switch.
+    # true took a one-line arithmetic prompt from 3 completion tokens to 130
+    # with 338 characters of reasoning_content, while reasoning_effort none, low
+    # and high all left it at 3 tokens with no trace.
     #
-    # chat_template_kwargs enable_thinking is inert here (verified with unique
-    # prompts to rule out cache hits). reasoning_effort "none" is the control
-    # that works, taking the same prompt from 98 reasoning tokens to 0, which
-    # is why `off` is mapped explicitly below rather than left unsent.
+    # LM Studio, 2026-08-08: exactly the reverse. enable_thinking was inert
+    # (verified with unique prompts to rule out cache hits) and reasoning_effort
+    # "none" was the control that worked, taking a prompt from 98 reasoning
+    # tokens to 0. Clio's on-off mechanism sends the template flag on every
+    # runtime and adds LM Studio's reasoning_effort spelling on that one, so
+    # both surfaces get a dial that reaches the model.
     #
     # reasoning=false previously resolved to mechanism "none", and mechanism
     # "none" makes openai-completions strip reasoning_effort from the payload.
@@ -882,25 +1147,75 @@
     maxTokens: 32768
   quirks:
     measuredUnder:
-      hardware: unknown
-      runtime: "lmstudio and its OpenAI-compatible port for the reasoning findings; an unrecorded llama.cpp host for the presence-penalty finding"
-      build: "unknown for both dates; the Windows LM Studio measurement host's bundled build string was first recorded on 2026-08-29 as llama.cpp-win-x86_64-nvidia-cuda12-avx2, which is three weeks after the reasoning measurement and cannot be backdated to it"
-      llamaCpp: "unknown; the reasoning measurement went through LM Studio's HTTP port, which does not expose the underlying server argv, and no argv survives for the dispatch that produced the presence-penalty value"
-      date: "2026-08-08 (reasoning_effort and enable_thinking), 2026-07-06 (presence penalty)"
-      source: "commits b3db3b82 and f7d76bf4"
+      hardware: "mini fleet node, AMD Radeon AI PRO R9700 32 GB, Vulkan backend, 4 request slots, for the 2026-09-02 run; unknown for both earlier dates"
+      runtime: "llamacpp for the 2026-09-02 on-off measurement; lmstudio and its OpenAI-compatible port for the 2026-08-08 reasoning findings; an unrecorded llama.cpp host for the presence-penalty finding"
+      build: "worker llama-server from /home/akougkas/src/llama.cpp-b10687/build-vulkan/bin behind a router reporting build_info b1-c841aee, for the 2026-09-02 run. Unknown for both earlier dates; the Windows LM Studio measurement host's bundled build string was first recorded on 2026-08-29 as llama.cpp-win-x86_64-nvidia-cuda12-avx2, which is three weeks after the reasoning measurement and cannot be backdated to it"
+      model: "Qwopus3.6-35B-A3B-Coder-MTP Q4_K_M (arch qwen35moe)"
+      llamaCpp: >-
+        2026-09-02 only: --ctx-size 786432 --parallel 4 --no-kv-unified
+        --cache-type-k q8_0 --cache-type-v q8_0 --batch-size 16384 --ubatch-size
+        2048 --flash-attn true --fit off --n-gpu-layers 99 --jinja --reasoning
+        off --chat-template-kwargs {"enable_thinking": false} --temperature 0.6
+        --top-k 20 --top-p 0.95 --min-p 0.0 --presence-penalty 0.0
+        --repeat-last-n 256 --repeat-penalty 1.0. Unknown for both earlier
+        dates; the reasoning measurement went through LM Studio's HTTP port,
+        which does not expose the underlying server argv, and no argv survives
+        for the dispatch that produced the presence-penalty value
+      date: "2026-09-02 (llama.cpp on-off switch), 2026-08-08 (reasoning_effort and enable_thinking on LM Studio), 2026-07-06 (presence penalty)"
+      source: "issue #263 thinking sweep for the llama.cpp date; commits b3db3b82 and f7d76bf4 for the two earlier ones"
       note: |
-        Three numbers here are measurements. On 2026-08-08 on that host this
-        model spent 98 of 103 completion tokens reasoning on "what is 17+25"
-        with no thinking field set; chat_template_kwargs enable_thinking was
-        inert, re-checked with unique prompts to rule out prompt-cache hits;
-        and reasoning_effort "none" took the same prompt to 0. The same session
-        recorded a wiki planning dispatch over a 1007-file repository going
-        from 89,501 reasoning tokens, 47 tool calls and exit 1 at 459 s to 0
-        reasoning tokens, 8 tool calls and exit 0 at 218 s. On 2026-07-06 a
-        live dispatch measured presence_penalty 1.5: without it a coder worker
-        repeated one code_nav call into the loop-guard abort on 3 of 3 runs,
-        and with it the same task passed 3 of 3. Neither commit records the
-        hardware or the server configuration, so both stay unknown.
+        Measured 2026-09-02 on llama.cpp: with no thinking field, and with
+        chat_template_kwargs.enable_thinking false, the model answered a
+        one-line arithmetic prompt in 3 completion tokens with no
+        reasoning_content; with enable_thinking true the same prompt spent 130
+        tokens, 338 characters of them reasoning. reasoning_effort none, low and
+        high all left it at 3 tokens with no trace. The silent baseline is the
+        router's doing rather than the checkpoint's: the preset launches this
+        model with --reasoning off and chat_template_kwargs enable_thinking
+        false, which is the opposite default from the LM Studio host below.
+
+        Three numbers from the earlier dates are also measurements. On
+        2026-08-08 on that LM Studio host this model spent 98 of 103 completion
+        tokens reasoning on "what is 17+25" with no thinking field set;
+        chat_template_kwargs enable_thinking was inert, re-checked with unique
+        prompts to rule out prompt-cache hits; and reasoning_effort "none" took
+        the same prompt to 0. The same session recorded a wiki planning dispatch
+        over a 1007-file repository going from 89,501 reasoning tokens, 47 tool
+        calls and exit 1 at 459 s to 0 reasoning tokens, 8 tool calls and exit 0
+        at 218 s. On 2026-07-06 a live dispatch measured presence_penalty 1.5:
+        without it a coder worker repeated one code_nav call into the loop-guard
+        abort on 3 of 3 runs, and with it the same task passed 3 of 3. Neither
+        commit records the hardware or the server configuration, so both stay
+        unknown.
+
+        The 2026-09-02 server also differs from the profile recommended below.
+        It ran ctx-size 786432 with --parallel 4 and --no-kv-unified, giving each
+        slot 196608 tokens, and served text-only with no mmproj and no MTP draft
+        where the block below recommends 262144 at parallel 1 with both active.
+        The vision capability stays the checkpoint's; that run simply did not
+        load the projector. That has a consequence for any projector-less
+        deployment, because the model Clio builds takes vision from this entry
+        and from a target-level override only: synthLocalModel merges the
+        runtime defaults, this entry and target.capabilities and passes no probe
+        layer, then sets input to ["text","image"] whenever vision is true. Once
+        this id resolves here rather than to the 27B preview it did before issue
+        #263, mini's text-only preset would be offered image input it rejects,
+        so a target serving this checkpoint without --mmproj needs
+        capabilities.vision false on the target entry.
+
+        Why on-off replaced the four-level effort map, and what it costs. The
+        map gave LM Studio a low/medium/high granularity nothing had measured,
+        and gave llama.cpp an active level that put reasoning_effort on the wire
+        and nothing else, which this template does not read. on-off is what both
+        measurements support. The cost is that resolveRequestCapability adds the
+        reasoning_effort spelling for an on-off family only when runtimeId is
+        literally `lmstudio` (REASONING_EFFORT_ONLY_RUNTIMES), while an
+        effort-levels family emitted it on every runtime. An LM Studio server
+        reached through a generic openai-compat, litellm or vllm target
+        therefore now receives only the template flag, which that host ignores,
+        and the model reasons at every level. Point such a target at the
+        `lmstudio` runtime, which is what the 2026-08-08 finding above was taken
+        on and what carries the spelling that reaches it.
     sampling:
       instruct:
         temperature: 0.2
@@ -913,7 +1228,6 @@
         # code_nav call into the loop-guard abort on 3 of 3 runs; with it the
         # same task passed 3 of 3 with clean edit-then-validate trajectories.
         presencePenalty: 1.5
-        maxTokens: 32768
     gpuTiers:
       "32gb": "Default 32 GB local profile: Q4_K_M, ctx=262144, q8 KV, flash attention, parallel=1, with F32 mmproj (vision) and draft-mtp speculative decoding both active. Leaves several GiB free at 262k on a 32 GiB class card."
     runtimePreference:
@@ -933,21 +1247,13 @@
       chatTemplateKwargs:
         enable_thinking: false
     thinking:
-      mechanism: effort-levels
-      effortByLevel:
-        # "none" is the only value that silences this family. It is carried on
-        # the wire because the model reasons by default, so sending nothing is
-        # in effect a request to keep reasoning.
-        off: none
-        low: low
-        medium: medium
-        high: high
+      mechanism: on-off
       guidance: |
-        Qwopus3.6 35B-A3B Coder reasons unless reasoning_effort is set to
-        "none", despite the creator's card describing the Coder-MTP variants as
-        thinking-off. That dial only reaches the wire on openai-compat; the LM
-        Studio native transport has no field for it, so a native-routed run
-        reasons at full rate whatever the configured thinking level says.
+        One switch, two spellings: chat_template_kwargs.enable_thinking on
+        llama.cpp, reasoning_effort ("none" off, "low" on) on LM Studio, each
+        inert on the other host. Intermediate levels coerce to on, and off has
+        to be carried on the wire because the checkpoint reasons unless told
+        not to, whatever its card says about thinking-off execution.
     serving: "Jackrong Qwopus3.6 35B-A3B Coder MTP, Q4_K_M. Thinking-off coder/agent tuned for tool-use loops. Served as a local router default with MTP speculative decoding (spec-type draft-mtp, n_max 2) and F32 mmproj vision both active at 262k."
 
 - family: qwopus3.6-27b-coder
@@ -994,7 +1300,6 @@
         # Same Coder-MTP non-thinking design as the 35B; see the measured
         # presence-penalty note there.
         presencePenalty: 1.5
-        maxTokens: 32768
     runtimePreference:
       llamaCpp: "A local llama.cpp router serves it with --jinja and enable_thinking false; coder sampler temperature 0.2, top_p 0.9, top_k 20."
       lmstudio: "Dense 27B Coder-MTP; configure MTP in LM Studio's per-model load settings."
@@ -1044,7 +1349,6 @@
         topP: 0.9
         topK: 20
         repeatPenalty: 1.05
-        maxTokens: 32768
     runtimePreference:
       llamaCpp: "A local llama.cpp router serves it with --jinja and enable_thinking false; coder sampler temperature 0.2, top_p 0.9, top_k 20."
     thinking:
@@ -1062,6 +1366,13 @@
     - gemma4-26b-a4b-it
     - gemma-4-26b-a4b-it-q4
     - gemma-4-26b-a4b-it-q4_k_m
+    # The mini router's short id for the same checkpoint; its alias is
+    # Gemma-4-26B-A4B-it-Q4_K_M-262K and it loads
+    # gemma-4-26B-A4B-it-qat-UD-Q4_K_XL.gguf. QAT changes how the weights are
+    # represented, not the chat template, so the on-off mechanism below carries
+    # over and there is no reason to split the family the way the 31B nvfp4 and
+    # qat builds are split (those declare different context windows).
+    - gemma4-26b-moe
   capabilities:
     chat: true
     tools: true
@@ -1077,12 +1388,48 @@
     contextWindow: 262144
     maxTokens: 65536
   quirks:
+    measuredUnder:
+      hardware: "mini fleet node, AMD Radeon AI PRO R9700 32 GB, Vulkan backend, 1 request slot"
+      runtime: llamacpp
+      build: "worker llama-server from /home/akougkas/src/llama.cpp-b10687/build-vulkan/bin, behind a router reporting build_info b1-c841aee"
+      model: "Gemma-4-26B-A4B-it-qat UD-Q4_K_XL with mmproj-F16, no draft model (arch gemma4)"
+      llamaCpp: >-
+        --ctx-size 262144 --parallel 1 --kv-unified --cache-type-k f16
+        --cache-type-v f16 --batch-size 16384 --ubatch-size 2048 --flash-attn
+        true --n-gpu-layers 99 --jinja --reasoning off --temperature 0.8
+        --top-k 64 --top-p 0.95 --repeat-penalty 1.05
+      date: "2026-09-02"
+      source: "issue #263 thinking sweep: one non-streaming POST /v1/chat/completions per variant on a one-line arithmetic prompt, max_tokens 4096"
+      note: |
+        The on-off mechanism this entry already claimed is now measured on
+        llama.cpp rather than inherited from the gemma-4 template. With no
+        thinking field, and with chat_template_kwargs.enable_thinking false, the
+        model answered in 3 completion tokens with no reasoning_content; with it
+        true the same prompt spent 99 tokens, 227 characters of them reasoning.
+        reasoning_effort is inert: none, low, and high all stayed at 3 tokens
+        with no trace, so this family is on-off and not effort-levels on this
+        runtime. The router serves it with --reasoning off, which is why the
+        no-field baseline is silent.
+
+        Three serving facts differ from the recommended profile below. The
+        measuring server ran f16 KV where the block below recommends q8_0, and
+        it ran at batch 16384 and ubatch 2048. It also ran without speculative
+        decoding: the preset's --tags string advertises
+        spec:mtp,spec_draft_n:4,draft:mtp-gemma-4-26B-A4B-it, but the argv
+        carries no --model-draft, --spec-type or --spec-draft-n-max, so the
+        token counts above come from a plain decode. The 31B qat sibling above
+        does pass those flags, which is where the tags were copied from.
+
+        Separately, the round-5 cache sweep on this same date measured the
+        Gemma 4 tokenizer rendering Clio's full-capability first prompt at 7,353
+        tokens against roughly 8.4k for the Qwen tokenizer on the identical
+        prompt, so this family buys about a thousand tokens of headroom per turn
+        over the Qwen-tokenized families.
     sampling:
       thinking:
         temperature: 0.3
         topP: 0.9
         topK: 20
-        maxTokens: 8192
         reasoningBudget: 4096
       instruct:
         temperature: 0.7  # 0.7 keeps tool-call JSON parseable; gemma card's 1.0 is for chat
@@ -1153,7 +1500,6 @@
         temperature: 0.3
         topP: 0.9
         topK: 20
-        maxTokens: 8192
     gpuTiers:
       "32gb": "Preferred local worker model when a text-only worker is enough."
     runtimePreference:
@@ -1221,13 +1567,11 @@
         temperature: 0.6
         topP: 0.95
         topK: 20
-        maxTokens: 16384
         reasoningBudget: 4096
       instruct:
         temperature: 0.2
         topP: 0.9
         topK: 20
-        maxTokens: 4096
     gpuTiers:
       "32gb": "Distilled Opus-style reasoning at Qwen3.5 35B A3B scale. Use as a strong local main agent when GPU room is available."
     runtimePreference:
@@ -1288,12 +1632,10 @@
         temperature: 0.2
         topP: 0.9
         topK: 20
-        maxTokens: 4096
       thinking:
         temperature: 0.6
         topP: 0.95
         topK: 20
-        maxTokens: 8192
         reasoningBudget: 2048
     gpuTiers:
       "16gb": "Use this as the local 16GB emulation target. It keeps the output cap below the 32GB-class MoE defaults."
@@ -1316,3 +1658,246 @@
         32GB-class MoE families because this is the 16GB emulation target;
         budget for a reasoning preamble ahead of every answer.
     serving: "LM Studio reports trained_for_tool_use for the local qwopus3.5-9b-v3 target. Configure residency through LM Studio or explicit lmstudio.load settings."
+
+# Meta's Muse Glimmer is a new vendor, architecture and template for this
+# catalog: no existing family shares its perception encoder, its DFlash drafter,
+# or its reasoning control. It is listed under always-on because its dial is not
+# representable, not because the model reasons unconditionally by design; the
+# measuredUnder note below records what was tried.
+- family: muse-glimmer-30b
+  matchPatterns:
+    # No bare `muse-glimmer`: it would swallow a future Muse Glimmer of another
+    # size, which shares neither this 131072 window nor a measured dial.
+    - unsloth/muse-glimmer-30b-gguf
+    - muse-glimmer-30b
+    # The mini router serves the same weights under both this short id and the
+    # unsloth repo name above.
+    - muse-30b-dense
+  capabilities:
+    chat: true
+    tools: true
+    toolCallFormat: openai
+    reasoning: true
+    thinkingFormat: qwen-chat-template
+    structuredOutputs: json-schema
+    vision: true
+    audio: false
+    embeddings: false
+    rerank: false
+    fim: false
+    contextWindow: 131072
+    maxTokens: 32768
+  quirks:
+    measuredUnder:
+      hardware: "mini fleet node, AMD Radeon AI PRO R9700 32 GB, ROCm backend, 1 request slot"
+      runtime: llamacpp
+      build: "worker llama-server from /home/akougkas/src/llama.cpp-b10687/build-rocm/bin, behind a router reporting build_info b1-c841aee"
+      model: "Muse-Glimmer-30B UD-Q4_K_XL with dflash-kquant draft and mmproj-Muse-Glimmer-30B-Q8_0 (arch muse-glimmer)"
+      llamaCpp: >-
+        --ctx-size 131072 --parallel 1 --kv-unified --cache-type-k q8_0
+        --cache-type-v q8_0 --batch-size 2048 --ubatch-size 512 --flash-attn
+        true --fit off --n-gpu-layers 99 --spec-type draft-dflash
+        --spec-draft-n-max 16 --n-gpu-layers-draft 99 --jinja
+        --chat-template-kwargs {"reasoning_strength": "xhigh"} --temperature 1.0
+        --top-k 64 --top-p 0.95
+      date: "2026-09-02"
+      source: "issue #263 thinking sweep: one non-streaming POST /v1/chat/completions per variant on a one-line arithmetic prompt, max_tokens 4096"
+      note: |
+        always-on is the measured classification, and it is measured against
+        both controls Clio can send. With no thinking field the model spent 99
+        completion tokens with 358 characters of reasoning_content;
+        chat_template_kwargs.enable_thinking false still produced 91 tokens with
+        332 characters, and true produced 90 with 312; reasoning_effort none,
+        low, and high produced 81, 83, and 61 tokens, every one of them with a
+        trace. Nothing Clio can put on the wire silences it.
+
+        The reason is the model's own design rather than a template bug. The
+        card's dial is reasoning strength with levels low, medium, high, and
+        xhigh, set through the system prompt as "Reasoning strength: <value>",
+        and the router passes it as chat_template_kwargs.reasoning_strength
+        pinned to xhigh. There is no off level on that scale, and no
+        ThinkingMechanism in this engine emits that key:
+        resolveRequestCapability populates chat_template_kwargs only with
+        enable_thinking or, on the harmony path, reasoning_effort. Encoding this
+        family as on-off would put a flag on the wire that this template does
+        not read while the TUI showed a working off/on dial, which is the
+        failure qwopus3.5-9b-v3 above was reclassified to stop. A real dial for
+        this family needs a per-family kwarg key in the engine, not a catalog
+        edit.
+
+        Sampler provenance: temperature 1.0, top_p 0.95 and top_k 64 are the
+        card's Best Practices values and the router preset uses the same three.
+        The preset sets no repeat penalty and the card recommends none, so this
+        entry declares none. contextWindow is the card's 131072, which the
+        router matches at ctx-size 131072. maxTokens has no card value and
+        follows the other 131072-context families here.
+    sampling:
+      thinking:
+        temperature: 1.0
+        topP: 0.95
+        topK: 64
+      instruct:
+        temperature: 1.0
+        topP: 0.95
+        topK: 64
+    gpuTiers:
+      "32gb": "Agent-role target on a 32 GB card: UD-Q4_K_XL weights, q8_0 KV at ctx=131072, parallel=1, with the Q8_0 perception encoder and the DFlash drafter both resident."
+    runtimePreference:
+      llamaCpp: "OpenAI-compatible chat completions with --jinja. Set the reasoning budget at launch with --chat-template-kwargs {\"reasoning_strength\": \"...\"}; there is no per-request field Clio can send for it."
+      openaiCompat: "Use when a gateway fronts the same llama.cpp chat-completions surface. Expect a reasoning trace on every response whatever the configured thinking level says."
+    llamaCpp:
+      ctxSize: 131072
+      cacheTypeK: q8_0
+      cacheTypeV: q8_0
+      flashAttn: true
+      nGpuLayers: 99
+      parallel: 1
+      parallelSlots: "Start the server with --parallel 1 so one agent gets the whole 131072-token context. Clio discovers total_slots 1 and admits one request at a time on this endpoint, so a dispatch raised while the orchestrator streams is refused rather than queued."
+      batchSize: 2048
+      ubatchSize: 512
+      mmproj: true
+      specType: draft-dflash
+      specDraftNMax: 16
+    thinking:
+      mechanism: always-on
+      chatTemplateKwargs:
+        byLevel:
+          # Card: reasoning strength low|medium|high|xhigh (scratchpad/catalog/models/unsloth-Muse-Glimmer-30B.md:35). The per-family kwarg key the note above asked for.
+          key: reasoning_strength
+          values:
+            low: low
+            medium: medium
+            high: high
+            xhigh: xhigh
+          lmstudio: unsupported
+        lmstudio: unsupported
+      guidance: |
+        Muse Glimmer reasons on every turn and its strength scale has no off
+        level, so Clio shows the dial as forced and budgets for a reasoning
+        preamble ahead of every answer. Change the amount of thinking at the
+        server, with the reasoning_strength chat-template kwarg; a per-request
+        thinking level cannot reach this model.
+    serving: "Meta Superintelligence Lab Muse-Glimmer-30B, a dense 29.6B causal transformer with a ~1.8B ViT-G/14 perception encoder, interleaved local and global attention with a 2048 sliding window, and a DFlash block-diffusion drafter that proposes 16 tokens per forward pass. Card context length is 131072. Input is text plus image; the mini router additionally tags video, which the card does not claim, so the catalog follows the card."
+
+# ThinkingCap is a brevity finetune of Qwen3.6-27B and not the base model, so it
+# keeps its own family rather than borrowing qwen3.6-27b's. No pattern here may
+# be a bare `qwen3.6` spelling: `lookup` is a substring test, so `qwen3.6` alone
+# would match every base-model id too and win or lose on length rather than on
+# which checkpoint the id names. The `thinkingcap-` prefixed spellings below all
+# contain `qwen3.6` harmlessly, because a longer pattern only matches ids that
+# contain the whole longer string.
+- family: thinkingcap-qwen3.6-27b
+  matchPatterns:
+    # No bare `thinkingcap`: it would swallow a ThinkingCap built on another
+    # base, such as a Qwen3.8 32B, and hand it this q4's sampler, its 131072
+    # window and a dial measured on a different checkpoint.
+    - thinkingcap-qwen3.6-27b
+    - thinkingcap-27b-dense-q4
+    - thinkingcap-27b-dense
+    - thinkingcap-27b
+  capabilities:
+    chat: true
+    tools: true
+    toolCallFormat: qwen
+    reasoning: true
+    thinkingFormat: qwen-chat-template
+    structuredOutputs: json-schema
+    vision: false
+    audio: false
+    embeddings: false
+    rerank: false
+    fim: false
+    contextWindow: 131072
+    maxTokens: 32768
+  quirks:
+    measuredUnder:
+      hardware: "mini fleet node, AMD Radeon AI PRO R9700 32 GB, ROCm backend, 1 request slot"
+      runtime: llamacpp
+      build: "worker llama-server from /home/akougkas/src/llama.cpp-b10687/build-rocm/bin, behind a router reporting build_info b1-c841aee"
+      model: "ThinkingCap-27B Q4_K_M with the MTP draft head baked into the gguf (arch qwen36)"
+      llamaCpp: >-
+        --ctx-size 131072 --parallel 1 --kv-unified --cache-type-k f16
+        --cache-type-v f16 --batch-size 2048 --ubatch-size 512 --flash-attn true
+        --fit off --n-gpu-layers 99 --spec-type draft-mtp --spec-draft-n-max 4
+        --n-gpu-layers-draft 99 --jinja --reasoning on --chat-template-kwargs
+        {"enable_thinking": true} --temperature 1.0 --top-k 20 --top-p 0.95
+        --min-p 0.0 --repeat-penalty 1.0
+      date: "2026-09-02"
+      source: "issue #263 thinking sweep: one non-streaming POST /v1/chat/completions per variant on a one-line arithmetic prompt, max_tokens 4096"
+      note: |
+        The router tag is wrong for this model and the measurement wins. It is
+        the only reasoning model on this router tagged reasoning:on rather than
+        reasoning:on_off, and it carries no default_reasoning tag at all, which
+        reads as always-on. It is not. On the q4 that this family is keyed to,
+        the no-field baseline spent 125 completion tokens with 354 characters of
+        reasoning_content, chat_template_kwargs.enable_thinking false answered
+        in 3 tokens with none, and true spent 130 tokens with 388 characters.
+        reasoning_effort is inert: none, low, and high produced 115, 140, and
+        131 tokens with a trace every time. The default-on behavior comes from
+        the preset rather than the checkpoint, because the
+        [thinkingcap-27b-dense-q4] block launches the server with reasoning = on
+        and chat-template-kwargs {"enable_thinking": true}.
+
+        Issue #263 names thinkingcap-27b-dense-q6, which no longer exists. That
+        quant was measured on this same date (baseline 99 tokens with 246
+        characters of reasoning, enable_thinking false 3 tokens with none, true
+        98 tokens with 291 characters, the same on-off shape as the q4) and then
+        deleted from mini: the weights are gone and the preset block is removed
+        from the router, which now serves the q4 as the only ThinkingCap id.
+        This family therefore names no q6 quant. The short `thinkingcap-27b`
+        pattern still covers that spelling if the operator ever restores it.
+
+        One serving observation was taken on the q6 before it was removed and
+        has not been repeated on the q4: in the round-5 cache sweep this model
+        re-prefilled 661 tokens on a `--continue` turn where the other eight
+        router models kept their prefix. Read it as a lead rather than a fact
+        about the q4.
+
+        Sampler provenance: temperature 0.6, top_p 0.95 and top_k 20 are the
+        quantizer's stated intended sampling for this finetune, which also warns
+        that greedy decoding makes it loop without closing the think block. The
+        router preset instead runs temperature 1.0 with the same top_p and
+        top_k, plus min_p 0.0 and repeat_penalty 1.0. The catalog follows the
+        card, as it does for every other family here, and records the preset
+        divergence above.
+    sampling:
+      thinking:
+        temperature: 0.6
+        topP: 0.95
+        topK: 20
+        minP: 0.0
+        repeatPenalty: 1.0
+      instruct:
+        temperature: 0.6
+        topP: 0.95
+        topK: 20
+        minP: 0.0
+        repeatPenalty: 1.0
+    gpuTiers:
+      "32gb": "Reasoning-role target on a 32 GB card: Q4_K_M weights (about 17 GB) with f16 KV at ctx=131072, parallel=1, and the baked-in MTP draft head active."
+    runtimePreference:
+      llamaCpp: "OpenAI-compatible chat completions with --jinja and --reasoning on. The MTP head ships inside the gguf, so --spec-type draft-mtp needs no second model. Never decode greedy: temperature 0 makes this finetune loop without closing its think block."
+      openaiCompat: "Use when a gateway fronts the same llama.cpp chat-completions surface with qwen reasoning and tool parsing."
+    llamaCpp:
+      ctxSize: 131072
+      cacheTypeK: f16
+      cacheTypeV: f16
+      flashAttn: true
+      nGpuLayers: 99
+      parallel: 1
+      parallelSlots: "Start the server with --parallel 1 so one agent gets the whole 131072-token context. Clio discovers total_slots 1 and admits one request at a time on this endpoint, so a dispatch raised while the orchestrator streams is refused rather than queued."
+      batchSize: 2048
+      ubatchSize: 512
+      specType: draft-mtp
+      specDraftNMax: 4
+      chatTemplateKwargs:
+        enable_thinking: true
+    thinking:
+      mechanism: on-off
+      guidance: |
+        ThinkingCap keeps the Qwen3.6 chat template's enable_thinking switch;
+        intermediate levels coerce to on. The router serves it with thinking on
+        by default, so off has to be carried on the wire rather than omitted,
+        and reasoning_effort is inert on llama.cpp. The finetune's point is a
+        shorter chain rather than no chain, so an active level is cheap.
+    serving: "BottleCapAI ThinkingCap-Qwen3.6-27B, a brevity finetune of Qwen3.6-27B that holds accuracy at roughly 40 to 46 percent fewer thinking tokens (the quantizer measured a 15-prompt reasoning set at mean 675 tokens against 1401 for base Qwen3.6-27B). Served on mini as the Q4_K_M gguf with the MTP draft head embedded in the file, so speculative decoding needs no second model."