@miller-tech/uap 1.115.0 → 1.118.0

package/config/model-profiles/claude-fable-5.json

@@ -0,0 +1,78 @@
+{
+  "_profile": "claude-fable-5",
+  "_description": "Anthropic Claude Fable 5 (claude-fable-5). Anthropic's most capable widely released model — the most demanding reasoning and long-horizon agentic work. 1M context (default), 128K max output. Thinking is ALWAYS on (omit the thinking parameter); the raw chain of thought is never returned. temperature/top_p/top_k and budget_tokens are removed. Requires 30-day data retention (not available under ZDR). Use only for the hardest tasks — pricing exceeds Opus tier.",
+
+  "model": "claude-fable-5",
+  "provider": "anthropic",
+  "api_base_url": "https://api.anthropic.com",
+  "max_tokens": 64000,
+  "timeout_ms": 900000,
+  "context_window": 1000000,
+
+  "sampling_parameters_supported": false,
+  "_sampling_comment": "temperature, top_p, and top_k are removed on Fable 5 and return HTTP 400. Steer with prompting and effort.",
+
+  "optimize_for_tool_calls": true,
+  "enable_thinking": true,
+
+  "reasoning": {
+    "mode": "always_on",
+    "effort": "high",
+    "display": "summarized",
+    "_comment": "Fable 5 thinking is always on — do NOT send a thinking config (an explicit {type:'disabled'} 400s; omit the param). Control depth with effort (low..max). Raw chain of thought is never returned; display=summarized surfaces a readable summary. Long turns are normal — plan timeouts/streaming."
+  },
+
+  "extended_thinking": {
+    "enabled": true,
+    "mode": "always_on",
+    "_comment": "Always-on adaptive thinking; no budget_tokens (removed). Do not pass a thinking parameter on the wire."
+  },
+
+  "tool_calling": {
+    "parallel_tool_calls": true,
+    "tool_choice": "auto",
+    "max_tool_calls_per_turn": 128,
+    "_comment": "Excellent parallel tool use and long-lived sub-agent coordination. No forced tool_choice needed."
+  },
+
+  "tool_call_batching": {
+    "enabled": true,
+    "system_prompt_suffix": "When multiple tools are needed, call ALL of them in a single response. Do not call one tool and wait - emit all tool calls together.",
+    "_comment": "Fable 5 natively supports parallel tool calls and asynchronous sub-agent delegation."
+  },
+
+  "refusal_fallback": {
+    "enabled": true,
+    "beta": "server-side-fallback-2026-06-01",
+    "fallback_model": "claude-opus-4-8",
+    "_comment": "Safety classifiers may return stop_reason=refusal (HTTP 200). Opt into server-side fallbacks so a benign-but-declined request is transparently re-served by Opus 4.8 in the same call."
+  },
+
+  "requirements": {
+    "data_retention_days": 30,
+    "_comment": "Fable 5 is unavailable under zero data retention — requests from a ZDR org 400 on every call."
+  },
+
+  "pricing": {
+    "input_per_1m": 10.0,
+    "output_per_1m": 50.0,
+    "cache_write_per_1m": 12.5,
+    "cache_read_per_1m": 1.0,
+    "currency": "USD",
+    "_comment": "Above Opus tier — reserve for the hardest reasoning/agentic work."
+  },
+
+  "rate_limits": {
+    "requests_per_minute": 500,
+    "tokens_per_minute": 60000,
+    "_comment": "Default tier limits. Higher tiers available via Anthropic."
+  },
+
+  "running_config": {
+    "description": "Anthropic hosted API. Requires ANTHROPIC_API_KEY and an org with >= 30-day data retention.",
+    "env_vars": {
+      "ANTHROPIC_API_KEY": "<your-api-key>",
+      "UAP_MODEL_PROFILE": "claude-fable-5"
+    }
+  }
+}

package/config/model-profiles/claude-haiku-4.5.json

@@ -0,0 +1,69 @@
+{
+  "_profile": "claude-haiku-4.5",
+  "_description": "Anthropic Claude Haiku 4.5 (claude-haiku-4-5-20251001). Fastest and most cost-effective current Claude model. 200K context, 64K max output, strong tool calling for high-throughput agentic work. Supports adaptive thinking; the effort parameter is NOT supported (it errors). Best for rapid iteration, simple-to-medium tool use, classification, and cost-sensitive workloads.",
+
+  "model": "claude-haiku-4-5-20251001",
+  "provider": "anthropic",
+  "api_base_url": "https://api.anthropic.com",
+  "max_tokens": 16384,
+  "temperature": 1.0,
+  "top_p": 1.0,
+  "timeout_ms": 120000,
+  "context_window": 200000,
+
+  "optimize_for_tool_calls": true,
+  "enable_thinking": true,
+
+  "reasoning": {
+    "mode": "adaptive",
+    "effort_supported": false,
+    "_comment": "Haiku 4.5 supports adaptive thinking but NOT the effort parameter (effort=* returns 400). Do not send output_config.effort. budget_tokens is not used on 4.5."
+  },
+
+  "extended_thinking": {
+    "enabled": true,
+    "mode": "adaptive",
+    "_comment": "Adaptive thinking supported. No effort control on Haiku 4.5."
+  },
+
+  "dynamic_temperature": {
+    "enabled": false,
+    "_comment": "Default temp=1.0 is recommended by Anthropic for most use cases."
+  },
+
+  "tool_calling": {
+    "parallel_tool_calls": true,
+    "tool_choice": "auto",
+    "max_tool_calls_per_turn": 128,
+    "_comment": "Strong tool calling for a fast/cheap model; good for high-throughput fan-out."
+  },
+
+  "tool_call_batching": {
+    "enabled": true,
+    "system_prompt_suffix": "When multiple tools are needed, call ALL of them in a single response. Do not call one tool and wait - emit all tool calls together.",
+    "_comment": "Haiku 4.5 supports parallel tool calls natively."
+  },
+
+  "pricing": {
+    "input_per_1m": 1.0,
+    "output_per_1m": 5.0,
+    "cache_write_per_1m": 1.25,
+    "cache_read_per_1m": 0.1,
+    "currency": "USD",
+    "_comment": "Cheapest current Claude. Ideal for cost-sensitive, latency-sensitive workloads."
+  },
+
+  "rate_limits": {
+    "requests_per_minute": 4000,
+    "tokens_per_minute": 400000,
+    "_comment": "Default tier limits. Separate rate-limit pool from Haiku 3.x."
+  },
+
+  "running_config": {
+    "description": "Anthropic hosted API. Requires ANTHROPIC_API_KEY environment variable.",
+    "env_vars": {
+      "ANTHROPIC_API_KEY": "<your-api-key>",
+      "UAP_MODEL_PROFILE": "claude-haiku-4.5"
+    }
+  }
+}

package/config/model-profiles/claude-opus-4.8.json

@@ -0,0 +1,72 @@
+{
+  "_profile": "claude-opus-4.8",
+  "_description": "Anthropic Claude Opus 4.8 (claude-opus-4-8). Most capable Opus-tier model — state-of-the-art long-horizon agentic execution, knowledge work, and memory. 1M context, 128K max output. Adaptive thinking + effort control (temperature/top_p/top_k and budget_tokens are removed on this model). Best for complex multi-step reasoning and autonomous coding where accuracy matters more than speed.",
+
+  "model": "claude-opus-4-8",
+  "provider": "anthropic",
+  "api_base_url": "https://api.anthropic.com",
+  "max_tokens": 64000,
+  "timeout_ms": 600000,
+  "context_window": 1000000,
+
+  "sampling_parameters_supported": false,
+  "_sampling_comment": "temperature, top_p, and top_k are removed on Opus 4.8 and return HTTP 400. Steer behaviour with prompting and the effort control instead.",
+
+  "optimize_for_tool_calls": true,
+  "enable_thinking": true,
+
+  "reasoning": {
+    "mode": "adaptive",
+    "effort": "xhigh",
+    "display": "summarized",
+    "_comment": "Opus 4.8 uses adaptive thinking only; budget_tokens is removed (400). effort=xhigh is the sweet spot for coding/agentic work; drop to low/medium for routine tasks. Streaming is required for max_tokens above ~16K."
+  },
+
+  "extended_thinking": {
+    "enabled": true,
+    "mode": "adaptive",
+    "_comment": "Adaptive thinking — no fixed budget_tokens (removed on Opus 4.8). Depth is controlled by reasoning.effort."
+  },
+
+  "tool_calling": {
+    "parallel_tool_calls": true,
+    "tool_choice": "auto",
+    "max_tool_calls_per_turn": 128,
+    "_comment": "Native parallel tool use with strong self-regulation. No forced tool_choice needed — Opus 4.8 uses tools reliably on its own."
+  },
+
+  "tool_call_batching": {
+    "enabled": true,
+    "system_prompt_suffix": "When multiple tools are needed, call ALL of them in a single response. Do not call one tool and wait - emit all tool calls together.",
+    "_comment": "Opus 4.8 natively supports parallel tool calls."
+  },
+
+  "fast_mode": {
+    "supported": true,
+    "beta": "fast-mode-2026-02-01",
+    "_comment": "Opus 4.8 is the durable fast-capable tier (speed=fast on client.beta.messages). Premium pricing; not available on Batch/Priority/third-party platforms."
+  },
+
+  "pricing": {
+    "input_per_1m": 5.0,
+    "output_per_1m": 25.0,
+    "cache_write_per_1m": 6.25,
+    "cache_read_per_1m": 0.5,
+    "currency": "USD",
+    "_comment": "1M context at standard pricing — no long-context premium. Prompt caching gives ~90% discount on cache reads."
+  },
+
+  "rate_limits": {
+    "requests_per_minute": 1000,
+    "tokens_per_minute": 80000,
+    "_comment": "Default tier limits. Higher tiers available via Anthropic."
+  },
+
+  "running_config": {
+    "description": "Anthropic hosted API. Requires ANTHROPIC_API_KEY environment variable (or an `ant auth login` profile / Claude Max OAuth via the proxy).",
+    "env_vars": {
+      "ANTHROPIC_API_KEY": "<your-api-key>",
+      "UAP_MODEL_PROFILE": "claude-opus-4.8"
+    }
+  }
+}

package/config/model-profiles/claude-sonnet-5.json

@@ -1,28 +1,66 @@
 {
-  "name": "Claude Sonnet 5",
-  "model": "claude-sonnet-5-20250514",
-  "max_tokens": 8192,
-  "temperature": 0.5,
-  "top_p": 0.9,
-  "top_k": 40,
-  "context_window": 200000,
-  "stop_sequences": [],
-  "presence_penalty": 0.0,
-  "frequency_penalty": 0.0,
+  "_profile": "claude-sonnet-5",
+  "_description": "Anthropic Claude Sonnet 5 (claude-sonnet-5). Best combination of speed and intelligence in the Sonnet tier — near-Opus quality on coding and agentic work at Sonnet cost. 1M context, 128K max output. Adaptive thinking on by default; non-default temperature/top_p/top_k are rejected; budget_tokens is removed. New tokenizer (~30% more tokens than Sonnet 4.6). Best value for everyday agentic work.",
+
+  "model": "claude-sonnet-5",
+  "provider": "anthropic",
+  "api_base_url": "https://api.anthropic.com",
+  "max_tokens": 32000,
+  "timeout_ms": 300000,
+  "context_window": 1000000,
+
+  "sampling_parameters_supported": false,
+  "_sampling_comment": "Non-default temperature/top_p/top_k are rejected on Sonnet 5 (400). Omit them; steer with prompting and effort.",
+
+  "optimize_for_tool_calls": true,
+  "enable_thinking": true,
+
   "reasoning": {
+    "mode": "adaptive",
+    "effort": "high",
+    "display": "summarized",
+    "_comment": "Adaptive thinking is on by default (omitting thinking runs adaptive). effort=high is the default; raise to xhigh for the hardest coding/agentic tasks. budget_tokens is removed. New tokenizer uses ~30% more tokens than Sonnet 4.6 — size max_tokens with headroom."
+  },
+
+  "extended_thinking": {
     "enabled": true,
-    "effort": "medium",
-    "budget": 16000
+    "mode": "adaptive",
+    "_comment": "Adaptive thinking — no fixed budget_tokens (removed). Depth is controlled by reasoning.effort (low..max, including xhigh)."
+  },
+
+  "tool_calling": {
+    "parallel_tool_calls": true,
+    "tool_choice": "auto",
+    "max_tool_calls_per_turn": 128,
+    "_comment": "More agentic than Sonnet 4.6 by default — reaches for tools and self-verification readily. No forced tool_choice needed."
   },
-  "tools": {
+
+  "tool_call_batching": {
     "enabled": true,
-    "max_concurrent": 10,
-    "auto_invoke": true
+    "system_prompt_suffix": "When multiple tools are needed, call ALL of them in a single response. Do not call one tool and wait - emit all tool calls together.",
+    "_comment": "Sonnet 5 natively supports parallel tool calls."
+  },
+
+  "pricing": {
+    "input_per_1m": 3.0,
+    "output_per_1m": 15.0,
+    "cache_write_per_1m": 3.75,
+    "cache_read_per_1m": 0.3,
+    "currency": "USD",
+    "_comment": "Introductory pricing $2/$10 per 1M through 2026-08-31, then $3/$15. Near-Opus quality at Sonnet cost."
+  },
+
+  "rate_limits": {
+    "requests_per_minute": 2000,
+    "tokens_per_minute": 160000,
+    "_comment": "Default tier limits. Higher throughput than Opus."
   },
-  "guardrails": {
-    "loop_detection": true,
-    "tool_narrowing": true,
-    "cycle_breaking": true,
-    "context_pruning": true
+
+  "running_config": {
+    "description": "Anthropic hosted API. Requires ANTHROPIC_API_KEY environment variable (or an `ant auth login` profile / Claude Max OAuth via the proxy).",
+    "env_vars": {
+      "ANTHROPIC_API_KEY": "<your-api-key>",
+      "UAP_MODEL_PROFILE": "claude-sonnet-5"
+    }
   }
 }

package/config/model-profiles/qwen36.json

@@ -0,0 +1,102 @@
+{
+  "_profile": "qwen36",
+  "_description": "Qwen3.6 35B A3B (IQ4_XS) — local llama.cpp deployment for agentic tool-calling workloads. Successor to qwen35; the default local executor in UAP routing. Routes through the anthropic-proxy (:4000) for <think>-stripping + tool/finalize guardrails.",
+  "model": "qwen36-35b-a3b-iq4xs",
+  "max_tokens": 81920,
+  "temperature": 0.3,
+  "top_p": 0.9,
+  "top_k": 20,
+  "min_p": 0.05,
+  "repetition_penalty": 1.0,
+  "stop_sequences": ["<|im_end|>"],
+  "timeout_ms": 300000,
+  "context_window": 262144,
+  "optimize_for_tool_calls": true,
+  "mode_switch_buffer_tokens": 500,
+  "enable_thinking": true,
+  "chat_template_kwargs": {
+    "enable_thinking": true
+  },
+
+  "dynamic_temperature": {
+    "enabled": true,
+    "decay": 0.5,
+    "floor": 0.2,
+    "_comment": "On retry, temp = max(floor, base_temp * decay^attempt). Reduces randomness for structured output."
+  },
+
+  "tool_call_batching": {
+    "enabled": true,
+    "system_prompt_suffix": "When multiple tools are needed, call ALL of them in a single response. Do not call one tool and wait - emit all tool calls together.",
+    "_comment": "Reduces round-trips by encouraging multi-tool emission in one turn."
+  },
+
+  "structured_output": {
+    "grammar_file": "tools/agents/config/tool-call.gbnf",
+    "json_schema_file": "tools/agents/config/tool-call-schema.json",
+    "_comment": "GBNF grammar for per-request use only. Do NOT use --grammar-file on server startup — it forces tool_call output on all requests including plain chat. llama.cpp's autoparser handles tool call detection lazily via the chat template."
+  },
+
+  "server_optimization": {
+    "flash_attention": true,
+    "kv_cache_type_k": "q8_0",
+    "kv_cache_type_v": "q4_0",
+    "prompt_cache": {
+      "enabled": true,
+      "slot_save_path": "./cache/slots"
+    },
+    "speculative_decoding": {
+      "enabled": true,
+      "type": "ngram-mod",
+      "ngram_size_n": 3,
+      "draft_max": 3,
+      "draft_min": 1,
+      "draft_p_min": 0.80,
+      "_comment": "ngram-mod self-speculation (no draft model). Draft-model spec crashes on the Qwen3.x hybrid memory (GDN) seq_add path; ngram-mod avoids this and gives stable throughput."
+    },
+    "_comment": "KV q8/q4 split saves ~60% KV VRAM. Flash attn gives 1.5-2x speed on long context."
+  },
+
+  "routing": {
+    "endpoint": "http://192.168.1.165:4000/v1",
+    "_comment": "Route through the anthropic-proxy (:4000), NOT llama :8080 directly — the proxy strips <think> blocks and applies the tool/finalize guardrails."
+  },
+
+  "quantization_profiles": {
+    "current": "iq4_xs",
+    "upgrade_path": [
+      { "quant": "IQ4_XS", "vram_gb": 17, "accuracy": "96%", "tool_calls": "94%" },
+      { "quant": "Q4_K_M", "vram_gb": 20, "accuracy": "95%", "tool_calls": "95%" },
+      { "quant": "Q5_K_M", "vram_gb": 24, "accuracy": "97%", "tool_calls": "97%" },
+      { "quant": "Q6_K", "vram_gb": 28, "accuracy": "98%", "tool_calls": "98%" }
+    ],
+    "_comment": "Each step up requires ~4-7GB more VRAM but improves instruction following."
+  },
+
+  "running_config": {
+    "description": "Dual-model llama.cpp setup: Qwen3.6 35B A3B for LLM + nomic-embed-text-v2-moe for embeddings. Run as two separate llama-server instances on different ports; the LLM is fronted by the anthropic-proxy on :4000.",
+    "llm_server": {
+      "description": "Instance 1: Main LLM (Qwen3.6 35B A3B) — port 8080, fronted by the proxy on :4000",
+      "flags": [
+        "--model /path/to/Qwen3.6-35B-A3B-UD-IQ4_XS.gguf",
+        "--chat-template-file chat_template.jinja",
+        "--n-predict 16384",
+        "--temp 0.3",
+        "--top-p 0.9",
+        "--top-k 20",
+        "--min-p 0.05",
+        "--ctx-size 131072",
+        "--gpu-layers 99",
+        "--cache-type-k q8_0",
+        "--cache-type-v q4_0",
+        "--mlock",
+        "-fa on",
+        "--metrics",
+        "--host 0.0.0.0",
+        "--port 8080"
+      ],
+      "vram_estimate": "~20GB (17GB model + KV cache)"
+    },
+    "env_var": "UAP_EMBEDDING_ENDPOINT=http://192.168.1.165:8081"
+  }
+}