@bendyline/gilde 0.1.18 → 0.1.20

package/README.md

@@ -16,7 +16,7 @@ against their inputs; repository tooling and CI files are not published.
 | Category | Path | What it is |
 | --- | --- | --- |
 | Craftbooks | data/craftbook-templates/ | Step-by-step task playbooks a gezel can follow, each with an eval sidecar (`test.json`) |
-| Chat models | data/chat-models/ | Local model manifests: engine sources (llama.cpp / MLX), pinned revisions and sha256s, tuning defaults |
+| Chat models | data/chat-models/ | Local model manifests: engine sources (llama.cpp / MLX / ds4), pinned revisions and sha256s, tuning defaults |
 | Roles | data/gezel-templates/ | Gezel role templates — the `about.md` prose that gives a gezel its character |
 | Project types | data/project-types/ | Ready-made project scaffolds (mission, about, pages, embedded craftbooks) |
 | Image models | data/image-models/ | Image generation model manifests |
@@ -36,6 +36,11 @@ data/<category>/index.json                          generated listing (do not ha
 
 where `<shard>` is the first two characters of the id.
 
+Chat-model introductions are Gilde-only: add an authoring recipe under
+`authoring/chat-models/`, run `npm run build-chat-model`, and commit the new
+identity, version payload, and regenerated index. See
+`authoring/chat-models/README.md` for the recipe contract.
+
 ## Contributing
 
 New craftbooks, model updates, roles — contributions are welcome, from

package/authoring/chat-models/README.md

@@ -0,0 +1,26 @@
+# Chat-model authoring recipes
+
+Each JSON file in this directory is the complete source recipe for one local
+chat model. Recipes combine stable catalog metadata with provider source
+pointers; `tools/build-chat-model.mjs` resolves Hugging Face revisions,
+checksums, file sizes, and MLX file lists and emits the released identity and
+version manifests under `data/chat-models/`.
+
+To introduce a model, add `<id>.json` here and run:
+
+```sh
+npm run build-chat-model -- --config authoring/chat-models/<id>.json --release
+npm run fix
+npm run check
+```
+
+At least one of `ollama`, `llamaCpp`, `mlx`, or `ds4` is required. The filename
+must match `id`. The generator will not overwrite an existing version: bump
+`version` and `updatedAt` for every release. A normal rebuild preserves tuning,
+behaviors, eval hints, and provider revision pins already evolved in the
+identity; `--reseed` deliberately makes the recipe authoritative for those
+editorial fields.
+
+The runtime never reads this directory. It consumes only the validated,
+generated manifests in `data/chat-models/`, so a published Gilde patch can add
+a model without a Gezel application release.

package/authoring/chat-models/btl4-35b-q4.json

@@ -0,0 +1,142 @@
+{
+  "id": "btl4-35b-q4",
+  "name": "BTL-4 (35B-A3B, Q4)",
+  "description": "Bad Theory Labs' agentic fine-tune of Ornith 1.0 35B, at 35B total parameters with ~2B active per token (Mixture-of-Experts: 256 experts, 8 routed). 30 of its 40 layers use linear attention, so the KV cache costs only ~20 KB/token and the full 256K context stays affordable. Q4_K_M quantization from bartowski — the full-quality companion to the 2.3-bit Compact build. The publisher's benchmark claims are not independently verified.",
+  "tags": [
+    "badtheorylabs",
+    "agentic",
+    "tools",
+    "code",
+    "reasoning",
+    "long-context",
+    "moe"
+  ],
+  "category": "coding",
+  "maintainer": {
+    "name": "Bad Theory Labs",
+    "url": "https://huggingface.co/badtheorylabs"
+  },
+  "version": "1.0.0",
+  "updatedAt": "2026-08-06T00:00:00Z",
+  "license": "Apache-2.0",
+  "licenseClass": "open",
+  "licenseShortName": "Apache 2.0",
+  "recoScore": 10,
+  "parameterSize": "35B",
+  "approxSizeBytes": 21391448192,
+  "supportsTools": true,
+  "contextWindow": 262144,
+  "kvBytesPerTokenF16": 20480,
+  "kvFixedBytesF16": 64389120,
+  "upstream": "https://huggingface.co/badtheorylabs/BTL-4",
+  "style": {
+    "family": "qwen",
+    "reasoningFormat": "think",
+    "toolCallFormat": "function-call"
+  },
+  "behaviors": [
+    "reasoning.strip-think-tags",
+    "prompt.private-reasoning-guidance",
+    "prompt.tool-cookbook-condensed",
+    "fabrication.detect-past-tense-no-tools",
+    "turn.ollama-num-predict-bumped",
+    "turn.preamble-folding",
+    "mcp.compact-tool-schemas"
+  ],
+  "tuning": {
+    "sampling": {
+      "temperature": 0.6,
+      "topP": 0.95,
+      "topK": 20,
+      "minP": 0,
+      "maxTokens": 16384
+    },
+    "samplingWhenThinking": {
+      "temperature": 0.6,
+      "topP": 0.95
+    },
+    "reasoning": {
+      "enableThinking": true,
+      "thinkingBudget": 8192
+    },
+    "promptTags": {
+      "enableThinkingTag": "/think",
+      "disableThinkingTag": "/no_think"
+    },
+    "profiles": {
+      "thinking-general": {
+        "sampling": {
+          "temperature": 0.6,
+          "topP": 0.95,
+          "topK": 20,
+          "minP": 0,
+          "repetitionPenalty": 1,
+          "maxTokens": 24576
+        },
+        "reasoning": {
+          "enableThinking": true,
+          "thinkingBudget": 8192
+        }
+      },
+      "thinking-coding": {
+        "sampling": {
+          "temperature": 0.6,
+          "topP": 0.95,
+          "topK": 20,
+          "minP": 0,
+          "repetitionPenalty": 1,
+          "maxTokens": 32768
+        },
+        "reasoning": {
+          "enableThinking": true,
+          "thinkingBudget": 12288
+        }
+      },
+      "thinking-precise": {
+        "sampling": {
+          "temperature": 0.4,
+          "topP": 0.9,
+          "topK": 20,
+          "minP": 0,
+          "seed": 0,
+          "maxTokens": 6144
+        },
+        "reasoning": {
+          "enableThinking": true,
+          "thinkingBudget": 4096
+        }
+      },
+      "instruct": {
+        "sampling": {
+          "temperature": 0.7,
+          "topP": 0.8,
+          "topK": 20,
+          "minP": 0,
+          "presencePenalty": 1.5,
+          "repetitionPenalty": 1,
+          "maxTokens": 4096
+        },
+        "reasoning": {
+          "enableThinking": false
+        }
+      },
+      "creative": {
+        "sampling": {
+          "temperature": 1.1,
+          "topP": 0.95,
+          "topK": 40,
+          "maxTokens": 8192
+        },
+        "reasoning": {
+          "enableThinking": false
+        }
+      }
+    }
+  },
+  "llamaCpp": {
+    "huggingfaceRepo": "bartowski/badtheorylabs_BTL-4-GGUF",
+    "filename": "badtheorylabs_BTL-4-Q4_K_M.gguf",
+    "quantization": "Q4_K_M",
+    "residentBytes": 25669737830
+  }
+}

package/authoring/chat-models/btl4-compact-iq2.json

@@ -0,0 +1,142 @@
+{
+  "id": "btl4-compact-iq2",
+  "name": "BTL-4 Compact (35B-A3B, IQ2_XXS)",
+  "description": "Bad Theory Labs' agentic fine-tune of Ornith 1.0 35B, at 35B total parameters with ~2B active per token (Mixture-of-Experts: 256 experts, 8 routed). 30 of its 40 layers use linear attention, so the KV cache costs only ~20 KB/token and the full 256K context stays affordable on a laptop. This is a 2.3 bit-per-weight imatrix quantization: expert tensors at IQ2_XXS, everything else at Q4_K_M, which fits the whole model in under 10 GB. The publisher's benchmark claims are not independently verified.",
+  "tags": [
+    "badtheorylabs",
+    "agentic",
+    "tools",
+    "code",
+    "reasoning",
+    "long-context",
+    "moe"
+  ],
+  "category": "coding",
+  "maintainer": {
+    "name": "Bad Theory Labs",
+    "url": "https://huggingface.co/badtheorylabs"
+  },
+  "version": "1.0.1",
+  "updatedAt": "2026-08-06T00:00:00Z",
+  "license": "Apache-2.0",
+  "licenseClass": "open",
+  "licenseShortName": "Apache 2.0",
+  "recoScore": 10,
+  "parameterSize": "35B",
+  "approxSizeBytes": 9967966240,
+  "supportsTools": true,
+  "contextWindow": 262144,
+  "kvBytesPerTokenF16": 20480,
+  "kvFixedBytesF16": 64389120,
+  "upstream": "https://huggingface.co/badtheorylabs/BTL-4-Compact",
+  "style": {
+    "family": "qwen",
+    "reasoningFormat": "think",
+    "toolCallFormat": "function-call"
+  },
+  "behaviors": [
+    "reasoning.strip-think-tags",
+    "prompt.private-reasoning-guidance",
+    "prompt.tool-cookbook-condensed",
+    "fabrication.detect-past-tense-no-tools",
+    "turn.ollama-num-predict-bumped",
+    "turn.preamble-folding",
+    "mcp.compact-tool-schemas"
+  ],
+  "tuning": {
+    "sampling": {
+      "temperature": 0.6,
+      "topP": 0.95,
+      "topK": 20,
+      "minP": 0,
+      "maxTokens": 16384
+    },
+    "samplingWhenThinking": {
+      "temperature": 0.6,
+      "topP": 0.95
+    },
+    "reasoning": {
+      "enableThinking": true,
+      "thinkingBudget": 8192
+    },
+    "promptTags": {
+      "enableThinkingTag": "/think",
+      "disableThinkingTag": "/no_think"
+    },
+    "profiles": {
+      "thinking-general": {
+        "sampling": {
+          "temperature": 0.6,
+          "topP": 0.95,
+          "topK": 20,
+          "minP": 0,
+          "repetitionPenalty": 1,
+          "maxTokens": 24576
+        },
+        "reasoning": {
+          "enableThinking": true,
+          "thinkingBudget": 8192
+        }
+      },
+      "thinking-coding": {
+        "sampling": {
+          "temperature": 0.6,
+          "topP": 0.95,
+          "topK": 20,
+          "minP": 0,
+          "repetitionPenalty": 1,
+          "maxTokens": 32768
+        },
+        "reasoning": {
+          "enableThinking": true,
+          "thinkingBudget": 12288
+        }
+      },
+      "thinking-precise": {
+        "sampling": {
+          "temperature": 0.4,
+          "topP": 0.9,
+          "topK": 20,
+          "minP": 0,
+          "seed": 0,
+          "maxTokens": 6144
+        },
+        "reasoning": {
+          "enableThinking": true,
+          "thinkingBudget": 4096
+        }
+      },
+      "instruct": {
+        "sampling": {
+          "temperature": 0.7,
+          "topP": 0.8,
+          "topK": 20,
+          "minP": 0,
+          "presencePenalty": 1.5,
+          "repetitionPenalty": 1,
+          "maxTokens": 4096
+        },
+        "reasoning": {
+          "enableThinking": false
+        }
+      },
+      "creative": {
+        "sampling": {
+          "temperature": 1.1,
+          "topP": 0.95,
+          "topK": 40,
+          "maxTokens": 8192
+        },
+        "reasoning": {
+          "enableThinking": false
+        }
+      }
+    }
+  },
+  "llamaCpp": {
+    "huggingfaceRepo": "badtheorylabs/BTL-4-Compact",
+    "filename": "BTL-4-IQ2_XXS.gguf",
+    "quantization": "IQ2_XXS",
+    "residentBytes": 11961559488
+  }
+}

package/authoring/chat-models/deepseek-r1-8b-q4.json

@@ -0,0 +1,53 @@
+{
+  "id": "deepseek-r1-8b-q4",
+  "name": "DeepSeek R1 (8B)",
+  "description": "Reasoning-tuned 8B model. Slower than a plain chat model, but thinks through its answers before replying.",
+  "tags": [
+    "deepseek",
+    "tools",
+    "reasoning"
+  ],
+  "category": "reasoning",
+  "maintainer": {
+    "name": "DeepSeek",
+    "url": "https://ollama.com/library/deepseek-r1"
+  },
+  "version": "1.1.0",
+  "updatedAt": "2026-04-26T00:00:00Z",
+  "license": "MIT",
+  "parameterSize": "8B",
+  "approxSizeBytes": 5200000000,
+  "supportsTools": true,
+  "contextWindow": 128000,
+  "upstream": "https://ollama.com/library/deepseek-r1",
+  "style": {
+    "family": "deepseek",
+    "reasoningFormat": "think",
+    "toolCallFormat": "function-call"
+  },
+  "behaviors": [
+    "reasoning.strip-think-tags",
+    "prompt.tool-cookbook-condensed",
+    "fabrication.detect-past-tense-no-tools",
+    "turn.preamble-folding",
+    {
+      "id": "turn.ramble-detection",
+      "config": {
+        "coldThreshold": 3000,
+        "postActionThreshold": 1200
+      }
+    }
+  ],
+  "ollama": {
+    "tag": "deepseek-r1:8b"
+  },
+  "llamaCpp": {
+    "huggingfaceRepo": "unsloth/DeepSeek-R1-0528-Qwen3-8B-GGUF",
+    "filename": "DeepSeek-R1-0528-Qwen3-8B-Q4_K_M.gguf",
+    "quantization": "Q4_K_M"
+  },
+  "mlx": {
+    "huggingfaceRepo": "mlx-community/DeepSeek-R1-0528-Qwen3-8B-4bit",
+    "quantization": "4bit"
+  }
+}

package/authoring/chat-models/deepseek-v4-flash-284b-mxfp4.json

@@ -0,0 +1,110 @@
+{
+  "id": "deepseek-v4-flash-284b-mxfp4",
+  "name": "DeepSeek V4 Flash (MXFP4)",
+  "description": "antirez's MXFP4 community GGUF of DeepSeek V4 Flash 0731 — the July 2026 re-post-trained 284B-parameter Mixture-of-Experts (~13B active per token) with 1M-token context and Compressed Sparse Attention. Routed experts use the MXFP4 micro-scaled 4-bit block format while attention projections, shared experts, router, embeddings, and output head stay at Q8_0 / F16, landing at ~156 GB — close to the native FP4 build's footprint and the highest-fidelity single-file ds4 variant. Needs a workstation or Mac Studio-class machine with ~200 GB of usable memory. Produced for antirez's `ds4` inference engine.",
+  "tags": [
+    "deepseek",
+    "reasoning",
+    "agentic",
+    "moe",
+    "tools",
+    "long-context",
+    "large",
+    "mxfp4",
+    "fp4"
+  ],
+  "category": "reasoning",
+  "maintainer": {
+    "name": "antirez",
+    "url": "https://huggingface.co/antirez/deepseek-v4-gguf"
+  },
+  "version": "1.0.0",
+  "updatedAt": "2026-08-02T00:00:00Z",
+  "license": "MIT",
+  "licenseClass": "open",
+  "licenseShortName": "MIT",
+  "parameterSize": "284B",
+  "approxSizeBytes": 155976458848,
+  "supportsTools": true,
+  "contextWindow": 1000000,
+  "upstream": "https://huggingface.co/deepseek-ai/DeepSeek-V4-Flash",
+  "style": {
+    "family": "deepseek",
+    "reasoningFormat": "think",
+    "toolCallFormat": "function-call"
+  },
+  "behaviors": [
+    "reasoning.strip-think-tags",
+    "prompt.tool-cookbook-condensed",
+    "prompt.meester-build-prelude",
+    "mcp.compact-tool-schemas",
+    "fabrication.detect-past-tense-no-tools",
+    "turn.preamble-folding"
+  ],
+  "tuning": {
+    "sampling": {
+      "maxTokens": 16384
+    },
+    "reasoning": {
+      "enableThinking": true,
+      "thinkingBudget": 4096
+    },
+    "profiles": {
+      "thinking-general": {
+        "sampling": {
+          "temperature": 0.6,
+          "topP": 0.95,
+          "maxTokens": 16384
+        }
+      },
+      "thinking-coding": {
+        "sampling": {
+          "temperature": 0.3,
+          "topP": 0.95,
+          "maxTokens": 16384
+        }
+      },
+      "thinking-precise": {
+        "sampling": {
+          "temperature": 0.1,
+          "topP": 0.9,
+          "maxTokens": 16384,
+          "seed": 0
+        }
+      },
+      "instruct": {
+        "sampling": {
+          "temperature": 0.7,
+          "topP": 0.95,
+          "maxTokens": 16384
+        },
+        "reasoning": {
+          "enableThinking": false
+        }
+      },
+      "creative": {
+        "sampling": {
+          "temperature": 1,
+          "topP": 1,
+          "maxTokens": 16384
+        },
+        "reasoning": {
+          "enableThinking": false
+        }
+      }
+    }
+  },
+  "llamaCpp": {
+    "huggingfaceRepo": "antirez/deepseek-v4-gguf",
+    "filename": "DeepSeek-V4-Flash-MXFP4Experts-F16HC-F16Compressor-F16Indexer-Q8Attn-Q8Shared-Q8Out-chat-v2-mxfp4-0731.gguf",
+    "quantization": "MXFP4"
+  },
+  "ds4": {
+    "huggingfaceRepo": "antirez/deepseek-v4-gguf",
+    "filename": "DeepSeek-V4-Flash-MXFP4Experts-F16HC-F16Compressor-F16Indexer-Q8Attn-Q8Shared-Q8Out-chat-v2-mxfp4-0731.gguf",
+    "quantization": "MXFP4",
+    "residentBytes": 85899345920,
+    "cacheExpertsBytes": 68719476736,
+    "ssdStreaming": true
+  }
+}

package/authoring/chat-models/deepseek-v4-flash-284b-q2.json

@@ -0,0 +1,110 @@
+{
+  "id": "deepseek-v4-flash-284b-q2",
+  "name": "DeepSeek V4 Flash (IQ2_XXS)",
+  "description": "antirez's 2-bit community GGUF of DeepSeek V4 Flash 0731 — the July 2026 re-post-trained checkpoint of the 284B-parameter Mixture-of-Experts (~13B active per token) with 1M-token context and Compressed Sparse Attention. Same architecture as the original V4 Flash, with substantially stronger agentic and coding behaviour. The routed experts are crushed to IQ2_XXS / Q2_K while attention projections, shared experts, router, embeddings, and output head stay at Q8_0 / F16, landing the whole model in ~81 GB — runnable on 128 GB-class machines versus ~156 GB for the native FP4 build. Note: these quants were produced for antirez's `ds4` inference engine; whether they load on stock llama depends on its DeepSeek-V4 support.",
+  "tags": [
+    "deepseek",
+    "reasoning",
+    "agentic",
+    "moe",
+    "tools",
+    "2-bit",
+    "iq2",
+    "long-context",
+    "large"
+  ],
+  "category": "reasoning",
+  "maintainer": {
+    "name": "antirez",
+    "url": "https://huggingface.co/antirez/deepseek-v4-gguf"
+  },
+  "version": "1.1.0",
+  "updatedAt": "2026-08-02T00:00:00Z",
+  "license": "MIT",
+  "licenseClass": "open",
+  "licenseShortName": "MIT",
+  "parameterSize": "284B",
+  "approxSizeBytes": 86720111488,
+  "supportsTools": true,
+  "contextWindow": 1000000,
+  "upstream": "https://huggingface.co/deepseek-ai/DeepSeek-V4-Flash",
+  "style": {
+    "family": "deepseek",
+    "reasoningFormat": "think",
+    "toolCallFormat": "function-call"
+  },
+  "behaviors": [
+    "reasoning.strip-think-tags",
+    "prompt.tool-cookbook-condensed",
+    "prompt.meester-build-prelude",
+    "mcp.compact-tool-schemas",
+    "fabrication.detect-past-tense-no-tools",
+    "turn.preamble-folding"
+  ],
+  "tuning": {
+    "sampling": {
+      "maxTokens": 16384
+    },
+    "reasoning": {
+      "enableThinking": true,
+      "thinkingBudget": 4096
+    },
+    "profiles": {
+      "thinking-general": {
+        "sampling": {
+          "temperature": 0.6,
+          "topP": 0.95,
+          "maxTokens": 16384
+        }
+      },
+      "thinking-coding": {
+        "sampling": {
+          "temperature": 0.3,
+          "topP": 0.95,
+          "maxTokens": 16384
+        }
+      },
+      "thinking-precise": {
+        "sampling": {
+          "temperature": 0.1,
+          "topP": 0.9,
+          "maxTokens": 16384,
+          "seed": 0
+        }
+      },
+      "instruct": {
+        "sampling": {
+          "temperature": 0.7,
+          "topP": 0.95,
+          "maxTokens": 16384
+        },
+        "reasoning": {
+          "enableThinking": false
+        }
+      },
+      "creative": {
+        "sampling": {
+          "temperature": 1,
+          "topP": 1,
+          "maxTokens": 16384
+        },
+        "reasoning": {
+          "enableThinking": false
+        }
+      }
+    }
+  },
+  "llamaCpp": {
+    "huggingfaceRepo": "antirez/deepseek-v4-gguf",
+    "filename": "DeepSeek-V4-Flash-IQ2XXS-w2Q2K-AProjQ8-SExpQ8-OutQ8-chat-v2-imatrix-0731.gguf",
+    "quantization": "IQ2_XXS"
+  },
+  "ds4": {
+    "huggingfaceRepo": "antirez/deepseek-v4-gguf",
+    "filename": "DeepSeek-V4-Flash-IQ2XXS-w2Q2K-AProjQ8-SExpQ8-OutQ8-chat-v2-imatrix-0731.gguf",
+    "quantization": "IQ2_XXS",
+    "residentBytes": 38654705664,
+    "cacheExpertsBytes": 34359738368,
+    "ssdStreaming": true
+  }
+}