@caupulican/pi-adaptative 0.85.3 → 0.85.8

package/dist/core/models/context-sizing.js.map

@@ -1 +1 @@
-{"version":3,"file":"context-sizing.js","sourceRoot":"","sources":["../../../src/core/models/context-sizing.ts"],"names":[],"mappings":"AAAA,MAAM,aAAa,GAAG,CAAC,KAAK,EAAE,KAAK,EAAE,MAAM,EAAE,MAAM,CAAU,CAAC;AAC9D,MAAM,yBAAyB,GAAG,GAAG,CAAC;AACtC,MAAM,SAAS,GAAG,CAAC,CAAC;AACpB,MAAM,QAAQ,GAAG,CAAC,CAAC;AAwBnB,MAAM,UAAU,wBAAwB,CAAC,IAKxC;IACA,MAAM,OAAO,GAAG,aAAa,CAAC,IAAI,CAAC,KAAK,CAAC,SAAS,CAAC,CAAC;IACpD,IAAI,CAAC,OAAO;QAAE,OAAO,SAAS,CAAC;IAC/B,MAAM,gBAAgB,GAAG,IAAI,CAAC,IAAI,CAAC,gBAAgB,IAAI,yBAAyB,CAAC;IACjF,MAAM,WAAW,GAAG,IAAI,CAAC,KAAK,CAAC,IAAI,CAAC,IAAI,CAAC,aAAa,GAAG,gBAAgB,CAAC,CAAC;IAC3E,MAAM,gBAAgB,GAAG,WAAW,GAAG,IAAI,CAAC,KAAK,CAAC,YAAY,CAAC;IAC/D,IAAI,gBAAgB,IAAI,CAAC;QAAE,OAAO,SAAS,CAAC;IAC5C,MAAM,KAAK,GAAG,IAAI,CAAC,KAAK,IAAI,aAAa,CAAC;IAC1C,MAAM,WAAW,GAAG,iBAAiB,CAAC,EAAE,gBAAgB,EAAE,eAAe,EAAE,SAAS,EAAE,WAAW,EAAE,OAAO,EAAE,KAAK,EAAE,CAAC,CAAC;IACrH,IAAI,CAAC,IAAI,CAAC,OAAO,EAAE,sBAAsB;QAAE,OAAO,WAAW,CAAC;IAC9D,MAAM,UAAU,GAAG,iBAAiB,CAAC;QACpC,gBAAgB;QAChB,eAAe,EAAE,QAAQ;QACzB,WAAW;QACX,WAAW,EAAE,MAAM;QACnB,OAAO;QACP,KAAK;KACL,CAAC,CAAC;IACH,IAAI,UAAU,IAAI,CAAC,CAAC,WAAW,IAAI,UAAU,CAAC,MAAM,GAAG,WAAW,CAAC,MAAM,CAAC;QAAE,OAAO,UAAU,CAAC;IAC9F,OAAO,WAAW,CAAC;AACpB,CAAC;AAED,MAAM,UAAU,4BAA4B,CAAC,IAAsC;IAClF,OAAO,QAAQ,IAAI,CAAC,IAAI,uBAAuB,IAAI,CAAC,MAAM,IAAI,CAAC;AAChE,CAAC;AAED,MAAM,UAAU,kBAAkB,CAAC,SAAiB,EAAE,MAAc;IACnE,MAAM,UAAU,GAAG,SAAS,CAAC,OAAO,CAAC,mBAAmB,EAAE,GAAG,CAAC,CAAC,OAAO,CAAC,UAAU,EAAE,EAAE,CAAC,IAAI,aAAa,CAAC;IACxG,OAAO,MAAM,UAAU,OAAO,MAAM,EAAE,CAAC;AACxC,CAAC;AAED,SAAS,iBAAiB,CAAC,IAO1B;IACA,MAAM,eAAe,GAAG,IAAI,CAAC,OAAO,CAAC,gBAAgB,GAAG,IAAI,CAAC,eAAe,CAAC;IAC7E,KAAK,MAAM,MAAM,IAAI,CAAC,GAAG,IAAI,CAAC,KAAK,CAAC,CAAC,IAAI,CAAC,CAAC,CAAC,EAAE,CAAC,EAAE,EAAE,CAAC,CAAC,GAAG,CAAC,CAAC,EAAE,CAAC;QAC5D,MAAM,cAAc,GAAG,eAAe,GAAG,MAAM,CAAC;QAChD,IAAI,cAAc,IAAI,IAAI,CAAC,gBAAgB,EAAE,CAAC;YAC7C,OAAO;gBACN,MAAM;gBACN,WAAW,EAAE,IAAI,CAAC,WAAW;gBAC7B,eAAe;gBACf,WAAW,EAAE,IAAI,CAAC,WAAW;gBAC7B,cAAc;aACd,CAAC;QACH,CAAC;IACF,CAAC;IACD,OAAO,SAAS,CAAC;AAClB,CAAC;AAED,SAAS,aAAa,CAAC,SAAkC;IACxD,MAAM,MAAM,GAAG,eAAe,CAAC,SAAS,EAAE,aAAa,CAAC,CAAC;IACzD,IAAI,CAAC,MAAM;QAAE,OAAO,SAAS,CAAC;IAC9B,MAAM,WAAW,GAAG,eAAe,CAAC,SAAS,EAAE,yBAAyB,CAAC,CAAC;IAC1E,MAAM,SAAS,GAAG,eAAe,CAAC,SAAS,EAAE,sBAAsB,CAAC,CAAC;IACrE,MAAM,WAAW,GAAG,eAAe,CAAC,SAAS,EAAE,wBAAwB,CAAC,CAAC;IACzE,IAAI,WAAW,IAAI,SAAS,IAAI,WAAW,EAAE,CAAC;QAC7C,OAAO,EAAE,gBAAgB,EAAE,MAAM,GAAG,WAAW,GAAG,CAAC,SAAS,GAAG,WAAW,CAAC,EAAE,CAAC;IAC/E,CAAC;IACD,MAAM,eAAe,GAAG,eAAe,CAAC,SAAS,EAAE,kBAAkB,CAAC,CAAC;IACvE,OAAO,eAAe,CAAC,CAAC,CAAC,EAAE,gBAAgB,EAAE,MAAM,GAAG,eAAe,GAAG,CAAC,EAAE,CAAC,CAAC,CAAC,SAAS,CAAC;AACzF,CAAC;AAED,SAAS,eAAe,CAAC,MAA+B,EAAE,MAAc;IACvE,MAAM,YAAY,GAAG,OAAO,MAAM,CAAC,sBAAsB,CAAC,KAAK,QAAQ,CAAC,CAAC,CAAC,MAAM,CAAC,sBAAsB,CAAC,CAAC,IAAI,EAAE,CAAC,CAAC,CAAC,EAAE,CAAC;IACrH,IAAI,YAAY;QAAE,OAAO,cAAc,CAAC,MAAM,CAAC,GAAG,YAAY,IAAI,MAAM,EAAE,CAAC,CAAC,CAAC;IAC7E,KAAK,MAAM,CAAC,GAAG,EAAE,KAAK,CAAC,IAAI,MAAM,CAAC,OAAO,CAAC,MAAM,CAAC,EAAE,CAAC;QACnD,IAAI,GAAG,CAAC,QAAQ,CAAC,IAAI,MAAM,EAAE,CAAC,EAAE,CAAC;YAChC,MAAM,OAAO,GAAG,cAAc,CAAC,KAAK,CAAC,CAAC;YACtC,IAAI,OAAO;gBAAE,OAAO,OAAO,CAAC;QAC7B,CAAC;IACF,CAAC;IACD,OAAO,SAAS,CAAC;AAClB,CAAC;AAED,SAAS,cAAc,CAAC,KAAc;IACrC,OAAO,OAAO,KAAK,KAAK,QAAQ,IAAI,MAAM,CAAC,QAAQ,CAAC,KAAK,CAAC,IAAI,KAAK,GAAG,CAAC,CAAC,CAAC,CAAC,KAAK,CAAC,CAAC,CAAC,SAAS,CAAC;AAC7F,CAAC","sourcesContent":["const CONTEXT_RUNGS = [4_096, 8_192, 16_384, 32_768] as const;\nconst DEFAULT_HEADROOM_FRACTION = 0.8;\nconst F16_BYTES = 2;\nconst Q8_BYTES = 1;\n\nexport interface ContextSizingModelInfo {\n\tmodelInfo: Record<string, unknown>;\n\tweightsBytes: number;\n}\n\nexport interface ContextSizingHostProfile {\n\ttotalMemBytes: number;\n\theadroomFraction?: number;\n}\n\nexport interface ContextSizingRuntimeCaps {\n\tsupportsKvQuantization?: boolean;\n}\n\nexport interface ContextSizingDecision {\n\tnumCtx: number;\n\tkvCacheType?: \"q8_0\";\n\tkvBytesPerToken: number;\n\tbudgetBytes: number;\n\testimatedBytes: number;\n}\n\nexport function deriveLocalContextSizing(args: {\n\thost: ContextSizingHostProfile;\n\tmodel: ContextSizingModelInfo;\n\truntime?: ContextSizingRuntimeCaps;\n\trungs?: readonly number[];\n}): ContextSizingDecision | undefined {\n\tconst kvShape = deriveKvShape(args.model.modelInfo);\n\tif (!kvShape) return undefined;\n\tconst headroomFraction = args.host.headroomFraction ?? DEFAULT_HEADROOM_FRACTION;\n\tconst budgetBytes = Math.floor(args.host.totalMemBytes * headroomFraction);\n\tconst availableKvBytes = budgetBytes - args.model.weightsBytes;\n\tif (availableKvBytes <= 0) return undefined;\n\tconst rungs = args.rungs ?? CONTEXT_RUNGS;\n\tconst f16Decision = chooseContextRung({ availableKvBytes, bytesPerElement: F16_BYTES, budgetBytes, kvShape, rungs });\n\tif (!args.runtime?.supportsKvQuantization) return f16Decision;\n\tconst q8Decision = chooseContextRung({\n\t\tavailableKvBytes,\n\t\tbytesPerElement: Q8_BYTES,\n\t\tbudgetBytes,\n\t\tkvCacheType: \"q8_0\",\n\t\tkvShape,\n\t\trungs,\n\t});\n\tif (q8Decision && (!f16Decision || q8Decision.numCtx > f16Decision.numCtx)) return q8Decision;\n\treturn f16Decision;\n}\n\nexport function renderOllamaContextModelfile(args: { from: string; numCtx: number }): string {\n\treturn `FROM ${args.from}\\nPARAMETER num_ctx ${args.numCtx}\\n`;\n}\n\nexport function sizedLocalModelRef(sourceRef: string, numCtx: number): string {\n\tconst safeSource = sourceRef.replace(/[^a-zA-Z0-9_.-]+/g, \"-\").replace(/^-+|-+$/g, \"\") || \"local-model\";\n\treturn `pi-${safeSource}:ctx${numCtx}`;\n}\n\nfunction chooseContextRung(args: {\n\tavailableKvBytes: number;\n\tbytesPerElement: number;\n\tbudgetBytes: number;\n\tkvCacheType?: \"q8_0\";\n\tkvShape: { elementsPerToken: number };\n\trungs: readonly number[];\n}): ContextSizingDecision | undefined {\n\tconst kvBytesPerToken = args.kvShape.elementsPerToken * args.bytesPerElement;\n\tfor (const numCtx of [...args.rungs].sort((a, b) => b - a)) {\n\t\tconst estimatedBytes = kvBytesPerToken * numCtx;\n\t\tif (estimatedBytes <= args.availableKvBytes) {\n\t\t\treturn {\n\t\t\t\tnumCtx,\n\t\t\t\tkvCacheType: args.kvCacheType,\n\t\t\t\tkvBytesPerToken,\n\t\t\t\tbudgetBytes: args.budgetBytes,\n\t\t\t\testimatedBytes,\n\t\t\t};\n\t\t}\n\t}\n\treturn undefined;\n}\n\nfunction deriveKvShape(modelInfo: Record<string, unknown>): { elementsPerToken: number } | undefined {\n\tconst layers = modelInfoNumber(modelInfo, \"block_count\");\n\tif (!layers) return undefined;\n\tconst headCountKv = modelInfoNumber(modelInfo, \"attention.head_count_kv\");\n\tconst keyLength = modelInfoNumber(modelInfo, \"attention.key_length\");\n\tconst valueLength = modelInfoNumber(modelInfo, \"attention.value_length\");\n\tif (headCountKv && keyLength && valueLength) {\n\t\treturn { elementsPerToken: layers * headCountKv * (keyLength + valueLength) };\n\t}\n\tconst embeddingLength = modelInfoNumber(modelInfo, \"embedding_length\");\n\treturn embeddingLength ? { elementsPerToken: layers * embeddingLength * 2 } : undefined;\n}\n\nfunction modelInfoNumber(record: Record<string, unknown>, suffix: string): number | undefined {\n\tconst architecture = typeof record[\"general.architecture\"] === \"string\" ? record[\"general.architecture\"].trim() : \"\";\n\tif (architecture) return positiveNumber(record[`${architecture}.${suffix}`]);\n\tfor (const [key, value] of Object.entries(record)) {\n\t\tif (key.endsWith(`.${suffix}`)) {\n\t\t\tconst matched = positiveNumber(value);\n\t\t\tif (matched) return matched;\n\t\t}\n\t}\n\treturn undefined;\n}\n\nfunction positiveNumber(value: unknown): number | undefined {\n\treturn typeof value === \"number\" && Number.isFinite(value) && value > 0 ? value : undefined;\n}\n"]}
+{"version":3,"file":"context-sizing.js","sourceRoot":"","sources":["../../../src/core/models/context-sizing.ts"],"names":[],"mappings":"AAAA,MAAM,aAAa,GAAG,CAAC,KAAK,EAAE,KAAK,EAAE,MAAM,EAAE,MAAM,CAAU,CAAC;AAC9D,MAAM,yBAAyB,GAAG,GAAG,CAAC;AACtC,MAAM,SAAS,GAAG,CAAC,CAAC;AACpB,MAAM,QAAQ,GAAG,CAAC,CAAC;AAwBnB,MAAM,UAAU,wBAAwB,CAAC,IAKxC;IACA,MAAM,OAAO,GAAG,aAAa,CAAC,IAAI,CAAC,KAAK,CAAC,SAAS,CAAC,CAAC;IACpD,IAAI,CAAC,OAAO;QAAE,OAAO,SAAS,CAAC;IAC/B,MAAM,gBAAgB,GAAG,IAAI,CAAC,IAAI,CAAC,gBAAgB,IAAI,yBAAyB,CAAC;IACjF,MAAM,WAAW,GAAG,IAAI,CAAC,KAAK,CAAC,IAAI,CAAC,IAAI,CAAC,aAAa,GAAG,gBAAgB,CAAC,CAAC;IAC3E,MAAM,gBAAgB,GAAG,WAAW,GAAG,IAAI,CAAC,KAAK,CAAC,YAAY,CAAC;IAC/D,IAAI,gBAAgB,IAAI,CAAC;QAAE,OAAO,SAAS,CAAC;IAC5C,MAAM,KAAK,GAAG,IAAI,CAAC,KAAK,IAAI,aAAa,CAAC;IAC1C,MAAM,WAAW,GAAG,iBAAiB,CAAC,EAAE,gBAAgB,EAAE,eAAe,EAAE,SAAS,EAAE,WAAW,EAAE,OAAO,EAAE,KAAK,EAAE,CAAC,CAAC;IACrH,IAAI,CAAC,IAAI,CAAC,OAAO,EAAE,sBAAsB;QAAE,OAAO,WAAW,CAAC;IAC9D,MAAM,UAAU,GAAG,iBAAiB,CAAC;QACpC,gBAAgB;QAChB,eAAe,EAAE,QAAQ;QACzB,WAAW;QACX,WAAW,EAAE,MAAM;QACnB,OAAO;QACP,KAAK;KACL,CAAC,CAAC;IACH,IAAI,UAAU,IAAI,CAAC,CAAC,WAAW,IAAI,UAAU,CAAC,MAAM,GAAG,WAAW,CAAC,MAAM,CAAC;QAAE,OAAO,UAAU,CAAC;IAC9F,OAAO,WAAW,CAAC;AACpB,CAAC;AAED,MAAM,UAAU,4BAA4B,CAAC,IAAsC;IAClF,OAAO,QAAQ,IAAI,CAAC,IAAI,uBAAuB,IAAI,CAAC,MAAM,IAAI,CAAC;AAChE,CAAC;AAED,MAAM,UAAU,kBAAkB,CAAC,SAAiB,EAAE,MAAc;IACnE,OAAO,oBAAoB,CAAC,SAAS,EAAE,MAAM,MAAM,EAAE,CAAC,CAAC;AACxD,CAAC;AAED,MAAM,UAAU,oBAAoB,CAAC,SAAiB,EAAE,GAAW;IAClE,MAAM,UAAU,GAAG,SAAS,CAAC,OAAO,CAAC,mBAAmB,EAAE,GAAG,CAAC,CAAC,OAAO,CAAC,UAAU,EAAE,EAAE,CAAC,IAAI,aAAa,CAAC;IACxG,OAAO,MAAM,UAAU,IAAI,GAAG,EAAE,CAAC;AAClC,CAAC;AAED,SAAS,iBAAiB,CAAC,IAO1B;IACA,MAAM,eAAe,GAAG,IAAI,CAAC,OAAO,CAAC,gBAAgB,GAAG,IAAI,CAAC,eAAe,CAAC;IAC7E,KAAK,MAAM,MAAM,IAAI,CAAC,GAAG,IAAI,CAAC,KAAK,CAAC,CAAC,IAAI,CAAC,CAAC,CAAC,EAAE,CAAC,EAAE,EAAE,CAAC,CAAC,GAAG,CAAC,CAAC,EAAE,CAAC;QAC5D,MAAM,cAAc,GAAG,eAAe,GAAG,MAAM,CAAC;QAChD,IAAI,cAAc,IAAI,IAAI,CAAC,gBAAgB,EAAE,CAAC;YAC7C,OAAO;gBACN,MAAM;gBACN,WAAW,EAAE,IAAI,CAAC,WAAW;gBAC7B,eAAe;gBACf,WAAW,EAAE,IAAI,CAAC,WAAW;gBAC7B,cAAc;aACd,CAAC;QACH,CAAC;IACF,CAAC;IACD,OAAO,SAAS,CAAC;AAClB,CAAC;AAED,SAAS,aAAa,CAAC,SAAkC;IACxD,MAAM,MAAM,GAAG,eAAe,CAAC,SAAS,EAAE,aAAa,CAAC,CAAC;IACzD,IAAI,CAAC,MAAM;QAAE,OAAO,SAAS,CAAC;IAC9B,MAAM,WAAW,GAAG,eAAe,CAAC,SAAS,EAAE,yBAAyB,CAAC,CAAC;IAC1E,MAAM,SAAS,GAAG,eAAe,CAAC,SAAS,EAAE,sBAAsB,CAAC,CAAC;IACrE,MAAM,WAAW,GAAG,eAAe,CAAC,SAAS,EAAE,wBAAwB,CAAC,CAAC;IACzE,IAAI,WAAW,IAAI,SAAS,IAAI,WAAW,EAAE,CAAC;QAC7C,OAAO,EAAE,gBAAgB,EAAE,MAAM,GAAG,WAAW,GAAG,CAAC,SAAS,GAAG,WAAW,CAAC,EAAE,CAAC;IAC/E,CAAC;IACD,MAAM,eAAe,GAAG,eAAe,CAAC,SAAS,EAAE,kBAAkB,CAAC,CAAC;IACvE,OAAO,eAAe,CAAC,CAAC,CAAC,EAAE,gBAAgB,EAAE,MAAM,GAAG,eAAe,GAAG,CAAC,EAAE,CAAC,CAAC,CAAC,SAAS,CAAC;AACzF,CAAC;AAED,SAAS,eAAe,CAAC,MAA+B,EAAE,MAAc;IACvE,MAAM,YAAY,GAAG,OAAO,MAAM,CAAC,sBAAsB,CAAC,KAAK,QAAQ,CAAC,CAAC,CAAC,MAAM,CAAC,sBAAsB,CAAC,CAAC,IAAI,EAAE,CAAC,CAAC,CAAC,EAAE,CAAC;IACrH,IAAI,YAAY;QAAE,OAAO,cAAc,CAAC,MAAM,CAAC,GAAG,YAAY,IAAI,MAAM,EAAE,CAAC,CAAC,CAAC;IAC7E,KAAK,MAAM,CAAC,GAAG,EAAE,KAAK,CAAC,IAAI,MAAM,CAAC,OAAO,CAAC,MAAM,CAAC,EAAE,CAAC;QACnD,IAAI,GAAG,CAAC,QAAQ,CAAC,IAAI,MAAM,EAAE,CAAC,EAAE,CAAC;YAChC,MAAM,OAAO,GAAG,cAAc,CAAC,KAAK,CAAC,CAAC;YACtC,IAAI,OAAO;gBAAE,OAAO,OAAO,CAAC;QAC7B,CAAC;IACF,CAAC;IACD,OAAO,SAAS,CAAC;AAClB,CAAC;AAED,SAAS,cAAc,CAAC,KAAc;IACrC,OAAO,OAAO,KAAK,KAAK,QAAQ,IAAI,MAAM,CAAC,QAAQ,CAAC,KAAK,CAAC,IAAI,KAAK,GAAG,CAAC,CAAC,CAAC,CAAC,KAAK,CAAC,CAAC,CAAC,SAAS,CAAC;AAC7F,CAAC","sourcesContent":["const CONTEXT_RUNGS = [4_096, 8_192, 16_384, 32_768] as const;\nconst DEFAULT_HEADROOM_FRACTION = 0.8;\nconst F16_BYTES = 2;\nconst Q8_BYTES = 1;\n\nexport interface ContextSizingModelInfo {\n\tmodelInfo: Record<string, unknown>;\n\tweightsBytes: number;\n}\n\nexport interface ContextSizingHostProfile {\n\ttotalMemBytes: number;\n\theadroomFraction?: number;\n}\n\nexport interface ContextSizingRuntimeCaps {\n\tsupportsKvQuantization?: boolean;\n}\n\nexport interface ContextSizingDecision {\n\tnumCtx: number;\n\tkvCacheType?: \"q8_0\";\n\tkvBytesPerToken: number;\n\tbudgetBytes: number;\n\testimatedBytes: number;\n}\n\nexport function deriveLocalContextSizing(args: {\n\thost: ContextSizingHostProfile;\n\tmodel: ContextSizingModelInfo;\n\truntime?: ContextSizingRuntimeCaps;\n\trungs?: readonly number[];\n}): ContextSizingDecision | undefined {\n\tconst kvShape = deriveKvShape(args.model.modelInfo);\n\tif (!kvShape) return undefined;\n\tconst headroomFraction = args.host.headroomFraction ?? DEFAULT_HEADROOM_FRACTION;\n\tconst budgetBytes = Math.floor(args.host.totalMemBytes * headroomFraction);\n\tconst availableKvBytes = budgetBytes - args.model.weightsBytes;\n\tif (availableKvBytes <= 0) return undefined;\n\tconst rungs = args.rungs ?? CONTEXT_RUNGS;\n\tconst f16Decision = chooseContextRung({ availableKvBytes, bytesPerElement: F16_BYTES, budgetBytes, kvShape, rungs });\n\tif (!args.runtime?.supportsKvQuantization) return f16Decision;\n\tconst q8Decision = chooseContextRung({\n\t\tavailableKvBytes,\n\t\tbytesPerElement: Q8_BYTES,\n\t\tbudgetBytes,\n\t\tkvCacheType: \"q8_0\",\n\t\tkvShape,\n\t\trungs,\n\t});\n\tif (q8Decision && (!f16Decision || q8Decision.numCtx > f16Decision.numCtx)) return q8Decision;\n\treturn f16Decision;\n}\n\nexport function renderOllamaContextModelfile(args: { from: string; numCtx: number }): string {\n\treturn `FROM ${args.from}\\nPARAMETER num_ctx ${args.numCtx}\\n`;\n}\n\nexport function sizedLocalModelRef(sourceRef: string, numCtx: number): string {\n\treturn derivedLocalModelRef(sourceRef, `ctx${numCtx}`);\n}\n\nexport function derivedLocalModelRef(sourceRef: string, tag: string): string {\n\tconst safeSource = sourceRef.replace(/[^a-zA-Z0-9_.-]+/g, \"-\").replace(/^-+|-+$/g, \"\") || \"local-model\";\n\treturn `pi-${safeSource}:${tag}`;\n}\n\nfunction chooseContextRung(args: {\n\tavailableKvBytes: number;\n\tbytesPerElement: number;\n\tbudgetBytes: number;\n\tkvCacheType?: \"q8_0\";\n\tkvShape: { elementsPerToken: number };\n\trungs: readonly number[];\n}): ContextSizingDecision | undefined {\n\tconst kvBytesPerToken = args.kvShape.elementsPerToken * args.bytesPerElement;\n\tfor (const numCtx of [...args.rungs].sort((a, b) => b - a)) {\n\t\tconst estimatedBytes = kvBytesPerToken * numCtx;\n\t\tif (estimatedBytes <= args.availableKvBytes) {\n\t\t\treturn {\n\t\t\t\tnumCtx,\n\t\t\t\tkvCacheType: args.kvCacheType,\n\t\t\t\tkvBytesPerToken,\n\t\t\t\tbudgetBytes: args.budgetBytes,\n\t\t\t\testimatedBytes,\n\t\t\t};\n\t\t}\n\t}\n\treturn undefined;\n}\n\nfunction deriveKvShape(modelInfo: Record<string, unknown>): { elementsPerToken: number } | undefined {\n\tconst layers = modelInfoNumber(modelInfo, \"block_count\");\n\tif (!layers) return undefined;\n\tconst headCountKv = modelInfoNumber(modelInfo, \"attention.head_count_kv\");\n\tconst keyLength = modelInfoNumber(modelInfo, \"attention.key_length\");\n\tconst valueLength = modelInfoNumber(modelInfo, \"attention.value_length\");\n\tif (headCountKv && keyLength && valueLength) {\n\t\treturn { elementsPerToken: layers * headCountKv * (keyLength + valueLength) };\n\t}\n\tconst embeddingLength = modelInfoNumber(modelInfo, \"embedding_length\");\n\treturn embeddingLength ? { elementsPerToken: layers * embeddingLength * 2 } : undefined;\n}\n\nfunction modelInfoNumber(record: Record<string, unknown>, suffix: string): number | undefined {\n\tconst architecture = typeof record[\"general.architecture\"] === \"string\" ? record[\"general.architecture\"].trim() : \"\";\n\tif (architecture) return positiveNumber(record[`${architecture}.${suffix}`]);\n\tfor (const [key, value] of Object.entries(record)) {\n\t\tif (key.endsWith(`.${suffix}`)) {\n\t\t\tconst matched = positiveNumber(value);\n\t\t\tif (matched) return matched;\n\t\t}\n\t}\n\treturn undefined;\n}\n\nfunction positiveNumber(value: unknown): number | undefined {\n\treturn typeof value === \"number\" && Number.isFinite(value) && value > 0 ? value : undefined;\n}\n"]}

package/dist/core/models/default-model-suggestions.d.ts.map

@@ -1 +1 @@
-{"version":3,"file":"default-model-suggestions.d.ts","sourceRoot":"","sources":["../../../src/core/models/default-model-suggestions.ts"],"names":[],"mappings":"AAAA,OAAO,KAAK,EAAE,WAAW,EAAE,MAAM,6DAA6D,CAAC;AAE/F;;;;;;;;;;;;GAYG;AAEH,MAAM,WAAW,eAAe;IAC/B,4CAA4C;IAC5C,IAAI,EAAE,MAAM,CAAC;IACb,6DAA6D;IAC7D,OAAO,EAAE,MAAM,CAAC;IAChB,gDAAgD;IAChD,IAAI,EAAE,MAAM,CAAC;IACb,oGAAoG;IACpG,WAAW,EAAE,OAAO,CAAC;IACrB,mGAAmG;IACnG,UAAU,CAAC,EAAE,WAAW,CAAC;IACzB,qFAAqF;IACrF,SAAS,EAAE,MAAM,CAAC;IAClB,oDAAoD;IACpD,IAAI,CAAC,EAAE,MAAM,CAAC;CACd;AAED,eAAO,MAAM,yBAAyB,EAAE,SAAS,eAAe,EA2G/D,CAAC;AAEF,gFAAgF;AAChF,wBAAgB,sBAAsB,CAAC,WAAW,GAAE,SAAS,eAAe,EAA8B,GAAG,MAAM,EAAE,CAcpH"}
+{"version":3,"file":"default-model-suggestions.d.ts","sourceRoot":"","sources":["../../../src/core/models/default-model-suggestions.ts"],"names":[],"mappings":"AAAA,OAAO,KAAK,EAAE,WAAW,EAAE,MAAM,6DAA6D,CAAC;AAG/F;;;;;;;;;;;;GAYG;AAEH,MAAM,WAAW,eAAe;IAC/B,4CAA4C;IAC5C,IAAI,EAAE,MAAM,CAAC;IACb,6DAA6D;IAC7D,OAAO,EAAE,MAAM,CAAC;IAChB,gDAAgD;IAChD,IAAI,EAAE,MAAM,CAAC;IACb,oGAAoG;IACpG,WAAW,EAAE,OAAO,CAAC;IACrB,mGAAmG;IACnG,UAAU,CAAC,EAAE,WAAW,CAAC;IACzB,qFAAqF;IACrF,SAAS,EAAE,MAAM,CAAC;IAClB,oDAAoD;IACpD,IAAI,CAAC,EAAE,MAAM,CAAC;CACd;AAED,eAAO,MAAM,yBAAyB,EAAE,SAAS,eAAe,EAqH/D,CAAC;AAEF,gFAAgF;AAChF,wBAAgB,sBAAsB,CAAC,WAAW,GAAE,SAAS,eAAe,EAA8B,GAAG,MAAM,EAAE,CAcpH"}

package/dist/core/models/default-model-suggestions.js

@@ -1,3 +1,4 @@
+import { ORNITH_9B_OLLAMA_REF } from "./managed-ollama-model.js";
 export const DEFAULT_MODEL_SUGGESTIONS = [
     {
         name: "qwen3:1.7b",
@@ -35,12 +36,21 @@ export const DEFAULT_MODEL_SUGGESTIONS = [
     },
     {
         name: "Ornith-1.0-9B",
-        pullRef: "hf.co/deepreinforce-ai/Ornith-1.0-9B-GGUF:Q4_K_M",
+        pullRef: ORNITH_9B_OLLAMA_REF,
         role: "Agentic-coding worker / router cheap tier",
         toolCalling: true,
         assignRole: "router-cheap",
         rationale: "External candidate (not from pi's own validation research): MIT, Qwen 3.5 base, RL-trained for agentic coding with native tool-calling — the strongest local worker SHAPE in the roster. /fitness on your hardware is the validator.",
-        note: "Q4_K_M ≈ 5.6 GB weights, ~7-8 GB peak with KV — on a 10 GB box run it as the ONLY local model. Qwen 3.5 arch: confirm pi's pinned Ollama supports it and probe tool-calls with /fitness before assigning (template derives from the GGUF). Larger boxes: consider router-medium after a passed worker lane.",
+        note: "Q4_K_M ≈ 5.6 GB weights, ~7-8 GB peak with KV — on a 10 GB box run it as the ONLY local model. Pi creates a managed Qwen 3.5 renderer/parser profile with the model's recommended sampling before probing. Larger boxes: consider router-medium after a passed worker lane.",
+    },
+    {
+        name: "LFM2.5-8B-A1B",
+        pullRef: "lfm2.5:8b-a1b-q4_K_M",
+        role: "Native-tool MoE worker / router cheap tier",
+        toolCalling: true,
+        assignRole: "router-cheap",
+        rationale: "Liquid AI's sparse MoE candidate is designed for agentic workflows. Pi probes native tool-calling first and only calibrates the phone protocol after native failure; /fitness must pass the task-scale tool lane on this host before assignment.",
+        note: "Q4_K_M ≈ 5.2 GB, 8.3B total / 1.5B active parameters, with a 128K-class model context. On a 10 GB host run it as the only large local resident; Pi still applies its bounded host profile and context budget.",
     },
     {
         name: "Bonsai-4B (GGUF Q1_0)",
@@ -49,7 +59,7 @@ export const DEFAULT_MODEL_SUGGESTIONS = [
         toolCalling: false,
         assignRole: "curator",
         rationale: "Lighter Bonsai lane model for context curation and structured local analysis; the exact Q1_0 artifact keeps /models suggest one-step.",
-        note: "Recommended when Bonsai-8B is too heavy. Pi pulls this exact GGUF through managed Ollama, probes /fitness on the host, then offers the curator role.",
+        note: "Recommended when Bonsai-8B is too heavy. Pi serves the exact Q1_0 GGUF through its bounded Prism llama.cpp adapter, probes /fitness on the host, then offers the curator role.",
     },
     {
         name: "Bonsai-27B (1-bit + vision)",
@@ -57,34 +67,34 @@ export const DEFAULT_MODEL_SUGGESTIONS = [
         role: "Context curator / reflex brain / vision-capable lane analyst",
         toolCalling: false,
         assignRole: "curator",
-        rationale: "External 1-bit 27B, binary weights, 262K-class context, vision via mmproj; served by pi through prism-ml's llama.cpp build (stock llama.cpp/Ollama cannot serve these weights). /fitness on your hardware is the validator — never pre-claim tool-calling.",
+        rationale: "External 1-bit 27B, binary weights, 262K-class context, vision via mmproj; served through Pi's pinned Prism llama.cpp adapter so runtime, artifact, and projector stay one validated unit. /fitness on your hardware is the validator — never pre-claim tool-calling.",
         note: "~3.8 GB weights + 629 MB vision projector + a one-time llama-server runtime download; CPU-served on GPU-less hosts. Pi installs the pinned prism llama.cpp runtime, downloads both files, starts llama-server, and registers the model in one step. Provider-recommended sampling is temp 0.7 / top-p 0.95 / top-k 20; models.json has no per-model default-sampling field yet, so pi cannot wire this — set it per-request/session if your workflow needs it.",
     },
     {
         name: "Ternary-Bonsai-1.7B",
-        pullRef: "hf.co/prism-ml/Ternary-Bonsai-1.7B-gguf",
+        pullRef: "hf.co/prism-ml/Ternary-Bonsai-1.7B-gguf:Q2_0",
         role: "Search scout (heavy-lifter)",
         toolCalling: false,
         rationale: "Ternary weights, very fast; strong at structured search plans. No tool-calling template — use it as a research/worker lane model, never an executor.",
-        note: "Pick a GGUF quant the runtime accepts (e.g. :Q8_0). The ternary Q2_0 build needs prism-ml's patched llama.cpp.",
+        note: "Pi pins the provider's group-128 Q2_0 artifact and its matched Prism llama.cpp runtime; no ambiguous quant selection is left to Ollama.",
     },
     {
         name: "Ternary-Bonsai-4B",
-        pullRef: "hf.co/prism-ml/Ternary-Bonsai-4B-gguf",
+        pullRef: "hf.co/prism-ml/Ternary-Bonsai-4B-gguf:Q2_0",
         role: "Context curator / reflex brain / lane analyst",
         toolCalling: false,
         assignRole: "curator",
         rationale: "Validated as the 'brain': strict-JSON interpretation and faithful digests. Drives context curation and the toolkit reflex interpreter. Not a tool-caller.",
-        note: "Pick a GGUF quant the runtime accepts (e.g. :Q8_0).",
+        note: "Pi pins the provider's group-128 Q2_0 artifact and its matched Prism llama.cpp runtime.",
     },
     {
         name: "Ternary-Bonsai-8B",
-        pullRef: "hf.co/prism-ml/Ternary-Bonsai-8B-gguf",
+        pullRef: "hf.co/prism-ml/Ternary-Bonsai-8B-gguf:Q2_0",
         role: "Routing judge (larger machines)",
         toolCalling: false,
         assignRole: "judge",
         rationale: "A judge candidate for machines with more headroom — too slow on ~16GB-class hardware in this research, kept for a bigger box.",
-        note: "Heavy: confirm tok/s with /fitness before committing. Ternary quant may need prism-ml's patched llama.cpp.",
+        note: "Heavy: confirm tok/s with /fitness before assigning it. Pi pins the Q2_0 artifact and matched Prism llama.cpp runtime.",
     },
     {
         name: "needle (function-call tester, 26M)",

package/dist/core/models/default-model-suggestions.js.map

@@ -1 +1 @@
-{"version":3,"file":"default-model-suggestions.js","sourceRoot":"","sources":["../../../src/core/models/default-model-suggestions.ts"],"names":[],"mappings":"AAiCA,MAAM,CAAC,MAAM,yBAAyB,GAA+B;IACpE;QACC,IAAI,EAAE,YAAY;QAClB,OAAO,EAAE,YAAY;QACrB,IAAI,EAAE,kCAAkC;QACxC,WAAW,EAAE,IAAI;QACjB,UAAU,EAAE,UAAU;QACtB,SAAS,EACR,0HAA0H;KAC3H;IACD;QACC,IAAI,EAAE,YAAY;QAClB,OAAO,EAAE,YAAY;QACrB,IAAI,EAAE,4BAA4B;QAClC,WAAW,EAAE,IAAI;QACjB,UAAU,EAAE,UAAU;QACtB,SAAS,EACR,4IAA4I;KAC7I;IACD;QACC,IAAI,EAAE,yBAAyB;QAC/B,OAAO,EAAE,2BAA2B;QACpC,IAAI,EAAE,qDAAqD;QAC3D,WAAW,EAAE,IAAI;QACjB,UAAU,EAAE,UAAU;QACtB,SAAS,EACR,wMAAwM;QACzM,IAAI,EAAE,oIAAoI;KAC1I;IACD;QACC,IAAI,EAAE,oBAAoB;QAC1B,OAAO,EAAE,qFAAqF;QAC9F,IAAI,EAAE,kCAAkC;QACxC,WAAW,EAAE,IAAI;QACjB,UAAU,EAAE,OAAO;QACnB,SAAS,EACR,gOAAgO;QACjO,IAAI,EAAE,2IAA2I;KACjJ;IACD;QACC,IAAI,EAAE,eAAe;QACrB,OAAO,EAAE,kDAAkD;QAC3D,IAAI,EAAE,2CAA2C;QACjD,WAAW,EAAE,IAAI;QACjB,UAAU,EAAE,cAAc;QAC1B,SAAS,EACR,sOAAsO;QACvO,IAAI,EAAE,6SAA6S;KACnT;IACD;QACC,IAAI,EAAE,uBAAuB;QAC7B,OAAO,EAAE,oCAAoC;QAC7C,IAAI,EAAE,uDAAuD;QAC7D,WAAW,EAAE,KAAK;QAClB,UAAU,EAAE,SAAS;QACrB,SAAS,EACR,uIAAuI;QACxI,IAAI,EAAE,sJAAsJ;KAC5J;IACD;QACC,IAAI,EAAE,6BAA6B;QACnC,OAAO,EAAE,qCAAqC;QAC9C,IAAI,EAAE,8DAA8D;QACpE,WAAW,EAAE,KAAK;QAClB,UAAU,EAAE,SAAS;QACrB,SAAS,EACR,4PAA4P;QAC7P,IAAI,EAAE,gcAAgc;KACtc;IACD;QACC,IAAI,EAAE,qBAAqB;QAC3B,OAAO,EAAE,yCAAyC;QAClD,IAAI,EAAE,6BAA6B;QACnC,WAAW,EAAE,KAAK;QAClB,SAAS,EACR,sJAAsJ;QACvJ,IAAI,EAAE,gHAAgH;KACtH;IACD;QACC,IAAI,EAAE,mBAAmB;QACzB,OAAO,EAAE,uCAAuC;QAChD,IAAI,EAAE,+CAA+C;QACrD,WAAW,EAAE,KAAK;QAClB,UAAU,EAAE,SAAS;QACrB,SAAS,EACR,2JAA2J;QAC5J,IAAI,EAAE,qDAAqD;KAC3D;IACD;QACC,IAAI,EAAE,mBAAmB;QACzB,OAAO,EAAE,uCAAuC;QAChD,IAAI,EAAE,iCAAiC;QACvC,WAAW,EAAE,KAAK;QAClB,UAAU,EAAE,OAAO;QACnB,SAAS,EACR,+HAA+H;QAChI,IAAI,EAAE,4GAA4G;KAClH;IACD;QACC,IAAI,EAAE,oCAAoC;QAC1C,OAAO,EAAE,6BAA6B;QACtC,IAAI,EAAE,6DAA6D;QACnE,WAAW,EAAE,KAAK;QAClB,SAAS,EACR,maAAma;QACpa,IAAI,EAAE,kMAAkM;KACxM;CACD,CAAC;AAEF,gFAAgF;AAChF,MAAM,UAAU,sBAAsB,CAAC,WAAW,GAA+B,yBAAyB;IACzG,MAAM,KAAK,GAAG;QACb,oHAAoH;KACpH,CAAC;IACF,KAAK,MAAM,UAAU,IAAI,WAAW,EAAE,CAAC;QACtC,KAAK,CAAC,IAAI,CACT,OAAO,UAAU,CAAC,IAAI,MAAM,UAAU,CAAC,IAAI,GAAG,UAAU,CAAC,WAAW,CAAC,CAAC,CAAC,EAAE,CAAC,CAAC,CAAC,oBAAoB,EAAE,EAClG,qBAAqB,UAAU,CAAC,OAAO,EAAE,EACzC,SAAS,UAAU,CAAC,SAAS,EAAE,CAC/B,CAAC;QACF,IAAI,UAAU,CAAC,IAAI;YAAE,KAAK,CAAC,IAAI,CAAC,eAAe,UAAU,CAAC,IAAI,EAAE,CAAC,CAAC;IACnE,CAAC;IACD,KAAK,CAAC,IAAI,CAAC,yFAAyF,CAAC,CAAC;IACtG,OAAO,KAAK,CAAC;AACd,CAAC","sourcesContent":["import type { FitnessRole } from \"../../modes/interactive/components/fitness-role-selector.ts\";\n\n/**\n * Curated local-model suggestions: a starting roster validated during pi-adaptative's own\n * small-model research so a user does not have to know WHICH model fits WHICH role. Each entry is\n * an install ref (accepted by /models add) plus the role it was validated for and whether it can use\n * tools. Native tool-calling remains preferred when it works; text-protocol models are listed only\n * when their runtime is pi-managed and the fitness/probe path can calibrate that fallback.\n *\n * Honesty: these are SUGGESTIONS, not guarantees. Fitness is a property of the model AND the host\n * hardware, so no per-model score is baked in here — `/models add` auto-probes on the actual\n * machine, and the roster only encodes what role each model was SHAPED for. `assignRole` is the\n * /fitness role the suggestion maps to (undefined = no single-setting slot yet; use it as a\n * research/worker lane model).\n */\n\nexport interface ModelSuggestion {\n\t/** Display name in the suggestions list. */\n\tname: string;\n\t/** Ref accepted by /models add and the source normalizer. */\n\tpullRef: string;\n\t/** Role this model was validated/shaped for. */\n\trole: string;\n\t/** True if the model can use tools after host-local probing/calibration (required for executor). */\n\ttoolCalling: boolean;\n\t/** The /fitness assignment target this suggestion maps to, when there is a single-setting slot. */\n\tassignRole?: FitnessRole;\n\t/** One-line rationale from the validation work; never a fabricated numeric score. */\n\trationale: string;\n\t/** Optional caveat (quant selection, RAM needs). */\n\tnote?: string;\n}\n\nexport const DEFAULT_MODEL_SUGGESTIONS: readonly ModelSuggestion[] = [\n\t{\n\t\tname: \"qwen3:1.7b\",\n\t\tpullRef: \"qwen3:1.7b\",\n\t\trole: \"Toolkit executor / reflex muscle\",\n\t\ttoolCalling: true,\n\t\tassignRole: \"executor\",\n\t\trationale:\n\t\t\t\"Validated as the executor: reliable native tool-calling and low latency, so direct toolkit commands run without retries.\",\n\t},\n\t{\n\t\tname: \"qwen3:0.6b\",\n\t\tpullRef: \"qwen3:0.6b\",\n\t\trole: \"Minimal executor (fastest)\",\n\t\ttoolCalling: true,\n\t\tassignRole: \"executor\",\n\t\trationale:\n\t\t\t\"Fastest local option; on harder requests it can narrate without executing, so prefer it only when speed dominates and requests are simple.\",\n\t},\n\t{\n\t\tname: \"MiniCPM5-1B (full-base)\",\n\t\tpullRef: \"hf.co/openbmb/MiniCPM5-1B\",\n\t\trole: \"Full-base Transformers executor / tiny local muscle\",\n\t\ttoolCalling: true,\n\t\tassignRole: \"executor\",\n\t\trationale:\n\t\t\t\"Full-base Hugging Face MiniCPM5-1B target: pi installs an isolated Transformers runtime and probes native tool-calling first, then calibrates the text protocol only if native tool calls do not work.\",\n\t\tnote: \"Not GGUF/quantized: weights and Python deps land under pi-owned runtime/cache directories, separate from Ollama and system Python.\",\n\t},\n\t{\n\t\tname: \"FastContext-1.0-4B\",\n\t\tpullRef: \"hf.co/KikoCis/FastContext-1.0-4B-longctx-imatrix-GGUF:fastcontext4b.Q4_K_M.imx.gguf\",\n\t\trole: \"Repository scout (context_scout)\",\n\t\ttoolCalling: true,\n\t\tassignRole: \"scout\",\n\t\trationale:\n\t\t\t\"The model the context_scout lane is built around (Plan 9): 4B Qwen3, 256K context, trained for read-only repo exploration returning file:line citations. /fitness must pass the tool-calls + research lanes before assignment.\",\n\t\tnote: \"Q4 ≈ 2.5 GB; ~5-6 GB peak with 32-64K KV. Qwen3 template caveat: if tool calls fail the probe, use the Modelfile recipe in docs/scout.md.\",\n\t},\n\t{\n\t\tname: \"Ornith-1.0-9B\",\n\t\tpullRef: \"hf.co/deepreinforce-ai/Ornith-1.0-9B-GGUF:Q4_K_M\",\n\t\trole: \"Agentic-coding worker / router cheap tier\",\n\t\ttoolCalling: true,\n\t\tassignRole: \"router-cheap\",\n\t\trationale:\n\t\t\t\"External candidate (not from pi's own validation research): MIT, Qwen 3.5 base, RL-trained for agentic coding with native tool-calling — the strongest local worker SHAPE in the roster. /fitness on your hardware is the validator.\",\n\t\tnote: \"Q4_K_M ≈ 5.6 GB weights, ~7-8 GB peak with KV — on a 10 GB box run it as the ONLY local model. Qwen 3.5 arch: confirm pi's pinned Ollama supports it and probe tool-calls with /fitness before assigning (template derives from the GGUF). Larger boxes: consider router-medium after a passed worker lane.\",\n\t},\n\t{\n\t\tname: \"Bonsai-4B (GGUF Q1_0)\",\n\t\tpullRef: \"hf.co/prism-ml/Bonsai-4B-gguf:Q1_0\",\n\t\trole: \"Context curator / reflex brain / lighter lane analyst\",\n\t\ttoolCalling: false,\n\t\tassignRole: \"curator\",\n\t\trationale:\n\t\t\t\"Lighter Bonsai lane model for context curation and structured local analysis; the exact Q1_0 artifact keeps /models suggest one-step.\",\n\t\tnote: \"Recommended when Bonsai-8B is too heavy. Pi pulls this exact GGUF through managed Ollama, probes /fitness on the host, then offers the curator role.\",\n\t},\n\t{\n\t\tname: \"Bonsai-27B (1-bit + vision)\",\n\t\tpullRef: \"hf.co/prism-ml/Bonsai-27B-gguf:Q1_0\",\n\t\trole: \"Context curator / reflex brain / vision-capable lane analyst\",\n\t\ttoolCalling: false,\n\t\tassignRole: \"curator\",\n\t\trationale:\n\t\t\t\"External 1-bit 27B, binary weights, 262K-class context, vision via mmproj; served by pi through prism-ml's llama.cpp build (stock llama.cpp/Ollama cannot serve these weights). /fitness on your hardware is the validator — never pre-claim tool-calling.\",\n\t\tnote: \"~3.8 GB weights + 629 MB vision projector + a one-time llama-server runtime download; CPU-served on GPU-less hosts. Pi installs the pinned prism llama.cpp runtime, downloads both files, starts llama-server, and registers the model in one step. Provider-recommended sampling is temp 0.7 / top-p 0.95 / top-k 20; models.json has no per-model default-sampling field yet, so pi cannot wire this — set it per-request/session if your workflow needs it.\",\n\t},\n\t{\n\t\tname: \"Ternary-Bonsai-1.7B\",\n\t\tpullRef: \"hf.co/prism-ml/Ternary-Bonsai-1.7B-gguf\",\n\t\trole: \"Search scout (heavy-lifter)\",\n\t\ttoolCalling: false,\n\t\trationale:\n\t\t\t\"Ternary weights, very fast; strong at structured search plans. No tool-calling template — use it as a research/worker lane model, never an executor.\",\n\t\tnote: \"Pick a GGUF quant the runtime accepts (e.g. :Q8_0). The ternary Q2_0 build needs prism-ml's patched llama.cpp.\",\n\t},\n\t{\n\t\tname: \"Ternary-Bonsai-4B\",\n\t\tpullRef: \"hf.co/prism-ml/Ternary-Bonsai-4B-gguf\",\n\t\trole: \"Context curator / reflex brain / lane analyst\",\n\t\ttoolCalling: false,\n\t\tassignRole: \"curator\",\n\t\trationale:\n\t\t\t\"Validated as the 'brain': strict-JSON interpretation and faithful digests. Drives context curation and the toolkit reflex interpreter. Not a tool-caller.\",\n\t\tnote: \"Pick a GGUF quant the runtime accepts (e.g. :Q8_0).\",\n\t},\n\t{\n\t\tname: \"Ternary-Bonsai-8B\",\n\t\tpullRef: \"hf.co/prism-ml/Ternary-Bonsai-8B-gguf\",\n\t\trole: \"Routing judge (larger machines)\",\n\t\ttoolCalling: false,\n\t\tassignRole: \"judge\",\n\t\trationale:\n\t\t\t\"A judge candidate for machines with more headroom — too slow on ~16GB-class hardware in this research, kept for a bigger box.\",\n\t\tnote: \"Heavy: confirm tok/s with /fitness before committing. Ternary quant may need prism-ml's patched llama.cpp.\",\n\t},\n\t{\n\t\tname: \"needle (function-call tester, 26M)\",\n\t\tpullRef: \"hf.co/Cactus-Compute/needle\",\n\t\trole: \"Standalone function-call test bench (not a chat/lane model)\",\n\t\ttoolCalling: false,\n\t\trationale:\n\t\t\t\"NOT a pi chat model, executor, or lane model — needle is a 26M-parameter single-shot function-call SPECIALIST invoked one query at a time through its own `needle run` CLI, useful only as a standalone bench for probing tool-call behavior in isolation. No OpenAI-compatible endpoint, no chat lanes, no models.json registration, and no /fitness probe: use /models needle <query> [tools-json] to test it directly instead.\",\n\t\tnote: \"Upstream weights (needle.pkl) are pickle-only — arbitrary code execution on load by construction, not something pi's pinned-sha256 verification makes safe in the abstract, only tamper-evident.\",\n\t},\n];\n\n/** Bounded plain-text roster for `/models suggest` and the empty-store hint. */\nexport function formatModelSuggestions(suggestions: readonly ModelSuggestion[] = DEFAULT_MODEL_SUGGESTIONS): string[] {\n\tconst lines = [\n\t\t\"Suggested local models (validated roles from pi's own small-model research; probe on YOUR hardware with /fitness):\",\n\t];\n\tfor (const suggestion of suggestions) {\n\t\tlines.push(\n\t\t\t`  - ${suggestion.name} → ${suggestion.role}${suggestion.toolCalling ? \"\" : \" [no tool-calling]\"}`,\n\t\t\t`      /models add ${suggestion.pullRef}`,\n\t\t\t`      ${suggestion.rationale}`,\n\t\t);\n\t\tif (suggestion.note) lines.push(`      note: ${suggestion.note}`);\n\t}\n\tlines.push(\"Add one with /models add <ref> — pi pulls it, probes it, and offers a role in one step.\");\n\treturn lines;\n}\n"]}
+{"version":3,"file":"default-model-suggestions.js","sourceRoot":"","sources":["../../../src/core/models/default-model-suggestions.ts"],"names":[],"mappings":"AACA,OAAO,EAAE,oBAAoB,EAAE,MAAM,2BAA2B,CAAC;AAiCjE,MAAM,CAAC,MAAM,yBAAyB,GAA+B;IACpE;QACC,IAAI,EAAE,YAAY;QAClB,OAAO,EAAE,YAAY;QACrB,IAAI,EAAE,kCAAkC;QACxC,WAAW,EAAE,IAAI;QACjB,UAAU,EAAE,UAAU;QACtB,SAAS,EACR,0HAA0H;KAC3H;IACD;QACC,IAAI,EAAE,YAAY;QAClB,OAAO,EAAE,YAAY;QACrB,IAAI,EAAE,4BAA4B;QAClC,WAAW,EAAE,IAAI;QACjB,UAAU,EAAE,UAAU;QACtB,SAAS,EACR,4IAA4I;KAC7I;IACD;QACC,IAAI,EAAE,yBAAyB;QAC/B,OAAO,EAAE,2BAA2B;QACpC,IAAI,EAAE,qDAAqD;QAC3D,WAAW,EAAE,IAAI;QACjB,UAAU,EAAE,UAAU;QACtB,SAAS,EACR,wMAAwM;QACzM,IAAI,EAAE,oIAAoI;KAC1I;IACD;QACC,IAAI,EAAE,oBAAoB;QAC1B,OAAO,EAAE,qFAAqF;QAC9F,IAAI,EAAE,kCAAkC;QACxC,WAAW,EAAE,IAAI;QACjB,UAAU,EAAE,OAAO;QACnB,SAAS,EACR,gOAAgO;QACjO,IAAI,EAAE,2IAA2I;KACjJ;IACD;QACC,IAAI,EAAE,eAAe;QACrB,OAAO,EAAE,oBAAoB;QAC7B,IAAI,EAAE,2CAA2C;QACjD,WAAW,EAAE,IAAI;QACjB,UAAU,EAAE,cAAc;QAC1B,SAAS,EACR,sOAAsO;QACvO,IAAI,EAAE,6QAA6Q;KACnR;IACD;QACC,IAAI,EAAE,eAAe;QACrB,OAAO,EAAE,sBAAsB;QAC/B,IAAI,EAAE,4CAA4C;QAClD,WAAW,EAAE,IAAI;QACjB,UAAU,EAAE,cAAc;QAC1B,SAAS,EACR,kPAAkP;QACnP,IAAI,EAAE,+MAA+M;KACrN;IACD;QACC,IAAI,EAAE,uBAAuB;QAC7B,OAAO,EAAE,oCAAoC;QAC7C,IAAI,EAAE,uDAAuD;QAC7D,WAAW,EAAE,KAAK;QAClB,UAAU,EAAE,SAAS;QACrB,SAAS,EACR,uIAAuI;QACxI,IAAI,EAAE,gLAAgL;KACtL;IACD;QACC,IAAI,EAAE,6BAA6B;QACnC,OAAO,EAAE,qCAAqC;QAC9C,IAAI,EAAE,8DAA8D;QACpE,WAAW,EAAE,KAAK;QAClB,UAAU,EAAE,SAAS;QACrB,SAAS,EACR,uQAAuQ;QACxQ,IAAI,EAAE,gcAAgc;KACtc;IACD;QACC,IAAI,EAAE,qBAAqB;QAC3B,OAAO,EAAE,8CAA8C;QACvD,IAAI,EAAE,6BAA6B;QACnC,WAAW,EAAE,KAAK;QAClB,SAAS,EACR,sJAAsJ;QACvJ,IAAI,EAAE,yIAAyI;KAC/I;IACD;QACC,IAAI,EAAE,mBAAmB;QACzB,OAAO,EAAE,4CAA4C;QACrD,IAAI,EAAE,+CAA+C;QACrD,WAAW,EAAE,KAAK;QAClB,UAAU,EAAE,SAAS;QACrB,SAAS,EACR,2JAA2J;QAC5J,IAAI,EAAE,yFAAyF;KAC/F;IACD;QACC,IAAI,EAAE,mBAAmB;QACzB,OAAO,EAAE,4CAA4C;QACrD,IAAI,EAAE,iCAAiC;QACvC,WAAW,EAAE,KAAK;QAClB,UAAU,EAAE,OAAO;QACnB,SAAS,EACR,+HAA+H;QAChI,IAAI,EAAE,wHAAwH;KAC9H;IACD;QACC,IAAI,EAAE,oCAAoC;QAC1C,OAAO,EAAE,6BAA6B;QACtC,IAAI,EAAE,6DAA6D;QACnE,WAAW,EAAE,KAAK;QAClB,SAAS,EACR,maAAma;QACpa,IAAI,EAAE,kMAAkM;KACxM;CACD,CAAC;AAEF,gFAAgF;AAChF,MAAM,UAAU,sBAAsB,CAAC,WAAW,GAA+B,yBAAyB;IACzG,MAAM,KAAK,GAAG;QACb,oHAAoH;KACpH,CAAC;IACF,KAAK,MAAM,UAAU,IAAI,WAAW,EAAE,CAAC;QACtC,KAAK,CAAC,IAAI,CACT,OAAO,UAAU,CAAC,IAAI,MAAM,UAAU,CAAC,IAAI,GAAG,UAAU,CAAC,WAAW,CAAC,CAAC,CAAC,EAAE,CAAC,CAAC,CAAC,oBAAoB,EAAE,EAClG,qBAAqB,UAAU,CAAC,OAAO,EAAE,EACzC,SAAS,UAAU,CAAC,SAAS,EAAE,CAC/B,CAAC;QACF,IAAI,UAAU,CAAC,IAAI;YAAE,KAAK,CAAC,IAAI,CAAC,eAAe,UAAU,CAAC,IAAI,EAAE,CAAC,CAAC;IACnE,CAAC;IACD,KAAK,CAAC,IAAI,CAAC,yFAAyF,CAAC,CAAC;IACtG,OAAO,KAAK,CAAC;AACd,CAAC","sourcesContent":["import type { FitnessRole } from \"../../modes/interactive/components/fitness-role-selector.ts\";\nimport { ORNITH_9B_OLLAMA_REF } from \"./managed-ollama-model.ts\";\n\n/**\n * Curated local-model suggestions: a starting roster validated during pi-adaptative's own\n * small-model research so a user does not have to know WHICH model fits WHICH role. Each entry is\n * an install ref (accepted by /models add) plus the role it was validated for and whether it can use\n * tools. Native tool-calling remains preferred when it works; text-protocol models are listed only\n * when their runtime is pi-managed and the fitness/probe path can calibrate that fallback.\n *\n * Honesty: these are SUGGESTIONS, not guarantees. Fitness is a property of the model AND the host\n * hardware, so no per-model score is baked in here — `/models add` auto-probes on the actual\n * machine, and the roster only encodes what role each model was SHAPED for. `assignRole` is the\n * /fitness role the suggestion maps to (undefined = no single-setting slot yet; use it as a\n * research/worker lane model).\n */\n\nexport interface ModelSuggestion {\n\t/** Display name in the suggestions list. */\n\tname: string;\n\t/** Ref accepted by /models add and the source normalizer. */\n\tpullRef: string;\n\t/** Role this model was validated/shaped for. */\n\trole: string;\n\t/** True if the model can use tools after host-local probing/calibration (required for executor). */\n\ttoolCalling: boolean;\n\t/** The /fitness assignment target this suggestion maps to, when there is a single-setting slot. */\n\tassignRole?: FitnessRole;\n\t/** One-line rationale from the validation work; never a fabricated numeric score. */\n\trationale: string;\n\t/** Optional caveat (quant selection, RAM needs). */\n\tnote?: string;\n}\n\nexport const DEFAULT_MODEL_SUGGESTIONS: readonly ModelSuggestion[] = [\n\t{\n\t\tname: \"qwen3:1.7b\",\n\t\tpullRef: \"qwen3:1.7b\",\n\t\trole: \"Toolkit executor / reflex muscle\",\n\t\ttoolCalling: true,\n\t\tassignRole: \"executor\",\n\t\trationale:\n\t\t\t\"Validated as the executor: reliable native tool-calling and low latency, so direct toolkit commands run without retries.\",\n\t},\n\t{\n\t\tname: \"qwen3:0.6b\",\n\t\tpullRef: \"qwen3:0.6b\",\n\t\trole: \"Minimal executor (fastest)\",\n\t\ttoolCalling: true,\n\t\tassignRole: \"executor\",\n\t\trationale:\n\t\t\t\"Fastest local option; on harder requests it can narrate without executing, so prefer it only when speed dominates and requests are simple.\",\n\t},\n\t{\n\t\tname: \"MiniCPM5-1B (full-base)\",\n\t\tpullRef: \"hf.co/openbmb/MiniCPM5-1B\",\n\t\trole: \"Full-base Transformers executor / tiny local muscle\",\n\t\ttoolCalling: true,\n\t\tassignRole: \"executor\",\n\t\trationale:\n\t\t\t\"Full-base Hugging Face MiniCPM5-1B target: pi installs an isolated Transformers runtime and probes native tool-calling first, then calibrates the text protocol only if native tool calls do not work.\",\n\t\tnote: \"Not GGUF/quantized: weights and Python deps land under pi-owned runtime/cache directories, separate from Ollama and system Python.\",\n\t},\n\t{\n\t\tname: \"FastContext-1.0-4B\",\n\t\tpullRef: \"hf.co/KikoCis/FastContext-1.0-4B-longctx-imatrix-GGUF:fastcontext4b.Q4_K_M.imx.gguf\",\n\t\trole: \"Repository scout (context_scout)\",\n\t\ttoolCalling: true,\n\t\tassignRole: \"scout\",\n\t\trationale:\n\t\t\t\"The model the context_scout lane is built around (Plan 9): 4B Qwen3, 256K context, trained for read-only repo exploration returning file:line citations. /fitness must pass the tool-calls + research lanes before assignment.\",\n\t\tnote: \"Q4 ≈ 2.5 GB; ~5-6 GB peak with 32-64K KV. Qwen3 template caveat: if tool calls fail the probe, use the Modelfile recipe in docs/scout.md.\",\n\t},\n\t{\n\t\tname: \"Ornith-1.0-9B\",\n\t\tpullRef: ORNITH_9B_OLLAMA_REF,\n\t\trole: \"Agentic-coding worker / router cheap tier\",\n\t\ttoolCalling: true,\n\t\tassignRole: \"router-cheap\",\n\t\trationale:\n\t\t\t\"External candidate (not from pi's own validation research): MIT, Qwen 3.5 base, RL-trained for agentic coding with native tool-calling — the strongest local worker SHAPE in the roster. /fitness on your hardware is the validator.\",\n\t\tnote: \"Q4_K_M ≈ 5.6 GB weights, ~7-8 GB peak with KV — on a 10 GB box run it as the ONLY local model. Pi creates a managed Qwen 3.5 renderer/parser profile with the model's recommended sampling before probing. Larger boxes: consider router-medium after a passed worker lane.\",\n\t},\n\t{\n\t\tname: \"LFM2.5-8B-A1B\",\n\t\tpullRef: \"lfm2.5:8b-a1b-q4_K_M\",\n\t\trole: \"Native-tool MoE worker / router cheap tier\",\n\t\ttoolCalling: true,\n\t\tassignRole: \"router-cheap\",\n\t\trationale:\n\t\t\t\"Liquid AI's sparse MoE candidate is designed for agentic workflows. Pi probes native tool-calling first and only calibrates the phone protocol after native failure; /fitness must pass the task-scale tool lane on this host before assignment.\",\n\t\tnote: \"Q4_K_M ≈ 5.2 GB, 8.3B total / 1.5B active parameters, with a 128K-class model context. On a 10 GB host run it as the only large local resident; Pi still applies its bounded host profile and context budget.\",\n\t},\n\t{\n\t\tname: \"Bonsai-4B (GGUF Q1_0)\",\n\t\tpullRef: \"hf.co/prism-ml/Bonsai-4B-gguf:Q1_0\",\n\t\trole: \"Context curator / reflex brain / lighter lane analyst\",\n\t\ttoolCalling: false,\n\t\tassignRole: \"curator\",\n\t\trationale:\n\t\t\t\"Lighter Bonsai lane model for context curation and structured local analysis; the exact Q1_0 artifact keeps /models suggest one-step.\",\n\t\tnote: \"Recommended when Bonsai-8B is too heavy. Pi serves the exact Q1_0 GGUF through its bounded Prism llama.cpp adapter, probes /fitness on the host, then offers the curator role.\",\n\t},\n\t{\n\t\tname: \"Bonsai-27B (1-bit + vision)\",\n\t\tpullRef: \"hf.co/prism-ml/Bonsai-27B-gguf:Q1_0\",\n\t\trole: \"Context curator / reflex brain / vision-capable lane analyst\",\n\t\ttoolCalling: false,\n\t\tassignRole: \"curator\",\n\t\trationale:\n\t\t\t\"External 1-bit 27B, binary weights, 262K-class context, vision via mmproj; served through Pi's pinned Prism llama.cpp adapter so runtime, artifact, and projector stay one validated unit. /fitness on your hardware is the validator — never pre-claim tool-calling.\",\n\t\tnote: \"~3.8 GB weights + 629 MB vision projector + a one-time llama-server runtime download; CPU-served on GPU-less hosts. Pi installs the pinned prism llama.cpp runtime, downloads both files, starts llama-server, and registers the model in one step. Provider-recommended sampling is temp 0.7 / top-p 0.95 / top-k 20; models.json has no per-model default-sampling field yet, so pi cannot wire this — set it per-request/session if your workflow needs it.\",\n\t},\n\t{\n\t\tname: \"Ternary-Bonsai-1.7B\",\n\t\tpullRef: \"hf.co/prism-ml/Ternary-Bonsai-1.7B-gguf:Q2_0\",\n\t\trole: \"Search scout (heavy-lifter)\",\n\t\ttoolCalling: false,\n\t\trationale:\n\t\t\t\"Ternary weights, very fast; strong at structured search plans. No tool-calling template — use it as a research/worker lane model, never an executor.\",\n\t\tnote: \"Pi pins the provider's group-128 Q2_0 artifact and its matched Prism llama.cpp runtime; no ambiguous quant selection is left to Ollama.\",\n\t},\n\t{\n\t\tname: \"Ternary-Bonsai-4B\",\n\t\tpullRef: \"hf.co/prism-ml/Ternary-Bonsai-4B-gguf:Q2_0\",\n\t\trole: \"Context curator / reflex brain / lane analyst\",\n\t\ttoolCalling: false,\n\t\tassignRole: \"curator\",\n\t\trationale:\n\t\t\t\"Validated as the 'brain': strict-JSON interpretation and faithful digests. Drives context curation and the toolkit reflex interpreter. Not a tool-caller.\",\n\t\tnote: \"Pi pins the provider's group-128 Q2_0 artifact and its matched Prism llama.cpp runtime.\",\n\t},\n\t{\n\t\tname: \"Ternary-Bonsai-8B\",\n\t\tpullRef: \"hf.co/prism-ml/Ternary-Bonsai-8B-gguf:Q2_0\",\n\t\trole: \"Routing judge (larger machines)\",\n\t\ttoolCalling: false,\n\t\tassignRole: \"judge\",\n\t\trationale:\n\t\t\t\"A judge candidate for machines with more headroom — too slow on ~16GB-class hardware in this research, kept for a bigger box.\",\n\t\tnote: \"Heavy: confirm tok/s with /fitness before assigning it. Pi pins the Q2_0 artifact and matched Prism llama.cpp runtime.\",\n\t},\n\t{\n\t\tname: \"needle (function-call tester, 26M)\",\n\t\tpullRef: \"hf.co/Cactus-Compute/needle\",\n\t\trole: \"Standalone function-call test bench (not a chat/lane model)\",\n\t\ttoolCalling: false,\n\t\trationale:\n\t\t\t\"NOT a pi chat model, executor, or lane model — needle is a 26M-parameter single-shot function-call SPECIALIST invoked one query at a time through its own `needle run` CLI, useful only as a standalone bench for probing tool-call behavior in isolation. No OpenAI-compatible endpoint, no chat lanes, no models.json registration, and no /fitness probe: use /models needle <query> [tools-json] to test it directly instead.\",\n\t\tnote: \"Upstream weights (needle.pkl) are pickle-only — arbitrary code execution on load by construction, not something pi's pinned-sha256 verification makes safe in the abstract, only tamper-evident.\",\n\t},\n];\n\n/** Bounded plain-text roster for `/models suggest` and the empty-store hint. */\nexport function formatModelSuggestions(suggestions: readonly ModelSuggestion[] = DEFAULT_MODEL_SUGGESTIONS): string[] {\n\tconst lines = [\n\t\t\"Suggested local models (validated roles from pi's own small-model research; probe on YOUR hardware with /fitness):\",\n\t];\n\tfor (const suggestion of suggestions) {\n\t\tlines.push(\n\t\t\t`  - ${suggestion.name} → ${suggestion.role}${suggestion.toolCalling ? \"\" : \" [no tool-calling]\"}`,\n\t\t\t`      /models add ${suggestion.pullRef}`,\n\t\t\t`      ${suggestion.rationale}`,\n\t\t);\n\t\tif (suggestion.note) lines.push(`      note: ${suggestion.note}`);\n\t}\n\tlines.push(\"Add one with /models add <ref> — pi pulls it, probes it, and offers a role in one step.\");\n\treturn lines;\n}\n"]}

package/dist/core/models/llamacpp-runtime.d.ts

@@ -1,10 +1,11 @@
 import type { Readable } from "node:stream";
+import { type LocalInferenceProfileMode } from "./local-inference-profile.ts";
 import { type ManagedRuntimeSpawn } from "./runtime-process.ts";
 /**
- * Managed runtime for prism-ml 1-bit GGUF models (Bonsai-27B first). The provider requires their
- * llama.cpp fork (Q1_0_g128 hybrid-attention kernels) — stock llama.cpp/Ollama cannot serve these
- * weights — so this module owns runtime install (prebuilt release download+extract), GGUF
- * downloads, and llama-server lifecycle. Mirrors local-runtime.ts's OllamaRuntime: injectable
+ * Managed runtime for Pi's curated Prism 1-bit and ternary GGUF models. Q1_0 is available in
+ * upstream llama.cpp, while the selected group-128 Q2_0 artifacts and paired DSpark acceleration
+ * still require Prism's validated build. This module therefore owns the pinned runtime install,
+ * GGUF downloads, and llama-server lifecycle. Mirrors local-runtime.ts's OllamaRuntime: injectable
  * seams for fetch/spawn/exists, pi-owned directories under agentDir (runtimes/, models/),
  * detached+tracked child processes, onProgress as best-effort UI feedback, honest error taxonomies
  * instead of silent fallbacks.
@@ -28,23 +29,38 @@ export interface PrismModelDescriptor {
     file: string;
     mmprojFile?: string;
     displayName: string;
+    architecture: "dense";
+    runtime: "prism-llamacpp";
+    family: "bonsai" | "ternary-bonsai";
+    parameterScale: "1.7B" | "4B" | "8B" | "27B";
+    weightFormat: "q1_0" | "q2_0";
+    matchedDrafter?: {
+        kind: "dspark";
+        file: string;
+        draftMax: 4;
+        minimumContext: 16_384;
+        validatedBackend: "cuda";
+    };
 }
-/** Curated first-model descriptor; the `/models add` wiring task consumes this. */
+/** Curated 27B descriptor shared by the full local execution catalog. */
 export declare const BONSAI_27B: PrismModelDescriptor;
 export type PrismBackend = "cpu" | "cuda";
 export type PrismAssetKind = "tar-gz" | "zip";
-export interface PrismLlamaAsset {
+export interface PrismArchiveAsset {
     name: string;
     kind: PrismAssetKind;
+}
+export interface PrismLlamaAsset extends PrismArchiveAsset {
     backend: PrismBackend;
+    companionAssets?: readonly PrismArchiveAsset[];
 }
 /**
  * Maps a platform/arch/GPU triple to the exact Prism llama.cpp release asset for
  * {@link PRISM_LLAMACPP_PINNED_RELEASE} — verified against the real GitHub release, not guessed.
- * CPU asset by default; the CUDA 12.4 variant is only offered for linux x64 with an NVIDIA GPU
- * present (a 12.8 variant also exists upstream but has no wired caller yet). Windows CUDA needs a
- * companion `cudart-*` archive and is out of scope for v1 — Windows always resolves to the CPU zip
- * regardless of `hasNvidiaGpu`. Pure and exported so it's independently testable.
+ * CPU asset by default; CUDA 12.4 is selected for x64 Linux or Windows with an NVIDIA GPU. The
+ * pinned Windows build is incomplete without its matching `cudart-*` archive, so the resolver
+ * returns both as one mandatory installation plan. Pure and exported so it's independently
+ * testable.
  */
 export declare function resolvePrismLlamaAsset(plat: string, architecture: string, hasNvidiaGpu: boolean): PrismLlamaAsset | undefined;
 export interface PrismDetectResult {
@@ -65,6 +81,11 @@ export type PrismServeResult = {
     ok: false;
     error: string;
 };
+export interface PrismProcessLifecycle {
+    track(pid: number): void;
+    untrack(pid: number): void;
+    terminate(pid: number): void;
+}
 type PrismExtractArchiveFn = (input: Readable, destDir: string, kind: PrismAssetKind) => Promise<{
     ok: boolean;
     error?: string;
@@ -81,6 +102,9 @@ export interface PrismLlamaCppDeps {
     hasNvidiaGpu?: () => boolean;
     platform?: () => string;
     arch?: () => string;
+    /** Host capacity probes used once to derive the bounded local inference profile. */
+    totalMemoryBytes?: () => number;
+    logicalCpuCount?: () => number;
     /** Runs the extraction step for a downloaded archive. Injectable so installManaged's
      * download->extract->scan orchestration is testable without a real tar/unzip pipeline; defaults
      * to the real spawn-based extractor. */
@@ -88,6 +112,7 @@ export interface PrismLlamaCppDeps {
     /** Health-poll bounds for serve(); default ~120s (240 * 500ms), both overridable for tests. */
     healthPollAttempts?: number;
     healthPollIntervalMs?: number;
+    processLifecycle?: PrismProcessLifecycle;
 }
 export declare class PrismLlamaCppRuntime {
     private readonly _agentDir;
@@ -102,9 +127,12 @@ export declare class PrismLlamaCppRuntime {
     private readonly _extractArchiveFn;
     private readonly _healthPollAttempts;
     private readonly _healthPollIntervalMs;
+    private readonly _profile;
+    private readonly _processLifecycle;
     private _child;
     constructor(args: {
         agentDir: string;
+        profileMode?: LocalInferenceProfileMode;
         deps?: PrismLlamaCppDeps;
     });
     runtimeDir(): string;
@@ -144,7 +172,6 @@ export declare class PrismLlamaCppRuntime {
         repo: string;
         file: string;
     }, onProgress?: (status: string) => void): Promise<PrismDownloadResult>;
-    private _healthUp;
     /**
      * Spawn `llama-server` detached+tracked (killed on parent shutdown even if pi crashes without
      * calling stop()) and poll `/health` until ready. `-ngl 99` is only added when the installed
@@ -152,6 +179,7 @@ export declare class PrismLlamaCppRuntime {
      */
     serve(args: {
         modelPath: string;
+        modelAlias: string;
         mmprojPath?: string;
         port: number;
         numCtx: number;

package/dist/core/models/llamacpp-runtime.d.ts.map

@@ -1 +1 @@
-{"version":3,"file":"llamacpp-runtime.d.ts","sourceRoot":"","sources":["../../../src/core/models/llamacpp-runtime.ts"],"names":[],"mappings":"AAIA,OAAO,KAAK,EAAE,QAAQ,EAAE,MAAM,aAAa,CAAC;AAK5C,OAAO,EAIN,KAAK,mBAAmB,EAOxB,MAAM,sBAAsB,CAAC;AAE9B;;;;;;;;;;;;;;GAcG;AAEH;;;;;GAKG;AACH,eAAO,MAAM,6BAA6B,wBAAwB,CAAC;AAEnE,eAAO,MAAM,gCAAgC,+DAA+D,CAAC;AAE7G,MAAM,WAAW,oBAAoB;IACpC,IAAI,EAAE,MAAM,CAAC;IACb,IAAI,EAAE,MAAM,CAAC;IACb,UAAU,CAAC,EAAE,MAAM,CAAC;IACpB,WAAW,EAAE,MAAM,CAAC;CACpB;AAED,mFAAmF;AACnF,eAAO,MAAM,UAAU,EAAE,oBAKxB,CAAC;AAEF,MAAM,MAAM,YAAY,GAAG,KAAK,GAAG,MAAM,CAAC;AAC1C,MAAM,MAAM,cAAc,GAAG,QAAQ,GAAG,KAAK,CAAC;AAE9C,MAAM,WAAW,eAAe;IAC/B,IAAI,EAAE,MAAM,CAAC;IACb,IAAI,EAAE,cAAc,CAAC;IACrB,OAAO,EAAE,YAAY,CAAC;CACtB;AAcD;;;;;;;GAOG;AACH,wBAAgB,sBAAsB,CACrC,IAAI,EAAE,MAAM,EACZ,YAAY,EAAE,MAAM,EACpB,YAAY,EAAE,OAAO,GACnB,eAAe,GAAG,SAAS,CAqB7B;AAED,MAAM,WAAW,iBAAiB;IACjC,gBAAgB,EAAE,OAAO,CAAC;IAC1B,UAAU,CAAC,EAAE,MAAM,CAAC;IACpB,OAAO,CAAC,EAAE,MAAM,CAAC;CACjB;AAED,MAAM,WAAW,mBAAmB;IACnC,EAAE,EAAE,OAAO,CAAC;IACZ,IAAI,CAAC,EAAE,MAAM,CAAC;IACd,OAAO,CAAC,EAAE,OAAO,CAAC;IAClB,KAAK,CAAC,EAAE,MAAM,CAAC;CACf;AAED,MAAM,MAAM,gBAAgB,GAAG;IAAE,EAAE,EAAE,IAAI,CAAC;IAAC,OAAO,EAAE,MAAM,CAAA;CAAE,GAAG;IAAE,EAAE,EAAE,KAAK,CAAC;IAAC,KAAK,EAAE,MAAM,CAAA;CAAE,CAAC;AAQ5F,KAAK,qBAAqB,GAAG,CAC5B,KAAK,EAAE,QAAQ,EACf,OAAO,EAAE,MAAM,EACf,IAAI,EAAE,cAAc,KAChB,OAAO,CAAC;IAAE,EAAE,EAAE,OAAO,CAAC;IAAC,KAAK,CAAC,EAAE,MAAM,CAAA;CAAE,CAAC,CAAC;AAE9C,MAAM,WAAW,iBAAiB;IACjC,OAAO,CAAC,EAAE,OAAO,KAAK,CAAC;IACvB,OAAO,CAAC,EAAE,mBAAmB,CAAC;IAC9B,QAAQ,CAAC,EAAE,CAAC,IAAI,EAAE,MAAM,KAAK,OAAO,CAAC;IACrC,OAAO,CAAC,EAAE,CAAC,EAAE,EAAE,MAAM,KAAK,OAAO,CAAC,IAAI,CAAC,CAAC;IACxC,gGAAgG;IAChG,UAAU,CAAC,EAAE,CAAC,OAAO,EAAE,MAAM,KAAK,OAAO,CAAC;IAC1C;8FAC0F;IAC1F,YAAY,CAAC,EAAE,MAAM,OAAO,CAAC;IAC7B,QAAQ,CAAC,EAAE,MAAM,MAAM,CAAC;IACxB,IAAI,CAAC,EAAE,MAAM,MAAM,CAAC;IACpB;;4CAEwC;IACxC,cAAc,CAAC,EAAE,qBAAqB,CAAC;IACvC,+FAA+F;IAC/F,kBAAkB,CAAC,EAAE,MAAM,CAAC;IAC5B,oBAAoB,CAAC,EAAE,MAAM,CAAC;CAC9B;AAeD,qBAAa,oBAAoB;IAChC,OAAO,CAAC,QAAQ,CAAC,SAAS,CAAS;IACnC,OAAO,CAAC,QAAQ,CAAC,MAAM,CAAe;IACtC,OAAO,CAAC,QAAQ,CAAC,MAAM,CAAsB;IAC7C,OAAO,CAAC,QAAQ,CAAC,OAAO,CAA4B;IACpD,OAAO,CAAC,QAAQ,CAAC,MAAM,CAAgC;IACvD,OAAO,CAAC,QAAQ,CAAC,WAAW,CAA+B;IAC3D,OAAO,CAAC,QAAQ,CAAC,aAAa,CAAgB;IAC9C,OAAO,CAAC,QAAQ,CAAC,SAAS,CAAe;IACzC,OAAO,CAAC,QAAQ,CAAC,KAAK,CAAe;IACrC,OAAO,CAAC,QAAQ,CAAC,iBAAiB,CAAwB;IAC1D,OAAO,CAAC,QAAQ,CAAC,mBAAmB,CAAS;IAC7C,OAAO,CAAC,QAAQ,CAAC,qBAAqB,CAAS;IAC/C,OAAO,CAAC,MAAM,CAAkE;IAEhF,YAAY,IAAI,EAAE;QAAE,QAAQ,EAAE,MAAM,CAAC;QAAC,IAAI,CAAC,EAAE,iBAAiB,CAAA;KAAE,EAW/D;IAED,UAAU,IAAI,MAAM,CAEnB;IAED,SAAS,IAAI,MAAM,CAElB;IAED,OAAO,CAAC,WAAW;IAInB,OAAO,CAAC,aAAa;IAIrB,OAAO,CAAC,aAAa;IAoBrB,OAAO,CAAC,cAAc;IAIhB,MAAM,IAAI,OAAO,CAAC,iBAAiB,CAAC,CAUzC;IAED;;;yDAGqD;IACrD,OAAO,CAAC,kBAAkB;IAqB1B;;;;;OAKG;IACG,cAAc,CAAC,UAAU,CAAC,EAAE,CAAC,MAAM,EAAE,MAAM,KAAK,IAAI,GAAG,OAAO,CAAC;QAAE,EAAE,EAAE,OAAO,CAAC;QAAC,KAAK,CAAC,EAAE,MAAM,CAAA;KAAE,CAAC,CAkBpG;IAED,OAAO,CAAC,eAAe;YAeT,eAAe;YAYf,aAAa;YA+Bb,WAAW;YAYX,oBAAoB;IAYlC;;;;;OAKG;IACG,aAAa,CAClB,IAAI,EAAE;QAAE,IAAI,EAAE,MAAM,CAAC;QAAC,IAAI,EAAE,MAAM,CAAA;KAAE,EACpC,UAAU,CAAC,EAAE,CAAC,MAAM,EAAE,MAAM,KAAK,IAAI,GACnC,OAAO,CAAC,mBAAmB,CAAC,CA6B9B;YAEa,SAAS;IAWvB;;;;OAIG;IACG,KAAK,CAAC,IAAI,EAAE;QACjB,SAAS,EAAE,MAAM,CAAC;QAClB,UAAU,CAAC,EAAE,MAAM,CAAC;QACpB,IAAI,EAAE,MAAM,CAAC;QACb,MAAM,EAAE,MAAM,CAAC;KACf,GAAG,OAAO,CAAC,gBAAgB,CAAC,CAsC5B;IAED,IAAI,IAAI;QAAE,OAAO,EAAE,OAAO,CAAA;KAAE,CAS3B;IAED,SAAS,IAAI,OAAO,CAEnB;CACD"}
+{"version":3,"file":"llamacpp-runtime.d.ts","sourceRoot":"","sources":["../../../src/core/models/llamacpp-runtime.ts"],"names":[],"mappings":"AAIA,OAAO,KAAK,EAAE,QAAQ,EAAE,MAAM,aAAa,CAAC;AAK5C,OAAO,EAGN,KAAK,yBAAyB,EAC9B,MAAM,8BAA8B,CAAC;AAEtC,OAAO,EAIN,KAAK,mBAAmB,EAOxB,MAAM,sBAAsB,CAAC;AAE9B;;;;;;;;;;;;;;GAcG;AAEH;;;;;GAKG;AACH,eAAO,MAAM,6BAA6B,wBAAwB,CAAC;AAEnE,eAAO,MAAM,gCAAgC,+DAA+D,CAAC;AAE7G,MAAM,WAAW,oBAAoB;IACpC,IAAI,EAAE,MAAM,CAAC;IACb,IAAI,EAAE,MAAM,CAAC;IACb,UAAU,CAAC,EAAE,MAAM,CAAC;IACpB,WAAW,EAAE,MAAM,CAAC;IACpB,YAAY,EAAE,OAAO,CAAC;IACtB,OAAO,EAAE,gBAAgB,CAAC;IAC1B,MAAM,EAAE,QAAQ,GAAG,gBAAgB,CAAC;IACpC,cAAc,EAAE,MAAM,GAAG,IAAI,GAAG,IAAI,GAAG,KAAK,CAAC;IAC7C,YAAY,EAAE,MAAM,GAAG,MAAM,CAAC;IAC9B,cAAc,CAAC,EAAE;QAChB,IAAI,EAAE,QAAQ,CAAC;QACf,IAAI,EAAE,MAAM,CAAC;QACb,QAAQ,EAAE,CAAC,CAAC;QACZ,cAAc,EAAE,MAAM,CAAC;QACvB,gBAAgB,EAAE,MAAM,CAAC;KACzB,CAAC;CACF;AAED,yEAAyE;AACzE,eAAO,MAAM,UAAU,EAAE,oBAiBxB,CAAC;AAEF,MAAM,MAAM,YAAY,GAAG,KAAK,GAAG,MAAM,CAAC;AAC1C,MAAM,MAAM,cAAc,GAAG,QAAQ,GAAG,KAAK,CAAC;AAE9C,MAAM,WAAW,iBAAiB;IACjC,IAAI,EAAE,MAAM,CAAC;IACb,IAAI,EAAE,cAAc,CAAC;CACrB;AAED,MAAM,WAAW,eAAgB,SAAQ,iBAAiB;IACzD,OAAO,EAAE,YAAY,CAAC;IACtB,eAAe,CAAC,EAAE,SAAS,iBAAiB,EAAE,CAAC;CAC/C;AAgBD;;;;;;;GAOG;AACH,wBAAgB,sBAAsB,CACrC,IAAI,EAAE,MAAM,EACZ,YAAY,EAAE,MAAM,EACpB,YAAY,EAAE,OAAO,GACnB,eAAe,GAAG,SAAS,CA8B7B;AAED,MAAM,WAAW,iBAAiB;IACjC,gBAAgB,EAAE,OAAO,CAAC;IAC1B,UAAU,CAAC,EAAE,MAAM,CAAC;IACpB,OAAO,CAAC,EAAE,MAAM,CAAC;CACjB;AAED,MAAM,WAAW,mBAAmB;IACnC,EAAE,EAAE,OAAO,CAAC;IACZ,IAAI,CAAC,EAAE,MAAM,CAAC;IACd,OAAO,CAAC,EAAE,OAAO,CAAC;IAClB,KAAK,CAAC,EAAE,MAAM,CAAC;CACf;AAED,MAAM,MAAM,gBAAgB,GAAG;IAAE,EAAE,EAAE,IAAI,CAAC;IAAC,OAAO,EAAE,MAAM,CAAA;CAAE,GAAG;IAAE,EAAE,EAAE,KAAK,CAAC;IAAC,KAAK,EAAE,MAAM,CAAA;CAAE,CAAC;AAE5F,MAAM,WAAW,qBAAqB;IACrC,KAAK,CAAC,GAAG,EAAE,MAAM,GAAG,IAAI,CAAC;IACzB,OAAO,CAAC,GAAG,EAAE,MAAM,GAAG,IAAI,CAAC;IAC3B,SAAS,CAAC,GAAG,EAAE,MAAM,GAAG,IAAI,CAAC;CAC7B;AAQD,KAAK,qBAAqB,GAAG,CAC5B,KAAK,EAAE,QAAQ,EACf,OAAO,EAAE,MAAM,EACf,IAAI,EAAE,cAAc,KAChB,OAAO,CAAC;IAAE,EAAE,EAAE,OAAO,CAAC;IAAC,KAAK,CAAC,EAAE,MAAM,CAAA;CAAE,CAAC,CAAC;AAE9C,MAAM,WAAW,iBAAiB;IACjC,OAAO,CAAC,EAAE,OAAO,KAAK,CAAC;IACvB,OAAO,CAAC,EAAE,mBAAmB,CAAC;IAC9B,QAAQ,CAAC,EAAE,CAAC,IAAI,EAAE,MAAM,KAAK,OAAO,CAAC;IACrC,OAAO,CAAC,EAAE,CAAC,EAAE,EAAE,MAAM,KAAK,OAAO,CAAC,IAAI,CAAC,CAAC;IACxC,gGAAgG;IAChG,UAAU,CAAC,EAAE,CAAC,OAAO,EAAE,MAAM,KAAK,OAAO,CAAC;IAC1C;8FAC0F;IAC1F,YAAY,CAAC,EAAE,MAAM,OAAO,CAAC;IAC7B,QAAQ,CAAC,EAAE,MAAM,MAAM,CAAC;IACxB,IAAI,CAAC,EAAE,MAAM,MAAM,CAAC;IACpB,oFAAoF;IACpF,gBAAgB,CAAC,EAAE,MAAM,MAAM,CAAC;IAChC,eAAe,CAAC,EAAE,MAAM,MAAM,CAAC;IAC/B;;4CAEwC;IACxC,cAAc,CAAC,EAAE,qBAAqB,CAAC;IACvC,+FAA+F;IAC/F,kBAAkB,CAAC,EAAE,MAAM,CAAC;IAC5B,oBAAoB,CAAC,EAAE,MAAM,CAAC;IAC9B,gBAAgB,CAAC,EAAE,qBAAqB,CAAC;CACzC;AAmBD,qBAAa,oBAAoB;IAChC,OAAO,CAAC,QAAQ,CAAC,SAAS,CAAS;IACnC,OAAO,CAAC,QAAQ,CAAC,MAAM,CAAe;IACtC,OAAO,CAAC,QAAQ,CAAC,MAAM,CAAsB;IAC7C,OAAO,CAAC,QAAQ,CAAC,OAAO,CAA4B;IACpD,OAAO,CAAC,QAAQ,CAAC,MAAM,CAAgC;IACvD,OAAO,CAAC,QAAQ,CAAC,WAAW,CAA+B;IAC3D,OAAO,CAAC,QAAQ,CAAC,aAAa,CAAgB;IAC9C,OAAO,CAAC,QAAQ,CAAC,SAAS,CAAe;IACzC,OAAO,CAAC,QAAQ,CAAC,KAAK,CAAe;IACrC,OAAO,CAAC,QAAQ,CAAC,iBAAiB,CAAwB;IAC1D,OAAO,CAAC,QAAQ,CAAC,mBAAmB,CAAS;IAC7C,OAAO,CAAC,QAAQ,CAAC,qBAAqB,CAAS;IAC/C,OAAO,CAAC,QAAQ,CAAC,QAAQ,CAAwB;IACjD,OAAO,CAAC,QAAQ,CAAC,iBAAiB,CAAwB;IAC1D,OAAO,CAAC,MAAM,CAAkE;IAEhF,YAAY,IAAI,EAAE;QACjB,QAAQ,EAAE,MAAM,CAAC;QACjB,WAAW,CAAC,EAAE,yBAAyB,CAAC;QACxC,IAAI,CAAC,EAAE,iBAAiB,CAAC;KACzB,EAiBA;IAED,UAAU,IAAI,MAAM,CAEnB;IAED,SAAS,IAAI,MAAM,CAElB;IAED,OAAO,CAAC,WAAW;IAInB,OAAO,CAAC,aAAa;IAIrB,OAAO,CAAC,aAAa;IAoBrB,OAAO,CAAC,cAAc;IAIhB,MAAM,IAAI,OAAO,CAAC,iBAAiB,CAAC,CAUzC;IAED;;;yDAGqD;IACrD,OAAO,CAAC,kBAAkB;IAqB1B;;;;;OAKG;IACG,cAAc,CAAC,UAAU,CAAC,EAAE,CAAC,MAAM,EAAE,MAAM,KAAK,IAAI,GAAG,OAAO,CAAC;QAAE,EAAE,EAAE,OAAO,CAAC;QAAC,KAAK,CAAC,EAAE,MAAM,CAAA;KAAE,CAAC,CAqBpG;IAED,OAAO,CAAC,eAAe;YAeT,eAAe;YAYf,aAAa;YA+Bb,WAAW;YAYX,oBAAoB;IAYlC;;;;;OAKG;IACG,aAAa,CAClB,IAAI,EAAE;QAAE,IAAI,EAAE,MAAM,CAAC;QAAC,IAAI,EAAE,MAAM,CAAA;KAAE,EACpC,UAAU,CAAC,EAAE,CAAC,MAAM,EAAE,MAAM,KAAK,IAAI,GACnC,OAAO,CAAC,mBAAmB,CAAC,CA6B9B;IAED;;;;OAIG;IACG,KAAK,CAAC,IAAI,EAAE;QACjB,SAAS,EAAE,MAAM,CAAC;QAClB,UAAU,EAAE,MAAM,CAAC;QACnB,UAAU,CAAC,EAAE,MAAM,CAAC;QACpB,IAAI,EAAE,MAAM,CAAC;QACb,MAAM,EAAE,MAAM,CAAC;KACf,GAAG,OAAO,CAAC,gBAAgB,CAAC,CAmE5B;IAED,IAAI,IAAI;QAAE,OAAO,EAAE,OAAO,CAAA;KAAE,CAS3B;IAED,SAAS,IAAI,OAAO,CAEnB;CACD"}

package/dist/core/models/llamacpp-runtime.js

@@ -5,12 +5,14 @@ import { pipeline } from "node:stream/promises";
 import { spawnProcess, waitForChildProcessWithTermination } from "../../utils/child-process.js";
 import { killProcessTree, trackDetachedChildPid, untrackDetachedChildPid } from "../../utils/shell.js";
 import { modelsDir as agentModelsDir, runtimesDir as agentRuntimesDir } from "../agent-paths.js";
+import { deriveHostLocalInferenceProfile, } from "./local-inference-profile.js";
+import { probePrismLlamaCppServer } from "./prism-llamacpp-server-probe.js";
 import { extractZipArchive, fetchRuntimeDownload, installRuntimeArchive, removePartialDownload, requireRuntimeStdin, resolveRuntimeLifecycleDependencies, runtimeCommandAvailable, tryFileSizeBytes, writeRuntimeDownload, } from "./runtime-process.js";
 /**
- * Managed runtime for prism-ml 1-bit GGUF models (Bonsai-27B first). The provider requires their
- * llama.cpp fork (Q1_0_g128 hybrid-attention kernels) — stock llama.cpp/Ollama cannot serve these
- * weights — so this module owns runtime install (prebuilt release download+extract), GGUF
- * downloads, and llama-server lifecycle. Mirrors local-runtime.ts's OllamaRuntime: injectable
+ * Managed runtime for Pi's curated Prism 1-bit and ternary GGUF models. Q1_0 is available in
+ * upstream llama.cpp, while the selected group-128 Q2_0 artifacts and paired DSpark acceleration
+ * still require Prism's validated build. This module therefore owns the pinned runtime install,
+ * GGUF downloads, and llama-server lifecycle. Mirrors local-runtime.ts's OllamaRuntime: injectable
  * seams for fetch/spawn/exists, pi-owned directories under agentDir (runtimes/, models/),
  * detached+tracked child processes, onProgress as best-effort UI feedback, honest error taxonomies
  * instead of silent fallbacks.
@@ -29,12 +31,24 @@ import { extractZipArchive, fetchRuntimeDownload, installRuntimeArchive, removeP
  */
 export const PRISM_LLAMACPP_PINNED_RELEASE = "prism-b9594-38c66ad";
 export const PRISM_LLAMACPP_RELEASES_BASE_URL = "https://github.com/PrismML-Eng/llama.cpp/releases/download";
-/** Curated first-model descriptor; the `/models add` wiring task consumes this. */
+/** Curated 27B descriptor shared by the full local execution catalog. */
 export const BONSAI_27B = {
     repo: "prism-ml/Bonsai-27B-gguf",
     file: "Bonsai-27B-Q1_0.gguf",
     mmprojFile: "Bonsai-27B-mmproj-Q8_0.gguf",
     displayName: "Bonsai-27B (1-bit Q1_0 + vision)",
+    architecture: "dense",
+    runtime: "prism-llamacpp",
+    family: "bonsai",
+    parameterScale: "27B",
+    weightFormat: "q1_0",
+    matchedDrafter: {
+        kind: "dspark",
+        file: "Bonsai-27B-dspark-Q4_1.gguf",
+        draftMax: 4,
+        minimumContext: 16_384,
+        validatedBackend: "cuda",
+    },
 };
 // Verbatim asset names from the pinned release (see PRISM_LLAMACPP_PINNED_RELEASE doc comment) —
 // hardcoded rather than templated from the tag because the provider's naming isn't fully
@@ -47,13 +61,15 @@ const MACOS_ARM64_ASSET = "llama-prism-b9594-38c66ad-bin-macos-arm64.tar.gz";
 const MACOS_X64_ASSET = "llama-prism-b9594-38c66ad-bin-macos-x64.tar.gz";
 const WIN_X64_CPU_ASSET = "llama-bin-win-cpu-x64.zip";
 const WIN_ARM64_CPU_ASSET = "llama-bin-win-cpu-arm64.zip";
+const WIN_X64_CUDA_ASSET = "llama-prism-b1-38c66ad-bin-win-cuda-12.4-x64.zip";
+const WIN_X64_CUDA_COMPANION_ASSET = "cudart-llama-bin-win-cuda-12.4-x64.zip";
 /**
  * Maps a platform/arch/GPU triple to the exact Prism llama.cpp release asset for
  * {@link PRISM_LLAMACPP_PINNED_RELEASE} — verified against the real GitHub release, not guessed.
- * CPU asset by default; the CUDA 12.4 variant is only offered for linux x64 with an NVIDIA GPU
- * present (a 12.8 variant also exists upstream but has no wired caller yet). Windows CUDA needs a
- * companion `cudart-*` archive and is out of scope for v1 — Windows always resolves to the CPU zip
- * regardless of `hasNvidiaGpu`. Pure and exported so it's independently testable.
+ * CPU asset by default; CUDA 12.4 is selected for x64 Linux or Windows with an NVIDIA GPU. The
+ * pinned Windows build is incomplete without its matching `cudart-*` archive, so the resolver
+ * returns both as one mandatory installation plan. Pure and exported so it's independently
+ * testable.
  */
 export function resolvePrismLlamaAsset(plat, architecture, hasNvidiaGpu) {
     if (plat === "linux") {
@@ -74,8 +90,16 @@ export function resolvePrismLlamaAsset(plat, architecture, hasNvidiaGpu) {
         return undefined;
     }
     if (plat === "win32") {
-        if (architecture === "x64")
-            return { name: WIN_X64_CPU_ASSET, kind: "zip", backend: "cpu" };
+        if (architecture === "x64") {
+            return hasNvidiaGpu
+                ? {
+                    name: WIN_X64_CUDA_ASSET,
+                    kind: "zip",
+                    backend: "cuda",
+                    companionAssets: [{ name: WIN_X64_CUDA_COMPANION_ASSET, kind: "zip" }],
+                }
+                : { name: WIN_X64_CPU_ASSET, kind: "zip", backend: "cpu" };
+        }
         if (architecture === "arm64")
             return { name: WIN_ARM64_CPU_ASSET, kind: "zip", backend: "cpu" };
         return undefined;
@@ -94,6 +118,9 @@ function parseContentLength(header) {
     const value = Number(header);
     return Number.isFinite(value) && value >= 0 ? value : undefined;
 }
+function modelIdentityConflictError(servedModelIds) {
+    return `model-identity-conflict:${servedModelIds.join(",") || "unknown"}`;
+}
 export class PrismLlamaCppRuntime {
     constructor(args) {
         this._agentDir = args.agentDir;
@@ -102,10 +129,16 @@ export class PrismLlamaCppRuntime {
         this._hasNvidiaGpu = args.deps?.hasNvidiaGpu ?? (() => this._hasCommand("nvidia-smi"));
         this._platform = args.deps?.platform ?? osPlatform;
         this._arch = args.deps?.arch ?? osArch;
+        this._profile = deriveHostLocalInferenceProfile(args.profileMode ?? "balanced", args.deps);
         this._extractArchiveFn =
             args.deps?.extractArchive ?? ((input, destDir, kind) => this._extractArchive(input, destDir, kind));
         this._healthPollAttempts = args.deps?.healthPollAttempts ?? DEFAULT_HEALTH_POLL_ATTEMPTS;
         this._healthPollIntervalMs = args.deps?.healthPollIntervalMs ?? DEFAULT_HEALTH_POLL_INTERVAL_MS;
+        this._processLifecycle = args.deps?.processLifecycle ?? {
+            track: trackDetachedChildPid,
+            untrack: untrackDetachedChildPid,
+            terminate: killProcessTree,
+        };
     }
     runtimeDir() {
         return agentRuntimesDir("prism-llamacpp", this._agentDir);
@@ -186,10 +219,13 @@ export class PrismLlamaCppRuntime {
         const asset = resolvePrismLlamaAsset(this._platform(), this._arch(), this._hasNvidiaGpu());
         if (!asset)
             return { ok: false, error: "unsupported-platform" };
-        const downloadUrl = `${PRISM_LLAMACPP_RELEASES_BASE_URL}/${PRISM_LLAMACPP_PINNED_RELEASE}/${asset.name}`;
-        const extracted = await this._installArchive(downloadUrl, asset, onProgress);
-        if (!extracted.ok)
-            return extracted;
+        const archives = [asset, ...(asset.companionAssets ?? [])];
+        for (const archive of archives) {
+            const downloadUrl = `${PRISM_LLAMACPP_RELEASES_BASE_URL}/${PRISM_LLAMACPP_PINNED_RELEASE}/${archive.name}`;
+            const extracted = await this._installArchive(downloadUrl, archive, onProgress);
+            if (!extracted.ok)
+                return extracted;
+        }
         onProgress?.("Locating llama-server binary…");
         const destDir = this.runtimeDir();
         const binaryRelPath = this._findBinaryRelPath(destDir);
@@ -297,17 +333,6 @@ export class PrismLlamaCppRuntime {
         onProgress?.(`${args.file} downloaded (${actualBytes ?? 0} bytes).`);
         return { ok: true, path: destPath };
     }
-    async _healthUp(baseUrl) {
-        try {
-            const response = await this._fetch(`${baseUrl}/health`, {
-                signal: AbortSignal.timeout(HEALTH_CHECK_TIMEOUT_MS),
-            });
-            return response.ok;
-        }
-        catch {
-            return false;
-        }
-    }
     /**
      * Spawn `llama-server` detached+tracked (killed on parent shutdown even if pi crashes without
      * calling stop()) and poll `/health` until ready. `-ngl 99` is only added when the installed
@@ -321,9 +346,19 @@ export class PrismLlamaCppRuntime {
         if (!this._exists(binaryPath))
             return { ok: false, error: "binary-missing" };
         const baseUrl = `http://127.0.0.1:${args.port}`;
+        const existing = await probePrismLlamaCppServer(baseUrl, args.modelAlias, this._fetch, HEALTH_CHECK_TIMEOUT_MS);
+        if (existing.status === "matching")
+            return { ok: true, baseUrl };
+        if (existing.status === "conflict" && !this._child) {
+            return { ok: false, error: modelIdentityConflictError(existing.servedModelIds) };
+        }
+        if (this._child)
+            this.stop();
         const argv = [
             "-m",
             args.modelPath,
+            "--alias",
+            args.modelAlias,
             ...(args.mmprojPath ? ["--mmproj", args.mmprojPath] : []),
             "--host",
             "127.0.0.1",
@@ -331,6 +366,20 @@ export class PrismLlamaCppRuntime {
             String(args.port),
             "-c",
             String(args.numCtx),
+            "--cache-ram",
+            String(this._profile.promptCacheMiB),
+            "-np",
+            String(this._profile.parallelRequests),
+            "-fa",
+            this._profile.flashAttention ? "on" : "off",
+            "-ctk",
+            this._profile.kvCacheType,
+            "-ctv",
+            this._profile.kvCacheType,
+            "-t",
+            String(this._profile.generationThreads),
+            "-tb",
+            String(this._profile.batchThreads),
             ...(manifest.backend === "cuda" ? ["-ngl", "99"] : []),
         ];
         const child = this._spawn(binaryPath, argv, {
@@ -339,15 +388,22 @@ export class PrismLlamaCppRuntime {
             env: process.env,
         });
         if (child.pid)
-            trackDetachedChildPid(child.pid);
+            this._processLifecycle.track(child.pid);
         child.unref?.();
         child.on("exit", () => {
-            this._child = undefined;
+            if (this._child === child) {
+                this._child = undefined;
+            }
         });
         this._child = child;
         for (let attempt = 0; attempt < this._healthPollAttempts; attempt++) {
-            if (await this._healthUp(baseUrl))
+            const probe = await probePrismLlamaCppServer(baseUrl, args.modelAlias, this._fetch, HEALTH_CHECK_TIMEOUT_MS);
+            if (probe.status === "matching")
                 return { ok: true, baseUrl };
+            if (probe.status === "conflict") {
+                this.stop();
+                return { ok: false, error: modelIdentityConflictError(probe.servedModelIds) };
+            }
             await this._sleep(this._healthPollIntervalMs);
         }
         this.stop();
@@ -359,8 +415,8 @@ export class PrismLlamaCppRuntime {
             return { stopped: false };
         this._child = undefined;
         if (child.pid) {
-            untrackDetachedChildPid(child.pid);
-            killProcessTree(child.pid);
+            this._processLifecycle.untrack(child.pid);
+            this._processLifecycle.terminate(child.pid);
         }
         return { stopped: true };
     }