@caupulican/pi-adaptative 0.85.0 → 0.85.4

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 {
@@ -81,6 +97,9 @@ 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. */
@@ -102,9 +121,11 @@ export declare class PrismLlamaCppRuntime {
     private readonly _extractArchiveFn;
     private readonly _healthPollAttempts;
     private readonly _healthPollIntervalMs;
+    private readonly _profile;
     private _child;
     constructor(args: {
         agentDir: string;
+        profileMode?: LocalInferenceProfileMode;
         deps?: PrismLlamaCppDeps;
     });
     runtimeDir(): string;
@@ -144,7 +165,6 @@ 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 +172,7 @@ 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;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,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;CAC9B;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,MAAM,CAAkE;IAEhF,YAAY,IAAI,EAAE;QACjB,QAAQ,EAAE,MAAM,CAAC;QACjB,WAAW,CAAC,EAAE,yBAAyB,CAAC;QACxC,IAAI,CAAC,EAAE,iBAAiB,CAAC;KACzB,EAYA;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:${servedModelIds.join(",") || "unknown"}`;
+}
 export class PrismLlamaCppRuntime {
     constructor(args) {
         this._agentDir = args.agentDir;
@@ -102,6 +129,7 @@ export class PrismLlamaCppRuntime {
         this._hasNvidiaGpu = args.deps?.hasNvidiaGpu ?? (() => this._hasCommand("nvidia-smi"));
         this._platform = args.deps?.platform ?? osPlatform;
         this._arch = args.deps?.arch ?? osArch;
+        this._profile = deriveHostLocalInferenceProfile(args.profileMode ?? "balanced", args.deps);
         this._extractArchiveFn =
             args.deps?.extractArchive ?? ((input, destDir, kind) => this._extractArchive(input, destDir, kind));
         this._healthPollAttempts = args.deps?.healthPollAttempts ?? DEFAULT_HEALTH_POLL_ATTEMPTS;
@@ -186,10 +214,13 @@ export class PrismLlamaCppRuntime {
         const asset = resolvePrismLlamaAsset(this._platform(), this._arch(), this._hasNvidiaGpu());
         if (!asset)
             return { ok: false, error: "unsupported-platform" };
-        const downloadUrl = `${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}/${archive.name}`;
+            const extracted = await this._installArchive(downloadUrl, archive, onProgress);
+            if (!extracted.ok)
+                return extracted;
+        }
         onProgress?.("Locating llama-server binary…");
         const destDir = this.runtimeDir();
         const binaryRelPath = this._findBinaryRelPath(destDir);
@@ -297,17 +328,6 @@ export class PrismLlamaCppRuntime {
         onProgress?.(`${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 +341,19 @@ 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 +361,20 @@ 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, {
@@ -342,12 +386,19 @@ export class PrismLlamaCppRuntime {
             trackDetachedChildPid(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();

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

@@ -1 +1 @@
-{"version":3,"file":"llamacpp-runtime.js","sourceRoot":"","sources":["../../../src/core/models/llamacpp-runtime.ts"],"names":[],"mappings":"AACA,OAAO,EAAe,SAAS,EAAE,WAAW,EAAE,YAAY,EAAE,aAAa,EAAE,MAAM,SAAS,CAAC;AAC3F,OAAO,EAAE,IAAI,IAAI,MAAM,EAAE,QAAQ,IAAI,UAAU,EAAE,MAAM,SAAS,CAAC;AACjE,OAAO,EAAE,OAAO,EAAE,IAAI,EAAE,QAAQ,EAAE,MAAM,WAAW,CAAC;AAEpD,OAAO,EAAE,QAAQ,EAAE,MAAM,sBAAsB,CAAC;AAChD,OAAO,EAAE,YAAY,EAAE,kCAAkC,EAAE,MAAM,8BAA8B,CAAC;AAChG,OAAO,EAAE,eAAe,EAAE,qBAAqB,EAAE,uBAAuB,EAAE,MAAM,sBAAsB,CAAC;AACvG,OAAO,EAAE,SAAS,IAAI,cAAc,EAAE,WAAW,IAAI,gBAAgB,EAAE,MAAM,mBAAmB,CAAC;AACjG,OAAO,EACN,iBAAiB,EACjB,oBAAoB,EACpB,qBAAqB,EAErB,qBAAqB,EACrB,mBAAmB,EACnB,mCAAmC,EACnC,uBAAuB,EACvB,gBAAgB,EAChB,oBAAoB,GACpB,MAAM,sBAAsB,CAAC;AAE9B;;;;;;;;;;;;;;GAcG;AAEH;;;;;GAKG;AACH,MAAM,CAAC,MAAM,6BAA6B,GAAG,qBAAqB,CAAC;AAEnE,MAAM,CAAC,MAAM,gCAAgC,GAAG,4DAA4D,CAAC;AAS7G,mFAAmF;AACnF,MAAM,CAAC,MAAM,UAAU,GAAyB;IAC/C,IAAI,EAAE,0BAA0B;IAChC,IAAI,EAAE,sBAAsB;IAC5B,UAAU,EAAE,6BAA6B;IACzC,WAAW,EAAE,kCAAkC;CAC/C,CAAC;AAWF,iGAAiG;AACjG,yFAAyF;AACzF,4FAA4F;AAC5F,+EAA+E;AAC/E,MAAM,mBAAmB,GAAG,iDAAiD,CAAC;AAC9E,MAAM,qBAAqB,GAAG,mDAAmD,CAAC;AAClF,MAAM,oBAAoB,GAAG,0DAA0D,CAAC;AACxF,MAAM,iBAAiB,GAAG,kDAAkD,CAAC;AAC7E,MAAM,eAAe,GAAG,gDAAgD,CAAC;AACzE,MAAM,iBAAiB,GAAG,2BAA2B,CAAC;AACtD,MAAM,mBAAmB,GAAG,6BAA6B,CAAC;AAE1D;;;;;;;GAOG;AACH,MAAM,UAAU,sBAAsB,CACrC,IAAY,EACZ,YAAoB,EACpB,YAAqB;IAErB,IAAI,IAAI,KAAK,OAAO,EAAE,CAAC;QACtB,IAAI,YAAY,KAAK,KAAK,EAAE,CAAC;YAC5B,OAAO,YAAY;gBAClB,CAAC,CAAC,EAAE,IAAI,EAAE,oBAAoB,EAAE,IAAI,EAAE,QAAQ,EAAE,OAAO,EAAE,MAAM,EAAE;gBACjE,CAAC,CAAC,EAAE,IAAI,EAAE,mBAAmB,EAAE,IAAI,EAAE,QAAQ,EAAE,OAAO,EAAE,KAAK,EAAE,CAAC;QAClE,CAAC;QACD,IAAI,YAAY,KAAK,OAAO;YAAE,OAAO,EAAE,IAAI,EAAE,qBAAqB,EAAE,IAAI,EAAE,QAAQ,EAAE,OAAO,EAAE,KAAK,EAAE,CAAC;QACrG,OAAO,SAAS,CAAC;IAClB,CAAC;IACD,IAAI,IAAI,KAAK,QAAQ,EAAE,CAAC;QACvB,IAAI,YAAY,KAAK,OAAO;YAAE,OAAO,EAAE,IAAI,EAAE,iBAAiB,EAAE,IAAI,EAAE,QAAQ,EAAE,OAAO,EAAE,KAAK,EAAE,CAAC;QACjG,IAAI,YAAY,KAAK,KAAK;YAAE,OAAO,EAAE,IAAI,EAAE,eAAe,EAAE,IAAI,EAAE,QAAQ,EAAE,OAAO,EAAE,KAAK,EAAE,CAAC;QAC7F,OAAO,SAAS,CAAC;IAClB,CAAC;IACD,IAAI,IAAI,KAAK,OAAO,EAAE,CAAC;QACtB,IAAI,YAAY,KAAK,KAAK;YAAE,OAAO,EAAE,IAAI,EAAE,iBAAiB,EAAE,IAAI,EAAE,KAAK,EAAE,OAAO,EAAE,KAAK,EAAE,CAAC;QAC5F,IAAI,YAAY,KAAK,OAAO;YAAE,OAAO,EAAE,IAAI,EAAE,mBAAmB,EAAE,IAAI,EAAE,KAAK,EAAE,OAAO,EAAE,KAAK,EAAE,CAAC;QAChG,OAAO,SAAS,CAAC;IAClB,CAAC;IACD,OAAO,SAAS,CAAC;AAClB,CAAC;AAkDD,MAAM,qBAAqB,GAAG,EAAE,GAAG,MAAM,CAAC;AAC1C,MAAM,qBAAqB,GAAG,KAAK,CAAC;AACpC,MAAM,4BAA4B,GAAG,GAAG,CAAC;AACzC,MAAM,+BAA+B,GAAG,GAAG,CAAC;AAC5C,MAAM,uBAAuB,GAAG,KAAK,CAAC;AACtC,MAAM,uBAAuB,GAAG,MAAM,CAAC;AAEvC,SAAS,kBAAkB,CAAC,MAAqB;IAChD,IAAI,CAAC,MAAM;QAAE,OAAO,SAAS,CAAC;IAC9B,MAAM,KAAK,GAAG,MAAM,CAAC,MAAM,CAAC,CAAC;IAC7B,OAAO,MAAM,CAAC,QAAQ,CAAC,KAAK,CAAC,IAAI,KAAK,IAAI,CAAC,CAAC,CAAC,CAAC,KAAK,CAAC,CAAC,CAAC,SAAS,CAAC;AACjE,CAAC;AAED,MAAM,OAAO,oBAAoB;IAehC,YAAY,IAAoD;QAC/D,IAAI,CAAC,SAAS,GAAG,IAAI,CAAC,QAAQ,CAAC;QAC/B,CAAC,IAAI,CAAC,MAAM,EAAE,IAAI,CAAC,MAAM,EAAE,IAAI,CAAC,OAAO,EAAE,IAAI,CAAC,MAAM,CAAC,GAAG,mCAAmC,CAAC,IAAI,CAAC,IAAI,CAAC,CAAC;QACvG,IAAI,CAAC,WAAW,GAAG,IAAI,CAAC,IAAI,EAAE,UAAU,IAAI,uBAAuB,CAAC;QACpE,IAAI,CAAC,aAAa,GAAG,IAAI,CAAC,IAAI,EAAE,YAAY,IAAI,CAAC,GAAG,EAAE,CAAC,IAAI,CAAC,WAAW,CAAC,YAAY,CAAC,CAAC,CAAC;QACvF,IAAI,CAAC,SAAS,GAAG,IAAI,CAAC,IAAI,EAAE,QAAQ,IAAI,UAAU,CAAC;QACnD,IAAI,CAAC,KAAK,GAAG,IAAI,CAAC,IAAI,EAAE,IAAI,IAAI,MAAM,CAAC;QACvC,IAAI,CAAC,iBAAiB;YACrB,IAAI,CAAC,IAAI,EAAE,cAAc,IAAI,CAAC,CAAC,KAAK,EAAE,OAAO,EAAE,IAAI,EAAE,EAAE,CAAC,IAAI,CAAC,eAAe,CAAC,KAAK,EAAE,OAAO,EAAE,IAAI,CAAC,CAAC,CAAC;QACrG,IAAI,CAAC,mBAAmB,GAAG,IAAI,CAAC,IAAI,EAAE,kBAAkB,IAAI,4BAA4B,CAAC;QACzF,IAAI,CAAC,qBAAqB,GAAG,IAAI,CAAC,IAAI,EAAE,oBAAoB,IAAI,+BAA+B,CAAC;IACjG,CAAC;IAED,UAAU;QACT,OAAO,gBAAgB,CAAC,gBAAgB,EAAE,IAAI,CAAC,SAAS,CAAC,CAAC;IAC3D,CAAC;IAED,SAAS;QACR,OAAO,cAAc,CAAC,UAAU,EAAE,IAAI,CAAC,SAAS,CAAC,CAAC;IACnD,CAAC;IAEO,WAAW;QAClB,OAAO,IAAI,CAAC,SAAS,EAAE,KAAK,OAAO,CAAC,CAAC,CAAC,kBAAkB,CAAC,CAAC,CAAC,cAAc,CAAC;IAC3E,CAAC;IAEO,aAAa;QACpB,OAAO,IAAI,CAAC,IAAI,CAAC,UAAU,EAAE,EAAE,cAAc,CAAC,CAAC;IAChD,CAAC;IAEO,aAAa;QACpB,IAAI,CAAC;YACJ,MAAM,MAAM,GAAG,IAAI,CAAC,KAAK,CAAC,YAAY,CAAC,IAAI,CAAC,aAAa,EAAE,EAAE,MAAM,CAAC,CAInE,CAAC;YACF,IACC,OAAO,MAAM,CAAC,OAAO,KAAK,QAAQ;gBAClC,OAAO,MAAM,CAAC,aAAa,KAAK,QAAQ;gBACxC,CAAC,MAAM,CAAC,OAAO,KAAK,KAAK,IAAI,MAAM,CAAC,OAAO,KAAK,MAAM,CAAC,EACtD,CAAC;gBACF,OAAO,EAAE,OAAO,EAAE,MAAM,CAAC,OAAO,EAAE,aAAa,EAAE,MAAM,CAAC,aAAa,EAAE,OAAO,EAAE,MAAM,CAAC,OAAO,EAAE,CAAC;YAClG,CAAC;YACD,OAAO,SAAS,CAAC;QAClB,CAAC;QAAC,MAAM,CAAC;YACR,OAAO,SAAS,CAAC;QAClB,CAAC;IACF,CAAC;IAEO,cAAc,CAAC,QAA8B;QACpD,aAAa,CAAC,IAAI,CAAC,aAAa,EAAE,EAAE,IAAI,CAAC,SAAS,CAAC,QAAQ,EAAE,IAAI,EAAE,IAAI,CAAC,CAAC,CAAC;IAC3E,CAAC;IAED,KAAK,CAAC,MAAM;QACX,MAAM,QAAQ,GAAG,IAAI,CAAC,aAAa,EAAE,CAAC;QACtC,IAAI,CAAC,QAAQ;YAAE,OAAO,EAAE,gBAAgB,EAAE,KAAK,EAAE,CAAC;QAClD,MAAM,UAAU,GAAG,IAAI,CAAC,IAAI,CAAC,UAAU,EAAE,EAAE,QAAQ,CAAC,aAAa,CAAC,CAAC;QACnE,MAAM,gBAAgB,GAAG,IAAI,CAAC,OAAO,CAAC,UAAU,CAAC,CAAC;QAClD,OAAO;YACN,gBAAgB;YAChB,UAAU,EAAE,gBAAgB,CAAC,CAAC,CAAC,UAAU,CAAC,CAAC,CAAC,SAAS;YACrD,OAAO,EAAE,QAAQ,CAAC,OAAO;SACzB,CAAC;IACH,CAAC;IAED;;;yDAGqD;IAC7C,kBAAkB,CAAC,IAAY;QACtC,MAAM,UAAU,GAAG,IAAI,CAAC,WAAW,EAAE,CAAC;QACtC,MAAM,KAAK,GAAa,CAAC,IAAI,CAAC,CAAC;QAC/B,OAAO,KAAK,CAAC,MAAM,GAAG,CAAC,EAAE,CAAC;YACzB,MAAM,GAAG,GAAG,KAAK,CAAC,GAAG,EAAE,CAAC;YACxB,IAAI,GAAG,KAAK,SAAS;gBAAE,SAAS;YAChC,IAAI,OAAiB,CAAC;YACtB,IAAI,CAAC;gBACJ,OAAO,GAAG,WAAW,CAAC,GAAG,EAAE,EAAE,aAAa,EAAE,IAAI,EAAE,CAAC,CAAC;YACrD,CAAC;YAAC,MAAM,CAAC;gBACR,SAAS;YACV,CAAC;YACD,KAAK,MAAM,KAAK,IAAI,OAAO,EAAE,CAAC;gBAC7B,MAAM,IAAI,GAAG,IAAI,CAAC,GAAG,EAAE,KAAK,CAAC,IAAI,CAAC,CAAC;gBACnC,IAAI,KAAK,CAAC,WAAW,EAAE;oBAAE,KAAK,CAAC,IAAI,CAAC,IAAI,CAAC,CAAC;qBACrC,IAAI,KAAK,CAAC,MAAM,EAAE,IAAI,KAAK,CAAC,IAAI,KAAK,UAAU;oBAAE,OAAO,QAAQ,CAAC,IAAI,EAAE,IAAI,CAAC,CAAC;YACnF,CAAC;QACF,CAAC;QACD,OAAO,SAAS,CAAC;IAClB,CAAC;IAED;;;;;OAKG;IACH,KAAK,CAAC,cAAc,CAAC,UAAqC;QACzD,MAAM,KAAK,GAAG,sBAAsB,CAAC,IAAI,CAAC,SAAS,EAAE,EAAE,IAAI,CAAC,KAAK,EAAE,EAAE,IAAI,CAAC,aAAa,EAAE,CAAC,CAAC;QAC3F,IAAI,CAAC,KAAK;YAAE,OAAO,EAAE,EAAE,EAAE,KAAK,EAAE,KAAK,EAAE,sBAAsB,EAAE,CAAC;QAEhE,MAAM,WAAW,GAAG,GAAG,gCAAgC,IAAI,6BAA6B,IAAI,KAAK,CAAC,IAAI,EAAE,CAAC;QACzG,MAAM,SAAS,GAAG,MAAM,IAAI,CAAC,eAAe,CAAC,WAAW,EAAE,KAAK,EAAE,UAAU,CAAC,CAAC;QAC7E,IAAI,CAAC,SAAS,CAAC,EAAE;YAAE,OAAO,SAAS,CAAC;QAEpC,UAAU,EAAE,CAAC,+BAA+B,CAAC,CAAC;QAC9C,MAAM,OAAO,GAAG,IAAI,CAAC,UAAU,EAAE,CAAC;QAClC,MAAM,aAAa,GAAG,IAAI,CAAC,kBAAkB,CAAC,OAAO,CAAC,CAAC;QACvD,IAAI,CAAC,aAAa,EAAE,CAAC;YACpB,OAAO,EAAE,EAAE,EAAE,KAAK,EAAE,KAAK,EAAE,uEAAuE,EAAE,CAAC;QACtG,CAAC;QAED,IAAI,CAAC,cAAc,CAAC,EAAE,OAAO,EAAE,6BAA6B,EAAE,aAAa,EAAE,OAAO,EAAE,KAAK,CAAC,OAAO,EAAE,CAAC,CAAC;QACvG,UAAU,EAAE,CAAC,oCAAoC,CAAC,CAAC;QACnD,OAAO,EAAE,EAAE,EAAE,IAAI,EAAE,CAAC;IACrB,CAAC;IAEO,eAAe,CACtB,WAAmB,EACnB,KAAsB,EACtB,UAAqC;QAErC,OAAO,qBAAqB,CAC3B,IAAI,CAAC,MAAM,EACX,WAAW,EACX,IAAI,CAAC,UAAU,EAAE,EACjB,KAAK,EACL,IAAI,CAAC,iBAAiB,EACtB,UAAU,CACV,CAAC;IACH,CAAC;IAEO,KAAK,CAAC,eAAe,CAC5B,KAAe,EACf,OAAe,EACf,IAAoB;QAEpB,IAAI,CAAC;YACJ,OAAO,IAAI,KAAK,KAAK,CAAC,CAAC,CAAC,MAAM,IAAI,CAAC,WAAW,CAAC,KAAK,EAAE,OAAO,CAAC,CAAC,CAAC,CAAC,MAAM,IAAI,CAAC,aAAa,CAAC,KAAK,EAAE,OAAO,CAAC,CAAC;QAC3G,CAAC;QAAC,OAAO,KAAK,EAAE,CAAC;YAChB,OAAO,EAAE,EAAE,EAAE,KAAK,EAAE,KAAK,EAAE,iBAAiB,KAAK,YAAY,KAAK,CAAC,CAAC,CAAC,KAAK,CAAC,OAAO,CAAC,CAAC,CAAC,MAAM,CAAC,KAAK,CAAC,EAAE,EAAE,CAAC;QACxG,CAAC;IACF,CAAC;IAEO,KAAK,CAAC,aAAa,CAAC,KAAe,EAAE,OAAe;QAC3D,MAAM,OAAO,GAAG,YAAY,CAAC,KAAK,EAAE,CAAC,MAAM,EAAE,GAAG,EAAE,IAAI,EAAE,OAAO,CAAC,EAAE;YACjE,QAAQ,EAAE,OAAO,CAAC,QAAQ,KAAK,OAAO;YACtC,KAAK,EAAE,CAAC,MAAM,EAAE,QAAQ,EAAE,QAAQ,CAAC;SACnC,CAAC,CAAC;QACH,MAAM,qBAAqB,GAAG,IAAI,eAAe,EAAE,CAAC;QACpD,MAAM,WAAW,GAAG,kCAAkC,CAAC,OAAO,EAAE;YAC/D,MAAM,EAAE,qBAAqB,CAAC,MAAM;YACpC,SAAS,EAAE,qBAAqB;YAChC,WAAW,EAAE,qBAAqB;SAClC,CAAC,CAAC;QACH,IAAI,CAAC;YACJ,MAAM,QAAQ,CAAC,KAAK,EAAE,mBAAmB,CAAC,OAAO,EAAE,KAAK,CAAC,CAAC,CAAC;QAC5D,CAAC;QAAC,OAAO,KAAK,EAAE,CAAC;YAChB,qBAAqB,CAAC,KAAK,EAAE,CAAC;YAC9B,MAAM,WAAW,CAAC,KAAK,CAAC,GAAG,EAAE,GAAE,CAAC,CAAC,CAAC;YAClC,MAAM,KAAK,CAAC;QACb,CAAC;QACD,MAAM,QAAQ,GAAG,MAAM,WAAW,CAAC;QACnC,IAAI,QAAQ,CAAC,MAAM,KAAK,SAAS,IAAI,QAAQ,CAAC,IAAI,KAAK,CAAC,EAAE,CAAC;YAC1D,OAAO;gBACN,EAAE,EAAE,KAAK;gBACT,KAAK,EACJ,QAAQ,CAAC,MAAM,KAAK,SAAS;oBAC5B,CAAC,CAAC,gDAAgD,qBAAqB,IAAI;oBAC3E,CAAC,CAAC,sCAAsC,QAAQ,CAAC,IAAI,IAAI,SAAS,EAAE;aACtE,CAAC;QACH,CAAC;QACD,OAAO,EAAE,EAAE,EAAE,IAAI,EAAE,CAAC;IACrB,CAAC;IAEO,KAAK,CAAC,WAAW,CAAC,KAAe,EAAE,OAAe;QACzD,OAAO,iBAAiB,CAAC;YACxB,KAAK;YACL,OAAO;YACP,UAAU,EAAE,yBAAyB;YACrC,QAAQ,EAAE,IAAI,CAAC,SAAS;YACxB,UAAU,EAAE,IAAI,CAAC,WAAW;YAC5B,SAAS,EAAE,qBAAqB;YAChC,WAAW,EAAE,qBAAqB;SAClC,CAAC,CAAC;IACJ,CAAC;IAEO,KAAK,CAAC,oBAAoB,CAAC,GAAW;QAC7C,IAAI,CAAC;YACJ,MAAM,QAAQ,GAAG,MAAM,IAAI,CAAC,MAAM,CAAC,GAAG,EAAE;gBACvC,MAAM,EAAE,MAAM;gBACd,MAAM,EAAE,WAAW,CAAC,OAAO,CAAC,uBAAuB,CAAC;aACpD,CAAC,CAAC;YACH,OAAO,QAAQ,CAAC,EAAE,CAAC,CAAC,CAAC,kBAAkB,CAAC,QAAQ,CAAC,OAAO,CAAC,GAAG,CAAC,gBAAgB,CAAC,CAAC,CAAC,CAAC,CAAC,SAAS,CAAC;QAC7F,CAAC;QAAC,MAAM,CAAC;YACR,OAAO,SAAS,CAAC;QAClB,CAAC;IACF,CAAC;IAED;;;;;OAKG;IACH,KAAK,CAAC,aAAa,CAClB,IAAoC,EACpC,UAAqC;QAErC,MAAM,QAAQ,GAAG,IAAI,CAAC,IAAI,CAAC,SAAS,EAAE,EAAE,IAAI,CAAC,IAAI,EAAE,IAAI,CAAC,IAAI,CAAC,CAAC;QAC9D,MAAM,GAAG,GAAG,0BAA0B,IAAI,CAAC,IAAI,iBAAiB,IAAI,CAAC,IAAI,EAAE,CAAC;QAC5E,SAAS,CAAC,OAAO,CAAC,QAAQ,CAAC,EAAE,EAAE,SAAS,EAAE,IAAI,EAAE,CAAC,CAAC;QAElD,IAAI,IAAI,CAAC,OAAO,CAAC,QAAQ,CAAC,EAAE,CAAC;YAC5B,MAAM,UAAU,GAAG,MAAM,IAAI,CAAC,oBAAoB,CAAC,GAAG,CAAC,CAAC;YACxD,IAAI,UAAU,KAAK,SAAS,IAAI,UAAU,KAAK,gBAAgB,CAAC,QAAQ,CAAC,EAAE,CAAC;gBAC3E,UAAU,EAAE,CAAC,GAAG,IAAI,CAAC,IAAI,wBAAwB,UAAU,qBAAqB,CAAC,CAAC;gBAClF,OAAO,EAAE,EAAE,EAAE,IAAI,EAAE,IAAI,EAAE,QAAQ,EAAE,OAAO,EAAE,IAAI,EAAE,CAAC;YACpD,CAAC;QACF,CAAC;QAED,UAAU,EAAE,CAAC,eAAe,IAAI,CAAC,IAAI,SAAS,IAAI,CAAC,IAAI,GAAG,CAAC,CAAC;QAC5D,MAAM,QAAQ,GAAG,MAAM,oBAAoB,CAAC,IAAI,CAAC,MAAM,EAAE,GAAG,CAAC,CAAC;QAC9D,IAAI,CAAC,QAAQ,CAAC,EAAE;YAAE,OAAO,QAAQ,CAAC;QAElC,MAAM,aAAa,GAAG,kBAAkB,CAAC,QAAQ,CAAC,QAAQ,CAAC,OAAO,CAAC,GAAG,CAAC,gBAAgB,CAAC,CAAC,CAAC;QAC1F,MAAM,OAAO,GAAG,MAAM,oBAAoB,CAAC,QAAQ,CAAC,IAA2B,EAAE,QAAQ,CAAC,CAAC;QAC3F,IAAI,CAAC,OAAO,CAAC,EAAE;YAAE,OAAO,OAAO,CAAC;QAEhC,MAAM,WAAW,GAAG,gBAAgB,CAAC,QAAQ,CAAC,CAAC;QAC/C,IAAI,aAAa,KAAK,SAAS,IAAI,WAAW,KAAK,aAAa,EAAE,CAAC;YAClE,qBAAqB,CAAC,QAAQ,CAAC,CAAC;YAChC,OAAO,EAAE,EAAE,EAAE,KAAK,EAAE,KAAK,EAAE,2BAA2B,aAAa,eAAe,WAAW,IAAI,CAAC,EAAE,EAAE,CAAC;QACxG,CAAC;QAED,UAAU,EAAE,CAAC,GAAG,IAAI,CAAC,IAAI,gBAAgB,WAAW,IAAI,CAAC,UAAU,CAAC,CAAC;QACrE,OAAO,EAAE,EAAE,EAAE,IAAI,EAAE,IAAI,EAAE,QAAQ,EAAE,CAAC;IACrC,CAAC;IAEO,KAAK,CAAC,SAAS,CAAC,OAAe;QACtC,IAAI,CAAC;YACJ,MAAM,QAAQ,GAAG,MAAM,IAAI,CAAC,MAAM,CAAC,GAAG,OAAO,SAAS,EAAE;gBACvD,MAAM,EAAE,WAAW,CAAC,OAAO,CAAC,uBAAuB,CAAC;aACpD,CAAC,CAAC;YACH,OAAO,QAAQ,CAAC,EAAE,CAAC;QACpB,CAAC;QAAC,MAAM,CAAC;YACR,OAAO,KAAK,CAAC;QACd,CAAC;IACF,CAAC;IAED;;;;OAIG;IACH,KAAK,CAAC,KAAK,CAAC,IAKX;QACA,MAAM,QAAQ,GAAG,IAAI,CAAC,aAAa,EAAE,CAAC;QACtC,IAAI,CAAC,QAAQ;YAAE,OAAO,EAAE,EAAE,EAAE,KAAK,EAAE,KAAK,EAAE,gBAAgB,EAAE,CAAC;QAC7D,MAAM,UAAU,GAAG,IAAI,CAAC,IAAI,CAAC,UAAU,EAAE,EAAE,QAAQ,CAAC,aAAa,CAAC,CAAC;QACnE,IAAI,CAAC,IAAI,CAAC,OAAO,CAAC,UAAU,CAAC;YAAE,OAAO,EAAE,EAAE,EAAE,KAAK,EAAE,KAAK,EAAE,gBAAgB,EAAE,CAAC;QAE7E,MAAM,OAAO,GAAG,oBAAoB,IAAI,CAAC,IAAI,EAAE,CAAC;QAChD,MAAM,IAAI,GAAG;YACZ,IAAI;YACJ,IAAI,CAAC,SAAS;YACd,GAAG,CAAC,IAAI,CAAC,UAAU,CAAC,CAAC,CAAC,CAAC,UAAU,EAAE,IAAI,CAAC,UAAU,CAAC,CAAC,CAAC,CAAC,EAAE,CAAC;YACzD,QAAQ;YACR,WAAW;YACX,QAAQ;YACR,MAAM,CAAC,IAAI,CAAC,IAAI,CAAC;YACjB,IAAI;YACJ,MAAM,CAAC,IAAI,CAAC,MAAM,CAAC;YACnB,GAAG,CAAC,QAAQ,CAAC,OAAO,KAAK,MAAM,CAAC,CAAC,CAAC,CAAC,MAAM,EAAE,IAAI,CAAC,CAAC,CAAC,CAAC,EAAE,CAAC;SACtD,CAAC;QAEF,MAAM,KAAK,GAAG,IAAI,CAAC,MAAM,CAAC,UAAU,EAAE,IAAI,EAAE;YAC3C,QAAQ,EAAE,OAAO,CAAC,QAAQ,KAAK,OAAO;YACtC,KAAK,EAAE,QAAQ;YACf,GAAG,EAAE,OAAO,CAAC,GAAG;SAChB,CAAC,CAAC;QACH,IAAI,KAAK,CAAC,GAAG;YAAE,qBAAqB,CAAC,KAAK,CAAC,GAAG,CAAC,CAAC;QAChD,KAAK,CAAC,KAAK,EAAE,EAAE,CAAC;QAChB,KAAK,CAAC,EAAE,CAAC,MAAM,EAAE,GAAG,EAAE;YACrB,IAAI,CAAC,MAAM,GAAG,SAAS,CAAC;QACzB,CAAC,CAAC,CAAC;QACH,IAAI,CAAC,MAAM,GAAG,KAAK,CAAC;QAEpB,KAAK,IAAI,OAAO,GAAG,CAAC,EAAE,OAAO,GAAG,IAAI,CAAC,mBAAmB,EAAE,OAAO,EAAE,EAAE,CAAC;YACrE,IAAI,MAAM,IAAI,CAAC,SAAS,CAAC,OAAO,CAAC;gBAAE,OAAO,EAAE,EAAE,EAAE,IAAI,EAAE,OAAO,EAAE,CAAC;YAChE,MAAM,IAAI,CAAC,MAAM,CAAC,IAAI,CAAC,qBAAqB,CAAC,CAAC;QAC/C,CAAC;QACD,IAAI,CAAC,IAAI,EAAE,CAAC;QACZ,OAAO,EAAE,EAAE,EAAE,KAAK,EAAE,KAAK,EAAE,gBAAgB,EAAE,CAAC;IAC/C,CAAC;IAED,IAAI;QACH,MAAM,KAAK,GAAG,IAAI,CAAC,MAAM,CAAC;QAC1B,IAAI,CAAC,KAAK;YAAE,OAAO,EAAE,OAAO,EAAE,KAAK,EAAE,CAAC;QACtC,IAAI,CAAC,MAAM,GAAG,SAAS,CAAC;QACxB,IAAI,KAAK,CAAC,GAAG,EAAE,CAAC;YACf,uBAAuB,CAAC,KAAK,CAAC,GAAG,CAAC,CAAC;YACnC,eAAe,CAAC,KAAK,CAAC,GAAG,CAAC,CAAC;QAC5B,CAAC;QACD,OAAO,EAAE,OAAO,EAAE,IAAI,EAAE,CAAC;IAC1B,CAAC;IAED,SAAS;QACR,OAAO,IAAI,CAAC,MAAM,KAAK,SAAS,CAAC;IAClC,CAAC;CACD","sourcesContent":["import type { ChildProcess } from \"node:child_process\";\nimport { type Dirent, mkdirSync, readdirSync, readFileSync, writeFileSync } from \"node:fs\";\nimport { arch as osArch, platform as osPlatform } from \"node:os\";\nimport { dirname, join, relative } from \"node:path\";\nimport type { Readable } from \"node:stream\";\nimport { pipeline } from \"node:stream/promises\";\nimport { spawnProcess, waitForChildProcessWithTermination } from \"../../utils/child-process.ts\";\nimport { killProcessTree, trackDetachedChildPid, untrackDetachedChildPid } from \"../../utils/shell.ts\";\nimport { modelsDir as agentModelsDir, runtimesDir as agentRuntimesDir } from \"../agent-paths.ts\";\nimport {\n\textractZipArchive,\n\tfetchRuntimeDownload,\n\tinstallRuntimeArchive,\n\ttype ManagedRuntimeSpawn,\n\tremovePartialDownload,\n\trequireRuntimeStdin,\n\tresolveRuntimeLifecycleDependencies,\n\truntimeCommandAvailable,\n\ttryFileSizeBytes,\n\twriteRuntimeDownload,\n} from \"./runtime-process.ts\";\n\n/**\n * Managed runtime for prism-ml 1-bit GGUF models (Bonsai-27B first). The provider requires their\n * llama.cpp fork (Q1_0_g128 hybrid-attention kernels) — stock llama.cpp/Ollama cannot serve these\n * weights — so this module owns runtime install (prebuilt release download+extract), GGUF\n * downloads, and llama-server lifecycle. Mirrors local-runtime.ts's OllamaRuntime: injectable\n * seams for fetch/spawn/exists, pi-owned directories under agentDir (runtimes/, models/),\n * detached+tracked child processes, onProgress as best-effort UI feedback, honest error taxonomies\n * instead of silent fallbacks.\n *\n * The fork publishes precompiled per-platform archives on GitHub Releases (no local build): pin a\n * release tag, download the matching asset, extract it, and locate the `llama-server` binary once\n * at install time (release archive layouts vary — bin/ vs build/bin/ at the top level — so this is\n * a deterministic one-time scan, not a fallback). The found relative path and backend (cpu/cuda)\n * are persisted to a manifest so later `detect()`/`serve()` calls never need to re-scan.\n */\n\n/**\n * Pinned Prism ML llama.cpp fork release — verified via\n * `GET https://api.github.com/repos/PrismML-Eng/llama.cpp/releases/tags/prism-b9594-38c66ad` on\n * 2026-07-18 (release exists, all assets referenced below are present on it). Bump here (and\n * re-verify with the same API call) when the provider ships new kernels.\n */\nexport const PRISM_LLAMACPP_PINNED_RELEASE = \"prism-b9594-38c66ad\";\n\nexport const PRISM_LLAMACPP_RELEASES_BASE_URL = \"https://github.com/PrismML-Eng/llama.cpp/releases/download\";\n\nexport interface PrismModelDescriptor {\n\trepo: string;\n\tfile: string;\n\tmmprojFile?: string;\n\tdisplayName: string;\n}\n\n/** Curated first-model descriptor; the `/models add` wiring task consumes this. */\nexport const BONSAI_27B: PrismModelDescriptor = {\n\trepo: \"prism-ml/Bonsai-27B-gguf\",\n\tfile: \"Bonsai-27B-Q1_0.gguf\",\n\tmmprojFile: \"Bonsai-27B-mmproj-Q8_0.gguf\",\n\tdisplayName: \"Bonsai-27B (1-bit Q1_0 + vision)\",\n};\n\nexport type PrismBackend = \"cpu\" | \"cuda\";\nexport type PrismAssetKind = \"tar-gz\" | \"zip\";\n\nexport interface PrismLlamaAsset {\n\tname: string;\n\tkind: PrismAssetKind;\n\tbackend: PrismBackend;\n}\n\n// Verbatim asset names from the pinned release (see PRISM_LLAMACPP_PINNED_RELEASE doc comment) —\n// hardcoded rather than templated from the tag because the provider's naming isn't fully\n// consistent across platforms (Windows CPU assets drop the tag/commit segment entirely; the\n// Windows CUDA asset uses a different build-number segment, \"b1\" not \"b9594\").\nconst LINUX_X64_CPU_ASSET = \"llama-prism-b9594-38c66ad-bin-ubuntu-x64.tar.gz\";\nconst LINUX_ARM64_CPU_ASSET = \"llama-prism-b9594-38c66ad-bin-ubuntu-arm64.tar.gz\";\nconst LINUX_X64_CUDA_ASSET = \"llama-prism-b9594-38c66ad-bin-linux-cuda-12.4-x64.tar.gz\";\nconst MACOS_ARM64_ASSET = \"llama-prism-b9594-38c66ad-bin-macos-arm64.tar.gz\";\nconst MACOS_X64_ASSET = \"llama-prism-b9594-38c66ad-bin-macos-x64.tar.gz\";\nconst WIN_X64_CPU_ASSET = \"llama-bin-win-cpu-x64.zip\";\nconst WIN_ARM64_CPU_ASSET = \"llama-bin-win-cpu-arm64.zip\";\n\n/**\n * Maps a platform/arch/GPU triple to the exact Prism llama.cpp release asset for\n * {@link PRISM_LLAMACPP_PINNED_RELEASE} — verified against the real GitHub release, not guessed.\n * CPU asset by default; the CUDA 12.4 variant is only offered for linux x64 with an NVIDIA GPU\n * present (a 12.8 variant also exists upstream but has no wired caller yet). Windows CUDA needs a\n * companion `cudart-*` archive and is out of scope for v1 — Windows always resolves to the CPU zip\n * regardless of `hasNvidiaGpu`. Pure and exported so it's independently testable.\n */\nexport function resolvePrismLlamaAsset(\n\tplat: string,\n\tarchitecture: string,\n\thasNvidiaGpu: boolean,\n): PrismLlamaAsset | undefined {\n\tif (plat === \"linux\") {\n\t\tif (architecture === \"x64\") {\n\t\t\treturn hasNvidiaGpu\n\t\t\t\t? { name: LINUX_X64_CUDA_ASSET, kind: \"tar-gz\", backend: \"cuda\" }\n\t\t\t\t: { name: LINUX_X64_CPU_ASSET, kind: \"tar-gz\", backend: \"cpu\" };\n\t\t}\n\t\tif (architecture === \"arm64\") return { name: LINUX_ARM64_CPU_ASSET, kind: \"tar-gz\", backend: \"cpu\" };\n\t\treturn undefined;\n\t}\n\tif (plat === \"darwin\") {\n\t\tif (architecture === \"arm64\") return { name: MACOS_ARM64_ASSET, kind: \"tar-gz\", backend: \"cpu\" };\n\t\tif (architecture === \"x64\") return { name: MACOS_X64_ASSET, kind: \"tar-gz\", backend: \"cpu\" };\n\t\treturn undefined;\n\t}\n\tif (plat === \"win32\") {\n\t\tif (architecture === \"x64\") return { name: WIN_X64_CPU_ASSET, kind: \"zip\", backend: \"cpu\" };\n\t\tif (architecture === \"arm64\") return { name: WIN_ARM64_CPU_ASSET, kind: \"zip\", backend: \"cpu\" };\n\t\treturn undefined;\n\t}\n\treturn undefined;\n}\n\nexport interface PrismDetectResult {\n\truntimeInstalled: boolean;\n\tbinaryPath?: string;\n\trelease?: string;\n}\n\nexport interface PrismDownloadResult {\n\tok: boolean;\n\tpath?: string;\n\tskipped?: boolean;\n\terror?: string;\n}\n\nexport type PrismServeResult = { ok: true; baseUrl: string } | { ok: false; error: string };\n\ninterface PrismInstallManifest {\n\trelease: string;\n\tbinaryRelPath: string;\n\tbackend: PrismBackend;\n}\n\ntype PrismExtractArchiveFn = (\n\tinput: Readable,\n\tdestDir: string,\n\tkind: PrismAssetKind,\n) => Promise<{ ok: boolean; error?: string }>;\n\nexport interface PrismLlamaCppDeps {\n\tfetchFn?: typeof fetch;\n\tspawnFn?: ManagedRuntimeSpawn;\n\texistsFn?: (path: string) => boolean;\n\tsleepFn?: (ms: number) => Promise<void>;\n\t/** Whether a named command exists on PATH (nvidia-smi, tar — for zip extraction on Windows). */\n\thasCommand?: (command: string) => boolean;\n\t/** Decided once at install time and persisted as `backend` — serve() reads the persisted value\n\t * so `-ngl 99` reflects what was actually installed, not the host's current GPU state. */\n\thasNvidiaGpu?: () => boolean;\n\tplatform?: () => string;\n\tarch?: () => string;\n\t/** Runs the extraction step for a downloaded archive. Injectable so installManaged's\n\t * download->extract->scan orchestration is testable without a real tar/unzip pipeline; defaults\n\t * to the real spawn-based extractor. */\n\textractArchive?: PrismExtractArchiveFn;\n\t/** Health-poll bounds for serve(); default ~120s (240 * 500ms), both overridable for tests. */\n\thealthPollAttempts?: number;\n\thealthPollIntervalMs?: number;\n}\n\nconst EXTRACTION_TIMEOUT_MS = 10 * 60_000;\nconst COMMAND_KILL_GRACE_MS = 2_000;\nconst DEFAULT_HEALTH_POLL_ATTEMPTS = 240;\nconst DEFAULT_HEALTH_POLL_INTERVAL_MS = 500;\nconst HEALTH_CHECK_TIMEOUT_MS = 2_000;\nconst HEAD_REQUEST_TIMEOUT_MS = 10_000;\n\nfunction parseContentLength(header: string | null): number | undefined {\n\tif (!header) return undefined;\n\tconst value = Number(header);\n\treturn Number.isFinite(value) && value >= 0 ? value : undefined;\n}\n\nexport class PrismLlamaCppRuntime {\n\tprivate readonly _agentDir: string;\n\tprivate readonly _fetch: typeof fetch;\n\tprivate readonly _spawn: ManagedRuntimeSpawn;\n\tprivate readonly _exists: (path: string) => boolean;\n\tprivate readonly _sleep: (ms: number) => Promise<void>;\n\tprivate readonly _hasCommand: (command: string) => boolean;\n\tprivate readonly _hasNvidiaGpu: () => boolean;\n\tprivate readonly _platform: () => string;\n\tprivate readonly _arch: () => string;\n\tprivate readonly _extractArchiveFn: PrismExtractArchiveFn;\n\tprivate readonly _healthPollAttempts: number;\n\tprivate readonly _healthPollIntervalMs: number;\n\tprivate _child: Pick<ChildProcess, \"pid\" | \"kill\" | \"unref\" | \"on\"> | undefined;\n\n\tconstructor(args: { agentDir: string; deps?: PrismLlamaCppDeps }) {\n\t\tthis._agentDir = args.agentDir;\n\t\t[this._fetch, this._spawn, this._exists, this._sleep] = resolveRuntimeLifecycleDependencies(args.deps);\n\t\tthis._hasCommand = args.deps?.hasCommand ?? runtimeCommandAvailable;\n\t\tthis._hasNvidiaGpu = args.deps?.hasNvidiaGpu ?? (() => this._hasCommand(\"nvidia-smi\"));\n\t\tthis._platform = args.deps?.platform ?? osPlatform;\n\t\tthis._arch = args.deps?.arch ?? osArch;\n\t\tthis._extractArchiveFn =\n\t\t\targs.deps?.extractArchive ?? ((input, destDir, kind) => this._extractArchive(input, destDir, kind));\n\t\tthis._healthPollAttempts = args.deps?.healthPollAttempts ?? DEFAULT_HEALTH_POLL_ATTEMPTS;\n\t\tthis._healthPollIntervalMs = args.deps?.healthPollIntervalMs ?? DEFAULT_HEALTH_POLL_INTERVAL_MS;\n\t}\n\n\truntimeDir(): string {\n\t\treturn agentRuntimesDir(\"prism-llamacpp\", this._agentDir);\n\t}\n\n\tmodelsDir(): string {\n\t\treturn agentModelsDir(\"llamacpp\", this._agentDir);\n\t}\n\n\tprivate _binaryName(): string {\n\t\treturn this._platform() === \"win32\" ? \"llama-server.exe\" : \"llama-server\";\n\t}\n\n\tprivate _manifestPath(): string {\n\t\treturn join(this.runtimeDir(), \"install.json\");\n\t}\n\n\tprivate _readManifest(): PrismInstallManifest | undefined {\n\t\ttry {\n\t\t\tconst parsed = JSON.parse(readFileSync(this._manifestPath(), \"utf8\")) as {\n\t\t\t\trelease?: unknown;\n\t\t\t\tbinaryRelPath?: unknown;\n\t\t\t\tbackend?: unknown;\n\t\t\t};\n\t\t\tif (\n\t\t\t\ttypeof parsed.release === \"string\" &&\n\t\t\t\ttypeof parsed.binaryRelPath === \"string\" &&\n\t\t\t\t(parsed.backend === \"cpu\" || parsed.backend === \"cuda\")\n\t\t\t) {\n\t\t\t\treturn { release: parsed.release, binaryRelPath: parsed.binaryRelPath, backend: parsed.backend };\n\t\t\t}\n\t\t\treturn undefined;\n\t\t} catch {\n\t\t\treturn undefined;\n\t\t}\n\t}\n\n\tprivate _writeManifest(manifest: PrismInstallManifest): void {\n\t\twriteFileSync(this._manifestPath(), JSON.stringify(manifest, null, \"\\t\"));\n\t}\n\n\tasync detect(): Promise<PrismDetectResult> {\n\t\tconst manifest = this._readManifest();\n\t\tif (!manifest) return { runtimeInstalled: false };\n\t\tconst binaryPath = join(this.runtimeDir(), manifest.binaryRelPath);\n\t\tconst runtimeInstalled = this._exists(binaryPath);\n\t\treturn {\n\t\t\truntimeInstalled,\n\t\t\tbinaryPath: runtimeInstalled ? binaryPath : undefined,\n\t\t\trelease: manifest.release,\n\t\t};\n\t}\n\n\t/** One-time recursive scan for the `llama-server` binary inside a freshly extracted release\n\t * archive — layouts vary (bin/ vs build/bin/ at the top level) across platforms/backends, so\n\t * this locates it deterministically instead of assuming a fixed path, then the result is\n\t * persisted so later calls never need to re-scan. */\n\tprivate _findBinaryRelPath(root: string): string | undefined {\n\t\tconst targetName = this._binaryName();\n\t\tconst stack: string[] = [root];\n\t\twhile (stack.length > 0) {\n\t\t\tconst dir = stack.pop();\n\t\t\tif (dir === undefined) continue;\n\t\t\tlet entries: Dirent[];\n\t\t\ttry {\n\t\t\t\tentries = readdirSync(dir, { withFileTypes: true });\n\t\t\t} catch {\n\t\t\t\tcontinue;\n\t\t\t}\n\t\t\tfor (const entry of entries) {\n\t\t\t\tconst full = join(dir, entry.name);\n\t\t\t\tif (entry.isDirectory()) stack.push(full);\n\t\t\t\telse if (entry.isFile() && entry.name === targetName) return relative(root, full);\n\t\t\t}\n\t\t}\n\t\treturn undefined;\n\t}\n\n\t/**\n\t * Download the pinned release asset for this host and extract it (consent-gated by the caller,\n\t * same contract as OllamaRuntime#installManaged — this method only does the mechanical\n\t * download+extract+locate and reports the outcome honestly). No compiler toolchain required:\n\t * the fork ships prebuilt binaries.\n\t */\n\tasync installManaged(onProgress?: (status: string) => void): Promise<{ ok: boolean; error?: string }> {\n\t\tconst asset = resolvePrismLlamaAsset(this._platform(), this._arch(), this._hasNvidiaGpu());\n\t\tif (!asset) return { ok: false, error: \"unsupported-platform\" };\n\n\t\tconst downloadUrl = `${PRISM_LLAMACPP_RELEASES_BASE_URL}/${PRISM_LLAMACPP_PINNED_RELEASE}/${asset.name}`;\n\t\tconst extracted = await this._installArchive(downloadUrl, asset, onProgress);\n\t\tif (!extracted.ok) return extracted;\n\n\t\tonProgress?.(\"Locating llama-server binary…\");\n\t\tconst destDir = this.runtimeDir();\n\t\tconst binaryRelPath = this._findBinaryRelPath(destDir);\n\t\tif (!binaryRelPath) {\n\t\t\treturn { ok: false, error: \"binary-missing: no llama-server binary found in the extracted archive\" };\n\t\t}\n\n\t\tthis._writeManifest({ release: PRISM_LLAMACPP_PINNED_RELEASE, binaryRelPath, backend: asset.backend });\n\t\tonProgress?.(\"Prism llama.cpp runtime installed.\");\n\t\treturn { ok: true };\n\t}\n\n\tprivate _installArchive(\n\t\tdownloadUrl: string,\n\t\tasset: PrismLlamaAsset,\n\t\tonProgress?: (status: string) => void,\n\t): Promise<{ ok: boolean; error?: string }> {\n\t\treturn installRuntimeArchive(\n\t\t\tthis._fetch,\n\t\t\tdownloadUrl,\n\t\t\tthis.runtimeDir(),\n\t\t\tasset,\n\t\t\tthis._extractArchiveFn,\n\t\t\tonProgress,\n\t\t);\n\t}\n\n\tprivate async _extractArchive(\n\t\tinput: Readable,\n\t\tdestDir: string,\n\t\tkind: PrismAssetKind,\n\t): Promise<{ ok: boolean; error?: string }> {\n\t\ttry {\n\t\t\treturn kind === \"zip\" ? await this._extractZip(input, destDir) : await this._extractTarGz(input, destDir);\n\t\t} catch (error) {\n\t\t\treturn { ok: false, error: `extract-fail: ${error instanceof Error ? error.message : String(error)}` };\n\t\t}\n\t}\n\n\tprivate async _extractTarGz(input: Readable, destDir: string): Promise<{ ok: boolean; error?: string }> {\n\t\tconst tarProc = spawnProcess(\"tar\", [\"-xzf\", \"-\", \"-C\", destDir], {\n\t\t\tdetached: process.platform !== \"win32\",\n\t\t\tstdio: [\"pipe\", \"ignore\", \"ignore\"],\n\t\t});\n\t\tconst terminationController = new AbortController();\n\t\tconst processWait = waitForChildProcessWithTermination(tarProc, {\n\t\t\tsignal: terminationController.signal,\n\t\t\ttimeoutMs: EXTRACTION_TIMEOUT_MS,\n\t\t\tkillGraceMs: COMMAND_KILL_GRACE_MS,\n\t\t});\n\t\ttry {\n\t\t\tawait pipeline(input, requireRuntimeStdin(tarProc, \"tar\"));\n\t\t} catch (error) {\n\t\t\tterminationController.abort();\n\t\t\tawait processWait.catch(() => {});\n\t\t\tthrow error;\n\t\t}\n\t\tconst terminal = await processWait;\n\t\tif (terminal.reason === \"timeout\" || terminal.code !== 0) {\n\t\t\treturn {\n\t\t\t\tok: false,\n\t\t\t\terror:\n\t\t\t\t\tterminal.reason === \"timeout\"\n\t\t\t\t\t\t? `extract-fail: tar extraction timed out after ${EXTRACTION_TIMEOUT_MS}ms`\n\t\t\t\t\t\t: `extract-fail: tar exited with code ${terminal.code ?? \"unknown\"}`,\n\t\t\t};\n\t\t}\n\t\treturn { ok: true };\n\t}\n\n\tprivate async _extractZip(input: Readable, destDir: string): Promise<{ ok: boolean; error?: string }> {\n\t\treturn extractZipArchive({\n\t\t\tinput,\n\t\t\tdestDir,\n\t\t\ttempPrefix: \"prism-llamacpp-download\",\n\t\t\tplatform: this._platform,\n\t\t\thasCommand: this._hasCommand,\n\t\t\ttimeoutMs: EXTRACTION_TIMEOUT_MS,\n\t\t\tkillGraceMs: COMMAND_KILL_GRACE_MS,\n\t\t});\n\t}\n\n\tprivate async _remoteContentLength(url: string): Promise<number | undefined> {\n\t\ttry {\n\t\t\tconst response = await this._fetch(url, {\n\t\t\t\tmethod: \"HEAD\",\n\t\t\t\tsignal: AbortSignal.timeout(HEAD_REQUEST_TIMEOUT_MS),\n\t\t\t});\n\t\t\treturn response.ok ? parseContentLength(response.headers.get(\"content-length\")) : undefined;\n\t\t} catch {\n\t\t\treturn undefined;\n\t\t}\n\t}\n\n\t/**\n\t * Stream a GGUF (or mmproj) file from Hugging Face into pi's owned models dir. Skips a re-download\n\t * when the local file already matches the remote size; verifies size when the response reports\n\t * `content-length` and deletes the partial file on any failure or mismatch — never leaves a\n\t * corrupt/truncated weight file behind for a later load to silently misread.\n\t */\n\tasync downloadModel(\n\t\targs: { repo: string; file: string },\n\t\tonProgress?: (status: string) => void,\n\t): Promise<PrismDownloadResult> {\n\t\tconst destPath = join(this.modelsDir(), args.repo, args.file);\n\t\tconst url = `https://huggingface.co/${args.repo}/resolve/main/${args.file}`;\n\t\tmkdirSync(dirname(destPath), { recursive: true });\n\n\t\tif (this._exists(destPath)) {\n\t\t\tconst remoteSize = await this._remoteContentLength(url);\n\t\t\tif (remoteSize !== undefined && remoteSize === tryFileSizeBytes(destPath)) {\n\t\t\t\tonProgress?.(`${args.file} already downloaded (${remoteSize} bytes) — skipping.`);\n\t\t\t\treturn { ok: true, path: destPath, skipped: true };\n\t\t\t}\n\t\t}\n\n\t\tonProgress?.(`Downloading ${args.file} from ${args.repo}…`);\n\t\tconst download = await fetchRuntimeDownload(this._fetch, url);\n\t\tif (!download.ok) return download;\n\n\t\tconst expectedBytes = parseContentLength(download.response.headers.get(\"content-length\"));\n\t\tconst written = await writeRuntimeDownload(download.body as unknown as Readable, destPath);\n\t\tif (!written.ok) return written;\n\n\t\tconst actualBytes = tryFileSizeBytes(destPath);\n\t\tif (expectedBytes !== undefined && actualBytes !== expectedBytes) {\n\t\t\tremovePartialDownload(destPath);\n\t\t\treturn { ok: false, error: `size-mismatch: expected ${expectedBytes} bytes, got ${actualBytes ?? 0}` };\n\t\t}\n\n\t\tonProgress?.(`${args.file} downloaded (${actualBytes ?? 0} bytes).`);\n\t\treturn { ok: true, path: destPath };\n\t}\n\n\tprivate async _healthUp(baseUrl: string): Promise<boolean> {\n\t\ttry {\n\t\t\tconst response = await this._fetch(`${baseUrl}/health`, {\n\t\t\t\tsignal: AbortSignal.timeout(HEALTH_CHECK_TIMEOUT_MS),\n\t\t\t});\n\t\t\treturn response.ok;\n\t\t} catch {\n\t\t\treturn false;\n\t\t}\n\t}\n\n\t/**\n\t * Spawn `llama-server` detached+tracked (killed on parent shutdown even if pi crashes without\n\t * calling stop()) and poll `/health` until ready. `-ngl 99` is only added when the installed\n\t * asset's persisted `backend` is `\"cuda\"`, not from the host's current GPU state.\n\t */\n\tasync serve(args: {\n\t\tmodelPath: string;\n\t\tmmprojPath?: string;\n\t\tport: number;\n\t\tnumCtx: number;\n\t}): Promise<PrismServeResult> {\n\t\tconst manifest = this._readManifest();\n\t\tif (!manifest) return { ok: false, error: \"binary-missing\" };\n\t\tconst binaryPath = join(this.runtimeDir(), manifest.binaryRelPath);\n\t\tif (!this._exists(binaryPath)) return { ok: false, error: \"binary-missing\" };\n\n\t\tconst baseUrl = `http://127.0.0.1:${args.port}`;\n\t\tconst argv = [\n\t\t\t\"-m\",\n\t\t\targs.modelPath,\n\t\t\t...(args.mmprojPath ? [\"--mmproj\", args.mmprojPath] : []),\n\t\t\t\"--host\",\n\t\t\t\"127.0.0.1\",\n\t\t\t\"--port\",\n\t\t\tString(args.port),\n\t\t\t\"-c\",\n\t\t\tString(args.numCtx),\n\t\t\t...(manifest.backend === \"cuda\" ? [\"-ngl\", \"99\"] : []),\n\t\t];\n\n\t\tconst child = this._spawn(binaryPath, argv, {\n\t\t\tdetached: process.platform !== \"win32\",\n\t\t\tstdio: \"ignore\",\n\t\t\tenv: process.env,\n\t\t});\n\t\tif (child.pid) trackDetachedChildPid(child.pid);\n\t\tchild.unref?.();\n\t\tchild.on(\"exit\", () => {\n\t\t\tthis._child = undefined;\n\t\t});\n\t\tthis._child = child;\n\n\t\tfor (let attempt = 0; attempt < this._healthPollAttempts; attempt++) {\n\t\t\tif (await this._healthUp(baseUrl)) return { ok: true, baseUrl };\n\t\t\tawait this._sleep(this._healthPollIntervalMs);\n\t\t}\n\t\tthis.stop();\n\t\treturn { ok: false, error: \"health-timeout\" };\n\t}\n\n\tstop(): { stopped: boolean } {\n\t\tconst child = this._child;\n\t\tif (!child) return { stopped: false };\n\t\tthis._child = undefined;\n\t\tif (child.pid) {\n\t\t\tuntrackDetachedChildPid(child.pid);\n\t\t\tkillProcessTree(child.pid);\n\t\t}\n\t\treturn { stopped: true };\n\t}\n\n\tisRunning(): boolean {\n\t\treturn this._child !== undefined;\n\t}\n}\n"]}
+{"version":3,"file":"llamacpp-runtime.js","sourceRoot":"","sources":["../../../src/core/models/llamacpp-runtime.ts"],"names":[],"mappings":"AACA,OAAO,EAAe,SAAS,EAAE,WAAW,EAAE,YAAY,EAAE,aAAa,EAAE,MAAM,SAAS,CAAC;AAC3F,OAAO,EAAE,IAAI,IAAI,MAAM,EAAE,QAAQ,IAAI,UAAU,EAAE,MAAM,SAAS,CAAC;AACjE,OAAO,EAAE,OAAO,EAAE,IAAI,EAAE,QAAQ,EAAE,MAAM,WAAW,CAAC;AAEpD,OAAO,EAAE,QAAQ,EAAE,MAAM,sBAAsB,CAAC;AAChD,OAAO,EAAE,YAAY,EAAE,kCAAkC,EAAE,MAAM,8BAA8B,CAAC;AAChG,OAAO,EAAE,eAAe,EAAE,qBAAqB,EAAE,uBAAuB,EAAE,MAAM,sBAAsB,CAAC;AACvG,OAAO,EAAE,SAAS,IAAI,cAAc,EAAE,WAAW,IAAI,gBAAgB,EAAE,MAAM,mBAAmB,CAAC;AACjG,OAAO,EACN,+BAA+B,GAG/B,MAAM,8BAA8B,CAAC;AACtC,OAAO,EAAE,wBAAwB,EAAE,MAAM,kCAAkC,CAAC;AAC5E,OAAO,EACN,iBAAiB,EACjB,oBAAoB,EACpB,qBAAqB,EAErB,qBAAqB,EACrB,mBAAmB,EACnB,mCAAmC,EACnC,uBAAuB,EACvB,gBAAgB,EAChB,oBAAoB,GACpB,MAAM,sBAAsB,CAAC;AAE9B;;;;;;;;;;;;;;GAcG;AAEH;;;;;GAKG;AACH,MAAM,CAAC,MAAM,6BAA6B,GAAG,qBAAqB,CAAC;AAEnE,MAAM,CAAC,MAAM,gCAAgC,GAAG,4DAA4D,CAAC;AAqB7G,yEAAyE;AACzE,MAAM,CAAC,MAAM,UAAU,GAAyB;IAC/C,IAAI,EAAE,0BAA0B;IAChC,IAAI,EAAE,sBAAsB;IAC5B,UAAU,EAAE,6BAA6B;IACzC,WAAW,EAAE,kCAAkC;IAC/C,YAAY,EAAE,OAAO;IACrB,OAAO,EAAE,gBAAgB;IACzB,MAAM,EAAE,QAAQ;IAChB,cAAc,EAAE,KAAK;IACrB,YAAY,EAAE,MAAM;IACpB,cAAc,EAAE;QACf,IAAI,EAAE,QAAQ;QACd,IAAI,EAAE,6BAA6B;QACnC,QAAQ,EAAE,CAAC;QACX,cAAc,EAAE,MAAM;QACtB,gBAAgB,EAAE,MAAM;KACxB;CACD,CAAC;AAeF,iGAAiG;AACjG,yFAAyF;AACzF,4FAA4F;AAC5F,+EAA+E;AAC/E,MAAM,mBAAmB,GAAG,iDAAiD,CAAC;AAC9E,MAAM,qBAAqB,GAAG,mDAAmD,CAAC;AAClF,MAAM,oBAAoB,GAAG,0DAA0D,CAAC;AACxF,MAAM,iBAAiB,GAAG,kDAAkD,CAAC;AAC7E,MAAM,eAAe,GAAG,gDAAgD,CAAC;AACzE,MAAM,iBAAiB,GAAG,2BAA2B,CAAC;AACtD,MAAM,mBAAmB,GAAG,6BAA6B,CAAC;AAC1D,MAAM,kBAAkB,GAAG,kDAAkD,CAAC;AAC9E,MAAM,4BAA4B,GAAG,wCAAwC,CAAC;AAE9E;;;;;;;GAOG;AACH,MAAM,UAAU,sBAAsB,CACrC,IAAY,EACZ,YAAoB,EACpB,YAAqB;IAErB,IAAI,IAAI,KAAK,OAAO,EAAE,CAAC;QACtB,IAAI,YAAY,KAAK,KAAK,EAAE,CAAC;YAC5B,OAAO,YAAY;gBAClB,CAAC,CAAC,EAAE,IAAI,EAAE,oBAAoB,EAAE,IAAI,EAAE,QAAQ,EAAE,OAAO,EAAE,MAAM,EAAE;gBACjE,CAAC,CAAC,EAAE,IAAI,EAAE,mBAAmB,EAAE,IAAI,EAAE,QAAQ,EAAE,OAAO,EAAE,KAAK,EAAE,CAAC;QAClE,CAAC;QACD,IAAI,YAAY,KAAK,OAAO;YAAE,OAAO,EAAE,IAAI,EAAE,qBAAqB,EAAE,IAAI,EAAE,QAAQ,EAAE,OAAO,EAAE,KAAK,EAAE,CAAC;QACrG,OAAO,SAAS,CAAC;IAClB,CAAC;IACD,IAAI,IAAI,KAAK,QAAQ,EAAE,CAAC;QACvB,IAAI,YAAY,KAAK,OAAO;YAAE,OAAO,EAAE,IAAI,EAAE,iBAAiB,EAAE,IAAI,EAAE,QAAQ,EAAE,OAAO,EAAE,KAAK,EAAE,CAAC;QACjG,IAAI,YAAY,KAAK,KAAK;YAAE,OAAO,EAAE,IAAI,EAAE,eAAe,EAAE,IAAI,EAAE,QAAQ,EAAE,OAAO,EAAE,KAAK,EAAE,CAAC;QAC7F,OAAO,SAAS,CAAC;IAClB,CAAC;IACD,IAAI,IAAI,KAAK,OAAO,EAAE,CAAC;QACtB,IAAI,YAAY,KAAK,KAAK,EAAE,CAAC;YAC5B,OAAO,YAAY;gBAClB,CAAC,CAAC;oBACA,IAAI,EAAE,kBAAkB;oBACxB,IAAI,EAAE,KAAK;oBACX,OAAO,EAAE,MAAM;oBACf,eAAe,EAAE,CAAC,EAAE,IAAI,EAAE,4BAA4B,EAAE,IAAI,EAAE,KAAK,EAAE,CAAC;iBACtE;gBACF,CAAC,CAAC,EAAE,IAAI,EAAE,iBAAiB,EAAE,IAAI,EAAE,KAAK,EAAE,OAAO,EAAE,KAAK,EAAE,CAAC;QAC7D,CAAC;QACD,IAAI,YAAY,KAAK,OAAO;YAAE,OAAO,EAAE,IAAI,EAAE,mBAAmB,EAAE,IAAI,EAAE,KAAK,EAAE,OAAO,EAAE,KAAK,EAAE,CAAC;QAChG,OAAO,SAAS,CAAC;IAClB,CAAC;IACD,OAAO,SAAS,CAAC;AAClB,CAAC;AAqDD,MAAM,qBAAqB,GAAG,EAAE,GAAG,MAAM,CAAC;AAC1C,MAAM,qBAAqB,GAAG,KAAK,CAAC;AACpC,MAAM,4BAA4B,GAAG,GAAG,CAAC;AACzC,MAAM,+BAA+B,GAAG,GAAG,CAAC;AAC5C,MAAM,uBAAuB,GAAG,KAAK,CAAC;AACtC,MAAM,uBAAuB,GAAG,MAAM,CAAC;AAEvC,SAAS,kBAAkB,CAAC,MAAqB;IAChD,IAAI,CAAC,MAAM;QAAE,OAAO,SAAS,CAAC;IAC9B,MAAM,KAAK,GAAG,MAAM,CAAC,MAAM,CAAC,CAAC;IAC7B,OAAO,MAAM,CAAC,QAAQ,CAAC,KAAK,CAAC,IAAI,KAAK,IAAI,CAAC,CAAC,CAAC,CAAC,KAAK,CAAC,CAAC,CAAC,SAAS,CAAC;AACjE,CAAC;AAED,SAAS,0BAA0B,CAAC,cAAiC;IACpE,OAAO,2BAA2B,cAAc,CAAC,IAAI,CAAC,GAAG,CAAC,IAAI,SAAS,EAAE,CAAC;AAC3E,CAAC;AAED,MAAM,OAAO,oBAAoB;IAgBhC,YAAY,IAIX;QACA,IAAI,CAAC,SAAS,GAAG,IAAI,CAAC,QAAQ,CAAC;QAC/B,CAAC,IAAI,CAAC,MAAM,EAAE,IAAI,CAAC,MAAM,EAAE,IAAI,CAAC,OAAO,EAAE,IAAI,CAAC,MAAM,CAAC,GAAG,mCAAmC,CAAC,IAAI,CAAC,IAAI,CAAC,CAAC;QACvG,IAAI,CAAC,WAAW,GAAG,IAAI,CAAC,IAAI,EAAE,UAAU,IAAI,uBAAuB,CAAC;QACpE,IAAI,CAAC,aAAa,GAAG,IAAI,CAAC,IAAI,EAAE,YAAY,IAAI,CAAC,GAAG,EAAE,CAAC,IAAI,CAAC,WAAW,CAAC,YAAY,CAAC,CAAC,CAAC;QACvF,IAAI,CAAC,SAAS,GAAG,IAAI,CAAC,IAAI,EAAE,QAAQ,IAAI,UAAU,CAAC;QACnD,IAAI,CAAC,KAAK,GAAG,IAAI,CAAC,IAAI,EAAE,IAAI,IAAI,MAAM,CAAC;QACvC,IAAI,CAAC,QAAQ,GAAG,+BAA+B,CAAC,IAAI,CAAC,WAAW,IAAI,UAAU,EAAE,IAAI,CAAC,IAAI,CAAC,CAAC;QAC3F,IAAI,CAAC,iBAAiB;YACrB,IAAI,CAAC,IAAI,EAAE,cAAc,IAAI,CAAC,CAAC,KAAK,EAAE,OAAO,EAAE,IAAI,EAAE,EAAE,CAAC,IAAI,CAAC,eAAe,CAAC,KAAK,EAAE,OAAO,EAAE,IAAI,CAAC,CAAC,CAAC;QACrG,IAAI,CAAC,mBAAmB,GAAG,IAAI,CAAC,IAAI,EAAE,kBAAkB,IAAI,4BAA4B,CAAC;QACzF,IAAI,CAAC,qBAAqB,GAAG,IAAI,CAAC,IAAI,EAAE,oBAAoB,IAAI,+BAA+B,CAAC;IACjG,CAAC;IAED,UAAU;QACT,OAAO,gBAAgB,CAAC,gBAAgB,EAAE,IAAI,CAAC,SAAS,CAAC,CAAC;IAC3D,CAAC;IAED,SAAS;QACR,OAAO,cAAc,CAAC,UAAU,EAAE,IAAI,CAAC,SAAS,CAAC,CAAC;IACnD,CAAC;IAEO,WAAW;QAClB,OAAO,IAAI,CAAC,SAAS,EAAE,KAAK,OAAO,CAAC,CAAC,CAAC,kBAAkB,CAAC,CAAC,CAAC,cAAc,CAAC;IAC3E,CAAC;IAEO,aAAa;QACpB,OAAO,IAAI,CAAC,IAAI,CAAC,UAAU,EAAE,EAAE,cAAc,CAAC,CAAC;IAChD,CAAC;IAEO,aAAa;QACpB,IAAI,CAAC;YACJ,MAAM,MAAM,GAAG,IAAI,CAAC,KAAK,CAAC,YAAY,CAAC,IAAI,CAAC,aAAa,EAAE,EAAE,MAAM,CAAC,CAInE,CAAC;YACF,IACC,OAAO,MAAM,CAAC,OAAO,KAAK,QAAQ;gBAClC,OAAO,MAAM,CAAC,aAAa,KAAK,QAAQ;gBACxC,CAAC,MAAM,CAAC,OAAO,KAAK,KAAK,IAAI,MAAM,CAAC,OAAO,KAAK,MAAM,CAAC,EACtD,CAAC;gBACF,OAAO,EAAE,OAAO,EAAE,MAAM,CAAC,OAAO,EAAE,aAAa,EAAE,MAAM,CAAC,aAAa,EAAE,OAAO,EAAE,MAAM,CAAC,OAAO,EAAE,CAAC;YAClG,CAAC;YACD,OAAO,SAAS,CAAC;QAClB,CAAC;QAAC,MAAM,CAAC;YACR,OAAO,SAAS,CAAC;QAClB,CAAC;IACF,CAAC;IAEO,cAAc,CAAC,QAA8B;QACpD,aAAa,CAAC,IAAI,CAAC,aAAa,EAAE,EAAE,IAAI,CAAC,SAAS,CAAC,QAAQ,EAAE,IAAI,EAAE,IAAI,CAAC,CAAC,CAAC;IAC3E,CAAC;IAED,KAAK,CAAC,MAAM;QACX,MAAM,QAAQ,GAAG,IAAI,CAAC,aAAa,EAAE,CAAC;QACtC,IAAI,CAAC,QAAQ;YAAE,OAAO,EAAE,gBAAgB,EAAE,KAAK,EAAE,CAAC;QAClD,MAAM,UAAU,GAAG,IAAI,CAAC,IAAI,CAAC,UAAU,EAAE,EAAE,QAAQ,CAAC,aAAa,CAAC,CAAC;QACnE,MAAM,gBAAgB,GAAG,IAAI,CAAC,OAAO,CAAC,UAAU,CAAC,CAAC;QAClD,OAAO;YACN,gBAAgB;YAChB,UAAU,EAAE,gBAAgB,CAAC,CAAC,CAAC,UAAU,CAAC,CAAC,CAAC,SAAS;YACrD,OAAO,EAAE,QAAQ,CAAC,OAAO;SACzB,CAAC;IACH,CAAC;IAED;;;yDAGqD;IAC7C,kBAAkB,CAAC,IAAY;QACtC,MAAM,UAAU,GAAG,IAAI,CAAC,WAAW,EAAE,CAAC;QACtC,MAAM,KAAK,GAAa,CAAC,IAAI,CAAC,CAAC;QAC/B,OAAO,KAAK,CAAC,MAAM,GAAG,CAAC,EAAE,CAAC;YACzB,MAAM,GAAG,GAAG,KAAK,CAAC,GAAG,EAAE,CAAC;YACxB,IAAI,GAAG,KAAK,SAAS;gBAAE,SAAS;YAChC,IAAI,OAAiB,CAAC;YACtB,IAAI,CAAC;gBACJ,OAAO,GAAG,WAAW,CAAC,GAAG,EAAE,EAAE,aAAa,EAAE,IAAI,EAAE,CAAC,CAAC;YACrD,CAAC;YAAC,MAAM,CAAC;gBACR,SAAS;YACV,CAAC;YACD,KAAK,MAAM,KAAK,IAAI,OAAO,EAAE,CAAC;gBAC7B,MAAM,IAAI,GAAG,IAAI,CAAC,GAAG,EAAE,KAAK,CAAC,IAAI,CAAC,CAAC;gBACnC,IAAI,KAAK,CAAC,WAAW,EAAE;oBAAE,KAAK,CAAC,IAAI,CAAC,IAAI,CAAC,CAAC;qBACrC,IAAI,KAAK,CAAC,MAAM,EAAE,IAAI,KAAK,CAAC,IAAI,KAAK,UAAU;oBAAE,OAAO,QAAQ,CAAC,IAAI,EAAE,IAAI,CAAC,CAAC;YACnF,CAAC;QACF,CAAC;QACD,OAAO,SAAS,CAAC;IAClB,CAAC;IAED;;;;;OAKG;IACH,KAAK,CAAC,cAAc,CAAC,UAAqC;QACzD,MAAM,KAAK,GAAG,sBAAsB,CAAC,IAAI,CAAC,SAAS,EAAE,EAAE,IAAI,CAAC,KAAK,EAAE,EAAE,IAAI,CAAC,aAAa,EAAE,CAAC,CAAC;QAC3F,IAAI,CAAC,KAAK;YAAE,OAAO,EAAE,EAAE,EAAE,KAAK,EAAE,KAAK,EAAE,sBAAsB,EAAE,CAAC;QAEhE,MAAM,QAAQ,GAAiC,CAAC,KAAK,EAAE,GAAG,CAAC,KAAK,CAAC,eAAe,IAAI,EAAE,CAAC,CAAC,CAAC;QACzF,KAAK,MAAM,OAAO,IAAI,QAAQ,EAAE,CAAC;YAChC,MAAM,WAAW,GAAG,GAAG,gCAAgC,IAAI,6BAA6B,IAAI,OAAO,CAAC,IAAI,EAAE,CAAC;YAC3G,MAAM,SAAS,GAAG,MAAM,IAAI,CAAC,eAAe,CAAC,WAAW,EAAE,OAAO,EAAE,UAAU,CAAC,CAAC;YAC/E,IAAI,CAAC,SAAS,CAAC,EAAE;gBAAE,OAAO,SAAS,CAAC;QACrC,CAAC;QAED,UAAU,EAAE,CAAC,+BAA+B,CAAC,CAAC;QAC9C,MAAM,OAAO,GAAG,IAAI,CAAC,UAAU,EAAE,CAAC;QAClC,MAAM,aAAa,GAAG,IAAI,CAAC,kBAAkB,CAAC,OAAO,CAAC,CAAC;QACvD,IAAI,CAAC,aAAa,EAAE,CAAC;YACpB,OAAO,EAAE,EAAE,EAAE,KAAK,EAAE,KAAK,EAAE,uEAAuE,EAAE,CAAC;QACtG,CAAC;QAED,IAAI,CAAC,cAAc,CAAC,EAAE,OAAO,EAAE,6BAA6B,EAAE,aAAa,EAAE,OAAO,EAAE,KAAK,CAAC,OAAO,EAAE,CAAC,CAAC;QACvG,UAAU,EAAE,CAAC,oCAAoC,CAAC,CAAC;QACnD,OAAO,EAAE,EAAE,EAAE,IAAI,EAAE,CAAC;IACrB,CAAC;IAEO,eAAe,CACtB,WAAmB,EACnB,KAAwB,EACxB,UAAqC;QAErC,OAAO,qBAAqB,CAC3B,IAAI,CAAC,MAAM,EACX,WAAW,EACX,IAAI,CAAC,UAAU,EAAE,EACjB,KAAK,EACL,IAAI,CAAC,iBAAiB,EACtB,UAAU,CACV,CAAC;IACH,CAAC;IAEO,KAAK,CAAC,eAAe,CAC5B,KAAe,EACf,OAAe,EACf,IAAoB;QAEpB,IAAI,CAAC;YACJ,OAAO,IAAI,KAAK,KAAK,CAAC,CAAC,CAAC,MAAM,IAAI,CAAC,WAAW,CAAC,KAAK,EAAE,OAAO,CAAC,CAAC,CAAC,CAAC,MAAM,IAAI,CAAC,aAAa,CAAC,KAAK,EAAE,OAAO,CAAC,CAAC;QAC3G,CAAC;QAAC,OAAO,KAAK,EAAE,CAAC;YAChB,OAAO,EAAE,EAAE,EAAE,KAAK,EAAE,KAAK,EAAE,iBAAiB,KAAK,YAAY,KAAK,CAAC,CAAC,CAAC,KAAK,CAAC,OAAO,CAAC,CAAC,CAAC,MAAM,CAAC,KAAK,CAAC,EAAE,EAAE,CAAC;QACxG,CAAC;IACF,CAAC;IAEO,KAAK,CAAC,aAAa,CAAC,KAAe,EAAE,OAAe;QAC3D,MAAM,OAAO,GAAG,YAAY,CAAC,KAAK,EAAE,CAAC,MAAM,EAAE,GAAG,EAAE,IAAI,EAAE,OAAO,CAAC,EAAE;YACjE,QAAQ,EAAE,OAAO,CAAC,QAAQ,KAAK,OAAO;YACtC,KAAK,EAAE,CAAC,MAAM,EAAE,QAAQ,EAAE,QAAQ,CAAC;SACnC,CAAC,CAAC;QACH,MAAM,qBAAqB,GAAG,IAAI,eAAe,EAAE,CAAC;QACpD,MAAM,WAAW,GAAG,kCAAkC,CAAC,OAAO,EAAE;YAC/D,MAAM,EAAE,qBAAqB,CAAC,MAAM;YACpC,SAAS,EAAE,qBAAqB;YAChC,WAAW,EAAE,qBAAqB;SAClC,CAAC,CAAC;QACH,IAAI,CAAC;YACJ,MAAM,QAAQ,CAAC,KAAK,EAAE,mBAAmB,CAAC,OAAO,EAAE,KAAK,CAAC,CAAC,CAAC;QAC5D,CAAC;QAAC,OAAO,KAAK,EAAE,CAAC;YAChB,qBAAqB,CAAC,KAAK,EAAE,CAAC;YAC9B,MAAM,WAAW,CAAC,KAAK,CAAC,GAAG,EAAE,GAAE,CAAC,CAAC,CAAC;YAClC,MAAM,KAAK,CAAC;QACb,CAAC;QACD,MAAM,QAAQ,GAAG,MAAM,WAAW,CAAC;QACnC,IAAI,QAAQ,CAAC,MAAM,KAAK,SAAS,IAAI,QAAQ,CAAC,IAAI,KAAK,CAAC,EAAE,CAAC;YAC1D,OAAO;gBACN,EAAE,EAAE,KAAK;gBACT,KAAK,EACJ,QAAQ,CAAC,MAAM,KAAK,SAAS;oBAC5B,CAAC,CAAC,gDAAgD,qBAAqB,IAAI;oBAC3E,CAAC,CAAC,sCAAsC,QAAQ,CAAC,IAAI,IAAI,SAAS,EAAE;aACtE,CAAC;QACH,CAAC;QACD,OAAO,EAAE,EAAE,EAAE,IAAI,EAAE,CAAC;IACrB,CAAC;IAEO,KAAK,CAAC,WAAW,CAAC,KAAe,EAAE,OAAe;QACzD,OAAO,iBAAiB,CAAC;YACxB,KAAK;YACL,OAAO;YACP,UAAU,EAAE,yBAAyB;YACrC,QAAQ,EAAE,IAAI,CAAC,SAAS;YACxB,UAAU,EAAE,IAAI,CAAC,WAAW;YAC5B,SAAS,EAAE,qBAAqB;YAChC,WAAW,EAAE,qBAAqB;SAClC,CAAC,CAAC;IACJ,CAAC;IAEO,KAAK,CAAC,oBAAoB,CAAC,GAAW;QAC7C,IAAI,CAAC;YACJ,MAAM,QAAQ,GAAG,MAAM,IAAI,CAAC,MAAM,CAAC,GAAG,EAAE;gBACvC,MAAM,EAAE,MAAM;gBACd,MAAM,EAAE,WAAW,CAAC,OAAO,CAAC,uBAAuB,CAAC;aACpD,CAAC,CAAC;YACH,OAAO,QAAQ,CAAC,EAAE,CAAC,CAAC,CAAC,kBAAkB,CAAC,QAAQ,CAAC,OAAO,CAAC,GAAG,CAAC,gBAAgB,CAAC,CAAC,CAAC,CAAC,CAAC,SAAS,CAAC;QAC7F,CAAC;QAAC,MAAM,CAAC;YACR,OAAO,SAAS,CAAC;QAClB,CAAC;IACF,CAAC;IAED;;;;;OAKG;IACH,KAAK,CAAC,aAAa,CAClB,IAAoC,EACpC,UAAqC;QAErC,MAAM,QAAQ,GAAG,IAAI,CAAC,IAAI,CAAC,SAAS,EAAE,EAAE,IAAI,CAAC,IAAI,EAAE,IAAI,CAAC,IAAI,CAAC,CAAC;QAC9D,MAAM,GAAG,GAAG,0BAA0B,IAAI,CAAC,IAAI,iBAAiB,IAAI,CAAC,IAAI,EAAE,CAAC;QAC5E,SAAS,CAAC,OAAO,CAAC,QAAQ,CAAC,EAAE,EAAE,SAAS,EAAE,IAAI,EAAE,CAAC,CAAC;QAElD,IAAI,IAAI,CAAC,OAAO,CAAC,QAAQ,CAAC,EAAE,CAAC;YAC5B,MAAM,UAAU,GAAG,MAAM,IAAI,CAAC,oBAAoB,CAAC,GAAG,CAAC,CAAC;YACxD,IAAI,UAAU,KAAK,SAAS,IAAI,UAAU,KAAK,gBAAgB,CAAC,QAAQ,CAAC,EAAE,CAAC;gBAC3E,UAAU,EAAE,CAAC,GAAG,IAAI,CAAC,IAAI,wBAAwB,UAAU,qBAAqB,CAAC,CAAC;gBAClF,OAAO,EAAE,EAAE,EAAE,IAAI,EAAE,IAAI,EAAE,QAAQ,EAAE,OAAO,EAAE,IAAI,EAAE,CAAC;YACpD,CAAC;QACF,CAAC;QAED,UAAU,EAAE,CAAC,eAAe,IAAI,CAAC,IAAI,SAAS,IAAI,CAAC,IAAI,GAAG,CAAC,CAAC;QAC5D,MAAM,QAAQ,GAAG,MAAM,oBAAoB,CAAC,IAAI,CAAC,MAAM,EAAE,GAAG,CAAC,CAAC;QAC9D,IAAI,CAAC,QAAQ,CAAC,EAAE;YAAE,OAAO,QAAQ,CAAC;QAElC,MAAM,aAAa,GAAG,kBAAkB,CAAC,QAAQ,CAAC,QAAQ,CAAC,OAAO,CAAC,GAAG,CAAC,gBAAgB,CAAC,CAAC,CAAC;QAC1F,MAAM,OAAO,GAAG,MAAM,oBAAoB,CAAC,QAAQ,CAAC,IAA2B,EAAE,QAAQ,CAAC,CAAC;QAC3F,IAAI,CAAC,OAAO,CAAC,EAAE;YAAE,OAAO,OAAO,CAAC;QAEhC,MAAM,WAAW,GAAG,gBAAgB,CAAC,QAAQ,CAAC,CAAC;QAC/C,IAAI,aAAa,KAAK,SAAS,IAAI,WAAW,KAAK,aAAa,EAAE,CAAC;YAClE,qBAAqB,CAAC,QAAQ,CAAC,CAAC;YAChC,OAAO,EAAE,EAAE,EAAE,KAAK,EAAE,KAAK,EAAE,2BAA2B,aAAa,eAAe,WAAW,IAAI,CAAC,EAAE,EAAE,CAAC;QACxG,CAAC;QAED,UAAU,EAAE,CAAC,GAAG,IAAI,CAAC,IAAI,gBAAgB,WAAW,IAAI,CAAC,UAAU,CAAC,CAAC;QACrE,OAAO,EAAE,EAAE,EAAE,IAAI,EAAE,IAAI,EAAE,QAAQ,EAAE,CAAC;IACrC,CAAC;IAED;;;;OAIG;IACH,KAAK,CAAC,KAAK,CAAC,IAMX;QACA,MAAM,QAAQ,GAAG,IAAI,CAAC,aAAa,EAAE,CAAC;QACtC,IAAI,CAAC,QAAQ;YAAE,OAAO,EAAE,EAAE,EAAE,KAAK,EAAE,KAAK,EAAE,gBAAgB,EAAE,CAAC;QAC7D,MAAM,UAAU,GAAG,IAAI,CAAC,IAAI,CAAC,UAAU,EAAE,EAAE,QAAQ,CAAC,aAAa,CAAC,CAAC;QACnE,IAAI,CAAC,IAAI,CAAC,OAAO,CAAC,UAAU,CAAC;YAAE,OAAO,EAAE,EAAE,EAAE,KAAK,EAAE,KAAK,EAAE,gBAAgB,EAAE,CAAC;QAE7E,MAAM,OAAO,GAAG,oBAAoB,IAAI,CAAC,IAAI,EAAE,CAAC;QAChD,MAAM,QAAQ,GAAG,MAAM,wBAAwB,CAAC,OAAO,EAAE,IAAI,CAAC,UAAU,EAAE,IAAI,CAAC,MAAM,EAAE,uBAAuB,CAAC,CAAC;QAChH,IAAI,QAAQ,CAAC,MAAM,KAAK,UAAU;YAAE,OAAO,EAAE,EAAE,EAAE,IAAI,EAAE,OAAO,EAAE,CAAC;QACjE,IAAI,QAAQ,CAAC,MAAM,KAAK,UAAU,IAAI,CAAC,IAAI,CAAC,MAAM,EAAE,CAAC;YACpD,OAAO,EAAE,EAAE,EAAE,KAAK,EAAE,KAAK,EAAE,0BAA0B,CAAC,QAAQ,CAAC,cAAc,CAAC,EAAE,CAAC;QAClF,CAAC;QACD,IAAI,IAAI,CAAC,MAAM;YAAE,IAAI,CAAC,IAAI,EAAE,CAAC;QAC7B,MAAM,IAAI,GAAG;YACZ,IAAI;YACJ,IAAI,CAAC,SAAS;YACd,SAAS;YACT,IAAI,CAAC,UAAU;YACf,GAAG,CAAC,IAAI,CAAC,UAAU,CAAC,CAAC,CAAC,CAAC,UAAU,EAAE,IAAI,CAAC,UAAU,CAAC,CAAC,CAAC,CAAC,EAAE,CAAC;YACzD,QAAQ;YACR,WAAW;YACX,QAAQ;YACR,MAAM,CAAC,IAAI,CAAC,IAAI,CAAC;YACjB,IAAI;YACJ,MAAM,CAAC,IAAI,CAAC,MAAM,CAAC;YACnB,aAAa;YACb,MAAM,CAAC,IAAI,CAAC,QAAQ,CAAC,cAAc,CAAC;YACpC,KAAK;YACL,MAAM,CAAC,IAAI,CAAC,QAAQ,CAAC,gBAAgB,CAAC;YACtC,KAAK;YACL,IAAI,CAAC,QAAQ,CAAC,cAAc,CAAC,CAAC,CAAC,IAAI,CAAC,CAAC,CAAC,KAAK;YAC3C,MAAM;YACN,IAAI,CAAC,QAAQ,CAAC,WAAW;YACzB,MAAM;YACN,IAAI,CAAC,QAAQ,CAAC,WAAW;YACzB,IAAI;YACJ,MAAM,CAAC,IAAI,CAAC,QAAQ,CAAC,iBAAiB,CAAC;YACvC,KAAK;YACL,MAAM,CAAC,IAAI,CAAC,QAAQ,CAAC,YAAY,CAAC;YAClC,GAAG,CAAC,QAAQ,CAAC,OAAO,KAAK,MAAM,CAAC,CAAC,CAAC,CAAC,MAAM,EAAE,IAAI,CAAC,CAAC,CAAC,CAAC,EAAE,CAAC;SACtD,CAAC;QAEF,MAAM,KAAK,GAAG,IAAI,CAAC,MAAM,CAAC,UAAU,EAAE,IAAI,EAAE;YAC3C,QAAQ,EAAE,OAAO,CAAC,QAAQ,KAAK,OAAO;YACtC,KAAK,EAAE,QAAQ;YACf,GAAG,EAAE,OAAO,CAAC,GAAG;SAChB,CAAC,CAAC;QACH,IAAI,KAAK,CAAC,GAAG;YAAE,qBAAqB,CAAC,KAAK,CAAC,GAAG,CAAC,CAAC;QAChD,KAAK,CAAC,KAAK,EAAE,EAAE,CAAC;QAChB,KAAK,CAAC,EAAE,CAAC,MAAM,EAAE,GAAG,EAAE;YACrB,IAAI,IAAI,CAAC,MAAM,KAAK,KAAK,EAAE,CAAC;gBAC3B,IAAI,CAAC,MAAM,GAAG,SAAS,CAAC;YACzB,CAAC;QACF,CAAC,CAAC,CAAC;QACH,IAAI,CAAC,MAAM,GAAG,KAAK,CAAC;QAEpB,KAAK,IAAI,OAAO,GAAG,CAAC,EAAE,OAAO,GAAG,IAAI,CAAC,mBAAmB,EAAE,OAAO,EAAE,EAAE,CAAC;YACrE,MAAM,KAAK,GAAG,MAAM,wBAAwB,CAAC,OAAO,EAAE,IAAI,CAAC,UAAU,EAAE,IAAI,CAAC,MAAM,EAAE,uBAAuB,CAAC,CAAC;YAC7G,IAAI,KAAK,CAAC,MAAM,KAAK,UAAU;gBAAE,OAAO,EAAE,EAAE,EAAE,IAAI,EAAE,OAAO,EAAE,CAAC;YAC9D,IAAI,KAAK,CAAC,MAAM,KAAK,UAAU,EAAE,CAAC;gBACjC,IAAI,CAAC,IAAI,EAAE,CAAC;gBACZ,OAAO,EAAE,EAAE,EAAE,KAAK,EAAE,KAAK,EAAE,0BAA0B,CAAC,KAAK,CAAC,cAAc,CAAC,EAAE,CAAC;YAC/E,CAAC;YACD,MAAM,IAAI,CAAC,MAAM,CAAC,IAAI,CAAC,qBAAqB,CAAC,CAAC;QAC/C,CAAC;QACD,IAAI,CAAC,IAAI,EAAE,CAAC;QACZ,OAAO,EAAE,EAAE,EAAE,KAAK,EAAE,KAAK,EAAE,gBAAgB,EAAE,CAAC;IAC/C,CAAC;IAED,IAAI;QACH,MAAM,KAAK,GAAG,IAAI,CAAC,MAAM,CAAC;QAC1B,IAAI,CAAC,KAAK;YAAE,OAAO,EAAE,OAAO,EAAE,KAAK,EAAE,CAAC;QACtC,IAAI,CAAC,MAAM,GAAG,SAAS,CAAC;QACxB,IAAI,KAAK,CAAC,GAAG,EAAE,CAAC;YACf,uBAAuB,CAAC,KAAK,CAAC,GAAG,CAAC,CAAC;YACnC,eAAe,CAAC,KAAK,CAAC,GAAG,CAAC,CAAC;QAC5B,CAAC;QACD,OAAO,EAAE,OAAO,EAAE,IAAI,EAAE,CAAC;IAC1B,CAAC;IAED,SAAS;QACR,OAAO,IAAI,CAAC,MAAM,KAAK,SAAS,CAAC;IAClC,CAAC;CACD","sourcesContent":["import type { ChildProcess } from \"node:child_process\";\nimport { type Dirent, mkdirSync, readdirSync, readFileSync, writeFileSync } from \"node:fs\";\nimport { arch as osArch, platform as osPlatform } from \"node:os\";\nimport { dirname, join, relative } from \"node:path\";\nimport type { Readable } from \"node:stream\";\nimport { pipeline } from \"node:stream/promises\";\nimport { spawnProcess, waitForChildProcessWithTermination } from \"../../utils/child-process.ts\";\nimport { killProcessTree, trackDetachedChildPid, untrackDetachedChildPid } from \"../../utils/shell.ts\";\nimport { modelsDir as agentModelsDir, runtimesDir as agentRuntimesDir } from \"../agent-paths.ts\";\nimport {\n\tderiveHostLocalInferenceProfile,\n\ttype LocalInferenceProfile,\n\ttype LocalInferenceProfileMode,\n} from \"./local-inference-profile.ts\";\nimport { probePrismLlamaCppServer } from \"./prism-llamacpp-server-probe.ts\";\nimport {\n\textractZipArchive,\n\tfetchRuntimeDownload,\n\tinstallRuntimeArchive,\n\ttype ManagedRuntimeSpawn,\n\tremovePartialDownload,\n\trequireRuntimeStdin,\n\tresolveRuntimeLifecycleDependencies,\n\truntimeCommandAvailable,\n\ttryFileSizeBytes,\n\twriteRuntimeDownload,\n} from \"./runtime-process.ts\";\n\n/**\n * Managed runtime for Pi's curated Prism 1-bit and ternary GGUF models. Q1_0 is available in\n * upstream llama.cpp, while the selected group-128 Q2_0 artifacts and paired DSpark acceleration\n * still require Prism's validated build. This module therefore owns the pinned runtime install,\n * GGUF downloads, and llama-server lifecycle. Mirrors local-runtime.ts's OllamaRuntime: injectable\n * seams for fetch/spawn/exists, pi-owned directories under agentDir (runtimes/, models/),\n * detached+tracked child processes, onProgress as best-effort UI feedback, honest error taxonomies\n * instead of silent fallbacks.\n *\n * The fork publishes precompiled per-platform archives on GitHub Releases (no local build): pin a\n * release tag, download the matching asset, extract it, and locate the `llama-server` binary once\n * at install time (release archive layouts vary — bin/ vs build/bin/ at the top level — so this is\n * a deterministic one-time scan, not a fallback). The found relative path and backend (cpu/cuda)\n * are persisted to a manifest so later `detect()`/`serve()` calls never need to re-scan.\n */\n\n/**\n * Pinned Prism ML llama.cpp fork release — verified via\n * `GET https://api.github.com/repos/PrismML-Eng/llama.cpp/releases/tags/prism-b9594-38c66ad` on\n * 2026-07-18 (release exists, all assets referenced below are present on it). Bump here (and\n * re-verify with the same API call) when the provider ships new kernels.\n */\nexport const PRISM_LLAMACPP_PINNED_RELEASE = \"prism-b9594-38c66ad\";\n\nexport const PRISM_LLAMACPP_RELEASES_BASE_URL = \"https://github.com/PrismML-Eng/llama.cpp/releases/download\";\n\nexport interface PrismModelDescriptor {\n\trepo: string;\n\tfile: string;\n\tmmprojFile?: string;\n\tdisplayName: string;\n\tarchitecture: \"dense\";\n\truntime: \"prism-llamacpp\";\n\tfamily: \"bonsai\" | \"ternary-bonsai\";\n\tparameterScale: \"1.7B\" | \"4B\" | \"8B\" | \"27B\";\n\tweightFormat: \"q1_0\" | \"q2_0\";\n\tmatchedDrafter?: {\n\t\tkind: \"dspark\";\n\t\tfile: string;\n\t\tdraftMax: 4;\n\t\tminimumContext: 16_384;\n\t\tvalidatedBackend: \"cuda\";\n\t};\n}\n\n/** Curated 27B descriptor shared by the full local execution catalog. */\nexport const BONSAI_27B: PrismModelDescriptor = {\n\trepo: \"prism-ml/Bonsai-27B-gguf\",\n\tfile: \"Bonsai-27B-Q1_0.gguf\",\n\tmmprojFile: \"Bonsai-27B-mmproj-Q8_0.gguf\",\n\tdisplayName: \"Bonsai-27B (1-bit Q1_0 + vision)\",\n\tarchitecture: \"dense\",\n\truntime: \"prism-llamacpp\",\n\tfamily: \"bonsai\",\n\tparameterScale: \"27B\",\n\tweightFormat: \"q1_0\",\n\tmatchedDrafter: {\n\t\tkind: \"dspark\",\n\t\tfile: \"Bonsai-27B-dspark-Q4_1.gguf\",\n\t\tdraftMax: 4,\n\t\tminimumContext: 16_384,\n\t\tvalidatedBackend: \"cuda\",\n\t},\n};\n\nexport type PrismBackend = \"cpu\" | \"cuda\";\nexport type PrismAssetKind = \"tar-gz\" | \"zip\";\n\nexport interface PrismArchiveAsset {\n\tname: string;\n\tkind: PrismAssetKind;\n}\n\nexport interface PrismLlamaAsset extends PrismArchiveAsset {\n\tbackend: PrismBackend;\n\tcompanionAssets?: readonly PrismArchiveAsset[];\n}\n\n// Verbatim asset names from the pinned release (see PRISM_LLAMACPP_PINNED_RELEASE doc comment) —\n// hardcoded rather than templated from the tag because the provider's naming isn't fully\n// consistent across platforms (Windows CPU assets drop the tag/commit segment entirely; the\n// Windows CUDA asset uses a different build-number segment, \"b1\" not \"b9594\").\nconst LINUX_X64_CPU_ASSET = \"llama-prism-b9594-38c66ad-bin-ubuntu-x64.tar.gz\";\nconst LINUX_ARM64_CPU_ASSET = \"llama-prism-b9594-38c66ad-bin-ubuntu-arm64.tar.gz\";\nconst LINUX_X64_CUDA_ASSET = \"llama-prism-b9594-38c66ad-bin-linux-cuda-12.4-x64.tar.gz\";\nconst MACOS_ARM64_ASSET = \"llama-prism-b9594-38c66ad-bin-macos-arm64.tar.gz\";\nconst MACOS_X64_ASSET = \"llama-prism-b9594-38c66ad-bin-macos-x64.tar.gz\";\nconst WIN_X64_CPU_ASSET = \"llama-bin-win-cpu-x64.zip\";\nconst WIN_ARM64_CPU_ASSET = \"llama-bin-win-cpu-arm64.zip\";\nconst WIN_X64_CUDA_ASSET = \"llama-prism-b1-38c66ad-bin-win-cuda-12.4-x64.zip\";\nconst WIN_X64_CUDA_COMPANION_ASSET = \"cudart-llama-bin-win-cuda-12.4-x64.zip\";\n\n/**\n * Maps a platform/arch/GPU triple to the exact Prism llama.cpp release asset for\n * {@link PRISM_LLAMACPP_PINNED_RELEASE} — verified against the real GitHub release, not guessed.\n * CPU asset by default; CUDA 12.4 is selected for x64 Linux or Windows with an NVIDIA GPU. The\n * pinned Windows build is incomplete without its matching `cudart-*` archive, so the resolver\n * returns both as one mandatory installation plan. Pure and exported so it's independently\n * testable.\n */\nexport function resolvePrismLlamaAsset(\n\tplat: string,\n\tarchitecture: string,\n\thasNvidiaGpu: boolean,\n): PrismLlamaAsset | undefined {\n\tif (plat === \"linux\") {\n\t\tif (architecture === \"x64\") {\n\t\t\treturn hasNvidiaGpu\n\t\t\t\t? { name: LINUX_X64_CUDA_ASSET, kind: \"tar-gz\", backend: \"cuda\" }\n\t\t\t\t: { name: LINUX_X64_CPU_ASSET, kind: \"tar-gz\", backend: \"cpu\" };\n\t\t}\n\t\tif (architecture === \"arm64\") return { name: LINUX_ARM64_CPU_ASSET, kind: \"tar-gz\", backend: \"cpu\" };\n\t\treturn undefined;\n\t}\n\tif (plat === \"darwin\") {\n\t\tif (architecture === \"arm64\") return { name: MACOS_ARM64_ASSET, kind: \"tar-gz\", backend: \"cpu\" };\n\t\tif (architecture === \"x64\") return { name: MACOS_X64_ASSET, kind: \"tar-gz\", backend: \"cpu\" };\n\t\treturn undefined;\n\t}\n\tif (plat === \"win32\") {\n\t\tif (architecture === \"x64\") {\n\t\t\treturn hasNvidiaGpu\n\t\t\t\t? {\n\t\t\t\t\t\tname: WIN_X64_CUDA_ASSET,\n\t\t\t\t\t\tkind: \"zip\",\n\t\t\t\t\t\tbackend: \"cuda\",\n\t\t\t\t\t\tcompanionAssets: [{ name: WIN_X64_CUDA_COMPANION_ASSET, kind: \"zip\" }],\n\t\t\t\t\t}\n\t\t\t\t: { name: WIN_X64_CPU_ASSET, kind: \"zip\", backend: \"cpu\" };\n\t\t}\n\t\tif (architecture === \"arm64\") return { name: WIN_ARM64_CPU_ASSET, kind: \"zip\", backend: \"cpu\" };\n\t\treturn undefined;\n\t}\n\treturn undefined;\n}\n\nexport interface PrismDetectResult {\n\truntimeInstalled: boolean;\n\tbinaryPath?: string;\n\trelease?: string;\n}\n\nexport interface PrismDownloadResult {\n\tok: boolean;\n\tpath?: string;\n\tskipped?: boolean;\n\terror?: string;\n}\n\nexport type PrismServeResult = { ok: true; baseUrl: string } | { ok: false; error: string };\n\ninterface PrismInstallManifest {\n\trelease: string;\n\tbinaryRelPath: string;\n\tbackend: PrismBackend;\n}\n\ntype PrismExtractArchiveFn = (\n\tinput: Readable,\n\tdestDir: string,\n\tkind: PrismAssetKind,\n) => Promise<{ ok: boolean; error?: string }>;\n\nexport interface PrismLlamaCppDeps {\n\tfetchFn?: typeof fetch;\n\tspawnFn?: ManagedRuntimeSpawn;\n\texistsFn?: (path: string) => boolean;\n\tsleepFn?: (ms: number) => Promise<void>;\n\t/** Whether a named command exists on PATH (nvidia-smi, tar — for zip extraction on Windows). */\n\thasCommand?: (command: string) => boolean;\n\t/** Decided once at install time and persisted as `backend` — serve() reads the persisted value\n\t * so `-ngl 99` reflects what was actually installed, not the host's current GPU state. */\n\thasNvidiaGpu?: () => boolean;\n\tplatform?: () => string;\n\tarch?: () => string;\n\t/** Host capacity probes used once to derive the bounded local inference profile. */\n\ttotalMemoryBytes?: () => number;\n\tlogicalCpuCount?: () => number;\n\t/** Runs the extraction step for a downloaded archive. Injectable so installManaged's\n\t * download->extract->scan orchestration is testable without a real tar/unzip pipeline; defaults\n\t * to the real spawn-based extractor. */\n\textractArchive?: PrismExtractArchiveFn;\n\t/** Health-poll bounds for serve(); default ~120s (240 * 500ms), both overridable for tests. */\n\thealthPollAttempts?: number;\n\thealthPollIntervalMs?: number;\n}\n\nconst EXTRACTION_TIMEOUT_MS = 10 * 60_000;\nconst COMMAND_KILL_GRACE_MS = 2_000;\nconst DEFAULT_HEALTH_POLL_ATTEMPTS = 240;\nconst DEFAULT_HEALTH_POLL_INTERVAL_MS = 500;\nconst HEALTH_CHECK_TIMEOUT_MS = 2_000;\nconst HEAD_REQUEST_TIMEOUT_MS = 10_000;\n\nfunction parseContentLength(header: string | null): number | undefined {\n\tif (!header) return undefined;\n\tconst value = Number(header);\n\treturn Number.isFinite(value) && value >= 0 ? value : undefined;\n}\n\nfunction modelIdentityConflictError(servedModelIds: readonly string[]): string {\n\treturn `model-identity-conflict:${servedModelIds.join(\",\") || \"unknown\"}`;\n}\n\nexport class PrismLlamaCppRuntime {\n\tprivate readonly _agentDir: string;\n\tprivate readonly _fetch: typeof fetch;\n\tprivate readonly _spawn: ManagedRuntimeSpawn;\n\tprivate readonly _exists: (path: string) => boolean;\n\tprivate readonly _sleep: (ms: number) => Promise<void>;\n\tprivate readonly _hasCommand: (command: string) => boolean;\n\tprivate readonly _hasNvidiaGpu: () => boolean;\n\tprivate readonly _platform: () => string;\n\tprivate readonly _arch: () => string;\n\tprivate readonly _extractArchiveFn: PrismExtractArchiveFn;\n\tprivate readonly _healthPollAttempts: number;\n\tprivate readonly _healthPollIntervalMs: number;\n\tprivate readonly _profile: LocalInferenceProfile;\n\tprivate _child: Pick<ChildProcess, \"pid\" | \"kill\" | \"unref\" | \"on\"> | undefined;\n\n\tconstructor(args: {\n\t\tagentDir: string;\n\t\tprofileMode?: LocalInferenceProfileMode;\n\t\tdeps?: PrismLlamaCppDeps;\n\t}) {\n\t\tthis._agentDir = args.agentDir;\n\t\t[this._fetch, this._spawn, this._exists, this._sleep] = resolveRuntimeLifecycleDependencies(args.deps);\n\t\tthis._hasCommand = args.deps?.hasCommand ?? runtimeCommandAvailable;\n\t\tthis._hasNvidiaGpu = args.deps?.hasNvidiaGpu ?? (() => this._hasCommand(\"nvidia-smi\"));\n\t\tthis._platform = args.deps?.platform ?? osPlatform;\n\t\tthis._arch = args.deps?.arch ?? osArch;\n\t\tthis._profile = deriveHostLocalInferenceProfile(args.profileMode ?? \"balanced\", args.deps);\n\t\tthis._extractArchiveFn =\n\t\t\targs.deps?.extractArchive ?? ((input, destDir, kind) => this._extractArchive(input, destDir, kind));\n\t\tthis._healthPollAttempts = args.deps?.healthPollAttempts ?? DEFAULT_HEALTH_POLL_ATTEMPTS;\n\t\tthis._healthPollIntervalMs = args.deps?.healthPollIntervalMs ?? DEFAULT_HEALTH_POLL_INTERVAL_MS;\n\t}\n\n\truntimeDir(): string {\n\t\treturn agentRuntimesDir(\"prism-llamacpp\", this._agentDir);\n\t}\n\n\tmodelsDir(): string {\n\t\treturn agentModelsDir(\"llamacpp\", this._agentDir);\n\t}\n\n\tprivate _binaryName(): string {\n\t\treturn this._platform() === \"win32\" ? \"llama-server.exe\" : \"llama-server\";\n\t}\n\n\tprivate _manifestPath(): string {\n\t\treturn join(this.runtimeDir(), \"install.json\");\n\t}\n\n\tprivate _readManifest(): PrismInstallManifest | undefined {\n\t\ttry {\n\t\t\tconst parsed = JSON.parse(readFileSync(this._manifestPath(), \"utf8\")) as {\n\t\t\t\trelease?: unknown;\n\t\t\t\tbinaryRelPath?: unknown;\n\t\t\t\tbackend?: unknown;\n\t\t\t};\n\t\t\tif (\n\t\t\t\ttypeof parsed.release === \"string\" &&\n\t\t\t\ttypeof parsed.binaryRelPath === \"string\" &&\n\t\t\t\t(parsed.backend === \"cpu\" || parsed.backend === \"cuda\")\n\t\t\t) {\n\t\t\t\treturn { release: parsed.release, binaryRelPath: parsed.binaryRelPath, backend: parsed.backend };\n\t\t\t}\n\t\t\treturn undefined;\n\t\t} catch {\n\t\t\treturn undefined;\n\t\t}\n\t}\n\n\tprivate _writeManifest(manifest: PrismInstallManifest): void {\n\t\twriteFileSync(this._manifestPath(), JSON.stringify(manifest, null, \"\\t\"));\n\t}\n\n\tasync detect(): Promise<PrismDetectResult> {\n\t\tconst manifest = this._readManifest();\n\t\tif (!manifest) return { runtimeInstalled: false };\n\t\tconst binaryPath = join(this.runtimeDir(), manifest.binaryRelPath);\n\t\tconst runtimeInstalled = this._exists(binaryPath);\n\t\treturn {\n\t\t\truntimeInstalled,\n\t\t\tbinaryPath: runtimeInstalled ? binaryPath : undefined,\n\t\t\trelease: manifest.release,\n\t\t};\n\t}\n\n\t/** One-time recursive scan for the `llama-server` binary inside a freshly extracted release\n\t * archive — layouts vary (bin/ vs build/bin/ at the top level) across platforms/backends, so\n\t * this locates it deterministically instead of assuming a fixed path, then the result is\n\t * persisted so later calls never need to re-scan. */\n\tprivate _findBinaryRelPath(root: string): string | undefined {\n\t\tconst targetName = this._binaryName();\n\t\tconst stack: string[] = [root];\n\t\twhile (stack.length > 0) {\n\t\t\tconst dir = stack.pop();\n\t\t\tif (dir === undefined) continue;\n\t\t\tlet entries: Dirent[];\n\t\t\ttry {\n\t\t\t\tentries = readdirSync(dir, { withFileTypes: true });\n\t\t\t} catch {\n\t\t\t\tcontinue;\n\t\t\t}\n\t\t\tfor (const entry of entries) {\n\t\t\t\tconst full = join(dir, entry.name);\n\t\t\t\tif (entry.isDirectory()) stack.push(full);\n\t\t\t\telse if (entry.isFile() && entry.name === targetName) return relative(root, full);\n\t\t\t}\n\t\t}\n\t\treturn undefined;\n\t}\n\n\t/**\n\t * Download the pinned release asset for this host and extract it (consent-gated by the caller,\n\t * same contract as OllamaRuntime#installManaged — this method only does the mechanical\n\t * download+extract+locate and reports the outcome honestly). No compiler toolchain required:\n\t * the fork ships prebuilt binaries.\n\t */\n\tasync installManaged(onProgress?: (status: string) => void): Promise<{ ok: boolean; error?: string }> {\n\t\tconst asset = resolvePrismLlamaAsset(this._platform(), this._arch(), this._hasNvidiaGpu());\n\t\tif (!asset) return { ok: false, error: \"unsupported-platform\" };\n\n\t\tconst archives: readonly PrismArchiveAsset[] = [asset, ...(asset.companionAssets ?? [])];\n\t\tfor (const archive of archives) {\n\t\t\tconst downloadUrl = `${PRISM_LLAMACPP_RELEASES_BASE_URL}/${PRISM_LLAMACPP_PINNED_RELEASE}/${archive.name}`;\n\t\t\tconst extracted = await this._installArchive(downloadUrl, archive, onProgress);\n\t\t\tif (!extracted.ok) return extracted;\n\t\t}\n\n\t\tonProgress?.(\"Locating llama-server binary…\");\n\t\tconst destDir = this.runtimeDir();\n\t\tconst binaryRelPath = this._findBinaryRelPath(destDir);\n\t\tif (!binaryRelPath) {\n\t\t\treturn { ok: false, error: \"binary-missing: no llama-server binary found in the extracted archive\" };\n\t\t}\n\n\t\tthis._writeManifest({ release: PRISM_LLAMACPP_PINNED_RELEASE, binaryRelPath, backend: asset.backend });\n\t\tonProgress?.(\"Prism llama.cpp runtime installed.\");\n\t\treturn { ok: true };\n\t}\n\n\tprivate _installArchive(\n\t\tdownloadUrl: string,\n\t\tasset: PrismArchiveAsset,\n\t\tonProgress?: (status: string) => void,\n\t): Promise<{ ok: boolean; error?: string }> {\n\t\treturn installRuntimeArchive(\n\t\t\tthis._fetch,\n\t\t\tdownloadUrl,\n\t\t\tthis.runtimeDir(),\n\t\t\tasset,\n\t\t\tthis._extractArchiveFn,\n\t\t\tonProgress,\n\t\t);\n\t}\n\n\tprivate async _extractArchive(\n\t\tinput: Readable,\n\t\tdestDir: string,\n\t\tkind: PrismAssetKind,\n\t): Promise<{ ok: boolean; error?: string }> {\n\t\ttry {\n\t\t\treturn kind === \"zip\" ? await this._extractZip(input, destDir) : await this._extractTarGz(input, destDir);\n\t\t} catch (error) {\n\t\t\treturn { ok: false, error: `extract-fail: ${error instanceof Error ? error.message : String(error)}` };\n\t\t}\n\t}\n\n\tprivate async _extractTarGz(input: Readable, destDir: string): Promise<{ ok: boolean; error?: string }> {\n\t\tconst tarProc = spawnProcess(\"tar\", [\"-xzf\", \"-\", \"-C\", destDir], {\n\t\t\tdetached: process.platform !== \"win32\",\n\t\t\tstdio: [\"pipe\", \"ignore\", \"ignore\"],\n\t\t});\n\t\tconst terminationController = new AbortController();\n\t\tconst processWait = waitForChildProcessWithTermination(tarProc, {\n\t\t\tsignal: terminationController.signal,\n\t\t\ttimeoutMs: EXTRACTION_TIMEOUT_MS,\n\t\t\tkillGraceMs: COMMAND_KILL_GRACE_MS,\n\t\t});\n\t\ttry {\n\t\t\tawait pipeline(input, requireRuntimeStdin(tarProc, \"tar\"));\n\t\t} catch (error) {\n\t\t\tterminationController.abort();\n\t\t\tawait processWait.catch(() => {});\n\t\t\tthrow error;\n\t\t}\n\t\tconst terminal = await processWait;\n\t\tif (terminal.reason === \"timeout\" || terminal.code !== 0) {\n\t\t\treturn {\n\t\t\t\tok: false,\n\t\t\t\terror:\n\t\t\t\t\tterminal.reason === \"timeout\"\n\t\t\t\t\t\t? `extract-fail: tar extraction timed out after ${EXTRACTION_TIMEOUT_MS}ms`\n\t\t\t\t\t\t: `extract-fail: tar exited with code ${terminal.code ?? \"unknown\"}`,\n\t\t\t};\n\t\t}\n\t\treturn { ok: true };\n\t}\n\n\tprivate async _extractZip(input: Readable, destDir: string): Promise<{ ok: boolean; error?: string }> {\n\t\treturn extractZipArchive({\n\t\t\tinput,\n\t\t\tdestDir,\n\t\t\ttempPrefix: \"prism-llamacpp-download\",\n\t\t\tplatform: this._platform,\n\t\t\thasCommand: this._hasCommand,\n\t\t\ttimeoutMs: EXTRACTION_TIMEOUT_MS,\n\t\t\tkillGraceMs: COMMAND_KILL_GRACE_MS,\n\t\t});\n\t}\n\n\tprivate async _remoteContentLength(url: string): Promise<number | undefined> {\n\t\ttry {\n\t\t\tconst response = await this._fetch(url, {\n\t\t\t\tmethod: \"HEAD\",\n\t\t\t\tsignal: AbortSignal.timeout(HEAD_REQUEST_TIMEOUT_MS),\n\t\t\t});\n\t\t\treturn response.ok ? parseContentLength(response.headers.get(\"content-length\")) : undefined;\n\t\t} catch {\n\t\t\treturn undefined;\n\t\t}\n\t}\n\n\t/**\n\t * Stream a GGUF (or mmproj) file from Hugging Face into pi's owned models dir. Skips a re-download\n\t * when the local file already matches the remote size; verifies size when the response reports\n\t * `content-length` and deletes the partial file on any failure or mismatch — never leaves a\n\t * corrupt/truncated weight file behind for a later load to silently misread.\n\t */\n\tasync downloadModel(\n\t\targs: { repo: string; file: string },\n\t\tonProgress?: (status: string) => void,\n\t): Promise<PrismDownloadResult> {\n\t\tconst destPath = join(this.modelsDir(), args.repo, args.file);\n\t\tconst url = `https://huggingface.co/${args.repo}/resolve/main/${args.file}`;\n\t\tmkdirSync(dirname(destPath), { recursive: true });\n\n\t\tif (this._exists(destPath)) {\n\t\t\tconst remoteSize = await this._remoteContentLength(url);\n\t\t\tif (remoteSize !== undefined && remoteSize === tryFileSizeBytes(destPath)) {\n\t\t\t\tonProgress?.(`${args.file} already downloaded (${remoteSize} bytes) — skipping.`);\n\t\t\t\treturn { ok: true, path: destPath, skipped: true };\n\t\t\t}\n\t\t}\n\n\t\tonProgress?.(`Downloading ${args.file} from ${args.repo}…`);\n\t\tconst download = await fetchRuntimeDownload(this._fetch, url);\n\t\tif (!download.ok) return download;\n\n\t\tconst expectedBytes = parseContentLength(download.response.headers.get(\"content-length\"));\n\t\tconst written = await writeRuntimeDownload(download.body as unknown as Readable, destPath);\n\t\tif (!written.ok) return written;\n\n\t\tconst actualBytes = tryFileSizeBytes(destPath);\n\t\tif (expectedBytes !== undefined && actualBytes !== expectedBytes) {\n\t\t\tremovePartialDownload(destPath);\n\t\t\treturn { ok: false, error: `size-mismatch: expected ${expectedBytes} bytes, got ${actualBytes ?? 0}` };\n\t\t}\n\n\t\tonProgress?.(`${args.file} downloaded (${actualBytes ?? 0} bytes).`);\n\t\treturn { ok: true, path: destPath };\n\t}\n\n\t/**\n\t * Spawn `llama-server` detached+tracked (killed on parent shutdown even if pi crashes without\n\t * calling stop()) and poll `/health` until ready. `-ngl 99` is only added when the installed\n\t * asset's persisted `backend` is `\"cuda\"`, not from the host's current GPU state.\n\t */\n\tasync serve(args: {\n\t\tmodelPath: string;\n\t\tmodelAlias: string;\n\t\tmmprojPath?: string;\n\t\tport: number;\n\t\tnumCtx: number;\n\t}): Promise<PrismServeResult> {\n\t\tconst manifest = this._readManifest();\n\t\tif (!manifest) return { ok: false, error: \"binary-missing\" };\n\t\tconst binaryPath = join(this.runtimeDir(), manifest.binaryRelPath);\n\t\tif (!this._exists(binaryPath)) return { ok: false, error: \"binary-missing\" };\n\n\t\tconst baseUrl = `http://127.0.0.1:${args.port}`;\n\t\tconst existing = await probePrismLlamaCppServer(baseUrl, args.modelAlias, this._fetch, HEALTH_CHECK_TIMEOUT_MS);\n\t\tif (existing.status === \"matching\") return { ok: true, baseUrl };\n\t\tif (existing.status === \"conflict\" && !this._child) {\n\t\t\treturn { ok: false, error: modelIdentityConflictError(existing.servedModelIds) };\n\t\t}\n\t\tif (this._child) this.stop();\n\t\tconst argv = [\n\t\t\t\"-m\",\n\t\t\targs.modelPath,\n\t\t\t\"--alias\",\n\t\t\targs.modelAlias,\n\t\t\t...(args.mmprojPath ? [\"--mmproj\", args.mmprojPath] : []),\n\t\t\t\"--host\",\n\t\t\t\"127.0.0.1\",\n\t\t\t\"--port\",\n\t\t\tString(args.port),\n\t\t\t\"-c\",\n\t\t\tString(args.numCtx),\n\t\t\t\"--cache-ram\",\n\t\t\tString(this._profile.promptCacheMiB),\n\t\t\t\"-np\",\n\t\t\tString(this._profile.parallelRequests),\n\t\t\t\"-fa\",\n\t\t\tthis._profile.flashAttention ? \"on\" : \"off\",\n\t\t\t\"-ctk\",\n\t\t\tthis._profile.kvCacheType,\n\t\t\t\"-ctv\",\n\t\t\tthis._profile.kvCacheType,\n\t\t\t\"-t\",\n\t\t\tString(this._profile.generationThreads),\n\t\t\t\"-tb\",\n\t\t\tString(this._profile.batchThreads),\n\t\t\t...(manifest.backend === \"cuda\" ? [\"-ngl\", \"99\"] : []),\n\t\t];\n\n\t\tconst child = this._spawn(binaryPath, argv, {\n\t\t\tdetached: process.platform !== \"win32\",\n\t\t\tstdio: \"ignore\",\n\t\t\tenv: process.env,\n\t\t});\n\t\tif (child.pid) trackDetachedChildPid(child.pid);\n\t\tchild.unref?.();\n\t\tchild.on(\"exit\", () => {\n\t\t\tif (this._child === child) {\n\t\t\t\tthis._child = undefined;\n\t\t\t}\n\t\t});\n\t\tthis._child = child;\n\n\t\tfor (let attempt = 0; attempt < this._healthPollAttempts; attempt++) {\n\t\t\tconst probe = await probePrismLlamaCppServer(baseUrl, args.modelAlias, this._fetch, HEALTH_CHECK_TIMEOUT_MS);\n\t\t\tif (probe.status === \"matching\") return { ok: true, baseUrl };\n\t\t\tif (probe.status === \"conflict\") {\n\t\t\t\tthis.stop();\n\t\t\t\treturn { ok: false, error: modelIdentityConflictError(probe.servedModelIds) };\n\t\t\t}\n\t\t\tawait this._sleep(this._healthPollIntervalMs);\n\t\t}\n\t\tthis.stop();\n\t\treturn { ok: false, error: \"health-timeout\" };\n\t}\n\n\tstop(): { stopped: boolean } {\n\t\tconst child = this._child;\n\t\tif (!child) return { stopped: false };\n\t\tthis._child = undefined;\n\t\tif (child.pid) {\n\t\t\tuntrackDetachedChildPid(child.pid);\n\t\t\tkillProcessTree(child.pid);\n\t\t}\n\t\treturn { stopped: true };\n\t}\n\n\tisRunning(): boolean {\n\t\treturn this._child !== undefined;\n\t}\n}\n"]}

package/dist/core/models/local-execution-planner.d.ts

@@ -0,0 +1,24 @@
+import { type PrismBackend, type PrismModelDescriptor } from "./llamacpp-runtime.ts";
+/**
+ * The complete model set Pi may manage through its pinned Prism llama.cpp adapter. Unknown models
+ * never enter this path. Every descriptor pins the exact GGUF artifact so user-provided quant
+ * suffixes cannot silently select a different format or runtime.
+ */
+export declare const PRISM_LLAMACPP_DESCRIPTORS: Readonly<Record<string, PrismModelDescriptor>>;
+export declare function resolvePrismLocalModelDescriptor(modelId: string): PrismModelDescriptor | undefined;
+export type PrismSpeculationPlan = {
+    enabled: true;
+    kind: "dspark";
+    drafterFile: string;
+    draftMax: 4;
+    minimumContext: 16_384;
+} | {
+    enabled: false;
+    reason: "not-requested" | "model-has-no-matched-drafter" | "backend-not-validated";
+};
+/** MTP is deliberately absent: Bonsai acceleration uses only the model's exact paired drafter. */
+export declare function planPrismSpeculation(descriptor: PrismModelDescriptor, args: {
+    requested: boolean;
+    backend: PrismBackend;
+}): PrismSpeculationPlan;
+//# sourceMappingURL=local-execution-planner.d.ts.map

package/dist/core/models/local-execution-planner.d.ts.map

@@ -0,0 +1 @@
+{"version":3,"file":"local-execution-planner.d.ts","sourceRoot":"","sources":["../../../src/core/models/local-execution-planner.ts"],"names":[],"mappings":"AAAA,OAAO,EAAc,KAAK,YAAY,EAAE,KAAK,oBAAoB,EAAE,MAAM,uBAAuB,CAAC;AA2FjG;;;;GAIG;AACH,eAAO,MAAM,0BAA0B,EAAE,QAAQ,CAAC,MAAM,CAAC,MAAM,EAAE,oBAAoB,CAAC,CASpF,CAAC;AAMH,wBAAgB,gCAAgC,CAAC,OAAO,EAAE,MAAM,GAAG,oBAAoB,GAAG,SAAS,CAElG;AAED,MAAM,MAAM,oBAAoB,GAC7B;IACA,OAAO,EAAE,IAAI,CAAC;IACd,IAAI,EAAE,QAAQ,CAAC;IACf,WAAW,EAAE,MAAM,CAAC;IACpB,QAAQ,EAAE,CAAC,CAAC;IACZ,cAAc,EAAE,MAAM,CAAC;CACtB,GACD;IAAE,OAAO,EAAE,KAAK,CAAC;IAAC,MAAM,EAAE,eAAe,GAAG,8BAA8B,GAAG,uBAAuB,CAAA;CAAE,CAAC;AAE1G,kGAAkG;AAClG,wBAAgB,oBAAoB,CACnC,UAAU,EAAE,oBAAoB,EAChC,IAAI,EAAE;IAAE,SAAS,EAAE,OAAO,CAAC;IAAC,OAAO,EAAE,YAAY,CAAA;CAAE,GACjD,oBAAoB,CAatB"}