@caupulican/pi-adaptative 0.81.19 → 0.81.21

package/dist/core/models/perf-profile.js.map

@@ -1 +1 @@
-{"version":3,"file":"perf-profile.js","sourceRoot":"","sources":["../../../src/core/models/perf-profile.ts"],"names":[],"mappings":"AACA,OAAO,EAIN,iCAAiC,GAGjC,MAAM,mBAAmB,CAAC;AAE3B,MAAM,eAAe,GAAG,GAAG,CAAC;AAC5B,MAAM,wBAAwB,GAAG,CAAC,CAAC;AACnC,MAAM,CAAC,MAAM,uCAAuC,GAAG,EAAE,GAAG,EAAE,GAAG,IAAI,CAAC;AAgCtE,MAAM,UAAU,kBAAkB,CAAC,KAAc,EAA6B;IAC7E,IAAI,OAAO,KAAK,KAAK,QAAQ,IAAI,KAAK,KAAK,IAAI,IAAI,KAAK,CAAC,OAAO,CAAC,KAAK,CAAC;QAAE,OAAO,KAAK,CAAC;IACtF,MAAM,MAAM,GAAG,KAAgC,CAAC;IAChD,MAAM,OAAO,GAAG,MAAM,CAAC,OAAO,CAAC;IAC/B,OAAO,CACN,CAAC,MAAM,CAAC,sBAAsB,KAAK,SAAS,IAAI,sBAAsB,CAAC,MAAM,CAAC,sBAAsB,CAAC,CAAC;QACtG,CAAC,MAAM,CAAC,qBAAqB,KAAK,SAAS,IAAI,sBAAsB,CAAC,MAAM,CAAC,qBAAqB,CAAC,CAAC;QACpG,OAAO,OAAO,KAAK,QAAQ;QAC3B,MAAM,CAAC,SAAS,CAAC,OAAO,CAAC;QACzB,OAAO,IAAI,CAAC;QACZ,OAAO,MAAM,CAAC,SAAS,KAAK,QAAQ,CACpC,CAAC;AAAA,CACF;AAED,MAAM,UAAU,wBAAwB,CAAC,MAAuB,EAAW;IAC1E,OAAO,qBAAqB,CAAC,MAAM,CAAC,KAAK,SAAS,IAAI,oBAAoB,CAAC,MAAM,CAAC,KAAK,SAAS,CAAC;AAAA,CACjG;AAED,MAAM,UAAU,sBAAsB,CACrC,OAAqC,EACrC,MAAuB,EACvB,EAAE,GAAW,MAAM,CAAC,EAAE,IAAI,IAAI,IAAI,EAAE,CAAC,WAAW,EAAE,EACnB;IAC/B,MAAM,WAAW,GAAG,qBAAqB,CAAC,MAAM,CAAC,CAAC;IAClD,MAAM,UAAU,GAAG,oBAAoB,CAAC,MAAM,CAAC,CAAC;IAChD,IAAI,WAAW,KAAK,SAAS,IAAI,UAAU,KAAK,SAAS;QAAE,OAAO,OAAO,CAAC;IAE1E,MAAM,eAAe,GAAG,OAAO,EAAE,OAAO,IAAI,CAAC,CAAC;IAC9C,OAAO;QACN,GAAG,CAAC,WAAW,KAAK,SAAS,IAAI;YAChC,sBAAsB,EAAE,UAAU,CAAC,OAAO,EAAE,sBAAsB,EAAE,WAAW,EAAE,eAAe,CAAC;SACjG,CAAC;QACF,GAAG,CAAC,WAAW,KAAK,SAAS,IAAI,OAAO,EAAE,sBAAsB,KAAK,SAAS;YAC7E,CAAC,CAAC,EAAE,sBAAsB,EAAE,OAAO,CAAC,sBAAsB,EAAE;YAC5D,CAAC,CAAC,EAAE,CAAC;QACN,GAAG,CAAC,UAAU,KAAK,SAAS,IAAI;YAC/B,qBAAqB,EAAE,UAAU,CAAC,OAAO,EAAE,qBAAqB,EAAE,UAAU,EAAE,eAAe,CAAC;SAC9F,CAAC;QACF,GAAG,CAAC,UAAU,KAAK,SAAS,IAAI,OAAO,EAAE,qBAAqB,KAAK,SAAS;YAC3E,CAAC,CAAC,EAAE,qBAAqB,EAAE,OAAO,CAAC,qBAAqB,EAAE;YAC1D,CAAC,CAAC,EAAE,CAAC;QACN,OAAO,EAAE,eAAe,GAAG,CAAC;QAC5B,SAAS,EAAE,EAAE;KACb,CAAC;AAAA,CACF;AAED,MAAM,UAAU,gCAAgC,CAAC,KAA8B,EAA8B;IAC5G,MAAM,OAAO,GAAG,KAAK,CAAC,OAAO,CAAC;IAC9B,MAAM,WAAW,GAAG,OAAO,EAAE,sBAAsB,CAAC;IACpD,IAAI,CAAC,OAAO,IAAI,CAAC,WAAW,IAAI,OAAO,CAAC,OAAO,GAAG,CAAC,IAAI,KAAK,CAAC,YAAY,IAAI,CAAC;QAAE,OAAO,EAAE,CAAC;IAE1F,MAAM,iBAAiB,GAAG,CAAC,KAAK,CAAC,YAAY,GAAG,WAAW,CAAC,GAAG,IAAI,CAAC;IACpE,IAAI,CAAC,MAAM,CAAC,QAAQ,CAAC,iBAAiB,CAAC,IAAI,iBAAiB,IAAI,CAAC;QAAE,OAAO,EAAE,CAAC;IAE7E,MAAM,QAAQ,GAAG,IAAI,CAAC,IAAI,CAAC,IAAI,CAAC,GAAG,CAAC,KAAK,CAAC,IAAI,CAAC,WAAW,EAAE,iBAAiB,GAAG,CAAC,CAAC,CAAC,CAAC;IACpF,MAAM,SAAS,GAAG,IAAI,CAAC,GAAG,CAAC,KAAK,CAAC,IAAI,CAAC,WAAW,EAAE,KAAK,CAAC,SAAS,IAAI,uCAAuC,CAAC,CAAC;IAC/G,MAAM,WAAW,GAAG,IAAI,CAAC,GAAG,CAAC,QAAQ,EAAE,SAAS,CAAC,CAAC;IAClD,OAAO,WAAW,GAAG,KAAK,CAAC,IAAI,CAAC,WAAW,CAAC,CAAC,CAAC,EAAE,WAAW,EAAE,CAAC,CAAC,CAAC,EAAE,CAAC;AAAA,CACnE;AAED,MAAM,UAAU,2BAA2B,CAAC,OAAgB,EAAU;IACrE,MAAM,UAAU,GAAG,IAAI,CAAC,SAAS,CAAC;QACjC,YAAY,EAAE,OAAO,CAAC,YAAY;QAClC,QAAQ,EAAE,OAAO,CAAC,QAAQ;QAC1B,KAAK,EAAE,OAAO,CAAC,KAAK;KACpB,CAAC,CAAC;IACH,IAAI,CAAC,UAAU;QAAE,OAAO,CAAC,CAAC;IAC1B,OAAO,IAAI,CAAC,IAAI,CAAC,UAAU,CAAC,MAAM,GAAG,wBAAwB,CAAC,CAAC;AAAA,CAC/D;AAED,MAAM,UAAU,oBAAoB,CAAC,QAAkB,EAAE,QAAwC,EAAY;IAC5G,OAAO,KAAK,EAAE,KAAK,EAAE,OAAO,EAAE,aAAa,EAAE,EAAE,CAAC;QAC/C,MAAM,KAAK,GAAG,QAAQ,CAAC,KAAK,IAAI,IAAI,CAAC,GAAG,CAAC;QACzC,MAAM,MAAM,GAAG,QAAQ,CAAC,MAAM,IAAI,CAAC,GAAG,EAAE,CAAC,IAAI,IAAI,EAAE,CAAC,WAAW,EAAE,CAAC,CAAC;QACnE,MAAM,kBAAkB,GAAG,KAAK,EAAE,CAAC;QACnC,IAAI,mBAAuC,CAAC;QAC5C,IAAI,cAAkC,CAAC;QACvC,MAAM,kBAAkB,GAAG,aAAa,EAAE,UAAU,CAAC;QACrD,MAAM,KAAK,GAAG,MAAM,QAAQ,CAAC,KAAK,EAAE,OAAO,EAAE;YAC5C,GAAG,aAAa;YAChB,UAAU,EAAE,KAAK,EAAE,QAA0B,EAAE,aAAyB,EAAE,EAAE,CAAC;gBAC5E,mBAAmB,KAAK,KAAK,EAAE,CAAC;gBAChC,MAAM,kBAAkB,EAAE,CAAC,QAAQ,EAAE,aAAa,CAAC,CAAC;YAAA,CACpD;SACD,CAAC,CAAC;QACH,MAAM,KAAK,GAAG,iCAAiC,EAAE,CAAC;QAElD,KAAK,CAAC,KAAK,IAAI,EAAE,CAAC;YACjB,IAAI,cAAc,GAAG,KAAK,CAAC;YAC3B,IAAI,KAAK,EAAE,MAAM,KAAK,IAAI,KAAK,EAAE,CAAC;gBACjC,IAAI,cAAc,KAAK,SAAS,IAAI,iBAAiB,CAAC,KAAK,CAAC,EAAE,CAAC;oBAC9D,cAAc,GAAG,KAAK,EAAE,CAAC;gBAC1B,CAAC;gBACD,IAAI,KAAK,CAAC,IAAI,KAAK,MAAM,EAAE,CAAC;oBAC3B,cAAc,GAAG,IAAI,CAAC;oBACtB,4BAA4B,CAAC;wBAC5B,gBAAgB,EAAE,KAAK,CAAC,OAAO,CAAC,KAAK,CAAC,MAAM;wBAC5C,QAAQ,EAAE,KAAK,EAAE;wBACjB,cAAc;wBACd,KAAK;wBACL,MAAM;wBACN,YAAY,EAAE,KAAK,CAAC,OAAO,CAAC,KAAK,CAAC,KAAK;wBACvC,QAAQ;wBACR,kBAAkB;wBAClB,mBAAmB;qBACnB,CAAC,CAAC;gBACJ,CAAC;qBAAM,IAAI,KAAK,CAAC,IAAI,KAAK,OAAO,EAAE,CAAC;oBACnC,cAAc,GAAG,IAAI,CAAC;gBACvB,CAAC;gBACD,KAAK,CAAC,IAAI,CAAC,KAAK,CAAC,CAAC;gBAClB,IAAI,cAAc;oBAAE,OAAO;YAC5B,CAAC;YACD,KAAK,CAAC,GAAG,EAAE,CAAC;QAAA,CACZ,CAAC,EAAE,CAAC;QAEL,OAAO,KAAK,CAAC;IAAA,CACb,CAAC;AAAA,CACF;AAED,SAAS,4BAA4B,CAAC,KAUrC,EAAQ;IACR,IAAI,KAAK,CAAC,cAAc,KAAK,SAAS;QAAE,OAAO;IAC/C,MAAM,QAAQ,GAAG,KAAK,CAAC,QAAQ,CAAC,QAAQ,CAAC,KAAK,CAAC,KAAK,CAAC,CAAC;IACtD,IAAI,CAAC,QAAQ;QAAE,OAAO;IACtB,MAAM,MAAM,GAAoB;QAC/B,YAAY,EAAE,KAAK,CAAC,YAAY;QAChC,gBAAgB,EAAE,KAAK,CAAC,gBAAgB;QACxC,qBAAqB,EAAE,KAAK,CAAC,cAAc,GAAG,KAAK,CAAC,kBAAkB;QACtE,kBAAkB,EAAE,KAAK,CAAC,QAAQ,GAAG,KAAK,CAAC,cAAc;QACzD,EAAE,EAAE,KAAK,CAAC,MAAM,EAAE;KAClB,CAAC;IACF,IAAI,KAAK,CAAC,mBAAmB,KAAK,SAAS,EAAE,CAAC;QAC7C,MAAM,CAAC,qBAAqB,GAAG,KAAK,CAAC,cAAc,GAAG,KAAK,CAAC,mBAAmB,CAAC;IACjF,CAAC;IACD,IAAI,CAAC,wBAAwB,CAAC,MAAM,CAAC;QAAE,OAAO;IAC9C,IAAI,CAAC;QACJ,KAAK,CAAC,QAAQ,CAAC,YAAY,CAAC,QAAQ,EAAE,MAAM,CAAC,CAAC;IAC/C,CAAC;IAAC,MAAM,CAAC;QACR,gDAAgD;IACjD,CAAC;AAAA,CACD;AAED,SAAS,iBAAiB,CAAC,KAA4B,EAAW;IACjE,OAAO,CACN,CAAC,KAAK,CAAC,IAAI,KAAK,YAAY,IAAI,KAAK,CAAC,IAAI,KAAK,gBAAgB,IAAI,KAAK,CAAC,IAAI,KAAK,gBAAgB,CAAC;QACnG,KAAK,CAAC,KAAK,CAAC,MAAM,GAAG,CAAC,CACtB,CAAC;AAAA,CACF;AAED,SAAS,qBAAqB,CAAC,MAAuB,EAAsB;IAC3E,MAAM,UAAU,GAAG,MAAM,CAAC,qBAAqB,IAAI,MAAM,CAAC,qBAAqB,CAAC;IAChF,OAAO,eAAe,CAAC,MAAM,CAAC,YAAY,EAAE,UAAU,CAAC,CAAC;AAAA,CACxD;AAED,SAAS,oBAAoB,CAAC,MAAuB,EAAsB;IAC1E,OAAO,eAAe,CAAC,MAAM,CAAC,gBAAgB,EAAE,MAAM,CAAC,kBAAkB,CAAC,CAAC;AAAA,CAC3E;AAED,SAAS,eAAe,CAAC,MAA0B,EAAE,UAA8B,EAAsB;IACxG,IAAI,CAAC,sBAAsB,CAAC,MAAM,CAAC,IAAI,CAAC,sBAAsB,CAAC,UAAU,CAAC;QAAE,OAAO,SAAS,CAAC;IAC7F,OAAO,MAAM,GAAG,CAAC,UAAU,GAAG,IAAI,CAAC,CAAC;AAAA,CACpC;AAED,SAAS,UAAU,CAAC,QAA4B,EAAE,IAAY,EAAE,eAAuB,EAAU;IAChG,IAAI,QAAQ,KAAK,SAAS,IAAI,eAAe,KAAK,CAAC;QAAE,OAAO,IAAI,CAAC;IACjE,OAAO,QAAQ,GAAG,CAAC,CAAC,GAAG,eAAe,CAAC,GAAG,IAAI,GAAG,eAAe,CAAC;AAAA,CACjE;AAED,SAAS,sBAAsB,CAAC,KAAc,EAAmB;IAChE,OAAO,OAAO,KAAK,KAAK,QAAQ,IAAI,MAAM,CAAC,QAAQ,CAAC,KAAK,CAAC,IAAI,KAAK,GAAG,CAAC,CAAC;AAAA,CACxE","sourcesContent":["import type { StreamFn, StreamIdleOptions } from \"@caupulican/pi-agent-core\";\nimport {\n\ttype Api,\n\ttype AssistantMessageEvent,\n\ttype Context,\n\tcreateAssistantMessageEventStream,\n\ttype Model,\n\ttype ProviderResponse,\n} from \"@caupulican/pi-ai\";\n\nconst PERF_EWMA_ALPHA = 0.3;\nconst CHARS_PER_TOKEN_ESTIMATE = 4;\nexport const DEFAULT_ADAPTIVE_STREAM_IDLE_CEILING_MS = 30 * 60 * 1000;\n\nexport interface ModelPerfProfile {\n\tprefillTokensPerSecond?: number;\n\tdecodeTokensPerSecond?: number;\n\tsamples: number;\n\tupdatedAt: string;\n}\n\nexport interface ModelPerfSample {\n\tpromptTokens?: number;\n\tcompletionTokens?: number;\n\theadersToFirstTokenMs?: number;\n\trequestToFirstTokenMs?: number;\n\tfirstTokenToDoneMs?: number;\n\tat?: string;\n}\n\nexport interface AdaptiveStreamIdleInput {\n\tbase: StreamIdleOptions;\n\tprofile?: ModelPerfProfile;\n\tpromptTokens: number;\n\tceilingMs?: number;\n}\n\nexport interface ModelPerfProfileStreamRecorder {\n\tmodelKey: (model: Model<Api>) => string | undefined;\n\trecordSample: (modelKey: string, sample: ModelPerfSample) => void;\n\tnowMs?: () => number;\n\tnowIso?: () => string;\n}\n\nexport function isModelPerfProfile(value: unknown): value is ModelPerfProfile {\n\tif (typeof value !== \"object\" || value === null || Array.isArray(value)) return false;\n\tconst record = value as Record<string, unknown>;\n\tconst samples = record.samples;\n\treturn (\n\t\t(record.prefillTokensPerSecond === undefined || isPositiveFiniteNumber(record.prefillTokensPerSecond)) &&\n\t\t(record.decodeTokensPerSecond === undefined || isPositiveFiniteNumber(record.decodeTokensPerSecond)) &&\n\t\ttypeof samples === \"number\" &&\n\t\tNumber.isInteger(samples) &&\n\t\tsamples >= 0 &&\n\t\ttypeof record.updatedAt === \"string\"\n\t);\n}\n\nexport function hasUsableModelPerfSample(sample: ModelPerfSample): boolean {\n\treturn prefillRateFromSample(sample) !== undefined || decodeRateFromSample(sample) !== undefined;\n}\n\nexport function updateModelPerfProfile(\n\tcurrent: ModelPerfProfile | undefined,\n\tsample: ModelPerfSample,\n\tat: string = sample.at ?? new Date().toISOString(),\n): ModelPerfProfile | undefined {\n\tconst prefillRate = prefillRateFromSample(sample);\n\tconst decodeRate = decodeRateFromSample(sample);\n\tif (prefillRate === undefined && decodeRate === undefined) return current;\n\n\tconst previousSamples = current?.samples ?? 0;\n\treturn {\n\t\t...(prefillRate !== undefined && {\n\t\t\tprefillTokensPerSecond: updateEwma(current?.prefillTokensPerSecond, prefillRate, previousSamples),\n\t\t}),\n\t\t...(prefillRate === undefined && current?.prefillTokensPerSecond !== undefined\n\t\t\t? { prefillTokensPerSecond: current.prefillTokensPerSecond }\n\t\t\t: {}),\n\t\t...(decodeRate !== undefined && {\n\t\t\tdecodeTokensPerSecond: updateEwma(current?.decodeTokensPerSecond, decodeRate, previousSamples),\n\t\t}),\n\t\t...(decodeRate === undefined && current?.decodeTokensPerSecond !== undefined\n\t\t\t? { decodeTokensPerSecond: current.decodeTokensPerSecond }\n\t\t\t: {}),\n\t\tsamples: previousSamples + 1,\n\t\tupdatedAt: at,\n\t};\n}\n\nexport function resolveAdaptiveStreamIdleOptions(input: AdaptiveStreamIdleInput): Partial<StreamIdleOptions> {\n\tconst profile = input.profile;\n\tconst prefillRate = profile?.prefillTokensPerSecond;\n\tif (!profile || !prefillRate || profile.samples < 1 || input.promptTokens <= 0) return {};\n\n\tconst expectedPrefillMs = (input.promptTokens / prefillRate) * 1000;\n\tif (!Number.isFinite(expectedPrefillMs) || expectedPrefillMs <= 0) return {};\n\n\tconst targetMs = Math.ceil(Math.max(input.base.quietIdleMs, expectedPrefillMs * 3));\n\tconst ceilingMs = Math.max(input.base.quietIdleMs, input.ceilingMs ?? DEFAULT_ADAPTIVE_STREAM_IDLE_CEILING_MS);\n\tconst quietIdleMs = Math.min(targetMs, ceilingMs);\n\treturn quietIdleMs > input.base.quietIdleMs ? { quietIdleMs } : {};\n}\n\nexport function estimateContextPromptTokens(context: Context): number {\n\tconst serialized = JSON.stringify({\n\t\tsystemPrompt: context.systemPrompt,\n\t\tmessages: context.messages,\n\t\ttools: context.tools,\n\t});\n\tif (!serialized) return 0;\n\treturn Math.ceil(serialized.length / CHARS_PER_TOKEN_ESTIMATE);\n}\n\nexport function withModelPerfProfile(streamFn: StreamFn, recorder: ModelPerfProfileStreamRecorder): StreamFn {\n\treturn async (model, context, streamOptions) => {\n\t\tconst nowMs = recorder.nowMs ?? Date.now;\n\t\tconst nowIso = recorder.nowIso ?? (() => new Date().toISOString());\n\t\tconst requestStartedAtMs = nowMs();\n\t\tlet responseHeadersAtMs: number | undefined;\n\t\tlet firstTokenAtMs: number | undefined;\n\t\tconst originalOnResponse = streamOptions?.onResponse;\n\t\tconst inner = await streamFn(model, context, {\n\t\t\t...streamOptions,\n\t\t\tonResponse: async (response: ProviderResponse, responseModel: Model<Api>) => {\n\t\t\t\tresponseHeadersAtMs ??= nowMs();\n\t\t\t\tawait originalOnResponse?.(response, responseModel);\n\t\t\t},\n\t\t});\n\t\tconst outer = createAssistantMessageEventStream();\n\n\t\tvoid (async () => {\n\t\t\tlet terminalPushed = false;\n\t\t\tfor await (const event of inner) {\n\t\t\t\tif (firstTokenAtMs === undefined && isFirstTokenEvent(event)) {\n\t\t\t\t\tfirstTokenAtMs = nowMs();\n\t\t\t\t}\n\t\t\t\tif (event.type === \"done\") {\n\t\t\t\t\tterminalPushed = true;\n\t\t\t\t\trecordSuccessfulStreamSample({\n\t\t\t\t\t\tcompletionTokens: event.message.usage.output,\n\t\t\t\t\t\tdoneAtMs: nowMs(),\n\t\t\t\t\t\tfirstTokenAtMs,\n\t\t\t\t\t\tmodel,\n\t\t\t\t\t\tnowIso,\n\t\t\t\t\t\tpromptTokens: event.message.usage.input,\n\t\t\t\t\t\trecorder,\n\t\t\t\t\t\trequestStartedAtMs,\n\t\t\t\t\t\tresponseHeadersAtMs,\n\t\t\t\t\t});\n\t\t\t\t} else if (event.type === \"error\") {\n\t\t\t\t\tterminalPushed = true;\n\t\t\t\t}\n\t\t\t\touter.push(event);\n\t\t\t\tif (terminalPushed) return;\n\t\t\t}\n\t\t\touter.end();\n\t\t})();\n\n\t\treturn outer;\n\t};\n}\n\nfunction recordSuccessfulStreamSample(input: {\n\tmodel: Model<Api>;\n\trecorder: ModelPerfProfileStreamRecorder;\n\trequestStartedAtMs: number;\n\tresponseHeadersAtMs: number | undefined;\n\tfirstTokenAtMs: number | undefined;\n\tdoneAtMs: number;\n\tpromptTokens: number;\n\tcompletionTokens: number;\n\tnowIso: () => string;\n}): void {\n\tif (input.firstTokenAtMs === undefined) return;\n\tconst modelKey = input.recorder.modelKey(input.model);\n\tif (!modelKey) return;\n\tconst sample: ModelPerfSample = {\n\t\tpromptTokens: input.promptTokens,\n\t\tcompletionTokens: input.completionTokens,\n\t\trequestToFirstTokenMs: input.firstTokenAtMs - input.requestStartedAtMs,\n\t\tfirstTokenToDoneMs: input.doneAtMs - input.firstTokenAtMs,\n\t\tat: input.nowIso(),\n\t};\n\tif (input.responseHeadersAtMs !== undefined) {\n\t\tsample.headersToFirstTokenMs = input.firstTokenAtMs - input.responseHeadersAtMs;\n\t}\n\tif (!hasUsableModelPerfSample(sample)) return;\n\ttry {\n\t\tinput.recorder.recordSample(modelKey, sample);\n\t} catch {\n\t\t// Perf profiling must never fail the user turn.\n\t}\n}\n\nfunction isFirstTokenEvent(event: AssistantMessageEvent): boolean {\n\treturn (\n\t\t(event.type === \"text_delta\" || event.type === \"thinking_delta\" || event.type === \"toolcall_delta\") &&\n\t\tevent.delta.length > 0\n\t);\n}\n\nfunction prefillRateFromSample(sample: ModelPerfSample): number | undefined {\n\tconst durationMs = sample.headersToFirstTokenMs ?? sample.requestToFirstTokenMs;\n\treturn tokensPerSecond(sample.promptTokens, durationMs);\n}\n\nfunction decodeRateFromSample(sample: ModelPerfSample): number | undefined {\n\treturn tokensPerSecond(sample.completionTokens, sample.firstTokenToDoneMs);\n}\n\nfunction tokensPerSecond(tokens: number | undefined, durationMs: number | undefined): number | undefined {\n\tif (!isPositiveFiniteNumber(tokens) || !isPositiveFiniteNumber(durationMs)) return undefined;\n\treturn tokens / (durationMs / 1000);\n}\n\nfunction updateEwma(previous: number | undefined, next: number, previousSamples: number): number {\n\tif (previous === undefined || previousSamples === 0) return next;\n\treturn previous * (1 - PERF_EWMA_ALPHA) + next * PERF_EWMA_ALPHA;\n}\n\nfunction isPositiveFiniteNumber(value: unknown): value is number {\n\treturn typeof value === \"number\" && Number.isFinite(value) && value > 0;\n}\n"]}
+{"version":3,"file":"perf-profile.js","sourceRoot":"","sources":["../../../src/core/models/perf-profile.ts"],"names":[],"mappings":"AACA,OAAO,EAIN,iCAAiC,GAGjC,MAAM,mBAAmB,CAAC;AAE3B,MAAM,eAAe,GAAG,GAAG,CAAC;AAC5B,MAAM,wBAAwB,GAAG,CAAC,CAAC;AACnC,MAAM,CAAC,MAAM,uCAAuC,GAAG,EAAE,GAAG,EAAE,GAAG,IAAI,CAAC;AAmCtE,MAAM,UAAU,kBAAkB,CAAC,KAAc,EAA6B;IAC7E,IAAI,OAAO,KAAK,KAAK,QAAQ,IAAI,KAAK,KAAK,IAAI,IAAI,KAAK,CAAC,OAAO,CAAC,KAAK,CAAC;QAAE,OAAO,KAAK,CAAC;IACtF,MAAM,MAAM,GAAG,KAAgC,CAAC;IAChD,MAAM,OAAO,GAAG,MAAM,CAAC,OAAO,CAAC;IAC/B,OAAO,CACN,CAAC,MAAM,CAAC,sBAAsB,KAAK,SAAS,IAAI,sBAAsB,CAAC,MAAM,CAAC,sBAAsB,CAAC,CAAC;QACtG,CAAC,MAAM,CAAC,qBAAqB,KAAK,SAAS,IAAI,sBAAsB,CAAC,MAAM,CAAC,qBAAqB,CAAC,CAAC;QACpG,CAAC,MAAM,CAAC,MAAM,KAAK,SAAS,IAAI,sBAAsB,CAAC,MAAM,CAAC,MAAM,CAAC,CAAC;QACtE,OAAO,OAAO,KAAK,QAAQ;QAC3B,MAAM,CAAC,SAAS,CAAC,OAAO,CAAC;QACzB,OAAO,IAAI,CAAC;QACZ,OAAO,MAAM,CAAC,SAAS,KAAK,QAAQ,CACpC,CAAC;AAAA,CACF;AAED,MAAM,UAAU,wBAAwB,CAAC,MAAuB,EAAW;IAC1E,OAAO,CACN,qBAAqB,CAAC,MAAM,CAAC,KAAK,SAAS;QAC3C,oBAAoB,CAAC,MAAM,CAAC,KAAK,SAAS;QAC1C,sBAAsB,CAAC,MAAM,CAAC,MAAM,CAAC,CACrC,CAAC;AAAA,CACF;AAED,MAAM,UAAU,sBAAsB,CACrC,OAAqC,EACrC,MAAuB,EACvB,EAAE,GAAW,MAAM,CAAC,EAAE,IAAI,IAAI,IAAI,EAAE,CAAC,WAAW,EAAE,EACnB;IAC/B,MAAM,WAAW,GAAG,qBAAqB,CAAC,MAAM,CAAC,CAAC;IAClD,MAAM,UAAU,GAAG,oBAAoB,CAAC,MAAM,CAAC,CAAC;IAChD,MAAM,MAAM,GAAG,sBAAsB,CAAC,MAAM,CAAC,MAAM,CAAC,CAAC,CAAC,CAAC,MAAM,CAAC,MAAM,CAAC,CAAC,CAAC,SAAS,CAAC;IACjF,IAAI,WAAW,KAAK,SAAS,IAAI,UAAU,KAAK,SAAS,IAAI,MAAM,KAAK,SAAS;QAAE,OAAO,OAAO,CAAC;IAElG,MAAM,eAAe,GAAG,OAAO,EAAE,OAAO,IAAI,CAAC,CAAC;IAC9C,OAAO;QACN,GAAG,CAAC,WAAW,KAAK,SAAS,IAAI;YAChC,sBAAsB,EAAE,UAAU,CAAC,OAAO,EAAE,sBAAsB,EAAE,WAAW,EAAE,eAAe,CAAC;SACjG,CAAC;QACF,GAAG,CAAC,WAAW,KAAK,SAAS,IAAI,OAAO,EAAE,sBAAsB,KAAK,SAAS;YAC7E,CAAC,CAAC,EAAE,sBAAsB,EAAE,OAAO,CAAC,sBAAsB,EAAE;YAC5D,CAAC,CAAC,EAAE,CAAC;QACN,GAAG,CAAC,UAAU,KAAK,SAAS,IAAI;YAC/B,qBAAqB,EAAE,UAAU,CAAC,OAAO,EAAE,qBAAqB,EAAE,UAAU,EAAE,eAAe,CAAC;SAC9F,CAAC;QACF,GAAG,CAAC,UAAU,KAAK,SAAS,IAAI,OAAO,EAAE,qBAAqB,KAAK,SAAS;YAC3E,CAAC,CAAC,EAAE,qBAAqB,EAAE,OAAO,CAAC,qBAAqB,EAAE;YAC1D,CAAC,CAAC,EAAE,CAAC;QACN,GAAG,CAAC,MAAM,KAAK,SAAS,IAAI,EAAE,MAAM,EAAE,UAAU,CAAC,OAAO,EAAE,MAAM,EAAE,MAAM,EAAE,eAAe,CAAC,EAAE,CAAC;QAC7F,GAAG,CAAC,MAAM,KAAK,SAAS,IAAI,OAAO,EAAE,MAAM,KAAK,SAAS,CAAC,CAAC,CAAC,EAAE,MAAM,EAAE,OAAO,CAAC,MAAM,EAAE,CAAC,CAAC,CAAC,EAAE,CAAC;QAC5F,OAAO,EAAE,eAAe,GAAG,CAAC;QAC5B,SAAS,EAAE,EAAE;KACb,CAAC;AAAA,CACF;AAED,MAAM,UAAU,gCAAgC,CAAC,KAA8B,EAA8B;IAC5G,MAAM,OAAO,GAAG,KAAK,CAAC,OAAO,CAAC;IAC9B,MAAM,MAAM,GAA+B,EAAE,CAAC;IAC9C,MAAM,qBAAqB,GAAG,KAAK,CAAC,UAAU;QAC7C,CAAC,CAAC,IAAI,CAAC,GAAG,CAAC,KAAK,CAAC,IAAI,CAAC,SAAS,EAAE,KAAK,CAAC,IAAI,CAAC,WAAW,CAAC;QACxD,CAAC,CAAC,KAAK,CAAC,IAAI,CAAC,SAAS,CAAC;IACxB,MAAM,cAAc,GAAG,IAAI,CAAC,GAAG,CAAC,KAAK,CAAC,IAAI,CAAC,WAAW,EAAE,KAAK,CAAC,SAAS,IAAI,uCAAuC,CAAC,CAAC;IACpH,MAAM,gBAAgB,GAAG,IAAI,CAAC,GAAG,CAAC,qBAAqB,EAAE,KAAK,CAAC,SAAS,IAAI,uCAAuC,CAAC,CAAC;IAErH,MAAM,iBAAiB,GAAG,0BAA0B,CAAC,OAAO,EAAE,KAAK,CAAC,YAAY,CAAC,CAAC;IAClF,IAAI,iBAAiB,KAAK,SAAS,EAAE,CAAC;QACrC,MAAM,WAAW,GAAG,aAAa,CAAC,KAAK,CAAC,IAAI,CAAC,WAAW,EAAE,cAAc,EAAE,iBAAiB,CAAC,CAAC;QAC7F,IAAI,WAAW,GAAG,KAAK,CAAC,IAAI,CAAC,WAAW;YAAE,MAAM,CAAC,WAAW,GAAG,WAAW,CAAC;IAC5E,CAAC;IAED,IAAI,KAAK,CAAC,UAAU,EAAE,CAAC;QACtB,MAAM,qBAAqB,GAC1B,iBAAiB,KAAK,SAAS,CAAC,CAAC,CAAC,iBAAiB,GAAG,CAAC,OAAO,EAAE,MAAM,IAAI,CAAC,CAAC,CAAC,CAAC,CAAC,OAAO,EAAE,MAAM,CAAC;QAChG,MAAM,SAAS,GACd,qBAAqB,KAAK,SAAS;YAClC,CAAC,CAAC,aAAa,CAAC,qBAAqB,EAAE,gBAAgB,EAAE,qBAAqB,CAAC;YAC/E,CAAC,CAAC,qBAAqB,CAAC;QAC1B,IAAI,SAAS,GAAG,KAAK,CAAC,IAAI,CAAC,SAAS;YAAE,MAAM,CAAC,SAAS,GAAG,SAAS,CAAC;IACpE,CAAC;IAED,OAAO,MAAM,CAAC;AAAA,CACd;AAED,MAAM,UAAU,2BAA2B,CAAC,OAAgB,EAAU;IACrE,MAAM,UAAU,GAAG,IAAI,CAAC,SAAS,CAAC;QACjC,YAAY,EAAE,OAAO,CAAC,YAAY;QAClC,QAAQ,EAAE,OAAO,CAAC,QAAQ;QAC1B,KAAK,EAAE,OAAO,CAAC,KAAK;KACpB,CAAC,CAAC;IACH,IAAI,CAAC,UAAU;QAAE,OAAO,CAAC,CAAC;IAC1B,OAAO,IAAI,CAAC,IAAI,CAAC,UAAU,CAAC,MAAM,GAAG,wBAAwB,CAAC,CAAC;AAAA,CAC/D;AAED,MAAM,UAAU,oBAAoB,CAAC,QAAkB,EAAE,QAAwC,EAAY;IAC5G,OAAO,KAAK,EAAE,KAAK,EAAE,OAAO,EAAE,aAAa,EAAE,EAAE,CAAC;QAC/C,MAAM,KAAK,GAAG,QAAQ,CAAC,KAAK,IAAI,IAAI,CAAC,GAAG,CAAC;QACzC,MAAM,MAAM,GAAG,QAAQ,CAAC,MAAM,IAAI,CAAC,GAAG,EAAE,CAAC,IAAI,IAAI,EAAE,CAAC,WAAW,EAAE,CAAC,CAAC;QACnE,MAAM,kBAAkB,GAAG,KAAK,EAAE,CAAC;QACnC,IAAI,mBAAuC,CAAC;QAC5C,IAAI,cAAkC,CAAC;QACvC,MAAM,kBAAkB,GAAG,aAAa,EAAE,UAAU,CAAC;QACrD,MAAM,KAAK,GAAG,MAAM,QAAQ,CAAC,KAAK,EAAE,OAAO,EAAE;YAC5C,GAAG,aAAa;YAChB,UAAU,EAAE,KAAK,EAAE,QAA0B,EAAE,aAAyB,EAAE,EAAE,CAAC;gBAC5E,mBAAmB,KAAK,KAAK,EAAE,CAAC;gBAChC,MAAM,kBAAkB,EAAE,CAAC,QAAQ,EAAE,aAAa,CAAC,CAAC;YAAA,CACpD;SACD,CAAC,CAAC;QACH,MAAM,KAAK,GAAG,iCAAiC,EAAE,CAAC;QAElD,KAAK,CAAC,KAAK,IAAI,EAAE,CAAC;YACjB,IAAI,cAAc,GAAG,KAAK,CAAC;YAC3B,IAAI,KAAK,EAAE,MAAM,KAAK,IAAI,KAAK,EAAE,CAAC;gBACjC,IAAI,cAAc,KAAK,SAAS,IAAI,iBAAiB,CAAC,KAAK,CAAC,EAAE,CAAC;oBAC9D,cAAc,GAAG,KAAK,EAAE,CAAC;gBAC1B,CAAC;gBACD,IAAI,KAAK,CAAC,IAAI,KAAK,MAAM,EAAE,CAAC;oBAC3B,cAAc,GAAG,IAAI,CAAC;oBACtB,4BAA4B,CAAC;wBAC5B,gBAAgB,EAAE,KAAK,CAAC,OAAO,CAAC,KAAK,CAAC,MAAM;wBAC5C,QAAQ,EAAE,KAAK,EAAE;wBACjB,cAAc;wBACd,KAAK;wBACL,MAAM;wBACN,YAAY,EAAE,KAAK,CAAC,OAAO,CAAC,KAAK,CAAC,KAAK;wBACvC,QAAQ;wBACR,kBAAkB;wBAClB,mBAAmB;qBACnB,CAAC,CAAC;gBACJ,CAAC;qBAAM,IAAI,KAAK,CAAC,IAAI,KAAK,OAAO,EAAE,CAAC;oBACnC,cAAc,GAAG,IAAI,CAAC;gBACvB,CAAC;gBACD,KAAK,CAAC,IAAI,CAAC,KAAK,CAAC,CAAC;gBAClB,IAAI,cAAc;oBAAE,OAAO;YAC5B,CAAC;YACD,KAAK,CAAC,GAAG,EAAE,CAAC;QAAA,CACZ,CAAC,EAAE,CAAC;QAEL,OAAO,KAAK,CAAC;IAAA,CACb,CAAC;AAAA,CACF;AAED,SAAS,4BAA4B,CAAC,KAUrC,EAAQ;IACR,IAAI,KAAK,CAAC,cAAc,KAAK,SAAS;QAAE,OAAO;IAC/C,MAAM,QAAQ,GAAG,KAAK,CAAC,QAAQ,CAAC,QAAQ,CAAC,KAAK,CAAC,KAAK,CAAC,CAAC;IACtD,IAAI,CAAC,QAAQ;QAAE,OAAO;IACtB,MAAM,MAAM,GAAoB;QAC/B,YAAY,EAAE,KAAK,CAAC,YAAY;QAChC,gBAAgB,EAAE,KAAK,CAAC,gBAAgB;QACxC,qBAAqB,EAAE,KAAK,CAAC,cAAc,GAAG,KAAK,CAAC,kBAAkB;QACtE,kBAAkB,EAAE,KAAK,CAAC,QAAQ,GAAG,KAAK,CAAC,cAAc;QACzD,EAAE,EAAE,KAAK,CAAC,MAAM,EAAE;KAClB,CAAC;IACF,IAAI,KAAK,CAAC,mBAAmB,KAAK,SAAS,EAAE,CAAC;QAC7C,MAAM,CAAC,qBAAqB,GAAG,KAAK,CAAC,cAAc,GAAG,KAAK,CAAC,mBAAmB,CAAC;IACjF,CAAC;IACD,IAAI,CAAC,wBAAwB,CAAC,MAAM,CAAC;QAAE,OAAO;IAC9C,IAAI,CAAC;QACJ,KAAK,CAAC,QAAQ,CAAC,YAAY,CAAC,QAAQ,EAAE,MAAM,CAAC,CAAC;IAC/C,CAAC;IAAC,MAAM,CAAC;QACR,gDAAgD;IACjD,CAAC;AAAA,CACD;AAED,SAAS,iBAAiB,CAAC,KAA4B,EAAW;IACjE,OAAO,CACN,CAAC,KAAK,CAAC,IAAI,KAAK,YAAY,IAAI,KAAK,CAAC,IAAI,KAAK,gBAAgB,IAAI,KAAK,CAAC,IAAI,KAAK,gBAAgB,CAAC;QACnG,KAAK,CAAC,KAAK,CAAC,MAAM,GAAG,CAAC,CACtB,CAAC;AAAA,CACF;AAED,SAAS,0BAA0B,CAAC,OAAqC,EAAE,YAAoB,EAAsB;IACpH,MAAM,WAAW,GAAG,OAAO,EAAE,sBAAsB,CAAC;IACpD,IAAI,CAAC,OAAO,IAAI,OAAO,CAAC,OAAO,GAAG,CAAC,IAAI,CAAC,WAAW,IAAI,YAAY,IAAI,CAAC;QAAE,OAAO,SAAS,CAAC;IAC3F,MAAM,iBAAiB,GAAG,CAAC,YAAY,GAAG,WAAW,CAAC,GAAG,IAAI,CAAC;IAC9D,OAAO,MAAM,CAAC,QAAQ,CAAC,iBAAiB,CAAC,IAAI,iBAAiB,GAAG,CAAC,CAAC,CAAC,CAAC,iBAAiB,CAAC,CAAC,CAAC,SAAS,CAAC;AAAA,CACnG;AAED,SAAS,aAAa,CAAC,SAAiB,EAAE,SAAiB,EAAE,UAAkB,EAAU;IACxF,OAAO,IAAI,CAAC,GAAG,CAAC,IAAI,CAAC,IAAI,CAAC,IAAI,CAAC,GAAG,CAAC,SAAS,EAAE,UAAU,GAAG,CAAC,CAAC,CAAC,EAAE,SAAS,CAAC,CAAC;AAAA,CAC3E;AAED,SAAS,qBAAqB,CAAC,MAAuB,EAAsB;IAC3E,MAAM,UAAU,GAAG,sBAAsB,CAAC,MAAM,CAAC,qBAAqB,CAAC;QACtE,CAAC,CAAC,MAAM,CAAC,qBAAqB;QAC9B,CAAC,CAAC,MAAM,CAAC,qBAAqB,CAAC;IAChC,OAAO,eAAe,CAAC,MAAM,CAAC,YAAY,EAAE,UAAU,CAAC,CAAC;AAAA,CACxD;AAED,SAAS,oBAAoB,CAAC,MAAuB,EAAsB;IAC1E,OAAO,eAAe,CAAC,MAAM,CAAC,gBAAgB,EAAE,MAAM,CAAC,kBAAkB,CAAC,CAAC;AAAA,CAC3E;AAED,SAAS,eAAe,CAAC,MAA0B,EAAE,UAA8B,EAAsB;IACxG,IAAI,CAAC,sBAAsB,CAAC,MAAM,CAAC,IAAI,CAAC,sBAAsB,CAAC,UAAU,CAAC;QAAE,OAAO,SAAS,CAAC;IAC7F,OAAO,MAAM,GAAG,CAAC,UAAU,GAAG,IAAI,CAAC,CAAC;AAAA,CACpC;AAED,SAAS,UAAU,CAAC,QAA4B,EAAE,IAAY,EAAE,eAAuB,EAAU;IAChG,IAAI,QAAQ,KAAK,SAAS,IAAI,eAAe,KAAK,CAAC;QAAE,OAAO,IAAI,CAAC;IACjE,OAAO,QAAQ,GAAG,CAAC,CAAC,GAAG,eAAe,CAAC,GAAG,IAAI,GAAG,eAAe,CAAC;AAAA,CACjE;AAED,SAAS,sBAAsB,CAAC,KAAc,EAAmB;IAChE,OAAO,OAAO,KAAK,KAAK,QAAQ,IAAI,MAAM,CAAC,QAAQ,CAAC,KAAK,CAAC,IAAI,KAAK,GAAG,CAAC,CAAC;AAAA,CACxE","sourcesContent":["import type { StreamFn, StreamIdleOptions } from \"@caupulican/pi-agent-core\";\nimport {\n\ttype Api,\n\ttype AssistantMessageEvent,\n\ttype Context,\n\tcreateAssistantMessageEventStream,\n\ttype Model,\n\ttype ProviderResponse,\n} from \"@caupulican/pi-ai\";\n\nconst PERF_EWMA_ALPHA = 0.3;\nconst CHARS_PER_TOKEN_ESTIMATE = 4;\nexport const DEFAULT_ADAPTIVE_STREAM_IDLE_CEILING_MS = 30 * 60 * 1000;\n\nexport interface ModelPerfProfile {\n\tprefillTokensPerSecond?: number;\n\tdecodeTokensPerSecond?: number;\n\tloadMs?: number;\n\tsamples: number;\n\tupdatedAt: string;\n}\n\nexport interface ModelPerfSample {\n\tpromptTokens?: number;\n\tcompletionTokens?: number;\n\theadersToFirstTokenMs?: number;\n\trequestToFirstTokenMs?: number;\n\tfirstTokenToDoneMs?: number;\n\tloadMs?: number;\n\tat?: string;\n}\n\nexport interface AdaptiveStreamIdleInput {\n\tbase: StreamIdleOptions;\n\tprofile?: ModelPerfProfile;\n\tpromptTokens: number;\n\tceilingMs?: number;\n\tlocalClass?: boolean;\n}\n\nexport interface ModelPerfProfileStreamRecorder {\n\tmodelKey: (model: Model<Api>) => string | undefined;\n\trecordSample: (modelKey: string, sample: ModelPerfSample) => void;\n\tnowMs?: () => number;\n\tnowIso?: () => string;\n}\n\nexport function isModelPerfProfile(value: unknown): value is ModelPerfProfile {\n\tif (typeof value !== \"object\" || value === null || Array.isArray(value)) return false;\n\tconst record = value as Record<string, unknown>;\n\tconst samples = record.samples;\n\treturn (\n\t\t(record.prefillTokensPerSecond === undefined || isPositiveFiniteNumber(record.prefillTokensPerSecond)) &&\n\t\t(record.decodeTokensPerSecond === undefined || isPositiveFiniteNumber(record.decodeTokensPerSecond)) &&\n\t\t(record.loadMs === undefined || isPositiveFiniteNumber(record.loadMs)) &&\n\t\ttypeof samples === \"number\" &&\n\t\tNumber.isInteger(samples) &&\n\t\tsamples >= 0 &&\n\t\ttypeof record.updatedAt === \"string\"\n\t);\n}\n\nexport function hasUsableModelPerfSample(sample: ModelPerfSample): boolean {\n\treturn (\n\t\tprefillRateFromSample(sample) !== undefined ||\n\t\tdecodeRateFromSample(sample) !== undefined ||\n\t\tisPositiveFiniteNumber(sample.loadMs)\n\t);\n}\n\nexport function updateModelPerfProfile(\n\tcurrent: ModelPerfProfile | undefined,\n\tsample: ModelPerfSample,\n\tat: string = sample.at ?? new Date().toISOString(),\n): ModelPerfProfile | undefined {\n\tconst prefillRate = prefillRateFromSample(sample);\n\tconst decodeRate = decodeRateFromSample(sample);\n\tconst loadMs = isPositiveFiniteNumber(sample.loadMs) ? sample.loadMs : undefined;\n\tif (prefillRate === undefined && decodeRate === undefined && loadMs === undefined) return current;\n\n\tconst previousSamples = current?.samples ?? 0;\n\treturn {\n\t\t...(prefillRate !== undefined && {\n\t\t\tprefillTokensPerSecond: updateEwma(current?.prefillTokensPerSecond, prefillRate, previousSamples),\n\t\t}),\n\t\t...(prefillRate === undefined && current?.prefillTokensPerSecond !== undefined\n\t\t\t? { prefillTokensPerSecond: current.prefillTokensPerSecond }\n\t\t\t: {}),\n\t\t...(decodeRate !== undefined && {\n\t\t\tdecodeTokensPerSecond: updateEwma(current?.decodeTokensPerSecond, decodeRate, previousSamples),\n\t\t}),\n\t\t...(decodeRate === undefined && current?.decodeTokensPerSecond !== undefined\n\t\t\t? { decodeTokensPerSecond: current.decodeTokensPerSecond }\n\t\t\t: {}),\n\t\t...(loadMs !== undefined && { loadMs: updateEwma(current?.loadMs, loadMs, previousSamples) }),\n\t\t...(loadMs === undefined && current?.loadMs !== undefined ? { loadMs: current.loadMs } : {}),\n\t\tsamples: previousSamples + 1,\n\t\tupdatedAt: at,\n\t};\n}\n\nexport function resolveAdaptiveStreamIdleOptions(input: AdaptiveStreamIdleInput): Partial<StreamIdleOptions> {\n\tconst profile = input.profile;\n\tconst result: Partial<StreamIdleOptions> = {};\n\tconst localConnectDefaultMs = input.localClass\n\t\t? Math.max(input.base.connectMs, input.base.quietIdleMs)\n\t\t: input.base.connectMs;\n\tconst quietCeilingMs = Math.max(input.base.quietIdleMs, input.ceilingMs ?? DEFAULT_ADAPTIVE_STREAM_IDLE_CEILING_MS);\n\tconst connectCeilingMs = Math.max(localConnectDefaultMs, input.ceilingMs ?? DEFAULT_ADAPTIVE_STREAM_IDLE_CEILING_MS);\n\n\tconst expectedPrefillMs = expectedPrefillFromProfile(profile, input.promptTokens);\n\tif (expectedPrefillMs !== undefined) {\n\t\tconst quietIdleMs = adaptiveBound(input.base.quietIdleMs, quietCeilingMs, expectedPrefillMs);\n\t\tif (quietIdleMs > input.base.quietIdleMs) result.quietIdleMs = quietIdleMs;\n\t}\n\n\tif (input.localClass) {\n\t\tconst expectedConnectWorkMs =\n\t\t\texpectedPrefillMs !== undefined ? expectedPrefillMs + (profile?.loadMs ?? 0) : profile?.loadMs;\n\t\tconst connectMs =\n\t\t\texpectedConnectWorkMs !== undefined\n\t\t\t\t? adaptiveBound(localConnectDefaultMs, connectCeilingMs, expectedConnectWorkMs)\n\t\t\t\t: localConnectDefaultMs;\n\t\tif (connectMs > input.base.connectMs) result.connectMs = connectMs;\n\t}\n\n\treturn result;\n}\n\nexport function estimateContextPromptTokens(context: Context): number {\n\tconst serialized = JSON.stringify({\n\t\tsystemPrompt: context.systemPrompt,\n\t\tmessages: context.messages,\n\t\ttools: context.tools,\n\t});\n\tif (!serialized) return 0;\n\treturn Math.ceil(serialized.length / CHARS_PER_TOKEN_ESTIMATE);\n}\n\nexport function withModelPerfProfile(streamFn: StreamFn, recorder: ModelPerfProfileStreamRecorder): StreamFn {\n\treturn async (model, context, streamOptions) => {\n\t\tconst nowMs = recorder.nowMs ?? Date.now;\n\t\tconst nowIso = recorder.nowIso ?? (() => new Date().toISOString());\n\t\tconst requestStartedAtMs = nowMs();\n\t\tlet responseHeadersAtMs: number | undefined;\n\t\tlet firstTokenAtMs: number | undefined;\n\t\tconst originalOnResponse = streamOptions?.onResponse;\n\t\tconst inner = await streamFn(model, context, {\n\t\t\t...streamOptions,\n\t\t\tonResponse: async (response: ProviderResponse, responseModel: Model<Api>) => {\n\t\t\t\tresponseHeadersAtMs ??= nowMs();\n\t\t\t\tawait originalOnResponse?.(response, responseModel);\n\t\t\t},\n\t\t});\n\t\tconst outer = createAssistantMessageEventStream();\n\n\t\tvoid (async () => {\n\t\t\tlet terminalPushed = false;\n\t\t\tfor await (const event of inner) {\n\t\t\t\tif (firstTokenAtMs === undefined && isFirstTokenEvent(event)) {\n\t\t\t\t\tfirstTokenAtMs = nowMs();\n\t\t\t\t}\n\t\t\t\tif (event.type === \"done\") {\n\t\t\t\t\tterminalPushed = true;\n\t\t\t\t\trecordSuccessfulStreamSample({\n\t\t\t\t\t\tcompletionTokens: event.message.usage.output,\n\t\t\t\t\t\tdoneAtMs: nowMs(),\n\t\t\t\t\t\tfirstTokenAtMs,\n\t\t\t\t\t\tmodel,\n\t\t\t\t\t\tnowIso,\n\t\t\t\t\t\tpromptTokens: event.message.usage.input,\n\t\t\t\t\t\trecorder,\n\t\t\t\t\t\trequestStartedAtMs,\n\t\t\t\t\t\tresponseHeadersAtMs,\n\t\t\t\t\t});\n\t\t\t\t} else if (event.type === \"error\") {\n\t\t\t\t\tterminalPushed = true;\n\t\t\t\t}\n\t\t\t\touter.push(event);\n\t\t\t\tif (terminalPushed) return;\n\t\t\t}\n\t\t\touter.end();\n\t\t})();\n\n\t\treturn outer;\n\t};\n}\n\nfunction recordSuccessfulStreamSample(input: {\n\tmodel: Model<Api>;\n\trecorder: ModelPerfProfileStreamRecorder;\n\trequestStartedAtMs: number;\n\tresponseHeadersAtMs: number | undefined;\n\tfirstTokenAtMs: number | undefined;\n\tdoneAtMs: number;\n\tpromptTokens: number;\n\tcompletionTokens: number;\n\tnowIso: () => string;\n}): void {\n\tif (input.firstTokenAtMs === undefined) return;\n\tconst modelKey = input.recorder.modelKey(input.model);\n\tif (!modelKey) return;\n\tconst sample: ModelPerfSample = {\n\t\tpromptTokens: input.promptTokens,\n\t\tcompletionTokens: input.completionTokens,\n\t\trequestToFirstTokenMs: input.firstTokenAtMs - input.requestStartedAtMs,\n\t\tfirstTokenToDoneMs: input.doneAtMs - input.firstTokenAtMs,\n\t\tat: input.nowIso(),\n\t};\n\tif (input.responseHeadersAtMs !== undefined) {\n\t\tsample.headersToFirstTokenMs = input.firstTokenAtMs - input.responseHeadersAtMs;\n\t}\n\tif (!hasUsableModelPerfSample(sample)) return;\n\ttry {\n\t\tinput.recorder.recordSample(modelKey, sample);\n\t} catch {\n\t\t// Perf profiling must never fail the user turn.\n\t}\n}\n\nfunction isFirstTokenEvent(event: AssistantMessageEvent): boolean {\n\treturn (\n\t\t(event.type === \"text_delta\" || event.type === \"thinking_delta\" || event.type === \"toolcall_delta\") &&\n\t\tevent.delta.length > 0\n\t);\n}\n\nfunction expectedPrefillFromProfile(profile: ModelPerfProfile | undefined, promptTokens: number): number | undefined {\n\tconst prefillRate = profile?.prefillTokensPerSecond;\n\tif (!profile || profile.samples < 1 || !prefillRate || promptTokens <= 0) return undefined;\n\tconst expectedPrefillMs = (promptTokens / prefillRate) * 1000;\n\treturn Number.isFinite(expectedPrefillMs) && expectedPrefillMs > 0 ? expectedPrefillMs : undefined;\n}\n\nfunction adaptiveBound(defaultMs: number, ceilingMs: number, expectedMs: number): number {\n\treturn Math.min(Math.ceil(Math.max(defaultMs, expectedMs * 3)), ceilingMs);\n}\n\nfunction prefillRateFromSample(sample: ModelPerfSample): number | undefined {\n\tconst durationMs = isPositiveFiniteNumber(sample.headersToFirstTokenMs)\n\t\t? sample.headersToFirstTokenMs\n\t\t: sample.requestToFirstTokenMs;\n\treturn tokensPerSecond(sample.promptTokens, durationMs);\n}\n\nfunction decodeRateFromSample(sample: ModelPerfSample): number | undefined {\n\treturn tokensPerSecond(sample.completionTokens, sample.firstTokenToDoneMs);\n}\n\nfunction tokensPerSecond(tokens: number | undefined, durationMs: number | undefined): number | undefined {\n\tif (!isPositiveFiniteNumber(tokens) || !isPositiveFiniteNumber(durationMs)) return undefined;\n\treturn tokens / (durationMs / 1000);\n}\n\nfunction updateEwma(previous: number | undefined, next: number, previousSamples: number): number {\n\tif (previous === undefined || previousSamples === 0) return next;\n\treturn previous * (1 - PERF_EWMA_ALPHA) + next * PERF_EWMA_ALPHA;\n}\n\nfunction isPositiveFiniteNumber(value: unknown): value is number {\n\treturn typeof value === \"number\" && Number.isFinite(value) && value > 0;\n}\n"]}

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

@@ -1,3 +1,4 @@
+import type { OllamaRuntime, TransformersRuntime } from "./local-runtime.ts";
 export type ResidencyControl = "full" | "keep-alive" | "advisory";
 export type RuntimeRole = "active" | "router" | "bench" | "probe";
 export interface RuntimeResidentModel {
@@ -49,6 +50,50 @@ export interface RuntimeResidencyAdapter {
     ensureResident(model: string): Promise<void>;
     release(model: string): Promise<void>;
 }
+export declare class RuntimeResidencyArbiter {
+    private readonly budgetBytes;
+    private readonly adapters;
+    constructor(args: {
+        budgetBytes: number;
+        adapters: RuntimeResidencyAdapter[];
+    });
+    ensureResident(adapterId: string, request: RuntimeLoadRequest): Promise<RuntimeResidencyPlan>;
+}
+export declare class OllamaRuntimeResidencyAdapter implements RuntimeResidencyAdapter {
+    readonly id: string;
+    readonly residencyControl: "keep-alive";
+    private readonly runtime;
+    private readonly nowMs;
+    constructor(id: string, runtime: Pick<OllamaRuntime, "listResidentModels" | "ensureResident" | "releaseResident">, options?: {
+        nowMs?: () => number;
+    });
+    list(): Promise<RuntimeResidentModel[]>;
+    ensureResident(model: string): Promise<void>;
+    release(model: string): Promise<void>;
+}
+export declare class TransformersRuntimeResidencyAdapter implements RuntimeResidencyAdapter {
+    readonly id: string;
+    readonly residencyControl: "full";
+    private readonly runtime;
+    private readonly model;
+    private readonly bytes;
+    private readonly nowMs;
+    constructor(id: string, runtime: Pick<TransformersRuntime, "detect" | "start" | "stop">, model: string, bytes: number, options?: {
+        nowMs?: () => number;
+    });
+    list(): Promise<RuntimeResidentModel[]>;
+    ensureResident(_model: string): Promise<void>;
+    release(_model: string): Promise<void>;
+}
+export declare class AdvisoryRuntimeResidencyAdapter implements RuntimeResidencyAdapter {
+    readonly id: string;
+    readonly residencyControl: "advisory";
+    private readonly residents;
+    constructor(id: string, residents: Omit<RuntimeResidentModel, "adapterId" | "residencyControl">[]);
+    list(): Promise<RuntimeResidentModel[]>;
+    ensureResident(_model: string): Promise<void>;
+    release(_model: string): Promise<void>;
+}
 export declare function planRuntimeResidency(args: {
     budgetBytes: number;
     residents: readonly RuntimeResidentModel[];

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

@@ -1 +1 @@
-{"version":3,"file":"runtime-arbiter.d.ts","sourceRoot":"","sources":["../../../src/core/models/runtime-arbiter.ts"],"names":[],"mappings":"AAAA,MAAM,MAAM,gBAAgB,GAAG,MAAM,GAAG,YAAY,GAAG,UAAU,CAAC;AAClE,MAAM,MAAM,WAAW,GAAG,QAAQ,GAAG,QAAQ,GAAG,OAAO,GAAG,OAAO,CAAC;AAElE,MAAM,WAAW,oBAAoB;IACpC,SAAS,EAAE,MAAM,CAAC;IAClB,KAAK,EAAE,MAAM,CAAC;IACd,KAAK,EAAE,MAAM,CAAC;IACd,YAAY,EAAE,MAAM,CAAC;IACrB,QAAQ,EAAE,MAAM,CAAC;IACjB,MAAM,CAAC,EAAE,OAAO,CAAC;IACjB,gBAAgB,EAAE,gBAAgB,CAAC;CACnC;AAED,MAAM,WAAW,kBAAkB;IAClC,KAAK,EAAE,MAAM,CAAC;IACd,KAAK,EAAE,MAAM,CAAC;IACd,IAAI,EAAE,WAAW,CAAC;IAClB,QAAQ,EAAE,MAAM,CAAC;IACjB,KAAK,EAAE,MAAM,CAAC;IACd,UAAU,CAAC,EAAE,MAAM,CAAC;IACpB,cAAc,CAAC,EAAE,MAAM,CAAC;IACxB,YAAY,CAAC,EAAE,kBAAkB,EAAE,CAAC;IACpC,eAAe,CAAC,EAAE,qBAAqB,EAAE,CAAC;CAC1C;AAED,MAAM,WAAW,kBAAkB;IAClC,KAAK,EAAE,MAAM,CAAC;IACd,KAAK,EAAE,MAAM,CAAC;IACd,QAAQ,EAAE,MAAM,CAAC;CACjB;AAED,MAAM,WAAW,qBAAqB;IACrC,OAAO,EAAE,MAAM,CAAC;IAChB,MAAM,EAAE,MAAM,CAAC;IACf,IAAI,EAAE,MAAM,CAAC;CACb;AAED,MAAM,MAAM,oBAAoB,GAC7B;IAAE,MAAM,EAAE,MAAM,CAAC;IAAC,KAAK,EAAE,oBAAoB,EAAE,CAAC;IAAC,aAAa,EAAE,MAAM,CAAC;IAAC,WAAW,EAAE,MAAM,CAAA;CAAE,GAC7F;IAAE,MAAM,EAAE,QAAQ,CAAC;IAAC,MAAM,EAAE,MAAM,CAAC;IAAC,KAAK,EAAE,oBAAoB,EAAE,CAAC;IAAC,aAAa,EAAE,MAAM,CAAC;IAAC,WAAW,EAAE,MAAM,CAAA;CAAE,CAAC;AAEnH,MAAM,WAAW,uBAAuB;IACvC,EAAE,EAAE,MAAM,CAAC;IACX,gBAAgB,EAAE,gBAAgB,CAAC;IACnC,IAAI,IAAI,OAAO,CAAC,oBAAoB,EAAE,CAAC,CAAC;IACxC,cAAc,CAAC,KAAK,EAAE,MAAM,GAAG,OAAO,CAAC,IAAI,CAAC,CAAC;IAC7C,OAAO,CAAC,KAAK,EAAE,MAAM,GAAG,OAAO,CAAC,IAAI,CAAC,CAAC;CACtC;AAED,wBAAgB,oBAAoB,CAAC,IAAI,EAAE;IAC1C,WAAW,EAAE,MAAM,CAAC;IACpB,SAAS,EAAE,SAAS,oBAAoB,EAAE,CAAC;IAC3C,OAAO,EAAE,kBAAkB,CAAC;CAC5B,GAAG,oBAAoB,CAkDvB","sourcesContent":["export type ResidencyControl = \"full\" | \"keep-alive\" | \"advisory\";\nexport type RuntimeRole = \"active\" | \"router\" | \"bench\" | \"probe\";\n\nexport interface RuntimeResidentModel {\n\tadapterId: string;\n\tmodel: string;\n\tbytes: number;\n\tlastUsedAtMs: number;\n\tpriority: number;\n\tpinned?: boolean;\n\tresidencyControl: ResidencyControl;\n}\n\nexport interface RuntimeLoadRequest {\n\tmodel: string;\n\tbytes: number;\n\trole: RuntimeRole;\n\tpriority: number;\n\tnowMs: number;\n\tminDwellMs?: number;\n\tpinActiveModel?: string;\n\treservations?: RuntimeReservation[];\n\trecentEvictions?: RuntimeEvictionRecord[];\n}\n\nexport interface RuntimeReservation {\n\tmodel: string;\n\tbytes: number;\n\tpriority: number;\n}\n\nexport interface RuntimeEvictionRecord {\n\tevicted: string;\n\tloaded: string;\n\tatMs: number;\n}\n\nexport type RuntimeResidencyPlan =\n\t| { status: \"fits\"; evict: RuntimeResidentModel[]; requiredBytes: number; budgetBytes: number }\n\t| { status: \"refuse\"; reason: string; evict: RuntimeResidentModel[]; requiredBytes: number; budgetBytes: number };\n\nexport interface RuntimeResidencyAdapter {\n\tid: string;\n\tresidencyControl: ResidencyControl;\n\tlist(): Promise<RuntimeResidentModel[]>;\n\tensureResident(model: string): Promise<void>;\n\trelease(model: string): Promise<void>;\n}\n\nexport function planRuntimeResidency(args: {\n\tbudgetBytes: number;\n\tresidents: readonly RuntimeResidentModel[];\n\trequest: RuntimeLoadRequest;\n}): RuntimeResidencyPlan {\n\tconst requestedModels = new Map<string, { bytes: number; priority: number }>();\n\trequestedModels.set(args.request.model, { bytes: args.request.bytes, priority: args.request.priority });\n\tfor (const reservation of args.request.reservations ?? []) {\n\t\tconst existing = requestedModels.get(reservation.model);\n\t\tif (!existing || reservation.bytes > existing.bytes) {\n\t\t\trequestedModels.set(reservation.model, { bytes: reservation.bytes, priority: reservation.priority });\n\t\t}\n\t}\n\n\tconst residentByModel = new Map(args.residents.map((resident) => [resident.model, resident]));\n\tconst residentBytes = args.residents.reduce((sum, resident) => sum + resident.bytes, 0);\n\tconst missingBytes = [...requestedModels].reduce((sum, [model, request]) => {\n\t\treturn residentByModel.has(model) ? sum : sum + request.bytes;\n\t}, 0);\n\tconst requiredBytes = residentBytes + missingBytes;\n\tif (requiredBytes <= args.budgetBytes) {\n\t\treturn { status: \"fits\", evict: [], requiredBytes, budgetBytes: args.budgetBytes };\n\t}\n\n\tif (hasPingPongRisk(args.request)) {\n\t\treturn { status: \"refuse\", reason: \"anti-thrash\", evict: [], requiredBytes, budgetBytes: args.budgetBytes };\n\t}\n\n\tlet projectedBytes = requiredBytes;\n\tconst evict: RuntimeResidentModel[] = [];\n\tconst candidates = args.residents\n\t\t.filter((resident) => !requestedModels.has(resident.model))\n\t\t.filter((resident) => !resident.pinned && resident.model !== args.request.pinActiveModel)\n\t\t.filter((resident) => resident.residencyControl !== \"advisory\")\n\t\t.filter(\n\t\t\t(resident) =>\n\t\t\t\targs.request.minDwellMs === undefined ||\n\t\t\t\targs.request.nowMs - resident.lastUsedAtMs >= args.request.minDwellMs,\n\t\t)\n\t\t.sort((left, right) => left.priority - right.priority || left.lastUsedAtMs - right.lastUsedAtMs);\n\tfor (const candidate of candidates) {\n\t\tevict.push(candidate);\n\t\tprojectedBytes -= candidate.bytes;\n\t\tif (projectedBytes <= args.budgetBytes) {\n\t\t\treturn { status: \"fits\", evict, requiredBytes, budgetBytes: args.budgetBytes };\n\t\t}\n\t}\n\treturn {\n\t\tstatus: \"refuse\",\n\t\treason: \"insufficient-evictable-memory\",\n\t\tevict,\n\t\trequiredBytes,\n\t\tbudgetBytes: args.budgetBytes,\n\t};\n}\n\nfunction hasPingPongRisk(request: RuntimeLoadRequest): boolean {\n\treturn (request.recentEvictions ?? []).some(\n\t\t(record) => record.evicted === request.model && request.nowMs - record.atMs <= (request.minDwellMs ?? 0),\n\t);\n}\n"]}
+{"version":3,"file":"runtime-arbiter.d.ts","sourceRoot":"","sources":["../../../src/core/models/runtime-arbiter.ts"],"names":[],"mappings":"AAAA,OAAO,KAAK,EAAE,aAAa,EAAE,mBAAmB,EAAE,MAAM,oBAAoB,CAAC;AAE7E,MAAM,MAAM,gBAAgB,GAAG,MAAM,GAAG,YAAY,GAAG,UAAU,CAAC;AAClE,MAAM,MAAM,WAAW,GAAG,QAAQ,GAAG,QAAQ,GAAG,OAAO,GAAG,OAAO,CAAC;AAElE,MAAM,WAAW,oBAAoB;IACpC,SAAS,EAAE,MAAM,CAAC;IAClB,KAAK,EAAE,MAAM,CAAC;IACd,KAAK,EAAE,MAAM,CAAC;IACd,YAAY,EAAE,MAAM,CAAC;IACrB,QAAQ,EAAE,MAAM,CAAC;IACjB,MAAM,CAAC,EAAE,OAAO,CAAC;IACjB,gBAAgB,EAAE,gBAAgB,CAAC;CACnC;AAED,MAAM,WAAW,kBAAkB;IAClC,KAAK,EAAE,MAAM,CAAC;IACd,KAAK,EAAE,MAAM,CAAC;IACd,IAAI,EAAE,WAAW,CAAC;IAClB,QAAQ,EAAE,MAAM,CAAC;IACjB,KAAK,EAAE,MAAM,CAAC;IACd,UAAU,CAAC,EAAE,MAAM,CAAC;IACpB,cAAc,CAAC,EAAE,MAAM,CAAC;IACxB,YAAY,CAAC,EAAE,kBAAkB,EAAE,CAAC;IACpC,eAAe,CAAC,EAAE,qBAAqB,EAAE,CAAC;CAC1C;AAED,MAAM,WAAW,kBAAkB;IAClC,KAAK,EAAE,MAAM,CAAC;IACd,KAAK,EAAE,MAAM,CAAC;IACd,QAAQ,EAAE,MAAM,CAAC;CACjB;AAED,MAAM,WAAW,qBAAqB;IACrC,OAAO,EAAE,MAAM,CAAC;IAChB,MAAM,EAAE,MAAM,CAAC;IACf,IAAI,EAAE,MAAM,CAAC;CACb;AAED,MAAM,MAAM,oBAAoB,GAC7B;IAAE,MAAM,EAAE,MAAM,CAAC;IAAC,KAAK,EAAE,oBAAoB,EAAE,CAAC;IAAC,aAAa,EAAE,MAAM,CAAC;IAAC,WAAW,EAAE,MAAM,CAAA;CAAE,GAC7F;IAAE,MAAM,EAAE,QAAQ,CAAC;IAAC,MAAM,EAAE,MAAM,CAAC;IAAC,KAAK,EAAE,oBAAoB,EAAE,CAAC;IAAC,aAAa,EAAE,MAAM,CAAC;IAAC,WAAW,EAAE,MAAM,CAAA;CAAE,CAAC;AAEnH,MAAM,WAAW,uBAAuB;IACvC,EAAE,EAAE,MAAM,CAAC;IACX,gBAAgB,EAAE,gBAAgB,CAAC;IACnC,IAAI,IAAI,OAAO,CAAC,oBAAoB,EAAE,CAAC,CAAC;IACxC,cAAc,CAAC,KAAK,EAAE,MAAM,GAAG,OAAO,CAAC,IAAI,CAAC,CAAC;IAC7C,OAAO,CAAC,KAAK,EAAE,MAAM,GAAG,OAAO,CAAC,IAAI,CAAC,CAAC;CACtC;AAED,qBAAa,uBAAuB;IACnC,OAAO,CAAC,QAAQ,CAAC,WAAW,CAAS;IACrC,OAAO,CAAC,QAAQ,CAAC,QAAQ,CAAuC;IAEhE,YAAY,IAAI,EAAE;QAAE,WAAW,EAAE,MAAM,CAAC;QAAC,QAAQ,EAAE,uBAAuB,EAAE,CAAA;KAAE,EAG7E;IAEK,cAAc,CAAC,SAAS,EAAE,MAAM,EAAE,OAAO,EAAE,kBAAkB,GAAG,OAAO,CAAC,oBAAoB,CAAC,CAmBlG;CACD;AAED,qBAAa,6BAA8B,YAAW,uBAAuB;IAC5E,QAAQ,CAAC,EAAE,EAAE,MAAM,CAAC;IACpB,QAAQ,CAAC,gBAAgB,eAAyB;IAClD,OAAO,CAAC,QAAQ,CAAC,OAAO,CAAmF;IAC3G,OAAO,CAAC,QAAQ,CAAC,KAAK,CAAe;IAErC,YACC,EAAE,EAAE,MAAM,EACV,OAAO,EAAE,IAAI,CAAC,aAAa,EAAE,oBAAoB,GAAG,gBAAgB,GAAG,iBAAiB,CAAC,EACzF,OAAO,GAAE;QAAE,KAAK,CAAC,EAAE,MAAM,MAAM,CAAA;KAAO,EAKtC;IAEK,IAAI,IAAI,OAAO,CAAC,oBAAoB,EAAE,CAAC,CAU5C;IAEK,cAAc,CAAC,KAAK,EAAE,MAAM,GAAG,OAAO,CAAC,IAAI,CAAC,CAGjD;IAEK,OAAO,CAAC,KAAK,EAAE,MAAM,GAAG,OAAO,CAAC,IAAI,CAAC,CAE1C;CACD;AAED,qBAAa,mCAAoC,YAAW,uBAAuB;IAClF,QAAQ,CAAC,EAAE,EAAE,MAAM,CAAC;IACpB,QAAQ,CAAC,gBAAgB,SAAmB;IAC5C,OAAO,CAAC,QAAQ,CAAC,OAAO,CAAyD;IACjF,OAAO,CAAC,QAAQ,CAAC,KAAK,CAAS;IAC/B,OAAO,CAAC,QAAQ,CAAC,KAAK,CAAS;IAC/B,OAAO,CAAC,QAAQ,CAAC,KAAK,CAAe;IAErC,YACC,EAAE,EAAE,MAAM,EACV,OAAO,EAAE,IAAI,CAAC,mBAAmB,EAAE,QAAQ,GAAG,OAAO,GAAG,MAAM,CAAC,EAC/D,KAAK,EAAE,MAAM,EACb,KAAK,EAAE,MAAM,EACb,OAAO,GAAE;QAAE,KAAK,CAAC,EAAE,MAAM,MAAM,CAAA;KAAO,EAOtC;IAEK,IAAI,IAAI,OAAO,CAAC,oBAAoB,EAAE,CAAC,CAa5C;IAEK,cAAc,CAAC,MAAM,EAAE,MAAM,GAAG,OAAO,CAAC,IAAI,CAAC,CAGlD;IAEK,OAAO,CAAC,MAAM,EAAE,MAAM,GAAG,OAAO,CAAC,IAAI,CAAC,CAE3C;CACD;AAED,qBAAa,+BAAgC,YAAW,uBAAuB;IAC9E,QAAQ,CAAC,EAAE,EAAE,MAAM,CAAC;IACpB,QAAQ,CAAC,gBAAgB,aAAuB;IAChD,OAAO,CAAC,QAAQ,CAAC,SAAS,CAAyB;IAEnD,YAAY,EAAE,EAAE,MAAM,EAAE,SAAS,EAAE,IAAI,CAAC,oBAAoB,EAAE,WAAW,GAAG,kBAAkB,CAAC,EAAE,EAOhG;IAEK,IAAI,IAAI,OAAO,CAAC,oBAAoB,EAAE,CAAC,CAE5C;IAEK,cAAc,CAAC,MAAM,EAAE,MAAM,GAAG,OAAO,CAAC,IAAI,CAAC,CAAG;IAEhD,OAAO,CAAC,MAAM,EAAE,MAAM,GAAG,OAAO,CAAC,IAAI,CAAC,CAAG;CAC/C;AAED,wBAAgB,oBAAoB,CAAC,IAAI,EAAE;IAC1C,WAAW,EAAE,MAAM,CAAC;IACpB,SAAS,EAAE,SAAS,oBAAoB,EAAE,CAAC;IAC3C,OAAO,EAAE,kBAAkB,CAAC;CAC5B,GAAG,oBAAoB,CAkDvB","sourcesContent":["import type { OllamaRuntime, TransformersRuntime } from \"./local-runtime.ts\";\n\nexport type ResidencyControl = \"full\" | \"keep-alive\" | \"advisory\";\nexport type RuntimeRole = \"active\" | \"router\" | \"bench\" | \"probe\";\n\nexport interface RuntimeResidentModel {\n\tadapterId: string;\n\tmodel: string;\n\tbytes: number;\n\tlastUsedAtMs: number;\n\tpriority: number;\n\tpinned?: boolean;\n\tresidencyControl: ResidencyControl;\n}\n\nexport interface RuntimeLoadRequest {\n\tmodel: string;\n\tbytes: number;\n\trole: RuntimeRole;\n\tpriority: number;\n\tnowMs: number;\n\tminDwellMs?: number;\n\tpinActiveModel?: string;\n\treservations?: RuntimeReservation[];\n\trecentEvictions?: RuntimeEvictionRecord[];\n}\n\nexport interface RuntimeReservation {\n\tmodel: string;\n\tbytes: number;\n\tpriority: number;\n}\n\nexport interface RuntimeEvictionRecord {\n\tevicted: string;\n\tloaded: string;\n\tatMs: number;\n}\n\nexport type RuntimeResidencyPlan =\n\t| { status: \"fits\"; evict: RuntimeResidentModel[]; requiredBytes: number; budgetBytes: number }\n\t| { status: \"refuse\"; reason: string; evict: RuntimeResidentModel[]; requiredBytes: number; budgetBytes: number };\n\nexport interface RuntimeResidencyAdapter {\n\tid: string;\n\tresidencyControl: ResidencyControl;\n\tlist(): Promise<RuntimeResidentModel[]>;\n\tensureResident(model: string): Promise<void>;\n\trelease(model: string): Promise<void>;\n}\n\nexport class RuntimeResidencyArbiter {\n\tprivate readonly budgetBytes: number;\n\tprivate readonly adapters: Map<string, RuntimeResidencyAdapter>;\n\n\tconstructor(args: { budgetBytes: number; adapters: RuntimeResidencyAdapter[] }) {\n\t\tthis.budgetBytes = args.budgetBytes;\n\t\tthis.adapters = new Map(args.adapters.map((adapter) => [adapter.id, adapter]));\n\t}\n\n\tasync ensureResident(adapterId: string, request: RuntimeLoadRequest): Promise<RuntimeResidencyPlan> {\n\t\tconst adapter = this.adapters.get(adapterId);\n\t\tconst residents = (await Promise.all([...this.adapters.values()].map((entry) => entry.list()))).flat();\n\t\tif (!adapter) {\n\t\t\treturn {\n\t\t\t\tstatus: \"refuse\",\n\t\t\t\treason: \"unknown-adapter\",\n\t\t\t\tevict: [],\n\t\t\t\trequiredBytes: request.bytes,\n\t\t\t\tbudgetBytes: this.budgetBytes,\n\t\t\t};\n\t\t}\n\t\tconst plan = planRuntimeResidency({ budgetBytes: this.budgetBytes, residents, request });\n\t\tif (plan.status !== \"fits\") return plan;\n\t\tfor (const resident of plan.evict) {\n\t\t\tawait this.adapters.get(resident.adapterId)?.release(resident.model);\n\t\t}\n\t\tawait adapter.ensureResident(request.model);\n\t\treturn plan;\n\t}\n}\n\nexport class OllamaRuntimeResidencyAdapter implements RuntimeResidencyAdapter {\n\treadonly id: string;\n\treadonly residencyControl = \"keep-alive\" as const;\n\tprivate readonly runtime: Pick<OllamaRuntime, \"listResidentModels\" | \"ensureResident\" | \"releaseResident\">;\n\tprivate readonly nowMs: () => number;\n\n\tconstructor(\n\t\tid: string,\n\t\truntime: Pick<OllamaRuntime, \"listResidentModels\" | \"ensureResident\" | \"releaseResident\">,\n\t\toptions: { nowMs?: () => number } = {},\n\t) {\n\t\tthis.id = id;\n\t\tthis.runtime = runtime;\n\t\tthis.nowMs = options.nowMs ?? Date.now;\n\t}\n\n\tasync list(): Promise<RuntimeResidentModel[]> {\n\t\tconst now = this.nowMs();\n\t\treturn (await this.runtime.listResidentModels()).map((model) => ({\n\t\t\tadapterId: this.id,\n\t\t\tmodel: model.name,\n\t\t\tbytes: model.sizeBytes,\n\t\t\tlastUsedAtMs: now,\n\t\t\tpriority: 10,\n\t\t\tresidencyControl: this.residencyControl,\n\t\t}));\n\t}\n\n\tasync ensureResident(model: string): Promise<void> {\n\t\tconst result = await this.runtime.ensureResident(model);\n\t\tif (!result.ok) throw new Error(result.error);\n\t}\n\n\tasync release(model: string): Promise<void> {\n\t\tawait this.runtime.releaseResident(model);\n\t}\n}\n\nexport class TransformersRuntimeResidencyAdapter implements RuntimeResidencyAdapter {\n\treadonly id: string;\n\treadonly residencyControl = \"full\" as const;\n\tprivate readonly runtime: Pick<TransformersRuntime, \"detect\" | \"start\" | \"stop\">;\n\tprivate readonly model: string;\n\tprivate readonly bytes: number;\n\tprivate readonly nowMs: () => number;\n\n\tconstructor(\n\t\tid: string,\n\t\truntime: Pick<TransformersRuntime, \"detect\" | \"start\" | \"stop\">,\n\t\tmodel: string,\n\t\tbytes: number,\n\t\toptions: { nowMs?: () => number } = {},\n\t) {\n\t\tthis.id = id;\n\t\tthis.runtime = runtime;\n\t\tthis.model = model;\n\t\tthis.bytes = bytes;\n\t\tthis.nowMs = options.nowMs ?? Date.now;\n\t}\n\n\tasync list(): Promise<RuntimeResidentModel[]> {\n\t\tconst status = await this.runtime.detect();\n\t\tif (!status.serverUp) return [];\n\t\treturn [\n\t\t\t{\n\t\t\t\tadapterId: this.id,\n\t\t\t\tmodel: this.model,\n\t\t\t\tbytes: this.bytes,\n\t\t\t\tlastUsedAtMs: this.nowMs(),\n\t\t\t\tpriority: 10,\n\t\t\t\tresidencyControl: this.residencyControl,\n\t\t\t},\n\t\t];\n\t}\n\n\tasync ensureResident(_model: string): Promise<void> {\n\t\tconst result = await this.runtime.start();\n\t\tif (!result.started && result.reason !== \"already_running\") throw new Error(result.reason);\n\t}\n\n\tasync release(_model: string): Promise<void> {\n\t\tthis.runtime.stop();\n\t}\n}\n\nexport class AdvisoryRuntimeResidencyAdapter implements RuntimeResidencyAdapter {\n\treadonly id: string;\n\treadonly residencyControl = \"advisory\" as const;\n\tprivate readonly residents: RuntimeResidentModel[];\n\n\tconstructor(id: string, residents: Omit<RuntimeResidentModel, \"adapterId\" | \"residencyControl\">[]) {\n\t\tthis.id = id;\n\t\tthis.residents = residents.map((resident) => ({\n\t\t\t...resident,\n\t\t\tadapterId: id,\n\t\t\tresidencyControl: this.residencyControl,\n\t\t}));\n\t}\n\n\tasync list(): Promise<RuntimeResidentModel[]> {\n\t\treturn [...this.residents];\n\t}\n\n\tasync ensureResident(_model: string): Promise<void> {}\n\n\tasync release(_model: string): Promise<void> {}\n}\n\nexport function planRuntimeResidency(args: {\n\tbudgetBytes: number;\n\tresidents: readonly RuntimeResidentModel[];\n\trequest: RuntimeLoadRequest;\n}): RuntimeResidencyPlan {\n\tconst requestedModels = new Map<string, { bytes: number; priority: number }>();\n\trequestedModels.set(args.request.model, { bytes: args.request.bytes, priority: args.request.priority });\n\tfor (const reservation of args.request.reservations ?? []) {\n\t\tconst existing = requestedModels.get(reservation.model);\n\t\tif (!existing || reservation.bytes > existing.bytes) {\n\t\t\trequestedModels.set(reservation.model, { bytes: reservation.bytes, priority: reservation.priority });\n\t\t}\n\t}\n\n\tconst residentByModel = new Map(args.residents.map((resident) => [resident.model, resident]));\n\tconst residentBytes = args.residents.reduce((sum, resident) => sum + resident.bytes, 0);\n\tconst missingBytes = [...requestedModels].reduce((sum, [model, request]) => {\n\t\treturn residentByModel.has(model) ? sum : sum + request.bytes;\n\t}, 0);\n\tconst requiredBytes = residentBytes + missingBytes;\n\tif (requiredBytes <= args.budgetBytes) {\n\t\treturn { status: \"fits\", evict: [], requiredBytes, budgetBytes: args.budgetBytes };\n\t}\n\n\tif (hasPingPongRisk(args.request)) {\n\t\treturn { status: \"refuse\", reason: \"anti-thrash\", evict: [], requiredBytes, budgetBytes: args.budgetBytes };\n\t}\n\n\tlet projectedBytes = requiredBytes;\n\tconst evict: RuntimeResidentModel[] = [];\n\tconst candidates = args.residents\n\t\t.filter((resident) => !requestedModels.has(resident.model))\n\t\t.filter((resident) => !resident.pinned && resident.model !== args.request.pinActiveModel)\n\t\t.filter((resident) => resident.residencyControl !== \"advisory\")\n\t\t.filter(\n\t\t\t(resident) =>\n\t\t\t\targs.request.minDwellMs === undefined ||\n\t\t\t\targs.request.nowMs - resident.lastUsedAtMs >= args.request.minDwellMs,\n\t\t)\n\t\t.sort((left, right) => left.priority - right.priority || left.lastUsedAtMs - right.lastUsedAtMs);\n\tfor (const candidate of candidates) {\n\t\tevict.push(candidate);\n\t\tprojectedBytes -= candidate.bytes;\n\t\tif (projectedBytes <= args.budgetBytes) {\n\t\t\treturn { status: \"fits\", evict, requiredBytes, budgetBytes: args.budgetBytes };\n\t\t}\n\t}\n\treturn {\n\t\tstatus: \"refuse\",\n\t\treason: \"insufficient-evictable-memory\",\n\t\tevict,\n\t\trequiredBytes,\n\t\tbudgetBytes: args.budgetBytes,\n\t};\n}\n\nfunction hasPingPongRisk(request: RuntimeLoadRequest): boolean {\n\treturn (request.recentEvictions ?? []).some(\n\t\t(record) => record.evicted === request.model && request.nowMs - record.atMs <= (request.minDwellMs ?? 0),\n\t);\n}\n"]}

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

@@ -1,3 +1,118 @@
+export class RuntimeResidencyArbiter {
+    budgetBytes;
+    adapters;
+    constructor(args) {
+        this.budgetBytes = args.budgetBytes;
+        this.adapters = new Map(args.adapters.map((adapter) => [adapter.id, adapter]));
+    }
+    async ensureResident(adapterId, request) {
+        const adapter = this.adapters.get(adapterId);
+        const residents = (await Promise.all([...this.adapters.values()].map((entry) => entry.list()))).flat();
+        if (!adapter) {
+            return {
+                status: "refuse",
+                reason: "unknown-adapter",
+                evict: [],
+                requiredBytes: request.bytes,
+                budgetBytes: this.budgetBytes,
+            };
+        }
+        const plan = planRuntimeResidency({ budgetBytes: this.budgetBytes, residents, request });
+        if (plan.status !== "fits")
+            return plan;
+        for (const resident of plan.evict) {
+            await this.adapters.get(resident.adapterId)?.release(resident.model);
+        }
+        await adapter.ensureResident(request.model);
+        return plan;
+    }
+}
+export class OllamaRuntimeResidencyAdapter {
+    id;
+    residencyControl = "keep-alive";
+    runtime;
+    nowMs;
+    constructor(id, runtime, options = {}) {
+        this.id = id;
+        this.runtime = runtime;
+        this.nowMs = options.nowMs ?? Date.now;
+    }
+    async list() {
+        const now = this.nowMs();
+        return (await this.runtime.listResidentModels()).map((model) => ({
+            adapterId: this.id,
+            model: model.name,
+            bytes: model.sizeBytes,
+            lastUsedAtMs: now,
+            priority: 10,
+            residencyControl: this.residencyControl,
+        }));
+    }
+    async ensureResident(model) {
+        const result = await this.runtime.ensureResident(model);
+        if (!result.ok)
+            throw new Error(result.error);
+    }
+    async release(model) {
+        await this.runtime.releaseResident(model);
+    }
+}
+export class TransformersRuntimeResidencyAdapter {
+    id;
+    residencyControl = "full";
+    runtime;
+    model;
+    bytes;
+    nowMs;
+    constructor(id, runtime, model, bytes, options = {}) {
+        this.id = id;
+        this.runtime = runtime;
+        this.model = model;
+        this.bytes = bytes;
+        this.nowMs = options.nowMs ?? Date.now;
+    }
+    async list() {
+        const status = await this.runtime.detect();
+        if (!status.serverUp)
+            return [];
+        return [
+            {
+                adapterId: this.id,
+                model: this.model,
+                bytes: this.bytes,
+                lastUsedAtMs: this.nowMs(),
+                priority: 10,
+                residencyControl: this.residencyControl,
+            },
+        ];
+    }
+    async ensureResident(_model) {
+        const result = await this.runtime.start();
+        if (!result.started && result.reason !== "already_running")
+            throw new Error(result.reason);
+    }
+    async release(_model) {
+        this.runtime.stop();
+    }
+}
+export class AdvisoryRuntimeResidencyAdapter {
+    id;
+    residencyControl = "advisory";
+    residents;
+    constructor(id, residents) {
+        this.id = id;
+        this.residents = residents.map((resident) => ({
+            ...resident,
+            adapterId: id,
+            residencyControl: this.residencyControl,
+        }));
+    }
+    async list() {
+        return [...this.residents];
+    }
+    async ensureResident(_model) { }
+    async release(_model) { }
+}
 export function planRuntimeResidency(args) {
     const requestedModels = new Map();
     requestedModels.set(args.request.model, { bytes: args.request.bytes, priority: args.request.priority });

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

@@ -1 +1 @@
-{"version":3,"file":"runtime-arbiter.js","sourceRoot":"","sources":["../../../src/core/models/runtime-arbiter.ts"],"names":[],"mappings":"AAiDA,MAAM,UAAU,oBAAoB,CAAC,IAIpC,EAAwB;IACxB,MAAM,eAAe,GAAG,IAAI,GAAG,EAA+C,CAAC;IAC/E,eAAe,CAAC,GAAG,CAAC,IAAI,CAAC,OAAO,CAAC,KAAK,EAAE,EAAE,KAAK,EAAE,IAAI,CAAC,OAAO,CAAC,KAAK,EAAE,QAAQ,EAAE,IAAI,CAAC,OAAO,CAAC,QAAQ,EAAE,CAAC,CAAC;IACxG,KAAK,MAAM,WAAW,IAAI,IAAI,CAAC,OAAO,CAAC,YAAY,IAAI,EAAE,EAAE,CAAC;QAC3D,MAAM,QAAQ,GAAG,eAAe,CAAC,GAAG,CAAC,WAAW,CAAC,KAAK,CAAC,CAAC;QACxD,IAAI,CAAC,QAAQ,IAAI,WAAW,CAAC,KAAK,GAAG,QAAQ,CAAC,KAAK,EAAE,CAAC;YACrD,eAAe,CAAC,GAAG,CAAC,WAAW,CAAC,KAAK,EAAE,EAAE,KAAK,EAAE,WAAW,CAAC,KAAK,EAAE,QAAQ,EAAE,WAAW,CAAC,QAAQ,EAAE,CAAC,CAAC;QACtG,CAAC;IACF,CAAC;IAED,MAAM,eAAe,GAAG,IAAI,GAAG,CAAC,IAAI,CAAC,SAAS,CAAC,GAAG,CAAC,CAAC,QAAQ,EAAE,EAAE,CAAC,CAAC,QAAQ,CAAC,KAAK,EAAE,QAAQ,CAAC,CAAC,CAAC,CAAC;IAC9F,MAAM,aAAa,GAAG,IAAI,CAAC,SAAS,CAAC,MAAM,CAAC,CAAC,GAAG,EAAE,QAAQ,EAAE,EAAE,CAAC,GAAG,GAAG,QAAQ,CAAC,KAAK,EAAE,CAAC,CAAC,CAAC;IACxF,MAAM,YAAY,GAAG,CAAC,GAAG,eAAe,CAAC,CAAC,MAAM,CAAC,CAAC,GAAG,EAAE,CAAC,KAAK,EAAE,OAAO,CAAC,EAAE,EAAE,CAAC;QAC3E,OAAO,eAAe,CAAC,GAAG,CAAC,KAAK,CAAC,CAAC,CAAC,CAAC,GAAG,CAAC,CAAC,CAAC,GAAG,GAAG,OAAO,CAAC,KAAK,CAAC;IAAA,CAC9D,EAAE,CAAC,CAAC,CAAC;IACN,MAAM,aAAa,GAAG,aAAa,GAAG,YAAY,CAAC;IACnD,IAAI,aAAa,IAAI,IAAI,CAAC,WAAW,EAAE,CAAC;QACvC,OAAO,EAAE,MAAM,EAAE,MAAM,EAAE,KAAK,EAAE,EAAE,EAAE,aAAa,EAAE,WAAW,EAAE,IAAI,CAAC,WAAW,EAAE,CAAC;IACpF,CAAC;IAED,IAAI,eAAe,CAAC,IAAI,CAAC,OAAO,CAAC,EAAE,CAAC;QACnC,OAAO,EAAE,MAAM,EAAE,QAAQ,EAAE,MAAM,EAAE,aAAa,EAAE,KAAK,EAAE,EAAE,EAAE,aAAa,EAAE,WAAW,EAAE,IAAI,CAAC,WAAW,EAAE,CAAC;IAC7G,CAAC;IAED,IAAI,cAAc,GAAG,aAAa,CAAC;IACnC,MAAM,KAAK,GAA2B,EAAE,CAAC;IACzC,MAAM,UAAU,GAAG,IAAI,CAAC,SAAS;SAC/B,MAAM,CAAC,CAAC,QAAQ,EAAE,EAAE,CAAC,CAAC,eAAe,CAAC,GAAG,CAAC,QAAQ,CAAC,KAAK,CAAC,CAAC;SAC1D,MAAM,CAAC,CAAC,QAAQ,EAAE,EAAE,CAAC,CAAC,QAAQ,CAAC,MAAM,IAAI,QAAQ,CAAC,KAAK,KAAK,IAAI,CAAC,OAAO,CAAC,cAAc,CAAC;SACxF,MAAM,CAAC,CAAC,QAAQ,EAAE,EAAE,CAAC,QAAQ,CAAC,gBAAgB,KAAK,UAAU,CAAC;SAC9D,MAAM,CACN,CAAC,QAAQ,EAAE,EAAE,CACZ,IAAI,CAAC,OAAO,CAAC,UAAU,KAAK,SAAS;QACrC,IAAI,CAAC,OAAO,CAAC,KAAK,GAAG,QAAQ,CAAC,YAAY,IAAI,IAAI,CAAC,OAAO,CAAC,UAAU,CACtE;SACA,IAAI,CAAC,CAAC,IAAI,EAAE,KAAK,EAAE,EAAE,CAAC,IAAI,CAAC,QAAQ,GAAG,KAAK,CAAC,QAAQ,IAAI,IAAI,CAAC,YAAY,GAAG,KAAK,CAAC,YAAY,CAAC,CAAC;IAClG,KAAK,MAAM,SAAS,IAAI,UAAU,EAAE,CAAC;QACpC,KAAK,CAAC,IAAI,CAAC,SAAS,CAAC,CAAC;QACtB,cAAc,IAAI,SAAS,CAAC,KAAK,CAAC;QAClC,IAAI,cAAc,IAAI,IAAI,CAAC,WAAW,EAAE,CAAC;YACxC,OAAO,EAAE,MAAM,EAAE,MAAM,EAAE,KAAK,EAAE,aAAa,EAAE,WAAW,EAAE,IAAI,CAAC,WAAW,EAAE,CAAC;QAChF,CAAC;IACF,CAAC;IACD,OAAO;QACN,MAAM,EAAE,QAAQ;QAChB,MAAM,EAAE,+BAA+B;QACvC,KAAK;QACL,aAAa;QACb,WAAW,EAAE,IAAI,CAAC,WAAW;KAC7B,CAAC;AAAA,CACF;AAED,SAAS,eAAe,CAAC,OAA2B,EAAW;IAC9D,OAAO,CAAC,OAAO,CAAC,eAAe,IAAI,EAAE,CAAC,CAAC,IAAI,CAC1C,CAAC,MAAM,EAAE,EAAE,CAAC,MAAM,CAAC,OAAO,KAAK,OAAO,CAAC,KAAK,IAAI,OAAO,CAAC,KAAK,GAAG,MAAM,CAAC,IAAI,IAAI,CAAC,OAAO,CAAC,UAAU,IAAI,CAAC,CAAC,CACxG,CAAC;AAAA,CACF","sourcesContent":["export type ResidencyControl = \"full\" | \"keep-alive\" | \"advisory\";\nexport type RuntimeRole = \"active\" | \"router\" | \"bench\" | \"probe\";\n\nexport interface RuntimeResidentModel {\n\tadapterId: string;\n\tmodel: string;\n\tbytes: number;\n\tlastUsedAtMs: number;\n\tpriority: number;\n\tpinned?: boolean;\n\tresidencyControl: ResidencyControl;\n}\n\nexport interface RuntimeLoadRequest {\n\tmodel: string;\n\tbytes: number;\n\trole: RuntimeRole;\n\tpriority: number;\n\tnowMs: number;\n\tminDwellMs?: number;\n\tpinActiveModel?: string;\n\treservations?: RuntimeReservation[];\n\trecentEvictions?: RuntimeEvictionRecord[];\n}\n\nexport interface RuntimeReservation {\n\tmodel: string;\n\tbytes: number;\n\tpriority: number;\n}\n\nexport interface RuntimeEvictionRecord {\n\tevicted: string;\n\tloaded: string;\n\tatMs: number;\n}\n\nexport type RuntimeResidencyPlan =\n\t| { status: \"fits\"; evict: RuntimeResidentModel[]; requiredBytes: number; budgetBytes: number }\n\t| { status: \"refuse\"; reason: string; evict: RuntimeResidentModel[]; requiredBytes: number; budgetBytes: number };\n\nexport interface RuntimeResidencyAdapter {\n\tid: string;\n\tresidencyControl: ResidencyControl;\n\tlist(): Promise<RuntimeResidentModel[]>;\n\tensureResident(model: string): Promise<void>;\n\trelease(model: string): Promise<void>;\n}\n\nexport function planRuntimeResidency(args: {\n\tbudgetBytes: number;\n\tresidents: readonly RuntimeResidentModel[];\n\trequest: RuntimeLoadRequest;\n}): RuntimeResidencyPlan {\n\tconst requestedModels = new Map<string, { bytes: number; priority: number }>();\n\trequestedModels.set(args.request.model, { bytes: args.request.bytes, priority: args.request.priority });\n\tfor (const reservation of args.request.reservations ?? []) {\n\t\tconst existing = requestedModels.get(reservation.model);\n\t\tif (!existing || reservation.bytes > existing.bytes) {\n\t\t\trequestedModels.set(reservation.model, { bytes: reservation.bytes, priority: reservation.priority });\n\t\t}\n\t}\n\n\tconst residentByModel = new Map(args.residents.map((resident) => [resident.model, resident]));\n\tconst residentBytes = args.residents.reduce((sum, resident) => sum + resident.bytes, 0);\n\tconst missingBytes = [...requestedModels].reduce((sum, [model, request]) => {\n\t\treturn residentByModel.has(model) ? sum : sum + request.bytes;\n\t}, 0);\n\tconst requiredBytes = residentBytes + missingBytes;\n\tif (requiredBytes <= args.budgetBytes) {\n\t\treturn { status: \"fits\", evict: [], requiredBytes, budgetBytes: args.budgetBytes };\n\t}\n\n\tif (hasPingPongRisk(args.request)) {\n\t\treturn { status: \"refuse\", reason: \"anti-thrash\", evict: [], requiredBytes, budgetBytes: args.budgetBytes };\n\t}\n\n\tlet projectedBytes = requiredBytes;\n\tconst evict: RuntimeResidentModel[] = [];\n\tconst candidates = args.residents\n\t\t.filter((resident) => !requestedModels.has(resident.model))\n\t\t.filter((resident) => !resident.pinned && resident.model !== args.request.pinActiveModel)\n\t\t.filter((resident) => resident.residencyControl !== \"advisory\")\n\t\t.filter(\n\t\t\t(resident) =>\n\t\t\t\targs.request.minDwellMs === undefined ||\n\t\t\t\targs.request.nowMs - resident.lastUsedAtMs >= args.request.minDwellMs,\n\t\t)\n\t\t.sort((left, right) => left.priority - right.priority || left.lastUsedAtMs - right.lastUsedAtMs);\n\tfor (const candidate of candidates) {\n\t\tevict.push(candidate);\n\t\tprojectedBytes -= candidate.bytes;\n\t\tif (projectedBytes <= args.budgetBytes) {\n\t\t\treturn { status: \"fits\", evict, requiredBytes, budgetBytes: args.budgetBytes };\n\t\t}\n\t}\n\treturn {\n\t\tstatus: \"refuse\",\n\t\treason: \"insufficient-evictable-memory\",\n\t\tevict,\n\t\trequiredBytes,\n\t\tbudgetBytes: args.budgetBytes,\n\t};\n}\n\nfunction hasPingPongRisk(request: RuntimeLoadRequest): boolean {\n\treturn (request.recentEvictions ?? []).some(\n\t\t(record) => record.evicted === request.model && request.nowMs - record.atMs <= (request.minDwellMs ?? 0),\n\t);\n}\n"]}
+{"version":3,"file":"runtime-arbiter.js","sourceRoot":"","sources":["../../../src/core/models/runtime-arbiter.ts"],"names":[],"mappings":"AAmDA,MAAM,OAAO,uBAAuB;IAClB,WAAW,CAAS;IACpB,QAAQ,CAAuC;IAEhE,YAAY,IAAkE,EAAE;QAC/E,IAAI,CAAC,WAAW,GAAG,IAAI,CAAC,WAAW,CAAC;QACpC,IAAI,CAAC,QAAQ,GAAG,IAAI,GAAG,CAAC,IAAI,CAAC,QAAQ,CAAC,GAAG,CAAC,CAAC,OAAO,EAAE,EAAE,CAAC,CAAC,OAAO,CAAC,EAAE,EAAE,OAAO,CAAC,CAAC,CAAC,CAAC;IAAA,CAC/E;IAED,KAAK,CAAC,cAAc,CAAC,SAAiB,EAAE,OAA2B,EAAiC;QACnG,MAAM,OAAO,GAAG,IAAI,CAAC,QAAQ,CAAC,GAAG,CAAC,SAAS,CAAC,CAAC;QAC7C,MAAM,SAAS,GAAG,CAAC,MAAM,OAAO,CAAC,GAAG,CAAC,CAAC,GAAG,IAAI,CAAC,QAAQ,CAAC,MAAM,EAAE,CAAC,CAAC,GAAG,CAAC,CAAC,KAAK,EAAE,EAAE,CAAC,KAAK,CAAC,IAAI,EAAE,CAAC,CAAC,CAAC,CAAC,IAAI,EAAE,CAAC;QACvG,IAAI,CAAC,OAAO,EAAE,CAAC;YACd,OAAO;gBACN,MAAM,EAAE,QAAQ;gBAChB,MAAM,EAAE,iBAAiB;gBACzB,KAAK,EAAE,EAAE;gBACT,aAAa,EAAE,OAAO,CAAC,KAAK;gBAC5B,WAAW,EAAE,IAAI,CAAC,WAAW;aAC7B,CAAC;QACH,CAAC;QACD,MAAM,IAAI,GAAG,oBAAoB,CAAC,EAAE,WAAW,EAAE,IAAI,CAAC,WAAW,EAAE,SAAS,EAAE,OAAO,EAAE,CAAC,CAAC;QACzF,IAAI,IAAI,CAAC,MAAM,KAAK,MAAM;YAAE,OAAO,IAAI,CAAC;QACxC,KAAK,MAAM,QAAQ,IAAI,IAAI,CAAC,KAAK,EAAE,CAAC;YACnC,MAAM,IAAI,CAAC,QAAQ,CAAC,GAAG,CAAC,QAAQ,CAAC,SAAS,CAAC,EAAE,OAAO,CAAC,QAAQ,CAAC,KAAK,CAAC,CAAC;QACtE,CAAC;QACD,MAAM,OAAO,CAAC,cAAc,CAAC,OAAO,CAAC,KAAK,CAAC,CAAC;QAC5C,OAAO,IAAI,CAAC;IAAA,CACZ;CACD;AAED,MAAM,OAAO,6BAA6B;IAChC,EAAE,CAAS;IACX,gBAAgB,GAAG,YAAqB,CAAC;IACjC,OAAO,CAAmF;IAC1F,KAAK,CAAe;IAErC,YACC,EAAU,EACV,OAAyF,EACzF,OAAO,GAA6B,EAAE,EACrC;QACD,IAAI,CAAC,EAAE,GAAG,EAAE,CAAC;QACb,IAAI,CAAC,OAAO,GAAG,OAAO,CAAC;QACvB,IAAI,CAAC,KAAK,GAAG,OAAO,CAAC,KAAK,IAAI,IAAI,CAAC,GAAG,CAAC;IAAA,CACvC;IAED,KAAK,CAAC,IAAI,GAAoC;QAC7C,MAAM,GAAG,GAAG,IAAI,CAAC,KAAK,EAAE,CAAC;QACzB,OAAO,CAAC,MAAM,IAAI,CAAC,OAAO,CAAC,kBAAkB,EAAE,CAAC,CAAC,GAAG,CAAC,CAAC,KAAK,EAAE,EAAE,CAAC,CAAC;YAChE,SAAS,EAAE,IAAI,CAAC,EAAE;YAClB,KAAK,EAAE,KAAK,CAAC,IAAI;YACjB,KAAK,EAAE,KAAK,CAAC,SAAS;YACtB,YAAY,EAAE,GAAG;YACjB,QAAQ,EAAE,EAAE;YACZ,gBAAgB,EAAE,IAAI,CAAC,gBAAgB;SACvC,CAAC,CAAC,CAAC;IAAA,CACJ;IAED,KAAK,CAAC,cAAc,CAAC,KAAa,EAAiB;QAClD,MAAM,MAAM,GAAG,MAAM,IAAI,CAAC,OAAO,CAAC,cAAc,CAAC,KAAK,CAAC,CAAC;QACxD,IAAI,CAAC,MAAM,CAAC,EAAE;YAAE,MAAM,IAAI,KAAK,CAAC,MAAM,CAAC,KAAK,CAAC,CAAC;IAAA,CAC9C;IAED,KAAK,CAAC,OAAO,CAAC,KAAa,EAAiB;QAC3C,MAAM,IAAI,CAAC,OAAO,CAAC,eAAe,CAAC,KAAK,CAAC,CAAC;IAAA,CAC1C;CACD;AAED,MAAM,OAAO,mCAAmC;IACtC,EAAE,CAAS;IACX,gBAAgB,GAAG,MAAe,CAAC;IAC3B,OAAO,CAAyD;IAChE,KAAK,CAAS;IACd,KAAK,CAAS;IACd,KAAK,CAAe;IAErC,YACC,EAAU,EACV,OAA+D,EAC/D,KAAa,EACb,KAAa,EACb,OAAO,GAA6B,EAAE,EACrC;QACD,IAAI,CAAC,EAAE,GAAG,EAAE,CAAC;QACb,IAAI,CAAC,OAAO,GAAG,OAAO,CAAC;QACvB,IAAI,CAAC,KAAK,GAAG,KAAK,CAAC;QACnB,IAAI,CAAC,KAAK,GAAG,KAAK,CAAC;QACnB,IAAI,CAAC,KAAK,GAAG,OAAO,CAAC,KAAK,IAAI,IAAI,CAAC,GAAG,CAAC;IAAA,CACvC;IAED,KAAK,CAAC,IAAI,GAAoC;QAC7C,MAAM,MAAM,GAAG,MAAM,IAAI,CAAC,OAAO,CAAC,MAAM,EAAE,CAAC;QAC3C,IAAI,CAAC,MAAM,CAAC,QAAQ;YAAE,OAAO,EAAE,CAAC;QAChC,OAAO;YACN;gBACC,SAAS,EAAE,IAAI,CAAC,EAAE;gBAClB,KAAK,EAAE,IAAI,CAAC,KAAK;gBACjB,KAAK,EAAE,IAAI,CAAC,KAAK;gBACjB,YAAY,EAAE,IAAI,CAAC,KAAK,EAAE;gBAC1B,QAAQ,EAAE,EAAE;gBACZ,gBAAgB,EAAE,IAAI,CAAC,gBAAgB;aACvC;SACD,CAAC;IAAA,CACF;IAED,KAAK,CAAC,cAAc,CAAC,MAAc,EAAiB;QACnD,MAAM,MAAM,GAAG,MAAM,IAAI,CAAC,OAAO,CAAC,KAAK,EAAE,CAAC;QAC1C,IAAI,CAAC,MAAM,CAAC,OAAO,IAAI,MAAM,CAAC,MAAM,KAAK,iBAAiB;YAAE,MAAM,IAAI,KAAK,CAAC,MAAM,CAAC,MAAM,CAAC,CAAC;IAAA,CAC3F;IAED,KAAK,CAAC,OAAO,CAAC,MAAc,EAAiB;QAC5C,IAAI,CAAC,OAAO,CAAC,IAAI,EAAE,CAAC;IAAA,CACpB;CACD;AAED,MAAM,OAAO,+BAA+B;IAClC,EAAE,CAAS;IACX,gBAAgB,GAAG,UAAmB,CAAC;IAC/B,SAAS,CAAyB;IAEnD,YAAY,EAAU,EAAE,SAAyE,EAAE;QAClG,IAAI,CAAC,EAAE,GAAG,EAAE,CAAC;QACb,IAAI,CAAC,SAAS,GAAG,SAAS,CAAC,GAAG,CAAC,CAAC,QAAQ,EAAE,EAAE,CAAC,CAAC;YAC7C,GAAG,QAAQ;YACX,SAAS,EAAE,EAAE;YACb,gBAAgB,EAAE,IAAI,CAAC,gBAAgB;SACvC,CAAC,CAAC,CAAC;IAAA,CACJ;IAED,KAAK,CAAC,IAAI,GAAoC;QAC7C,OAAO,CAAC,GAAG,IAAI,CAAC,SAAS,CAAC,CAAC;IAAA,CAC3B;IAED,KAAK,CAAC,cAAc,CAAC,MAAc,EAAiB,EAAC,CAAC;IAEtD,KAAK,CAAC,OAAO,CAAC,MAAc,EAAiB,EAAC,CAAC;CAC/C;AAED,MAAM,UAAU,oBAAoB,CAAC,IAIpC,EAAwB;IACxB,MAAM,eAAe,GAAG,IAAI,GAAG,EAA+C,CAAC;IAC/E,eAAe,CAAC,GAAG,CAAC,IAAI,CAAC,OAAO,CAAC,KAAK,EAAE,EAAE,KAAK,EAAE,IAAI,CAAC,OAAO,CAAC,KAAK,EAAE,QAAQ,EAAE,IAAI,CAAC,OAAO,CAAC,QAAQ,EAAE,CAAC,CAAC;IACxG,KAAK,MAAM,WAAW,IAAI,IAAI,CAAC,OAAO,CAAC,YAAY,IAAI,EAAE,EAAE,CAAC;QAC3D,MAAM,QAAQ,GAAG,eAAe,CAAC,GAAG,CAAC,WAAW,CAAC,KAAK,CAAC,CAAC;QACxD,IAAI,CAAC,QAAQ,IAAI,WAAW,CAAC,KAAK,GAAG,QAAQ,CAAC,KAAK,EAAE,CAAC;YACrD,eAAe,CAAC,GAAG,CAAC,WAAW,CAAC,KAAK,EAAE,EAAE,KAAK,EAAE,WAAW,CAAC,KAAK,EAAE,QAAQ,EAAE,WAAW,CAAC,QAAQ,EAAE,CAAC,CAAC;QACtG,CAAC;IACF,CAAC;IAED,MAAM,eAAe,GAAG,IAAI,GAAG,CAAC,IAAI,CAAC,SAAS,CAAC,GAAG,CAAC,CAAC,QAAQ,EAAE,EAAE,CAAC,CAAC,QAAQ,CAAC,KAAK,EAAE,QAAQ,CAAC,CAAC,CAAC,CAAC;IAC9F,MAAM,aAAa,GAAG,IAAI,CAAC,SAAS,CAAC,MAAM,CAAC,CAAC,GAAG,EAAE,QAAQ,EAAE,EAAE,CAAC,GAAG,GAAG,QAAQ,CAAC,KAAK,EAAE,CAAC,CAAC,CAAC;IACxF,MAAM,YAAY,GAAG,CAAC,GAAG,eAAe,CAAC,CAAC,MAAM,CAAC,CAAC,GAAG,EAAE,CAAC,KAAK,EAAE,OAAO,CAAC,EAAE,EAAE,CAAC;QAC3E,OAAO,eAAe,CAAC,GAAG,CAAC,KAAK,CAAC,CAAC,CAAC,CAAC,GAAG,CAAC,CAAC,CAAC,GAAG,GAAG,OAAO,CAAC,KAAK,CAAC;IAAA,CAC9D,EAAE,CAAC,CAAC,CAAC;IACN,MAAM,aAAa,GAAG,aAAa,GAAG,YAAY,CAAC;IACnD,IAAI,aAAa,IAAI,IAAI,CAAC,WAAW,EAAE,CAAC;QACvC,OAAO,EAAE,MAAM,EAAE,MAAM,EAAE,KAAK,EAAE,EAAE,EAAE,aAAa,EAAE,WAAW,EAAE,IAAI,CAAC,WAAW,EAAE,CAAC;IACpF,CAAC;IAED,IAAI,eAAe,CAAC,IAAI,CAAC,OAAO,CAAC,EAAE,CAAC;QACnC,OAAO,EAAE,MAAM,EAAE,QAAQ,EAAE,MAAM,EAAE,aAAa,EAAE,KAAK,EAAE,EAAE,EAAE,aAAa,EAAE,WAAW,EAAE,IAAI,CAAC,WAAW,EAAE,CAAC;IAC7G,CAAC;IAED,IAAI,cAAc,GAAG,aAAa,CAAC;IACnC,MAAM,KAAK,GAA2B,EAAE,CAAC;IACzC,MAAM,UAAU,GAAG,IAAI,CAAC,SAAS;SAC/B,MAAM,CAAC,CAAC,QAAQ,EAAE,EAAE,CAAC,CAAC,eAAe,CAAC,GAAG,CAAC,QAAQ,CAAC,KAAK,CAAC,CAAC;SAC1D,MAAM,CAAC,CAAC,QAAQ,EAAE,EAAE,CAAC,CAAC,QAAQ,CAAC,MAAM,IAAI,QAAQ,CAAC,KAAK,KAAK,IAAI,CAAC,OAAO,CAAC,cAAc,CAAC;SACxF,MAAM,CAAC,CAAC,QAAQ,EAAE,EAAE,CAAC,QAAQ,CAAC,gBAAgB,KAAK,UAAU,CAAC;SAC9D,MAAM,CACN,CAAC,QAAQ,EAAE,EAAE,CACZ,IAAI,CAAC,OAAO,CAAC,UAAU,KAAK,SAAS;QACrC,IAAI,CAAC,OAAO,CAAC,KAAK,GAAG,QAAQ,CAAC,YAAY,IAAI,IAAI,CAAC,OAAO,CAAC,UAAU,CACtE;SACA,IAAI,CAAC,CAAC,IAAI,EAAE,KAAK,EAAE,EAAE,CAAC,IAAI,CAAC,QAAQ,GAAG,KAAK,CAAC,QAAQ,IAAI,IAAI,CAAC,YAAY,GAAG,KAAK,CAAC,YAAY,CAAC,CAAC;IAClG,KAAK,MAAM,SAAS,IAAI,UAAU,EAAE,CAAC;QACpC,KAAK,CAAC,IAAI,CAAC,SAAS,CAAC,CAAC;QACtB,cAAc,IAAI,SAAS,CAAC,KAAK,CAAC;QAClC,IAAI,cAAc,IAAI,IAAI,CAAC,WAAW,EAAE,CAAC;YACxC,OAAO,EAAE,MAAM,EAAE,MAAM,EAAE,KAAK,EAAE,aAAa,EAAE,WAAW,EAAE,IAAI,CAAC,WAAW,EAAE,CAAC;QAChF,CAAC;IACF,CAAC;IACD,OAAO;QACN,MAAM,EAAE,QAAQ;QAChB,MAAM,EAAE,+BAA+B;QACvC,KAAK;QACL,aAAa;QACb,WAAW,EAAE,IAAI,CAAC,WAAW;KAC7B,CAAC;AAAA,CACF;AAED,SAAS,eAAe,CAAC,OAA2B,EAAW;IAC9D,OAAO,CAAC,OAAO,CAAC,eAAe,IAAI,EAAE,CAAC,CAAC,IAAI,CAC1C,CAAC,MAAM,EAAE,EAAE,CAAC,MAAM,CAAC,OAAO,KAAK,OAAO,CAAC,KAAK,IAAI,OAAO,CAAC,KAAK,GAAG,MAAM,CAAC,IAAI,IAAI,CAAC,OAAO,CAAC,UAAU,IAAI,CAAC,CAAC,CACxG,CAAC;AAAA,CACF","sourcesContent":["import type { OllamaRuntime, TransformersRuntime } from \"./local-runtime.ts\";\n\nexport type ResidencyControl = \"full\" | \"keep-alive\" | \"advisory\";\nexport type RuntimeRole = \"active\" | \"router\" | \"bench\" | \"probe\";\n\nexport interface RuntimeResidentModel {\n\tadapterId: string;\n\tmodel: string;\n\tbytes: number;\n\tlastUsedAtMs: number;\n\tpriority: number;\n\tpinned?: boolean;\n\tresidencyControl: ResidencyControl;\n}\n\nexport interface RuntimeLoadRequest {\n\tmodel: string;\n\tbytes: number;\n\trole: RuntimeRole;\n\tpriority: number;\n\tnowMs: number;\n\tminDwellMs?: number;\n\tpinActiveModel?: string;\n\treservations?: RuntimeReservation[];\n\trecentEvictions?: RuntimeEvictionRecord[];\n}\n\nexport interface RuntimeReservation {\n\tmodel: string;\n\tbytes: number;\n\tpriority: number;\n}\n\nexport interface RuntimeEvictionRecord {\n\tevicted: string;\n\tloaded: string;\n\tatMs: number;\n}\n\nexport type RuntimeResidencyPlan =\n\t| { status: \"fits\"; evict: RuntimeResidentModel[]; requiredBytes: number; budgetBytes: number }\n\t| { status: \"refuse\"; reason: string; evict: RuntimeResidentModel[]; requiredBytes: number; budgetBytes: number };\n\nexport interface RuntimeResidencyAdapter {\n\tid: string;\n\tresidencyControl: ResidencyControl;\n\tlist(): Promise<RuntimeResidentModel[]>;\n\tensureResident(model: string): Promise<void>;\n\trelease(model: string): Promise<void>;\n}\n\nexport class RuntimeResidencyArbiter {\n\tprivate readonly budgetBytes: number;\n\tprivate readonly adapters: Map<string, RuntimeResidencyAdapter>;\n\n\tconstructor(args: { budgetBytes: number; adapters: RuntimeResidencyAdapter[] }) {\n\t\tthis.budgetBytes = args.budgetBytes;\n\t\tthis.adapters = new Map(args.adapters.map((adapter) => [adapter.id, adapter]));\n\t}\n\n\tasync ensureResident(adapterId: string, request: RuntimeLoadRequest): Promise<RuntimeResidencyPlan> {\n\t\tconst adapter = this.adapters.get(adapterId);\n\t\tconst residents = (await Promise.all([...this.adapters.values()].map((entry) => entry.list()))).flat();\n\t\tif (!adapter) {\n\t\t\treturn {\n\t\t\t\tstatus: \"refuse\",\n\t\t\t\treason: \"unknown-adapter\",\n\t\t\t\tevict: [],\n\t\t\t\trequiredBytes: request.bytes,\n\t\t\t\tbudgetBytes: this.budgetBytes,\n\t\t\t};\n\t\t}\n\t\tconst plan = planRuntimeResidency({ budgetBytes: this.budgetBytes, residents, request });\n\t\tif (plan.status !== \"fits\") return plan;\n\t\tfor (const resident of plan.evict) {\n\t\t\tawait this.adapters.get(resident.adapterId)?.release(resident.model);\n\t\t}\n\t\tawait adapter.ensureResident(request.model);\n\t\treturn plan;\n\t}\n}\n\nexport class OllamaRuntimeResidencyAdapter implements RuntimeResidencyAdapter {\n\treadonly id: string;\n\treadonly residencyControl = \"keep-alive\" as const;\n\tprivate readonly runtime: Pick<OllamaRuntime, \"listResidentModels\" | \"ensureResident\" | \"releaseResident\">;\n\tprivate readonly nowMs: () => number;\n\n\tconstructor(\n\t\tid: string,\n\t\truntime: Pick<OllamaRuntime, \"listResidentModels\" | \"ensureResident\" | \"releaseResident\">,\n\t\toptions: { nowMs?: () => number } = {},\n\t) {\n\t\tthis.id = id;\n\t\tthis.runtime = runtime;\n\t\tthis.nowMs = options.nowMs ?? Date.now;\n\t}\n\n\tasync list(): Promise<RuntimeResidentModel[]> {\n\t\tconst now = this.nowMs();\n\t\treturn (await this.runtime.listResidentModels()).map((model) => ({\n\t\t\tadapterId: this.id,\n\t\t\tmodel: model.name,\n\t\t\tbytes: model.sizeBytes,\n\t\t\tlastUsedAtMs: now,\n\t\t\tpriority: 10,\n\t\t\tresidencyControl: this.residencyControl,\n\t\t}));\n\t}\n\n\tasync ensureResident(model: string): Promise<void> {\n\t\tconst result = await this.runtime.ensureResident(model);\n\t\tif (!result.ok) throw new Error(result.error);\n\t}\n\n\tasync release(model: string): Promise<void> {\n\t\tawait this.runtime.releaseResident(model);\n\t}\n}\n\nexport class TransformersRuntimeResidencyAdapter implements RuntimeResidencyAdapter {\n\treadonly id: string;\n\treadonly residencyControl = \"full\" as const;\n\tprivate readonly runtime: Pick<TransformersRuntime, \"detect\" | \"start\" | \"stop\">;\n\tprivate readonly model: string;\n\tprivate readonly bytes: number;\n\tprivate readonly nowMs: () => number;\n\n\tconstructor(\n\t\tid: string,\n\t\truntime: Pick<TransformersRuntime, \"detect\" | \"start\" | \"stop\">,\n\t\tmodel: string,\n\t\tbytes: number,\n\t\toptions: { nowMs?: () => number } = {},\n\t) {\n\t\tthis.id = id;\n\t\tthis.runtime = runtime;\n\t\tthis.model = model;\n\t\tthis.bytes = bytes;\n\t\tthis.nowMs = options.nowMs ?? Date.now;\n\t}\n\n\tasync list(): Promise<RuntimeResidentModel[]> {\n\t\tconst status = await this.runtime.detect();\n\t\tif (!status.serverUp) return [];\n\t\treturn [\n\t\t\t{\n\t\t\t\tadapterId: this.id,\n\t\t\t\tmodel: this.model,\n\t\t\t\tbytes: this.bytes,\n\t\t\t\tlastUsedAtMs: this.nowMs(),\n\t\t\t\tpriority: 10,\n\t\t\t\tresidencyControl: this.residencyControl,\n\t\t\t},\n\t\t];\n\t}\n\n\tasync ensureResident(_model: string): Promise<void> {\n\t\tconst result = await this.runtime.start();\n\t\tif (!result.started && result.reason !== \"already_running\") throw new Error(result.reason);\n\t}\n\n\tasync release(_model: string): Promise<void> {\n\t\tthis.runtime.stop();\n\t}\n}\n\nexport class AdvisoryRuntimeResidencyAdapter implements RuntimeResidencyAdapter {\n\treadonly id: string;\n\treadonly residencyControl = \"advisory\" as const;\n\tprivate readonly residents: RuntimeResidentModel[];\n\n\tconstructor(id: string, residents: Omit<RuntimeResidentModel, \"adapterId\" | \"residencyControl\">[]) {\n\t\tthis.id = id;\n\t\tthis.residents = residents.map((resident) => ({\n\t\t\t...resident,\n\t\t\tadapterId: id,\n\t\t\tresidencyControl: this.residencyControl,\n\t\t}));\n\t}\n\n\tasync list(): Promise<RuntimeResidentModel[]> {\n\t\treturn [...this.residents];\n\t}\n\n\tasync ensureResident(_model: string): Promise<void> {}\n\n\tasync release(_model: string): Promise<void> {}\n}\n\nexport function planRuntimeResidency(args: {\n\tbudgetBytes: number;\n\tresidents: readonly RuntimeResidentModel[];\n\trequest: RuntimeLoadRequest;\n}): RuntimeResidencyPlan {\n\tconst requestedModels = new Map<string, { bytes: number; priority: number }>();\n\trequestedModels.set(args.request.model, { bytes: args.request.bytes, priority: args.request.priority });\n\tfor (const reservation of args.request.reservations ?? []) {\n\t\tconst existing = requestedModels.get(reservation.model);\n\t\tif (!existing || reservation.bytes > existing.bytes) {\n\t\t\trequestedModels.set(reservation.model, { bytes: reservation.bytes, priority: reservation.priority });\n\t\t}\n\t}\n\n\tconst residentByModel = new Map(args.residents.map((resident) => [resident.model, resident]));\n\tconst residentBytes = args.residents.reduce((sum, resident) => sum + resident.bytes, 0);\n\tconst missingBytes = [...requestedModels].reduce((sum, [model, request]) => {\n\t\treturn residentByModel.has(model) ? sum : sum + request.bytes;\n\t}, 0);\n\tconst requiredBytes = residentBytes + missingBytes;\n\tif (requiredBytes <= args.budgetBytes) {\n\t\treturn { status: \"fits\", evict: [], requiredBytes, budgetBytes: args.budgetBytes };\n\t}\n\n\tif (hasPingPongRisk(args.request)) {\n\t\treturn { status: \"refuse\", reason: \"anti-thrash\", evict: [], requiredBytes, budgetBytes: args.budgetBytes };\n\t}\n\n\tlet projectedBytes = requiredBytes;\n\tconst evict: RuntimeResidentModel[] = [];\n\tconst candidates = args.residents\n\t\t.filter((resident) => !requestedModels.has(resident.model))\n\t\t.filter((resident) => !resident.pinned && resident.model !== args.request.pinActiveModel)\n\t\t.filter((resident) => resident.residencyControl !== \"advisory\")\n\t\t.filter(\n\t\t\t(resident) =>\n\t\t\t\targs.request.minDwellMs === undefined ||\n\t\t\t\targs.request.nowMs - resident.lastUsedAtMs >= args.request.minDwellMs,\n\t\t)\n\t\t.sort((left, right) => left.priority - right.priority || left.lastUsedAtMs - right.lastUsedAtMs);\n\tfor (const candidate of candidates) {\n\t\tevict.push(candidate);\n\t\tprojectedBytes -= candidate.bytes;\n\t\tif (projectedBytes <= args.budgetBytes) {\n\t\t\treturn { status: \"fits\", evict, requiredBytes, budgetBytes: args.budgetBytes };\n\t\t}\n\t}\n\treturn {\n\t\tstatus: \"refuse\",\n\t\treason: \"insufficient-evictable-memory\",\n\t\tevict,\n\t\trequiredBytes,\n\t\tbudgetBytes: args.budgetBytes,\n\t};\n}\n\nfunction hasPingPongRisk(request: RuntimeLoadRequest): boolean {\n\treturn (request.recentEvictions ?? []).some(\n\t\t(record) => record.evicted === request.model && request.nowMs - record.atMs <= (request.minDwellMs ?? 0),\n\t);\n}\n"]}

package/dist/modes/interactive/local-model-commands.d.ts.map

@@ -1 +1 @@
-{"version":3,"file":"local-model-commands.d.ts","sourceRoot":"","sources":["../../../src/modes/interactive/local-model-commands.ts"],"names":[],"mappings":"AAAA;;;;;;;;;;;GAWG;AAKH,OAAO,EACN,KAAK,SAAS,EACd,SAAS,EAMT,KAAK,GAAG,EACR,MAAM,oBAAoB,CAAC;AAE5B,OAAO,KAAK,EAAE,YAAY,EAAE,MAAM,6BAA6B,CAAC;AAChE,OAAO,KAAK,EAAE,aAAa,EAAE,MAAM,8BAA8B,CAAC;AAiBlE,OAAO,KAAK,EAAE,aAAa,EAAE,mBAAmB,EAAE,MAAM,oCAAoC,CAAC;AAG7F,OAAO,KAAK,EAAE,eAAe,EAAE,MAAM,gCAAgC,CAAC;AAEtE,OAAO,EAAE,KAAK,WAAW,EAAgC,MAAM,uCAAuC,CAAC;AAKvG,KAAK,eAAe,GAAG,CAAC,IAAI,EAAE,MAAM,IAAI,KAAK;IAAE,SAAS,EAAE,SAAS,CAAC;IAAC,KAAK,EAAE,SAAS,CAAA;CAAE,CAAC;AAoCxF,0FAAwF;AACxF,MAAM,WAAW,cAAc;IAC9B,QAAQ,CAAC,eAAe,EAAE,eAAe,CAAC;IAC1C,UAAU,CAAC,OAAO,EAAE,MAAM,GAAG,IAAI,CAAC;CAClC;AAED,+FAA6F;AAC7F,MAAM,WAAW,cAAc;IAC9B,QAAQ,CAAC,OAAO,EAAE,IAAI,CAAC,YAAY,EAAE,iBAAiB,CAAC,GAAG;QAAE,aAAa,CAAC,EAAE,aAAa,CAAA;KAAE,CAAC;IAC5F,QAAQ,CAAC,eAAe,EAAE,eAAe,CAAC;IAC1C,QAAQ,CAAC,aAAa,EAAE,SAAS,CAAC;IAClC,QAAQ,CAAC,EAAE,EAAE,GAAG,CAAC;IACjB,UAAU,CAAC,OAAO,EAAE,MAAM,GAAG,IAAI,CAAC;IAClC,SAAS,CAAC,OAAO,EAAE,MAAM,GAAG,IAAI,CAAC;IACjC,YAAY,CAAC,MAAM,EAAE,eAAe,GAAG,IAAI,CAAC;CAC5C;AAED,yDAAyD;AACzD,MAAM,WAAW,cAAc;IAC9B,QAAQ,CAAC,YAAY,EAAE,aAAa,CAAC;IACrC,QAAQ,CAAC,OAAO,EAAE,YAAY,CAAC;IAC/B,QAAQ,CAAC,eAAe,EAAE,eAAe,CAAC;IAC1C,QAAQ,CAAC,EAAE,EAAE,GAAG,CAAC;IACjB,QAAQ,CAAC,aAAa,EAAE,SAAS,CAAC;IAClC,sBAAsB,CAAC,OAAO,EAAE,MAAM,EAAE,OAAO,CAAC,EAAE,MAAM,GAAG,mBAAmB,CAAC;IAC/E,UAAU,CAAC,OAAO,EAAE,MAAM,GAAG,IAAI,CAAC;IAClC,SAAS,CAAC,OAAO,EAAE,MAAM,GAAG,IAAI,CAAC;IACjC,YAAY,CAAC,MAAM,EAAE,eAAe,GAAG,IAAI,CAAC;CAC5C;AAmDD;;;;GAIG;AACH,wBAAsB,mBAAmB,CAAC,IAAI,EAAE,cAAc,EAAE,QAAQ,EAAE,MAAM,GAAG,OAAO,CAAC,IAAI,CAAC,CAyE/F;AAED,wBAAsB,iBAAiB,CAAC,IAAI,EAAE,cAAc,GAAG,OAAO,CAAC,OAAO,CAAC,CAgB9E;AAED,wBAAsB,eAAe,CAAC,IAAI,EAAE,cAAc,GAAG,OAAO,CAAC,IAAI,CAAC,CAqDzE;AAED,wBAAsB,aAAa,CAAC,IAAI,EAAE,cAAc,EAAE,OAAO,EAAE,MAAM,EAAE,aAAa,CAAC,EAAE,WAAW,GAAG,OAAO,CAAC,IAAI,CAAC,CAwDrH;AAED,wBAAsB,oBAAoB,CACzC,IAAI,EAAE,cAAc,EACpB,OAAO,EAAE,MAAM,EACf,aAAa,CAAC,EAAE,WAAW,GACzB,OAAO,CAAC,IAAI,CAAC,CA8Cf;AAED,wBAAsB,gBAAgB,CAAC,IAAI,EAAE,cAAc,EAAE,GAAG,EAAE,MAAM,EAAE,SAAS,EAAE,OAAO,GAAG,OAAO,CAAC,IAAI,CAAC,CAqC3G;AA8BD,iGAAiG;AACjG,kGAA4F;AAC5F,wBAAgB,2BAA2B,CAAC,IAAI,EAAE,cAAc,GAAG,IAAI,CA4BtE;AAED,wBAAgB,wBAAwB,CAAC,IAAI,EAAE,cAAc,GAAG,IAAI,CAmBnE;AAoED;kGACkG;AAClG,wBAAsB,mBAAmB,CACxC,IAAI,EAAE,cAAc,EACpB,QAAQ,EAAE,MAAM,EAChB,aAAa,CAAC,EAAE,WAAW,GACzB,OAAO,CAAC,IAAI,CAAC,CAkDf;AAED,yFAAyF;AACzF,wBAAgB,iBAAiB,CAAC,IAAI,EAAE,cAAc,EAAE,QAAQ,EAAE,MAAM,EAAE,IAAI,EAAE,WAAW,GAAG,IAAI,CAqCjG","sourcesContent":["/**\n * `/models` local-model lifecycle and `/fitness` probe/role-assignment flows\n * extracted from interactive-mode.\n *\n * `/models` is a USER-invoked local-model manager (list/add/remove/stop, never a\n * model-invokable tool); the fitness helpers probe a model on the current host\n * and land a shaped role. The server/selector flows share a `LocalModelHost`\n * seam, while `runFitnessAndAssign`/`assignFitnessRole` take narrow host shapes\n * matching their prototype-driven behaviour tests (fitness-probe-gate,\n * fitness-role-assignment), which keep exercising interactive-mode's thin\n * wrappers unchanged.\n */\n\nimport { totalmem } from \"node:os\";\nimport type { ThinkingLevel } from \"@caupulican/pi-agent-core\";\nimport { getSupportedThinkingLevels } from \"@caupulican/pi-ai\";\nimport {\n\ttype Component,\n\tContainer,\n\ttype SelectItem,\n\tSelectList,\n\ttype SelectListLayoutOptions,\n\tSpacer,\n\tText,\n\ttype TUI,\n} from \"@caupulican/pi-tui\";\nimport { getAgentDir } from \"../../config.ts\";\nimport type { AgentSession } from \"../../core/agent-session.ts\";\nimport type { ModelRegistry } from \"../../core/model-registry.ts\";\nimport { resolveCliModel } from \"../../core/model-resolver.ts\";\nimport { evaluateSurfaceFitness } from \"../../core/model-router/fitness-gate.ts\";\nimport {\n\tderiveLocalContextSizing,\n\trenderOllamaContextModelfile,\n\tsizedLocalModelRef,\n} from \"../../core/models/context-sizing.ts\";\nimport { DEFAULT_MODEL_SUGGESTIONS } from \"../../core/models/default-model-suggestions.ts\";\nimport { FitnessStore } from \"../../core/models/fitness-store.ts\";\nimport {\n\tHF_TRANSFORMERS_PROVIDER,\n\tregisterLocalModel,\n\tregisterTransformersModel,\n\tunregisterLocalModel,\n\tunregisterTransformersModel,\n} from \"../../core/models/local-registration.ts\";\nimport type { OllamaRuntime, TransformersRuntime } from \"../../core/models/local-runtime.ts\";\nimport { matchesInstalledLocalModel, normalizeModelSource } from \"../../core/models/model-ref.ts\";\nimport { formatModelFitnessReport, isProbeAllFailed } from \"../../core/research/model-fitness.ts\";\nimport type { SettingsManager } from \"../../core/settings-manager.ts\";\nimport { DynamicBorder } from \"./components/dynamic-border.ts\";\nimport { type FitnessRole, FitnessRoleSelectorComponent } from \"./components/fitness-role-selector.ts\";\nimport { ModelSelectorComponent } from \"./components/model-selector.ts\";\nimport { ModelSuggestionSelectorComponent } from \"./components/model-suggestion-selector.ts\";\nimport { getSelectListTheme } from \"./theme/theme.ts\";\n\ntype SelectorFactory = (done: () => void) => { component: Component; focus: Component };\n\nconst MODEL_ROUTER_THINKING_SELECT_LIST_LAYOUT: SelectListLayoutOptions = {\n\tminPrimaryColumnWidth: 12,\n\tmaxPrimaryColumnWidth: 32,\n};\n\nconst MODEL_ROUTER_THINKING_INHERIT_VALUE = \"__inherit_model_router_thinking__\";\nconst MODEL_ROUTER_THINKING_INHERIT_LABEL = \"(inherit)\";\n\nconst THINKING_LEVEL_DESCRIPTIONS: Record<ThinkingLevel, string> = {\n\toff: \"No reasoning\",\n\tminimal: \"Very brief reasoning (~1k tokens)\",\n\tlow: \"Light reasoning (~2k tokens)\",\n\tmedium: \"Moderate reasoning (~8k tokens)\",\n\thigh: \"Deep reasoning (~16k tokens)\",\n\txhigh: \"Maximum reasoning (~32k tokens)\",\n};\n\nconst MODEL_ROUTER_THINKING_FALLBACK_LEVELS: readonly ThinkingLevel[] = [\n\t\"off\",\n\t\"minimal\",\n\t\"low\",\n\t\"medium\",\n\t\"high\",\n\t\"xhigh\",\n];\n\ntype RouterThinkingLevel = ThinkingLevel | typeof MODEL_ROUTER_THINKING_INHERIT_VALUE;\ntype ModelRouterThinkingField =\n\t| \"cheapThinking\"\n\t| \"mediumThinking\"\n\t| \"expensiveThinking\"\n\t| \"executorThinking\"\n\t| \"judgeThinking\";\n\n/** Narrow seam for persisting a probed model's role — matches the fitness-role test. */\nexport interface AssignRoleHost {\n\treadonly settingsManager: SettingsManager;\n\tshowStatus(message: string): void;\n}\n\n/** Seam for the fitness probe + role-selector flow — matches the fitness-probe-gate test. */\nexport interface RunFitnessHost {\n\treadonly session: Pick<AgentSession, \"runModelFitness\"> & { modelRegistry?: ModelRegistry };\n\treadonly settingsManager: SettingsManager;\n\treadonly chatContainer: Container;\n\treadonly ui: TUI;\n\tshowStatus(message: string): void;\n\tshowError(message: string): void;\n\tshowSelector(create: SelectorFactory): void;\n}\n\n/** Full seam for the `/models` server/selector flows. */\nexport interface LocalModelHost {\n\treadonly localRuntime: OllamaRuntime;\n\treadonly session: AgentSession;\n\treadonly settingsManager: SettingsManager;\n\treadonly ui: TUI;\n\treadonly chatContainer: Container;\n\tgetTransformersRuntime(modelId: string, baseUrl?: string): TransformersRuntime;\n\tshowStatus(message: string): void;\n\tshowError(message: string): void;\n\tshowSelector(create: SelectorFactory): void;\n}\n\nclass ModelRouterThinkingSelectorComponent extends Container {\n\tprivate selectList: SelectList;\n\n\tconstructor(\n\t\tcurrentLevel: RouterThinkingLevel,\n\t\tavailableLevels: ThinkingLevel[],\n\t\tonSelect: (level: RouterThinkingLevel) => void,\n\t\tonCancel: () => void,\n\t) {\n\t\tsuper();\n\n\t\tconst thinkingOptions: SelectItem[] = [\n\t\t\t{\n\t\t\t\tvalue: MODEL_ROUTER_THINKING_INHERIT_VALUE,\n\t\t\t\tlabel: MODEL_ROUTER_THINKING_INHERIT_LABEL,\n\t\t\t\tdescription: \"Use current session default for this tier.\",\n\t\t\t},\n\t\t\t...availableLevels.map((level) => ({\n\t\t\t\tvalue: level,\n\t\t\t\tlabel: level,\n\t\t\t\tdescription: THINKING_LEVEL_DESCRIPTIONS[level],\n\t\t\t})),\n\t\t];\n\n\t\tthis.addChild(new DynamicBorder());\n\n\t\tthis.selectList = new SelectList(\n\t\t\tthinkingOptions,\n\t\t\tthinkingOptions.length,\n\t\t\tgetSelectListTheme(),\n\t\t\tMODEL_ROUTER_THINKING_SELECT_LIST_LAYOUT,\n\t\t);\n\n\t\tconst index = thinkingOptions.findIndex((item) => item.value === currentLevel);\n\t\tif (index !== -1) {\n\t\t\tthis.selectList.setSelectedIndex(index);\n\t\t}\n\n\t\tthis.selectList.onSelect = (item) => {\n\t\t\tonSelect(item.value as RouterThinkingLevel);\n\t\t};\n\t\tthis.selectList.onCancel = () => {\n\t\t\tonCancel();\n\t\t};\n\n\t\tthis.addChild(this.selectList);\n\t\tthis.addChild(new DynamicBorder());\n\t}\n}\n/**\n * /models — USER-invoked local model lifecycle (never a model-invokable tool):\n * list/add/remove/stop per local-model-lifecycle-design.md. Removal is explicit-only with\n * full disclosure; a pasted install command is parsed for its ref, never executed.\n */\nexport async function handleModelsCommand(host: LocalModelHost, argsText: string): Promise<void> {\n\tconst [action = \"list\", ...rest] = argsText.split(/\\s+/).filter(Boolean);\n\ttry {\n\t\tif (action === \"suggest\" || action === \"suggestions\") {\n\t\t\tshowModelSuggestionSelector(host);\n\t\t\treturn;\n\t\t}\n\t\tif (action === \"stop\") {\n\t\t\tconst stopped = host.localRuntime.stop();\n\t\t\tlet stoppedTransformers = 0;\n\t\t\tfor (const model of host.session.modelRegistry\n\t\t\t\t.getAll()\n\t\t\t\t.filter((entry) => entry.provider === HF_TRANSFORMERS_PROVIDER)) {\n\t\t\t\tconst serverUrl = model.baseUrl.replace(/\\/v1\\/?$/, \"\");\n\t\t\t\tif (host.getTransformersRuntime(model.id, serverUrl).stop().stopped) stoppedTransformers++;\n\t\t\t}\n\t\t\tconst stoppedAny = stopped.stopped || stoppedTransformers > 0;\n\t\t\thost.showStatus(\n\t\t\t\tstoppedAny\n\t\t\t\t\t? `Pi-managed local model server stopped${stoppedTransformers > 0 ? ` (${stoppedTransformers} Transformers sidecar(s))` : \"\"}; models remain installed.`\n\t\t\t\t\t: \"No pi-managed server running (a system server, if any, is not pi's to stop).\",\n\t\t\t);\n\t\t\treturn;\n\t\t}\n\n\t\tif (action === \"add\") {\n\t\t\tconst rawRef = rest.join(\" \");\n\t\t\tif (!rawRef) {\n\t\t\t\thost.showStatus(\n\t\t\t\t\t\"Usage: /models add <ollama-tag | hf.co/org/repo[:quant] | curated HF full-base ref | huggingface URL | pasted install command>\",\n\t\t\t\t);\n\t\t\t\thost.showStatus(\"Or start from a validated suggestion: /models suggest\");\n\t\t\t\treturn;\n\t\t\t}\n\t\t\tconst source = normalizeModelSource(rawRef);\n\t\t\tif (source.type === \"rejected\") {\n\t\t\t\thost.showStatus(`Not added: ${source.reason}`);\n\t\t\t\treturn;\n\t\t\t}\n\t\t\tif (source.type === \"api\") {\n\t\t\t\thost.showStatus(\n\t\t\t\t\t`${source.ref} is an API model — nothing to install. Configure auth for the provider, then probe it with /fitness ${source.ref}.`,\n\t\t\t\t);\n\t\t\t\treturn;\n\t\t\t}\n\t\t\tif (source.type === \"transformers\") {\n\t\t\t\tawait addTransformersModel(host, source.modelId);\n\t\t\t\treturn;\n\t\t\t}\n\t\t\tawait addLocalModel(host, source.pullRef);\n\t\t\treturn;\n\t\t}\n\n\t\tif (action === \"remove\") {\n\t\t\tconst ref = rest[0];\n\t\t\tconst confirmed = rest[1] === \"confirm\";\n\t\t\tif (!ref) {\n\t\t\t\thost.showStatus(\"Usage: /models remove <ref> confirm\");\n\t\t\t\treturn;\n\t\t\t}\n\t\t\tconst source = normalizeModelSource(ref);\n\t\t\tif (source.type === \"transformers\") {\n\t\t\t\tawait removeTransformersModel(host, source.modelId, confirmed);\n\t\t\t\treturn;\n\t\t\t}\n\t\t\tawait removeLocalModel(host, ref, confirmed);\n\t\t\treturn;\n\t\t}\n\n\t\tawait listLocalModels(host);\n\t} catch (error) {\n\t\thost.showError(error instanceof Error ? error.message : String(error));\n\t}\n}\n\nexport async function ensureLocalServer(host: LocalModelHost): Promise<boolean> {\n\tconst status = await host.localRuntime.detect();\n\tif (status.serverUp) return true;\n\tif (!status.binaryPath) {\n\t\tfor (const line of host.localRuntime.installGuide()) host.showStatus(line);\n\t\treturn false;\n\t}\n\thost.showStatus(\n\t\t`Starting local model server (${status.binarySource} binary, owned storage: ${status.ownedModelsDir})…`,\n\t);\n\tconst started = await host.localRuntime.start();\n\tif (!started.started) {\n\t\thost.showStatus(`Could not start the local server: ${started.reason}`);\n\t\treturn false;\n\t}\n\treturn true;\n}\n\nexport async function listLocalModels(host: LocalModelHost): Promise<void> {\n\tconst status = await host.localRuntime.detect();\n\tconst models = status.serverUp ? await host.localRuntime.list() : [];\n\tconst ollamaLines = status.serverUp\n\t\t? [\n\t\t\t\t`Ollama models (${status.managedByPi ? `pi-managed server, storage: ${status.ownedModelsDir}` : \"system server — storage owned by the system daemon\"}):`,\n\t\t\t\t...(models.length === 0\n\t\t\t\t\t? [\"  (none installed — /models add <ref>, or /models suggest for a validated roster)\"]\n\t\t\t\t\t: []),\n\t\t\t]\n\t\t: [\n\t\t\t\t\"Ollama models:\",\n\t\t\t\t...(status.binaryPath\n\t\t\t\t\t? [\n\t\t\t\t\t\t\t`  server not running (binary: ${status.binarySource}). /models add starts it on demand; /fitness probes registered models.`,\n\t\t\t\t\t\t]\n\t\t\t\t\t: host.localRuntime.installGuide().map((line) => `  ${line}`)),\n\t\t\t];\n\tconst fitness = FitnessStore.forAgentDir(getAgentDir()).getForHost();\n\tconst transformersModels = host.session.modelRegistry\n\t\t.getAll()\n\t\t.filter((model) => model.provider === HF_TRANSFORMERS_PROVIDER);\n\tconst transformersLines = await Promise.all(\n\t\ttransformersModels.map(async (model) => {\n\t\t\tconst serverUrl = model.baseUrl.replace(/\\/v1\\/?$/, \"\");\n\t\t\tconst runtime = host.getTransformersRuntime(model.id, serverUrl);\n\t\t\tconst runtimeStatus = await runtime.detect();\n\t\t\tconst report = fitness.find((entry) => entry.model === `${HF_TRANSFORMERS_PROVIDER}/${model.id}`);\n\t\t\tconst readiness = runtimeStatus.serverUp\n\t\t\t\t? \"server running\"\n\t\t\t\t: runtimeStatus.runtimeInstalled\n\t\t\t\t\t? \"runtime installed\"\n\t\t\t\t\t: \"runtime missing\";\n\t\t\tconst probe = report\n\t\t\t\t? `probed ${report.at.slice(0, 10)}: digest ${report.report.digest?.succeeded ?? \"?\"}/${report.report.digest?.total ?? \"?\"}, tool-calls ${report.report.toolCall.succeeded}/${report.report.toolCall.total}${report.report.tokensPerSecond ? `, ~${report.report.tokensPerSecond} tok/s` : \"\"}`\n\t\t\t\t: `unprobed — run /fitness ${HF_TRANSFORMERS_PROVIDER}/${model.id}`;\n\t\t\treturn `  - ${model.id} (${readiness}, cache: ${runtimeStatus.cacheDir}) · ${probe}`;\n\t\t}),\n\t);\n\tconst lines = [\n\t\t...ollamaLines,\n\t\t...models.map((model) => {\n\t\t\tconst report = fitness.find((entry) => entry.model === `ollama/${model.name}`);\n\t\t\tconst gb = (model.sizeBytes / 1e9).toFixed(2);\n\t\t\tconst probe = report\n\t\t\t\t? `probed ${report.at.slice(0, 10)}: digest ${report.report.digest?.succeeded ?? \"?\"}/${report.report.digest?.total ?? \"?\"}, tool-calls ${report.report.toolCall.succeeded}/${report.report.toolCall.total}${report.report.tokensPerSecond ? `, ~${report.report.tokensPerSecond} tok/s` : \"\"}`\n\t\t\t\t: `unprobed — run /fitness ollama/${model.name}`;\n\t\t\treturn `  - ${model.name} (${gb} GB) · ${probe}`;\n\t\t}),\n\t\t...(transformersLines.length > 0 ? [\"Pi-managed Transformers models:\", ...transformersLines] : []),\n\t\t\"Commands: /models add <ref> · /models remove <ref> confirm · /models stop\",\n\t];\n\tfor (const line of lines) host.showStatus(line);\n}\n\nexport async function addLocalModel(host: LocalModelHost, pullRef: string, preselectRole?: FitnessRole): Promise<void> {\n\tif (!(await ensureLocalServer(host))) return;\n\thost.showStatus(`Pulling ${pullRef}… (weights land in the server's model storage)`);\n\tlet lastShown = 0;\n\tconst pulled = await host.localRuntime.pull(pullRef, (progress) => {\n\t\tconst now = Date.now();\n\t\tif (now - lastShown > 2000) {\n\t\t\tlastShown = now;\n\t\t\thost.showStatus(`  ${pullRef}: ${progress}`);\n\t\t}\n\t});\n\tif (!pulled.ok) {\n\t\thost.showStatus(`Pull failed: ${pulled.error}`);\n\t\treturn;\n\t}\n\tconst installed = await host.localRuntime.list().catch(() => []);\n\tconst installedEntry = installed.find((entry) => matchesInstalledLocalModel(pullRef, entry.name));\n\tconst shown = await host.localRuntime.show(pullRef);\n\tconst sizing =\n\t\tinstalledEntry && shown.ok\n\t\t\t? deriveLocalContextSizing({\n\t\t\t\t\thost: { totalMemBytes: totalmem() },\n\t\t\t\t\tmodel: { modelInfo: shown.info.modelInfo, weightsBytes: installedEntry.sizeBytes },\n\t\t\t\t\truntime: { supportsKvQuantization: true },\n\t\t\t\t})\n\t\t\t: undefined;\n\tlet registeredRef = pullRef;\n\tif (sizing) {\n\t\tconst sizedRef = sizedLocalModelRef(pullRef, sizing.numCtx);\n\t\tconst created = await host.localRuntime.createFromModelfile({\n\t\t\tname: sizedRef,\n\t\t\tmodelfile: renderOllamaContextModelfile({ from: pullRef, numCtx: sizing.numCtx }),\n\t\t});\n\t\tif (!created.ok) {\n\t\t\thost.showStatus(`Pulled, but could not create sized context model: ${created.error}`);\n\t\t\treturn;\n\t\t}\n\t\tregisteredRef = sizedRef;\n\t}\n\tconst registration = registerLocalModel({\n\t\tagentDir: getAgentDir(),\n\t\tref: registeredRef,\n\t\tbaseUrl: host.localRuntime.baseUrl,\n\t\tcontextWindow: sizing?.numCtx,\n\t\tservedContextWindow: sizing?.numCtx,\n\t});\n\tif (!registration.ok) {\n\t\thost.showStatus(`Pulled, but not auto-registered: ${registration.reason}`);\n\t\tif (registration.manualSnippet) {\n\t\t\thost.showStatus(`Add this to ${registration.modelsJsonPath} yourself:\\n${registration.manualSnippet}`);\n\t\t}\n\t\treturn;\n\t}\n\thost.session.modelRegistry.refresh();\n\thost.showStatus(`${pullRef} installed and registered as ollama/${registeredRef}. Probing fitness…`);\n\tawait runFitnessAndAssign(host, `ollama/${registeredRef}`, preselectRole);\n}\n\nexport async function addTransformersModel(\n\thost: LocalModelHost,\n\tmodelId: string,\n\tpreselectRole?: FitnessRole,\n): Promise<void> {\n\tconst runtime = host.getTransformersRuntime(modelId);\n\tlet status = await runtime.detect();\n\tif (!status.runtimeInstalled) {\n\t\thost.showStatus(`Installing isolated Transformers runtime for ${modelId} at ${status.venvDir}…`);\n\t\tconst installed = await runtime.installManaged((progress) => host.showStatus(`  transformers: ${progress}`));\n\t\tif (!installed.ok) {\n\t\t\thost.showStatus(`Transformers runtime install failed: ${installed.error}`);\n\t\t\treturn;\n\t\t}\n\t\tstatus = await runtime.detect();\n\t}\n\tif (!status.runtimeInstalled) {\n\t\thost.showStatus(`Transformers runtime is still unavailable at ${status.venvDir}.`);\n\t\treturn;\n\t}\n\n\tconst downloaded = await runtime.downloadModel((progress) => host.showStatus(`  ${modelId}: ${progress}`));\n\tif (!downloaded.ok) {\n\t\thost.showStatus(`Model download failed: ${downloaded.error}`);\n\t\treturn;\n\t}\n\n\tconst started = await runtime.start();\n\tif (!started.started && started.reason !== \"already_running\") {\n\t\thost.showStatus(`Could not start Transformers sidecar: ${started.reason}`);\n\t\treturn;\n\t}\n\n\tconst registration = registerTransformersModel({\n\t\tagentDir: getAgentDir(),\n\t\tmodelId,\n\t\tbaseUrl: runtime.baseUrl,\n\t});\n\tif (!registration.ok) {\n\t\thost.showStatus(`Installed, but not auto-registered: ${registration.reason}`);\n\t\tif (registration.manualSnippet) {\n\t\t\thost.showStatus(`Add this to ${registration.modelsJsonPath} yourself:\\n${registration.manualSnippet}`);\n\t\t}\n\t\treturn;\n\t}\n\thost.session.modelRegistry.refresh();\n\thost.showStatus(\n\t\t`${modelId} installed in pi-managed Transformers and registered as ${HF_TRANSFORMERS_PROVIDER}/${modelId}. Probing fitness…`,\n\t);\n\tawait runFitnessAndAssign(host, `${HF_TRANSFORMERS_PROVIDER}/${modelId}`, preselectRole);\n}\n\nexport async function removeLocalModel(host: LocalModelHost, ref: string, confirmed: boolean): Promise<void> {\n\tconst status = await host.localRuntime.detect();\n\tif (!status.serverUp) {\n\t\thost.showStatus(\"Local server not running — start it (any /models action) before removing.\");\n\t\treturn;\n\t}\n\tconst models = await host.localRuntime.list();\n\tconst target = models.find((model) => matchesInstalledLocalModel(ref, model.name));\n\tif (!target) {\n\t\thost.showStatus(\n\t\t\t`${ref} is not installed. Installed: ${models.map((model) => model.name).join(\", \") || \"(none)\"}`,\n\t\t);\n\t\treturn;\n\t}\n\tif (!confirmed) {\n\t\t// EXPLICIT USER ACTION ONLY: full disclosure, then require the confirm token.\n\t\tconst gb = (target.sizeBytes / 1e9).toFixed(2);\n\t\thost.showStatus(\n\t\t\t[\n\t\t\t\t`Removing ${ref} will delete:`,\n\t\t\t\t`  - model weights (${gb} GB) from ${status.managedByPi ? status.ownedModelsDir : \"the system server's storage\"}`,\n\t\t\t\t`  - the ollama/${ref} entry in models.json`,\n\t\t\t\t`  - its cached fitness report for this host`,\n\t\t\t\t`Run: /models remove ${ref} confirm`,\n\t\t\t].join(\"\\n\"),\n\t\t);\n\t\treturn;\n\t}\n\tconst removed = await host.localRuntime.remove(ref);\n\tif (!removed.ok) {\n\t\thost.showStatus(`Remove failed: ${removed.error}`);\n\t\treturn;\n\t}\n\tunregisterLocalModel({ agentDir: getAgentDir(), ref });\n\tFitnessStore.forAgentDir(getAgentDir()).remove(`ollama/${ref}`);\n\thost.session.modelRegistry.refresh();\n\thost.showStatus(`${ref} removed: weights deleted, registration and fitness report dropped.`);\n}\n\nasync function removeTransformersModel(host: LocalModelHost, modelId: string, confirmed: boolean): Promise<void> {\n\tconst runtime = host.getTransformersRuntime(modelId);\n\tconst status = await runtime.detect();\n\tif (!confirmed) {\n\t\thost.showStatus(\n\t\t\t[\n\t\t\t\t`Removing ${modelId} will delete:`,\n\t\t\t\t`  - the ${HF_TRANSFORMERS_PROVIDER}/${modelId} entry in models.json`,\n\t\t\t\t`  - its cached fitness report for this host`,\n\t\t\t\t`It will stop this session's Transformers sidecar if running. Cached weights remain under ${status.cacheDir}.`,\n\t\t\t\t`Run: /models remove hf.co/${modelId} confirm`,\n\t\t\t].join(\"\\n\"),\n\t\t);\n\t\treturn;\n\t}\n\truntime.stop();\n\tconst registration = unregisterTransformersModel({ agentDir: getAgentDir(), modelId });\n\tif (!registration.ok) {\n\t\thost.showStatus(`Remove failed: ${registration.reason}`);\n\t\treturn;\n\t}\n\tFitnessStore.forAgentDir(getAgentDir()).remove(`${HF_TRANSFORMERS_PROVIDER}/${modelId}`);\n\thost.session.modelRegistry.refresh();\n\thost.showStatus(\n\t\t`${modelId} registration and fitness report dropped; cached weights remain under ${status.cacheDir}.`,\n\t);\n}\n\n/** /fitness with no args: pick a model from the configured registry, probe it, assign a role. */\n/** Pick a validated suggestion → install it → probe on this host → land its shaped role. */\nexport function showModelSuggestionSelector(host: LocalModelHost): void {\n\thost.showSelector((done) => {\n\t\tconst selector = new ModelSuggestionSelectorComponent(\n\t\t\tDEFAULT_MODEL_SUGGESTIONS,\n\t\t\tasync (suggestion) => {\n\t\t\t\tdone();\n\t\t\t\t// The shaped role rides along so the post-probe selector lands on it pre-selected;\n\t\t\t\t// non-tool-callers carry curator/judge/none, never executor, so this can't footgun.\n\t\t\t\tconst source = normalizeModelSource(suggestion.pullRef);\n\t\t\t\tif (source.type === \"transformers\") {\n\t\t\t\t\tawait addTransformersModel(host, source.modelId, suggestion.assignRole);\n\t\t\t\t\treturn;\n\t\t\t\t}\n\t\t\t\tif (source.type === \"local\") {\n\t\t\t\t\tawait addLocalModel(host, source.pullRef, suggestion.assignRole);\n\t\t\t\t\treturn;\n\t\t\t\t}\n\t\t\t\thost.showStatus(\n\t\t\t\t\t`Suggestion ${suggestion.name} is not installable: ${source.type === \"rejected\" ? source.reason : source.ref}`,\n\t\t\t\t);\n\t\t\t},\n\t\t\t() => {\n\t\t\t\tdone();\n\t\t\t\thost.ui.requestRender();\n\t\t\t},\n\t\t);\n\t\treturn { component: selector, focus: selector };\n\t});\n}\n\nexport function showFitnessModelSelector(host: LocalModelHost): void {\n\thost.showSelector((done) => {\n\t\tconst selector = new ModelSelectorComponent(\n\t\t\thost.ui,\n\t\t\thost.session.model,\n\t\t\thost.settingsManager,\n\t\t\thost.session.modelRegistry,\n\t\t\thost.session.scopedModels,\n\t\t\tasync (model) => {\n\t\t\t\tdone();\n\t\t\t\tawait runFitnessAndAssign(host, `${model.provider}/${model.id}`);\n\t\t\t},\n\t\t\t() => {\n\t\t\t\tdone();\n\t\t\t\thost.ui.requestRender();\n\t\t\t},\n\t\t);\n\t\treturn { component: selector, focus: selector };\n\t});\n}\n\n/** Resolve the routed model's supported thinking levels for a fitness-assignment flow, with a safe fallback. */\nfunction getModelThinkingLevels(modelRegistry: ModelRegistry | undefined, modelRef: string): ThinkingLevel[] {\n\tif (!modelRegistry) {\n\t\treturn [...MODEL_ROUTER_THINKING_FALLBACK_LEVELS];\n\t}\n\tconst resolved = resolveCliModel({ cliModel: modelRef, modelRegistry });\n\tif (!resolved.model) {\n\t\treturn [...MODEL_ROUTER_THINKING_FALLBACK_LEVELS];\n\t}\n\treturn getSupportedThinkingLevels(resolved.model);\n}\n\nfunction getModelRouterThinkingField(role: FitnessRole): ModelRouterThinkingField | undefined {\n\tif (role === \"router-cheap\") {\n\t\treturn \"cheapThinking\";\n\t}\n\tif (role === \"router-medium\") {\n\t\treturn \"mediumThinking\";\n\t}\n\tif (role === \"router-expensive\") {\n\t\treturn \"expensiveThinking\";\n\t}\n\tif (role === \"executor\") {\n\t\treturn \"executorThinking\";\n\t}\n\tif (role === \"judge\") {\n\t\treturn \"judgeThinking\";\n\t}\n\treturn undefined;\n}\n\nfunction promptForModelRouterThinking(host: RunFitnessHost, modelRef: string, role: FitnessRole): void {\n\tconst thinkingField = getModelRouterThinkingField(role);\n\tif (!thinkingField) {\n\t\treturn;\n\t}\n\n\tconst modelRegistry = host.session.modelRegistry;\n\tconst availableLevels = getModelThinkingLevels(modelRegistry, modelRef);\n\tconst settings = host.settingsManager.getModelRouterSettings();\n\tconst configuredThinking = settings[thinkingField];\n\tconst currentThinking =\n\t\tconfiguredThinking && availableLevels.includes(configuredThinking)\n\t\t\t? configuredThinking\n\t\t\t: MODEL_ROUTER_THINKING_INHERIT_VALUE;\n\n\thost.showSelector((done) => {\n\t\tconst selector = new ModelRouterThinkingSelectorComponent(\n\t\t\tcurrentThinking,\n\t\t\tavailableLevels,\n\t\t\t(level) => {\n\t\t\t\tdone();\n\t\t\t\thost.settingsManager.setModelRouterSettings({\n\t\t\t\t\t...settings,\n\t\t\t\t\t[thinkingField]: level === MODEL_ROUTER_THINKING_INHERIT_VALUE ? undefined : level,\n\t\t\t\t});\n\t\t\t},\n\t\t\t() => {\n\t\t\t\tdone();\n\t\t\t\thost.ui.requestRender();\n\t\t\t},\n\t\t);\n\t\treturn { component: selector, focus: selector };\n\t});\n}\n\n/** Probe a model's fitness, show the report, then offer one-step role assignment. When the model\n * came from a validated suggestion, `preselectRole` lands its shaped role already highlighted. */\nexport async function runFitnessAndAssign(\n\thost: RunFitnessHost,\n\tmodelRef: string,\n\tpreselectRole?: FitnessRole,\n): Promise<void> {\n\thost.showStatus(`Model fitness probe running on ${modelRef}… (6 surfaces; local models may take a few minutes)`);\n\ttry {\n\t\tconst outcome = await host.session.runModelFitness({ model: modelRef });\n\t\tif (!outcome.started) {\n\t\t\thost.showStatus(`Model fitness skipped: ${outcome.skipReason}`);\n\t\t\treturn;\n\t\t}\n\t\thost.chatContainer.addChild(new Spacer(1));\n\t\thost.chatContainer.addChild(new Text(formatModelFitnessReport(outcome.model, outcome.report), 1, 0));\n\t\thost.ui.requestRender();\n\t\t// Validate-before-load: zero successes on every probed surface means the model cannot\n\t\t// drive any of the harness's subagent contracts on this host — refuse adoption instead\n\t\t// of landing a role selector the user might reflexively confirm (this is the reported\n\t\t// bug: a 0/3-everywhere model still got set as judge model and saved to Model Router).\n\t\tif (isProbeAllFailed(outcome.report)) {\n\t\t\thost.showStatus(\n\t\t\t\t`${outcome.model} failed the fitness probe on all surfaces — not configured. Use /model to set it manually if you accept the risk.`,\n\t\t\t);\n\t\t\treturn;\n\t\t}\n\t\thost.showSelector((done) => {\n\t\t\tconst selector = new FitnessRoleSelectorComponent(\n\t\t\t\toutcome.model,\n\t\t\t\t(role) => {\n\t\t\t\t\tdone();\n\t\t\t\t\tif (role === \"scout\") {\n\t\t\t\t\t\tconst verdict = evaluateSurfaceFitness(\"scout_auto\", outcome.report);\n\t\t\t\t\t\tif (!verdict.fit) {\n\t\t\t\t\t\t\tconst reason = verdict.reason === \"lane_failed\" ? `failed ${verdict.lane}` : \"was not probed\";\n\t\t\t\t\t\t\thost.showStatus(\n\t\t\t\t\t\t\t\t`${outcome.model} not assigned as scout: ${reason} on the scout_auto fitness exam. Use the docs/scout.md Modelfile recipe, then rerun /fitness.`,\n\t\t\t\t\t\t\t);\n\t\t\t\t\t\t\treturn;\n\t\t\t\t\t\t}\n\t\t\t\t\t}\n\t\t\t\t\tassignFitnessRole(host, outcome.model, role);\n\t\t\t\t\tpromptForModelRouterThinking(host, outcome.model, role);\n\t\t\t\t},\n\t\t\t\t() => {\n\t\t\t\t\tdone();\n\t\t\t\t\thost.ui.requestRender();\n\t\t\t\t},\n\t\t\t\tpreselectRole,\n\t\t\t);\n\t\t\treturn { component: selector, focus: selector };\n\t\t});\n\t} catch (error) {\n\t\thost.showError(error instanceof Error ? error.message : String(error));\n\t}\n}\n\n/** Persist a role assignment from the post-probe selector into the matching settings. */\nexport function assignFitnessRole(host: AssignRoleHost, modelRef: string, role: FitnessRole): void {\n\tif (role === \"none\") {\n\t\thost.showStatus(`Fitness result for ${modelRef} saved. Assign a role later from /settings.`);\n\t\treturn;\n\t}\n\tif (role === \"curator\") {\n\t\tconst current = host.settingsManager.getContextCurationSettings();\n\t\thost.settingsManager.setContextCurationSettings({ ...current, enabled: true, model: modelRef });\n\t\thost.showStatus(`Context curation enabled with ${modelRef} as the curator.`);\n\t\treturn;\n\t}\n\tif (role === \"executor\") {\n\t\tconst router = host.settingsManager.getModelRouterSettings();\n\t\thost.settingsManager.setModelRouterSettings({ ...router, executorModel: modelRef });\n\t\tconst hint = router.enabled ? \"\" : \" Model router is currently disabled — enable it in /settings → Model Router.\";\n\t\thost.showStatus(`${modelRef} set as the toolkit executor (direct Level-0 hits route to it).${hint}`);\n\t\treturn;\n\t}\n\tif (role === \"scout\") {\n\t\thost.settingsManager.setScoutSettings({ enabled: true, model: modelRef });\n\t\thost.showStatus(`${modelRef} set as the repository scout (context_scout enabled).`);\n\t\treturn;\n\t}\n\tconst router = host.settingsManager.getModelRouterSettings();\n\tconst field =\n\t\trole === \"router-cheap\"\n\t\t\t? \"cheapModel\"\n\t\t\t: role === \"router-medium\"\n\t\t\t\t? \"mediumModel\"\n\t\t\t\t: role === \"router-expensive\"\n\t\t\t\t\t? \"expensiveModel\"\n\t\t\t\t\t: role === \"judge\"\n\t\t\t\t\t\t? \"judgeModel\"\n\t\t\t\t\t\t: \"learningModel\";\n\thost.settingsManager.setModelRouterSettings({ ...router, [field]: modelRef });\n\tconst hint = router.enabled ? \"\" : \" Model router is currently disabled — enable it in /settings → Model Router.\";\n\thost.showStatus(`${modelRef} set as ${role.replace(\"router-\", \"router \")} model.${hint}`);\n}\n"]}
+{"version":3,"file":"local-model-commands.d.ts","sourceRoot":"","sources":["../../../src/modes/interactive/local-model-commands.ts"],"names":[],"mappings":"AAAA;;;;;;;;;;;GAWG;AAKH,OAAO,EACN,KAAK,SAAS,EACd,SAAS,EAMT,KAAK,GAAG,EACR,MAAM,oBAAoB,CAAC;AAE5B,OAAO,KAAK,EAAE,YAAY,EAAE,MAAM,6BAA6B,CAAC;AAChE,OAAO,KAAK,EAAE,aAAa,EAAE,MAAM,8BAA8B,CAAC;AAiBlE,OAAO,KAAK,EAAE,aAAa,EAAE,mBAAmB,EAAE,MAAM,oCAAoC,CAAC;AAG7F,OAAO,KAAK,EAAE,eAAe,EAAE,MAAM,gCAAgC,CAAC;AAEtE,OAAO,EAAE,KAAK,WAAW,EAAgC,MAAM,uCAAuC,CAAC;AAKvG,KAAK,eAAe,GAAG,CAAC,IAAI,EAAE,MAAM,IAAI,KAAK;IAAE,SAAS,EAAE,SAAS,CAAC;IAAC,KAAK,EAAE,SAAS,CAAA;CAAE,CAAC;AAoCxF,0FAAwF;AACxF,MAAM,WAAW,cAAc;IAC9B,QAAQ,CAAC,eAAe,EAAE,eAAe,CAAC;IAC1C,UAAU,CAAC,OAAO,EAAE,MAAM,GAAG,IAAI,CAAC;CAClC;AAED,+FAA6F;AAC7F,MAAM,WAAW,cAAc;IAC9B,QAAQ,CAAC,OAAO,EAAE,IAAI,CAAC,YAAY,EAAE,iBAAiB,CAAC,GAAG;QAAE,aAAa,CAAC,EAAE,aAAa,CAAA;KAAE,CAAC;IAC5F,QAAQ,CAAC,eAAe,EAAE,eAAe,CAAC;IAC1C,QAAQ,CAAC,aAAa,EAAE,SAAS,CAAC;IAClC,QAAQ,CAAC,EAAE,EAAE,GAAG,CAAC;IACjB,UAAU,CAAC,OAAO,EAAE,MAAM,GAAG,IAAI,CAAC;IAClC,SAAS,CAAC,OAAO,EAAE,MAAM,GAAG,IAAI,CAAC;IACjC,YAAY,CAAC,MAAM,EAAE,eAAe,GAAG,IAAI,CAAC;CAC5C;AAED,yDAAyD;AACzD,MAAM,WAAW,cAAc;IAC9B,QAAQ,CAAC,YAAY,EAAE,aAAa,CAAC;IACrC,QAAQ,CAAC,OAAO,EAAE,YAAY,CAAC;IAC/B,QAAQ,CAAC,eAAe,EAAE,eAAe,CAAC;IAC1C,QAAQ,CAAC,EAAE,EAAE,GAAG,CAAC;IACjB,QAAQ,CAAC,aAAa,EAAE,SAAS,CAAC;IAClC,sBAAsB,CAAC,OAAO,EAAE,MAAM,EAAE,OAAO,CAAC,EAAE,MAAM,GAAG,mBAAmB,CAAC;IAC/E,UAAU,CAAC,OAAO,EAAE,MAAM,GAAG,IAAI,CAAC;IAClC,SAAS,CAAC,OAAO,EAAE,MAAM,GAAG,IAAI,CAAC;IACjC,YAAY,CAAC,MAAM,EAAE,eAAe,GAAG,IAAI,CAAC;CAC5C;AAmDD;;;;GAIG;AACH,wBAAsB,mBAAmB,CAAC,IAAI,EAAE,cAAc,EAAE,QAAQ,EAAE,MAAM,GAAG,OAAO,CAAC,IAAI,CAAC,CAmF/F;AAOD,wBAAsB,iBAAiB,CAAC,IAAI,EAAE,cAAc,GAAG,OAAO,CAAC,OAAO,CAAC,CAwB9E;AAED,wBAAsB,eAAe,CAAC,IAAI,EAAE,cAAc,GAAG,OAAO,CAAC,IAAI,CAAC,CAqDzE;AAED,wBAAsB,aAAa,CAAC,IAAI,EAAE,cAAc,EAAE,OAAO,EAAE,MAAM,EAAE,aAAa,CAAC,EAAE,WAAW,GAAG,OAAO,CAAC,IAAI,CAAC,CAwDrH;AAED,wBAAsB,oBAAoB,CACzC,IAAI,EAAE,cAAc,EACpB,OAAO,EAAE,MAAM,EACf,aAAa,CAAC,EAAE,WAAW,GACzB,OAAO,CAAC,IAAI,CAAC,CA8Cf;AAED,wBAAsB,gBAAgB,CAAC,IAAI,EAAE,cAAc,EAAE,GAAG,EAAE,MAAM,EAAE,SAAS,EAAE,OAAO,GAAG,OAAO,CAAC,IAAI,CAAC,CAqC3G;AA8BD,iGAAiG;AACjG,kGAA4F;AAC5F,wBAAgB,2BAA2B,CAAC,IAAI,EAAE,cAAc,GAAG,IAAI,CA4BtE;AAED,wBAAgB,wBAAwB,CAAC,IAAI,EAAE,cAAc,GAAG,IAAI,CAmBnE;AAoED;kGACkG;AAClG,wBAAsB,mBAAmB,CACxC,IAAI,EAAE,cAAc,EACpB,QAAQ,EAAE,MAAM,EAChB,aAAa,CAAC,EAAE,WAAW,GACzB,OAAO,CAAC,IAAI,CAAC,CAkDf;AAED,yFAAyF;AACzF,wBAAgB,iBAAiB,CAAC,IAAI,EAAE,cAAc,EAAE,QAAQ,EAAE,MAAM,EAAE,IAAI,EAAE,WAAW,GAAG,IAAI,CAqCjG","sourcesContent":["/**\n * `/models` local-model lifecycle and `/fitness` probe/role-assignment flows\n * extracted from interactive-mode.\n *\n * `/models` is a USER-invoked local-model manager (list/add/remove/stop, never a\n * model-invokable tool); the fitness helpers probe a model on the current host\n * and land a shaped role. The server/selector flows share a `LocalModelHost`\n * seam, while `runFitnessAndAssign`/`assignFitnessRole` take narrow host shapes\n * matching their prototype-driven behaviour tests (fitness-probe-gate,\n * fitness-role-assignment), which keep exercising interactive-mode's thin\n * wrappers unchanged.\n */\n\nimport { totalmem } from \"node:os\";\nimport type { ThinkingLevel } from \"@caupulican/pi-agent-core\";\nimport { getSupportedThinkingLevels } from \"@caupulican/pi-ai\";\nimport {\n\ttype Component,\n\tContainer,\n\ttype SelectItem,\n\tSelectList,\n\ttype SelectListLayoutOptions,\n\tSpacer,\n\tText,\n\ttype TUI,\n} from \"@caupulican/pi-tui\";\nimport { getAgentDir } from \"../../config.ts\";\nimport type { AgentSession } from \"../../core/agent-session.ts\";\nimport type { ModelRegistry } from \"../../core/model-registry.ts\";\nimport { resolveCliModel } from \"../../core/model-resolver.ts\";\nimport { evaluateSurfaceFitness } from \"../../core/model-router/fitness-gate.ts\";\nimport {\n\tderiveLocalContextSizing,\n\trenderOllamaContextModelfile,\n\tsizedLocalModelRef,\n} from \"../../core/models/context-sizing.ts\";\nimport { DEFAULT_MODEL_SUGGESTIONS } from \"../../core/models/default-model-suggestions.ts\";\nimport { FitnessStore } from \"../../core/models/fitness-store.ts\";\nimport {\n\tHF_TRANSFORMERS_PROVIDER,\n\tregisterLocalModel,\n\tregisterTransformersModel,\n\tunregisterLocalModel,\n\tunregisterTransformersModel,\n} from \"../../core/models/local-registration.ts\";\nimport type { OllamaRuntime, TransformersRuntime } from \"../../core/models/local-runtime.ts\";\nimport { matchesInstalledLocalModel, normalizeModelSource } from \"../../core/models/model-ref.ts\";\nimport { formatModelFitnessReport, isProbeAllFailed } from \"../../core/research/model-fitness.ts\";\nimport type { SettingsManager } from \"../../core/settings-manager.ts\";\nimport { DynamicBorder } from \"./components/dynamic-border.ts\";\nimport { type FitnessRole, FitnessRoleSelectorComponent } from \"./components/fitness-role-selector.ts\";\nimport { ModelSelectorComponent } from \"./components/model-selector.ts\";\nimport { ModelSuggestionSelectorComponent } from \"./components/model-suggestion-selector.ts\";\nimport { getSelectListTheme } from \"./theme/theme.ts\";\n\ntype SelectorFactory = (done: () => void) => { component: Component; focus: Component };\n\nconst MODEL_ROUTER_THINKING_SELECT_LIST_LAYOUT: SelectListLayoutOptions = {\n\tminPrimaryColumnWidth: 12,\n\tmaxPrimaryColumnWidth: 32,\n};\n\nconst MODEL_ROUTER_THINKING_INHERIT_VALUE = \"__inherit_model_router_thinking__\";\nconst MODEL_ROUTER_THINKING_INHERIT_LABEL = \"(inherit)\";\n\nconst THINKING_LEVEL_DESCRIPTIONS: Record<ThinkingLevel, string> = {\n\toff: \"No reasoning\",\n\tminimal: \"Very brief reasoning (~1k tokens)\",\n\tlow: \"Light reasoning (~2k tokens)\",\n\tmedium: \"Moderate reasoning (~8k tokens)\",\n\thigh: \"Deep reasoning (~16k tokens)\",\n\txhigh: \"Maximum reasoning (~32k tokens)\",\n};\n\nconst MODEL_ROUTER_THINKING_FALLBACK_LEVELS: readonly ThinkingLevel[] = [\n\t\"off\",\n\t\"minimal\",\n\t\"low\",\n\t\"medium\",\n\t\"high\",\n\t\"xhigh\",\n];\n\ntype RouterThinkingLevel = ThinkingLevel | typeof MODEL_ROUTER_THINKING_INHERIT_VALUE;\ntype ModelRouterThinkingField =\n\t| \"cheapThinking\"\n\t| \"mediumThinking\"\n\t| \"expensiveThinking\"\n\t| \"executorThinking\"\n\t| \"judgeThinking\";\n\n/** Narrow seam for persisting a probed model's role — matches the fitness-role test. */\nexport interface AssignRoleHost {\n\treadonly settingsManager: SettingsManager;\n\tshowStatus(message: string): void;\n}\n\n/** Seam for the fitness probe + role-selector flow — matches the fitness-probe-gate test. */\nexport interface RunFitnessHost {\n\treadonly session: Pick<AgentSession, \"runModelFitness\"> & { modelRegistry?: ModelRegistry };\n\treadonly settingsManager: SettingsManager;\n\treadonly chatContainer: Container;\n\treadonly ui: TUI;\n\tshowStatus(message: string): void;\n\tshowError(message: string): void;\n\tshowSelector(create: SelectorFactory): void;\n}\n\n/** Full seam for the `/models` server/selector flows. */\nexport interface LocalModelHost {\n\treadonly localRuntime: OllamaRuntime;\n\treadonly session: AgentSession;\n\treadonly settingsManager: SettingsManager;\n\treadonly ui: TUI;\n\treadonly chatContainer: Container;\n\tgetTransformersRuntime(modelId: string, baseUrl?: string): TransformersRuntime;\n\tshowStatus(message: string): void;\n\tshowError(message: string): void;\n\tshowSelector(create: SelectorFactory): void;\n}\n\nclass ModelRouterThinkingSelectorComponent extends Container {\n\tprivate selectList: SelectList;\n\n\tconstructor(\n\t\tcurrentLevel: RouterThinkingLevel,\n\t\tavailableLevels: ThinkingLevel[],\n\t\tonSelect: (level: RouterThinkingLevel) => void,\n\t\tonCancel: () => void,\n\t) {\n\t\tsuper();\n\n\t\tconst thinkingOptions: SelectItem[] = [\n\t\t\t{\n\t\t\t\tvalue: MODEL_ROUTER_THINKING_INHERIT_VALUE,\n\t\t\t\tlabel: MODEL_ROUTER_THINKING_INHERIT_LABEL,\n\t\t\t\tdescription: \"Use current session default for this tier.\",\n\t\t\t},\n\t\t\t...availableLevels.map((level) => ({\n\t\t\t\tvalue: level,\n\t\t\t\tlabel: level,\n\t\t\t\tdescription: THINKING_LEVEL_DESCRIPTIONS[level],\n\t\t\t})),\n\t\t];\n\n\t\tthis.addChild(new DynamicBorder());\n\n\t\tthis.selectList = new SelectList(\n\t\t\tthinkingOptions,\n\t\t\tthinkingOptions.length,\n\t\t\tgetSelectListTheme(),\n\t\t\tMODEL_ROUTER_THINKING_SELECT_LIST_LAYOUT,\n\t\t);\n\n\t\tconst index = thinkingOptions.findIndex((item) => item.value === currentLevel);\n\t\tif (index !== -1) {\n\t\t\tthis.selectList.setSelectedIndex(index);\n\t\t}\n\n\t\tthis.selectList.onSelect = (item) => {\n\t\t\tonSelect(item.value as RouterThinkingLevel);\n\t\t};\n\t\tthis.selectList.onCancel = () => {\n\t\t\tonCancel();\n\t\t};\n\n\t\tthis.addChild(this.selectList);\n\t\tthis.addChild(new DynamicBorder());\n\t}\n}\n/**\n * /models — USER-invoked local model lifecycle (never a model-invokable tool):\n * list/add/remove/stop per local-model-lifecycle-design.md. Removal is explicit-only with\n * full disclosure; a pasted install command is parsed for its ref, never executed.\n */\nexport async function handleModelsCommand(host: LocalModelHost, argsText: string): Promise<void> {\n\tconst [action = \"list\", ...rest] = argsText.split(/\\s+/).filter(Boolean);\n\ttry {\n\t\tif (action === \"suggest\" || action === \"suggestions\") {\n\t\t\tshowModelSuggestionSelector(host);\n\t\t\treturn;\n\t\t}\n\t\tif (action === \"stop\") {\n\t\t\tconst stopped = host.localRuntime.stop();\n\t\t\tlet stoppedTransformers = 0;\n\t\t\tfor (const model of host.session.modelRegistry\n\t\t\t\t.getAll()\n\t\t\t\t.filter((entry) => entry.provider === HF_TRANSFORMERS_PROVIDER)) {\n\t\t\t\tconst serverUrl = model.baseUrl.replace(/\\/v1\\/?$/, \"\");\n\t\t\t\tif (host.getTransformersRuntime(model.id, serverUrl).stop().stopped) stoppedTransformers++;\n\t\t\t}\n\t\t\tconst stoppedAny = stopped.stopped || stoppedTransformers > 0;\n\t\t\thost.showStatus(\n\t\t\t\tstoppedAny\n\t\t\t\t\t? `Pi-managed local model server stopped${stoppedTransformers > 0 ? ` (${stoppedTransformers} Transformers sidecar(s))` : \"\"}; models remain installed.`\n\t\t\t\t\t: \"No pi-managed server running (a system server, if any, is not pi's to stop).\",\n\t\t\t);\n\t\t\treturn;\n\t\t}\n\n\t\tif (action === \"import\") {\n\t\t\tconst imported = host.localRuntime.importUserModels();\n\t\t\thost.showStatus(\n\t\t\t\t`Imported Ollama models from ${imported.sourceDir} into pi-owned store ${imported.targetDir}: ` +\n\t\t\t\t\t`${imported.manifestsImported} manifest(s), ${imported.blobsHardlinked} blob(s) hardlinked, ` +\n\t\t\t\t\t`${imported.blobsCopied} blob(s) copied, ${imported.blobsSkipped + imported.manifestsSkipped} existing file(s) skipped.`,\n\t\t\t);\n\t\t\treturn;\n\t\t}\n\n\t\tif (action === \"add\") {\n\t\t\tconst rawRef = rest.join(\" \");\n\t\t\tif (!rawRef) {\n\t\t\t\thost.showStatus(\n\t\t\t\t\t\"Usage: /models add <ollama-tag | hf.co/org/repo[:quant] | curated HF full-base ref | huggingface URL | pasted install command>\",\n\t\t\t\t);\n\t\t\t\thost.showStatus(\"Or start from a validated suggestion: /models suggest\");\n\t\t\t\treturn;\n\t\t\t}\n\t\t\tconst source = normalizeModelSource(rawRef);\n\t\t\tif (source.type === \"rejected\") {\n\t\t\t\thost.showStatus(`Not added: ${source.reason}`);\n\t\t\t\treturn;\n\t\t\t}\n\t\t\tif (source.type === \"api\") {\n\t\t\t\thost.showStatus(\n\t\t\t\t\t`${source.ref} is an API model — nothing to install. Configure auth for the provider, then probe it with /fitness ${source.ref}.`,\n\t\t\t\t);\n\t\t\t\treturn;\n\t\t\t}\n\t\t\tif (source.type === \"transformers\") {\n\t\t\t\tawait addTransformersModel(host, source.modelId);\n\t\t\t\treturn;\n\t\t\t}\n\t\t\tawait addLocalModel(host, source.pullRef);\n\t\t\treturn;\n\t\t}\n\n\t\tif (action === \"remove\") {\n\t\t\tconst ref = rest[0];\n\t\t\tconst confirmed = rest[1] === \"confirm\";\n\t\t\tif (!ref) {\n\t\t\t\thost.showStatus(\"Usage: /models remove <ref> confirm\");\n\t\t\t\treturn;\n\t\t\t}\n\t\t\tconst source = normalizeModelSource(ref);\n\t\t\tif (source.type === \"transformers\") {\n\t\t\t\tawait removeTransformersModel(host, source.modelId, confirmed);\n\t\t\t\treturn;\n\t\t\t}\n\t\t\tawait removeLocalModel(host, ref, confirmed);\n\t\t\treturn;\n\t\t}\n\n\t\tawait listLocalModels(host);\n\t} catch (error) {\n\t\thost.showError(error instanceof Error ? error.message : String(error));\n\t}\n}\n\nfunction formatOllamaStore(store: { kind: string; path: string; modelCount: number } | undefined): string {\n\tif (!store) return \"unknown store (0 model(s))\";\n\treturn `${store.path} [${store.kind}, ${store.modelCount} model(s)]`;\n}\n\nexport async function ensureLocalServer(host: LocalModelHost): Promise<boolean> {\n\tconst status = await host.localRuntime.detect();\n\tif (status.serverUp) {\n\t\tif (status.activeStore?.kind === \"pi-owned\") return true;\n\t\thost.showStatus(\n\t\t\t`Ollama is already running on ${status.serverUrl} with store ${formatOllamaStore(status.activeStore)}; ` +\n\t\t\t\t`pi's canonical store is ${status.ownedModelsDir} (${status.ownedStore.modelCount} model(s)). ` +\n\t\t\t\t\"Stop the other serve or run /models import before adding models.\",\n\t\t);\n\t\treturn false;\n\t}\n\tif (!status.binaryPath) {\n\t\tfor (const line of host.localRuntime.installGuide()) host.showStatus(line);\n\t\treturn false;\n\t}\n\thost.showStatus(\n\t\t`Starting local model server (${status.binarySource} binary, owned storage: ${status.ownedModelsDir})…`,\n\t);\n\tconst started = await host.localRuntime.start();\n\tif (!started.started) {\n\t\thost.showStatus(`Could not start the local server: ${started.reason}`);\n\t\treturn false;\n\t}\n\treturn true;\n}\n\nexport async function listLocalModels(host: LocalModelHost): Promise<void> {\n\tconst status = await host.localRuntime.detect();\n\tconst models = status.serverUp ? await host.localRuntime.list() : [];\n\tconst ollamaLines = status.serverUp\n\t\t? [\n\t\t\t\t`Ollama models (active store: ${formatOllamaStore(status.activeStore)}; pi-owned store: ${status.ownedModelsDir} with ${status.ownedStore.modelCount} model(s)):`,\n\t\t\t\t...(models.length === 0\n\t\t\t\t\t? [\"  (none installed — /models add <ref>, /models import, or /models suggest for a validated roster)\"]\n\t\t\t\t\t: []),\n\t\t\t]\n\t\t: [\n\t\t\t\t\"Ollama models:\",\n\t\t\t\t...(status.binaryPath\n\t\t\t\t\t? [\n\t\t\t\t\t\t\t`  server not running (binary: ${status.binarySource}). /models add starts it on demand; /fitness probes registered models.`,\n\t\t\t\t\t\t]\n\t\t\t\t\t: host.localRuntime.installGuide().map((line) => `  ${line}`)),\n\t\t\t];\n\tconst fitness = FitnessStore.forAgentDir(getAgentDir()).getForHost();\n\tconst transformersModels = host.session.modelRegistry\n\t\t.getAll()\n\t\t.filter((model) => model.provider === HF_TRANSFORMERS_PROVIDER);\n\tconst transformersLines = await Promise.all(\n\t\ttransformersModels.map(async (model) => {\n\t\t\tconst serverUrl = model.baseUrl.replace(/\\/v1\\/?$/, \"\");\n\t\t\tconst runtime = host.getTransformersRuntime(model.id, serverUrl);\n\t\t\tconst runtimeStatus = await runtime.detect();\n\t\t\tconst report = fitness.find((entry) => entry.model === `${HF_TRANSFORMERS_PROVIDER}/${model.id}`);\n\t\t\tconst readiness = runtimeStatus.serverUp\n\t\t\t\t? \"server running\"\n\t\t\t\t: runtimeStatus.runtimeInstalled\n\t\t\t\t\t? \"runtime installed\"\n\t\t\t\t\t: \"runtime missing\";\n\t\t\tconst probe = report\n\t\t\t\t? `probed ${report.at.slice(0, 10)}: digest ${report.report.digest?.succeeded ?? \"?\"}/${report.report.digest?.total ?? \"?\"}, tool-calls ${report.report.toolCall.succeeded}/${report.report.toolCall.total}${report.report.tokensPerSecond ? `, ~${report.report.tokensPerSecond} tok/s` : \"\"}`\n\t\t\t\t: `unprobed — run /fitness ${HF_TRANSFORMERS_PROVIDER}/${model.id}`;\n\t\t\treturn `  - ${model.id} (${readiness}, cache: ${runtimeStatus.cacheDir}) · ${probe}`;\n\t\t}),\n\t);\n\tconst lines = [\n\t\t...ollamaLines,\n\t\t...models.map((model) => {\n\t\t\tconst report = fitness.find((entry) => entry.model === `ollama/${model.name}`);\n\t\t\tconst gb = (model.sizeBytes / 1e9).toFixed(2);\n\t\t\tconst probe = report\n\t\t\t\t? `probed ${report.at.slice(0, 10)}: digest ${report.report.digest?.succeeded ?? \"?\"}/${report.report.digest?.total ?? \"?\"}, tool-calls ${report.report.toolCall.succeeded}/${report.report.toolCall.total}${report.report.tokensPerSecond ? `, ~${report.report.tokensPerSecond} tok/s` : \"\"}`\n\t\t\t\t: `unprobed — run /fitness ollama/${model.name}`;\n\t\t\treturn `  - ${model.name} (${gb} GB) · ${probe}`;\n\t\t}),\n\t\t...(transformersLines.length > 0 ? [\"Pi-managed Transformers models:\", ...transformersLines] : []),\n\t\t\"Commands: /models import · /models add <ref> · /models remove <ref> confirm · /models stop\",\n\t];\n\tfor (const line of lines) host.showStatus(line);\n}\n\nexport async function addLocalModel(host: LocalModelHost, pullRef: string, preselectRole?: FitnessRole): Promise<void> {\n\tif (!(await ensureLocalServer(host))) return;\n\thost.showStatus(`Pulling ${pullRef}… (weights land in pi's owned Ollama store)`);\n\tlet lastShown = 0;\n\tconst pulled = await host.localRuntime.pull(pullRef, (progress) => {\n\t\tconst now = Date.now();\n\t\tif (now - lastShown > 2000) {\n\t\t\tlastShown = now;\n\t\t\thost.showStatus(`  ${pullRef}: ${progress}`);\n\t\t}\n\t});\n\tif (!pulled.ok) {\n\t\thost.showStatus(`Pull failed: ${pulled.error}`);\n\t\treturn;\n\t}\n\tconst installed = await host.localRuntime.list().catch(() => []);\n\tconst installedEntry = installed.find((entry) => matchesInstalledLocalModel(pullRef, entry.name));\n\tconst shown = await host.localRuntime.show(pullRef);\n\tconst sizing =\n\t\tinstalledEntry && shown.ok\n\t\t\t? deriveLocalContextSizing({\n\t\t\t\t\thost: { totalMemBytes: totalmem() },\n\t\t\t\t\tmodel: { modelInfo: shown.info.modelInfo, weightsBytes: installedEntry.sizeBytes },\n\t\t\t\t\truntime: { supportsKvQuantization: true },\n\t\t\t\t})\n\t\t\t: undefined;\n\tlet registeredRef = pullRef;\n\tif (sizing) {\n\t\tconst sizedRef = sizedLocalModelRef(pullRef, sizing.numCtx);\n\t\tconst created = await host.localRuntime.createFromModelfile({\n\t\t\tname: sizedRef,\n\t\t\tmodelfile: renderOllamaContextModelfile({ from: pullRef, numCtx: sizing.numCtx }),\n\t\t});\n\t\tif (!created.ok) {\n\t\t\thost.showStatus(`Pulled, but could not create sized context model: ${created.error}`);\n\t\t\treturn;\n\t\t}\n\t\tregisteredRef = sizedRef;\n\t}\n\tconst registration = registerLocalModel({\n\t\tagentDir: getAgentDir(),\n\t\tref: registeredRef,\n\t\tbaseUrl: host.localRuntime.baseUrl,\n\t\tcontextWindow: sizing?.numCtx,\n\t\tservedContextWindow: sizing?.numCtx,\n\t});\n\tif (!registration.ok) {\n\t\thost.showStatus(`Pulled, but not auto-registered: ${registration.reason}`);\n\t\tif (registration.manualSnippet) {\n\t\t\thost.showStatus(`Add this to ${registration.modelsJsonPath} yourself:\\n${registration.manualSnippet}`);\n\t\t}\n\t\treturn;\n\t}\n\thost.session.modelRegistry.refresh();\n\thost.showStatus(`${pullRef} installed and registered as ollama/${registeredRef}. Probing fitness…`);\n\tawait runFitnessAndAssign(host, `ollama/${registeredRef}`, preselectRole);\n}\n\nexport async function addTransformersModel(\n\thost: LocalModelHost,\n\tmodelId: string,\n\tpreselectRole?: FitnessRole,\n): Promise<void> {\n\tconst runtime = host.getTransformersRuntime(modelId);\n\tlet status = await runtime.detect();\n\tif (!status.runtimeInstalled) {\n\t\thost.showStatus(`Installing isolated Transformers runtime for ${modelId} at ${status.venvDir}…`);\n\t\tconst installed = await runtime.installManaged((progress) => host.showStatus(`  transformers: ${progress}`));\n\t\tif (!installed.ok) {\n\t\t\thost.showStatus(`Transformers runtime install failed: ${installed.error}`);\n\t\t\treturn;\n\t\t}\n\t\tstatus = await runtime.detect();\n\t}\n\tif (!status.runtimeInstalled) {\n\t\thost.showStatus(`Transformers runtime is still unavailable at ${status.venvDir}.`);\n\t\treturn;\n\t}\n\n\tconst downloaded = await runtime.downloadModel((progress) => host.showStatus(`  ${modelId}: ${progress}`));\n\tif (!downloaded.ok) {\n\t\thost.showStatus(`Model download failed: ${downloaded.error}`);\n\t\treturn;\n\t}\n\n\tconst started = await runtime.start();\n\tif (!started.started && started.reason !== \"already_running\") {\n\t\thost.showStatus(`Could not start Transformers sidecar: ${started.reason}`);\n\t\treturn;\n\t}\n\n\tconst registration = registerTransformersModel({\n\t\tagentDir: getAgentDir(),\n\t\tmodelId,\n\t\tbaseUrl: runtime.baseUrl,\n\t});\n\tif (!registration.ok) {\n\t\thost.showStatus(`Installed, but not auto-registered: ${registration.reason}`);\n\t\tif (registration.manualSnippet) {\n\t\t\thost.showStatus(`Add this to ${registration.modelsJsonPath} yourself:\\n${registration.manualSnippet}`);\n\t\t}\n\t\treturn;\n\t}\n\thost.session.modelRegistry.refresh();\n\thost.showStatus(\n\t\t`${modelId} installed in pi-managed Transformers and registered as ${HF_TRANSFORMERS_PROVIDER}/${modelId}. Probing fitness…`,\n\t);\n\tawait runFitnessAndAssign(host, `${HF_TRANSFORMERS_PROVIDER}/${modelId}`, preselectRole);\n}\n\nexport async function removeLocalModel(host: LocalModelHost, ref: string, confirmed: boolean): Promise<void> {\n\tconst status = await host.localRuntime.detect();\n\tif (!status.serverUp) {\n\t\thost.showStatus(\"Local server not running — start it (any /models action) before removing.\");\n\t\treturn;\n\t}\n\tconst models = await host.localRuntime.list();\n\tconst target = models.find((model) => matchesInstalledLocalModel(ref, model.name));\n\tif (!target) {\n\t\thost.showStatus(\n\t\t\t`${ref} is not installed. Installed: ${models.map((model) => model.name).join(\", \") || \"(none)\"}`,\n\t\t);\n\t\treturn;\n\t}\n\tif (!confirmed) {\n\t\t// EXPLICIT USER ACTION ONLY: full disclosure, then require the confirm token.\n\t\tconst gb = (target.sizeBytes / 1e9).toFixed(2);\n\t\thost.showStatus(\n\t\t\t[\n\t\t\t\t`Removing ${ref} will delete:`,\n\t\t\t\t`  - model weights (${gb} GB) from ${status.managedByPi ? status.ownedModelsDir : \"the system server's storage\"}`,\n\t\t\t\t`  - the ollama/${ref} entry in models.json`,\n\t\t\t\t`  - its cached fitness report for this host`,\n\t\t\t\t`Run: /models remove ${ref} confirm`,\n\t\t\t].join(\"\\n\"),\n\t\t);\n\t\treturn;\n\t}\n\tconst removed = await host.localRuntime.remove(ref);\n\tif (!removed.ok) {\n\t\thost.showStatus(`Remove failed: ${removed.error}`);\n\t\treturn;\n\t}\n\tunregisterLocalModel({ agentDir: getAgentDir(), ref });\n\tFitnessStore.forAgentDir(getAgentDir()).remove(`ollama/${ref}`);\n\thost.session.modelRegistry.refresh();\n\thost.showStatus(`${ref} removed: weights deleted, registration and fitness report dropped.`);\n}\n\nasync function removeTransformersModel(host: LocalModelHost, modelId: string, confirmed: boolean): Promise<void> {\n\tconst runtime = host.getTransformersRuntime(modelId);\n\tconst status = await runtime.detect();\n\tif (!confirmed) {\n\t\thost.showStatus(\n\t\t\t[\n\t\t\t\t`Removing ${modelId} will delete:`,\n\t\t\t\t`  - the ${HF_TRANSFORMERS_PROVIDER}/${modelId} entry in models.json`,\n\t\t\t\t`  - its cached fitness report for this host`,\n\t\t\t\t`It will stop this session's Transformers sidecar if running. Cached weights remain under ${status.cacheDir}.`,\n\t\t\t\t`Run: /models remove hf.co/${modelId} confirm`,\n\t\t\t].join(\"\\n\"),\n\t\t);\n\t\treturn;\n\t}\n\truntime.stop();\n\tconst registration = unregisterTransformersModel({ agentDir: getAgentDir(), modelId });\n\tif (!registration.ok) {\n\t\thost.showStatus(`Remove failed: ${registration.reason}`);\n\t\treturn;\n\t}\n\tFitnessStore.forAgentDir(getAgentDir()).remove(`${HF_TRANSFORMERS_PROVIDER}/${modelId}`);\n\thost.session.modelRegistry.refresh();\n\thost.showStatus(\n\t\t`${modelId} registration and fitness report dropped; cached weights remain under ${status.cacheDir}.`,\n\t);\n}\n\n/** /fitness with no args: pick a model from the configured registry, probe it, assign a role. */\n/** Pick a validated suggestion → install it → probe on this host → land its shaped role. */\nexport function showModelSuggestionSelector(host: LocalModelHost): void {\n\thost.showSelector((done) => {\n\t\tconst selector = new ModelSuggestionSelectorComponent(\n\t\t\tDEFAULT_MODEL_SUGGESTIONS,\n\t\t\tasync (suggestion) => {\n\t\t\t\tdone();\n\t\t\t\t// The shaped role rides along so the post-probe selector lands on it pre-selected;\n\t\t\t\t// non-tool-callers carry curator/judge/none, never executor, so this can't footgun.\n\t\t\t\tconst source = normalizeModelSource(suggestion.pullRef);\n\t\t\t\tif (source.type === \"transformers\") {\n\t\t\t\t\tawait addTransformersModel(host, source.modelId, suggestion.assignRole);\n\t\t\t\t\treturn;\n\t\t\t\t}\n\t\t\t\tif (source.type === \"local\") {\n\t\t\t\t\tawait addLocalModel(host, source.pullRef, suggestion.assignRole);\n\t\t\t\t\treturn;\n\t\t\t\t}\n\t\t\t\thost.showStatus(\n\t\t\t\t\t`Suggestion ${suggestion.name} is not installable: ${source.type === \"rejected\" ? source.reason : source.ref}`,\n\t\t\t\t);\n\t\t\t},\n\t\t\t() => {\n\t\t\t\tdone();\n\t\t\t\thost.ui.requestRender();\n\t\t\t},\n\t\t);\n\t\treturn { component: selector, focus: selector };\n\t});\n}\n\nexport function showFitnessModelSelector(host: LocalModelHost): void {\n\thost.showSelector((done) => {\n\t\tconst selector = new ModelSelectorComponent(\n\t\t\thost.ui,\n\t\t\thost.session.model,\n\t\t\thost.settingsManager,\n\t\t\thost.session.modelRegistry,\n\t\t\thost.session.scopedModels,\n\t\t\tasync (model) => {\n\t\t\t\tdone();\n\t\t\t\tawait runFitnessAndAssign(host, `${model.provider}/${model.id}`);\n\t\t\t},\n\t\t\t() => {\n\t\t\t\tdone();\n\t\t\t\thost.ui.requestRender();\n\t\t\t},\n\t\t);\n\t\treturn { component: selector, focus: selector };\n\t});\n}\n\n/** Resolve the routed model's supported thinking levels for a fitness-assignment flow, with a safe fallback. */\nfunction getModelThinkingLevels(modelRegistry: ModelRegistry | undefined, modelRef: string): ThinkingLevel[] {\n\tif (!modelRegistry) {\n\t\treturn [...MODEL_ROUTER_THINKING_FALLBACK_LEVELS];\n\t}\n\tconst resolved = resolveCliModel({ cliModel: modelRef, modelRegistry });\n\tif (!resolved.model) {\n\t\treturn [...MODEL_ROUTER_THINKING_FALLBACK_LEVELS];\n\t}\n\treturn getSupportedThinkingLevels(resolved.model);\n}\n\nfunction getModelRouterThinkingField(role: FitnessRole): ModelRouterThinkingField | undefined {\n\tif (role === \"router-cheap\") {\n\t\treturn \"cheapThinking\";\n\t}\n\tif (role === \"router-medium\") {\n\t\treturn \"mediumThinking\";\n\t}\n\tif (role === \"router-expensive\") {\n\t\treturn \"expensiveThinking\";\n\t}\n\tif (role === \"executor\") {\n\t\treturn \"executorThinking\";\n\t}\n\tif (role === \"judge\") {\n\t\treturn \"judgeThinking\";\n\t}\n\treturn undefined;\n}\n\nfunction promptForModelRouterThinking(host: RunFitnessHost, modelRef: string, role: FitnessRole): void {\n\tconst thinkingField = getModelRouterThinkingField(role);\n\tif (!thinkingField) {\n\t\treturn;\n\t}\n\n\tconst modelRegistry = host.session.modelRegistry;\n\tconst availableLevels = getModelThinkingLevels(modelRegistry, modelRef);\n\tconst settings = host.settingsManager.getModelRouterSettings();\n\tconst configuredThinking = settings[thinkingField];\n\tconst currentThinking =\n\t\tconfiguredThinking && availableLevels.includes(configuredThinking)\n\t\t\t? configuredThinking\n\t\t\t: MODEL_ROUTER_THINKING_INHERIT_VALUE;\n\n\thost.showSelector((done) => {\n\t\tconst selector = new ModelRouterThinkingSelectorComponent(\n\t\t\tcurrentThinking,\n\t\t\tavailableLevels,\n\t\t\t(level) => {\n\t\t\t\tdone();\n\t\t\t\thost.settingsManager.setModelRouterSettings({\n\t\t\t\t\t...settings,\n\t\t\t\t\t[thinkingField]: level === MODEL_ROUTER_THINKING_INHERIT_VALUE ? undefined : level,\n\t\t\t\t});\n\t\t\t},\n\t\t\t() => {\n\t\t\t\tdone();\n\t\t\t\thost.ui.requestRender();\n\t\t\t},\n\t\t);\n\t\treturn { component: selector, focus: selector };\n\t});\n}\n\n/** Probe a model's fitness, show the report, then offer one-step role assignment. When the model\n * came from a validated suggestion, `preselectRole` lands its shaped role already highlighted. */\nexport async function runFitnessAndAssign(\n\thost: RunFitnessHost,\n\tmodelRef: string,\n\tpreselectRole?: FitnessRole,\n): Promise<void> {\n\thost.showStatus(`Model fitness probe running on ${modelRef}… (6 surfaces; local models may take a few minutes)`);\n\ttry {\n\t\tconst outcome = await host.session.runModelFitness({ model: modelRef });\n\t\tif (!outcome.started) {\n\t\t\thost.showStatus(`Model fitness skipped: ${outcome.skipReason}`);\n\t\t\treturn;\n\t\t}\n\t\thost.chatContainer.addChild(new Spacer(1));\n\t\thost.chatContainer.addChild(new Text(formatModelFitnessReport(outcome.model, outcome.report), 1, 0));\n\t\thost.ui.requestRender();\n\t\t// Validate-before-load: zero successes on every probed surface means the model cannot\n\t\t// drive any of the harness's subagent contracts on this host — refuse adoption instead\n\t\t// of landing a role selector the user might reflexively confirm (this is the reported\n\t\t// bug: a 0/3-everywhere model still got set as judge model and saved to Model Router).\n\t\tif (isProbeAllFailed(outcome.report)) {\n\t\t\thost.showStatus(\n\t\t\t\t`${outcome.model} failed the fitness probe on all surfaces — not configured. Use /model to set it manually if you accept the risk.`,\n\t\t\t);\n\t\t\treturn;\n\t\t}\n\t\thost.showSelector((done) => {\n\t\t\tconst selector = new FitnessRoleSelectorComponent(\n\t\t\t\toutcome.model,\n\t\t\t\t(role) => {\n\t\t\t\t\tdone();\n\t\t\t\t\tif (role === \"scout\") {\n\t\t\t\t\t\tconst verdict = evaluateSurfaceFitness(\"scout_auto\", outcome.report);\n\t\t\t\t\t\tif (!verdict.fit) {\n\t\t\t\t\t\t\tconst reason = verdict.reason === \"lane_failed\" ? `failed ${verdict.lane}` : \"was not probed\";\n\t\t\t\t\t\t\thost.showStatus(\n\t\t\t\t\t\t\t\t`${outcome.model} not assigned as scout: ${reason} on the scout_auto fitness exam. Use the docs/scout.md Modelfile recipe, then rerun /fitness.`,\n\t\t\t\t\t\t\t);\n\t\t\t\t\t\t\treturn;\n\t\t\t\t\t\t}\n\t\t\t\t\t}\n\t\t\t\t\tassignFitnessRole(host, outcome.model, role);\n\t\t\t\t\tpromptForModelRouterThinking(host, outcome.model, role);\n\t\t\t\t},\n\t\t\t\t() => {\n\t\t\t\t\tdone();\n\t\t\t\t\thost.ui.requestRender();\n\t\t\t\t},\n\t\t\t\tpreselectRole,\n\t\t\t);\n\t\t\treturn { component: selector, focus: selector };\n\t\t});\n\t} catch (error) {\n\t\thost.showError(error instanceof Error ? error.message : String(error));\n\t}\n}\n\n/** Persist a role assignment from the post-probe selector into the matching settings. */\nexport function assignFitnessRole(host: AssignRoleHost, modelRef: string, role: FitnessRole): void {\n\tif (role === \"none\") {\n\t\thost.showStatus(`Fitness result for ${modelRef} saved. Assign a role later from /settings.`);\n\t\treturn;\n\t}\n\tif (role === \"curator\") {\n\t\tconst current = host.settingsManager.getContextCurationSettings();\n\t\thost.settingsManager.setContextCurationSettings({ ...current, enabled: true, model: modelRef });\n\t\thost.showStatus(`Context curation enabled with ${modelRef} as the curator.`);\n\t\treturn;\n\t}\n\tif (role === \"executor\") {\n\t\tconst router = host.settingsManager.getModelRouterSettings();\n\t\thost.settingsManager.setModelRouterSettings({ ...router, executorModel: modelRef });\n\t\tconst hint = router.enabled ? \"\" : \" Model router is currently disabled — enable it in /settings → Model Router.\";\n\t\thost.showStatus(`${modelRef} set as the toolkit executor (direct Level-0 hits route to it).${hint}`);\n\t\treturn;\n\t}\n\tif (role === \"scout\") {\n\t\thost.settingsManager.setScoutSettings({ enabled: true, model: modelRef });\n\t\thost.showStatus(`${modelRef} set as the repository scout (context_scout enabled).`);\n\t\treturn;\n\t}\n\tconst router = host.settingsManager.getModelRouterSettings();\n\tconst field =\n\t\trole === \"router-cheap\"\n\t\t\t? \"cheapModel\"\n\t\t\t: role === \"router-medium\"\n\t\t\t\t? \"mediumModel\"\n\t\t\t\t: role === \"router-expensive\"\n\t\t\t\t\t? \"expensiveModel\"\n\t\t\t\t\t: role === \"judge\"\n\t\t\t\t\t\t? \"judgeModel\"\n\t\t\t\t\t\t: \"learningModel\";\n\thost.settingsManager.setModelRouterSettings({ ...router, [field]: modelRef });\n\tconst hint = router.enabled ? \"\" : \" Model router is currently disabled — enable it in /settings → Model Router.\";\n\thost.showStatus(`${modelRef} set as ${role.replace(\"router-\", \"router \")} model.${hint}`);\n}\n"]}

package/dist/modes/interactive/local-model-commands.js

@@ -109,6 +109,13 @@ export async function handleModelsCommand(host, argsText) {
                 : "No pi-managed server running (a system server, if any, is not pi's to stop).");
             return;
         }
+        if (action === "import") {
+            const imported = host.localRuntime.importUserModels();
+            host.showStatus(`Imported Ollama models from ${imported.sourceDir} into pi-owned store ${imported.targetDir}: ` +
+                `${imported.manifestsImported} manifest(s), ${imported.blobsHardlinked} blob(s) hardlinked, ` +
+                `${imported.blobsCopied} blob(s) copied, ${imported.blobsSkipped + imported.manifestsSkipped} existing file(s) skipped.`);
+            return;
+        }
         if (action === "add") {
             const rawRef = rest.join(" ");
             if (!rawRef) {
@@ -153,10 +160,21 @@ export async function handleModelsCommand(host, argsText) {
         host.showError(error instanceof Error ? error.message : String(error));
     }
 }
+function formatOllamaStore(store) {
+    if (!store)
+        return "unknown store (0 model(s))";
+    return `${store.path} [${store.kind}, ${store.modelCount} model(s)]`;
+}
 export async function ensureLocalServer(host) {
     const status = await host.localRuntime.detect();
-    if (status.serverUp)
-        return true;
+    if (status.serverUp) {
+        if (status.activeStore?.kind === "pi-owned")
+            return true;
+        host.showStatus(`Ollama is already running on ${status.serverUrl} with store ${formatOllamaStore(status.activeStore)}; ` +
+            `pi's canonical store is ${status.ownedModelsDir} (${status.ownedStore.modelCount} model(s)). ` +
+            "Stop the other serve or run /models import before adding models.");
+        return false;
+    }
     if (!status.binaryPath) {
         for (const line of host.localRuntime.installGuide())
             host.showStatus(line);
@@ -175,9 +193,9 @@ export async function listLocalModels(host) {
     const models = status.serverUp ? await host.localRuntime.list() : [];
     const ollamaLines = status.serverUp
         ? [
-            `Ollama models (${status.managedByPi ? `pi-managed server, storage: ${status.ownedModelsDir}` : "system server — storage owned by the system daemon"}):`,
+            `Ollama models (active store: ${formatOllamaStore(status.activeStore)}; pi-owned store: ${status.ownedModelsDir} with ${status.ownedStore.modelCount} model(s)):`,
             ...(models.length === 0
-                ? ["  (none installed — /models add <ref>, or /models suggest for a validated roster)"]
+                ? ["  (none installed — /models add <ref>, /models import, or /models suggest for a validated roster)"]
                 : []),
         ]
         : [
@@ -218,7 +236,7 @@ export async function listLocalModels(host) {
             return `  - ${model.name} (${gb} GB) · ${probe}`;
         }),
         ...(transformersLines.length > 0 ? ["Pi-managed Transformers models:", ...transformersLines] : []),
-        "Commands: /models add <ref> · /models remove <ref> confirm · /models stop",
+        "Commands: /models import · /models add <ref> · /models remove <ref> confirm · /models stop",
     ];
     for (const line of lines)
         host.showStatus(line);
@@ -226,7 +244,7 @@ export async function listLocalModels(host) {
 export async function addLocalModel(host, pullRef, preselectRole) {
     if (!(await ensureLocalServer(host)))
         return;
-    host.showStatus(`Pulling ${pullRef}… (weights land in the server's model storage)`);
+    host.showStatus(`Pulling ${pullRef}… (weights land in pi's owned Ollama store)`);
     let lastShown = 0;
     const pulled = await host.localRuntime.pull(pullRef, (progress) => {
         const now = Date.now();