@nemesis-oss/ollama-sdk 1.13.0 → 1.15.0

package/dist/testing.js.map

@@ -0,0 +1 @@
+{"version":3,"sources":["../src/testing/mock-server.ts"],"names":[],"mappings":";;;AAmPA,SAAS,MAAA,GAAiB;AACxB,EAAA,OAAA,iBAAO,IAAI,IAAA,EAAK,EAAE,WAAA,EAAY;AAChC;AAgBO,SAAS,cAAc,KAAA,EAYnB;AACT,EAAA,MAAM,IAAA,GAAO,MAAM,IAAA,IAAQ,KAAA;AAC3B,EAAA,MAAM,OAAA,GAAmC;AAAA,IACvC,IAAA,EAAM,WAAA;AAAA,IACN,OAAA,EAAS,MAAM,OAAA,IAAW;AAAA,GAC5B;AACA,EAAA,IAAI,MAAM,SAAA,KAAc,MAAA,IAAa,KAAA,CAAM,SAAA,CAAU,SAAS,CAAA,EAAG;AAC/D,IAAA,OAAA,CAAQ,YAAY,CAAA,GAAI,KAAA,CAAM,UAAU,GAAA,CAAI,CAAC,MAAM,CAAA,MAAO;AAAA,MACxD,EAAA,EAAI,IAAA,CAAK,EAAA,IAAM,CAAA,KAAA,EAAQ,CAAC,CAAA,CAAA;AAAA,MACxB,QAAA,EAAU;AAAA,QACR,MAAM,IAAA,CAAK,IAAA;AAAA,QACX,SAAA,EACE,OAAO,IAAA,CAAK,SAAA,KAAc,QAAA,GAAW,KAAK,SAAA,GAAY,IAAA,CAAK,SAAA,CAAU,IAAA,CAAK,SAAS;AAAA;AACvF,KACF,CAAE,CAAA;AAAA,EACJ;AACA,EAAA,MAAM,OAAA,GAAmC;AAAA,IACvC,KAAA,EAAO,MAAM,KAAA,IAAS,YAAA;AAAA,IACtB,YAAY,MAAA,EAAO;AAAA,IACnB,OAAA;AAAA,IACA;AAAA,GACF;AACA,EAAA,IAAI,IAAA,EAAM;AAGR,IAAA,OAAA,CAAQ,gBAAgB,CAAA,GAAI,KAAA,CAAM,eAAA,IAAmB,GAAA;AACrD,IAAA,IAAI,MAAM,cAAA,KAAmB,MAAA,EAAW,OAAA,CAAQ,eAAe,IAAI,KAAA,CAAM,cAAA;AACzE,IAAA,IAAI,MAAM,eAAA,KAAoB,MAAA,EAAW,OAAA,CAAQ,mBAAmB,IAAI,KAAA,CAAM,eAAA;AAC9E,IAAA,IAAI,KAAA,CAAM,yBAAyB,MAAA,EAAW;AAC5C,MAAA,OAAA,CAAQ,sBAAsB,IAAI,KAAA,CAAM,oBAAA;AAAA,IAC1C;AACA,IAAA,IAAI,KAAA,CAAM,0BAA0B,MAAA,EAAW;AAC7C,MAAA,OAAA,CAAQ,0BAA0B,IAAI,KAAA,CAAM,qBAAA;AAAA,IAC9C;AACA,IAAA,OAAA,CAAQ,YAAY,CAAA,GAAI,KAAA,CAAM,SAAA,IAAa,CAAA;AAC3C,IAAA,OAAA,CAAQ,eAAe,CAAA,GAAI,KAAA,CAAM,cAAA,IAAkB,GAAA;AAAA,EACrD;AACA,EAAA,OAAO,IAAA,CAAK,UAAU,OAAO,CAAA;AAC/B;AAaO,SAAS,cAAA,CAAe,IAAA,EAAc,IAAA,GAA8B,EAAC,EAAa;AACvF,EAAA,MAAM,KAAA,GAAQ,KAAK,KAAA,IAAS,YAAA;AAC5B,EAAA,MAAM,IAAA,GAAO,KAAK,SAAA,IAAa,IAAA,CAAK,YAAY,CAAA,GAAI,IAAA,CAAK,SAAA,GAAY,IAAA,CAAK,MAAA,IAAU,CAAA;AACpF,EAAA,MAAM,SAAmB,EAAC;AAC1B,EAAA,IAAI,IAAA,CAAK,SAAS,CAAA,EAAG;AACnB,IAAA,KAAA,IAAS,IAAI,CAAA,EAAG,CAAA,GAAI,IAAA,CAAK,MAAA,EAAQ,KAAK,IAAA,EAAM;AAC1C,MAAA,MAAA,CAAO,IAAA,CAAK,aAAA,CAAc,EAAE,OAAA,EAAS,IAAA,CAAK,KAAA,CAAM,CAAA,EAAG,CAAA,GAAI,IAAI,CAAA,EAAG,KAAA,EAAO,CAAC,CAAA;AAAA,IACxE;AAAA,EACF;AACA,EAAA,MAAA,CAAO,IAAA;AAAA,IACL,aAAA,CAAc;AAAA,MACZ,KAAA;AAAA,MACA,IAAA,EAAM,IAAA;AAAA,MACN,SAAA,EAAW,IAAA,CAAK,GAAA,CAAI,CAAA,EAAG,IAAA,CAAK,IAAA,CAAK,IAAA,CAAK,MAAA,GAAS,IAAA,CAAK,GAAA,CAAI,CAAA,EAAG,IAAI,CAAC,CAAC,CAAA;AAAA,MACjE,cAAA,EAAgB,GAAA;AAAA,MAChB,eAAA,EAAiB;AAAA,KAClB;AAAA,GACH;AACA,EAAA,OAAO,MAAA;AACT;AAkBO,SAAS,kBAAA,CACd,SAAA,EACA,IAAA,GAAkC,EAAC,EACzB;AACV,EAAA,OAAO;AAAA,IACL,aAAA,CAAc,EAAE,KAAA,EAAO,IAAA,CAAK,OAAO,OAAA,EAAS,IAAA,CAAK,OAAA,EAAS,SAAA,EAAW,CAAA;AAAA,IACrE,aAAA,CAAc,EAAE,KAAA,EAAO,IAAA,CAAK,KAAA,EAAO,IAAA,EAAM,IAAA,EAAM,SAAA,EAAW,CAAA,EAAG,cAAA,EAAgB,GAAA,EAAY;AAAA,GAC3F;AACF;AAkBO,IAAM,mBAAN,MAAuB;AAAA,EACX,MAAA;AAAA,EACA,MAAA,uBAAa,GAAA,EAAiC;AAAA,EAC9C,aAAkC,EAAC;AAAA,EACnC,eAAA;AAAA,EACA,cAAA;AAAA,EACT,IAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA,EAQR,WAAA,CAAY,gBAA4C,CAAA,EAAG;AACzD,IAAA,MAAM,UACJ,OAAO,aAAA,KAAkB,WAAW,EAAE,IAAA,EAAM,eAAc,GAAI,aAAA;AAChE,IAAA,IAAA,CAAK,IAAA,GAAO,QAAQ,IAAA,IAAQ,CAAA;AAC5B,IAAA,IAAA,CAAK,eAAA,GAAkB,QAAQ,aAAA,IAAiB,IAAA;AAChD,IAAA,IAAA,CAAK,cAAA,GAAiB,QAAQ,OAAA,IAAW,QAAA;AACzC,IAAA,IAAA,CAAK,MAAA,GAAS,YAAA,CAAa,CAAC,GAAA,EAAK,GAAA,KAAQ;AACvC,MAAA,KAAK,IAAA,CAAK,aAAA,CAAc,GAAA,EAAK,GAAG,CAAA;AAAA,IAClC,CAAC,CAAA;AAGD,IAAA,IAAA,CAAK,MAAA,CAAO,EAAA,CAAG,YAAA,EAAc,CAAC,MAAA,KAAW;AACvC,MAAA,MAAA,CAAO,WAAW,IAAI,CAAA;AAAA,IACxB,CAAC,CAAA;AAAA,EACH;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA,EAQA,QAAA,CAAS,MAAc,OAAA,EAAoC;AACzD,IAAA,IAAA,CAAK,MAAA,CAAO,GAAA,CAAI,IAAA,EAAM,OAAO,CAAA;AAAA,EAC/B;AAAA;AAAA,EAGA,KAAA,GAAc;AACZ,IAAA,IAAA,CAAK,OAAO,KAAA,EAAM;AAAA,EACpB;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA,EAQA,IAAI,QAAA,GAAyC;AAC3C,IAAA,OAAO,IAAA,CAAK,UAAA;AAAA,EACd;AAAA;AAAA,EAGA,aAAA,GAAsB;AACpB,IAAA,IAAA,CAAK,WAAW,MAAA,GAAS,CAAA;AAAA,EAC3B;AAAA;AAAA;AAAA;AAAA;AAAA,EAMA,IAAI,OAAA,GAAkB;AACpB,IAAA,MAAM,OAAA,GAAU,IAAA,CAAK,MAAA,CAAO,OAAA,EAAQ;AACpC,IAAA,MAAM,IAAA,GAAO,OAAA,EAAS,IAAA,IAAQ,IAAA,CAAK,IAAA;AACnC,IAAA,OAAO,oBAAoB,IAAI,CAAA,CAAA;AAAA,EACjC;AAAA;AAAA;AAAA;AAAA;AAAA,EAMA,KAAA,GAAuB;AACrB,IAAA,OAAO,IAAI,OAAA,CAAQ,CAAC,OAAA,EAAS,MAAA,KAAW;AACtC,MAAA,IAAA,CAAK,MAAA,CAAO,IAAA,CAAK,OAAA,EAAS,MAAM,CAAA;AAChC,MAAA,IAAA,CAAK,MAAA,CAAO,MAAA,CAAO,IAAA,CAAK,IAAA,EAAM,aAAa,MAAM;AAC/C,QAAA,IAAA,CAAK,MAAA,CAAO,cAAA,CAAe,OAAA,EAAS,MAAM,CAAA;AAC1C,QAAA,MAAM,OAAA,GAAU,IAAA,CAAK,MAAA,CAAO,OAAA,EAAQ;AACpC,QAAA,IAAI,YAAY,IAAA,EAAM;AACpB,UAAA,IAAA,CAAK,OAAO,OAAA,CAAQ,IAAA;AAAA,QACtB;AACA,QAAA,OAAA,EAAQ;AAAA,MACV,CAAC,CAAA;AAAA,IACH,CAAC,CAAA;AAAA,EACH;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA,EAQA,IAAA,GAAsB;AACpB,IAAA,OAAO,IAAI,OAAA,CAAQ,CAAC,OAAA,EAAS,MAAA,KAAW;AAKtC,MAAA,IAAI,CAAC,IAAA,CAAK,MAAA,CAAO,SAAA,EAAW;AAC1B,QAAA,OAAA,EAAQ;AACR,QAAA;AAAA,MACF;AACA,MAAA,IAAA,CAAK,MAAA,CAAO,KAAA,CAAM,CAAC,GAAA,KAAQ;AACzB,QAAA,IAAI,GAAA,IAAQ,GAAA,CAA8B,IAAA,KAAS,wBAAA,EAA0B;AAC3E,UAAA,MAAA,CAAO,GAAG,CAAA;AACV,UAAA;AAAA,QACF;AACA,QAAA,OAAA,EAAQ;AAAA,MACV,CAAC,CAAA;AAAA,IACH,CAAC,CAAA;AAAA,EACH;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA,EAOQ,eAAA,CAAgB,QAAgB,IAAA,EAA+C;AACrF,IAAA,IAAI,CAAC,KAAK,eAAA,EAAiB;AACzB,MAAA,OAAO,MAAA;AAAA,IACT;AAIA,IAAA,IAAI,IAAA,KAAS,GAAA,KAAQ,MAAA,KAAW,MAAA,IAAU,WAAW,KAAA,CAAA,EAAQ;AAC3D,MAAA,OAAO;AAAA,QACL,MAAA,EAAQ,GAAA;AAAA,QACR,OAAA,EAAS,EAAE,cAAA,EAAgB,YAAA,EAAa;AAAA;AAAA;AAAA;AAAA,QAIxC,GAAI,MAAA,KAAW,KAAA,GAAQ,EAAE,OAAA,EAAS,mBAAA,KAAwB;AAAC,OAC7D;AAAA,IACF;AACA,IAAA,IAAI,IAAA,KAAS,cAAA,IAAkB,MAAA,KAAW,KAAA,EAAO;AAC/C,MAAA,OAAO;AAAA,QACL,MAAA,EAAQ,GAAA;AAAA,QACR,OAAA,EAAS,EAAE,cAAA,EAAgB,kBAAA,EAAmB;AAAA,QAC9C,MAAA,EAAQ,CAAC,IAAA,CAAK,SAAA,CAAU,EAAE,OAAA,EAAS,IAAA,CAAK,cAAA,EAAgB,CAAC;AAAA,OAC3D;AAAA,IACF;AACA,IAAA,OAAO,MAAA;AAAA,EACT;AAAA,EAEA,MAAc,aAAA,CAAc,GAAA,EAAsB,GAAA,EAAoC;AACpF,IAAA,MAAM,GAAA,GAAM,IAAI,GAAA,IAAO,GAAA;AAIvB,IAAA,MAAM,OAAO,GAAA,CAAI,KAAA,CAAM,GAAG,CAAA,CAAE,CAAC,CAAA,IAAK,GAAA;AAClC,IAAA,MAAM,MAAA,GAAS,IAAI,MAAA,IAAU,KAAA;AAG7B,IAAA,IAAI,UAAA;AACJ,IAAA,IAAI,MAAA,KAAW,MAAA,IAAU,MAAA,KAAW,KAAA,EAAO;AACzC,MAAA,MAAM,aAAuB,EAAC;AAC9B,MAAA,WAAA,MAAiB,SAAS,GAAA,EAAK;AAC7B,QAAA,UAAA,CAAW,KAAK,KAAe,CAAA;AAAA,MACjC;AACA,MAAA,MAAM,WAAW,MAAA,CAAO,MAAA,CAAO,UAAU,CAAA,CAAE,SAAS,OAAO,CAAA;AAC3D,MAAA,IAAI,QAAA,EAAU;AACZ,QAAA,IAAI;AACF,UAAA,UAAA,GAAa,IAAA,CAAK,MAAM,QAAQ,CAAA;AAAA,QAClC,CAAA,CAAA,MAAQ;AACN,UAAA,UAAA,GAAa,QAAA;AAAA,QACf;AAAA,MACF;AAAA,IACF;AAEA,IAAA,MAAM,MAAA,GAA4B;AAAA,MAChC,MAAA;AAAA,MACA,IAAA;AAAA,MACA,GAAI,UAAA,KAAe,MAAA,GAAY,EAAE,IAAA,EAAM,UAAA,KAAe,EAAC;AAAA,MACvD,WAAW,MAAA;AAAO,KACpB;AACA,IAAA,IAAA,CAAK,UAAA,CAAW,KAAK,MAAM,CAAA;AAE3B,IAAA,MAAM,KAAA,GAAQ,KAAK,MAAA,CAAO,GAAA,CAAI,IAAI,CAAA,IAAK,IAAA,CAAK,eAAA,CAAgB,MAAA,EAAQ,IAAI,CAAA;AAExE,IAAA,IAAI,CAAC,KAAA,EAAO;AACV,MAAA,GAAA,CAAI,SAAA,CAAU,GAAA,EAAK,EAAE,cAAA,EAAgB,oBAAoB,CAAA;AACzD,MAAA,GAAA,CAAI,GAAA,CAAI,KAAK,SAAA,CAAU,EAAE,OAAO,CAAA,kBAAA,EAAqB,IAAI,CAAA,CAAA,EAAI,CAAC,CAAA;AAC9D,MAAA;AAAA,IACF;AAEA,IAAA,KAAA,CAAM,YAAY,UAAU,CAAA;AAK5B,IAAA,MAAM,QAAA,GACJ,MAAM,SAAA,KAAc,MAAA,GAAa,MAAM,SAAA,CAAU,MAAM,KAAK,KAAA,GAAS,KAAA;AAEvE,IAAA,MAAM,MAAA,GAAS,SAAS,MAAA,IAAU,GAAA;AAClC,IAAA,MAAM,MAAA,GAAS,QAAA,CAAS,MAAA,IAAU,EAAC;AACnC,IAAA,MAAM,YAAA,GAAe,SAAS,YAAA,IAAgB,CAAA;AAC9C,IAAA,MAAM,OAAA,GAAU,QAAA,CAAS,OAAA,IAAW,EAAE,gBAAgB,sBAAA,EAAuB;AAE7E,IAAA,GAAA,CAAI,SAAA,CAAU,QAAQ,OAAO,CAAA;AAE7B,IAAA,IAAI,QAAA,CAAS,YAAY,MAAA,EAAW;AAClC,MAAA,IAAI,eAAe,CAAA,EAAG;AACpB,QAAA,MAAM,IAAI,OAAA,CAAQ,CAAC,MAAM,UAAA,CAAW,CAAA,EAAG,YAAY,CAAC,CAAA;AAAA,MACtD;AACA,MAAA,GAAA,CAAI,KAAA,CAAM,SAAS,OAAO,CAAA;AAC1B,MAAA,GAAA,CAAI,GAAA,EAAI;AACR,MAAA;AAAA,IACF;AAEA,IAAA,IAAI,SAAS,uBAAA,EAAyB;AAGpC,MAAA,IAAI,MAAA,CAAO,SAAS,CAAA,EAAG;AACrB,QAAA,IAAI,eAAe,CAAA,EAAG;AACpB,UAAA,MAAM,IAAI,OAAA,CAAQ,CAAC,MAAM,UAAA,CAAW,CAAA,EAAG,YAAY,CAAC,CAAA;AAAA,QACtD;AACA,QAAA,GAAA,CAAI,KAAA,CAAM,MAAA,CAAO,CAAC,CAAA,GAAI,IAAI,CAAA;AAAA,MAC5B;AACA,MAAA,GAAA,CAAI,QAAQ,OAAA,EAAQ;AACpB,MAAA;AAAA,IACF;AAIA,IAAA,MAAM,OAAA,GAAU,QAAA,CAAS,gBAAA,GACpB,QAAA,CAAS,gBAAA,CAAiB,KAAA,IAAS,IAAA,CAAK,GAAA,CAAI,CAAA,EAAG,MAAA,CAAO,MAAA,GAAS,CAAC,IACjE,MAAA,CAAO,iBAAA;AACX,IAAA,MAAM,YAAA,GAAe,SAAS,gBAAA,EAAkB,OAAA;AAEhD,IAAA,KAAA,IAAS,CAAA,GAAI,CAAA,EAAG,CAAA,GAAI,MAAA,CAAO,QAAQ,CAAA,EAAA,EAAK;AACtC,MAAA,IAAI,CAAA,KAAM,OAAA,IAAW,YAAA,KAAiB,MAAA,EAAW;AAG/C,QAAA,GAAA,CAAI,KAAA,CAAM,KAAK,SAAA,CAAU,EAAE,OAAO,YAAA,EAAc,IAAI,IAAI,CAAA;AACxD,QAAA,GAAA,CAAI,GAAA,EAAI;AACR,QAAA;AAAA,MACF;AACA,MAAA,IAAI,eAAe,CAAA,EAAG;AACpB,QAAA,MAAM,IAAI,OAAA,CAAQ,CAAC,MAAM,UAAA,CAAW,CAAA,EAAG,YAAY,CAAC,CAAA;AAAA,MACtD;AACA,MAAA,GAAA,CAAI,KAAA,CAAM,MAAA,CAAO,CAAC,CAAA,GAAI,IAAI,CAAA;AAAA,IAC5B;AAEA,IAAA,IAAI,YAAA,KAAiB,MAAA,IAAa,MAAA,CAAO,MAAA,IAAU,OAAA,EAAS;AAC1D,MAAA,GAAA,CAAI,KAAA,CAAM,KAAK,SAAA,CAAU,EAAE,OAAO,YAAA,EAAc,IAAI,IAAI,CAAA;AAAA,IAC1D;AACA,IAAA,GAAA,CAAI,GAAA,EAAI;AAAA,EACV;AACF;AAgBO,SAAS,sBAAA,CAAuB,OAAA,GAA6B,EAAC,EAAqB;AACxF,EAAA,OAAO,IAAI,iBAAiB,OAAO,CAAA;AACrC","file":"testing.js","sourcesContent":["/**\n * In-memory Ollama mock server for deterministic, zero-network CI testing.\n *\n * This is the SDK's own mock engine, promoted from the internal test fixture\n * (`test/mocks/ollama-mock-server.ts`) to a public entry point\n * (`@nemesis-oss/ollama-sdk/testing`) so consumers don't have to hand-roll\n * ad-hoc HTTP stubs to test their Ollama integrations — see ADR 0029.\n * The internal fixture re-exports this module, so the SDK's own test suite\n * exercises the exact same code consumers get.\n *\n * The VCR cassette system (`test/vcr.ts` in the repo) replays recorded HTTP\n * interactions, but cassettes are static. This mock server covers what\n * cassettes can't:\n *\n *   - programmable inter-chunk delays (backpressure handling)\n *   - mid-stream **in-band error injection** — an `{\"error\": \"...\"}` NDJSON\n *     frame emitted mid-response with HTTP 200 still, exactly how a real\n *     Ollama daemon reports GPU OOM / driver crashes mid-generation\n *   - connection drops mid-stream (`dropConnectionMidStream`)\n *   - 503/502/429 error status codes with custom bodies\n *   - arbitrary NDJSON chunk payloads (tool calls, thinking, partial frames)\n *\n * On top of the raw engine, the public surface adds the conveniences a\n * consumer test suite wants on day one:\n *\n *   - {@link createMockOllamaServer} factory (the recommended entry point)\n *   - built-in **default routes** — the official liveness probe (`HEAD /`\n *     and `GET /` → \"Ollama is running\") and `GET /api/version` — so\n *     `client.ping()` and friends work before any route is registered.\n *     Explicit `register()` calls always win.\n *   - **request recording** (`server.requests`, `clearRequests()`) for\n *     wire-level assertions without per-route callbacks\n *   - **request-aware routes** (`responder`) — per-request responses computed\n *     from the recorded method/path/body: fail-twice-then-succeed retry\n *     drills, body-dependent payloads, call-count scenarios\n *   - NDJSON **chunk builders** ({@link mockChatChunk}, {@link mockTextChunks},\n *     {@link mockToolCallChunks}) that produce correctly-shaped `/api/chat`\n *     frames — including parallel tool calls in one frame, the documented\n *     native behavior\n *\n * ## Usage\n *\n * ```ts\n * import { createMockOllamaServer, mockToolCallChunks } from '@nemesis-oss/ollama-sdk/testing';\n * import { OllamaClient } from '@nemesis-oss/ollama-sdk';\n *\n * const server = createMockOllamaServer(); // ephemeral port, default routes on\n * await server.start();\n *\n * server.register('/api/chat', {\n *   chunks: mockToolCallChunks([\n *     { name: 'get_weather', arguments: { city: 'Tokyo' } },\n *     { name: 'get_time', arguments: { tz: 'Asia/Tokyo' } },\n *   ]),\n * });\n *\n * const client = new OllamaClient({ baseUrl: server.baseUrl });\n * // ... exercise the client; assert on server.requests ...\n *\n * await server.stop();\n * ```\n *\n * ## Environment\n *\n * Node.js only — the module builds on `node:http`, which is why it lives in\n * a dedicated subpath export: importing it from the browser/Edge-safe main\n * entry would drag a Node built-in into every bundle. Never re-export this\n * module from `src/index.ts`; the Edge Runtime verification\n * (`npm run verify:edge-runtime`) exists to catch exactly that.\n *\n * Designed for Vitest/Jest: `start()` resolves once the server listens and\n * `stop()` is idempotent, so `beforeEach`/`afterEach` hooks manage the\n * lifecycle without tracking whether `start()` ran.\n */\n\nimport { createServer, type Server, type IncomingMessage, type ServerResponse } from 'node:http';\nimport type { AddressInfo } from 'node:net';\n\n/** One recorded incoming request, for wire-level test assertions. */\nexport interface MockRequestRecord {\n  /** HTTP method (`GET`, `POST`, `HEAD`, ...). */\n  readonly method: string;\n  /** Request path with the query string stripped. */\n  readonly path: string;\n  /** Parsed request body (JSON-parsed when possible, raw string otherwise); `undefined` for body-less methods. */\n  readonly body?: unknown;\n  /** ISO-8601 timestamp of when the request arrived. */\n  readonly timestamp: string;\n}\n\n/**\n * The response a mock route emits. {@link MockEndpointOptions} extends this\n * with hooks, and a {@link MockEndpointOptions.responder} computes one of\n * these per request. All fields optional — `chunks` defaults to `[]` (useful\n * for testing the \"stream closed without data\" case).\n */\nexport interface MockRouteResponse {\n  /** HTTP status code. Defaults to 200. */\n  readonly status?: number | undefined;\n  /**\n   * NDJSON chunks to emit, in order. Each chunk is emitted as a separate\n   * `res.write()` call followed by `\\n`. Pass `[]` to emit no body (useful\n   * for testing the \"stream closed without data\" case).\n   */\n  readonly chunks?: readonly string[] | undefined;\n  /**\n   * Delay (ms) before emitting each chunk. Defaults to 0 (no delay). Set to\n   * a small value (5-50ms) to test backpressure handling; set to a large\n   * value (10_000ms) to test timeout/cancellation.\n   */\n  readonly chunkDelayMs?: number | undefined;\n  /**\n   * Custom response headers. Defaults to\n   * `{ 'content-type': 'application/x-ndjson' }` for streaming endpoints.\n   * Override for non-JSON responses (e.g. blob uploads).\n   */\n  readonly headers?: Record<string, string> | undefined;\n  /**\n   * When `true`, the server closes the connection mid-stream after emitting\n   * the first chunk. Use this to test connection-drop recovery. Defaults to\n   * `false`.\n   */\n  readonly dropConnectionMidStream?: boolean | undefined;\n  /**\n   * Emit this exact string as the whole response body — no NDJSON newline\n   * framing. Overrides `chunks` (use it for non-JSON responses such as the\n   * plain-text liveness body or raw blob bytes).\n   */\n  readonly rawBody?: string | undefined;\n  /**\n   * Inject an in-band error frame — `{\"error\": \"<message>\"}` — into the\n   * NDJSON response while the HTTP status stays 200. This is how a real\n   * Ollama daemon reports mid-generation failures (GPU OOM, driver crash,\n   * context overflow): the status code never changes, the parser has to\n   * catch the frame. The SDK's stream readers raise `OllamaStreamError` on\n   * it.\n   *\n   * `after` is the number of chunks emitted before the error frame. The\n   * default is `chunks.length - 1` — every content chunk, then the error\n   * frame, with the final `done` frame never emitted, exactly what a\n   * mid-generation crash looks like on the wire (an error frame *after* the\n   * done frame would be unread: the stream has already completed cleanly).\n   * `after: 0` errors before any content; an explicit `after >=\n   * chunks.length` emits every chunk and then the error. Chunks at or after\n   * the injection point are never emitted: the server \"crashed\".\n   */\n  readonly injectErrorFrame?:\n    | {\n        readonly after?: number | undefined;\n        readonly message: string;\n      }\n    | undefined;\n}\n\n/**\n * A single mock route handler. The server matches on the request path\n * (query string stripped — Ollama's API doesn't use query params for the\n * routes we care about).\n */\nexport interface MockEndpointOptions extends MockRouteResponse {\n  /**\n   * Optional callback fired when a request arrives at this route. Receives\n   * the parsed request body (or `undefined` if the body wasn't valid JSON).\n   * Useful for assertions like \"the client sent the right `model` field\".\n   * For global capture across routes see {@link OllamaMockServer.requests}.\n   */\n  readonly onRequest?: ((body: unknown) => void) | undefined;\n  /**\n   * Compute this route's response **per request** — request-aware scenario\n   * scripting without subclassing. Receives the same {@link MockRequestRecord}\n   * that lands in {@link OllamaMockServer.requests} (method, path, parsed\n   * body, timestamp); the return value is the complete response for that\n   * request, or `undefined` to fall back to the route's static fields.\n   * Typical scenarios:\n   *\n   * ```ts\n   * // Fail twice, then serve — retry/backoff drills:\n   * let hits = 0;\n   * server.register('/api/chat', {\n   *   chunks: mockTextChunks('recovered'),\n   *   responder: () =>\n   *     hits++ < 2 ? { status: 503, chunks: ['{\"error\":\"server queue is full\"}'] } : undefined,\n   * });\n   *\n   * // Body-dependent — answer differently by request payload:\n   * server.register('/api/generate', {\n   *   responder: (record) => ({\n   *     chunks: [frameFor((record.body as { options?: { draft_num_predict?: number } })\n   *       .options?.draft_num_predict ?? 0)],\n   *   }),\n   * });\n   * ```\n   *\n   * The return value replaces the static response fields for that request\n   * (the fallback only happens on `undefined`), and it feeds the full\n   * emission pipeline — `chunkDelayMs`, `injectErrorFrame`,\n   * `dropConnectionMidStream`, and `rawBody` all keep their semantics.\n   */\n  readonly responder?: ((record: MockRequestRecord) => MockRouteResponse | undefined) | undefined;\n}\n\n/** Options for {@link createMockOllamaServer} and the `OllamaMockServer` constructor. */\nexport interface MockServerOptions {\n  /** TCP port to listen on. `0` (the default) picks an ephemeral port. */\n  readonly port?: number | undefined;\n  /**\n   * Serve the built-in default routes when no route is explicitly registered\n   * for a path: `HEAD /` and `GET /` → `200` (the official liveness probe —\n   * `GET /` answers the plain-text body \"Ollama is running\") and\n   * `GET /api/version` → `200 {\"version\": \"<version>\"}`. Defaults to `true`.\n   * Explicit `register()` calls always take precedence.\n   */\n  readonly defaultRoutes?: boolean | undefined;\n  /** Version reported by the default `GET /api/version` route. Defaults to `'0.14.0'`. */\n  readonly version?: string | undefined;\n}\n\n/** Shape accepted by {@link mockToolCallChunks} for each tool call. */\nexport interface MockToolCallSpec {\n  /** Tool-call id; auto-generated (`call-0`, `call-1`, …) when omitted. */\n  readonly id?: string | undefined;\n  /** Tool name, as the model would emit it. */\n  readonly name: string;\n  /** Tool arguments — an object (stringified by the builder) or a pre-encoded string. */\n  readonly arguments: Record<string, unknown> | string;\n}\n\n/** Options for {@link mockTextChunks}. */\nexport interface MockTextChunksOptions {\n  /** Split the text into content frames of at most this many characters. Default: one frame with the whole text. */\n  readonly chunkSize?: number | undefined;\n  /** Model name stamped on each frame. Defaults to `'mock-model'`. */\n  readonly model?: string | undefined;\n}\n\n/** Options for {@link mockToolCallChunks}. */\nexport interface MockToolCallChunksOptions {\n  /** Model name stamped on each frame. Defaults to `'mock-model'`. */\n  readonly model?: string | undefined;\n  /** Assistant text to carry on the tool-call frame (some models emit a preamble). Defaults to `''`. */\n  readonly content?: string | undefined;\n}\n\nfunction nowIso(): string {\n  return new Date().toISOString();\n}\n\n/**\n * One `/api/chat` NDJSON frame as a ready-to-emit string. Sensible defaults:\n * `model: 'mock-model'`, a fresh `created_at`, `done: false`.\n *\n * ```ts\n * server.register('/api/chat', {\n *   chunks: [\n *     mockChatChunk({ content: 'Hel' }),\n *     mockChatChunk({ content: 'lo' }),\n *     mockChatChunk({ done: true, evalCount: 2, evalDurationNs: 20_000_000 }),\n *   ],\n * });\n * ```\n */\nexport function mockChatChunk(frame: {\n  readonly content?: string | undefined;\n  readonly toolCalls?: readonly MockToolCallSpec[] | undefined;\n  readonly done?: boolean | undefined;\n  readonly model?: string | undefined;\n  readonly evalCount?: number | undefined;\n  readonly evalDurationNs?: number | undefined;\n  readonly promptEvalCount?: number | undefined;\n  readonly promptEvalDurationNs?: number | undefined;\n  readonly promptEvalCachedCount?: number | undefined;\n  readonly totalDurationNs?: number | undefined;\n  readonly loadDurationNs?: number | undefined;\n}): string {\n  const done = frame.done ?? false;\n  const message: Record<string, unknown> = {\n    role: 'assistant',\n    content: frame.content ?? '',\n  };\n  if (frame.toolCalls !== undefined && frame.toolCalls.length > 0) {\n    message['tool_calls'] = frame.toolCalls.map((call, i) => ({\n      id: call.id ?? `call-${i}`,\n      function: {\n        name: call.name,\n        arguments:\n          typeof call.arguments === 'string' ? call.arguments : JSON.stringify(call.arguments),\n      },\n    }));\n  }\n  const payload: Record<string, unknown> = {\n    model: frame.model ?? 'mock-model',\n    created_at: nowIso(),\n    message,\n    done,\n  };\n  if (done) {\n    // Counters on the final frame — shaped exactly like the official\n    // /api/chat streaming examples so telemetry code paths see real fields.\n    payload['total_duration'] = frame.totalDurationNs ?? 50_000_000;\n    if (frame.loadDurationNs !== undefined) payload['load_duration'] = frame.loadDurationNs;\n    if (frame.promptEvalCount !== undefined) payload['prompt_eval_count'] = frame.promptEvalCount;\n    if (frame.promptEvalDurationNs !== undefined) {\n      payload['prompt_eval_duration'] = frame.promptEvalDurationNs;\n    }\n    if (frame.promptEvalCachedCount !== undefined) {\n      payload['prompt_eval_cached_count'] = frame.promptEvalCachedCount;\n    }\n    payload['eval_count'] = frame.evalCount ?? 2;\n    payload['eval_duration'] = frame.evalDurationNs ?? 20_000_000;\n  }\n  return JSON.stringify(payload);\n}\n\n/**\n * Content frames + a final `done` frame for a plain-text `/api/chat` reply.\n * With `chunkSize`, the text is split into that many characters per frame —\n * the shape streaming code needs to exercise incremental assembly.\n *\n * ```ts\n * server.register('/api/chat', {\n *   chunks: mockTextChunks('Hello, world!', { chunkSize: 5 }),\n * });\n * ```\n */\nexport function mockTextChunks(text: string, opts: MockTextChunksOptions = {}): string[] {\n  const model = opts.model ?? 'mock-model';\n  const size = opts.chunkSize && opts.chunkSize > 0 ? opts.chunkSize : text.length || 1;\n  const frames: string[] = [];\n  if (text.length > 0) {\n    for (let i = 0; i < text.length; i += size) {\n      frames.push(mockChatChunk({ content: text.slice(i, i + size), model }));\n    }\n  }\n  frames.push(\n    mockChatChunk({\n      model,\n      done: true,\n      evalCount: Math.max(1, Math.ceil(text.length / Math.max(1, size))),\n      evalDurationNs: 20_000_000,\n      totalDurationNs: 50_000_000,\n    }),\n  );\n  return frames;\n}\n\n/**\n * Tool-call frames for `/api/chat` — the documented native shape: tool calls\n * arrive **complete in one frame** (parallel calls = one array with several\n * entries), followed by a final `done` frame. Pair with an `Agent` or tool\n * loop for deterministic tool-execution tests.\n *\n * ```ts\n * // Parallel tool calls, one round trip:\n * server.register('/api/chat', {\n *   chunks: mockToolCallChunks([\n *     { name: 'get_weather', arguments: { city: 'Tokyo' } },\n *     { name: 'get_time', arguments: { tz: 'Asia/Tokyo' } },\n *   ]),\n * });\n * ```\n */\nexport function mockToolCallChunks(\n  toolCalls: readonly MockToolCallSpec[],\n  opts: MockToolCallChunksOptions = {},\n): string[] {\n  return [\n    mockChatChunk({ model: opts.model, content: opts.content, toolCalls }),\n    mockChatChunk({ model: opts.model, done: true, evalCount: 1, evalDurationNs: 10_000_000 }),\n  ];\n}\n\n/**\n * Lightweight in-memory Ollama mock server.\n *\n * Spawns a real `node:http` server on a free port (or a fixed port if you\n * pass one). Tests register route handlers via `register()`, then point an\n * `OllamaClient` at `server.baseUrl`. The server emits the registered chunks\n * with the configured delay, status, and connection behavior; when no route\n * is registered the built-in liveness + version defaults answer (see\n * {@link MockServerOptions.defaultRoutes}).\n *\n * Prefer the {@link createMockOllamaServer} factory; the class is exported\n * for typed references and back-compatibility with the SDK's own suite.\n *\n * The server is **single-process** — it doesn't fork or use workers. It's\n * intended for unit/integration tests, not load testing.\n */\nexport class OllamaMockServer {\n  private readonly server: Server;\n  private readonly routes = new Map<string, MockEndpointOptions>();\n  private readonly requestLog: MockRequestRecord[] = [];\n  private readonly defaultsEnabled: boolean;\n  private readonly defaultVersion: string;\n  private port: number;\n\n  /**\n   * @param portOrOptions Either a TCP port (pass `0` for an ephemeral port —\n   *        recommended for parallel test runs) or a {@link MockServerOptions}\n   *        object. The actual port is exposed via {@link baseUrl} after\n   *        {@link start}.\n   */\n  constructor(portOrOptions: number | MockServerOptions = 0) {\n    const options: MockServerOptions =\n      typeof portOrOptions === 'number' ? { port: portOrOptions } : portOrOptions;\n    this.port = options.port ?? 0;\n    this.defaultsEnabled = options.defaultRoutes ?? true;\n    this.defaultVersion = options.version ?? '0.14.0';\n    this.server = createServer((req, res) => {\n      void this.handleRequest(req, res);\n    });\n    // Disable Nagle's algorithm so chunks are emitted immediately rather\n    // than being coalesced — makes delay/fracture behavior deterministic.\n    this.server.on('connection', (socket) => {\n      socket.setNoDelay(true);\n    });\n  }\n\n  /**\n   * Register a route handler. Replaces any existing handler for the same\n   * path (including a built-in default). The path is matched against\n   * `req.url` exactly (query string stripped, no template matching — for\n   * `/api/blobs/{digest}`, register each specific digest value).\n   */\n  register(path: string, options: MockEndpointOptions): void {\n    this.routes.set(path, options);\n  }\n\n  /** Clear all registered routes (built-in defaults remain). Useful in `beforeEach` hooks. */\n  clear(): void {\n    this.routes.clear();\n  }\n\n  /**\n   * Every request this server has served, in arrival order — method, path,\n   * parsed body, timestamp — including 404s and default-route hits. Assert\n   * on what the client actually put on the wire; clear with\n   * {@link clearRequests} (e.g. in `beforeEach`).\n   */\n  get requests(): readonly MockRequestRecord[] {\n    return this.requestLog;\n  }\n\n  /** Empty the request log ({@link requests}). Routes are unaffected. */\n  clearRequests(): void {\n    this.requestLog.length = 0;\n  }\n\n  /**\n   * Base URL to point an `OllamaClient` at. Only meaningful after\n   * {@link start} has resolved.\n   */\n  get baseUrl(): string {\n    const address = this.server.address() as AddressInfo | null;\n    const port = address?.port ?? this.port;\n    return `http://127.0.0.1:${port}`;\n  }\n\n  /**\n   * Start listening. Resolves once the server is ready to accept\n   * connections.\n   */\n  start(): Promise<void> {\n    return new Promise((resolve, reject) => {\n      this.server.once('error', reject);\n      this.server.listen(this.port, '127.0.0.1', () => {\n        this.server.removeListener('error', reject);\n        const address = this.server.address() as AddressInfo | null;\n        if (address !== null) {\n          this.port = address.port;\n        }\n        resolve();\n      });\n    });\n  }\n\n  /**\n   * Stop listening and close all active connections. Resolves once the\n   * server is fully closed (no lingering sockets). Idempotent — calling\n   * `stop()` on an already-stopped server resolves cleanly rather than\n   * throwing `ERR_SERVER_NOT_RUNNING`.\n   */\n  stop(): Promise<void> {\n    return new Promise((resolve, reject) => {\n      // `server.close()` errors with ERR_SERVER_NOT_RUNNING if the server\n      // isn't currently listening. Treat that as a successful no-op so\n      // callers can use stop() in afterEach hooks without tracking whether\n      // start() was actually called.\n      if (!this.server.listening) {\n        resolve();\n        return;\n      }\n      this.server.close((err) => {\n        if (err && (err as NodeJS.ErrnoException).code !== 'ERR_SERVER_NOT_RUNNING') {\n          reject(err);\n          return;\n        }\n        resolve();\n      });\n    });\n  }\n\n  /**\n   * Built-in default handler for a path, or `undefined` when defaults are\n   * disabled or the path isn't one of the official liveness surfaces.\n   * Registered routes are checked first and always win.\n   */\n  private defaultRouteFor(method: string, path: string): MockEndpointOptions | undefined {\n    if (!this.defaultsEnabled) {\n      return undefined;\n    }\n    // The official liveness probe: HEAD / → 200 (no body — Node suppresses\n    // writes for HEAD responses); GET / → the plain-text body\n    // \"Ollama is running\" (docs: the root endpoint answers exactly this).\n    if (path === '/' && (method === 'HEAD' || method === 'GET')) {\n      return {\n        status: 200,\n        headers: { 'content-type': 'text/plain' },\n        // GET / answers the plain-text body byte-for-byte — no trailing\n        // newline, hence rawBody rather than the NDJSON chunk emitter.\n        // HEAD / carries no body at all (Node suppresses HEAD writes).\n        ...(method === 'GET' ? { rawBody: 'Ollama is running' } : {}),\n      };\n    }\n    if (path === '/api/version' && method === 'GET') {\n      return {\n        status: 200,\n        headers: { 'content-type': 'application/json' },\n        chunks: [JSON.stringify({ version: this.defaultVersion })],\n      };\n    }\n    return undefined;\n  }\n\n  private async handleRequest(req: IncomingMessage, res: ServerResponse): Promise<void> {\n    const url = req.url ?? '/';\n    // Strip query string for route matching — Ollama's API doesn't use\n    // query params for the routes we care about, and stripping makes the\n    // test setup simpler.\n    const path = url.split('?')[0] ?? '/';\n    const method = req.method ?? 'GET';\n\n    // Capture the request body for recording + the route's onRequest.\n    let parsedBody: unknown;\n    if (method === 'POST' || method === 'PUT') {\n      const bodyChunks: Buffer[] = [];\n      for await (const chunk of req) {\n        bodyChunks.push(chunk as Buffer);\n      }\n      const bodyText = Buffer.concat(bodyChunks).toString('utf-8');\n      if (bodyText) {\n        try {\n          parsedBody = JSON.parse(bodyText);\n        } catch {\n          parsedBody = bodyText;\n        }\n      }\n    }\n\n    const record: MockRequestRecord = {\n      method,\n      path,\n      ...(parsedBody !== undefined ? { body: parsedBody } : {}),\n      timestamp: nowIso(),\n    };\n    this.requestLog.push(record);\n\n    const route = this.routes.get(path) ?? this.defaultRouteFor(method, path);\n\n    if (!route) {\n      res.writeHead(404, { 'content-type': 'application/json' });\n      res.end(JSON.stringify({ error: `Route not mocked: ${path}` }));\n      return;\n    }\n\n    route.onRequest?.(parsedBody);\n\n    // Request-aware routes: the responder computes the response for THIS\n    // request (method/path/body in, response out); returning `undefined`\n    // falls back to the route's static fields.\n    const response: MockRouteResponse =\n      route.responder !== undefined ? (route.responder(record) ?? route) : route;\n\n    const status = response.status ?? 200;\n    const chunks = response.chunks ?? [];\n    const chunkDelayMs = response.chunkDelayMs ?? 0;\n    const headers = response.headers ?? { 'content-type': 'application/x-ndjson' };\n\n    res.writeHead(status, headers);\n\n    if (response.rawBody !== undefined) {\n      if (chunkDelayMs > 0) {\n        await new Promise((r) => setTimeout(r, chunkDelayMs));\n      }\n      res.write(response.rawBody);\n      res.end();\n      return;\n    }\n\n    if (response.dropConnectionMidStream) {\n      // Emit the first chunk (if any), then destroy the socket without\n      // ending the response — a network drop mid-stream.\n      if (chunks.length > 0) {\n        if (chunkDelayMs > 0) {\n          await new Promise((r) => setTimeout(r, chunkDelayMs));\n        }\n        res.write(chunks[0] + '\\n');\n      }\n      res.socket?.destroy();\n      return;\n    }\n\n    // In-band injection point, measured in already-emitted chunks. See\n    // MockRouteResponse.injectErrorFrame for the default rationale.\n    const errorAt = response.injectErrorFrame\n      ? (response.injectErrorFrame.after ?? Math.max(0, chunks.length - 1))\n      : Number.POSITIVE_INFINITY;\n    const errorMessage = response.injectErrorFrame?.message;\n\n    for (let i = 0; i < chunks.length; i++) {\n      if (i === errorAt && errorMessage !== undefined) {\n        // The server \"crashed\": emit the error frame, never the remaining\n        // chunks (no clean done frame — the stream just ends, HTTP 200).\n        res.write(JSON.stringify({ error: errorMessage }) + '\\n');\n        res.end();\n        return;\n      }\n      if (chunkDelayMs > 0) {\n        await new Promise((r) => setTimeout(r, chunkDelayMs));\n      }\n      res.write(chunks[i] + '\\n');\n    }\n    // Explicit `after` at or beyond the chunk count: everything, then error.\n    if (errorMessage !== undefined && chunks.length <= errorAt) {\n      res.write(JSON.stringify({ error: errorMessage }) + '\\n');\n    }\n    res.end();\n  }\n}\n\n/**\n * Create an in-memory Ollama mock server — the recommended entry point for\n * consumer test suites. Ephemeral port, built-in liveness + version\n * defaults on, empty request log. Call `await server.start()` before\n * pointing an `OllamaClient` at `server.baseUrl`, and `await server.stop()`\n * in teardown.\n *\n * ```ts\n * const server = createMockOllamaServer();\n * await server.start();\n * // defaults already answer: HEAD / (ping) and GET /api/version\n * server.register('/api/chat', { chunks: mockTextChunks('Hi!') });\n * ```\n */\nexport function createMockOllamaServer(options: MockServerOptions = {}): OllamaMockServer {\n  return new OllamaMockServer(options);\n}\n"]}

package/dist/{types-B57ZZfp3.d.cts → types-DbPnfVh-.d.cts}

@@ -410,11 +410,22 @@ interface CreateRequestOptions extends RequestCancellationOptions {
     readonly modelfile?: string | undefined;
     readonly stream?: boolean | undefined;
     readonly quantize?: string | undefined;
-    /** Quantization level to apply to draft weights during import. */
+    /**
+     * Quantization level to apply to draft weights during import, independently of the
+     * primary `quantize` (e.g. `'nvfp4'`, `'mxfp8'`, `'q8_0'`). Speculative decoding
+     * only — draft tokens are verified, not returned, so aggressive compression of the
+     * draft network is usually safe.
+     */
     readonly draft_quantize?: string | undefined;
     readonly from?: string | undefined;
     readonly files?: Record<string, string> | undefined;
-    /** Draft source file names mapped to their SHA-256 digests. */
+    /**
+     * Draft source file names mapped to their SHA-256 digests — the speculative-decoding
+     * draft network's shards. Ollama runs the small draft network first and batch-verifies
+     * its tokens against the primary weights in a single forward pass. Use
+     * {@link ModelsClient.createModelFromGguf}'s `draftGguf` to upload draft shards from
+     * local paths directly.
+     */
     readonly draft_files?: Record<string, string> | undefined;
     readonly adapters?: Record<string, string> | undefined;
     readonly template?: string | undefined;
@@ -422,7 +433,13 @@ interface CreateRequestOptions extends RequestCancellationOptions {
     readonly parser?: string | undefined;
     readonly license?: string | readonly string[] | undefined;
     readonly system?: string | undefined;
-    /** Modelfile `PARAMETER` instructions, e.g. `{ temperature: 0.7, stop: ['\n'] }`. */
+    /**
+     * Modelfile `PARAMETER` instructions, e.g. `{ temperature: 0.7, stop: ['\n'] }`.
+     * For speculative decoding, set `draft_num_predict` here — how many tokens
+     * the draft network predicts per step (Ollama defaults separate draft
+     * models to 4; `0` disables drafting; models with embedded MTP tensors
+     * require setting it explicitly).
+     */
     readonly parameters?: Record<string, unknown> | undefined;
     readonly messages?: readonly Message[] | undefined;
     /** Minimum Ollama version required by the model. */

package/dist/{types-B57ZZfp3.d.ts → types-DbPnfVh-.d.ts}

@@ -410,11 +410,22 @@ interface CreateRequestOptions extends RequestCancellationOptions {
     readonly modelfile?: string | undefined;
     readonly stream?: boolean | undefined;
     readonly quantize?: string | undefined;
-    /** Quantization level to apply to draft weights during import. */
+    /**
+     * Quantization level to apply to draft weights during import, independently of the
+     * primary `quantize` (e.g. `'nvfp4'`, `'mxfp8'`, `'q8_0'`). Speculative decoding
+     * only — draft tokens are verified, not returned, so aggressive compression of the
+     * draft network is usually safe.
+     */
     readonly draft_quantize?: string | undefined;
     readonly from?: string | undefined;
     readonly files?: Record<string, string> | undefined;
-    /** Draft source file names mapped to their SHA-256 digests. */
+    /**
+     * Draft source file names mapped to their SHA-256 digests — the speculative-decoding
+     * draft network's shards. Ollama runs the small draft network first and batch-verifies
+     * its tokens against the primary weights in a single forward pass. Use
+     * {@link ModelsClient.createModelFromGguf}'s `draftGguf` to upload draft shards from
+     * local paths directly.
+     */
     readonly draft_files?: Record<string, string> | undefined;
     readonly adapters?: Record<string, string> | undefined;
     readonly template?: string | undefined;
@@ -422,7 +433,13 @@ interface CreateRequestOptions extends RequestCancellationOptions {
     readonly parser?: string | undefined;
     readonly license?: string | readonly string[] | undefined;
     readonly system?: string | undefined;
-    /** Modelfile `PARAMETER` instructions, e.g. `{ temperature: 0.7, stop: ['\n'] }`. */
+    /**
+     * Modelfile `PARAMETER` instructions, e.g. `{ temperature: 0.7, stop: ['\n'] }`.
+     * For speculative decoding, set `draft_num_predict` here — how many tokens
+     * the draft network predicts per step (Ollama defaults separate draft
+     * models to 4; `0` disables drafting; models with embedded MTP tensors
+     * require setting it explicitly).
+     */
     readonly parameters?: Record<string, unknown> | undefined;
     readonly messages?: readonly Message[] | undefined;
     /** Minimum Ollama version required by the model. */

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@nemesis-oss/ollama-sdk",
-  "version": "1.13.0",
+  "version": "1.15.0",
   "description": "Production-grade TypeScript SDK for Ollama with native fetch, high availability failover, multi-turn tool calling, structured outputs, reasoning stream tokens, OpenAI/Anthropic bridges, MCP, Web Stream adapters, and a contract-first hybrid architecture that generates TypeScript types, API classes, MCP tools, and Zod schemas from a single canonical IR.",
   "keywords": [
     "ollama",
@@ -59,6 +59,16 @@
         "default": "./dist/skills.cjs"
       }
     },
+    "./testing": {
+      "import": {
+        "types": "./dist/testing.d.ts",
+        "default": "./dist/testing.js"
+      },
+      "require": {
+        "types": "./dist/testing.d.cts",
+        "default": "./dist/testing.cjs"
+      }
+    },
     "./package.json": "./package.json",
     "./mcp/stdio": {
       "import": {
@@ -126,6 +136,9 @@
       "skills": [
         "dist/skills.d.ts"
       ],
+      "testing": [
+        "dist/testing.d.ts"
+      ],
       "mcp/stdio": [
         "dist/mcp-stdio.d.ts"
       ],