summaryrefslogtreecommitdiffhomepage
path: root/src/ollama-client.ts
blob: 0eaae7414a74b1129513f6658ac7aecb13153442 (plain)
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
340
341
342
343
344
345
346
347
348
349
350
351
352
353
354
355
356
357
358
359
360
361
362
363
364
365
366
367
368
369
370
371
372
373
374
375
376
377
378
379
380
381
382
383
384
385
386
387
388
389
390
391
392
393
394
395
396
397
398
399
400
401
402
403
404
405
406
407
408
409
410
411
412
413
414
415
416
417
418
419
420
421
422
423
424
425
426
427
428
429
430
431
432
433
434
435
436
437
438
439
440
441
442
443
444
445
446
447
448
449
450
451
452
453
454
455
456
457
458
459
460
461
462
463
464
465
466
467
468
469
470
471
472
473
474
475
476
477
478
479
480
481
482
483
484
485
486
487
488
489
490
491
492
493
494
495
496
497
498
499
500
501
502
503
504
505
506
507
508
509
510
511
512
513
514
515
516
517
518
519
520
521
522
523
524
525
526
527
528
529
530
531
532
533
534
535
536
537
538
539
540
541
542
543
544
545
546
547
548
549
550
551
552
553
554
555
556
557
558
559
560
561
562
563
564
565
566
567
568
569
570
571
572
573
574
575
576
577
578
579
580
581
582
583
584
585
586
587
588
589
590
591
592
593
594
595
596
597
598
599
600
601
602
603
604
605
606
607
608
609
610
611
612
613
614
615
616
617
618
619
620
621
622
623
624
625
626
627
628
629
630
631
632
633
634
635
636
637
638
639
640
641
642
643
644
645
646
647
648
649
650
651
652
653
654
655
656
657
658
659
660
661
662
663
664
665
666
667
668
669
670
671
672
673
674
675
676
677
678
679
680
681
682
683
684
685
686
687
688
689
690
691
692
693
694
695
696
697
698
699
700
701
702
703
704
705
706
707
708
709
710
711
712
713
714
715
716
717
718
719
720
721
722
723
724
725
726
727
728
729
730
731
732
733
734
735
736
737
738
739
740
741
742
743
744
745
746
747
748
749
750
751
752
753
754
755
756
757
758
759
760
761
762
763
764
765
import { Platform, requestUrl } from "obsidian";
import type { App } from "obsidian";
import type { OllamaToolDefinition } from "./tools";
import { findToolByName } from "./tools";

export interface ChatMessage {
	role: "system" | "user" | "assistant" | "tool";
	content: string;
	tool_calls?: ToolCallResponse[];
	tool_name?: string;
}

export interface ToolCallResponse {
	type?: string;
	function: {
		index?: number;
		name: string;
		arguments: Record<string, unknown>;
	};
}

export interface ToolCallEvent {
	toolName: string;
	friendlyName: string;
	summary: string;
	resultSummary: string;
	args: Record<string, unknown>;
	result: string;
}

/**
 * Approval request event for tools that require user confirmation.
 */
export interface ApprovalRequestEvent {
	toolName: string;
	friendlyName: string;
	message: string;
	args: Record<string, unknown>;
}

export interface ModelOptions {
	temperature?: number;
	num_ctx?: number;
	num_predict?: number;
}

/**
 * Validate that a value looks like a ToolCallResponse[].
 * Ollama returns untyped JSON, so we narrow manually.
 */
function parseToolCalls(value: unknown): ToolCallResponse[] {
	if (!Array.isArray(value)) return [];
	const result: ToolCallResponse[] = [];
	for (const item of value) {
		if (typeof item !== "object" || item === null) continue;
		const obj = item as Record<string, unknown>;
		const fn = obj.function;
		if (typeof fn !== "object" || fn === null) continue;
		const fnObj = fn as Record<string, unknown>;
		if (typeof fnObj.name !== "string") continue;
		result.push({
			...(typeof obj.type === "string" ? { type: obj.type } : {}),
			function: {
				...(typeof fnObj.index === "number" ? { index: fnObj.index } : {}),
				name: fnObj.name,
				arguments: typeof fnObj.arguments === "object" && fnObj.arguments !== null
					? fnObj.arguments as Record<string, unknown>
					: {},
			},
		});
	}
	return result;
}

/**
 * Result returned by a chat request strategy.
 */
interface ChatRequestResult {
	content: string;
	toolCalls: ToolCallResponse[];
}

/**
 * A strategy function that performs a single HTTP request to the Ollama chat API.
 * Different implementations handle non-streaming, mobile fallback, and desktop streaming.
 */
type ChatRequestStrategy = (
	workingMessages: ChatMessage[],
) => Promise<ChatRequestResult>;

/**
 * Options for the shared agent loop.
 */
interface AgentLoopOptions {
	messages: ChatMessage[];
	tools?: OllamaToolDefinition[];
	app?: App;
	userSystemPrompt?: string;
	vaultContext?: string;
	onToolCall?: (event: ToolCallEvent) => void;
	onApprovalRequest?: (event: ApprovalRequestEvent) => Promise<boolean>;
	sendRequest: ChatRequestStrategy;
}

/**
 * System prompt injected when tools are available.
 */
const TOOL_SYSTEM_PROMPT =
	"You are a helpful assistant with access to tools for interacting with an Obsidian vault. " +
	"When you use the search_files tool, the results contain exact file paths. " +
	"You MUST use these exact paths when calling read_file, edit_file, or referencing files. " +
	"NEVER guess or modify file paths — always use the paths returned by search_files or get_current_note verbatim.\n\n" +
	"LINKING TO NOTES — MANDATORY FORMAT:\n" +
	"When referencing any note that exists in the vault, you MUST use Obsidian wiki-link syntax.\n" +
	"FORMAT: [[exact file path without .md extension]]\n" +
	"RULES:\n" +
	"1. ALWAYS use the full vault-relative path minus the .md extension.\n" +
	"   Example: a file at 'projects/2024/my-note.md' MUST be linked as [[projects/2024/my-note]].\n" +
	"2. NEVER use just the basename when the file is inside a subfolder.\n" +
	"   WRONG: [[my-note]]  CORRECT: [[projects/2024/my-note]]\n" +
	"3. For files in the vault root (no folder), use just the name: [[my-note]].\n" +
	"4. NEVER include the .md extension in the link: WRONG: [[my-note.md]]  CORRECT: [[my-note]]\n" +
	"5. To show different display text, use a pipe: [[projects/2024/my-note|My Note]].\n" +
	"6. Get the exact path from search_files, read_file, or get_current_note output, strip the .md extension, and use that as the link target.\n" +
	"7. Link to notes whenever helpful — search results, related notes, files you read or edited. Links let the user click to navigate directly.\n\n" +
	"EDITING FILES — MANDATORY WORKFLOW:\n" +
	"The edit_file tool performs a find-and-replace. You provide old_text (the exact text currently in the file) and new_text (what to replace it with). " +
	"If old_text does not match the file contents exactly, the edit WILL FAIL.\n" +
	"Therefore you MUST follow this sequence every time you edit a file:\n" +
	"1. Get the file path (use search_files or get_current_note).\n" +
	"2. Call read_file to see the CURRENT content of the file.\n" +
	"3. Copy the exact text you want to change from the read_file output and use it as old_text.\n" +
	"4. Call edit_file with the correct old_text and your new_text.\n" +
	"NEVER skip step 2. NEVER guess what the file contains — always read it first.\n" +
	"If the file is empty (read_file returned no content), you may set old_text to an empty string to write initial content.\n" +
	"If the file is NOT empty, old_text MUST NOT be empty — copy the exact passage you want to change from the read_file output.\n" +
	"old_text must include enough surrounding context (a few lines) to uniquely identify the location in the file. " +
	"Preserve the exact whitespace, indentation, and newlines from the read_file output.\n\n" +
	"CREATING FILES:\n" +
	"Use create_file to make new notes. It will fail if the file already exists — use edit_file for existing files. " +
	"Parent folders are created automatically.\n\n" +
	"MOVING/RENAMING FILES:\n" +
	"Use move_file to move or rename a file. All [[wiki-links]] across the vault are automatically updated.\n\n" +
	"SEARCHING FILE CONTENTS:\n" +
	"Use grep_search to find text inside file contents (like grep). " +
	"Use search_files to find files by name/path. Use grep_search to find files containing specific text.\n\n" +
	"FRONTMATTER MANAGEMENT:\n" +
	"When you read a file with read_file, its YAML frontmatter is automatically included as a parsed JSON block at the top of the output. " +
	"Use set_frontmatter to add, update, or remove frontmatter properties (tags, aliases, categories, etc.). " +
	"set_frontmatter is MUCH safer than edit_file for metadata changes \u2014 it preserves YAML formatting. " +
	"ALWAYS prefer set_frontmatter over edit_file when modifying tags, aliases, or other frontmatter fields. " +
	"RECOMMENDED: Read the file first to see existing frontmatter before calling set_frontmatter.\n\n" +
	"Some tools (such as delete_file, edit_file, create_file, and move_file) require user approval before they execute. " +
	"If the user declines an action, ask them why so you can better assist them.\n\n" +
	"BATCH TOOLS:\n" +
	"When you need to perform the same type of operation on multiple files, prefer batch tools over calling individual tools repeatedly. " +
	"Available batch tools: batch_search_files, batch_grep_search, batch_delete_file, batch_move_file, batch_set_frontmatter, batch_edit_file. " +
	"Batch tools accept an array of operations and execute them all in one call, reporting per-item success/failure. " +
	"Batch tools that modify files (delete, move, edit, set_frontmatter) require a single user approval for the entire batch. " +
	"The parameters for batch tools use JSON arrays passed as strings. " +
	"IMPORTANT: For batch_edit_file, you MUST still read each file first to get exact content before editing.";

/**
 * Shared agent loop: injects the system prompt, calls the strategy for each
 * iteration, executes tool calls, and loops until the model returns a final
 * text response or the iteration cap is reached.
 */
async function chatAgentLoop(opts: AgentLoopOptions): Promise<string> {
	const { messages, tools, app, userSystemPrompt, vaultContext, onToolCall, onApprovalRequest, sendRequest } = opts;
	const maxIterations = 10;
	let iterations = 0;

	const workingMessages = messages.map((m) => ({ ...m }));

	// Build combined system prompt from tool instructions + vault context + user custom prompt
	const hasTools = tools !== undefined && tools.length > 0;
	const hasUserPrompt = userSystemPrompt !== undefined && userSystemPrompt.trim() !== "";
	const hasVaultContext = vaultContext !== undefined && vaultContext.trim() !== "";

	if (hasTools || hasUserPrompt || hasVaultContext) {
		const parts: string[] = [];
		if (hasTools) {
			parts.push(TOOL_SYSTEM_PROMPT);
		}
		if (vaultContext !== undefined && vaultContext.trim() !== "") {
			parts.push(vaultContext);
		}
		if (userSystemPrompt !== undefined && userSystemPrompt.trim() !== "") {
			parts.push("USER INSTRUCTIONS:\n" + userSystemPrompt.trim());
		}
		workingMessages.unshift({ role: "system", content: parts.join("\n\n") });
	}

	while (iterations < maxIterations) {
		iterations++;

		const { content, toolCalls } = await sendRequest(workingMessages);

		// No tool calls — return the final content
		if (toolCalls.length === 0) {
			return content;
		}

		// Append assistant message with tool_calls to working history
		workingMessages.push({
			role: "assistant",
			content,
			tool_calls: toolCalls,
		});

		if (app === undefined) {
			throw new Error("App reference required for tool execution.");
		}

		// Execute each tool call and append results
		for (const tc of toolCalls) {
			const fnName = tc.function.name;
			const fnArgs = tc.function.arguments;
			const toolEntry = findToolByName(fnName);

			let result: string;
			if (toolEntry === undefined) {
				result = `Error: Unknown tool "${fnName}".`;
			} else if (toolEntry.requiresApproval) {
				let approved = false;
				if (onApprovalRequest !== undefined) {
					const message = toolEntry.approvalMessage !== undefined
						? toolEntry.approvalMessage(fnArgs)
						: `Allow ${toolEntry.friendlyName}?`;
					approved = await onApprovalRequest({
						toolName: fnName,
						friendlyName: toolEntry.friendlyName,
						message,
						args: fnArgs,
					});
				}
				result = approved
					? await toolEntry.execute(app, fnArgs)
					: `Action declined by user: ${toolEntry.friendlyName} was not approved.`;
			} else {
				result = await toolEntry.execute(app, fnArgs);
			}

			if (onToolCall !== undefined) {
				const friendlyName = toolEntry !== undefined ? toolEntry.friendlyName : fnName;
				const summary = toolEntry !== undefined ? toolEntry.summarize(fnArgs) : `Called ${fnName}`;
				const resultSummary = toolEntry !== undefined ? toolEntry.summarizeResult(result) : "";
				onToolCall({ toolName: fnName, friendlyName, summary, resultSummary, args: fnArgs, result });
			}

			workingMessages.push({
				role: "tool",
				tool_name: fnName,
				content: result,
			});
		}

		// Loop continues — model sees tool results
	}

	throw new Error("Tool calling loop exceeded maximum iterations.");
}

// ---------------------------------------------------------------------------
// Utility functions
// ---------------------------------------------------------------------------

export async function testConnection(ollamaUrl: string): Promise<string> {
	try {
		const response = await requestUrl({
			url: `${ollamaUrl}/api/version`,
			method: "GET",
			throw: false,
		});

		if (response.status === 200) {
			const version = (response.json as Record<string, unknown>).version;
			if (typeof version === "string") {
				return version;
			}
			throw new Error("Unexpected response format: missing version field.");
		}

		throw new Error(`Ollama returned status ${response.status}.`);
	} catch (err: unknown) {
		if (err instanceof Error) {
			const msg = err.message.toLowerCase();
			if (msg.includes("net") || msg.includes("fetch") || msg.includes("failed to fetch") || msg.includes("load failed")) {
				if (Platform.isMobile) {
					throw new Error(
						"Ollama is unreachable. On mobile, use your computer's LAN IP " +
						"(e.g. http://192.168.1.x:11434) instead of localhost."
					);
				}
				throw new Error("Ollama is unreachable. Is the server running?");
			}
			throw err;
		}
		throw new Error("Ollama is unreachable. Is the server running?");
	}
}

export async function listModels(ollamaUrl: string): Promise<string[]> {
	try {
		const response = await requestUrl({
			url: `${ollamaUrl}/api/tags`,
			method: "GET",
		});

		const models = (response.json as Record<string, unknown>).models;
		if (!Array.isArray(models)) {
			throw new Error("Unexpected response format: missing models array.");
		}

		return models.map((m: unknown) => {
			if (typeof m === "object" && m !== null && "name" in m) {
				const name = (m as Record<string, unknown>).name;
				if (typeof name === "string") {
					return name;
				}
				return String(name);
			}
			return String(m);
		});
	} catch (err: unknown) {
		if (err instanceof Error) {
			throw new Error(`Failed to list models: ${err.message}`);
		}
		throw new Error("Failed to list models: unknown error.");
	}
}

/**
 * Model info returned by /api/show.
 */
export interface ModelInfo {
	contextLength: number;
}

/**
 * Query Ollama for model details, extracting the context length.
 * The context length is found in model_info under keys like
 * "<family>.context_length" or "context_length".
 */
export async function showModel(ollamaUrl: string, model: string): Promise<ModelInfo> {
	try {
		const response = await requestUrl({
			url: `${ollamaUrl}/api/show`,
			method: "POST",
			headers: { "Content-Type": "application/json" },
			body: JSON.stringify({ model }),
		});

		const json = response.json as Record<string, unknown>;
		let contextLength = 4096; // fallback default

		const modelInfo = json.model_info as Record<string, unknown> | undefined;
		if (modelInfo !== undefined && modelInfo !== null) {
			for (const key of Object.keys(modelInfo)) {
				if (key.endsWith(".context_length") || key === "context_length") {
					const val = modelInfo[key];
					if (typeof val === "number" && val > 0) {
						contextLength = val;
						break;
					}
				}
			}
		}

		return { contextLength };
	} catch (err: unknown) {
		if (err instanceof Error) {
			throw new Error(`Failed to get model info: ${err.message}`);
		}
		throw new Error("Failed to get model info: unknown error.");
	}
}

// ---------------------------------------------------------------------------
// Non-streaming chat (requestUrl, no UI callbacks)
// ---------------------------------------------------------------------------

/**
 * Options for a non-streaming chat request.
 */
export interface ChatMessageOptions {
	ollamaUrl: string;
	model: string;
	messages: ChatMessage[];
	tools?: OllamaToolDefinition[];
	app?: App;
	onToolCall?: (event: ToolCallEvent) => void;
	onApprovalRequest?: (event: ApprovalRequestEvent) => Promise<boolean>;
	userSystemPrompt?: string;
	vaultContext?: string;
}

/**
 * Send a chat message with optional tool-calling agent loop.
 * When tools are provided, the function handles the multi-turn tool
 * execution loop automatically and calls onToolCall for each invocation.
 */
export async function sendChatMessage(
	opts: ChatMessageOptions,
): Promise<string> {
	const { ollamaUrl, model, tools, app, userSystemPrompt, vaultContext, onToolCall, onApprovalRequest } = opts;

	const sendRequest: ChatRequestStrategy = async (workingMessages) => {
		const body: Record<string, unknown> = {
			model,
			messages: workingMessages,
			stream: false,
		};

		if (tools !== undefined && tools.length > 0) {
			body.tools = tools;
		}

		try {
			const response = await requestUrl({
				url: `${ollamaUrl}/api/chat`,
				method: "POST",
				headers: { "Content-Type": "application/json" },
				body: JSON.stringify(body),
			});

			const messageObj = (response.json as Record<string, unknown>).message;
			if (typeof messageObj !== "object" || messageObj === null) {
				throw new Error("Unexpected response format: missing message.");
			}

			const msg = messageObj as Record<string, unknown>;
			const content = typeof msg.content === "string" ? msg.content : "";
			const toolCalls = parseToolCalls(msg.tool_calls);

			return { content, toolCalls };
		} catch (err: unknown) {
			if (err instanceof DOMException && err.name === "AbortError") {
				throw err;
			}
			if (err instanceof Error) {
				throw new Error(`Chat request failed: ${err.message}`);
			}
			throw new Error("Chat request failed: unknown error.");
		}
	};

	return chatAgentLoop({
		messages: opts.messages,
		tools,
		app,
		userSystemPrompt,
		vaultContext,
		onToolCall,
		onApprovalRequest,
		sendRequest,
	});
}

// ---------------------------------------------------------------------------
// Streaming chat
// ---------------------------------------------------------------------------

/**
 * Streaming chat options.
 */
export interface StreamingChatOptions {
	ollamaUrl: string;
	model: string;
	messages: ChatMessage[];
	tools?: OllamaToolDefinition[];
	app?: App;
	options?: ModelOptions;
	userSystemPrompt?: string;
	vaultContext?: string;
	onChunk: (text: string) => void;
	onToolCall?: (event: ToolCallEvent) => void;
	onApprovalRequest?: (event: ApprovalRequestEvent) => Promise<boolean>;
	onCreateBubble: () => void;
	abortSignal?: AbortSignal;
}

/**
 * Parse ndjson lines from a streamed response body.
 * Handles partial lines that may span across chunks from the reader.
 */
async function* readNdjsonStream(
	reader: ReadableStreamDefaultReader<Uint8Array>,
	decoder: TextDecoder,
): AsyncGenerator<Record<string, unknown>> {
	let buffer = "";

	while (true) {
		const { done, value } = await reader.read();
		if (done) break;

		buffer += decoder.decode(value, { stream: true });
		const lines = buffer.split("\n");
		// Last element may be incomplete — keep it in buffer
		buffer = lines.pop() ?? "";

		for (const line of lines) {
			const trimmed = line.trim();
			if (trimmed === "") continue;
			yield JSON.parse(trimmed) as Record<string, unknown>;
		}
	}

	// Process any remaining data in buffer
	const trimmed = buffer.trim();
	if (trimmed !== "") {
		yield JSON.parse(trimmed) as Record<string, unknown>;
	}
}

/**
 * Wraps an async generator with a per-iteration idle timeout.
 * The timer resets on every yielded value. If no value arrives
 * within `timeoutMs`, an error is thrown.
 *
 * This handles cold model starts (long initial load) as well as
 * mid-stream stalls where the connection goes silent.
 */
async function* withIdleTimeout<T>(
	source: AsyncGenerator<T>,
	timeoutMs: number,
): AsyncGenerator<T> {
	while (true) {
		const result = await Promise.race([
			source.next(),
			new Promise<never>((_resolve, reject) => {
				setTimeout(() => {
					reject(new Error(
						`No response from Ollama for ${Math.round(timeoutMs / 1000)} seconds. ` +
						"The model may still be loading — try again in a moment.",
					));
				}, timeoutMs);
			}),
		]);

		if (result.done === true) return;
		yield result.value;
	}
}

/**
 * Send a chat message with streaming.
 * Streams text chunks via onChunk callback. Supports tool-calling agent loop.
 * Returns the full accumulated response text.
 *
 * On mobile platforms, falls back to non-streaming via Obsidian's requestUrl()
 * because native fetch() cannot reach local network addresses from the mobile
 * WebView sandbox.
 */
export async function sendChatMessageStreaming(
	opts: StreamingChatOptions,
): Promise<string> {
	const { ollamaUrl, model, tools, app, options, userSystemPrompt, vaultContext, onChunk, onToolCall, onApprovalRequest, onCreateBubble, abortSignal } = opts;

	const sendRequest: ChatRequestStrategy = Platform.isMobile
		? buildMobileStrategy(ollamaUrl, model, tools, options, onChunk, onCreateBubble, abortSignal)
		: buildDesktopStreamingStrategy(ollamaUrl, model, tools, options, onChunk, onCreateBubble, abortSignal);

	return chatAgentLoop({
		messages: opts.messages,
		tools,
		app,
		userSystemPrompt,
		vaultContext,
		onToolCall,
		onApprovalRequest,
		sendRequest,
	});
}

/**
 * Mobile strategy: uses Obsidian's requestUrl() (non-streaming) so the request
 * goes through the native networking layer and can reach localhost / LAN.
 * Delivers the full response as a single chunk.
 *
 * Since requestUrl() cannot be natively aborted, we race it against the
 * AbortSignal and check the signal before delivering content.
 */
function buildMobileStrategy(
	ollamaUrl: string,
	model: string,
	tools: OllamaToolDefinition[] | undefined,
	options: ModelOptions | undefined,
	onChunk: (text: string) => void,
	onCreateBubble: () => void,
	abortSignal?: AbortSignal,
): ChatRequestStrategy {
	return async (workingMessages) => {
		// Bail out immediately if already aborted
		if (abortSignal?.aborted === true) {
			throw new DOMException("The operation was aborted.", "AbortError");
		}

		onCreateBubble();

		const body: Record<string, unknown> = {
			model,
			messages: workingMessages,
			stream: false,
		};

		if (tools !== undefined && tools.length > 0) {
			body.tools = tools;
		}

		if (options !== undefined) {
			body.options = options;
		}

		try {
			// Race requestUrl against the abort signal so the user gets
			// immediate feedback even though the HTTP request completes
			// in the background.
			const requestPromise = requestUrl({
				url: `${ollamaUrl}/api/chat`,
				method: "POST",
				headers: { "Content-Type": "application/json" },
				body: JSON.stringify(body),
			});

			let response: Awaited<ReturnType<typeof requestUrl>>;
			if (abortSignal !== undefined) {
				const abortPromise = new Promise<never>((_resolve, reject) => {
					if (abortSignal.aborted) {
						reject(new DOMException("The operation was aborted.", "AbortError"));
						return;
					}
					abortSignal.addEventListener("abort", () => {
						reject(new DOMException("The operation was aborted.", "AbortError"));
					}, { once: true });
				});
				response = await Promise.race([requestPromise, abortPromise]);
			} else {
				response = await requestPromise;
			}

			const messageObj = (response.json as Record<string, unknown>).message;
			if (typeof messageObj !== "object" || messageObj === null) {
				throw new Error("Unexpected response format: missing message.");
			}

			const msg = messageObj as Record<string, unknown>;
			const content = typeof msg.content === "string" ? msg.content : "";
			const toolCalls = parseToolCalls(msg.tool_calls);

			// Check abort before delivering content to the UI
			if (abortSignal?.aborted === true) {
				throw new DOMException("The operation was aborted.", "AbortError");
			}

			if (content !== "") {
				onChunk(content);
			}

			return { content, toolCalls };
		} catch (err: unknown) {
			if (err instanceof DOMException && err.name === "AbortError") {
				throw err;
			}
			if (err instanceof Error) {
				const errMsg = err.message.toLowerCase();
				if (errMsg.includes("net") || errMsg.includes("fetch") || errMsg.includes("load") || errMsg.includes("failed")) {
					throw new Error(
						`Cannot reach Ollama at ${ollamaUrl}. ` +
						"On mobile, Ollama must be accessible over your network (not localhost). " +
						"Set the Ollama URL to your computer's LAN IP (e.g. http://192.168.1.x:11434)."
					);
				}
				throw new Error(`Chat request failed: ${err.message}`);
			}
			throw new Error("Chat request failed: unknown error.");
		}
	};
}

/**
 * Desktop streaming strategy: uses native fetch() for real token-by-token streaming.
 */
function buildDesktopStreamingStrategy(
	ollamaUrl: string,
	model: string,
	tools: OllamaToolDefinition[] | undefined,
	options: ModelOptions | undefined,
	onChunk: (text: string) => void,
	onCreateBubble: () => void,
	abortSignal?: AbortSignal,
): ChatRequestStrategy {
	return async (workingMessages) => {
		onCreateBubble();

		const body: Record<string, unknown> = {
			model,
			messages: workingMessages,
			stream: true,
		};

		if (tools !== undefined && tools.length > 0) {
			body.tools = tools;
		}

		if (options !== undefined) {
			body.options = options;
		}

		const response = await fetch(`${ollamaUrl}/api/chat`, {
			method: "POST",
			headers: { "Content-Type": "application/json" },
			body: JSON.stringify(body),
			signal: abortSignal,
		});

		if (!response.ok) {
			throw new Error(`Ollama returned status ${response.status}.`);
		}

		if (response.body === null) {
			throw new Error("Response body is null \u2014 streaming not supported.");
		}

		const reader = response.body.getReader();
		const decoder = new TextDecoder();

		let content = "";
		const toolCalls: ToolCallResponse[] = [];

		// 5 minute idle timeout per chunk — generous enough for cold model
		// loads, but catches silent connection drops.
		const IDLE_TIMEOUT_MS = 300_000;

		try {
			for await (const chunk of withIdleTimeout(readNdjsonStream(reader, decoder), IDLE_TIMEOUT_MS)) {
				// Check for mid-stream errors from Ollama
				if (typeof chunk.error === "string") {
					throw new Error(`Ollama error: ${chunk.error}`);
				}

				const rawMsg: unknown = chunk.message;
				const msg = typeof rawMsg === "object" && rawMsg !== null
					? rawMsg as Record<string, unknown>
					: undefined;
				if (msg !== undefined) {
					if (typeof msg.content === "string" && msg.content !== "") {
						content += msg.content;
						onChunk(msg.content);
					}
					if (Array.isArray(msg.tool_calls)) {
						toolCalls.push(...parseToolCalls(msg.tool_calls));
					}
				}
			}
		} catch (err: unknown) {
			if (err instanceof DOMException && err.name === "AbortError") {
				throw err;
			}
			throw err;
		}

		return { content, toolCalls };
	};
}