feat: add Ollama Cloud compatibility

README.md

@@ -54,6 +54,7 @@ You can also set `commitollama.model` manually in VS Code settings.
 | `commitollama.useLowerCase` | Lowercase the first letter of the commit message. | `false` |
 | `commitollama.language` | Language preset (`English`, `Spanish`, `Custom`, …). | `English` |
 | `commitollama.promptTemperature` | Model temperature (`0`–`1`). Higher = more creative. | `0.2` |
+| `commitollama.cloudCompatibilityMode` | Always use JSON-only prompting for Ollama Cloud models that do not support Ollama structured output. Commitollama also falls back to this mode automatically after an invalid structured response. | `false` |
 | `commitollama.commitTemplate` | Final commit format. Placeholders: `{{type}}`, `{{emoji}}`, `{{message}}`. | `{{type}} {{emoji}}: {{message}}` |
 
 ### Custom overrides
@@ -69,6 +70,20 @@ You can also set `commitollama.model` manually in VS Code settings.
 | `commitollama.custom.descriptionPrompt` | Custom prompt for the commit description. |
 | `commitollama.custom.requestHeaders` | Extra HTTP headers for Ollama requests (e.g. auth). |
 
+### Ollama Cloud compatibility
+
+For Ollama Cloud, configure its endpoint and authentication headers. Commitollama automatically retries a failed or incomplete structured response as JSON-only text, extracts a JSON object, and validates it against the commit schema. Enable `commitollama.cloudCompatibilityMode` to use this compatibility path on the first request instead.
+
+```json
+{
+  "commitollama.custom.endpoint": "<your Ollama Cloud endpoint>",
+  "commitollama.custom.requestHeaders": {
+    "Authorization": "Bearer <your token>"
+  },
+  "commitollama.cloudCompatibilityMode": true
+}
+```
+
 Example emoji map:
 
 ```json

package.json

@@ -115,11 +115,17 @@
 						"default": 0.2,
 						"order": 7
 					},
+					"commitollama.cloudCompatibilityMode": {
+						"type": "boolean",
+						"description": "Use JSON-only compatibility mode immediately for Ollama Cloud models that do not support structured output. Failed or incomplete structured responses automatically retry with this mode.",
+						"default": false,
+						"order": 8
+					},
 					"commitollama.commitTemplate": {
 						"type": "string",
 						"description": "Custom template for commit messages.",
 						"default": "{{type}} {{emoji}}: {{message}}",
-						"order": 8
+						"order": 9
 					},
 					"commitollama.custom.language": {
 						"type": "string",

sampleWorkspace/.vscode/settings.json

@@ -3,5 +3,6 @@
     "commitollama.useDescription": false,
     "commitollama.useLowerCase": false,
     "commitollama.commitTemplate": "{{type}} {{emoji}}: {{message}}",
-    "commitollama.custom.emojis": {}
+    "commitollama.custom.emojis": {},
+    "commitollama.cloudCompatibilityMode": false
 }

src/config.ts

@@ -17,6 +17,7 @@ export const defaultConfig = {
 	language: Languages.English,
 	commitTemplate: '{{type}} {{emoji}}: {{message}}',
 	promptTemperature: 0.2,
+	cloudCompatibilityMode: false,
 	requestHeaders: {},
 	emojis: {
 		feat: '✨',
@@ -78,6 +79,9 @@ class Config {
 
 		const promptTemperature =
 			getConfig('promptTemperature') || defaultConfig.promptTemperature
+		const cloudCompatibilityMode =
+			getConfig('cloudCompatibilityMode') ??
+			defaultConfig.cloudCompatibilityMode
 
 		const customPrompt = getConfig('custom.prompt')
 		const customTypeRules = getConfig('custom.typeRules')
@@ -94,6 +98,7 @@ class Config {
 
 		return {
 			commitEmojis,
+			cloudCompatibilityMode,
 			promptTemperature,
 			commitTemplate,
 			customCommitMessageRules,

src/generator.ts

@@ -1,12 +1,15 @@
 import * as vscode from 'vscode'
+import { ZodError } from 'zod'
 import * as ai from './ai'
 import { isLowQualityCommit } from './commitQuality'
 import { config } from './config'
 import { OLLAMA_LIBRARY_URL } from './constants'
+import { buildCommitModelOptions } from './modelOptions'
 import { createChatAdapter } from './ollamaAdapter'
 import {
 	buildCommitSchema,
 	type CommitStructure,
+	extractJsonObject,
 	parseCommitResponse,
 } from './schemas/commit'
 import { formatExtensionError, logExtensionError } from './security/log'
@@ -55,6 +58,15 @@ function isStructuredOutputCompatibilityError(error: unknown): boolean {
 	)
 }
 
+function isCloudResponseValidationError(error: unknown): boolean {
+	return (
+		error instanceof ZodError ||
+		(error instanceof Error &&
+			(error.message === 'Could not find a JSON object in the model response' ||
+				isStructuredOutputCompatibilityError(error)))
+	)
+}
+
 function formatChangeSummaries(summaries: ChangeSummary[]): string {
 	return summaries
 		.map(({ file, summary }) => `- ${file}: ${summary}`)
@@ -93,6 +105,7 @@ function buildStructuredPrompt(options: {
 	descriptionPrompt: string
 	customPrompt?: string
 	extraInstruction?: string
+	jsonOnly?: boolean
 }): string {
 	const {
 		typeRules,
@@ -102,6 +115,7 @@ function buildStructuredPrompt(options: {
 		descriptionPrompt,
 		customPrompt,
 		extraInstruction,
+		jsonOnly,
 	} = options
 
 	const basePrompt =
@@ -126,29 +140,40 @@ function buildStructuredPrompt(options: {
 	${useDescription ? descriptionPrompt : ''}
 	Respond using JSON`
 
-	if (!extraInstruction) {
+	const instructions = [
+		extraInstruction,
+		jsonOnly
+			? `Respond with exactly one valid JSON object and nothing else. Do not use Markdown, code fences, or explanatory text. Required fields: ${useDescription ? '{"type":"feat","message":"Brief commit subject","summary":"One to three sentence commit description"}' : '{"type":"feat","message":"Brief commit subject"}'}. The type value must be one of: feat, fix, docs, style, test, chore, revert, refactor.`
+			: undefined,
+	].filter(Boolean)
+
+	if (instructions.length === 0) {
 		return basePrompt
 	}
 
-	return `${basePrompt}\n\n${extraInstruction}`
+	return `${basePrompt}\n\n${instructions.join('\n\n')}`
 }
 
 async function requestStructuredCommit(
 	summaries: ChangeSummary[],
 	options?: {
 		extraInstruction?: string
 		branchName?: string | null
+		jsonOnly?: boolean
 	},
 ): Promise<CommitStructure> {
 	const {
+		model,
 		promptTemperature,
 		language,
 		useDescription,
 		customPrompt,
 		customTypeRules,
 		customCommitMessageRules,
 		customDescriptionPrompt,
+		cloudCompatibilityMode,
 	} = config.inference
+	const jsonOnly = options?.jsonOnly ?? cloudCompatibilityMode
 
 	const typeRules =
 		customTypeRules ||
@@ -180,28 +205,34 @@ async function requestStructuredCommit(
 		descriptionPrompt,
 		customPrompt,
 		extraInstruction: options?.extraInstruction,
+		jsonOnly,
 	})
 
 	const outputSchema = buildCommitSchema(useDescription, language)
 
-	const result = await ai.chat({
+	const chatOptions = {
 		adapter: createChatAdapter(),
 		systemPrompts: [structuredPrompt],
 		messages: [
 			{
-				role: 'user',
+				role: 'user' as const,
 				content: buildCommitUserContent(summaries, options?.branchName),
 			},
 		],
-		outputSchema,
-		modelOptions: {
-			options: {
-				temperature: promptTemperature,
-				num_predict: 256,
-			},
-			think: false,
-		} as never,
-	})
+		modelOptions: buildCommitModelOptions(model, promptTemperature),
+	}
+
+	if (jsonOnly) {
+		const result = await ai.chat({ ...chatOptions, stream: false })
+		const commit = parseCommitResponse(
+			extractJsonObject(result),
+			useDescription,
+			language,
+		)
+		return commit
+	}
+
+	const result = await ai.chat({ ...chatOptions, outputSchema })
 
 	return parseCommitResponse(result, useDescription, language)
 }
@@ -211,11 +242,36 @@ export async function generateStructuredCommit(
 	branchName?: string | null,
 ): Promise<CommitStructure> {
 	try {
-		let commit = await requestStructuredCommit(summaries, { branchName })
+		let commit: CommitStructure
+		let jsonOnly = config.inference.cloudCompatibilityMode
+
+		try {
+			commit = await requestStructuredCommit(summaries, {
+				branchName,
+				jsonOnly,
+			})
+		} catch (error) {
+			if (!isCloudResponseValidationError(error)) {
+				throw error
+			}
+
+			// Cloud models frequently accept chat requests but either reject
+			// Ollama's structured-output format or omit an optional-looking field
+			// such as the enabled commit description. Retry as JSON-only text even
+			// when the compatibility setting was not enabled in advance.
+			jsonOnly = true
+			commit = await requestStructuredCommit(summaries, {
+				branchName,
+				jsonOnly,
+				extraInstruction:
+					'Your previous response was invalid. Return exactly one JSON object with every required field from the requested format.',
+			})
+		}
 
 		if (isLowQualityCommit(commit.message, commit.type)) {
 			commit = await requestStructuredCommit(summaries, {
 				branchName,
+				jsonOnly,
 				extraInstruction:
 					'The previous response was invalid because it did not describe the staged changes. Use the summaries exactly and describe the real code changes.',
 			})

src/modelOptions.ts

@@ -0,0 +1,51 @@
+export function isGptOssModel(model: string): boolean {
+	return /^gpt-oss(?::|$)/i.test(model)
+}
+
+export interface OllamaGenerationOptions {
+	temperature: number
+	num_predict?: number
+}
+
+export interface OllamaModelOptions {
+	options: OllamaGenerationOptions
+	think?: 'low' | false
+}
+
+export function buildCommitModelOptions(
+	model: string,
+	promptTemperature: number,
+): OllamaModelOptions {
+	const isGptOss = isGptOssModel(model)
+
+	return {
+		options: {
+			temperature: promptTemperature,
+			// GPT-OSS cannot disable reasoning. Give its low-effort trace enough
+			// room to finish before it emits the short, visible JSON response.
+			num_predict: isGptOss ? 2048 : 256,
+		},
+		think: isGptOss ? 'low' : false,
+	}
+}
+
+export function buildSummarizeModelOptions(
+	model: string,
+	promptTemperature: number,
+): OllamaModelOptions {
+	if (!isGptOssModel(model)) {
+		return {
+			options: { temperature: promptTemperature },
+		}
+	}
+
+	return {
+		options: {
+			temperature: promptTemperature,
+			// The summarize adapter otherwise maps maxLength directly to an
+			// 80-token generation cap, which GPT-OSS can exhaust on reasoning.
+			num_predict: 1024,
+		},
+		think: 'low',
+	}
+}

src/ollamaAdapter.ts

@@ -1,5 +1,9 @@
-import type { AnySummarizeAdapter, AnyTextAdapter } from '@tanstack/ai'
-import { createOllamaChat, createOllamaSummarize } from '@tanstack/ai-ollama'
+import {
+	ChatStreamSummarizeAdapter,
+	type AnySummarizeAdapter,
+	type AnyTextAdapter,
+} from '@tanstack/ai/adapters'
+import { createOllamaChat } from '@tanstack/ai-ollama'
 import { config } from './config'
 
 export interface OllamaConnectionConfig {
@@ -38,6 +42,15 @@ export function createChatAdapter(): AnyTextAdapter {
 
 export function createSummarizeAdapter() {
 	const { model } = config.inference
-	const { host } = getOllamaConnectionConfig()
-	return createOllamaSummarize(model, host) as unknown as AnySummarizeAdapter
+	const { host, headers } = getOllamaConnectionConfig()
+
+	// createOllamaSummarize accepts only a host, so it silently loses custom
+	// headers (including the authorization header required by Ollama Cloud).
+	// Wrap the configured chat adapter instead so summaries and commit messages
+	// use the identical connection configuration.
+	return new ChatStreamSummarizeAdapter(
+		createOllamaChat(model, { host, headers }),
+		model,
+		'ollama',
+	) as unknown as AnySummarizeAdapter
 }