fix(ai): openai-completions - throw error on missing finish-reason

- require \ before treating \ streams as successful
- add regression coverage for truncated streams without \
- closes #4345

packages/ai/CHANGELOG.md

@@ -9,6 +9,7 @@
 ### Fixed
 
 - Fixed GitHub Copilot model availability to ignore generic `GH_TOKEN` and `GITHUB_TOKEN` environment variables, requiring OAuth login or `COPILOT_GITHUB_TOKEN` instead ([#4485](https://github.com/earendil-works/pi/issues/4485)).
+- Fixed `openai-completions` streams to surface an error when the stream ends before any terminal `finish_reason`, so truncated responses can retry instead of being accepted as success ([#4345](https://github.com/earendil-works/pi/issues/4345)).
 - Fixed Bedrock proxy handling to preserve `NO_PROXY` exclusions while using HTTP(S)-only proxy agents.
 - Fixed GitHub Copilot Claude test coverage to use the current Claude Sonnet 4.6 model ID.
 - Fixed OpenAI Responses requests for models that support disabling reasoning to send `reasoning.effort: "none"` when thinking is off.

packages/ai/src/providers/openai-completions.ts

@@ -165,6 +165,7 @@ export const streamOpenAICompletions: StreamFunction<"openai-completions", OpenA
 
 			let textBlock: TextContent | null = null;
 			let thinkingBlock: ThinkingContent | null = null;
+			let hasFinishReason = false;
 			const toolCallBlocksByIndex = new Map<number, StreamingToolCallBlock>();
 			const toolCallBlocksById = new Map<string, StreamingToolCallBlock>();
 			const blocks = output.content as StreamingBlock[];
@@ -288,6 +289,7 @@ export const streamOpenAICompletions: StreamFunction<"openai-completions", OpenA
 					if (finishReasonResult.errorMessage) {
 						output.errorMessage = finishReasonResult.errorMessage;
 					}
+					hasFinishReason = true;
 				}
 
 				if (choice.delta) {
@@ -390,6 +392,9 @@ export const streamOpenAICompletions: StreamFunction<"openai-completions", OpenA
 			if (output.stopReason === "error") {
 				throw new Error(output.errorMessage || "Provider returned an error stop reason");
 			}
+			if (!hasFinishReason) {
+				throw new Error("Stream ended without finish_reason");
+			}
 
 			stream.push({ type: "done", reason: output.stopReason, message: output });
 			stream.end();

packages/ai/test/openai-completions-tool-choice.test.ts

@@ -441,6 +441,38 @@ describe("openai-completions tool_choice", () => {
 		expect(response.content).toEqual([{ type: "text", text: "OK" }]);
 	});
 
+	it("errors when a stream ends after only null finish_reason chunks", async () => {
+		mockState.chunks = [
+			{
+				id: "chatcmpl-truncated",
+				choices: [{ delta: { content: "partial answer" }, finish_reason: null }],
+			},
+			{
+				id: "chatcmpl-truncated",
+				choices: [{ delta: { content: "partial answer" }, finish_reason: null }],
+			},
+		];
+
+		const { compat: _compat, ...baseModel } = getModel("openai", "gpt-4o-mini")!;
+		const model = { ...baseModel, api: "openai-completions" } as const;
+		const response = await streamSimple(
+			model,
+			{
+				messages: [
+					{
+						role: "user",
+						content: "Reply with a longer sentence",
+						timestamp: Date.now(),
+					},
+				],
+			},
+			{ apiKey: "test" },
+		).result();
+
+		expect(response.stopReason).toBe("error");
+		expect(response.errorMessage).toBe("Stream ended without finish_reason");
+	});
+
 	it("coalesces tool call deltas by stable index when provider mutates ids mid-stream", async () => {
 		mockState.chunks = [
 			{