@oh-my-pi/pi-ai 18.8.7 → 18.8.9

package/CHANGELOG.md

@@ -2,6 +2,21 @@
 
 ## [Unreleased]
 
+## [18.8.9] - 2026-10-10
+
+### Changed
+
+- Updated state field rendering to trim trailing whitespace and normalize empty collection formatting for consistent prompt output
+
+## [18.8.8] - 2026-10-10
+
+### Fixed
+
+- Fixed OpenAI Responses and Codex tool calls running with empty `{}` arguments when a compatible host or proxy ends the call with an empty arguments payload after streaming the real ones ([#15099](https://github.com/can1357/oh-my-pi/pull/15099) by [@will-bogusz](https://github.com/will-bogusz))
+- Fixed truncated Factory Droid Gemini responses ending as a hard error or a half-received tool call instead of being retried ([#15103](https://github.com/can1357/oh-my-pi/pull/15103) by [@will-bogusz](https://github.com/will-bogusz))
+- Fixed Factory Droid Gemini ignoring an error the server reports mid-response, which hid its status and retried errors that cannot succeed ([#15103](https://github.com/can1357/oh-my-pi/pull/15103) by [@will-bogusz](https://github.com/will-bogusz))
+- Fixed resumed OpenAI Responses sessions (xAI, Factory Droid, OpenAI and other hosts) dropping all earlier encrypted reasoning on their first request; GitHub Copilot still rebuilds history until its first response ([#15148](https://github.com/can1357/oh-my-pi/pull/15148) by [@will-bogusz](https://github.com/will-bogusz))
+
 ## [18.8.7] - 2026-10-09
 
 ### Added

package/dist/types/providers/google-shared.d.ts

@@ -153,6 +153,13 @@ export declare function pushToolCallEvents(toolCall: ToolCall, contentIndex: num
 export declare function startTextOrThinkingBlock(isThinking: true, output: AssistantMessage, stream: AssistantMessageEventStream, onBeforeStartEvent?: () => void): ThinkingContent;
 export declare function startTextOrThinkingBlock(isThinking: false, output: AssistantMessage, stream: AssistantMessageEventStream, onBeforeStartEvent?: () => void): TextContent;
 export declare function startTextOrThinkingBlock(isThinking: boolean, output: AssistantMessage, stream: AssistantMessageEventStream, onBeforeStartEvent?: () => void): TextContent | ThinkingContent;
+/**
+ * Error for an in-band `error` frame in a 200 generateContent stream. Keeps the
+ * RPC status alongside the message: an in-band quota failure is classified
+ * from this text, and `RESOURCE_EXHAUSTED` is the only account-exhaustion
+ * signal some of these chunks carry (#13090).
+ */
+export declare function googleStreamChunkError(error: NonNullable<GenerateContentResponse["error"]>, provider: string): Error;
 /**
  * Drives the chunked `generateContentStream` iterator into an `AssistantMessage` and
  * the corresponding `AssistantMessageEventStream`. Shared between `streamGoogle` and

package/dist/types/providers/openai-responses.d.ts

@@ -82,7 +82,14 @@ export declare function pollOpenAIResponsesResultForCompletion(args: {
     wait?: (ms: number, signal?: AbortSignal) => Promise<void>;
 }): Promise<Record<string, unknown> | undefined>;
 interface OpenAIResponsesProviderSessionState extends ProviderSessionState, OpenAIStrictToolsState, OpenAIReasoningEffortFallbackState {
+    /**
+     * Replay native history items. Starts false only on connection-bound hosts;
+     * `close()` clears it so the next request rebuilds history from message
+     * content; the first successful response sets it.
+     */
     nativeHistoryReplayWarmed: boolean;
+    /** Bumped by `close()`; a response warms the state only if no close happened since its request started. */
+    closeCount: number;
     /** Stateful `previous_response_id` chain baselines, keyed by baseUrl/model/session. */
     chains: Map<string, OpenAIResponsesChainState>;
     /** `configuration_update` effort baselines, keyed by baseUrl/model/session. */

package/dist/types/providers/openai-shared.d.ts

@@ -622,7 +622,8 @@ export declare function accumulateToolCallArgumentsDelta(block: ResponsesToolCal
  * Finalize streamed function-call arguments from the authoritative `.done`
  * payload. The caller owns the `argumentsDone` flag (generic Responses sets it;
  * Codex's block shape has no such field), so this only rewrites `arguments` and
- * drops the transient accumulation fields.
+ * drops the transient accumulation fields. A blank payload is not authoritative:
+ * it parses the buffered deltas and keeps them so the final item can still win.
  */
 export declare function finalizeToolCallArgumentsDone(block: ResponsesToolCallBlock, args: string): void;
 export declare function accumulateCustomToolCallInputDelta(block: ResponsesToolCallBlock, delta: string, stream: AssistantMessageEventStream, output: AssistantMessage, contentIndex: number): void;

package/package.json

@@ -1,6 +1,6 @@
 {
 	"name": "@oh-my-pi/pi-ai",
-	"version": "18.8.7",
+	"version": "18.8.9",
 	"description": "Unified LLM API with automatic model discovery and provider configuration",
 	"keywords": [
 		"ai",
@@ -155,11 +155,11 @@
 		"fmt": "oxfmt --no-error-on-unmatched-pattern 'src/**/*.{ts,tsx}' '{test,bench,examples,scripts}/**/*.ts' '*.ts'"
 	},
 	"dependencies": {
-		"@oh-my-pi/omptype": "18.8.7",
-		"@oh-my-pi/pi-catalog": "18.8.7",
-		"@oh-my-pi/pi-natives": "18.8.7",
-		"@oh-my-pi/pi-utils": "18.8.7",
-		"@oh-my-pi/pi-wire": "18.8.7"
+		"@oh-my-pi/omptype": "18.8.9",
+		"@oh-my-pi/pi-catalog": "18.8.9",
+		"@oh-my-pi/pi-natives": "18.8.9",
+		"@oh-my-pi/pi-utils": "18.8.9",
+		"@oh-my-pi/pi-wire": "18.8.9"
 	},
 	"devDependencies": {
 		"@types/bun": "^1.3.14"

package/src/judgment/text.ts

@@ -93,7 +93,15 @@ function renderStateField(key: string, value: JsonValue): string {
 	if (typeof value !== "object" || value === null) {
 		return `${open}${escapeXmlText(value === null ? "null" : String(value))}${close}`;
 	}
-	const yaml = YAML.stringify(value, null, 2).trimEnd();
+	const yaml = YAML.stringify(value, null, 2)
+		// Bun <=1.4.2 emits `key: ` before a nested block and puts an empty
+		// collection on its own line; 1.4.3 emits `key:` and `key: {}`. Pin the
+		// newer form so prompt bytes (and prompt-cache keys) do not depend on the
+		// Bun build. Values with edge whitespace are always quoted, so trimming
+		// line ends never alters data.
+		.replace(/[ \t]+$/gm, "")
+		.replace(/:\n +(\{\}|\[\])$/gm, ": $1")
+		.trimEnd();
 	return `${open}\n${escapeXmlText(yaml)}\n${close}`;
 }
 

package/src/providers/factory-droid/gemini.ts

@@ -15,6 +15,7 @@ import { notifyProviderResponse } from "../../utils/provider-response";
 import { dereferenceJsonSchema, normalizeSchemaForFactoryDroid, toolWireSchema } from "../../utils/schema";
 import {
 	extractGoogleErrorMessage,
+	googleStreamChunkError,
 	mapGoogleUsage,
 	mapStopReasonString,
 	nextToolCallId,
@@ -87,22 +88,22 @@ const FACTORY_DROID_BLOCK_REASONS: Record<string, true> = {
  * "unknown" bucket has no StopReason equivalent, so unknown terminators
  * surface as errors instead of masquerading as a clean stop.
  */
-function mapFactoryDroidFinishReason(reason: string | undefined): {
+function mapFactoryDroidFinishReason(reason: string): {
 	stopReason: "stop" | "length" | "error";
 	errorMessage?: string;
 } {
 	// mapStopReasonString already implements the CLI's outcome table
 	// (STOP→stop, MAX_TOKENS→length, everything else→error); the cast narrows
 	// its wide StopReason return to the three values it ever produces.
-	const stopReason = mapStopReasonString(reason ?? "") as "stop" | "length" | "error";
+	const stopReason = mapStopReasonString(reason) as "stop" | "length" | "error";
 	if (stopReason !== "error") return { stopReason };
-	if (reason && FACTORY_DROID_BLOCK_REASONS[reason]) {
+	if (FACTORY_DROID_BLOCK_REASONS[reason]) {
 		return { stopReason: "error", errorMessage: `Generation was blocked by content filters (${reason})` };
 	}
 	if (reason === "MALFORMED_FUNCTION_CALL") {
 		return { stopReason: "error", errorMessage: `Generation failed with finish reason: ${reason}` };
 	}
-	return { stopReason: "error", errorMessage: `Unknown finish reason: ${reason ?? "none"}` };
+	return { stopReason: "error", errorMessage: `Unknown finish reason: ${reason}` };
 }
 
 /** Transports whose outputs carry Gemini-verifiable signatures (Factory's own outputs are stamped google-generative-ai). */
@@ -414,6 +415,9 @@ export function streamFactoryDroidGemini(
 				},
 			);
 			for await (const chunk of chunks) {
+				// A server-declared error keeps its status and the same classification
+				// as the Google provider; only a stream that just stops is a premature close.
+				if (chunk.error) throw googleStreamChunkError(chunk.error, model.provider);
 				if (firstTokenTime === undefined && chunk.candidates?.[0]?.content?.parts?.some(part => part.text)) {
 					firstTokenTime = performance.now();
 				}
@@ -484,6 +488,20 @@ export function streamFactoryDroidGemini(
 			if (firstTokenTime !== undefined) output.ttft = firstTokenTime - startTime;
 
 			closeBlock();
+			// A stream that reaches EOF without a finishReason (and without a
+			// promptFeedback block) was truncated, even when it already carried
+			// functionCall parts; fail it as retryable instead of finishing the turn.
+			if (finishReason === undefined && !blockReason) {
+				// Worded like the other incomplete-stream errors so turn recovery
+				// continues a stream that already rendered text.
+				throw new AIError.ProviderResponseError(
+					"Factory Droid Gemini stream closed before a finish_reason was received",
+					{
+						provider: model.provider,
+						kind: "incomplete-stream",
+					},
+				);
+			}
 			for (const contentIndex of toolCallIndices) {
 				const toolCall = output.content[contentIndex] as ToolCall;
 				stream.push({ type: "toolcall_end", contentIndex, toolCall, partial: output });
@@ -499,12 +517,12 @@ export function streamFactoryDroidGemini(
 				output.errorMessage = `Generation was blocked by content filters (${blockReason})`;
 				output.stopDetails = { type: "content_filter", category: blockReason };
 				stream.push({ type: "error", reason: "error", error: output });
-			} else {
+			} else if (finishReason !== undefined) {
 				const mapped = mapFactoryDroidFinishReason(finishReason);
 				output.stopReason = mapped.stopReason;
 				if (mapped.errorMessage) {
 					output.errorMessage = mapped.errorMessage;
-					if (mapped.stopReason === "error" && finishReason && FACTORY_DROID_BLOCK_REASONS[finishReason]) {
+					if (mapped.stopReason === "error" && FACTORY_DROID_BLOCK_REASONS[finishReason]) {
 						output.stopDetails = { type: "content_filter", category: finishReason };
 					}
 				}

package/src/providers/google-shared.ts

@@ -595,6 +595,23 @@ export function startTextOrThinkingBlock(
 	return block;
 }
 
+/**
+ * Error for an in-band `error` frame in a 200 generateContent stream. Keeps the
+ * RPC status alongside the message: an in-band quota failure is classified
+ * from this text, and `RESOURCE_EXHAUSTED` is the only account-exhaustion
+ * signal some of these chunks carry (#13090).
+ */
+export function googleStreamChunkError(error: NonNullable<GenerateContentResponse["error"]>, provider: string): Error {
+	const detail =
+		error.message && error.status
+			? `${error.message} (${error.status})`
+			: error.message || error.status || "unknown error";
+	const message = `Google API stream error: ${detail}`;
+	return typeof error.code === "number" && error.code >= 400
+		? new AIError.GoogleApiError(message, error.code)
+		: new AIError.ProviderResponseError(message, { provider, kind: "output" });
+}
+
 /**
  * Drives the chunked `generateContentStream` iterator into an `AssistantMessage` and
  * the corresponding `AssistantMessageEventStream`. Shared between `streamGoogle` and
@@ -645,19 +662,7 @@ export async function consumeGoogleStream<T extends GoogleApiType>(args: {
 	};
 
 	for await (const chunk of googleStream) {
-		if (chunk.error) {
-			// Keep the RPC status alongside the message: an in-band quota failure
-			// is classified from this text, and `RESOURCE_EXHAUSTED` is the only
-			// account-exhaustion signal some of these chunks carry (#13090).
-			const detail =
-				chunk.error.message && chunk.error.status
-					? `${chunk.error.message} (${chunk.error.status})`
-					: chunk.error.message || chunk.error.status || "unknown error";
-			const message = `Google API stream error: ${detail}`;
-			throw typeof chunk.error.code === "number" && chunk.error.code >= 400
-				? new AIError.GoogleApiError(message, chunk.error.code)
-				: new AIError.ProviderResponseError(message, { provider: model.provider, kind: "output" });
-		}
+		if (chunk.error) throw googleStreamChunkError(chunk.error, model.provider);
 		if (!chunk.candidates?.length && chunk.promptFeedback?.blockReason) {
 			const detail = chunk.promptFeedback.blockReasonMessage;
 			throw new AIError.ProviderResponseError(

package/src/providers/openai-codex-responses.ts

@@ -2637,11 +2637,17 @@ class CodexStreamProcessor {
 		}
 
 		if (item.type === "function_call") {
+			const partial = block?.type === "toolCall" ? block[kStreamingPartialJson] : undefined;
 			const toolCall: ToolCall = {
 				type: "toolCall",
 				id: encodeResponsesToolCallId(item.call_id, item.id),
 				name: item.name,
-				arguments: parseToolCallArguments(item.arguments),
+				// An empty terminal item falls back to what `.delta`/`.done` already delivered.
+				arguments: item.arguments
+					? parseToolCallArguments(item.arguments)
+					: block?.type === "toolCall" && !partial
+						? block.arguments
+						: parseToolCallArguments(partial),
 			};
 			item.arguments = replayableToolCallArguments(item.arguments, toolCall.arguments);
 			if (block?.type === "toolCall") {

package/src/providers/openai-responses.ts

@@ -395,7 +395,14 @@ async function* resumeOpenAIResponsesEventStream(
 
 interface OpenAIResponsesProviderSessionState
 	extends ProviderSessionState, OpenAIStrictToolsState, OpenAIReasoningEffortFallbackState {
+	/**
+	 * Replay native history items. Starts false only on connection-bound hosts;
+	 * `close()` clears it so the next request rebuilds history from message
+	 * content; the first successful response sets it.
+	 */
 	nativeHistoryReplayWarmed: boolean;
+	/** Bumped by `close()`; a response warms the state only if no close happened since its request started. */
+	closeCount: number;
 	/** Stateful `previous_response_id` chain baselines, keyed by baseUrl/model/session. */
 	chains: Map<string, OpenAIResponsesChainState>;
 	/** `configuration_update` effort baselines, keyed by baseUrl/model/session. */
@@ -423,13 +430,16 @@ interface OpenAIResponsesChainState {
 	disabled: boolean;
 }
 
-function createOpenAIResponsesProviderSessionState(): OpenAIResponsesProviderSessionState {
+function createOpenAIResponsesProviderSessionState(
+	model: Model<"openai-responses">,
+): OpenAIResponsesProviderSessionState {
 	const strictToolsState = createOpenAIStrictToolsState();
 	const reasoningEffortFallbackState = createOpenAIReasoningEffortFallbackState();
 	const state: OpenAIResponsesProviderSessionState = {
 		...strictToolsState,
 		...reasoningEffortFallbackState,
-		nativeHistoryReplayWarmed: false,
+		nativeHistoryReplayWarmed: !model.compat.connectionBoundNativeHistory,
+		closeCount: 0,
 		chains: new Map(),
 		effortControls: new Map(),
 		releaseSession: sessionId => {
@@ -442,6 +452,7 @@ function createOpenAIResponsesProviderSessionState(): OpenAIResponsesProviderSes
 		},
 		close: () => {
 			state.nativeHistoryReplayWarmed = false;
+			state.closeCount++;
 			state.chains.clear();
 			state.effortControls.clear();
 			clearOpenAIStrictToolsState(state);
@@ -459,7 +470,7 @@ function getOpenAIResponsesProviderSessionState(
 	const key = `${OPENAI_RESPONSES_PROVIDER_SESSION_STATE_PREFIX}${model.provider}`;
 	const existing = providerSessionState.get(key) as OpenAIResponsesProviderSessionState | undefined;
 	if (existing) return existing;
-	const created = createOpenAIResponsesProviderSessionState();
+	const created = createOpenAIResponsesProviderSessionState(model);
 	providerSessionState.set(key, created);
 	return created;
 }
@@ -733,6 +744,7 @@ const streamOpenAIResponsesOnce = (
 				});
 			const premiumRequestsTotal = copilotPremiumRequests;
 			const providerSessionState = getOpenAIResponsesProviderSessionState(model, options?.providerSessionState);
+			const closeCountAtStart = providerSessionState?.closeCount;
 			const strictToolsScope = getOpenAIStrictToolsScope(model, baseUrl);
 			const promptCacheBreakpointPolicy =
 				resolveCacheRetention(options?.cacheRetention) !== "none" && options?.promptCache?.mode === "explicit"
@@ -1250,7 +1262,9 @@ const streamOpenAIResponsesOnce = (
 				{ supportsImageDetailOriginal: model.compat.supportsImageDetailOriginal },
 			);
 			if (replayableResponseItems) {
-				if (providerSessionState) providerSessionState.nativeHistoryReplayWarmed = true;
+				if (providerSessionState && providerSessionState.closeCount === closeCountAtStart) {
+					providerSessionState.nativeHistoryReplayWarmed = true;
+				}
 				if (chainState) {
 					chainState.lastParams = cloneJsonTree(
 						activeTrailingScaffoldingItems > 0 && Array.isArray(activeParams.input)

package/src/providers/openai-shared.ts

@@ -3052,7 +3052,8 @@ export function accumulateToolCallArgumentsDelta(
 	contentIndex: number,
 ): void {
 	delta = optionalResponsesText(delta, "function call arguments delta") ?? "";
-	if (!delta) return;
+	// A non-blank `.done` consumed the buffer; a late delta must not reopen it.
+	if (!delta || block[kStreamingPartialJson] === undefined) return;
 	block[kStreamingPartialJson] += delta;
 	const throttled = parseStreamingJsonThrottled(block[kStreamingPartialJson], block[kStreamingLastParseLen] ?? 0);
 	if (throttled) {
@@ -3066,9 +3067,14 @@ export function accumulateToolCallArgumentsDelta(
  * Finalize streamed function-call arguments from the authoritative `.done`
  * payload. The caller owns the `argumentsDone` flag (generic Responses sets it;
  * Codex's block shape has no such field), so this only rewrites `arguments` and
- * drops the transient accumulation fields.
+ * drops the transient accumulation fields. A blank payload is not authoritative:
+ * it parses the buffered deltas and keeps them so the final item can still win.
  */
 export function finalizeToolCallArgumentsDone(block: ResponsesToolCallBlock, args: string): void {
+	if (!args.trim()) {
+		block.arguments = parseToolCallArguments(block[kStreamingPartialJson]);
+		return;
+	}
 	block[kStreamingPartialJson] = args;
 	block.arguments = parseToolCallArguments(block[kStreamingPartialJson]);
 	clearStreamingPartialJson(block);
@@ -3638,11 +3644,13 @@ export async function processResponsesStream<TApi extends Api>(
 				closeOpenItem(event.output_index, item.id, entry);
 			} else if (item.type === "function_call") {
 				const block = entry?.block.type === "toolCall" ? entry.block : undefined;
-				const args = block?.[kStreamingArgumentsDone]
-					? block.arguments
-					: item.arguments
-						? parseToolCallArguments(item.arguments)
-						: parseToolCallArguments(block?.[kStreamingPartialJson]);
+				// A blank `.done` leaves the delta buffer in place, so the item's own arguments still win.
+				const args =
+					block?.[kStreamingArgumentsDone] && block[kStreamingPartialJson] === undefined
+						? block.arguments
+						: item.arguments
+							? parseToolCallArguments(item.arguments)
+							: parseToolCallArguments(block?.[kStreamingPartialJson]);
 				item.arguments = replayableToolCallArguments(item.arguments, args);
 				const toolCall: ToolCall = {
 					type: "toolCall",