@victor-software-house/pi-openai-proxy 4.6.0 → 4.7.0

package/dist/index.mjs

@@ -179,6 +179,17 @@ function logDisconnect(ctx) {
 	};
 	console.error(JSON.stringify(entry));
 }
+function logUpstreamOverload(ctx, status, code) {
+	const entry = {
+		ts: timestamp(),
+		level: "warn",
+		event: "upstream_overload",
+		requestId: ctx.requestId,
+		upstreamStatus: status,
+		code
+	};
+	console.error(JSON.stringify(entry));
+}
 function logStartup(host, port, modelCount) {
 	const entry = {
 		ts: timestamp(),
@@ -1001,14 +1012,16 @@ function convertTools(openaiTools) {
 /**
 * Zod schemas for the OpenAI chat-completions request subset.
 *
-* Phase 2 contract:
-* - Phase 1 supported fields: model, messages, stream, temperature,
-*   max_tokens, max_completion_tokens, stop, user, stream_options
-* - Phase 2 additions: tools, tool_choice, reasoning_effort,
-*   top_p, frequency_penalty, presence_penalty, seed, response_format
+* Supported fields (cumulative through Phase 3D):
+* - Core: model, messages, stream, temperature, max_tokens,
+*   max_completion_tokens, stop, user, stream_options
+* - Tools: tools, tool_choice, parallel_tool_calls
+* - Model control: reasoning_effort, top_p, frequency_penalty,
+*   presence_penalty, seed, response_format
+* - Client interop: metadata, prediction
 * - Unknown top-level fields are rejected with 422
-* - `n > 1` is rejected
-* - `logprobs` is rejected
+* - Explicitly rejected: n, logprobs, top_logprobs, logit_bias,
+*   functions (deprecated), function_call (deprecated)
 */
 const textContentPartSchema = z.object({
 	type: z.literal("text"),
@@ -1125,6 +1138,7 @@ const chatCompletionRequestSchema = z.object({
 	stream_options: streamOptionsSchema.nullable().optional(),
 	tools: z.array(functionToolSchema).optional(),
 	tool_choice: toolChoiceSchema.optional(),
+	parallel_tool_calls: z.boolean().optional(),
 	reasoning_effort: z.enum([
 		"none",
 		"minimal",
@@ -1137,11 +1151,25 @@ const chatCompletionRequestSchema = z.object({
 	frequency_penalty: z.number().min(-2).max(2).optional(),
 	presence_penalty: z.number().min(-2).max(2).optional(),
 	seed: z.int().optional(),
-	response_format: responseFormatSchema.optional()
+	response_format: responseFormatSchema.optional(),
+	metadata: z.record(z.string().trim(), z.unknown()).optional(),
+	prediction: z.object({
+		type: z.literal("content"),
+		content: z.union([z.string().trim(), z.array(z.object({
+			type: z.literal("text"),
+			text: z.string().trim()
+		}))])
+	}).optional()
 }).strict();
 /**
 * Fields that are explicitly rejected with a helpful error.
-* These are not supported and won't be promoted.
+*
+* `n`, `logprobs`, `top_logprobs`, `logit_bias`: not supported by the pi SDK's
+* simple completion interface. The proxy returns a single choice with no token
+* probability data. Promoting these would require response-side changes.
+*
+* `functions`, `function_call`: deprecated OpenAI fields, superseded by `tools`
+* and `tool_choice`. Clients should migrate to the current API.
 */
 const rejectedFields = [
 	"n",
@@ -1149,8 +1177,7 @@ const rejectedFields = [
 	"top_logprobs",
 	"logit_bias",
 	"functions",
-	"function_call",
-	"parallel_tool_calls"
+	"function_call"
 ];
 //#endregion
 //#region src/openai/validate.ts
@@ -1227,6 +1254,8 @@ const SKIP_PAYLOAD_PASSTHROUGH_APIS = new Set(["openai-codex-responses"]);
 /**
 * Collect fields that need to be injected via onPayload.
 * Skips passthrough for APIs that use non-standard request formats.
+*
+* @internal Exported for unit testing only.
 */
 function collectPayloadFields(request, api) {
 	if (SKIP_PAYLOAD_PASSTHROUGH_APIS.has(api)) return;
@@ -1260,9 +1289,67 @@ function collectPayloadFields(request, api) {
 		fields["response_format"] = request.response_format;
 		hasFields = true;
 	}
+	if (request.tool_choice !== void 0) {
+		fields["tool_choice"] = request.tool_choice;
+		hasFields = true;
+	}
+	if (request.parallel_tool_calls !== void 0) {
+		fields["parallel_tool_calls"] = request.parallel_tool_calls;
+		hasFields = true;
+	}
+	if (request.metadata !== void 0) {
+		fields["metadata"] = request.metadata;
+		hasFields = true;
+	}
+	if (request.prediction !== void 0) {
+		fields["prediction"] = request.prediction;
+		hasFields = true;
+	}
 	return hasFields ? fields : void 0;
 }
 /**
+* Collect tool strict flags from the original OpenAI request.
+*
+* The pi SDK's `Tool` interface has no `strict` field, so the SDK always sets
+* `strict: false` when building the upstream payload. This function extracts
+* the per-tool strict flags from the original request so they can be restored
+* via `onPayload` after the SDK builds the payload.
+*
+* Returns a map of tool index -> true for tools that requested strict mode,
+* or undefined if no tools use strict mode.
+*
+* @internal Exported for unit testing only.
+*/
+function collectToolStrictFlags(tools) {
+	if (tools === void 0 || tools.length === 0) return;
+	let flags;
+	for (let i = 0; i < tools.length; i++) if (tools[i]?.function.strict === true) {
+		flags ??= /* @__PURE__ */ new Map();
+		flags.set(i, true);
+	}
+	return flags;
+}
+/**
+* Apply strict flags to tool definitions in the upstream payload.
+*
+* The pi SDK always sets `strict: false` on tool definitions. This function
+* patches the payload's `tools` array to restore the client's requested
+* `strict: true` flags on the matching tool definitions.
+*
+* @internal Exported for unit testing only.
+*/
+function applyToolStrictFlags(payload, strictFlags) {
+	const tools = payload["tools"];
+	if (!Array.isArray(tools)) return;
+	for (const [index, _flag] of strictFlags) {
+		const tool = tools[index];
+		if (isRecord(tool)) {
+			const fn = tool["function"];
+			if (isRecord(fn)) fn["strict"] = true;
+		}
+	}
+}
+/**
 * Combine a client disconnect signal with an upstream timeout into a single signal.
 * Returns the combined signal, or undefined if neither is provided.
 */
@@ -1293,8 +1380,12 @@ async function buildStreamOptions(model, request, options) {
 		if (apiKey !== void 0) opts.apiKey = apiKey;
 	}
 	const payloadFields = collectPayloadFields(request, model.api);
-	if (payloadFields !== void 0) opts.onPayload = (payload) => {
-		if (isRecord(payload)) for (const [key, value] of Object.entries(payloadFields)) payload[key] = value;
+	const strictFlags = collectToolStrictFlags(request.tools);
+	if (payloadFields !== void 0 || strictFlags !== void 0) opts.onPayload = (payload) => {
+		if (isRecord(payload)) {
+			if (payloadFields !== void 0) for (const [key, value] of Object.entries(payloadFields)) payload[key] = value;
+			if (strictFlags !== void 0) applyToolStrictFlags(payload, strictFlags);
+		}
 		return payload;
 	};
 	return opts;
@@ -1329,6 +1420,15 @@ function createRoutes(config, configReader = fileConfigReader) {
 		if (!outcome.ok) throw new Error(`Model exposure configuration error: ${outcome.message}`);
 		return outcome;
 	}
+	/**
+	* Map upstream error, log it, and emit a structured warn for rate limit / overload.
+	*/
+	function handleUpstreamError(ctx, err) {
+		const mapped = mapUpstreamError(err);
+		logError(ctx, mapped.body.error.message, err instanceof Error ? err.message : void 0);
+		if (mapped.status === 429 || mapped.status === 503) logUpstreamOverload(ctx, mapped.status, mapped.body.error.code ?? "unknown");
+		return mapped;
+	}
 	const routes = new Hono();
 	routes.get("/v1/models", (c) => {
 		const exposure = getExposure();
@@ -1390,12 +1490,11 @@ function createRoutes(config, configReader = fileConfigReader) {
 					const eventStream = await piStream(model, context, request, completionOptions);
 					for await (const frame of streamToSSE(eventStream, requestId, canonicalModelId, includeUsage)) await stream.write(frame);
 				} catch (err) {
-					const mapped = mapUpstreamError(err);
-					logError({
+					const mapped = handleUpstreamError({
 						requestId,
 						method: "POST",
 						path: "/v1/chat/completions"
-					}, mapped.body.error.message, err instanceof Error ? err.message : void 0);
+					}, err);
 					const errorChunk = JSON.stringify({ error: mapped.body.error });
 					await stream.write(`data: ${errorChunk}\n\n`);
 					await stream.write("data: [DONE]\n\n");
@@ -1406,22 +1505,20 @@ function createRoutes(config, configReader = fileConfigReader) {
 			const message = await piComplete(model, context, request, completionOptions);
 			if (message.stopReason === "error" || message.stopReason === "aborted") {
 				const errorMessage = message.errorMessage ?? "Upstream provider error";
-				const mapped = mapUpstreamError(new Error(errorMessage));
-				logError({
+				const mapped = handleUpstreamError({
 					requestId,
 					method: "POST",
 					path: "/v1/chat/completions"
-				}, errorMessage);
+				}, new Error(errorMessage));
 				return c.json(mapped.body, mapped.status);
 			}
 			return c.json(buildChatCompletion(requestId, canonicalModelId, message));
 		} catch (err) {
-			const mapped = mapUpstreamError(err);
-			logError({
+			const mapped = handleUpstreamError({
 				requestId,
 				method: "POST",
 				path: "/v1/chat/completions"
-			}, mapped.body.error.message, err instanceof Error ? err.message : void 0);
+			}, err);
 			return c.json(mapped.body, mapped.status);
 		}
 	});

package/dist/sync-zed.mjs

@@ -48,7 +48,7 @@ function toZedModel(exposed) {
 		capabilities: {
 			tools: true,
 			images: model.input.includes("image"),
-			parallel_tool_calls: false,
+			parallel_tool_calls: true,
 			prompt_cache_key: false,
 			chat_completions: true
 		}

package/package.json

@@ -1,6 +1,6 @@
 {
 	"name": "@victor-software-house/pi-openai-proxy",
-	"version": "4.6.0",
+	"version": "4.7.0",
 	"description": "OpenAI-compatible HTTP proxy for pi's multi-provider model registry",
 	"license": "MIT",
 	"author": "Victor Software House",