@oh-my-pi/pi-ai 16.3.9 → 16.3.10

package/CHANGELOG.md

@@ -2,6 +2,13 @@
 
 ## [Unreleased]
 
+## [16.3.10] - 2026-07-06
+
+### Fixed
+
+- Fixed Ollama/Ollama Cloud EOS-only completions to retry empty stops with a single output token before the agent loop can halt silently. ([#4659](https://github.com/can1357/oh-my-pi/issues/4659))
+- Fixed Claude Sonnet 5 failing every request on feature-gated gateways (Azure Foundry, OpenAI-compatible relays) that reject strict tools with "structured_outputs not supported" — the rejection is now classified as a strict-tool rejection, so the request retries without strict tools and the session remembers the downgrade.
+
 ## [16.3.7] - 2026-07-05
 
 ### Fixed

package/dist/types/error/flags.d.ts

@@ -13,7 +13,7 @@ export declare const Flag: {
     readonly SilentAbort: 33554432;
     readonly UserInterrupt: 67108864;
     readonly Abort: 134217728;
-    /** Anthropic strict-tool grammar too large / schema too complex to compile (400). */
+    /** Strict-tool rejection (400): grammar too large, schema too complex, or structured outputs unsupported by the model/endpoint. */
     readonly Grammar: 268435456;
     /** Anthropic model/account does not support fast mode / the `speed` parameter. */
     readonly FastModeUnsupported: 536870912;
@@ -48,7 +48,8 @@ export declare function classify(error: unknown, api?: Api): number;
  */
 export declare function isUsageLimit(error: unknown, api?: Api): boolean;
 /**
- * Anthropic strict-tool grammar too large / schema too complex to compile.
+ * Strict-tool rejection: grammar too large, schema too complex, or structured
+ * outputs unsupported by the model/endpoint.
  * Accessor for {@link Flag.Grammar}.
  */
 export declare function isGrammarError(error: unknown): boolean;

package/dist/types/providers/ollama.d.ts

@@ -4,4 +4,5 @@ export interface OllamaChatOptions extends StreamOptions {
     disableReasoning?: boolean;
     toolChoice?: ToolChoice;
 }
+/** Retry EOS-only Ollama completions before the agent loop sees an empty stop. */
 export declare const streamOllama: StreamFunction<"ollama-chat">;

package/package.json

@@ -1,7 +1,7 @@
 {
 	"type": "module",
 	"name": "@oh-my-pi/pi-ai",
-	"version": "16.3.9",
+	"version": "16.3.10",
 	"description": "Unified LLM API with automatic model discovery and provider configuration",
 	"homepage": "https://omp.sh",
 	"author": "Can Boluk",
@@ -38,9 +38,9 @@
 	},
 	"dependencies": {
 		"@bufbuild/protobuf": "^2.12.0",
-		"@oh-my-pi/pi-catalog": "16.3.9",
-		"@oh-my-pi/pi-utils": "16.3.9",
-		"@oh-my-pi/pi-wire": "16.3.9",
+		"@oh-my-pi/pi-catalog": "16.3.10",
+		"@oh-my-pi/pi-utils": "16.3.10",
+		"@oh-my-pi/pi-wire": "16.3.10",
 		"arktype": "^2.2.0",
 		"zod": "^4"
 	},

package/src/error/flags.ts

@@ -22,7 +22,7 @@ export const Flag = {
 	SilentAbort: 0x0200_0000,
 	UserInterrupt: 0x0400_0000,
 	Abort: 0x0800_0000,
-	/** Anthropic strict-tool grammar too large / schema too complex to compile (400). */
+	/** Strict-tool rejection (400): grammar too large, schema too complex, or structured outputs unsupported by the model/endpoint. */
 	Grammar: 0x1000_0000,
 	/** Anthropic model/account does not support fast mode / the `speed` parameter. */
 	FastModeUnsupported: 0x2000_0000,
@@ -108,12 +108,17 @@ export const LLAMA_CPP_TOOL_CALL_PARSE_PATTERN =
 // on a backend that has the model.
 const COPILOT_MODEL_NOT_SUPPORTED_PATTERN = /model_not_supported/i;
 // Anthropic strict-tool grammar too large / schema too complex (400 invalid_request_error).
+// Feature-gated deployments (Azure Foundry, Baseten, …) reject `strict: true`
+// tools outright when the hosted model lacks structured outputs, e.g.
+// "structured_outputs not supported" — without an invalid_request_error wrapper.
 const GRAMMAR_TOO_LARGE_PATTERN = /compiled grammar/i;
 const GRAMMAR_TOO_LARGE_DETAIL_PATTERN = /too large/i;
 const SCHEMA_TOO_COMPLEX_PATTERN = /schema/i;
 const SCHEMA_TOO_COMPLEX_DETAIL_PATTERN = /too complex/i;
 const SCHEMA_COMPILE_PATTERN = /compil/i;
 const INVALID_REQUEST_PATTERN = /invalid_request_error/i;
+const STRUCTURED_OUTPUTS_PATTERN = /structured[_ -]?outputs?/i;
+const FEATURE_NOT_SUPPORTED_PATTERN = /not (?:supported|available|enabled)|unsupported|does(?: not|n'?t) support/i;
 // Anthropic fast-mode unsupported: 400 rejecting `speed`, or 429 rate_limit_error
 // because the account lacks the extra-usage entitlement fast mode requires.
 const FAST_MODE_SPEED_PARAM_PATTERN = /\bspeed\b/i;
@@ -127,8 +132,9 @@ const OAUTH_TRANSIENT_FAILURE_PATTERN =
 	/timeout|network|fetch failed|ECONN(?:REFUSED|RESET)|ETIMEDOUT|EAI_AGAIN|socket hang up|\b(?:408|425|429|5\d{2})\b|rate.?limit|too many requests|temporar|unavailable|forbidden|permission_denied|cloudflare|captcha/i;
 const OAUTH_HTTP_AUTH_PATTERN = /\b401\b/;
 
-function matchesGrammarTooLarge(message: string, errorStatus: number | undefined): boolean {
+function matchesStrictToolsRejection(message: string, errorStatus: number | undefined): boolean {
 	if (errorStatus !== 400) return false;
+	if (STRUCTURED_OUTPUTS_PATTERN.test(message) && FEATURE_NOT_SUPPORTED_PATTERN.test(message)) return true;
 	if (!INVALID_REQUEST_PATTERN.test(message)) return false;
 	const grammarTooLarge = GRAMMAR_TOO_LARGE_PATTERN.test(message) && GRAMMAR_TOO_LARGE_DETAIL_PATTERN.test(message);
 	const schemaTooComplex =
@@ -317,7 +323,7 @@ function classifyText(errorMessage: string | undefined, errorStatus: number | un
 
 		// Copilot per-client routing flap is transient.
 		if (statusClean === 400 && COPILOT_MODEL_NOT_SUPPORTED_PATTERN.test(cleanMessage)) kinds |= Flag.Transient;
-		if (matchesGrammarTooLarge(cleanMessage, statusClean)) kinds |= Flag.Grammar;
+		if (matchesStrictToolsRejection(cleanMessage, statusClean)) kinds |= Flag.Grammar;
 		if (matchesFastModeUnsupported(cleanMessage, statusClean)) kinds |= Flag.FastModeUnsupported;
 	}
 	if (kinds !== 0) return create(kinds);
@@ -411,7 +417,8 @@ export function isUsageLimit(error: unknown, api?: Api): boolean {
 }
 
 /**
- * Anthropic strict-tool grammar too large / schema too complex to compile.
+ * Strict-tool rejection: grammar too large, schema too complex, or structured
+ * outputs unsupported by the model/endpoint.
  * Accessor for {@link Flag.Grammar}.
  */
 export function isGrammarError(error: unknown): boolean {

package/src/providers/anthropic.ts

@@ -2396,7 +2396,7 @@ const streamAnthropicOnce = (
 					) {
 						// Log-only: the retried turn must not carry an errorMessage on
 						// success (consumers treat its presence as failure).
-						logger.warn("anthropic: strict tool grammar rejected, retrying without strict tools", {
+						logger.warn("anthropic: strict tools rejected, retrying without strict tools", {
 							model: model.id,
 							error: await finalizeErrorMessage(streamFailure, rawRequestDump),
 						});

package/src/providers/ollama.ts

@@ -16,6 +16,7 @@ import type {
 } from "../types";
 import { normalizeSystemPrompts } from "../utils";
 import { clearStreamingPartialJson, kStreamingPartialJson } from "../utils/block-symbols";
+import { withEmptyCompletionRetry } from "../utils/empty-completion-retry";
 import { AssistantMessageEventStream } from "../utils/event-stream";
 import type { CapturedHttpErrorResponse, RawHttpRequestDump } from "../utils/http-inspector";
 import {
@@ -449,10 +450,10 @@ function hasVisibleAssistantContent(output: AssistantMessage): boolean {
 
 const OLLAMA_RETRY_DELAYS_MS = [2_000, 5_000, 10_000];
 
-export const streamOllama: StreamFunction<"ollama-chat"> = (
+const streamOllamaOnce = (
 	model: Model<"ollama-chat">,
 	context: Context,
-	options: OllamaChatOptions,
+	options: OllamaChatOptions = {},
 ): AssistantMessageEventStream => {
 	const stream = new AssistantMessageEventStream();
 	void (async () => {
@@ -771,3 +772,7 @@ export const streamOllama: StreamFunction<"ollama-chat"> = (
 	})();
 	return stream;
 };
+
+/** Retry EOS-only Ollama completions before the agent loop sees an empty stop. */
+export const streamOllama: StreamFunction<"ollama-chat"> = (model, context, options) =>
+	withEmptyCompletionRetry(model, context, options, streamOllamaOnce);

package/src/providers/openai-shared.ts

@@ -1041,7 +1041,7 @@ export function shouldRetryWithoutStrictTools(
 	const messageParts = [error instanceof Error ? error.message : undefined, capturedErrorResponse?.bodyText]
 		.filter((value): value is string => typeof value === "string" && value.trim().length > 0)
 		.join("\n");
-	return /wrong_api_format|mixed values for 'strict'|tool[s]?\b.*strict|\bstrict\b.*tool|tool parameters? schema|invalid schema for function/i.test(
+	return /wrong_api_format|mixed values for 'strict'|tool[s]?\b.*strict|\bstrict\b.*tool|tool parameters? schema|invalid schema for function|structured[_ -]?outputs?\b[^\n]*(?:not (?:supported|available|enabled)|unsupported)|(?:not support|unsupported)[^\n]*structured[_ -]?outputs?\b/i.test(
 		messageParts,
 	);
 }

package/src/utils/empty-completion-retry.ts

@@ -109,17 +109,16 @@ export function withEmptyCompletionRetry<M, O extends EmptyCompletionRetryOption
 			}
 
 			// Retry only a genuinely degenerate completion: a normal stop that
-			// produced no visible content AND billed no output tokens (the flaky
-			// gateway signature — charged nothing, returned nothing). A stop that
-			// reports output tokens spent its budget somewhere (e.g. thinking) and
-			// is left alone.
+			// produced no visible content and reported no generated content tokens.
+			// Some providers count the terminal EOS as one output token, so a
+			// one-token invisible stop is still the same empty-completion failure.
 			const message = terminal?.type === "done" ? terminal.message : undefined;
 			const isRetryableEmpty =
 				!committed &&
 				message !== undefined &&
 				message.stopReason === "stop" &&
 				!message.errorMessage &&
-				(message.usage?.output ?? 0) <= 0 &&
+				(message.usage?.output ?? 0) <= 1 &&
 				!hasVisibleAssistantContent(message);
 
 			if (isRetryableEmpty && emptyAttempt < MAX_EMPTY_COMPLETION_RETRIES && !signal?.aborted) {