@gajae-code/ai 0.17.1 → 0.17.4

package/src/providers/openai-completions.ts

@@ -18,6 +18,7 @@ import {
 } from "../adapter-internals/provider-safety-stop";
 import {
 	type Effort,
+	getMiniMaxThinkingMode,
 	getSupportedEfforts,
 	isGroqCompoundReasoningUnsupported,
 	modelSupportsReasoningControl,
@@ -33,6 +34,7 @@ import {
 	type Model,
 	type OpenAICompat,
 	type ProviderSessionState,
+	type RepetitionGuardOptions,
 	resolveServiceTier,
 	type ServiceTier,
 	type StopReason,
@@ -78,6 +80,13 @@ import { callWithCopilotModelRetry } from "../utils/retry";
 import { resolveRetryBudget } from "../utils/retry-budget";
 import { adaptSchemaForStrict, flattenToolRootCombinators, NO_STRICT, toolWireSchema } from "../utils/schema";
 import { wrapFetchForSseDebug } from "../utils/sse-debug";
+import {
+	DEFAULT_REPETITION_THRESHOLD,
+	REPETITION_GUARD_ERROR_CODE,
+	REPETITION_GUARD_STOP_MESSAGE,
+	StreamRepetitionGuard,
+	type StreamRepetitionTrip,
+} from "../utils/stream-repetition-guard";
 import { type HealedToolCall, modelMayLeakKimiToolCalls, ToolCallHealer } from "../utils/tool-call-healing";
 import { isForcedToolChoice, mapToOpenAICompletionsToolChoice } from "../utils/tool-choice";
 import {
@@ -85,6 +94,7 @@ import {
 	markToolChoiceIncapability,
 	resolveToolChoice,
 } from "../utils/tool-choice-capability";
+import { ToolFenceStripper } from "../utils/tool-fence-strip";
 import { COMPOSER_EDIT_DISCIPLINE_PROMPT, isComposerHarnessModel } from "./composer-discipline";
 import { mergeDashScopeTokenPlanHeaders } from "./dashscope-token-plan-headers";
 import {
@@ -358,13 +368,21 @@ export interface OpenAICompletionsOptions extends StreamOptions {
 	/** Force-disable reasoning where supported, or request the lowest effort on generic effort endpoints. */
 	disableReasoning?: boolean;
 	serviceTier?: ServiceTier;
+	/**
+	 * Runaway-repetition guard thresholds, per stream channel. A number sets the
+	 * consecutive-repeat threshold; `false` disables the channel's guard.
+	 * Defaults: thinking = DEFAULT_REPETITION_THRESHOLD, text = false — visible
+	 * output is a deliverable and intentional repetition there (logs, fixtures,
+	 * tables, generated code) must survive byte for byte (#5627).
+	 */
+	repetitionGuard?: RepetitionGuardOptions;
 }
 
 type OpenAICompletionsParams = Omit<OpenAI.Chat.Completions.ChatCompletionCreateParamsStreaming, "reasoning_effort"> & {
 	top_k?: number;
 	min_p?: number;
 	repetition_penalty?: number;
-	thinking?: { type: "enabled" | "disabled" };
+	thinking?: { type: "enabled" | "disabled" | "adaptive" };
 	enable_thinking?: boolean;
 	chat_template_kwargs?: { enable_thinking: boolean };
 	reasoning?: { effort?: string } | { enabled: false };
@@ -529,6 +547,20 @@ const OPENAI_COMPLETIONS_EMPTY_RESPONSE_MESSAGE = "Provider returned an empty re
 const OPENAI_COMPLETIONS_NETWORK_ERROR_RETRY_MAX_RETRIES = 3;
 const OPENAI_COMPLETIONS_NETWORK_ERROR_RETRY_BASE_DELAY_MS = 2000;
 
+// A tripped repetition guard stops *emitting* immediately, so the user-visible
+// symptom is already fixed at the trip. Aborting the stream right then would
+// also drop `tool_calls` frames a provider emits *after* the repeats, losing a
+// valid invocation (#5627). The stream is drained for a bounded window instead;
+// the only thing the abort still buys is not burning provider budget, and that
+// can wait this long.
+const REPETITION_DRAIN_MAX_CHUNKS = 64;
+const REPETITION_DRAIN_MAX_MS = 2_000;
+// A tool call whose accumulated arguments are not yet complete JSON is worth
+// waiting longer for — but not forever, or a call whose arguments never
+// complete would hold the stream open for the rest of the turn's budget.
+const REPETITION_DRAIN_PENDING_TOOL_MAX_CHUNKS = 256;
+const REPETITION_DRAIN_PENDING_TOOL_MAX_MS = 8_000;
+
 function hasReplayUnsafeOpenAICompletionsDelta(chunk: ChatCompletionChunk): boolean {
 	const choice = Array.isArray(chunk.choices) ? chunk.choices[0] : undefined;
 	const delta = choice?.delta;
@@ -583,6 +615,48 @@ export const streamOpenAICompletions: StreamFunction<"openai-completions"> = (
 		const abortTracker = createAbortSourceTracker(options?.signal);
 		const { requestAbortController, requestSignal } = abortTracker;
 
+		// Declared outside the try so the catch block — which is where the abort
+		// below lands — can tell a runaway-repetition stop from a transport error.
+		let repetitionTrip: (StreamRepetitionTrip & { channel: "text" | "thinking" }) | undefined;
+		// Drain-window bookkeeping: when the trip happened, and how many chunks have
+		// been consumed since. Both are only meaningful once `repetitionTrip` is set.
+		let repetitionTrippedAt: number | undefined;
+		let repetitionDrainedChunks = 0;
+		// Set immediately before the guard's own abort and nowhere else. The catch
+		// block must be able to tell OUR abort from a provider stall or a transport
+		// failure that merely happened to land inside the drain window: `repetitionTrip`
+		// alone is true for all three, and using it there discarded the real
+		// timeout/transport facts and flipped the retry classification (#5627 review r4).
+		let repetitionSelfAbort = false;
+		const finalizeRepetitionGuardStop = (): void => {
+			if (!repetitionTrip) return;
+			// `error`, not `aborted`: this is a provider-side failure we detected
+			// locally, and `aborted` is the wire for *client cancellation* — the
+			// auth gateway maps it to 499/`request_aborted` and telemetry counts it
+			// as a user cancel, so borrowing it misreports the turn (#5627). No new
+			// StopReason variant: the union is switched on exhaustively everywhere.
+			// `errorCode` stays the bounded classifier for *why* (#5624).
+			output.stopReason = "error";
+			output.errorCode = REPETITION_GUARD_ERROR_CODE;
+			// A fixed literal, never the observed sample/channel/count: the auth
+			// gateway forwards `errorMessage` to API clients on the streaming path,
+			// so interpolating here publishes raw model output and feeds it to a
+			// keyword classifier that picks HTTP status from message text. The
+			// diagnostic detail lives in the `logger.debug` at the trip site
+			// instead (#5627 review r5).
+			output.errorMessage = REPETITION_GUARD_STOP_MESSAGE;
+			output.duration = Date.now() - startTime;
+			if (firstTokenTime) output.ttft = firstTokenTime - startTime;
+			// No `transportFailure`: this is a local decision, not a retryable
+			// transport fault, and the agent loop's retry admission keys on that
+			// field. The session-layer classifier does not — absent transport facts
+			// it defaults to a bounded retry — so it branches on this `errorCode`
+			// and treats the trip as terminal instead (#5627). Retrying is pointless
+			// anyway: a decode loop is deterministic for the submitted context.
+			stream.push({ type: "error", reason: "error", error: output });
+			stream.end();
+		};
+
 		try {
 			const apiKey = options?.apiKey || getEnvApiKey(model.provider) || "";
 			const idleTimeoutMs = options?.streamIdleTimeoutMs ?? getOpenAIStreamIdleTimeoutMs(model.provider, model.id);
@@ -869,15 +943,117 @@ export const streamOpenAICompletions: StreamFunction<"openai-completions"> = (
 
 			let taggedTextBuffer = "";
 			let insideTaggedThinking = false;
+			// One guard per channel: interleaving visible text and reasoning through
+			// a single instance would splice unrelated tokens into the same window.
+			// Tool-call frames are never fed through either guard.
+			//
+			// A disabled channel gets no guard at all rather than a lenient one, so
+			// it is structurally impossible for it to set `repetitionTrip`. Visible
+			// text is disabled by default: a decode loop there is not the reported
+			// failure (#5624 was reasoning-channel), and truncating deliverable
+			// output — a log dump, a fixture, a table — corrupts the answer (#5627).
+			const createRepetitionGuard = (
+				setting: number | false | undefined,
+				fallback: number | false,
+			): StreamRepetitionGuard | undefined => {
+				const threshold = setting ?? fallback;
+				return threshold === false ? undefined : new StreamRepetitionGuard({ threshold });
+			};
+			const textRepetitionGuard = createRepetitionGuard(options?.repetitionGuard?.text, false);
+			const thinkingRepetitionGuard = createRepetitionGuard(
+				options?.repetitionGuard?.thinking,
+				DEFAULT_REPETITION_THRESHOLD,
+			);
+			const noteRepetitionTrip = (guard: StreamRepetitionGuard, channel: "text" | "thinking") => {
+				// `takeTrip()` latches once per guard; this latches once per request,
+				// so the drain window below opens exactly once no matter which
+				// channel loops.
+				if (repetitionTrip) return;
+				const trip = guard.takeTrip();
+				if (!trip) return;
+				repetitionTrip = { ...trip, channel };
+				// The repeated unit is never logged. `logger`'s default transport is a
+				// rotating file under `~/.gjc/logs` and `makeLogFormat` JSON-stringifies
+				// every metadata key verbatim — no redaction — so a sample would persist
+				// raw model output to disk and carry it into log rotation, support
+				// bundles and backups. If the loop swallowed a secret or a private
+				// fragment of the prompt, that is where it would land (#5627 review r6).
+				//
+				// Only bounded metadata the model cannot control the *content* of goes
+				// out: an id, two enums and two counts. `sampleLength` is deliberately a
+				// number, not a hash — a hash of a short secret is a probe oracle and
+				// buys nothing for debugging a decode loop. `trip.sample` itself stays on
+				// the in-memory trip for callers; this is only the logging contract.
+				//
+				// Separately, `errorMessage` stays a fixed literal because the gateway
+				// forwards it to API clients (#5627 review r5). Both hold at once.
+				logger.debug("openai-completions: repetition guard tripped", {
+					model: model.id,
+					channel,
+					kind: trip.kind,
+					repeats: trip.repeats,
+					sampleLength: trip.sample.length,
+				});
+				// Deliberately no abort here — see the REPETITION_DRAIN_* constants.
+				// The main loop closes the window once late tool-call frames have had
+				// their chance to land.
+				repetitionTrippedAt = Date.now();
+				repetitionDrainedChunks = 0;
+			};
+			/**
+			 * Closes the post-trip drain window. Called once per consumed chunk after
+			 * that chunk is fully processed, so the frames it carried are finalized
+			 * before the stream is cut.
+			 */
+			const maybeAbortAfterRepetitionDrain = (): void => {
+				if (repetitionTrippedAt === undefined || requestSignal.aborted) return;
+				// Reuses the `stopReason === "length"` truncation check below: an open
+				// tool call whose `partialArgs` will not parse is still mid-flight.
+				const toolCallPending =
+					currentBlock?.type === "toolCall" &&
+					!isCompleteJson((currentBlock as { partialArgs?: string }).partialArgs);
+				const maxChunks = toolCallPending ? REPETITION_DRAIN_PENDING_TOOL_MAX_CHUNKS : REPETITION_DRAIN_MAX_CHUNKS;
+				const maxMs = toolCallPending ? REPETITION_DRAIN_PENDING_TOOL_MAX_MS : REPETITION_DRAIN_MAX_MS;
+				if (repetitionDrainedChunks >= maxChunks || Date.now() - repetitionTrippedAt >= maxMs) {
+					repetitionSelfAbort = true;
+					requestAbortController.abort();
+				}
+			};
+
+			// Reasoning-channel only — a fence token in visible prose must survive
+			// as text (CHANGELOG.md:1094), and the Kimi healer must not see this
+			// channel at all or its holdback buffer corrupts.
+			const thinkingFenceStripper = new ToolFenceStripper();
+			let lastThinkingSignature: string | undefined;
+
+			/** Returns the portion safe to emit — the whole chunk when the channel is unguarded. */
+			const feedRepetitionGuard = (
+				guard: StreamRepetitionGuard | undefined,
+				text: string,
+				channel: "text" | "thinking",
+			): string => {
+				if (!guard) return text;
+				const emit = guard.feed(text);
+				noteRepetitionTrip(guard, channel);
+				return emit;
+			};
+
 			const appendTextDelta = (text: string) => {
 				if (!text) return;
 				if (!firstTokenTime) firstTokenTime = Date.now();
-				appendText(output, stream, text);
+				const emit = feedRepetitionGuard(textRepetitionGuard, text, "text");
+				if (emit) appendText(output, stream, emit);
+			};
+			const emitThinkingText = (thinking: string, signature?: string) => {
+				if (!thinking) return;
+				const emit = feedRepetitionGuard(thinkingRepetitionGuard, thinking, "thinking");
+				if (emit) appendThinking(output, stream, emit, signature);
 			};
 			const appendThinkingDelta = (thinking: string, signature?: string) => {
 				if (!thinking) return;
 				if (!firstTokenTime) firstTokenTime = Date.now();
-				appendThinking(output, stream, thinking, signature);
+				lastThinkingSignature = signature;
+				emitThinkingText(thinkingFenceStripper.feed(thinking), signature);
 			};
 
 			const flushTaggedTextBuffer = () => {
@@ -1000,150 +1176,180 @@ export const streamOpenAICompletions: StreamFunction<"openai-completions"> = (
 			};
 
 			for await (const chunk of iterateWithNetworkErrorRetry()) {
-				if (!chunk || typeof chunk !== "object") continue;
-
-				// OpenAI documents ChatCompletionChunk.id as the unique chat completion identifier,
-				// and each chunk in a streamed completion carries the same id.
-				output.responseId ||= chunk.id;
-
-				if (chunk.usage) {
-					applyUsage(chunk.usage);
-				}
-
-				const choice = Array.isArray(chunk.choices) ? chunk.choices[0] : undefined;
-				if (!choice) continue;
-
-				if (!chunk.usage) {
-					const choiceUsage = getChoiceUsage(choice);
-					if (choiceUsage) {
-						applyUsage(choiceUsage);
+				// Counted before the guarded blocks below so chunks carrying no
+				// delta still spend the drain budget rather than extending it.
+				if (repetitionTrippedAt !== undefined) repetitionDrainedChunks += 1;
+
+				// Positive-form guards instead of early `continue`s: the drain check
+				// at the bottom of the body has to be reached on *every* path, or a
+				// provider that keeps emitting usage-only, keepalive-shaped,
+				// `choices`-less or malformed chunks after a trip never spends the
+				// budget and the request hangs open (#5627 review r5).
+				if (chunk && typeof chunk === "object") {
+					// OpenAI documents ChatCompletionChunk.id as the unique chat completion identifier,
+					// and each chunk in a streamed completion carries the same id.
+					output.responseId ||= chunk.id;
+
+					if (chunk.usage) {
+						applyUsage(chunk.usage);
 					}
-				}
 
-				if (choice.finish_reason) {
-					const finishReasonResult = mapStopReason(choice.finish_reason);
-					if (choice.finish_reason === "content_filter") {
-						markProviderSafetyStop(finishReasonResult.errorMessage);
-					} else if (!providerSafetyStop) {
-						output.stopReason = finishReasonResult.stopReason;
-						if (finishReasonResult.errorMessage) {
-							output.errorMessage = finishReasonResult.errorMessage;
+					const choice = Array.isArray(chunk.choices) ? chunk.choices[0] : undefined;
+					if (choice) {
+						if (!chunk.usage) {
+							const choiceUsage = getChoiceUsage(choice);
+							if (choiceUsage) {
+								applyUsage(choiceUsage);
+							}
 						}
-					}
-				}
 
-				if (choice.delta) {
-					if (typeof choice.delta.refusal === "string" && choice.delta.refusal.length > 0) {
-						appendTextDelta(choice.delta.refusal);
-						if (!providerSafetyStop) {
-							markProviderSafetyStop("Provider returned a safety refusal");
-						}
-					}
-					const normalizedDeltaText = normalizeStreamingContentText(choice.delta.content);
-					if (normalizedDeltaText.length > 0) {
-						if (!firstTokenTime) firstTokenTime = Date.now();
-						if (parseMiniMaxThinkTags) {
-							taggedTextBuffer += normalizedDeltaText;
-							flushTaggedTextBuffer();
-						} else if (stripDeepseekChatTemplateTokens) {
-							deepseekStripBuffer += normalizedDeltaText;
-							flushDeepseekStripBuffer(false);
-						} else if (kimiHealer) {
-							const hasStructuredToolCalls =
-								Array.isArray(choice.delta.tool_calls) && choice.delta.tool_calls.length > 0;
-							if (hasStructuredToolCalls) {
-								// Same chunk leaks markers AND carries structured tool_calls.
-								// Strip the marker text from visible output, but drop any
-								// synthesized calls so the structured payload stays the
-								// single source of truth (avoids double-dispatch).
-								const clean = kimiHealer.consumeWithoutCalls(normalizedDeltaText);
-								if (clean.length > 0) appendTextDelta(clean);
-							} else {
-								const clean = kimiHealer.feed(normalizedDeltaText);
-								if (clean.length > 0) appendTextDelta(clean);
-								flushHealedToolCalls();
+						if (choice.finish_reason) {
+							const finishReasonResult = mapStopReason(choice.finish_reason);
+							if (choice.finish_reason === "content_filter") {
+								markProviderSafetyStop(finishReasonResult.errorMessage);
+							} else if (!providerSafetyStop) {
+								output.stopReason = finishReasonResult.stopReason;
+								if (finishReasonResult.errorMessage) {
+									output.errorMessage = finishReasonResult.errorMessage;
+								}
 							}
-						} else {
-							appendTextDelta(normalizedDeltaText);
 						}
-					}
 
-					// Some endpoints return reasoning in reasoning_content (llama.cpp),
-					// or reasoning (other openai compatible endpoints)
-					// Use the first non-empty reasoning field to avoid duplication
-					// (e.g., chutes.ai returns both reasoning_content and reasoning with same content)
-					const reasoningFields = ["reasoning_content", "reasoning", "reasoning_text"];
-					let foundReasoningField: string | null = null;
-					for (const field of reasoningFields) {
-						if (
-							(choice.delta as any)[field] !== null &&
-							(choice.delta as any)[field] !== undefined &&
-							(choice.delta as any)[field].length > 0
-						) {
-							if (!foundReasoningField) {
-								foundReasoningField = field;
-								break;
+						if (choice.delta) {
+							if (typeof choice.delta.refusal === "string" && choice.delta.refusal.length > 0) {
+								appendTextDelta(choice.delta.refusal);
+								if (!providerSafetyStop) {
+									markProviderSafetyStop("Provider returned a safety refusal");
+								}
+							}
+							const normalizedDeltaText = normalizeStreamingContentText(choice.delta.content);
+							if (normalizedDeltaText.length > 0) {
+								if (!firstTokenTime) firstTokenTime = Date.now();
+								if (parseMiniMaxThinkTags) {
+									taggedTextBuffer += normalizedDeltaText;
+									flushTaggedTextBuffer();
+								} else if (stripDeepseekChatTemplateTokens) {
+									deepseekStripBuffer += normalizedDeltaText;
+									flushDeepseekStripBuffer(false);
+								} else if (kimiHealer) {
+									const hasStructuredToolCalls =
+										Array.isArray(choice.delta.tool_calls) && choice.delta.tool_calls.length > 0;
+									if (hasStructuredToolCalls) {
+										// Same chunk leaks markers AND carries structured tool_calls.
+										// Strip the marker text from visible output, but drop any
+										// synthesized calls so the structured payload stays the
+										// single source of truth (avoids double-dispatch).
+										const clean = kimiHealer.consumeWithoutCalls(normalizedDeltaText);
+										if (clean.length > 0) appendTextDelta(clean);
+									} else {
+										const clean = kimiHealer.feed(normalizedDeltaText);
+										if (clean.length > 0) appendTextDelta(clean);
+										flushHealedToolCalls();
+									}
+								} else {
+									appendTextDelta(normalizedDeltaText);
+								}
 							}
-						}
-					}
 
-					if (foundReasoningField) {
-						const delta = (choice.delta as any)[foundReasoningField];
-						appendThinkingDelta(delta, foundReasoningField);
-					}
+							// Some endpoints return reasoning in reasoning_content (llama.cpp),
+							// or reasoning (other openai compatible endpoints)
+							// Use the first non-empty reasoning field to avoid duplication
+							// (e.g., chutes.ai returns both reasoning_content and reasoning with same content)
+							const reasoningFields = ["reasoning_content", "reasoning", "reasoning_text"];
+							let foundReasoningField: string | null = null;
+							for (const field of reasoningFields) {
+								if (
+									(choice.delta as any)[field] !== null &&
+									(choice.delta as any)[field] !== undefined &&
+									(choice.delta as any)[field].length > 0
+								) {
+									if (!foundReasoningField) {
+										foundReasoningField = field;
+										break;
+									}
+								}
+							}
 
-					if (choice?.delta?.tool_calls && choice.delta.tool_calls.length > 0) {
-						for (const toolCall of choice.delta.tool_calls) {
-							if (currentBlock?.type !== "toolCall" || (toolCall.id && currentBlock.id !== toolCall.id)) {
-								finishCurrentBlock(currentBlock);
-								currentBlock = {
-									type: "toolCall",
-									id: toolCall.id || "",
-									name: toolCall.function?.name || "",
-									arguments: {},
-									partialArgs: "",
-								};
-								output.content.push(currentBlock);
-								stream.push({
-									type: "toolcall_start",
-									contentIndex: blockIndex(currentBlock),
-									partial: output,
-								});
+							if (foundReasoningField) {
+								const delta = (choice.delta as any)[foundReasoningField];
+								appendThinkingDelta(delta, foundReasoningField);
 							}
 
-							if (currentBlock.type === "toolCall") {
-								if (toolCall.id) currentBlock.id = toolCall.id;
-								if (toolCall.function?.name) currentBlock.name = toolCall.function.name;
-								let delta = "";
-								if (toolCall.function?.arguments) {
-									delta = toolCall.function.arguments;
-									currentBlock.partialArgs += toolCall.function.arguments;
-									currentBlock.arguments = parseStreamingJson(currentBlock.partialArgs);
+							if (choice?.delta?.tool_calls && choice.delta.tool_calls.length > 0) {
+								for (const toolCall of choice.delta.tool_calls) {
+									if (currentBlock?.type !== "toolCall" || (toolCall.id && currentBlock.id !== toolCall.id)) {
+										finishCurrentBlock(currentBlock);
+										currentBlock = {
+											type: "toolCall",
+											id: toolCall.id || "",
+											name: toolCall.function?.name || "",
+											arguments: {},
+											partialArgs: "",
+										};
+										output.content.push(currentBlock);
+										stream.push({
+											type: "toolcall_start",
+											contentIndex: blockIndex(currentBlock),
+											partial: output,
+										});
+									}
+
+									if (currentBlock.type === "toolCall") {
+										if (toolCall.id) currentBlock.id = toolCall.id;
+										if (toolCall.function?.name) currentBlock.name = toolCall.function.name;
+										let delta = "";
+										if (toolCall.function?.arguments) {
+											delta = toolCall.function.arguments;
+											currentBlock.partialArgs += toolCall.function.arguments;
+											currentBlock.arguments = parseStreamingJson(currentBlock.partialArgs);
+										}
+										stream.push({
+											type: "toolcall_delta",
+											contentIndex: blockIndex(currentBlock),
+											delta,
+											partial: output,
+										});
+									}
 								}
-								stream.push({
-									type: "toolcall_delta",
-									contentIndex: blockIndex(currentBlock),
-									delta,
-									partial: output,
-								});
 							}
-						}
-					}
 
-					const reasoningDetails = (choice.delta as any).reasoning_details;
-					if (reasoningDetails && Array.isArray(reasoningDetails)) {
-						for (const detail of reasoningDetails) {
-							if (detail.type === "reasoning.encrypted" && detail.id && detail.data) {
-								const matchingToolCall = output.content.find(
-									b => b.type === "toolCall" && b.id === detail.id,
-								) as ToolCall | undefined;
-								if (matchingToolCall) {
-									matchingToolCall.thoughtSignature = JSON.stringify(detail);
+							const reasoningDetails = (choice.delta as any).reasoning_details;
+							if (reasoningDetails && Array.isArray(reasoningDetails)) {
+								for (const detail of reasoningDetails) {
+									if (detail.type === "reasoning.encrypted" && detail.id && detail.data) {
+										const matchingToolCall = output.content.find(
+											b => b.type === "toolCall" && b.id === detail.id,
+										) as ToolCall | undefined;
+										if (matchingToolCall) {
+											matchingToolCall.thoughtSignature = JSON.stringify(detail);
+										}
+									}
 								}
 							}
 						}
 					}
 				}
+
+				// Single invariant: exactly one evaluation per consumed chunk, on
+				// every path. Two properties ride on this being the last *statement*
+				// of the body rather than a `finally`:
+				//
+				// (a) This chunk is fully processed — anything it carried (including
+				//     `tool_calls` frames) has landed. Only now may the drain window
+				//     close and cut the stream. The check must never run *before*
+				//     the chunk's processing.
+				// (b) A throwing chunk keeps its own transport facts. A `finally`
+				//     would also run when the body throws, and this check sets
+				//     `repetitionSelfAbort` — which the catch below branches on to
+				//     call `finalizeRepetitionGuardStop()`. A malformed payload
+				//     throwing inside the drain window would then be re-labelled as
+				//     the guard's own abort, discarding the real
+				//     errorMessage/errorStatus/transportFailure and flipping retry
+				//     admission — exactly the defect fixed in e577268e.
+				//
+				// Caller-abort priority is unchanged: a genuine caller abort still
+				// wins over the guard's self-abort, and a trip is still not a cancel.
+				maybeAbortAfterRepetitionDrain();
 			}
 
 			if (parseMiniMaxThinkTags && taggedTextBuffer.length > 0) {
@@ -1159,6 +1365,29 @@ export const streamOpenAICompletions: StreamFunction<"openai-completions"> = (
 				flushDeepseekStripBuffer(true);
 			}
 
+			// A partial fence held back at the last chunk never completed, so it was
+			// ordinary thinking text after all.
+			emitThinkingText(thinkingFenceStripper.flush(), lastThinkingSignature);
+
+			// Close each guard's in-progress unit now that no more text is coming:
+			// a final repeat with no trailing newline would otherwise go uncounted
+			// and the runaway turn would read as a healthy completion. Must run
+			// after the fence flush above, whose output feeds the thinking guard.
+			//
+			// Normal-completion path ONLY. Never finalize in the catch block: a
+			// stream that threw mid-repeat must keep its own transport facts rather
+			// than be reclassified as a decode loop (#5627 r4, commit c2aa25d30).
+			// No abort either — the stream has already ended, so aborting would set
+			// `repetitionSelfAbort` for nothing.
+			for (const [guard, channel] of [
+				[textRepetitionGuard, "text"],
+				[thinkingRepetitionGuard, "thinking"],
+			] as const) {
+				if (!guard) continue;
+				guard.finalize();
+				noteRepetitionTrip(guard, channel);
+			}
+
 			if (kimiHealer) {
 				const trailing = kimiHealer.flushPending();
 				if (trailing.length > 0) appendTextDelta(trailing);
@@ -1186,6 +1415,14 @@ export const streamOpenAICompletions: StreamFunction<"openai-completions"> = (
 
 			finishCurrentBlock(currentBlock);
 
+			// A repetition abort usually surfaces as a throw from the stream
+			// iterator, but a host that had already buffered the rest of the
+			// response finishes the loop normally instead. Same outcome either way.
+			if (repetitionTrip) {
+				finalizeRepetitionGuardStop();
+				return;
+			}
+
 			const firstEventTimeoutError = abortTracker.getLocalAbortReason();
 			if (firstEventTimeoutError) {
 				throw firstEventTimeoutError;
@@ -1220,6 +1457,15 @@ export const streamOpenAICompletions: StreamFunction<"openai-completions"> = (
 			stream.end();
 		} catch (error) {
 			for (const block of output.content) delete (block as any).index;
+			// Our own abort landed here. Classify it before the generic transport
+			// path turns it into a retryable provider error. A caller abort still
+			// wins: the user's cancel is the more meaningful intent. Keyed on the
+			// self-abort flag, not on `repetitionTrip`: a stall or transport error
+			// during the drain window must keep its own facts.
+			if (repetitionSelfAbort && !abortTracker.wasCallerAbort()) {
+				finalizeRepetitionGuardStop();
+				return;
+			}
 			const localAbortReason = abortTracker.getLocalAbortReason();
 			const normalizedError =
 				!streamConnected && model.provider === "alibaba-token-plan" && error instanceof APIConnectionTimeoutError
@@ -1576,7 +1822,12 @@ function buildParams(
 		delete params.tool_choice;
 	}
 
-	if (supportsReasoningParams && compat.thinkingFormat === "zai" && model.reasoning) {
+	if (supportsReasoningParams && getMiniMaxThinkingMode(model, resolvedBaseUrl) === "toggle") {
+		// MiniMax-M3 accepts an on/off switch, not reasoning_effort. Omitting
+		// the switch preserves the OpenAI-compatible endpoint's default (on).
+		if (options?.disableReasoning) params.thinking = { type: "disabled" };
+		else if (options?.reasoning) params.thinking = { type: "adaptive" };
+	} else if (supportsReasoningParams && compat.thinkingFormat === "zai" && model.reasoning) {
 		// Z.ai uses binary thinking: { type: "enabled" | "disabled" }
 		// Must explicitly disable since z.ai defaults to thinking enabled.
 		const enabled = options?.reasoning && !options?.disableReasoning;