@gajae-code/ai 0.12.8 → 0.12.10

package/src/models.json

@@ -1,5 +1,40 @@
 {
 	"alibaba-token-plan": {
+		"deepseek-v4-flash-0731": {
+			"id": "deepseek-v4-flash-0731",
+			"name": "DeepSeek V4 Flash 0731",
+			"api": "openai-completions",
+			"provider": "alibaba-token-plan",
+			"baseUrl": "https://token-plan.ap-southeast-1.maas.aliyuncs.com/compatible-mode/v1",
+			"reasoning": true,
+			"input": [
+				"text"
+			],
+			"cost": {
+				"input": 0,
+				"output": 0,
+				"cacheRead": 0,
+				"cacheWrite": 0
+			},
+			"contextWindow": 1000000,
+			"maxTokens": 384000,
+			"compat": {
+				"supportsDeveloperRole": false,
+				"supportsReasoningEffort": true,
+				"reasoningContentField": "reasoning_content",
+				"requiresReasoningContentForToolCalls": true
+			},
+			"thinking": {
+				"mode": "effort",
+				"minLevel": "low",
+				"maxLevel": "max",
+				"levels": [
+					"low",
+					"high",
+					"max"
+				]
+			}
+		},
 		"deepseek-v4-pro": {
 			"id": "deepseek-v4-pro",
 			"name": "DeepSeek V4 Pro",
@@ -58436,7 +58471,16 @@
 				"minLevel": "low",
 				"maxLevel": "max"
 			},
-			"applyPatchToolType": "freeform"
+			"applyPatchToolType": "freeform",
+			"longContextPricing": {
+				"threshold": 272000,
+				"cost": {
+					"input": 10,
+					"output": 45,
+					"cacheRead": 1,
+					"cacheWrite": 12.5
+				}
+			}
 		},
 		"gpt-5.6-luna": {
 			"id": "gpt-5.6-luna",
@@ -58450,10 +58494,10 @@
 				"image"
 			],
 			"cost": {
-				"input": 1,
-				"output": 6,
-				"cacheRead": 0.1,
-				"cacheWrite": 1.25
+				"input": 0.2,
+				"output": 1.2,
+				"cacheRead": 0.02,
+				"cacheWrite": 0.25
 			},
 			"contextWindow": 1050000,
 			"maxTokens": 128000,
@@ -58462,7 +58506,16 @@
 				"minLevel": "low",
 				"maxLevel": "max"
 			},
-			"applyPatchToolType": "freeform"
+			"applyPatchToolType": "freeform",
+			"longContextPricing": {
+				"threshold": 272000,
+				"cost": {
+					"input": 0.4,
+					"output": 1.8,
+					"cacheRead": 0.04,
+					"cacheWrite": 0.5
+				}
+			}
 		},
 		"gpt-5.6-sol": {
 			"id": "gpt-5.6-sol",
@@ -58488,7 +58541,16 @@
 				"minLevel": "low",
 				"maxLevel": "max"
 			},
-			"applyPatchToolType": "freeform"
+			"applyPatchToolType": "freeform",
+			"longContextPricing": {
+				"threshold": 272000,
+				"cost": {
+					"input": 10,
+					"output": 45,
+					"cacheRead": 1,
+					"cacheWrite": 12.5
+				}
+			}
 		},
 		"gpt-5.6-terra": {
 			"id": "gpt-5.6-terra",
@@ -58502,10 +58564,10 @@
 				"image"
 			],
 			"cost": {
-				"input": 2.5,
-				"output": 15,
-				"cacheRead": 0.25,
-				"cacheWrite": 3.125
+				"input": 2,
+				"output": 12,
+				"cacheRead": 0.2,
+				"cacheWrite": 2.5
 			},
 			"contextWindow": 1050000,
 			"maxTokens": 128000,
@@ -58514,7 +58576,16 @@
 				"minLevel": "low",
 				"maxLevel": "max"
 			},
-			"applyPatchToolType": "freeform"
+			"applyPatchToolType": "freeform",
+			"longContextPricing": {
+				"threshold": 272000,
+				"cost": {
+					"input": 4,
+					"output": 18,
+					"cacheRead": 0.4,
+					"cacheWrite": 5
+				}
+			}
 		},
 		"gpt-image-2": {
 			"id": "gpt-image-2",
@@ -59224,10 +59295,10 @@
 				"image"
 			],
 			"cost": {
-				"input": 1,
-				"output": 6,
-				"cacheRead": 0.1,
-				"cacheWrite": 1.25
+				"input": 0.2,
+				"output": 1.2,
+				"cacheRead": 0.02,
+				"cacheWrite": 0.25
 			},
 			"contextWindow": 272000,
 			"maxTokens": 128000,
@@ -59238,7 +59309,16 @@
 				"minLevel": "low",
 				"maxLevel": "max"
 			},
-			"applyPatchToolType": "freeform"
+			"applyPatchToolType": "freeform",
+			"longContextPricing": {
+				"threshold": 272000,
+				"cost": {
+					"input": 0.4,
+					"output": 1.8,
+					"cacheRead": 0.04,
+					"cacheWrite": 0.5
+				}
+			}
 		},
 		"gpt-5.6-sol": {
 			"id": "gpt-5.6-sol",
@@ -59266,7 +59346,16 @@
 				"minLevel": "low",
 				"maxLevel": "max"
 			},
-			"applyPatchToolType": "freeform"
+			"applyPatchToolType": "freeform",
+			"longContextPricing": {
+				"threshold": 272000,
+				"cost": {
+					"input": 10,
+					"output": 45,
+					"cacheRead": 1,
+					"cacheWrite": 12.5
+				}
+			}
 		},
 		"gpt-5.6-terra": {
 			"id": "gpt-5.6-terra",
@@ -59280,10 +59369,10 @@
 				"image"
 			],
 			"cost": {
-				"input": 2.5,
-				"output": 15,
-				"cacheRead": 0.25,
-				"cacheWrite": 3.125
+				"input": 2,
+				"output": 12,
+				"cacheRead": 0.2,
+				"cacheWrite": 2.5
 			},
 			"contextWindow": 272000,
 			"maxTokens": 128000,
@@ -59294,7 +59383,16 @@
 				"minLevel": "low",
 				"maxLevel": "max"
 			},
-			"applyPatchToolType": "freeform"
+			"applyPatchToolType": "freeform",
+			"longContextPricing": {
+				"threshold": 272000,
+				"cost": {
+					"input": 4,
+					"output": 18,
+					"cacheRead": 0.4,
+					"cacheWrite": 5
+				}
+			}
 		},
 		"gpt-image-2": {
 			"id": "gpt-image-2",
@@ -85511,4 +85609,4 @@
 			}
 		}
 	}
-}
+}

package/src/models.ts

@@ -1,4 +1,5 @@
 import { readFileSync } from "node:fs";
+import { getOpenAIModelCost } from "./model-pricing";
 import { isRetiredModelKey } from "./model-retirements";
 import { applyGeneratedModelPolicies, enrichModelThinking } from "./model-thinking";
 // `with { type: "file" }` is embedded by `bun build --compile` and resolves to
@@ -92,10 +93,12 @@ export function getBundledModels(provider: GeneratedProvider): Model<Api>[] {
 }
 
 export function calculateCost<TApi extends Api>(model: Model<TApi>, usage: Usage): Usage["cost"] {
-	usage.cost.input = (model.cost.input / 1000000) * usage.input;
-	usage.cost.output = (model.cost.output / 1000000) * usage.output;
-	usage.cost.cacheRead = (model.cost.cacheRead / 1000000) * usage.cacheRead;
-	usage.cost.cacheWrite = (model.cost.cacheWrite / 1000000) * usage.cacheWrite;
+	const inputTokens = usage.input + usage.cacheRead + usage.cacheWrite;
+	const pricing = getOpenAIModelCost(model, inputTokens) ?? model.cost;
+	usage.cost.input = (pricing.input / 1000000) * usage.input;
+	usage.cost.output = (pricing.output / 1000000) * usage.output;
+	usage.cost.cacheRead = (pricing.cacheRead / 1000000) * usage.cacheRead;
+	usage.cost.cacheWrite = (pricing.cacheWrite / 1000000) * usage.cacheWrite;
 	usage.cost.total = usage.cost.input + usage.cost.output + usage.cost.cacheRead + usage.cost.cacheWrite;
 	return usage.cost;
 }

package/src/prompts/composer-bash-policy-recovery.md

@@ -0,0 +1 @@
+A Composer bash policy block interrupted a shell attempt. This is not a terminal condition: continue the same task now. Do not retry repository file I/O in bash. Use the dedicated find, search, read, and edit tools for repository work, then continue with the next safe step.

package/src/prompts/cursor-composer-bash-policy-recovery.md

@@ -0,0 +1 @@
+A Composer bash policy block interrupted a shell attempt. This is not a terminal condition: continue the same task now. Do not retry repository file I/O in shell. Use Cursor-native read for files or directories, grep for search or globs, write for changes, and delete only when deletion is required; then continue with the next safe step.

package/src/prompts/cursor-composer-edit-discipline.md

@@ -0,0 +1,7 @@
+File-editing discipline for this Cursor Composer harness (this OVERRIDES contrary habits from your training):
+
+- Inspect repository files ONLY with Cursor-native read and grep: use read for file bodies or directories, and grep for content search or glob discovery. NEVER inspect repository files through shell commands (ls, find, fd, cat, sed, awk, grep, rg, head, tail, less, more) or scripts.
+- Modify files ONLY with Cursor-native write, or delete only when deletion is required. NEVER mutate files through shell redirection, tee, sed -i, perl -pi, inline python/node/bun scripts, or other out-of-band writes.
+- Re-read a file after any write before relying on its contents again. Do not fabricate line anchors, paths, tool names, or tool-call arguments.
+- Tool-call arguments must be the exact schema object requested by the native tool. Do not include Markdown, commentary, analysis text, or invented fields inside tool arguments.
+- Use shell only for terminal operations such as tests, builds, package scripts, and git commands. A shell command string must contain only the command itself; NEVER interleave reasoning or commentary into command strings or heredocs.

package/src/providers/composer-discipline.ts

@@ -1,3 +1,9 @@
+import composerBashPolicyRecoveryPrompt from "../prompts/composer-bash-policy-recovery.md" with { type: "text" };
+import cursorComposerBashPolicyRecoveryPrompt from "../prompts/cursor-composer-bash-policy-recovery.md" with {
+	type: "text",
+};
+import cursorComposerEditDisciplinePrompt from "../prompts/cursor-composer-edit-discipline.md" with { type: "text" };
+
 /**
  * Anchor/edit discipline for composer-harness models (xai grok-composer-*,
  * cursor composer-*).
@@ -30,6 +36,47 @@ export function isComposerHarnessModel(modelId: string): boolean {
 	return COMPOSER_MODEL_ID_PATTERN.test(modelId);
 }
 
+/** Stable text contract for a local shell rejection caused by Composer file-I/O discipline. */
+export const COMPOSER_BASH_POLICY_ERROR_PREFIX = "Composer bash policy blocked repository file I/O.";
+export const COMPOSER_BASH_POLICY_ERROR_CODE = "composer-bash-policy:repository-file-io";
+
+export type ComposerBashPolicyToolSurface = "generic" | "cursor";
+
+/**
+ * Format the model-visible policy rejection with a stable marker and the tool
+ * vocabulary the model actually receives on this provider surface.
+ */
+export function formatComposerBashPolicyError(surface: ComposerBashPolicyToolSurface = "generic"): string {
+	const recovery =
+		surface === "cursor"
+			? "Continue the same task with Cursor-native read, grep, write, or delete tools; do not retry repository file I/O through shell."
+			: "Continue the same task with find, search, read, and edit tools; do not retry repository file I/O through bash.";
+	return `${COMPOSER_BASH_POLICY_ERROR_PREFIX} [${COMPOSER_BASH_POLICY_ERROR_CODE}] Recovery required: ${recovery}`;
+}
+
+/**
+ * Matches both the structured current error and the original prefix so a
+ * resumed session can recover after an upgrade without string-version skew.
+ */
+export function isComposerBashPolicyBlockedError(text: string): boolean {
+	return text.includes(COMPOSER_BASH_POLICY_ERROR_PREFIX);
+}
+
+/**
+ * Matches only errors emitted directly by the current policy implementation.
+ * Live recovery must use this strict form so failed shell output that merely
+ * quotes a policy error cannot masquerade as the policy gate itself.
+ */
+export function isCurrentComposerBashPolicyBlockedError(text: string): boolean {
+	return text === formatComposerBashPolicyError("generic") || text === formatComposerBashPolicyError("cursor");
+}
+
+/** One bounded, tool-enabled retry instruction for generic Composer agent loops. */
+export const COMPOSER_BASH_POLICY_RECOVERY_PROMPT = composerBashPolicyRecoveryPrompt;
+
+/** One bounded, tool-enabled retry instruction for Cursor's native remote tool surface. */
+export const CURSOR_COMPOSER_BASH_POLICY_RECOVERY_PROMPT = cursorComposerBashPolicyRecoveryPrompt;
+
 export const COMPOSER_EDIT_DISCIPLINE_PROMPT = `File-editing discipline for this Composer harness (this OVERRIDES contrary habits from your training):
 
 - Discover file names ONLY with the find tool; search file contents ONLY with the search tool; read file bodies or line ranges ONLY with the read tool. NEVER inspect repository files through shell commands (ls, find, fd, cat, sed, awk, grep, rg, head, tail, less, more) or scripts — that output carries no hashline anchors and bypasses the agent's safety limits.
@@ -39,3 +86,10 @@ export const COMPOSER_EDIT_DISCIPLINE_PROMPT = `File-editing discipline for this
 - If an edit is rejected with "anchors do not match", the rejection message prints the current lines WITH fresh anchors. Retry using exactly those printed anchors.
 - Tool-call arguments must be the exact JSON/schema object requested by the tool. Do not include Markdown, commentary, analysis text, or invented fields inside tool arguments.
 - Use bash only for terminal operations such as tests, builds, package scripts, and git commands. A shell command string must contain only the command itself; NEVER interleave reasoning or commentary into command strings or heredocs.`;
+
+/**
+ * Cursor executes a different native tool vocabulary from the generic agent
+ * loop. Keep this prompt separate so Composer is never told to call `edit`,
+ * `find`, or `search` when those names are unavailable remotely.
+ */
+export const CURSOR_COMPOSER_EDIT_DISCIPLINE_PROMPT = cursorComposerEditDisciplinePrompt;

package/src/providers/cursor.ts

@@ -30,7 +30,7 @@ import { AssistantMessageEventStream } from "../utils/event-stream";
 import { parseStreamingJson } from "../utils/json-parse";
 import { formatErrorMessageWithRetryAfter } from "../utils/retry-after";
 import { flattenToolRootCombinators, toolWireSchema } from "../utils/schema";
-import { COMPOSER_EDIT_DISCIPLINE_PROMPT, isComposerHarnessModel } from "./composer-discipline";
+import { CURSOR_COMPOSER_EDIT_DISCIPLINE_PROMPT, isComposerHarnessModel } from "./composer-discipline";
 import { CURSOR_CLIENT_VERSION } from "./cursor/client-version";
 import type { McpToolDefinition } from "./cursor/gen/agent_pb";
 import {
@@ -2329,7 +2329,7 @@ export function buildCursorSystemPromptJsons(systemPrompt: readonly string[] | u
 	// Composer-harness models need anchor/edit discipline pinned ahead of any
 	// host/default prompt (see composer-discipline.ts for the observed failure modes).
 	if (modelId !== undefined && isComposerHarnessModel(modelId)) {
-		jsons.unshift(JSON.stringify({ role: "system", content: COMPOSER_EDIT_DISCIPLINE_PROMPT }));
+		jsons.unshift(JSON.stringify({ role: "system", content: CURSOR_COMPOSER_EDIT_DISCIPLINE_PROMPT }));
 	}
 	return jsons;
 }

package/src/providers/openai-completions.ts

@@ -1,5 +1,5 @@
 import { $credentialEnv, $env, extractHttpStatusFromError, logger } from "@gajae-code/utils";
-import OpenAI from "openai";
+import OpenAI, { APIConnectionTimeoutError } from "openai";
 import type {
 	ChatCompletionAssistantMessageParam,
 	ChatCompletionChunk,
@@ -47,6 +47,7 @@ import {
 	rewriteCopilotError,
 } from "../utils/http-inspector";
 import {
+	FirstEventTimeoutError,
 	getOpenAIStreamIdleTimeoutMs,
 	getProviderFirstEventTimeoutFallbackMs,
 	getStreamFirstEventTimeoutMs,
@@ -500,8 +501,6 @@ function getTrailingPartialDeepseekToken(text: string): string {
 	return tail;
 }
 
-const ALIBABA_TOKEN_PLAN_FIRST_EVENT_TIMEOUT_MS = 300_000;
-
 const OPENAI_COMPLETIONS_FIRST_EVENT_TIMEOUT_MESSAGE =
 	"OpenAI completions stream timed out while waiting for the first event";
 
@@ -515,6 +514,7 @@ export const streamOpenAICompletions: StreamFunction<"openai-completions"> = (
 	(async () => {
 		const startTime = Date.now();
 		let firstTokenTime: number | undefined;
+		let streamConnected = false;
 		let getCapturedErrorResponse: (() => CapturedHttpErrorResponse | undefined) | undefined;
 
 		const output: AssistantMessage = createInitialResponsesAssistantMessage(model.api, model.provider, model.id);
@@ -640,10 +640,8 @@ export const streamOpenAICompletions: StreamFunction<"openai-completions"> = (
 					openaiStream = await createCompletionsStream("none");
 				}
 			}
-			const firstEventFallbackMs =
-				model.provider === "alibaba-token-plan"
-					? ALIBABA_TOKEN_PLAN_FIRST_EVENT_TIMEOUT_MS
-					: getProviderFirstEventTimeoutFallbackMs(model.provider);
+			streamConnected = true;
+			const firstEventFallbackMs = getProviderFirstEventTimeoutFallbackMs(model.provider);
 			const firstEventTimeoutMs =
 				options?.streamFirstEventTimeoutMs ?? getStreamFirstEventTimeoutMs(idleTimeoutMs, firstEventFallbackMs);
 			if (premiumRequestsTotal !== undefined) {
@@ -1049,20 +1047,25 @@ export const streamOpenAICompletions: StreamFunction<"openai-completions"> = (
 		} catch (error) {
 			for (const block of output.content) delete (block as any).index;
 			const localAbortReason = abortTracker.getLocalAbortReason();
+			const normalizedError =
+				!streamConnected && model.provider === "alibaba-token-plan" && error instanceof APIConnectionTimeoutError
+					? new FirstEventTimeoutError(OPENAI_COMPLETIONS_FIRST_EVENT_TIMEOUT_MESSAGE)
+					: error;
 			const capturedErrorResponse = getCapturedErrorResponse?.();
 			output.stopReason = abortTracker.wasCallerAbort() ? "aborted" : "error";
 			output.errorStatus =
-				extractHttpStatusFromError(localAbortReason ?? error) ??
+				extractHttpStatusFromError(localAbortReason ?? normalizedError) ??
 				(localAbortReason ? undefined : capturedErrorResponse?.status);
 			output.transportFailure = localAbortReason
 				? transportFailureFacts(localAbortReason)
-				: transportFailureFacts(error, capturedErrorResponse);
+				: transportFailureFacts(normalizedError, capturedErrorResponse);
 			output.errorMessage =
-				localAbortReason?.message ?? (await finalizeErrorMessage(error, rawRequestDump, capturedErrorResponse));
+				localAbortReason?.message ??
+				(await finalizeErrorMessage(normalizedError, rawRequestDump, capturedErrorResponse));
 			// Some providers via OpenRouter include extra details here.
-			const rawMetadata = (error as { error?: { metadata?: { raw?: string } } })?.error?.metadata?.raw;
+			const rawMetadata = (normalizedError as { error?: { metadata?: { raw?: string } } })?.error?.metadata?.raw;
 			if (rawMetadata) output.errorMessage += `\n${rawMetadata}`;
-			output.errorMessage = rewriteCopilotError(output.errorMessage, error, model.provider);
+			output.errorMessage = rewriteCopilotError(output.errorMessage, normalizedError, model.provider);
 			if (hasContentFilterSafetyCode(capturedErrorResponse)) {
 				output.errorKind = "provider_safety_stop";
 			}
@@ -1235,15 +1238,20 @@ async function createClient(
 	// in the IIFE.
 	// A caller may raise `StreamOptions.streamFirstEventTimeoutMs` for a slow-
 	// before-headers provider; respect it so the SDK doesn't give up before the
-	// wrapping watchdog arms. An explicit `0` disables the first-event watchdog,
+	// wrapping watchdog arms. Provider-specific fallbacks apply only when the
+	// caller does not pin a value, so an explicit nonzero override must beat that
+	// fallback even when it is shorter. An explicit `0` disables the watchdog,
 	// and the SDK treats `timeout: 0` as an immediate timeout, so do not pass a
 	// request timeout in that case.
-	const envSdkTimeoutMs = getStreamFirstEventTimeoutMs(getOpenAIStreamIdleTimeoutMs());
+	const providerFirstEventFallbackMs = getProviderFirstEventTimeoutFallbackMs(model.provider);
+	const envSdkTimeoutMs = getStreamFirstEventTimeoutMs(getOpenAIStreamIdleTimeoutMs(), providerFirstEventFallbackMs);
 	const sdkTimeoutMs =
 		streamFirstEventTimeoutOverride === 0
 			? undefined
 			: streamFirstEventTimeoutOverride !== undefined
-				? Math.max(envSdkTimeoutMs ?? 0, streamFirstEventTimeoutOverride)
+				? providerFirstEventFallbackMs !== undefined
+					? streamFirstEventTimeoutOverride
+					: Math.max(envSdkTimeoutMs ?? 0, streamFirstEventTimeoutOverride)
 				: envSdkTimeoutMs;
 	return {
 		client: new OpenAI({

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

@@ -960,20 +960,31 @@ export function populateResponsesUsageFromResponse(
 				input_tokens?: number | null;
 				output_tokens?: number | null;
 				total_tokens?: number | null;
-				input_tokens_details?: { cached_tokens?: number | null } | null;
+				input_tokens_details?: {
+					cached_tokens?: number | null;
+					cache_write_tokens?: number | null;
+				} | null;
 				output_tokens_details?: { reasoning_tokens?: number | null } | null;
 		  }
 		| null
 		| undefined,
 ): void {
 	if (!usage) return;
+	const inputTokens = usage.input_tokens || 0;
 	const cachedTokens = usage.input_tokens_details?.cached_tokens || 0;
+	const reportedCacheWrite = usage.input_tokens_details?.cache_write_tokens || 0;
+	const cacheWriteTokens =
+		Number.isSafeInteger(reportedCacheWrite) &&
+		reportedCacheWrite >= 0 &&
+		cachedTokens + reportedCacheWrite <= inputTokens
+			? reportedCacheWrite
+			: 0;
 	const reasoningTokens = usage.output_tokens_details?.reasoning_tokens || 0;
 	output.usage = {
-		input: (usage.input_tokens || 0) - cachedTokens,
+		input: Math.max(0, inputTokens - cachedTokens - cacheWriteTokens),
 		output: usage.output_tokens || 0,
 		cacheRead: cachedTokens,
-		cacheWrite: 0,
+		cacheWrite: cacheWriteTokens,
 		totalTokens: usage.total_tokens || 0,
 		...(reasoningTokens > 0 ? { reasoningTokens } : {}),
 		cost: { input: 0, output: 0, cacheRead: 0, cacheWrite: 0, total: 0 },

package/src/providers/openai-responses.ts

@@ -1,5 +1,5 @@
 import { $credentialEnv, extractHttpStatusFromError, logger, structuredCloneJSON } from "@gajae-code/utils";
-import OpenAI from "openai";
+import OpenAI, { APIConnectionTimeoutError } from "openai";
 import type {
 	Tool as OpenAITool,
 	ResponseCreateParamsStreaming,
@@ -38,7 +38,9 @@ import { AssistantMessageEventStream } from "../utils/event-stream";
 import { transportFailureFacts } from "../utils/fallback-transport";
 import { finalizeErrorMessage, type RawHttpRequestDump, rewriteCopilotError } from "../utils/http-inspector";
 import {
+	FirstEventTimeoutError,
 	getOpenAIStreamIdleTimeoutMs,
+	getProviderFirstEventTimeoutFallbackMs,
 	getStreamFirstEventTimeoutMs,
 	iterateWithIdleTimeout,
 } from "../utils/idle-iterator";
@@ -124,7 +126,6 @@ export interface OpenAIResponsesOptions extends StreamOptions {
 }
 
 const OPENAI_RESPONSES_PROVIDER_SESSION_STATE_PREFIX = "openai-responses:";
-const ALIBABA_TOKEN_PLAN_FIRST_EVENT_TIMEOUT_MS = 300_000;
 const OPENAI_RESPONSES_FIRST_EVENT_TIMEOUT_MESSAGE =
 	"OpenAI responses stream timed out while waiting for the first event";
 const OPENAI_DEFAULT_BASE_URL = "https://api.openai.com/v1";
@@ -314,6 +315,7 @@ export const streamOpenAIResponses: StreamFunction<"openai-responses"> = (
 	(async () => {
 		const startTime = Date.now();
 		let firstTokenTime: number | undefined;
+		let streamConnected = false;
 
 		const output: AssistantMessage = createInitialResponsesAssistantMessage(
 			"openai-responses",
@@ -392,8 +394,8 @@ export const streamOpenAIResponses: StreamFunction<"openai-responses"> = (
 				await notifyProviderResponse(options, response, model, request_id);
 				return data;
 			});
-			const firstEventFallbackMs =
-				model.provider === "alibaba-token-plan" ? ALIBABA_TOKEN_PLAN_FIRST_EVENT_TIMEOUT_MS : undefined;
+			streamConnected = true;
+			const firstEventFallbackMs = getProviderFirstEventTimeoutFallbackMs(model.provider);
 			const firstEventTimeoutMs =
 				options?.streamFirstEventTimeoutMs ?? getStreamFirstEventTimeoutMs(idleTimeoutMs, firstEventFallbackMs);
 			if (premiumRequestsTotal !== undefined) output.usage.premiumRequests = premiumRequestsTotal;
@@ -447,11 +449,16 @@ export const streamOpenAIResponses: StreamFunction<"openai-responses"> = (
 		} catch (error) {
 			for (const block of output.content) delete (block as { index?: number }).index;
 			const localAbortReason = abortTracker.getLocalAbortReason();
+			const normalizedError =
+				!streamConnected && model.provider === "alibaba-token-plan" && error instanceof APIConnectionTimeoutError
+					? new FirstEventTimeoutError(OPENAI_RESPONSES_FIRST_EVENT_TIMEOUT_MESSAGE)
+					: error;
 			output.stopReason = abortTracker.wasCallerAbort() ? "aborted" : "error";
-			output.errorStatus = extractHttpStatusFromError(localAbortReason ?? error);
-			output.transportFailure = transportFailureFacts(localAbortReason ?? error);
-			output.errorMessage = localAbortReason?.message ?? (await finalizeErrorMessage(error, rawRequestDump));
-			output.errorMessage = rewriteCopilotError(output.errorMessage, error, model.provider);
+			output.errorStatus = extractHttpStatusFromError(localAbortReason ?? normalizedError);
+			output.transportFailure = transportFailureFacts(localAbortReason ?? normalizedError);
+			output.errorMessage =
+				localAbortReason?.message ?? (await finalizeErrorMessage(normalizedError, rawRequestDump));
+			output.errorMessage = rewriteCopilotError(output.errorMessage, normalizedError, model.provider);
 			// Explicitly mark the poisoned-history rejection so the shared
 			// `invalid_prompt` contract is present even when the SDK error surfaces
 			// only a message (no structured code). This keeps the responses

package/src/providers/register-builtins.ts

@@ -25,6 +25,7 @@ import { AssistantMessageEventStream as EventStreamImpl } from "../utils/event-s
 import { transportFailureFacts } from "../utils/fallback-transport";
 import {
 	FirstEventTimeoutError,
+	getProviderFirstEventTimeoutFallbackMs,
 	getStreamFirstEventTimeoutMs,
 	getStreamIdleTimeoutMs,
 	iterateWithIdleTimeout,
@@ -197,13 +198,12 @@ interface LazyStreamLimits {
 const GOOGLE_GEMINI_CLI_LAZY_STREAM_LIMITS: LazyStreamLimits = {
 	defaultFirstEventTimeoutMs: 300_000,
 };
-const SLOW_FIRST_EVENT_PROVIDERS = new Set(["alibaba-token-plan", "kimi-code"]);
 
 /**
  * Resolves the first-event timeout fallback for the outer lazy-stream watchdog.
  * A configured wrapper-specific fallback (from `LazyStreamLimits`) always wins;
- * otherwise providers known to have slow first events get a five-minute floor
- * matching their inner provider-level override. Returns `undefined` for
+ * otherwise providers known to have slow first events use the same centralized
+ * fallback as their inner provider-level watchdog. Returns `undefined` for
  * providers that should use the shared default.
  */
 export function resolveLazyStreamFirstEventFallbackMs(
@@ -211,7 +211,7 @@ export function resolveLazyStreamFirstEventFallbackMs(
 	configuredFallbackMs?: number,
 ): number | undefined {
 	if (configuredFallbackMs !== undefined) return configuredFallbackMs;
-	return SLOW_FIRST_EVENT_PROVIDERS.has(provider) ? 300_000 : undefined;
+	return getProviderFirstEventTimeoutFallbackMs(provider);
 }
 
 function forwardStream<TApi extends Api>(