@eddyskywalker/dsh-chatgpt-subscription 0.11.4 → 0.11.7

package/lib/index.js

@@ -5629,6 +5629,32 @@ function withWritable(value) {
 	};
 }
 //#endregion
+//#region src/host/common/context-overflow.ts
+/** Stable code understood by Harness overflow recovery, including older hosts. */
+const CONTEXT_OVERFLOW_CODE = "CONTEXT_WINDOW_EXCEEDED";
+/** Inspect provider error fields only; never classify request bodies or successful output. */
+function isContextOverflow(error) {
+	if (typeof error === "string") {
+		try {
+			const parsed = JSON.parse(error);
+			if (typeof parsed === "object" && parsed !== null) return isContextOverflow(parsed);
+		} catch {}
+		return /\bcontext[_ -](?:length|window)[_ -](?:exceeded|overflow|limit[_ -]exceeded)\b/i.test(error) || /\b(?:maximum|max)(?: allowed| supported)? context (?:length|window)\b/i.test(error) && /\b(?:exceed(?:s|ed)?|requested|resulted in|reduce|too (?:long|large))\b/i.test(error) || /\b(?:prompt|input|request) (?:is )?too (?:long|large) for (?:the |this )?(?:model(?:'s)? )?context\b/i.test(error) || /\b(?:prompt|input) (?:is )?too long:\s*[\d,]+ tokens?\s*>\s*[\d,]+\s*(?:tokens?\s*)?maximum\b/i.test(error) || /\brequest exceeded model token limit\b/i.test(error) || /\b(?:input|prompt|request|messages?)\b.{0,40}\bexceed(?:s|ed)?\b.{0,40}\b(?:model(?:'s)? )?context (?:length|window)\b/i.test(error);
+	}
+	if (typeof error !== "object" || error === null || Array.isArray(error)) return false;
+	const fields = error;
+	return [
+		fields.code,
+		fields.type,
+		fields.message,
+		fields.error
+	].some((value) => typeof value === "string" ? isContextOverflow(value) : value !== error && value !== void 0 && isContextOverflow(value));
+}
+/** Only request-rejection statuses can be reclassified as context overflow. */
+function isHttpContextOverflow(status, error) {
+	return (status === 400 || status === 413 || status === 422) && isContextOverflow(error);
+}
+//#endregion
 //#region src/host/common/brand-compat.ts
 const toToolCallId = (id) => {
 	const mod = LlmModule;
@@ -6578,7 +6604,7 @@ async function* parseResponsesStream(response, signal, hiddenSandboxControls = /
 			const message = string(error?.message) ?? "Codex generation failed.";
 			const rawCode = string(error?.code)?.toLowerCase();
 			const isOverload = message.toLowerCase().includes("overload") || message.toLowerCase().includes("server error") || rawCode === "server_error" || rawCode === "service_unavailable" || rawCode === "internal_error";
-			throw new LlmError(message, message.toLowerCase().includes("rate limit") || rawCode === "rate_limit" ? "RATE_LIMIT" : isOverload ? "SERVER_ERROR" : string(error?.code)?.toUpperCase() ?? "PROVIDER_ERROR");
+			throw new LlmError(message, message.toLowerCase().includes("rate limit") || rawCode === "rate_limit" ? "RATE_LIMIT" : isOverload ? "SERVER_ERROR" : isContextOverflow(error) ? CONTEXT_OVERFLOW_CODE : string(error?.code)?.toUpperCase() ?? "PROVIDER_ERROR");
 		}
 	};
 	try {
@@ -6746,6 +6772,7 @@ async function responseError(response) {
 		...requestId ? { requestId: ProviderRequestId(requestId) } : {},
 		...response.status === 429 ? { providerRetryAfterMs: retryAfterMs$1(response.headers) } : {}
 	};
+	if (isHttpContextOverflow(response.status, detail)) return new LlmError(`Codex context window exceeded: ${detail}`, CONTEXT_OVERFLOW_CODE, options);
 	if (response.status === 401) return new LlmError("ChatGPT sign-in has expired. Sign in again.", "AUTH", options);
 	if (response.status === 404) return new LlmError(`Codex model or resource not found (${response.status})${detail ? `: ${detail}` : "."}`, "NOT_FOUND", options);
 	if (response.status === 429) return new LlmError("Codex rate limit reached.", "RATE_LIMIT", options);
@@ -9142,6 +9169,8 @@ function processStreamLine$2(line, state) {
 	const chunk = safeJsonParse$5(json);
 	if (!isRecord$21(chunk)) return [];
 	const responseData = isRecord$21(chunk.response) ? chunk.response : chunk;
+	const error = isRecord$21(chunk.error) ? chunk.error : isRecord$21(responseData.error) ? responseData.error : void 0;
+	if (error !== void 0) throw new LlmError$1(`Antigravity stream error: ${asString$14(error.message) ?? "unknown error"}`, isContextOverflow(error) ? CONTEXT_OVERFLOW_CODE : "PROVIDER_ERROR");
 	const candidates = Array.isArray(responseData.candidates) ? responseData.candidates : [];
 	const candidate = isRecord$21(candidates[0]) ? candidates[0] : void 0;
 	const content = isRecord$21(candidate?.content) ? candidate.content : void 0;
@@ -9663,7 +9692,7 @@ var AntigravityAdapter = class extends LlmAdapter {
 			const status = response?.status ?? 500;
 			const errText = await response?.text().catch(() => "");
 			if (status === 429) throw new LlmError(`Antigravity 账号配额已耗尽或请求受限 (429 RESOURCE_EXHAUSTED)。请在插件设置页查看配额剩余百分比及重置倒计时。原始响应: ${errText || "No response"}`, "RATE_LIMIT", { status: 429 });
-			throw new LlmError(`Antigravity API error (${status}): ${errText || "No response"}`, "PROVIDER_ERROR", { status });
+			throw new LlmError(`Antigravity API error (${status}): ${errText || "No response"}`, isHttpContextOverflow(status, errText) ? CONTEXT_OVERFLOW_CODE : "PROVIDER_ERROR", { status });
 		}
 		if (!response.body) throw new LlmError("Antigravity returned empty response body", "PROVIDER_ERROR");
 		const reader = response.body.getReader();
@@ -9950,6 +9979,103 @@ function registerAntigravityRoutes(ctx, store, modelSettings, preferences, fetch
 	});
 }
 //#endregion
+//#region src/host/common/context-budget-fetch.ts
+/**
+* Recover a rejected input+output budget without dropping model-visible history.
+* Only explicit wire output controls and provider-confirmed counts qualify.
+* HTTP success/SSE, auth, quota, and byte-limit failures retain their old paths.
+*/
+function fetchWithContextBudgetRecovery(fetchFn) {
+	return async (input, init) => {
+		const response = await fetchFn(input, init);
+		if (response.status !== 400 && response.status !== 422) return response;
+		if (typeof init?.body !== "string") return response;
+		let body;
+		try {
+			const parsed = JSON.parse(init.body);
+			if (typeof parsed !== "object" || parsed === null || Array.isArray(parsed)) return response;
+			body = parsed;
+		} catch {
+			return response;
+		}
+		const keys = [
+			"max_tokens",
+			"max_completion_tokens",
+			"max_output_tokens"
+		].filter((key) => key in body);
+		if (keys.length !== 1) return response;
+		const key = keys[0];
+		const requested = body[key];
+		if (typeof requested !== "number" || !Number.isSafeInteger(requested) || requested <= 0) return response;
+		const reduced = reducedOutputBudget(await boundedDiagnostic(response), requested);
+		if (reduced === void 0) return response;
+		const thinking = body.thinking;
+		if (typeof thinking === "object" && thinking !== null) {
+			const budget = thinking.budget_tokens;
+			if (typeof budget === "number" && reduced <= budget) return response;
+		}
+		init.signal?.throwIfAborted();
+		await response.body?.cancel().catch(() => void 0);
+		return fetchFn(input, {
+			...init,
+			body: JSON.stringify({
+				...body,
+				[key]: reduced
+			})
+		});
+	};
+}
+/** Read a bounded clone so a non-recoverable refusal remains readable by its owner. */
+async function boundedDiagnostic(response) {
+	const reader = response.clone().body?.getReader();
+	if (reader === void 0) return "";
+	const decoder = new TextDecoder();
+	let size = 0;
+	let text = "";
+	try {
+		for (;;) {
+			const { done, value } = await reader.read();
+			if (done) return text + decoder.decode();
+			size += value.byteLength;
+			if (size > 16384) return "";
+			text += decoder.decode(value, { stream: true });
+		}
+	} catch {
+		return "";
+	} finally {
+		reader.cancel().catch(() => void 0);
+	}
+}
+function diagnosticMessage(value, depth = 0) {
+	if (depth > 4) return void 0;
+	if (typeof value === "string") return value;
+	if (typeof value !== "object" || value === null || Array.isArray(value)) return void 0;
+	const fields = value;
+	return diagnosticMessage(fields.error, depth + 1) ?? diagnosticMessage(fields.message, depth + 1);
+}
+/** Never infer capacity from a model name or from an echoed request. */
+function reducedOutputBudget(detail, requested) {
+	let message;
+	try {
+		message = diagnosticMessage(JSON.parse(detail));
+	} catch {
+		message = detail;
+	}
+	if (message === void 0) return void 0;
+	const match = /maximum context length is ([\d,]+) tokens[\s\S]*?requested ([\d,]+) tokens[\s\S]*?\(([\d,]+) in the messages, ([\d,]+) in the completion\)/i.exec(message);
+	if (match === null) return void 0;
+	const [window, total, prompt, completion] = match.slice(1).map((value) => Number(value.replaceAll(",", "")));
+	if (![
+		window,
+		total,
+		prompt,
+		completion
+	].every((value) => Number.isSafeInteger(value) && value > 0)) return void 0;
+	if (completion !== requested || prompt + completion !== total || total <= window) return void 0;
+	const available = window - prompt - 1024;
+	return available >= 1024 && available < requested ? available : void 0;
+}
+//#endregion
 //#region src/host/command-code/model-catalog.ts
 /** Every model the Command Code registry describes, in registry order. */
 const COMMAND_CODE_MODELS = [
@@ -13126,6 +13252,7 @@ function processOpenAIStreamLine$1(line, state) {
 	const chunk = safeJsonParse$4(payload);
 	if (!isRecord$19(chunk)) return [];
 	const out = [];
+	if (isRecord$19(chunk.error)) throw new LlmError$1(`Command Code stream error: ${asString$12(chunk.error.message) ?? "unknown error"}`, isContextOverflow(chunk.error) ? CONTEXT_OVERFLOW_CODE : "PROVIDER_ERROR");
 	const usage = isRecord$19(chunk.usage) ? chunk.usage : void 0;
 	if (usage) {
 		state.sawUsage = true;
@@ -13395,7 +13522,10 @@ function processAnthropicStreamLine$1(line, state) {
 		state.done = true;
 		return closeStream$4(state);
 	}
-	if (type === "error") throw new LlmError$1(`Command Code stream error: ${asString$12((isRecord$19(event.error) ? event.error : {}).message) ?? "unknown error"}`, "PROVIDER_ERROR");
+	if (type === "error") {
+		const error = isRecord$19(event.error) ? event.error : {};
+		throw new LlmError$1(`Command Code stream error: ${asString$12(error.message) ?? "unknown error"}`, isContextOverflow(error) ? CONTEXT_OVERFLOW_CODE : "PROVIDER_ERROR");
+	}
 	return out;
 }
 function tokenUsage$4(state) {
@@ -13506,7 +13636,7 @@ function processResponsesStreamLine(line, state) {
 		state.done = true;
 		const errorObj = isRecord$19(parsed.error) ? parsed.error : isRecord$19(parsed.response) && isRecord$19(parsed.response.error) ? parsed.response.error : null;
 		const detail = (typeof parsed.message === "string" ? parsed.message : "") || (errorObj && typeof errorObj.message === "string" ? errorObj.message : "") || (isRecord$19(parsed.response) && typeof parsed.response.error === "string" ? parsed.response.error : typeof parsed.error === "string" ? parsed.error : "");
-		throw new LlmError$1(detail === "" ? "Command Code responses stream failed" : `Command Code responses stream failed: ${detail}`, "PROVIDER_ERROR");
+		throw new LlmError$1(detail === "" ? "Command Code responses stream failed" : `Command Code responses stream failed: ${detail}`, isContextOverflow(errorObj ?? detail) ? CONTEXT_OVERFLOW_CODE : "PROVIDER_ERROR");
 	}
 	return [];
 }
@@ -14059,7 +14189,7 @@ var CommandCodeAdapter = class extends LlmAdapter {
 				"anthropic-version": "2023-06-01"
 			} : baseHeaders;
 			try {
-				response = await fetchFn(endpoint, {
+				response = await fetchWithContextBudgetRecovery(fetchFn)(endpoint, {
 					method: "POST",
 					headers,
 					body,
@@ -14106,7 +14236,7 @@ var CommandCodeAdapter = class extends LlmAdapter {
 				});
 			}
 			if (status >= 500) throw new LlmError(`${PROVIDER_NAME$5} upstream server error (${status}): ${detail || "No response"}`, "SERVER", { status });
-			throw new LlmError(`${PROVIDER_NAME$5} API error (${status}): ${detail || "No response"}`, "PROVIDER_ERROR", { status });
+			throw new LlmError(`${PROVIDER_NAME$5} API error (${status}): ${detail || "No response"}`, isHttpContextOverflow(status, detail) ? CONTEXT_OVERFLOW_CODE : "PROVIDER_ERROR", { status });
 		}
 		if (response.body === null) throw new LlmError("Command Code returned an empty response body", "PROVIDER_ERROR");
 		const reader = response.body.getReader();
@@ -15860,16 +15990,57 @@ function accumulateToolCall(raw, pending, openaiShape) {
 * making the line unusable.
 */
 async function loadCatalog$1(fetchFn, credentials) {
-	const response = await fetchFn(`${CLOUD_BASE_URL}${NATIVE_TAGS_PATH}`, {
-		method: "GET",
-		headers: headersFor(credentials, "application/json"),
-		signal: AbortSignal.timeout(CATALOG_TIMEOUT_MS)
-	});
-	if (!response.ok) return [];
-	const data = await response.json();
-	if (typeof data !== "object" || data === null) return [];
+	return (await fetchCatalog(fetchFn, credentials)).models;
+}
+/**
+* The same read as {@link loadCatalog}, keeping the reason it failed.
+*
+* Never throws, for the reason its wrapper has always given. What it adds is that
+* "failed" stops being one opaque state, so the settings card can name the
+* cause instead of asking the user to guess.
+*/
+async function fetchCatalog(fetchFn, credentials) {
+	let response;
+	try {
+		response = await fetchFn(`${CLOUD_BASE_URL}${NATIVE_TAGS_PATH}`, {
+			method: "GET",
+			headers: headersFor(credentials, "application/json"),
+			signal: AbortSignal.timeout(CATALOG_TIMEOUT_MS)
+		});
+	} catch {
+		return {
+			models: [],
+			failure: "unreachable"
+		};
+	}
+	if (!response.ok) return response.status === 401 || response.status === 403 ? {
+		models: [],
+		failure: "auth",
+		status: response.status
+	} : {
+		models: [],
+		failure: "upstream",
+		status: response.status
+	};
+	let data;
+	try {
+		data = await response.json();
+	} catch {
+		return {
+			models: [],
+			failure: "malformed"
+		};
+	}
+	if (typeof data !== "object" || data === null) return {
+		models: [],
+		failure: "malformed"
+	};
 	const record = data;
-	const models = Array.isArray(record.models) ? record.models : [];
+	if (!Array.isArray(record.models)) return {
+		models: [],
+		failure: "malformed"
+	};
+	const models = record.models;
 	const result = [];
 	for (const entry of models) {
 		if (typeof entry !== "object" || entry === null) continue;
@@ -15884,7 +16055,7 @@ async function loadCatalog$1(fetchFn, credentials) {
 			fetchedAt: Date.now()
 		});
 	}
-	return result;
+	return { models: result };
 }
 //#endregion
 //#region src/host/ollama/mapper.ts
@@ -16482,6 +16653,8 @@ function textOf$4(content) {
 //#endregion
 //#region src/host/ollama/routes.ts
 const ROUTE_PREFIX$3 = "/ollama/api";
+/** Where a user creates the keys this line stores. */
+const OLLAMA_KEYS_URL = "https://ollama.com/settings/keys";
 function sendJson$5(response, status, body) {
 	const payload = JSON.stringify(body);
 	response.writeHead(status, { "content-type": "application/json; charset=utf-8" });
@@ -16507,6 +16680,22 @@ async function readRequestJson$5(request) {
 	if (typeof parsed !== "object" || parsed === null || Array.isArray(parsed)) throw new Error("Request body must be a JSON object.");
 	return parsed;
 }
+/**
+* The sentence the card shows for a catalog read that produced nothing.
+*
+* Every one of these used to collapse into "Could not read the model list from
+* Ollama", which told a user nothing they could act on — the complaint in issue
+* #36. The four causes have four different remedies, so they are named apart.
+*/
+function catalogFailureMessage(result) {
+	switch (result.failure) {
+		case "auth": return `Ollama rejected the stored API key (HTTP ${result.status ?? 401}). Create or re-paste a key at ${OLLAMA_KEYS_URL}.`;
+		case "upstream": return `Ollama answered HTTP ${result.status ?? "?"} for ${CLOUD_BASE_URL}/api/tags.`;
+		case "unreachable": return "Could not reach Ollama Cloud. Check the network or proxy, then sync again.";
+		case "malformed": return "Ollama answered with something that is not a /api/tags document (a proxy or filter may be rewriting the reply).";
+		default: return "Could not read the model list from Ollama.";
+	}
+}
 /** Whether one account can serve a request right now. */
 function isUsable(account, now) {
 	if (account.authStatus !== void 0 && account.authStatus !== "ok") return false;
@@ -16557,123 +16746,144 @@ function registerOllamaRoutes(ctx, options) {
 		handler: async (request, response) => {
 			const path = new URL(request.url || "/", "http://dsh.local").pathname.replace(/^\/ollama\/api\/?/, "");
 			const method = request.method ?? "GET";
-			if (path === "" || path === "status") return sendJson$5(response, 200, {
-				ok: true,
-				value: await getOllamaWebStatus(options)
-			});
-			if (path === "accounts") {
-				if (method === "GET") return sendJson$5(response, 200, {
+			try {
+				if (path === "" || path === "status") return sendJson$5(response, 200, {
 					ok: true,
 					value: await getOllamaWebStatus(options)
 				});
-				if (method !== "POST") return sendMethodNotAllowed$4(response);
-				if (!isSameOriginMutation(request)) return sendJson$5(response, 403, {
-					ok: false,
-					error: "Cross-origin request rejected."
-				});
-				let body;
-				try {
-					body = await readRequestJson$5(request);
-				} catch (error) {
-					return sendJson$5(response, 400, {
+				if (path === "accounts") {
+					if (method === "GET") return sendJson$5(response, 200, {
+						ok: true,
+						value: await getOllamaWebStatus(options)
+					});
+					if (method !== "POST") return sendMethodNotAllowed$4(response);
+					if (!isSameOriginMutation(request)) return sendJson$5(response, 403, {
 						ok: false,
-						error: error instanceof Error ? error.message : String(error)
+						error: "Cross-origin request rejected."
 					});
-				}
-				const action = typeof body.action === "string" ? body.action : "";
-				const accountId = typeof body.accountId === "string" ? body.accountId : void 0;
-				try {
-					if (action === "add") {
-						const apiKey = typeof body.apiKey === "string" ? body.apiKey.trim() : "";
-						if (apiKey === "") return sendJson$5(response, 400, {
+					let body;
+					try {
+						body = await readRequestJson$5(request);
+					} catch (error) {
+						return sendJson$5(response, 400, {
 							ok: false,
-							error: "API key is required."
+							error: error instanceof Error ? error.message : String(error)
 						});
-						const alias = typeof body.alias === "string" && body.alias.trim() !== "" ? body.alias.trim() : void 0;
-						const credentials = {
-							apiKey,
-							addedAt: Date.now()
-						};
-						if (alias !== void 0) credentials.alias = alias;
+					}
+					const action = typeof body.action === "string" ? body.action : "";
+					const accountId = typeof body.accountId === "string" ? body.accountId : void 0;
+					try {
+						if (action === "add") {
+							const apiKey = typeof body.apiKey === "string" ? body.apiKey.trim() : "";
+							if (apiKey === "") return sendJson$5(response, 400, {
+								ok: false,
+								error: "API key is required."
+							});
+							const alias = typeof body.alias === "string" && body.alias.trim() !== "" ? body.alias.trim() : void 0;
+							const credentials = {
+								apiKey,
+								addedAt: Date.now()
+							};
+							if (alias !== void 0) credentials.alias = alias;
+							return sendJson$5(response, 200, {
+								ok: true,
+								value: { id: (await accountPool.addAccount(credentials, alias)).id }
+							});
+						}
+						if (accountId === void 0) return sendJson$5(response, 400, {
+							ok: false,
+							error: "accountId is required."
+						});
+						if (action === "set-primary") await accountPool.setPrimary(accountId);
+						else if (action === "set-alias" && typeof body.alias === "string") await accountPool.setAlias(accountId, body.alias);
+						else if (action === "delete") await accountPool.deleteAccount(accountId);
+						else if (action === "clear-cooldown") await accountPool.clearCooldown(accountId);
+						else if (action === "strategy" && (body.strategy === "sequential" || body.strategy === "round-robin" || body.strategy === "sticky")) await accountPool.setStrategy(body.strategy);
+						else return sendJson$5(response, 400, {
+							ok: false,
+							error: "Unsupported account action."
+						});
+					} catch (error) {
+						return sendJson$5(response, 400, {
+							ok: false,
+							error: error instanceof Error ? error.message : String(error)
+						});
+					}
+					return sendJson$5(response, 200, {
+						ok: true,
+						value: await getOllamaWebStatus(options)
+					});
+				}
+				if (path === "models") {
+					if (method === "GET") {
+						const settings = await modelSettings.read();
 						return sendJson$5(response, 200, {
 							ok: true,
-							value: { id: (await accountPool.addAccount(credentials, alias)).id }
+							value: {
+								models: settings.catalogModels,
+								enabled: settings.enabled,
+								enabledModelIds: settings.enabledModelIds
+							}
 						});
 					}
-					if (accountId === void 0) return sendJson$5(response, 400, {
+					if (method !== "POST") return sendMethodNotAllowed$4(response);
+					if (!isSameOriginMutation(request)) return sendJson$5(response, 403, {
 						ok: false,
-						error: "accountId is required."
+						error: "Cross-origin request rejected."
 					});
-					if (action === "set-primary") await accountPool.setPrimary(accountId);
-					else if (action === "set-alias" && typeof body.alias === "string") await accountPool.setAlias(accountId, body.alias);
-					else if (action === "delete") await accountPool.deleteAccount(accountId);
-					else if (action === "clear-cooldown") await accountPool.clearCooldown(accountId);
-					else if (action === "strategy" && (body.strategy === "sequential" || body.strategy === "round-robin" || body.strategy === "sticky")) await accountPool.setStrategy(body.strategy);
-					else return sendJson$5(response, 400, {
+					const body = await readRequestJson$5(request).catch(() => ({}));
+					if (typeof body.enabled === "boolean") await modelSettings.update({ enabled: body.enabled });
+					if (Array.isArray(body.enabledModelIds)) await modelSettings.update({ enabledModelIds: body.enabledModelIds.filter((id) => typeof id === "string") });
+					return sendJson$5(response, 200, {
+						ok: true,
+						value: await modelSettings.read()
+					});
+				}
+				if (path === "catalog/refresh") {
+					if (method !== "POST") return sendMethodNotAllowed$4(response);
+					if (!isSameOriginMutation(request)) return sendJson$5(response, 403, {
 						ok: false,
-						error: "Unsupported account action."
+						error: "Cross-origin request rejected."
 					});
-				} catch (error) {
-					return sendJson$5(response, 400, {
+					let credentials = null;
+					let poolBusy = false;
+					try {
+						credentials = (await accountPool.getEffectiveCredential(void 0, fetchFn)).credentials;
+					} catch {
+						poolBusy = await accountPool.read().then((data) => data.accounts.length > 0 && data.accounts.every((account) => isUsable(account, Date.now()) === false)).catch(() => false);
+						credentials = null;
+					}
+					if (credentials === null) return sendJson$5(response, 400, {
 						ok: false,
-						error: error instanceof Error ? error.message : String(error)
+						error: poolBusy ? "Every stored Ollama key is cooling down right now. Wait for the cooldown, then sync again." : "Add an API key before syncing models."
 					});
-				}
-				return sendJson$5(response, 200, {
-					ok: true,
-					value: await getOllamaWebStatus(options)
-				});
-			}
-			if (path === "models") {
-				if (method === "GET") {
-					const settings = await modelSettings.read();
+					const result = await fetchCatalog(fetchFn, credentials);
+					if (result.models.length === 0) return sendJson$5(response, 502, {
+						ok: false,
+						error: catalogFailureMessage(result)
+					});
+					try {
+						await modelSettings.storeCatalog(result.models);
+					} catch (error) {
+						const detail = error instanceof Error ? error.message : String(error);
+						return sendJson$5(response, 500, {
+							ok: false,
+							error: `Read ${result.models.length} models from Ollama but could not save them locally: ${detail}`
+						});
+					}
 					return sendJson$5(response, 200, {
 						ok: true,
-						value: {
-							models: settings.catalogModels,
-							enabled: settings.enabled,
-							enabledModelIds: settings.enabledModelIds
-						}
+						value: { models: result.models }
 					});
 				}
-				if (method !== "POST") return sendMethodNotAllowed$4(response);
-				if (!isSameOriginMutation(request)) return sendJson$5(response, 403, {
-					ok: false,
-					error: "Cross-origin request rejected."
-				});
-				const body = await readRequestJson$5(request).catch(() => ({}));
-				if (typeof body.enabled === "boolean") await modelSettings.update({ enabled: body.enabled });
-				if (Array.isArray(body.enabledModelIds)) await modelSettings.update({ enabledModelIds: body.enabledModelIds.filter((id) => typeof id === "string") });
-				return sendJson$5(response, 200, {
-					ok: true,
-					value: await modelSettings.read()
-				});
-			}
-			if (path === "catalog/refresh") {
-				if (method !== "POST") return sendMethodNotAllowed$4(response);
-				if (!isSameOriginMutation(request)) return sendJson$5(response, 403, {
+				return sendJson$5(response, 404, {
 					ok: false,
-					error: "Cross-origin request rejected."
-				});
-				let credentials = null;
-				try {
-					credentials = (await accountPool.getEffectiveCredential(void 0, fetchFn)).credentials;
-				} catch {
-					credentials = null;
-				}
-				if (credentials === null) return sendJson$5(response, 400, {
-					ok: false,
-					error: "Add an API key before syncing models."
+					error: "not-found"
 				});
-				const models = await loadCatalog$1(fetchFn, credentials).catch(() => []);
-				if (models.length === 0) return sendJson$5(response, 502, {
+			} catch (error) {
+				return sendJson$5(response, 500, {
 					ok: false,
-					error: "Could not read the model list from Ollama."
-				});
-				await modelSettings.storeCatalog(models);
-				return sendJson$5(response, 200, {
-					ok: true,
-					value: { models }
+					error: error instanceof Error ? error.message : String(error)
 				});
 			}
 		}
@@ -19810,7 +20020,20 @@ function estimatedInputTokens(options) {
 			characters += JSON.stringify(tool.parameters).length;
 		} catch {}
 	}
-	for (const message of options.messages) characters += textOf$3(message.content).length;
+	const contentCharacters = (content) => {
+		if (!Array.isArray(content)) return typeof content === "string" ? content.length : 0;
+		return content.reduce((total, block) => {
+			if (!isRecord$16(block)) return total;
+			if ((block.type === "text" || block.type === "reasoning") && typeof block.text === "string") return total + block.text.length;
+			if (block.type === "tool-result") return total + contentCharacters(block.content);
+			if (block.type === "tool-call") {
+				const args = toolCallArguments$3(block.arguments);
+				return total + String(block.name ?? "").length + args.length;
+			}
+			return total;
+		}, 0);
+	};
+	for (const message of options.messages) characters += contentCharacters(message.content);
 	if (characters === 0) return void 0;
 	return Math.ceil(characters / 4);
 }
@@ -20244,7 +20467,7 @@ function processOpenAIStreamLine(line, state) {
 	if (!isRecord$16(chunk)) return [];
 	const out = [];
 	const errorPayload = isRecord$16(chunk.error) ? chunk.error : void 0;
-	if (errorPayload !== void 0) throw new LlmError$1(`Kimi Code stream error: ${asString$9(errorPayload.message) ?? "unknown error"}`, "PROVIDER_ERROR");
+	if (errorPayload !== void 0) throw new LlmError$1(`Kimi Code stream error: ${asString$9(errorPayload.message) ?? "unknown error"}`, isContextOverflow(errorPayload) ? CONTEXT_OVERFLOW_CODE : "PROVIDER_ERROR");
 	const usage = isRecord$16(chunk.usage) ? chunk.usage : void 0;
 	if (usage) {
 		state.sawUsage = true;
@@ -20511,7 +20734,10 @@ function processAnthropicStreamLine(line, state) {
 		state.done = true;
 		return closeStream$2(state);
 	}
-	if (type === "error") throw new LlmError$1(`Kimi Code stream error: ${asString$9((isRecord$16(event.error) ? event.error : {}).message) ?? "unknown error"}`, "PROVIDER_ERROR");
+	if (type === "error") {
+		const error = isRecord$16(event.error) ? event.error : {};
+		throw new LlmError$1(`Kimi Code stream error: ${asString$9(error.message) ?? "unknown error"}`, isContextOverflow(error) ? CONTEXT_OVERFLOW_CODE : "PROVIDER_ERROR");
+	}
 	return out;
 }
 /**
@@ -20744,6 +20970,11 @@ function summarizeFailureBody$1(raw) {
 */
 function classifyKimiFailure(status, bodyText) {
 	const detail = summarizeFailureBody$1(bodyText);
+	if (isHttpContextOverflow(status, bodyText)) return {
+		code: CONTEXT_OVERFLOW_CODE,
+		retryable: false,
+		message: `${PROVIDER_NAME$3} context window exceeded: ${detail}`
+	};
 	if (status === 402) return {
 		code: "SERVER",
 		retryable: true,
@@ -20961,7 +21192,7 @@ var KimiCodeAdapter = class extends LlmAdapter {
 			const endpoint = wire === "anthropic" ? `${base}/v1/messages?beta=true` : `${base}/v1/chat/completions`;
 			const headers = await modelRequestHeaders$1(apiCredentials.accessToken, wire);
 			try {
-				response = await fetchFn(endpoint, {
+				response = await fetchWithContextBudgetRecovery(fetchFn)(endpoint, {
 					method: "POST",
 					headers,
 					body,
@@ -23379,7 +23610,10 @@ function processMinimaxStreamLine(line, state) {
 		state.done = true;
 		return closeMinimaxStream(state);
 	}
-	if (type === "error") throw new LlmError$1("MiniMax Code（编程订阅） stream error: " + (asString$8((isRecord$15(event.error) ? event.error : {}).message) ?? "unknown error"), "PROVIDER_ERROR");
+	if (type === "error") {
+		const error = isRecord$15(event.error) ? event.error : {};
+		throw new LlmError$1("MiniMax Code（编程订阅） stream error: " + (asString$8(error.message) ?? "unknown error"), isContextOverflow(error) ? CONTEXT_OVERFLOW_CODE : "PROVIDER_ERROR");
+	}
 	return out;
 }
 //#endregion
@@ -25407,6 +25641,11 @@ function matchesAny(text, patterns) {
 */
 function classifyMinimaxFailure(status, bodyText) {
 	const detail = summarizeFailureBody(bodyText);
+	if (isHttpContextOverflow(status, bodyText)) return {
+		code: CONTEXT_OVERFLOW_CODE,
+		retryable: false,
+		message: `${PROVIDER_NAME$2} context window exceeded: ${detail}`
+	};
 	if (status === 401 || status === 403) return {
 		code: "INVALID_CREDENTIAL",
 		retryable: false,
@@ -25670,7 +25909,7 @@ var MinimaxCodeAdapter = class extends LlmAdapter {
 		let response;
 		for (let attempt = 0;; attempt += 1) {
 			try {
-				response = await fetchFn(messagesUrl(credentials.region), {
+				response = await fetchWithContextBudgetRecovery(fetchFn)(messagesUrl(credentials.region), {
 					method: "POST",
 					headers: modelRequestHeaders(credentials.accessToken),
 					body,
@@ -30052,7 +30291,7 @@ function processStreamLine$1(line, state) {
 	if (!isRecord$8(chunk)) return [];
 	const out = [];
 	const errorPayload = isRecord$8(chunk.error) ? chunk.error : void 0;
-	if (errorPayload !== void 0) throw new LlmError$1(`WorkBuddy stream error: ${asString$4(errorPayload.message) ?? "unknown error"}`, "PROVIDER_ERROR");
+	if (errorPayload !== void 0) throw new LlmError$1(`WorkBuddy stream error: ${asString$4(errorPayload.message) ?? "unknown error"}`, isContextOverflow(errorPayload) ? CONTEXT_OVERFLOW_CODE : "PROVIDER_ERROR");
 	const usage = isRecord$8(chunk.usage) ? chunk.usage : void 0;
 	if (usage) {
 		state.sawUsage = true;
@@ -30410,7 +30649,7 @@ var WorkBuddyAdapter = class extends LlmAdapter {
 				credentials = effective.credentials;
 			}
 			try {
-				response = await fetchFn(`${credentials.backend}${CHAT_PATH}`, {
+				response = await fetchWithContextBudgetRecovery(fetchFn)(`${credentials.backend}${CHAT_PATH}`, {
 					method: "POST",
 					headers: workBuddyHeaders(credentials, { accept: "text/event-stream" }),
 					body,
@@ -30485,7 +30724,7 @@ function classifyFailure$1(status, detail, headers) {
 	if (code === 11135 || code === 11133) return new LlmError(`${PROVIDER_NAME$1} rejected the request parameters (${code}). An image may be unsupported by this model or unreadable.${detail ? ` ${detail}` : ""}`, "PROVIDER_ERROR", { status });
 	if (code === 11128) return new LlmError(`${PROVIDER_NAME$1} rejected the message history: the conversation must open with a system prompt.${detail ? ` ${detail}` : ""}`, "PROVIDER_ERROR", { status });
 	if (status >= 500 || code === 11134) return new LlmError(`${PROVIDER_NAME$1} upstream server error (${status}): ${detail || "No response"}`, "SERVER", { status });
-	return new LlmError(`${PROVIDER_NAME$1} API error (${status}): ${detail || "No response"}`, "PROVIDER_ERROR", { status });
+	return new LlmError(`${PROVIDER_NAME$1} API error (${status}): ${detail || "No response"}`, isHttpContextOverflow(status, detail) ? CONTEXT_OVERFLOW_CODE : "PROVIDER_ERROR", { status });
 }
 /** Read the subscription's own numeric error code out of a response body. */
 function readErrorCode(detail) {
@@ -35029,7 +35268,7 @@ function processStreamLine(line, state) {
 		const error = isRecord$5(event.error) ? event.error : {};
 		const message = asString$2(error.message) ?? "unknown error";
 		const kind = asString$2(error.type);
-		throw new LlmError$1("Claude stream error" + (kind === void 0 ? "" : " (" + kind + ")") + ": " + message, "PROVIDER_ERROR");
+		throw new LlmError$1("Claude stream error" + (kind === void 0 ? "" : " (" + kind + ")") + ": " + message, isContextOverflow(error) ? CONTEXT_OVERFLOW_CODE : "PROVIDER_ERROR");
 	}
 	return out;
 }
@@ -36558,7 +36797,7 @@ var ClaudeAdapter = class ClaudeAdapter extends LlmAdapter {
 			cacheTtl
 		});
 		try {
-			return await fetchFn("https://api.anthropic.com/v1/messages?beta=true", {
+			return await fetchWithContextBudgetRecovery(fetchFn)("https://api.anthropic.com/v1/messages?beta=true", {
 				method: "POST",
 				headers: buildClaudeHeaders(credentials.accessToken, {
 					model: requestOptions.model,
@@ -36754,6 +36993,7 @@ function toLlmError(failure) {
 	};
 	const detail = failure.message === "" ? "" : " " + failure.message;
 	if (failure.clientVersionTooOld) return new LlmError("Claude（订阅） refused this request because the client version it was told is below the server's floor (claude_code_version_too_old). Raise it with the DSH_CLAUDE_CLI_VERSION environment variable (or a setClaudeCliVersion pin), then retry — the stored sign-in is still valid." + detail, "PROVIDER_ERROR", options);
+	if (failure.kind === "request" && isHttpContextOverflow(failure.status, failure.message)) return new LlmError("Claude（订阅） context window exceeded." + detail, CONTEXT_OVERFLOW_CODE, options);
 	switch (failure.kind) {
 		case "credential": return new LlmError("Claude（订阅） rejected the stored credential. Sign in again from Settings > Claude." + detail, "INVALID_CREDENTIAL", options);
 		case "rate_limit_account":