@eddyskywalker/dsh-chatgpt-subscription 0.12.3 → 0.12.6

package/lib/index.js

@@ -791,7 +791,7 @@ function parseListing(payload) {
 function parseListingEntry$1(value) {
 	const record = asRecord$10(value);
 	if (record === void 0) return [];
-	const id = asString$17(record.slug) ?? asString$17(record.id);
+	const id = asString$18(record.slug) ?? asString$18(record.id);
 	if (id === void 0) return [];
 	const contextWindow = asNumber$4(record.context_window) ?? asNumber$4(record.contextWindow);
 	const modalities = parseModalities(record.input_modalities);
@@ -799,18 +799,18 @@ function parseListingEntry$1(value) {
 	const fallback = CODEX_MODEL_CATALOG.find((entry) => entry.id === id);
 	return [{
 		id,
-		name: asString$17(record.display_name) ?? asString$17(record.displayName) ?? fallback?.name ?? id,
+		name: asString$18(record.display_name) ?? asString$18(record.displayName) ?? fallback?.name ?? id,
 		contextWindow: contextWindow !== void 0 && contextWindow > 0 ? contextWindow : null,
 		inputModalities: modalities ?? (fallback ? [...fallback.inputModalities] : ["text"]),
 		...efforts !== null ? { reasoningEfforts: efforts } : {},
-		defaultReasoningEffort: asString$17(record.default_reasoning_level) ?? fallback?.defaultReasoningEffort ?? null
+		defaultReasoningEffort: asString$18(record.default_reasoning_level) ?? fallback?.defaultReasoningEffort ?? null
 	}];
 }
 function parseModalities(value) {
 	if (!Array.isArray(value)) return null;
 	const out = [];
 	for (const item of value) {
-		const name = asString$17(item);
+		const name = asString$18(item);
 		if (name === "text" || name === "image") out.push(name);
 	}
 	return out.length > 0 ? out : null;
@@ -825,7 +825,7 @@ function parseEfforts(value) {
 	if (!Array.isArray(value)) return null;
 	const out = [];
 	for (const item of value) {
-		const effort = asString$17(item) ?? asString$17(asRecord$10(item)?.effort);
+		const effort = asString$18(item) ?? asString$18(asRecord$10(item)?.effort);
 		if (effort !== void 0 && !out.includes(effort)) out.push(effort);
 	}
 	return out.length > 0 ? out : null;
@@ -843,7 +843,7 @@ function accountKeyFor$1(credentials) {
 function asRecord$10(value) {
 	return typeof value === "object" && value !== null && !Array.isArray(value) ? value : void 0;
 }
-function asString$17(value) {
+function asString$18(value) {
 	return typeof value === "string" && value !== "" ? value : void 0;
 }
 function asNumber$4(value) {
@@ -1085,6 +1085,22 @@ var WindowsDpapiTokenStore = class extends WindowsDpapiCredentialStore {
 		super(path, parseStoredCredentials);
 	}
 };
+/**
+* How long a helper may run before it is killed and the read fails.
+*
+* A healthy spawn costs ~200 ms (see the cache note on this class), so the guard
+* exists only for a child that never exits, and it was set at 10 s - fifty times
+* the healthy cost. That turned out to be too tight for one real environment: a
+* cold Windows runner whose freshly-installed dependency tree is being scanned
+* on first spawn, where the helper legitimately took longer than 10 s and the
+* kill turned a working credential read into `DPAPI helper timed out`.
+*
+* The guard is kept, and only its headroom widened: 30 s is still bounded, still
+* far above anything a warm machine needs, and a hung helper now blocks one
+* credential read for 30 s rather than 10 - a worse wait in a case that is already
+* a failure, in exchange for not failing a machine that is merely cold.
+*/
+const POWERSHELL_HELPER_TIMEOUT_MS = 3e4;
 function runPowerShell(script, path, stdin) {
 	return new Promise((resolve, reject) => {
 		const child = spawn("powershell.exe", [
@@ -1110,7 +1126,7 @@ function runPowerShell(script, path, stdin) {
 		const timer = setTimeout(() => {
 			child.kill();
 			reject(/* @__PURE__ */ new Error("DPAPI helper timed out"));
-		}, 1e4);
+		}, POWERSHELL_HELPER_TIMEOUT_MS);
 		child.stdout.setEncoding("utf8");
 		child.stdout.on("data", (chunk) => {
 			stdout += chunk;
@@ -3309,13 +3325,13 @@ function createCodexFetchProvider(options = {}) {
 //#region src/host/common/llm-compat.ts
 /** This package's kind in the harness message-source map. */
 const PLUGIN_MESSAGE_SOURCE_KIND = "dsh-chatgpt-subscription";
-function isRecord$29(value) {
+function isRecord$30(value) {
 	return typeof value === "object" && value !== null && !Array.isArray(value);
 }
 /** Read a message body's blocks, keeping every entry that carries a block tag. */
 function readBlocks(raw) {
 	if (!Array.isArray(raw)) return [];
-	return raw.filter((block) => isRecord$29(block) && typeof block.type === "string");
+	return raw.filter((block) => isRecord$30(block) && typeof block.type === "string");
 }
 /**
 * Normalize one harness message list into {@link Message}.
@@ -3330,10 +3346,10 @@ function readBlocks(raw) {
 function normalizeMessages(messages) {
 	const normalized = [];
 	for (const raw of messages ?? []) {
-		if (!isRecord$29(raw)) continue;
+		if (!isRecord$30(raw)) continue;
 		const content = readBlocks(raw.content);
 		if (raw.role === "tool") {
-			const source = isRecord$29(raw.source) ? raw.source : void 0;
+			const source = isRecord$30(raw.source) ? raw.source : void 0;
 			const callId = String(raw.toolCallId ?? source?.callId ?? "");
 			normalized.push({
 				id: raw.id,
@@ -3355,7 +3371,7 @@ function normalizeMessages(messages) {
 			id: raw.id,
 			role: typeof raw.role === "string" ? raw.role : "user",
 			content,
-			...isRecord$29(raw.source) ? { source: raw.source } : {}
+			...isRecord$30(raw.source) ? { source: raw.source } : {}
 		});
 	}
 	return normalized;
@@ -7055,7 +7071,7 @@ function field(value, name) {
 	const candidate = value[name];
 	return typeof candidate === "string" && candidate !== "" ? candidate : null;
 }
-function isRecord$28(value) {
+function isRecord$29(value) {
 	return typeof value === "object" && value !== null && !Array.isArray(value);
 }
 function readPreferencesUpdate(value, current) {
@@ -7089,7 +7105,7 @@ function readPreferencesUpdate(value, current) {
 		patch.searchProvider = value.searchProvider;
 	}
 	if ("contextWindowOverrides" in value) {
-		if (!isRecord$28(value.contextWindowOverrides)) throw new PreferenceError("contextWindowOverrides must be an object.");
+		if (!isRecord$29(value.contextWindowOverrides)) throw new PreferenceError("contextWindowOverrides must be an object.");
 		const overrides = {};
 		for (const [model, contextWindow] of Object.entries(value.contextWindowOverrides)) {
 			if (!isCodexModelId(model)) throw new PreferenceError("Unknown Codex model for a context window override.");
@@ -8589,7 +8605,7 @@ const schemaFields = /* @__PURE__ */ new Set([
 	"required",
 	"propertyOrdering"
 ]);
-function isRecord$27(value) {
+function isRecord$28(value) {
 	return typeof value === "object" && value !== null && !Array.isArray(value);
 }
 function mergeSchemas(left, right) {
@@ -8597,7 +8613,7 @@ function mergeSchemas(left, right) {
 		...left,
 		...right
 	};
-	if (isRecord$27(left.properties) && isRecord$27(right.properties)) merged.properties = {
+	if (isRecord$28(left.properties) && isRecord$28(right.properties)) merged.properties = {
 		...left.properties,
 		...right.properties
 	};
@@ -8633,9 +8649,9 @@ const mergedTypeKeys = /* @__PURE__ */ new Set([
 * model with strictly less than the schema it started from.
 */
 function describeBranch(branch) {
-	const type = typeof branch.type === "string" ? branch.type : isRecord$27(branch.properties) ? "object" : isRecord$27(branch.items) ? "array" : "value";
+	const type = typeof branch.type === "string" ? branch.type : isRecord$28(branch.properties) ? "object" : isRecord$28(branch.items) ? "array" : "value";
 	const details = [];
-	if (isRecord$27(branch.properties)) {
+	if (isRecord$28(branch.properties)) {
 		const names = Object.keys(branch.properties);
 		if (names.length > 0) details.push(`{${names.join(",")}}`);
 	}
@@ -8649,7 +8665,7 @@ function describeBranch(branch) {
 }
 /** How well one branch can stand in for a whole union; a named object shape wins. */
 function branchRank(branch) {
-	if (isRecord$27(branch.properties)) return 3;
+	if (isRecord$28(branch.properties)) return 3;
 	if (branch.type === "object") return 2;
 	if (branch.type === "array") return 1;
 	return 0;
@@ -8678,7 +8694,7 @@ function unionShape(branches) {
 		shape: branches[0],
 		folded: false
 	};
-	if (new Set(branches.map((branch) => branch.type)).size === 1 && branches.every((branch) => !isRecord$27(branch.properties) && !isRecord$27(branch.items))) {
+	if (new Set(branches.map((branch) => branch.type)).size === 1 && branches.every((branch) => !isRecord$28(branch.properties) && !isRecord$28(branch.items))) {
 		const shape = { type: branches[0].type };
 		if (branches.every((branch) => Array.isArray(branch.enum))) {
 			const values = [...new Set(branches.flatMap((branch) => branch.enum))];
@@ -8715,7 +8731,7 @@ function foldOwnFields(node) {
 	const out = {};
 	for (const [key, value] of Object.entries(node)) {
 		if (key === "anyOf" || key === "oneOf" || key === "allOf") continue;
-		if (key === "properties" && isRecord$27(value)) {
+		if (key === "properties" && isRecord$28(value)) {
 			out.properties = Object.fromEntries(Object.entries(value).map(([name, child]) => [name, foldSchemaUnions(child)]));
 			continue;
 		}
@@ -8744,15 +8760,15 @@ function foldOwnFields(node) {
 * argument the model sends anyway fails there, naming the actual problem.
 */
 function foldSchemaUnions(node) {
-	if (!isRecord$27(node)) return node;
+	if (!isRecord$28(node)) return node;
 	let own = foldOwnFields(node);
 	if (Array.isArray(node.allOf)) for (const branch of node.allOf) {
 		const folded = foldSchemaUnions(branch);
-		if (isRecord$27(folded)) own = mergeSchemas(own, folded);
+		if (isRecord$28(folded)) own = mergeSchemas(own, folded);
 	}
 	const alternatives = Array.isArray(node.anyOf) ? node.anyOf : Array.isArray(node.oneOf) ? node.oneOf : [];
 	if (alternatives.length === 0) return own;
-	const branches = alternatives.map((branch) => foldSchemaUnions(branch)).filter(isRecord$27);
+	const branches = alternatives.map((branch) => foldSchemaUnions(branch)).filter(isRecord$28);
 	if (branches.length === 0) return own;
 	const shapes = branches.filter((branch) => branch.type !== "null");
 	const acceptsNull = branches.some((branch) => branch.type === "null");
@@ -8768,7 +8784,7 @@ function foldSchemaUnions(node) {
 	return merged;
 }
 function toAntigravityToolSchema(schema, options = {}) {
-	if (!isRecord$27(schema)) return schema;
+	if (!isRecord$28(schema)) return schema;
 	const root = schema;
 	function resolveReference(reference) {
 		let target = root;
@@ -8776,13 +8792,13 @@ function toAntigravityToolSchema(schema, options = {}) {
 		const path = reference === "#" ? [] : reference.slice(2).split("/");
 		for (const segment of path) {
 			const key = segment.replace(/~1/g, "/").replace(/~0/g, "~");
-			target = isRecord$27(target) && Object.hasOwn(target, key) ? target[key] : void 0;
+			target = isRecord$28(target) && Object.hasOwn(target, key) ? target[key] : void 0;
 		}
-		if (!isRecord$27(target)) throw new LlmError(`Antigravity tool schema reference was not found: ${reference}`, "PROVIDER_ERROR");
+		if (!isRecord$28(target)) throw new LlmError(`Antigravity tool schema reference was not found: ${reference}`, "PROVIDER_ERROR");
 		return target;
 	}
 	function convert(node, references) {
-		if (!isRecord$27(node)) return {};
+		if (!isRecord$28(node)) return {};
 		let inherited = {};
 		if (typeof node.$ref === "string") {
 			if (references.has(node.$ref)) throw new LlmError(`Antigravity tool schema contains a recursive reference: ${node.$ref}`, "PROVIDER_ERROR");
@@ -8791,8 +8807,8 @@ function toAntigravityToolSchema(schema, options = {}) {
 		if (Array.isArray(node.allOf)) for (const branch of node.allOf) inherited = mergeSchemas(inherited, convert(branch, references));
 		const out = {};
 		for (const [key, value] of Object.entries(node)) if (schemaFields.has(key)) out[key] = value;
-		if (isRecord$27(node.properties)) out.properties = Object.fromEntries(Object.entries(node.properties).map(([name, child]) => [name, convert(child, references)]));
-		if (isRecord$27(node.items)) out.items = convert(node.items, references);
+		if (isRecord$28(node.properties)) out.properties = Object.fromEntries(Object.entries(node.properties).map(([name, child]) => [name, convert(child, references)]));
+		if (isRecord$28(node.items)) out.items = convert(node.items, references);
 		const alternatives = Array.isArray(node.anyOf) ? node.anyOf : node.oneOf;
 		if (Array.isArray(alternatives)) out.anyOf = alternatives.map((child) => convert(child, references));
 		if (Array.isArray(node.type)) {
@@ -8828,10 +8844,10 @@ let toolCallCounter = 0;
 function sanitizeText$5(text) {
 	return text.replace(/\0/g, "");
 }
-function isRecord$26(value) {
+function isRecord$27(value) {
 	return typeof value === "object" && value !== null && !Array.isArray(value);
 }
-function asString$16(value) {
+function asString$17(value) {
 	return typeof value === "string" ? value : void 0;
 }
 function safeJsonParse$5(text) {
@@ -8848,25 +8864,25 @@ function toolCallIdNeeded(modelId, runtimeModel) {
 	return modelId.startsWith("claude-") || modelId.startsWith("gpt-oss-") || runtimeModel.startsWith("claude-") || runtimeModel.startsWith("gpt-oss-");
 }
 function parseArguments$1(raw) {
-	if (isRecord$26(raw)) return raw;
+	if (isRecord$27(raw)) return raw;
 	if (raw === void 0 || raw === null || raw === "") return {};
 	const parsed = typeof raw === "string" ? safeJsonParse$5(raw) : raw;
-	return isRecord$26(parsed) ? parsed : {};
+	return isRecord$27(parsed) ? parsed : {};
 }
 const NO_RESOLVED_IMAGES$5 = /* @__PURE__ */ new Map();
 function attachmentOf$5(block) {
 	const attachment = block.attachment;
-	if (!isRecord$26(attachment)) return void 0;
+	if (!isRecord$27(attachment)) return void 0;
 	return typeof attachment.attachmentId === "string" ? attachment : void 0;
 }
 function attachmentLabel$6(block) {
-	const attachment = isRecord$26(block.attachment) ? block.attachment : void 0;
-	return asString$16(attachment?.name) || asString$16(attachment?.attachmentId);
+	const attachment = isRecord$27(block.attachment) ? block.attachment : void 0;
+	return asString$17(attachment?.name) || asString$17(attachment?.attachmentId);
 }
 function collectImageRefs$4(content, refs) {
 	if (!Array.isArray(content)) return;
 	for (const block of content) {
-		if (!isRecord$26(block)) continue;
+		if (!isRecord$27(block)) continue;
 		if (block.type === "image") {
 			const attachment = attachmentOf$5(block);
 			if (attachment) refs.set(attachment.attachmentId, attachment);
@@ -8904,13 +8920,13 @@ function base64Length$4(bytes) {
 function requestImageBytes$4(block) {
 	const attachment = attachmentOf$5(block);
 	if (attachment) return base64Length$4(attachment.bytes);
-	const inline = asString$16(block.data) || asString$16(block.base64);
+	const inline = asString$17(block.data) || asString$17(block.base64);
 	return inline ? inline.length : void 0;
 }
 function collectRequestImageBytes$4(content, lengths) {
 	if (!Array.isArray(content)) return;
 	for (const block of content) {
-		if (!isRecord$26(block) || block.type !== "image") continue;
+		if (!isRecord$27(block) || block.type !== "image") continue;
 		const bytes = requestImageBytes$4(block);
 		if (bytes !== void 0) lengths.push(bytes);
 	}
@@ -8945,7 +8961,7 @@ function offloadOldestRequestImages$4(options) {
 		if (remaining.count === 0 || !Array.isArray(message.content)) return message;
 		let replaced = false;
 		const content = message.content.map((block) => {
-			if (remaining.count === 0 || !isRecord$26(block) || block.type !== "image") return block;
+			if (remaining.count === 0 || !isRecord$27(block) || block.type !== "image") return block;
 			if (requestImageBytes$4(block) === void 0) return block;
 			remaining.count -= 1;
 			replaced = true;
@@ -9008,10 +9024,10 @@ function unavailableImageText$6(block) {
 	return `[image unavailable: ${label ? `${label} could not be read` : "the image could not be read"}; ask the user to attach it again if the image is needed]`;
 }
 function imageBlockToPart(block, images) {
-	let data = asString$16(block.data) || asString$16(block.base64);
-	const source = isRecord$26(block.source) ? block.source : void 0;
-	if (!data && source) data = asString$16(source.data) || asString$16(source.base64);
-	let mimeType = asString$16(block.mimeType) || asString$16(block.mediaType) || (source ? asString$16(source.mimeType) || asString$16(source.mediaType) : void 0) || "image/png";
+	let data = asString$17(block.data) || asString$17(block.base64);
+	const source = isRecord$27(block.source) ? block.source : void 0;
+	if (!data && source) data = asString$17(source.data) || asString$17(source.base64);
+	let mimeType = asString$17(block.mimeType) || asString$17(block.mediaType) || (source ? asString$17(source.mimeType) || asString$17(source.mediaType) : void 0) || "image/png";
 	if (data?.startsWith("data:")) {
 		const match = data.match(/^data:([^;,]+);base64,(.*)$/s);
 		if (match) {
@@ -9034,8 +9050,8 @@ function contentToUserParts(content, images) {
 	if (typeof content === "string") return [{ text: sanitizeText$5(content) }];
 	if (!Array.isArray(content)) return [];
 	const parts = [];
-	for (const block of content) if (isRecord$26(block) && block.type === "text" && typeof block.text === "string") parts.push({ text: sanitizeText$5(block.text) });
-	else if (isRecord$26(block) && block.type === "image") {
+	for (const block of content) if (isRecord$27(block) && block.type === "text" && typeof block.text === "string") parts.push({ text: sanitizeText$5(block.text) });
+	else if (isRecord$27(block) && block.type === "image") {
 		const img = imageBlockToPart(block, images);
 		parts.push(img ?? { text: unavailableImageText$6(block) });
 	}
@@ -9050,7 +9066,7 @@ function toolResultImageParts(blocks, images) {
 	if (!Array.isArray(blocks)) return [];
 	const parts = [];
 	for (const block of blocks) {
-		if (!isRecord$26(block)) continue;
+		if (!isRecord$27(block)) continue;
 		if (block.type === "image") {
 			parts.push(imageBlockToPart(block, images) ?? { text: unavailableImageText$6(block) });
 			continue;
@@ -9062,7 +9078,7 @@ function toolResultImageParts(blocks, images) {
 function toolResultText$5(blocks) {
 	if (!Array.isArray(blocks)) return "";
 	return blocks.map((block) => {
-		if (!isRecord$26(block)) return "";
+		if (!isRecord$27(block)) return "";
 		if (block.type === "text" && typeof block.text === "string") return sanitizeText$5(block.text);
 		if (block.type === "tool-result") return toolResultText$5(block.content);
 		if (block.type === "image") {
@@ -9076,16 +9092,16 @@ function replayBlockFor$1(message, index) {
 	const source = message.source;
 	if (!source || source.kind !== "model" || source.provider !== "antigravity") return void 0;
 	const state = source.replayState;
-	if (!isRecord$26(state)) return void 0;
+	if (!isRecord$27(state)) return void 0;
 	if (Array.isArray(state.blocks)) return state.blocks[index];
-	const resp = isRecord$26(state.response) ? state.response : void 0;
+	const resp = isRecord$27(state.response) ? state.response : void 0;
 	if (resp) {
 		if (Array.isArray(resp.outputItems)) return resp.outputItems[index];
 		if (Array.isArray(resp.blocks)) return resp.blocks[index];
 	}
 }
 function thoughtSignature(part) {
-	return asString$16(part?.thoughtSignature) || asString$16(part?.thought_signature) || asString$16(part?.thinkingSignature) || asString$16(part?.textSignature);
+	return asString$17(part?.thoughtSignature) || asString$17(part?.thought_signature) || asString$17(part?.thinkingSignature) || asString$17(part?.textSignature);
 }
 function replayPart(part) {
 	const copy = { ...part };
@@ -9138,11 +9154,11 @@ function assistantParts(message, model, runtimeModel, toolCalls) {
 	if (!Array.isArray(message.content)) return parts;
 	for (let index = 0; index < message.content.length; index++) {
 		const block = message.content[index];
-		if (!isRecord$26(block)) continue;
+		if (!isRecord$27(block)) continue;
 		if (block.type === "reasoning" && runtimeModel.startsWith("claude-") && (message.source?.kind !== "model" || message.source.provider !== "antigravity" || !("model" in message.source) || message.source.model !== model.id)) continue;
 		const replay = replayBlockFor$1(message, index);
-		const originalParts = Array.isArray(replay?.parts) ? replay.parts.filter(isRecord$26) : [];
-		if ((block.type === "text" || block.type === "reasoning") && originalParts.length > 0 && originalParts.every((part) => !part.functionCall) && originalParts.map((part) => asString$16(part.text) || "").join("") === sanitizeText$5(String(block.text || ""))) {
+		const originalParts = Array.isArray(replay?.parts) ? replay.parts.filter(isRecord$27) : [];
+		if ((block.type === "text" || block.type === "reasoning") && originalParts.length > 0 && originalParts.every((part) => !part.functionCall) && originalParts.map((part) => asString$17(part.text) || "").join("") === sanitizeText$5(String(block.text || ""))) {
 			parts.push(...originalParts.map(replayPart));
 			continue;
 		}
@@ -9162,8 +9178,8 @@ function assistantParts(message, model, runtimeModel, toolCalls) {
 		} else if (block.type === "tool-call") {
 			const toolId = String(block.id || "");
 			const toolName = String(block.name || "");
-			const originalCall = originalParts.find((part) => isRecord$26(part.functionCall));
-			const wireId = asString$16((isRecord$26(originalCall?.functionCall) ? originalCall.functionCall : void 0)?.id) || (toolCallIdNeeded(model.id, runtimeModel) ? sanitizeToolCallId(toolId, toolName) : originalCall ? void 0 : toolId || void 0);
+			const originalCall = originalParts.find((part) => isRecord$27(part.functionCall));
+			const wireId = asString$17((isRecord$27(originalCall?.functionCall) ? originalCall.functionCall : void 0)?.id) || (toolCallIdNeeded(model.id, runtimeModel) ? sanitizeToolCallId(toolId, toolName) : originalCall ? void 0 : toolId || void 0);
 			toolCalls.set(toolId, {
 				name: toolName,
 				id: wireId
@@ -9215,7 +9231,7 @@ function convertMessages(options, model, runtimeModel, images = NO_RESOLVED_IMAG
 			continue;
 		}
 		const content = Array.isArray(message.content) ? message.content : [];
-		const userParts = contentToUserParts(content.filter((b) => !isRecord$26(b) || b.type !== "tool-result"), images);
+		const userParts = contentToUserParts(content.filter((b) => !isRecord$27(b) || b.type !== "tool-result"), images);
 		if (role === "system") {
 			if (userParts.length) contents.push({
 				role: GEMINI_ROLE.user,
@@ -9227,7 +9243,7 @@ function convertMessages(options, model, runtimeModel, images = NO_RESOLVED_IMAG
 			role: GEMINI_ROLE.user,
 			parts: userParts
 		});
-		for (const b of content) if (isRecord$26(b) && b.type === "tool-result") pushToolResult(contents, b, toolCalls, model, runtimeModel, images);
+		for (const b of content) if (isRecord$27(b) && b.type === "tool-result") pushToolResult(contents, b, toolCalls, model, runtimeModel, images);
 	}
 	return contents;
 }
@@ -9362,7 +9378,7 @@ const USAGE_FIELDS = [
 	"totalTokenCount"
 ];
 function collectUsage(value, state) {
-	if (!isRecord$26(value)) return;
+	if (!isRecord$27(value)) return;
 	for (const key of USAGE_FIELDS) {
 		const count = value[key];
 		if (typeof count !== "number" || !Number.isSafeInteger(count) || count < 0) continue;
@@ -9392,17 +9408,17 @@ function processStreamLine$2(line, state) {
 	}
 	if (!json) return [];
 	const chunk = safeJsonParse$5(json);
-	if (!isRecord$26(chunk)) return [];
-	const responseData = isRecord$26(chunk.response) ? chunk.response : chunk;
-	const error = isRecord$26(chunk.error) ? chunk.error : isRecord$26(responseData.error) ? responseData.error : void 0;
-	if (error !== void 0) throw new LlmError$1(`Antigravity stream error: ${asString$16(error.message) ?? "unknown error"}`, isContextOverflow(error) ? CONTEXT_OVERFLOW_CODE : "PROVIDER_ERROR");
+	if (!isRecord$27(chunk)) return [];
+	const responseData = isRecord$27(chunk.response) ? chunk.response : chunk;
+	const error = isRecord$27(chunk.error) ? chunk.error : isRecord$27(responseData.error) ? responseData.error : void 0;
+	if (error !== void 0) throw new LlmError$1(`Antigravity stream error: ${asString$17(error.message) ?? "unknown error"}`, isContextOverflow(error) ? CONTEXT_OVERFLOW_CODE : "PROVIDER_ERROR");
 	const candidates = Array.isArray(responseData.candidates) ? responseData.candidates : [];
-	const candidate = isRecord$26(candidates[0]) ? candidates[0] : void 0;
-	const content = isRecord$26(candidate?.content) ? candidate.content : void 0;
+	const candidate = isRecord$27(candidates[0]) ? candidates[0] : void 0;
+	const content = isRecord$27(candidate?.content) ? candidate.content : void 0;
 	const parts = Array.isArray(content?.parts) ? content.parts : [];
 	const out = [];
 	for (const part of parts) {
-		if (!isRecord$26(part)) continue;
+		if (!isRecord$27(part)) continue;
 		if (typeof part.text === "string" && part.text !== "") {
 			const isThinking = Boolean(part.thought);
 			const blockType = isThinking ? "reasoning" : "text";
@@ -9434,7 +9450,7 @@ function processStreamLine$2(line, state) {
 				index: state.currentBlock.index,
 				text: delta
 			});
-		} else if (!isRecord$26(part.functionCall) && thoughtSignature(part)) {
+		} else if (!isRecord$27(part.functionCall) && thoughtSignature(part)) {
 			if (state.replayBlocks.length === 0) {
 				const type = part.thought ? "reasoning" : "text";
 				state.blocks.push({
@@ -9458,12 +9474,12 @@ function processStreamLine$2(line, state) {
 			}
 			state.replayBlocks[state.replayBlocks.length - 1].parts.push(replayPart(part));
 		}
-		if (isRecord$26(part.functionCall)) {
+		if (isRecord$27(part.functionCall)) {
 			out.push(...closeCurrentBlock(state));
 			const fc = part.functionCall;
-			const toolName = asString$16(fc.name) || "";
-			const toolId = asString$16(fc.id) || sanitizeToolCallId("", toolName);
-			const argsText = JSON.stringify(isRecord$26(fc.args) ? fc.args : {});
+			const toolName = asString$17(fc.name) || "";
+			const toolId = asString$17(fc.id) || sanitizeToolCallId("", toolName);
+			const argsText = JSON.stringify(isRecord$27(fc.args) ? fc.args : {});
 			const index = state.blocks.length;
 			const block = {
 				type: "tool-call",
@@ -9500,7 +9516,7 @@ function processStreamLine$2(line, state) {
 	}
 	collectUsage(chunk.usageMetadata, state);
 	if (responseData !== chunk) collectUsage(responseData.usageMetadata, state);
-	const finishReason = asString$16(candidate?.finishReason) || asString$16(responseData.finishReason);
+	const finishReason = asString$17(candidate?.finishReason) || asString$17(responseData.finishReason);
 	if (finishReason) {
 		state.finishReason = finishReason;
 		out.push(...closeCurrentBlock(state));
@@ -11859,13 +11875,13 @@ function commandCodeHeaders(apiKey, extra = {}) {
 function timeoutSignal$3(signal, ms) {
 	return signal ? AbortSignal.any([signal, AbortSignal.timeout(ms)]) : AbortSignal.timeout(ms);
 }
-function isRecord$25(value) {
+function isRecord$26(value) {
 	return typeof value === "object" && value !== null && !Array.isArray(value);
 }
 function asRecord$9(value) {
-	return isRecord$25(value) ? value : void 0;
+	return isRecord$26(value) ? value : void 0;
 }
-function asString$15(value) {
+function asString$16(value) {
 	if (typeof value === "string" && value.trim() !== "") return value;
 	if (typeof value === "number" && Number.isFinite(value)) return String(value);
 }
@@ -11878,7 +11894,7 @@ function asNumber$3(value) {
 }
 function firstString$1(record, keys) {
 	for (const key of keys) {
-		const value = asString$15(record[key]);
+		const value = asString$16(record[key]);
 		if (value !== void 0) return value;
 	}
 }
@@ -11907,7 +11923,7 @@ function deepValue(value, keys, depth = 0) {
 		}
 		return;
 	}
-	if (!isRecord$25(value)) return void 0;
+	if (!isRecord$26(value)) return void 0;
 	for (const key of keys) if (value[key] !== void 0 && value[key] !== null) return value[key];
 	for (const nested of Object.values(value)) {
 		const hit = deepValue(nested, keys, depth + 1);
@@ -11921,7 +11937,7 @@ function collectRecords(value, keys, depth = 0, out = []) {
 		for (const entry of value) collectRecords(entry, keys, depth + 1, out);
 		return out;
 	}
-	if (!isRecord$25(value)) return out;
+	if (!isRecord$26(value)) return out;
 	if (keys.some((key) => value[key] !== void 0)) out.push(value);
 	for (const nested of Object.values(value)) collectRecords(nested, keys, depth + 1, out);
 	return out;
@@ -12125,7 +12141,7 @@ function parseCreditBalances(payload, consumed = []) {
 /** Resolve the plan id the subscription or credits payload reports, when either does. */
 function parsePlanId(payloads) {
 	for (const payload of payloads) {
-		const id = asString$15(deepValue(payload, [
+		const id = asString$16(deepValue(payload, [
 			"planId",
 			"plan_id",
 			"priceId",
@@ -12144,7 +12160,7 @@ function parsePlanId(payloads) {
 */
 function parseSubscriptionStatus(payload) {
 	const root = asRecord$9(payload) ?? {};
-	return asString$15(deepValue(asRecord$9(root.data) ?? root, ["status"])) ?? null;
+	return asString$16(deepValue(asRecord$9(root.data) ?? root, ["status"])) ?? null;
 }
 /** End of the current billing period, in Unix milliseconds. */
 function parseSubscriptionPeriodEnd(payload) {
@@ -12696,10 +12712,10 @@ function buildModelOptions$2(catalog, enabledModelIds, overrides) {
 *
 * https://commandcode.ai/blog/command-code-provider-api
 */
-function isRecord$24(value) {
+function isRecord$25(value) {
 	return typeof value === "object" && value !== null && !Array.isArray(value);
 }
-function asString$14(value) {
+function asString$15(value) {
 	return typeof value === "string" ? value : void 0;
 }
 function safeJsonParse$4(text) {
@@ -12736,17 +12752,17 @@ const SUPPORTED_IMAGE_MEDIA_TYPES$4 = /* @__PURE__ */ new Set([
 ]);
 function attachmentOf$4(block) {
 	const attachment = block.attachment;
-	if (!isRecord$24(attachment)) return void 0;
+	if (!isRecord$25(attachment)) return void 0;
 	return typeof attachment.attachmentId === "string" ? attachment : void 0;
 }
 function attachmentLabel$5(block) {
-	const attachment = isRecord$24(block.attachment) ? block.attachment : void 0;
-	return asString$14(attachment?.name) || asString$14(attachment?.attachmentId);
+	const attachment = isRecord$25(block.attachment) ? block.attachment : void 0;
+	return asString$15(attachment?.name) || asString$15(attachment?.attachmentId);
 }
 function collectImageRefs$3(content, refs) {
 	if (!Array.isArray(content)) return;
 	for (const block of content) {
-		if (!isRecord$24(block)) continue;
+		if (!isRecord$25(block)) continue;
 		if (block.type === "image") {
 			const attachment = attachmentOf$4(block);
 			if (attachment) refs.set(attachment.attachmentId, attachment);
@@ -12761,13 +12777,13 @@ function base64Length$3(bytes) {
 function requestImageBytes$3(block) {
 	const attachment = attachmentOf$4(block);
 	if (attachment) return base64Length$3(attachment.bytes);
-	const inline = asString$14(block.data) || asString$14(block.base64);
+	const inline = asString$15(block.data) || asString$15(block.base64);
 	return inline ? inline.length : void 0;
 }
 function collectRequestImageBytes$3(content, lengths) {
 	if (!Array.isArray(content)) return;
 	for (const block of content) {
-		if (!isRecord$24(block) || block.type !== "image") continue;
+		if (!isRecord$25(block) || block.type !== "image") continue;
 		const bytes = requestImageBytes$3(block);
 		if (bytes !== void 0) lengths.push(bytes);
 	}
@@ -12794,7 +12810,7 @@ function offloadOldestRequestImages$3(options) {
 		if (remaining.count === 0 || !Array.isArray(message.content)) return message;
 		let replaced = false;
 		const content = message.content.map((block) => {
-			if (remaining.count === 0 || !isRecord$24(block) || block.type !== "image") return block;
+			if (remaining.count === 0 || !isRecord$25(block) || block.type !== "image") return block;
 			if (requestImageBytes$3(block) === void 0) return block;
 			remaining.count -= 1;
 			replaced = true;
@@ -12847,10 +12863,10 @@ function unavailableImageText$5(block) {
 	return `[image unavailable: ${label ? `${label} could not be read` : "the image could not be read"}; ask the user to attach it again if the image is needed]`;
 }
 function imageBlockToInline$3(block, images) {
-	let data = asString$14(block.data) || asString$14(block.base64);
-	const source = isRecord$24(block.source) ? block.source : void 0;
-	if (!data && source) data = asString$14(source.data) || asString$14(source.base64);
-	let mediaType = asString$14(block.mimeType) || asString$14(block.mediaType) || (source ? asString$14(source.mimeType) || asString$14(source.mediaType) : void 0) || "image/png";
+	let data = asString$15(block.data) || asString$15(block.base64);
+	const source = isRecord$25(block.source) ? block.source : void 0;
+	if (!data && source) data = asString$15(source.data) || asString$15(source.base64);
+	let mediaType = asString$15(block.mimeType) || asString$15(block.mediaType) || (source ? asString$15(source.mimeType) || asString$15(source.mediaType) : void 0) || "image/png";
 	if (data?.startsWith("data:")) {
 		const matched = data.match(/^data:([^;,]+);base64,(.*)$/s);
 		if (matched) {
@@ -12874,7 +12890,7 @@ function textOf$5(content) {
 	if (!Array.isArray(content)) return "";
 	const parts = [];
 	for (const block of content) {
-		if (!isRecord$24(block)) continue;
+		if (!isRecord$25(block)) continue;
 		if (block.type === "text" && typeof block.text === "string") parts.push(sanitizeText$4(block.text));
 		else if (block.type === "tool-result") parts.push(textOf$5(block.content));
 	}
@@ -12883,7 +12899,7 @@ function textOf$5(content) {
 function toolResultText$4(blocks) {
 	if (!Array.isArray(blocks)) return "";
 	return blocks.map((block) => {
-		if (!isRecord$24(block)) return "";
+		if (!isRecord$25(block)) return "";
 		if (block.type === "text" && typeof block.text === "string") return sanitizeText$4(block.text);
 		if (block.type === "tool-result") return toolResultText$4(block.content);
 		if (block.type === "image") return `[image: ${attachmentLabel$5(block) ?? "attached image"}]`;
@@ -12901,7 +12917,7 @@ function toolResultBlocks$2(blocks, images) {
 	const out = [];
 	let hasImage = false;
 	for (const block of blocks) {
-		if (!isRecord$24(block)) continue;
+		if (!isRecord$25(block)) continue;
 		if (block.type === "text" && typeof block.text === "string") {
 			out.push({
 				type: "text",
@@ -12950,7 +12966,7 @@ function toolResultImageBlocks$2(blocks, images) {
 	if (!Array.isArray(blocks)) return [];
 	const out = [];
 	for (const block of blocks) {
-		if (!isRecord$24(block)) continue;
+		if (!isRecord$25(block)) continue;
 		if (block.type === "image") {
 			const inline = imageBlockToInline$3(block, images);
 			if (inline && SUPPORTED_IMAGE_MEDIA_TYPES$4.has(inline.mediaType)) out.push({
@@ -12994,7 +13010,7 @@ function nonSystemMessages$4(options) {
 }
 /** Drop the JSON-Schema keywords provider gateways reject or ignore. */
 function stripMetaSchema$2(schema) {
-	if (!isRecord$24(schema)) return {
+	if (!isRecord$25(schema)) return {
 		type: "object",
 		properties: {}
 	};
@@ -13011,7 +13027,7 @@ function openAIUserContent$2(message, images) {
 	const parts = [];
 	let hasImage = false;
 	for (const block of message.content) {
-		if (!isRecord$24(block)) continue;
+		if (!isRecord$25(block)) continue;
 		if (block.type === "text" && typeof block.text === "string") {
 			const text = sanitizeText$4(block.text);
 			if (text !== "") parts.push({
@@ -13039,7 +13055,7 @@ function openAIAssistantContent$2(message) {
 	const textParts = [];
 	const toolCalls = [];
 	for (const block of message.content) {
-		if (!isRecord$24(block)) continue;
+		if (!isRecord$25(block)) continue;
 		if (block.type === "text" && typeof block.text === "string") textParts.push(sanitizeText$4(block.text));
 		else if (block.type === "tool-call" && typeof block.name === "string") toolCalls.push({
 			id: typeof block.id === "string" && block.id !== "" ? block.id : `call_${toolCalls.length}`,
@@ -13071,7 +13087,7 @@ function buildOpenAIRequest$1(options, images = NO_RESOLVED_IMAGES$4) {
 			while (index < conversation.length && isToolResultMessage$4(conversation[index])) {
 				const current = conversation[index];
 				const block = current.content[0];
-				const callId = isRecord$24(block) && typeof block.toolCallId === "string" ? block.toolCallId : "";
+				const callId = isRecord$25(block) && typeof block.toolCallId === "string" ? block.toolCallId : "";
 				messages.push({
 					role: "tool",
 					tool_call_id: callId,
@@ -13132,7 +13148,7 @@ function anthropicUserContent$2(message, images) {
 	if (!Array.isArray(message.content)) return [];
 	const blocks = [];
 	for (const block of message.content) {
-		if (!isRecord$24(block)) continue;
+		if (!isRecord$25(block)) continue;
 		if (block.type === "text" && typeof block.text === "string") {
 			const text = sanitizeText$4(block.text);
 			if (text !== "") blocks.push({
@@ -13169,7 +13185,7 @@ function anthropicUserContent$2(message, images) {
 function anthropicAssistantContent$2(message) {
 	const blocks = [];
 	for (const block of message.content) {
-		if (!isRecord$24(block)) continue;
+		if (!isRecord$25(block)) continue;
 		if (block.type === "text" && typeof block.text === "string") {
 			const text = sanitizeText$4(block.text);
 			if (text !== "") blocks.push({
@@ -13182,7 +13198,7 @@ function anthropicAssistantContent$2(message) {
 				type: "tool_use",
 				id: typeof block.id === "string" && block.id !== "" ? block.id : `toolu_${blocks.length}`,
 				name: block.name,
-				input: isRecord$24(parsed) ? parsed : {}
+				input: isRecord$25(parsed) ? parsed : {}
 			});
 		}
 	}
@@ -13276,7 +13292,7 @@ function responsesUserContent(message, images) {
 	if (!Array.isArray(message.content)) return [];
 	const parts = [];
 	for (const block of message.content) {
-		if (!isRecord$24(block)) continue;
+		if (!isRecord$25(block)) continue;
 		if (block.type === "text" && typeof block.text === "string") {
 			const text = sanitizeText$4(block.text);
 			if (text !== "") parts.push({
@@ -13301,7 +13317,7 @@ function responsesToolResultImageBlocks(blocks, images) {
 	if (!Array.isArray(blocks)) return [];
 	const out = [];
 	for (const block of blocks) {
-		if (!isRecord$24(block)) continue;
+		if (!isRecord$25(block)) continue;
 		if (block.type === "image") {
 			const inline = imageBlockToInline$3(block, images);
 			if (inline && SUPPORTED_IMAGE_MEDIA_TYPES$4.has(inline.mediaType)) out.push({
@@ -13328,8 +13344,8 @@ function buildResponsesRequest(options, images = NO_RESOLVED_IMAGES$4) {
 			const imageBlocks = [];
 			while (index < conversation.length && isToolResultMessage$4(conversation[index])) {
 				const current = conversation[index];
-				const toolResult = Array.isArray(current.content) ? current.content.find((b) => isRecord$24(b) && b.type === "tool-result") : void 0;
-				const callId = (isRecord$24(toolResult) && typeof toolResult.toolCallId === "string" ? toolResult.toolCallId : "") || (typeof current.source?.callId === "string" ? current.source.callId : "");
+				const toolResult = Array.isArray(current.content) ? current.content.find((b) => isRecord$25(b) && b.type === "tool-result") : void 0;
+				const callId = (isRecord$25(toolResult) && typeof toolResult.toolCallId === "string" ? toolResult.toolCallId : "") || (typeof current.source?.callId === "string" ? current.source.callId : "");
 				input.push({
 					type: "function_call_output",
 					call_id: callId,
@@ -13352,7 +13368,7 @@ function buildResponsesRequest(options, images = NO_RESOLVED_IMAGES$4) {
 				content
 			});
 			for (const call of toolCalls) {
-				const fn = isRecord$24(call.function) ? call.function : void 0;
+				const fn = isRecord$25(call.function) ? call.function : void 0;
 				input.push({
 					type: "function_call",
 					call_id: call.id,
@@ -13475,25 +13491,33 @@ function processOpenAIStreamLine$1(line, state) {
 	}
 	if (payload === "") return [];
 	const chunk = safeJsonParse$4(payload);
-	if (!isRecord$24(chunk)) return [];
+	if (!isRecord$25(chunk)) return [];
 	const out = [];
-	if (isRecord$24(chunk.error)) throw new LlmError$1(`Command Code stream error: ${asString$14(chunk.error.message) ?? "unknown error"}`, isContextOverflow(chunk.error) ? CONTEXT_OVERFLOW_CODE : "PROVIDER_ERROR");
-	const usage = isRecord$24(chunk.usage) ? chunk.usage : void 0;
+	if (isRecord$25(chunk.error)) {
+		const message = asString$15(chunk.error.message) ?? "unknown error";
+		state.streamError = {
+			vocabulary: "openai",
+			error: chunk.error,
+			message
+		};
+		throw new LlmError$1(`Command Code stream error: ${message}`, isContextOverflow(chunk.error) ? CONTEXT_OVERFLOW_CODE : "PROVIDER_ERROR");
+	}
+	const usage = isRecord$25(chunk.usage) ? chunk.usage : void 0;
 	if (usage) {
 		state.sawUsage = true;
 		const prompt = numberOr$5(usage.prompt_tokens, 0);
-		const details = isRecord$24(usage.prompt_tokens_details) ? usage.prompt_tokens_details : void 0;
+		const details = isRecord$25(usage.prompt_tokens_details) ? usage.prompt_tokens_details : void 0;
 		const cached = details ? numberOr$5(details.cached_tokens, 0) : 0;
 		state.inputTokens = Math.max(0, prompt - cached);
 		state.cacheReadTokens = cached;
 		state.outputTokens = numberOr$5(usage.completion_tokens, state.outputTokens);
-		if (isRecord$24(usage.completion_tokens_details)) state.reasoningTokens = numberOr$5(usage.completion_tokens_details.reasoning_tokens, state.reasoningTokens);
+		if (isRecord$25(usage.completion_tokens_details)) state.reasoningTokens = numberOr$5(usage.completion_tokens_details.reasoning_tokens, state.reasoningTokens);
 	}
 	const choices = Array.isArray(chunk.choices) ? chunk.choices : [];
-	const choice = isRecord$24(choices[0]) ? choices[0] : void 0;
-	const delta = isRecord$24(choice?.delta) ? choice.delta : void 0;
+	const choice = isRecord$25(choices[0]) ? choices[0] : void 0;
+	const delta = isRecord$25(choice?.delta) ? choice.delta : void 0;
 	if (delta) {
-		const reasoning = asString$14(delta.reasoning_content) ?? asString$14(delta.reasoning);
+		const reasoning = asString$15(delta.reasoning_content) ?? asString$15(delta.reasoning);
 		if (reasoning !== void 0 && reasoning !== "") {
 			out.push(...closeToolCalls$4(state));
 			if (state.current === null || state.current.type !== "reasoning") out.push(...openTextBlock$3(state, "reasoning"));
@@ -13505,7 +13529,7 @@ function processOpenAIStreamLine$1(line, state) {
 				text: sanitizeText$4(reasoning)
 			});
 		}
-		const content = asString$14(delta.content);
+		const content = asString$15(delta.content);
 		if (content !== void 0 && content !== "") {
 			out.push(...closeToolCalls$4(state));
 			if (state.current === null || state.current.type !== "text") out.push(...openTextBlock$3(state, "text"));
@@ -13519,11 +13543,11 @@ function processOpenAIStreamLine$1(line, state) {
 		}
 		const toolDeltas = Array.isArray(delta.tool_calls) ? delta.tool_calls : [];
 		for (const entry of toolDeltas) {
-			if (!isRecord$24(entry)) continue;
+			if (!isRecord$25(entry)) continue;
 			out.push(...applyOpenAIToolDelta$1(entry, state));
 		}
 	}
-	const finish = asString$14(choice?.finish_reason);
+	const finish = asString$15(choice?.finish_reason);
 	if (finish !== void 0 && finish !== "") {
 		state.finishReason = finish;
 		out.push(...closeCurrent$3(state));
@@ -13533,15 +13557,15 @@ function processOpenAIStreamLine$1(line, state) {
 }
 function applyOpenAIToolDelta$1(entry, state) {
 	const wireIndex = typeof entry.index === "number" ? entry.index : 0;
-	const fn = isRecord$24(entry.function) ? entry.function : {};
+	const fn = isRecord$25(entry.function) ? entry.function : {};
 	const out = [];
 	let call = state.toolCalls.get(wireIndex);
 	if (call === void 0) {
 		out.push(...closeCurrent$3(state));
 		call = {
 			blockIndex: state.blocks.length,
-			id: asString$14(entry.id) ?? `call_${wireIndex}`,
-			name: asString$14(fn.name) ?? "",
+			id: asString$15(entry.id) ?? `call_${wireIndex}`,
+			name: asString$15(fn.name) ?? "",
 			arguments: "",
 			started: false
 		};
@@ -13554,13 +13578,13 @@ function applyOpenAIToolDelta$1(entry, state) {
 		state.toolCalls.set(wireIndex, call);
 	} else {
 		if (call.id === `call_${wireIndex}`) {
-			const id = asString$14(entry.id);
+			const id = asString$15(entry.id);
 			if (id !== void 0) call.id = id;
 		}
-		const name = asString$14(fn.name);
+		const name = asString$15(fn.name);
 		if (name !== void 0 && name !== "") call.name = name;
 	}
-	const argsDelta = asString$14(fn.arguments) ?? "";
+	const argsDelta = asString$15(fn.arguments) ?? "";
 	if (argsDelta !== "") call.arguments += argsDelta;
 	if (!call.started) {
 		call.started = true;
@@ -13591,12 +13615,12 @@ function processAnthropicStreamLine$1(line, state) {
 	const payload = trimmed.slice(5).trim();
 	if (payload === "" || payload === "[DONE]") return [];
 	const event = safeJsonParse$4(payload);
-	if (!isRecord$24(event)) return [];
-	const type = asString$14(event.type);
+	if (!isRecord$25(event)) return [];
+	const type = asString$15(event.type);
 	const out = [];
 	if (type === "message_start") {
-		const message = isRecord$24(event.message) ? event.message : void 0;
-		const usage = message && isRecord$24(message.usage) ? message.usage : void 0;
+		const message = isRecord$25(event.message) ? event.message : void 0;
+		const usage = message && isRecord$25(message.usage) ? message.usage : void 0;
 		if (usage) {
 			state.sawUsage = true;
 			state.inputTokens = numberOr$5(usage.input_tokens, 0);
@@ -13604,21 +13628,21 @@ function processAnthropicStreamLine$1(line, state) {
 			state.cacheWriteTokens = numberOr$5(usage.cache_creation_input_tokens, 0);
 			state.outputTokens = numberOr$5(usage.output_tokens, 0);
 		}
-		const stop = message ? asString$14(message.stop_reason) : void 0;
+		const stop = message ? asString$15(message.stop_reason) : void 0;
 		if (stop !== void 0 && stop !== null) state.finishReason = stop;
 		return out;
 	}
 	if (type === "content_block_start") {
 		const contentIndex = numberOr$5(event.index, 0);
-		const block = isRecord$24(event.content_block) ? event.content_block : {};
-		const blockType = asString$14(block.type);
+		const block = isRecord$25(event.content_block) ? event.content_block : {};
+		const blockType = asString$15(block.type);
 		out.push(...closeCurrent$3(state));
 		if (blockType === "tool_use") {
 			const index = state.blocks.length;
 			const pending = {
 				blockIndex: index,
-				id: asString$14(block.id) ?? `toolu_${contentIndex}`,
-				name: asString$14(block.name) ?? "",
+				id: asString$15(block.id) ?? `toolu_${contentIndex}`,
+				name: asString$15(block.name) ?? "",
 				arguments: "",
 				started: true
 			};
@@ -13660,11 +13684,11 @@ function processAnthropicStreamLine$1(line, state) {
 	}
 	if (type === "content_block_delta") {
 		const contentIndex = numberOr$5(event.index, 0);
-		const delta = isRecord$24(event.delta) ? event.delta : {};
-		const deltaType = asString$14(delta.type);
+		const delta = isRecord$25(event.delta) ? event.delta : {};
+		const deltaType = asString$15(delta.type);
 		if (deltaType === "input_json_delta") {
 			const pending = state.toolCalls.get(contentIndex);
-			const partial = asString$14(delta.partial_json) ?? "";
+			const partial = asString$15(delta.partial_json) ?? "";
 			if (pending !== void 0) {
 				pending.arguments += partial;
 				out.push({
@@ -13677,7 +13701,7 @@ function processAnthropicStreamLine$1(line, state) {
 			}
 			return out;
 		}
-		const text = deltaType === "thinking_delta" ? asString$14(delta.thinking) : asString$14(delta.text);
+		const text = deltaType === "thinking_delta" ? asString$15(delta.thinking) : asString$15(delta.text);
 		if (text !== void 0 && text !== "") {
 			const index = state.contentIndexes.get(contentIndex) ?? state.current?.index;
 			const kind = deltaType === "thinking_delta" ? "reasoning" : "text";
@@ -13733,10 +13757,10 @@ function processAnthropicStreamLine$1(line, state) {
 		return out;
 	}
 	if (type === "message_delta") {
-		const delta = isRecord$24(event.delta) ? event.delta : void 0;
-		const stop = delta ? asString$14(delta.stop_reason) : void 0;
+		const delta = isRecord$25(event.delta) ? event.delta : void 0;
+		const stop = delta ? asString$15(delta.stop_reason) : void 0;
 		if (stop !== void 0 && stop !== "") state.finishReason = stop;
-		const usage = isRecord$24(event.usage) ? event.usage : void 0;
+		const usage = isRecord$25(event.usage) ? event.usage : void 0;
 		if (usage) {
 			state.sawUsage = true;
 			state.outputTokens = numberOr$5(usage.output_tokens, state.outputTokens);
@@ -13748,8 +13772,14 @@ function processAnthropicStreamLine$1(line, state) {
 		return closeStream$4(state);
 	}
 	if (type === "error") {
-		const error = isRecord$24(event.error) ? event.error : {};
-		throw new LlmError$1(`Command Code stream error: ${asString$14(error.message) ?? "unknown error"}`, isContextOverflow(error) ? CONTEXT_OVERFLOW_CODE : "PROVIDER_ERROR");
+		const error = isRecord$25(event.error) ? event.error : {};
+		const message = asString$15(error.message) ?? "unknown error";
+		state.streamError = {
+			vocabulary: "anthropic",
+			error,
+			message
+		};
+		throw new LlmError$1(`Command Code stream error: ${message}`, isContextOverflow(error) ? CONTEXT_OVERFLOW_CODE : "PROVIDER_ERROR");
 	}
 	return out;
 }
@@ -13788,7 +13818,7 @@ function processResponsesStreamLine(line, state) {
 	} catch {
 		return [];
 	}
-	if (!isRecord$24(parsed)) return [];
+	if (!isRecord$25(parsed)) return [];
 	const type = typeof parsed.type === "string" ? parsed.type : "";
 	if (type === "response.output_text.delta") {
 		const delta = typeof parsed.delta === "string" ? parsed.delta : "";
@@ -13815,7 +13845,7 @@ function processResponsesStreamLine(line, state) {
 		}];
 	}
 	if (type === "response.output_item.added") {
-		const item = isRecord$24(parsed.item) ? parsed.item : null;
+		const item = isRecord$25(parsed.item) ? parsed.item : null;
 		if (item === null || item.type !== "function_call") return [];
 		return startResponsesToolCall(state, parsed, item);
 	}
@@ -13839,19 +13869,24 @@ function processResponsesStreamLine(line, state) {
 		return [];
 	}
 	if (type === "response.completed" || type === "response.incomplete") {
-		const response = isRecord$24(parsed.response) ? parsed.response : null;
+		const response = isRecord$25(parsed.response) ? parsed.response : null;
 		if (response !== null) {
 			if (response.status === "failed") {
 				state.done = true;
-				const errorObj = isRecord$24(response.error) ? response.error : null;
+				const errorObj = isRecord$25(response.error) ? response.error : null;
 				const detail = errorObj && typeof errorObj.message === "string" ? errorObj.message : typeof response.error === "string" ? response.error : "";
+				state.streamError = {
+					vocabulary: "responses",
+					error: errorObj,
+					message: detail
+				};
 				throw new LlmError$1(detail === "" ? "Command Code responses stream failed" : `Command Code responses stream failed: ${detail}`, "PROVIDER_ERROR");
 			}
 			readResponsesUsage(response, state);
 		}
 		state.done = true;
 		if (type === "response.incomplete") {
-			const details = response && isRecord$24(response.incomplete_details) ? response.incomplete_details : null;
+			const details = response && isRecord$25(response.incomplete_details) ? response.incomplete_details : null;
 			const reason = details && typeof details.reason === "string" ? details.reason : "";
 			state.finishReason = reason === "max_output_tokens" || reason === "length" || reason === "max_tokens" ? "length" : reason || "length";
 		} else state.finishReason = "stop";
@@ -13859,8 +13894,13 @@ function processResponsesStreamLine(line, state) {
 	}
 	if (type === "response.failed" || type === "error") {
 		state.done = true;
-		const errorObj = isRecord$24(parsed.error) ? parsed.error : isRecord$24(parsed.response) && isRecord$24(parsed.response.error) ? parsed.response.error : null;
-		const detail = (typeof parsed.message === "string" ? parsed.message : "") || (errorObj && typeof errorObj.message === "string" ? errorObj.message : "") || (isRecord$24(parsed.response) && typeof parsed.response.error === "string" ? parsed.response.error : typeof parsed.error === "string" ? parsed.error : "");
+		const errorObj = isRecord$25(parsed.error) ? parsed.error : isRecord$25(parsed.response) && isRecord$25(parsed.response.error) ? parsed.response.error : null;
+		const detail = (typeof parsed.message === "string" ? parsed.message : "") || (errorObj && typeof errorObj.message === "string" ? errorObj.message : "") || (isRecord$25(parsed.response) && typeof parsed.response.error === "string" ? parsed.response.error : typeof parsed.error === "string" ? parsed.error : "");
+		state.streamError = {
+			vocabulary: "responses",
+			error: errorObj,
+			message: detail
+		};
 		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 [];
@@ -13943,15 +13983,15 @@ function responsesToolCall(state, key) {
 	for (const call of state.toolCalls.values()) if (call.id === toToolCallId(key)) return call;
 }
 function readResponsesUsage(response, state) {
-	const usage = isRecord$24(response.usage) ? response.usage : null;
+	const usage = isRecord$25(response.usage) ? response.usage : null;
 	if (usage === null) return;
-	const cached = isRecord$24(usage.input_tokens_details) ? numberOrZero(usage.input_tokens_details.cached_tokens) : 0;
-	const written = isRecord$24(usage.input_tokens_details) ? numberOrZero(usage.input_tokens_details.cache_write_tokens) : 0;
+	const cached = isRecord$25(usage.input_tokens_details) ? numberOrZero(usage.input_tokens_details.cached_tokens) : 0;
+	const written = isRecord$25(usage.input_tokens_details) ? numberOrZero(usage.input_tokens_details.cache_write_tokens) : 0;
 	state.inputTokens = Math.max(0, numberOrZero(usage.input_tokens) - cached - written);
 	state.cacheReadTokens = cached;
 	state.cacheWriteTokens = written;
 	state.outputTokens = numberOrZero(usage.output_tokens);
-	const outputDetails = isRecord$24(usage.output_tokens_details) ? usage.output_tokens_details : null;
+	const outputDetails = isRecord$25(usage.output_tokens_details) ? usage.output_tokens_details : null;
 	if (outputDetails !== null) state.reasoningTokens = numberOrZero(outputDetails.reasoning_tokens);
 	state.sawUsage = true;
 }
@@ -13978,6 +14018,136 @@ function assertStreamComplete$4(state) {
 	if (!state.done && state.finishReason === null) throw new LlmError$1("Command Code stream ended before its terminal event", "PROVIDER_ERROR");
 }
 //#endregion
+//#region src/host/common/stream-error.ts
+/**
+* A failure the provider reported INSIDE a 200 stream.
+*
+* A service can answer 200 and then report the failure in the stream, in the
+* same `{ error: { type, message } }` envelope a non-2xx body carries. A mapper
+* has to type that event with one code, and the code that is right in general is
+* PROVIDER_ERROR - it cannot know whether output has started - which is outside
+* every retry set. So a transient failure delivered this way ends the turn on
+* the first try while the identical failure delivered as a status is retried.
+*
+* The reclassification itself is per line, because only the adapter knows
+* whether anything has reached the caller yet, and because every line owns rules
+* its own HTTP path already owns. What is shared is the question each line asks
+* its own classifier: what status WOULD this failure have carried?
+*/
+/**
+* The status an in-band `error` event stands for, or null for a type this
+* vocabulary does not name.
+*
+* The Anthropic-compatible shims (MiniMax Code, Kimi Code, and Command Code's
+* messages stream) speak one vocabulary, and each of these types already has a
+* status: 529 is Anthropic's documented overload, 429 the rate limit, and a 5xx
+* the server-side failure. Naming the STATUS rather than the verdict is what lets
+* each line hand the answer to its own classifier, so a rule that line already
+* owns - a 429 that names an exhausted plan, a credential refusal reported as
+* 403 - keeps owning it here.
+*
+* Null for everything else: a request the provider rejected, a context
+* overflow, a credential refusal, a type from the future. Null is what keeps
+* the mapper's own verdict, and that is the safe direction to fail in - a
+* fallback to some made-up status would file an unknown type as a retryable
+* server error, which is worse than not retrying at all.
+*/
+function inBandAnthropicStatus(error) {
+	const type = readErrorType(error);
+	if (type === null) return null;
+	return ANTHROPIC_IN_BAND_STATUS[type] ?? null;
+}
+/** The status each transient Anthropic-compatible in-band type stands for. */
+const ANTHROPIC_IN_BAND_STATUS = {
+	overloaded_error: 529,
+	rate_limit_error: 429,
+	api_error: 500
+};
+/**
+* The DSH code an in-band Responses-vocabulary failure becomes, or null when it
+* names nothing transient.
+*
+* The Responses stream reports failure as `response.failed` or `error`, whose
+* message is free text and whose structured evidence is split across TWO fields:
+* this plugin's own routes document `{"error":{"message":"Upstream model
+* provider is temporarily unavailable. Please try again in a moment.","type":
+* "server_error"}}` - `type`, not `code` - so reading only `code` would miss
+* the exact body this exists for, and its message names nothing the heuristic
+* would catch. Both fields are read, and the message is read too, because a
+* deployment that says "overloaded" in prose is the whole signal there.
+*
+* Anything not recognized keeps the mapper's verdict.
+*/
+function inBandResponsesCode(error, message) {
+	const fields = isRecord$24(error) ? error : {};
+	const rawCode = (asString$14(fields.code) ?? asString$14(fields.type) ?? "").toLowerCase();
+	const text = message.toLowerCase();
+	if (text.includes("rate limit") || rawCode === "rate_limit" || rawCode === "rate_limit_exceeded") return "RATE_LIMIT";
+	if (text.includes("overload") || text.includes("server error") || rawCode === "server_error" || rawCode === "service_unavailable" || rawCode === "internal_error") return "SERVER";
+	return null;
+}
+/**
+* The error one in-band stream failure becomes, or `thrown` itself when the
+* line's own verdict stands.
+*
+* Call this only while nothing has reached the caller: the whole reason a fresh
+* request is still free is what makes the retry safe (see module note 1 in the
+* Claude adapter, which this generalizes). After the first chunk the mapper's
+* verdict stands and this must not be called at all.
+*
+* Only the CODE changes. The message stays the mapper's, which names the wire
+* type the user saw, and no `status` is attached: the response really was a
+* 200, and a synthetic one would misreport what the provider said.
+*
+* @param thrown - the mapper's own verdict, passed through when it stands.
+* @param rawError - the event's wire `error` object, from the stream state.
+* @param classify - this line's own failure classifier, so every rule its HTTP
+*   path owns still owns the verdict reached here.
+* @param statusFor - the vocabulary's status table; Responses lines pass
+*   {@link inBandResponsesCode} through {@link reclassifyInBandResponsesError}.
+*/
+function reclassifyInBandError(thrown, rawError, classify, statusFor = inBandAnthropicStatus) {
+	const status = statusFor(rawError);
+	if (status === null) return thrown;
+	const failure = classify(status, JSON.stringify({ error: rawError }));
+	if (!failure.retryable) return thrown;
+	return new LlmError(thrown.message, failure.code, { cause: thrown });
+}
+/**
+* {@link reclassifyInBandError} for a line whose classifier reads a status, fed
+* by the Responses vocabulary instead of the Anthropic one.
+*
+* The transient verdict found here is a QUESTION, not the answer: a
+* `rate_limit` code says only that the failure is worth retrying, and whether
+* it actually is - a 429 naming a spent balance is this route's PROVIDER_ERROR,
+* not a RATE_LIMIT - is a rule the line's own classifier owns. So the answer is
+* asked of `classify` at the status the verdict stands for, exactly as the
+* Anthropic sibling does, and the CODE that comes back is the one the HTTP path
+* would have produced for the same body. That agreement is the whole point: the
+* in-band delivery must not be the one delivery with its own opinion.
+*/
+function reclassifyInBandResponsesError(thrown, rawError, message, classify) {
+	const code = inBandResponsesCode(rawError, message);
+	if (code === null) return thrown;
+	const failure = classify(code === "RATE_LIMIT" ? 429 : 500, JSON.stringify({
+		error: rawError,
+		message
+	}));
+	if (!failure.retryable) return thrown;
+	return new LlmError(thrown.message, failure.code, { cause: thrown });
+}
+/** The wire `error` object of an in-band event, read defensively. */
+function readErrorType(error) {
+	if (!isRecord$24(error)) return null;
+	return typeof error.type === "string" ? error.type : null;
+}
+function asString$14(value) {
+	return typeof value === "string" ? value : void 0;
+}
+function isRecord$24(value) {
+	return typeof value === "object" && value !== null && !Array.isArray(value);
+}
+//#endregion
 //#region src/host/common/capabilities.ts
 /** Error thrown by requireCapability when a capability contract fails. */
 var CapabilityError = class extends Error {
@@ -14253,6 +14423,35 @@ function isCommandCodeCredentialInvalid(status, detail) {
 	}
 	return false;
 }
+/**
+* The verdict one failed response becomes on this route.
+*
+* An in-band `error` event is a status the provider chose not to send as a
+* status, so the chain that answers a non-2xx body has to answer it too rather
+* than a second, looser one: every rule this route already owns - a model the
+* plan does not include, a rejected key, a spent quota - keeps owning the
+* answer, and the codes below are the ONLY place the HTTP path reads them from.
+*
+* `retryable` is membership of this route's own policy rather than a second
+* judgement, because a code is worth repeating exactly when RETRY_POLICY repeats
+* it; a verdict the policy would not act on must not be filed as transient.
+*
+* @param status - the response status, or the status an in-band type stands for.
+* @param detail - the response body, or the serialized in-band error envelope.
+*/
+function classifyCommandCodeFailure(status, detail) {
+	let code;
+	if (status === 422) code = "PROVIDER_ERROR";
+	else if (isCommandCodeModelAccessDenied(status, detail)) code = "PROVIDER_ERROR";
+	else if (isCommandCodeCredentialInvalid(status, detail) || status === 401) code = "INVALID_CREDENTIAL";
+	else if (status === 429) code = "RATE_LIMIT";
+	else if (status >= 500) code = "SERVER";
+	else code = isHttpContextOverflow(status, detail) ? CONTEXT_OVERFLOW_CODE : "PROVIDER_ERROR";
+	return {
+		code,
+		retryable: RETRY_POLICY$5.mode === "normal" && RETRY_POLICY$5.retryableCodes.includes(code)
+	};
+}
 /** Cooldown one rate-limited key takes when the provider states no delay. */
 const POOL_COOLDOWN_MS$4 = 15 * 6e4;
 var CommandCodeAdapter = class extends LlmAdapter {
@@ -14447,49 +14646,105 @@ var CommandCodeAdapter = class extends LlmAdapter {
 		}
 		if (response === void 0 || !response.ok) {
 			const status = response?.status ?? 500;
+			const { code } = classifyCommandCodeFailure(status, detail);
 			if (status === 422) {
-				if (detail.includes("cmd_zdr_no_providers")) throw new LlmError(`${PROVIDER_NAME$5} rejected request under Zero Data Retention: no ZDR-capable upstream is available for this model (${detail || "cmd_zdr_no_providers"}).`, "PROVIDER_ERROR", { status: 422 });
-				throw new LlmError(`${PROVIDER_NAME$5} validation error (422): ${detail || "Unprocessable Entity"}`, "PROVIDER_ERROR", { status: 422 });
+				if (detail.includes("cmd_zdr_no_providers")) throw new LlmError(`${PROVIDER_NAME$5} rejected request under Zero Data Retention: no ZDR-capable upstream is available for this model (${detail || "cmd_zdr_no_providers"}).`, code, { status: 422 });
+				throw new LlmError(`${PROVIDER_NAME$5} validation error (422): ${detail || "Unprocessable Entity"}`, code, { status: 422 });
 			}
-			if (isCommandCodeModelAccessDenied(status, detail)) throw new LlmError(`${PROVIDER_NAME$5} access denied for model ${options.model}: this model is not included in the plan or requires higher entitlement (${status}).${detail ? ` ${detail}` : ""}`, "PROVIDER_ERROR", { status });
-			if (isCommandCodeCredentialInvalid(status, detail) || status === 401) throw new LlmError(`${PROVIDER_NAME$5} rejected the stored API key (${status}). Sign in again from Settings > Command Code.${detail ? ` ${detail}` : ""}`, "INVALID_CREDENTIAL", { status });
+			if (isCommandCodeModelAccessDenied(status, detail)) throw new LlmError(`${PROVIDER_NAME$5} access denied for model ${options.model}: this model is not included in the plan or requires higher entitlement (${status}).${detail ? ` ${detail}` : ""}`, code, { status });
+			if (isCommandCodeCredentialInvalid(status, detail) || status === 401) throw new LlmError(`${PROVIDER_NAME$5} rejected the stored API key (${status}). Sign in again from Settings > Command Code.${detail ? ` ${detail}` : ""}`, code, { status });
 			if (status === 429) {
 				const after = response === void 0 ? void 0 : retryAfterMs$1(response.headers);
-				throw new LlmError(`${PROVIDER_NAME$5} rate limit or plan quota reached (429). Check the quota card in Settings > Command Code.${detail ? ` ${detail}` : ""}`, "RATE_LIMIT", {
+				throw new LlmError(`${PROVIDER_NAME$5} rate limit or plan quota reached (429). Check the quota card in Settings > Command Code.${detail ? ` ${detail}` : ""}`, code, {
 					status: 429,
 					...after === void 0 ? {} : { providerRetryAfterMs: after }
 				});
 			}
-			if (status >= 500) throw new LlmError(`${PROVIDER_NAME$5} upstream server error (${status}): ${detail || "No response"}`, "SERVER", { status });
-			throw new LlmError(`${PROVIDER_NAME$5} API error (${status}): ${detail || "No response"}`, isHttpContextOverflow(status, detail) ? CONTEXT_OVERFLOW_CODE : "PROVIDER_ERROR", { status });
+			if (status >= 500) throw new LlmError(`${PROVIDER_NAME$5} upstream server error (${status}): ${detail || "No response"}`, code, { status });
+			throw new LlmError(`${PROVIDER_NAME$5} API error (${status}): ${detail || "No response"}`, code, { status });
 		}
 		if (response.body === null) throw new LlmError("Command Code returned an empty response body", "PROVIDER_ERROR");
 		const reader = response.body.getReader();
 		const decoder = new TextDecoder();
 		const state = createStreamState$5(wire);
 		let buffer = "";
+		/**
+		* Set the moment ANY chunk reaches the caller, and never cleared.
+		*
+		* Deliberately broader than `a text delta or a tool call`: a retry would repeat
+		* a block-start the caller has already seen just as surely, and could re-issue
+		* a tool call the agent has already run, so the conservative reading is the
+		* one that cannot be wrong. It is what keeps the in-band reclassification
+		* below pre-output only, and before that point a fresh request is still free.
+		*/
+		let outputStarted = false;
 		try {
-			while (true) {
-				const { done, value } = await reader.read();
-				if (done) break;
-				buffer += decoder.decode(value, { stream: true });
-				const lines = buffer.split("\n");
-				buffer = lines.pop() ?? "";
-				for (const line of lines) {
-					for (const chunk of processLine$1(line, state, wire)) yield chunk;
-					if (state.finished) return;
+			try {
+				while (true) {
+					const { done, value } = await reader.read();
+					if (done) break;
+					buffer += decoder.decode(value, { stream: true });
+					const lines = buffer.split("\n");
+					buffer = lines.pop() ?? "";
+					for (const line of lines) {
+						for (const chunk of processLine$1(line, state, wire)) {
+							outputStarted = true;
+							yield chunk;
+						}
+						if (state.finished) return;
+					}
+				}
+				buffer += decoder.decode();
+				if (buffer.trim() !== "") for (const line of buffer.split("\n")) for (const chunk of processLine$1(line, state, wire)) {
+					outputStarted = true;
+					yield chunk;
+				}
+				if (state.finished) return;
+				assertStreamComplete$4(state);
+				for (const chunk of closeStream$4(state)) {
+					outputStarted = true;
+					yield chunk;
+				}
+			} catch (error) {
+				const inBand = state.streamError;
+				if (error instanceof LlmError && !outputStarted && inBand !== void 0 && error.code === "PROVIDER_ERROR") {
+					const reclassified = reclassifyInBandStreamError(error, inBand);
+					if (reclassified.code === "RATE_LIMIT" && pool !== null && accountId !== void 0) {
+						const after = retryAfterMs$1(response.headers);
+						await pool.markCooldown(accountId, after ?? POOL_COOLDOWN_MS$4, `${PROVIDER_NAME$5} in-stream 429`).catch(() => void 0);
+					}
+					throw reclassified;
 				}
+				throw error;
 			}
-			buffer += decoder.decode();
-			if (buffer.trim() !== "") for (const line of buffer.split("\n")) for (const chunk of processLine$1(line, state, wire)) yield chunk;
-			if (state.finished) return;
-			assertStreamComplete$4(state);
-			for (const chunk of closeStream$4(state)) yield chunk;
 		} finally {
 			reader.cancel().catch(() => void 0);
 		}
 	}
 };
+/**
+* The error one recorded in-band failure becomes, or the mapper's own verdict.
+*
+* Two rules, because this line speaks three in-band vocabularies and they do
+* not carry the same evidence. The messages wire names a transient failure with
+* a `type` that STANDS FOR A STATUS - an `overloaded_error` is the 529 it would
+* have sent - so its envelope is answered by the status table. The
+* chat-completions and Responses wires have no such type: their evidence is a
+* `code`, a `type` in the looser `server_error` sense, or plain prose, and the
+* status-reading helper is what turns that into the question this line's own
+* classifier answers.
+*
+* The switch is exhaustive on purpose. A fourth vocabulary must be a compile
+* error here rather than a silent fall into the rule above, which is the one
+* shape of this bug that cannot be caught by a test.
+*/
+function reclassifyInBandStreamError(thrown, inBand) {
+	switch (inBand.vocabulary) {
+		case "anthropic": return reclassifyInBandError(thrown, inBand.error, classifyCommandCodeFailure);
+		case "openai":
+		case "responses": return reclassifyInBandResponsesError(thrown, inBand.error, inBand.message, classifyCommandCodeFailure);
+	}
+}
 /** Path suffix the chosen route answers on. */
 function endpointPathFor(wire) {
 	if (wire === "anthropic") return "/messages";
@@ -15541,6 +15796,24 @@ const DEFAULT_CONTEXT_WINDOW$4 = 128e3;
 * ceiling rather than a claim about what the model can do.
 */
 const DEFAULT_MAX_OUTPUT_TOKENS = 8192;
+/**
+* Raise a cap that is too small for this model's forced thinking.
+*
+* Left alone: a model whose thinking CAN be disabled - its caller may have asked
+* for thinking off on purpose, and raising the cap would not enable it while the
+* number read back would be misleading - and an absent cap, which
+* `DEFAULT_MAX_OUTPUT_TOKENS` already owns at a size nothing is short of.
+*
+* @param modelId - the routed model id.
+* @param requested - the caller's cap, undefined when it stated none.
+* @returns the cap to request, raised only when that is the difference between an
+* empty truncated answer and a usable one.
+*/
+function floorForcedThinkingTokens$1(modelId, requested) {
+	if (requested === void 0) return void 0;
+	if (thinkingModeForModel(modelId) !== "levels") return requested;
+	return Math.max(requested, 512);
+}
 /** How long a catalog sync may take before the cached list is used instead. */
 const CATALOG_TIMEOUT_MS = 15e3;
 /**
@@ -16586,6 +16859,23 @@ var OllamaAdapter = class extends LlmAdapter {
 		let lastStatus;
 		let lastDetail = "";
 		let accountId;
+		/**
+		* Set the moment ANY chunk reaches the caller, and never cleared.
+		*
+		* Deliberately broader than "some text arrived": a block start or a tool
+		* call the caller has already seen is exactly as visible as the text, so the
+		* conservative reading is the one that cannot be wrong. It also survives the
+		* pool rotation below, because output an account produced before failing is
+		* still output.
+		*/
+		let outputStarted = false;
+		/**
+		* The code a transient in-band failure becomes, or null for this attempt.
+		*
+		* Assigned per failure rather than accumulated, so a later attempt that
+		* failed for a reason of its own is never judged by an earlier one's wording.
+		*/
+		let lastInBandCode = null;
 		for (;;) {
 			let credentials;
 			if (pool === null) {
@@ -16613,6 +16903,7 @@ var OllamaAdapter = class extends LlmAdapter {
 				if (event.type === "error") {
 					lastStatus = openedStatus;
 					lastDetail = event.message;
+					lastInBandCode = outputStarted || !(openedStatus !== void 0 && openedStatus >= 200 && openedStatus < 300) ? null : inBandTransientCode(event.message);
 					failed = true;
 					break;
 				}
@@ -16621,12 +16912,18 @@ var OllamaAdapter = class extends LlmAdapter {
 					continue;
 				}
 				if (event.type === "text" || event.type === "thinking" || event.type === "tool_call") {
-					for (const chunk of applyEvent(state, event)) yield chunk;
+					for (const chunk of applyEvent(state, event)) {
+						outputStarted = true;
+						yield chunk;
+					}
 					continue;
 				}
 			}
 			if (!failed) {
-				for (const chunk of closeStream$3(state)) yield chunk;
+				for (const chunk of closeStream$3(state)) {
+					outputStarted = true;
+					yield chunk;
+				}
 				if (pool !== null && accountId !== void 0 && spent !== null) await pool.recordUsage(accountId, spent.inputTokens, spent.outputTokens);
 				return;
 			}
@@ -16646,9 +16943,76 @@ var OllamaAdapter = class extends LlmAdapter {
 		const status = lastStatus;
 		if (status === 401 || status === 403) throw new LlmError(`${PROVIDER_NAME$4} rejected the API key (${status}). Replace it from Settings > ${PROVIDER_NAME$4}.${lastDetail ? ` ${lastDetail}` : ""}`, "INVALID_CREDENTIAL", { status });
 		if (status === 429) throw new LlmError(`${PROVIDER_NAME$4} rate limit or plan quota reached (429). Add another key, or wait for the cooldown shown in Settings > ${PROVIDER_NAME$4}.${lastDetail ? ` ${lastDetail}` : ""}`, "RATE_LIMIT", { status: 429 });
-		throw new LlmError(`${PROVIDER_NAME$4} request failed${status === void 0 ? "" : ` (${status})`}.${lastDetail}`, status !== void 0 && status >= 500 ? "SERVER" : "INVALID_REQUEST", status === void 0 ? {} : { status });
+		throw new LlmError(`${PROVIDER_NAME$4} request failed${status === void 0 ? "" : ` (${status})`}.${lastDetail}`, lastInBandCode ?? (status !== void 0 && status >= 500 ? "SERVER" : "INVALID_REQUEST"), status === void 0 ? {} : { status });
 	}
 };
+/**
+* The DSH code an in-band Ollama failure becomes, or null when the message names
+* nothing transient.
+*
+* Ollama reports a failure inside a 200 as `{"error": "<message>"}` - one string,
+* no type - so the shared helper's Anthropic table cannot apply and there is no
+* status to name either. The wording is the whole signal, which is the same
+* message heuristic the Codex line already reads on `response.failed` in
+* `responses-client.ts` (generalized for the Responses vocabulary by
+* `inBandResponsesCode` in `common/stream-error.ts`); it is kept local here
+* because these are Ollama's own phrasings, not a vocabulary lines share.
+*
+* Every phrase below names a condition Ollama reports that clears on its own: a
+* model still being pulled into memory, a runner that would not start, a server
+* with no slot for this request, an overloaded or throttled front door. That is
+* what makes a retry worth taking - the request did not change, so a later one
+* is genuinely a different bet.
+*
+* Deliberately narrow, because the cost of guessing is paid by the user in
+* backoff. A refusal the wording does not name is usually the service objecting
+* to the caller's message, and retrying repeats it verbatim: three identical
+* requests ending in the same verdict, with the real fault unreported. Falling
+* back to some invented status instead would be worse on both counts, filing an
+* unrecognized refusal as a retryable server error and inviting the pool chain
+* to cool an account down for something nothing observed.
+*/
+function inBandTransientCode(message) {
+	const text = message.toLowerCase();
+	if (IN_BAND_RATE_LIMIT.some((phrase) => text.includes(phrase))) return "RATE_LIMIT";
+	if (IN_BAND_SERVER.some((phrase) => text.includes(phrase))) return "SERVER";
+	return null;
+}
+/**
+* Throttling wording, read first because it is the narrower of the two verdicts.
+*
+* The same conditions this line already types RATE_LIMIT for when they arrive as
+* a 429; a gateway that reports one inside the stream is describing the same
+* state, and the harness backoff that 429 earns is what the turn should get.
+*/
+const IN_BAND_RATE_LIMIT = [
+	"rate limit",
+	"rate-limit",
+	"rate_limit",
+	"too many requests",
+	"quota"
+];
+/**
+* Server-side conditions, in no particular order.
+*
+* The model-loading and runner wording covers the two failures Ollama raises while
+* it is still getting ready to serve - the request was well formed and the
+* service was not ready for it - and the rest is what its front door says when it
+* is saturated.
+*/
+const IN_BAND_SERVER = [
+	"overload",
+	"server busy",
+	"model is loading",
+	"loading model",
+	"couldn't load",
+	"could not load",
+	"unable to load",
+	"failed to load",
+	"runner",
+	"service unavailable",
+	"internal server error"
+];
 function isAbort$3(error, signal) {
 	return signal?.aborted === true || error instanceof Error && error.name === "AbortError";
 }
@@ -16812,7 +17176,8 @@ async function toOllamaRequest(options, attachments, signal) {
 		model: options.model,
 		messages
 	};
-	if (options.maxTokens !== void 0) request.maxOutputTokens = options.maxTokens;
+	const flooredMax = floorForcedThinkingTokens$1(options.model, options.maxTokens);
+	if (flooredMax !== void 0) request.maxOutputTokens = flooredMax;
 	if (options.temperature !== void 0) request.temperature = options.temperature;
 	const think = thinkForModel(options.model, options.reasoningEffort);
 	if (think !== void 0) request.think = think;
@@ -21104,7 +21469,15 @@ function processOpenAIStreamLine(line, state) {
 	if (!isRecord$17(chunk)) return [];
 	const out = [];
 	const errorPayload = isRecord$17(chunk.error) ? chunk.error : void 0;
-	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");
+	if (errorPayload !== void 0) {
+		const message = asString$9(errorPayload.message) ?? "unknown error";
+		state.streamError = {
+			vocabulary: "openai",
+			error: errorPayload,
+			message
+		};
+		throw new LlmError$1(`Kimi Code stream error: ${message}`, isContextOverflow(errorPayload) ? CONTEXT_OVERFLOW_CODE : "PROVIDER_ERROR");
+	}
 	const usage = isRecord$17(chunk.usage) ? chunk.usage : void 0;
 	if (usage) {
 		state.sawUsage = true;
@@ -21373,7 +21746,13 @@ function processAnthropicStreamLine(line, state) {
 	}
 	if (type === "error") {
 		const error = isRecord$17(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");
+		const message = asString$9(error.message) ?? "unknown error";
+		state.streamError = {
+			vocabulary: "anthropic",
+			error,
+			message
+		};
+		throw new LlmError$1(`Kimi Code stream error: ${message}`, isContextOverflow(error) ? CONTEXT_OVERFLOW_CODE : "PROVIDER_ERROR");
 	}
 	return out;
 }
@@ -21610,8 +21989,10 @@ function assertStreamComplete$3(state) {
 * - `INVALID_CREDENTIAL` — a rejected access token fails identically on every
 *   attempt;
 * - `PROVIDER_ERROR` — a 400, a 401 that is really a plan-entitlement refusal,
-*   or a 403 quota limit. Retrying a quota that resets in hours only burns
-*   requests and delays the message the user needs to see;
+*   or a 403 quota limit. Retrying a quota that resets in hours against the
+*   SAME account only burns requests and delays the message the user needs to
+*   see, so none of it is retried; a 403 limit still rotates to another account
+*   when the pool holds one, which is a routing decision rather than a retry;
 * - `ABORTED` — the caller already cancelled.
 *
 * The DSH normal defaults would apply anyway; stating the values here pins them
@@ -21695,17 +22076,20 @@ function classifyKimiFailure(status, bodyText) {
 	if (isHttpContextOverflow(status, bodyText)) return {
 		code: CONTEXT_OVERFLOW_CODE,
 		retryable: false,
+		accountScoped: false,
 		message: `${PROVIDER_NAME$3} context window exceeded: ${detail}`
 	};
 	if (status === 402) return {
 		code: "SERVER",
 		retryable: true,
+		accountScoped: false,
 		message: `${PROVIDER_NAME$3} could not verify the subscription tier (402). Retrying; if it persists, confirm the membership is active.${detail ? ` ${detail}` : ""}`
 	};
 	if (status === 401 || status === 403) {
 		if (matchesAny$1(detail, ENTITLEMENT_PATTERNS)) return {
 			code: "PROVIDER_ERROR",
 			retryable: false,
+			accountScoped: false,
 			message: `${PROVIDER_NAME$3} refused this request for the current plan: ${detail || "the requested model or context is not included"}. Switch to a model the plan includes, lower the context-window override, or upgrade the subscription.`
 		};
 		if (status === 403) {
@@ -21713,12 +22097,14 @@ function classifyKimiFailure(status, bodyText) {
 			return {
 				code: "PROVIDER_ERROR",
 				retryable: false,
+				accountScoped: true,
 				message: `${PROVIDER_NAME$3} blocked the request on an account limit (403): ${detail || (limitReached ? "the account limit was reached" : "the account refused the request")}. The quota refreshes on its own schedule — check the Kimi Code card in Settings for the reset time.`
 			};
 		}
 		return {
 			code: "INVALID_CREDENTIAL",
 			retryable: false,
+			accountScoped: false,
 			message: `${PROVIDER_NAME$3} rejected the stored credential (401). Sign in again from Settings > Kimi Code.${detail ? ` ${detail}` : ""}`
 		};
 	}
@@ -21726,32 +22112,81 @@ function classifyKimiFailure(status, bodyText) {
 		if (matchesAny$1(detail, QUOTA_EXHAUSTED_PATTERNS$1)) return {
 			code: "PROVIDER_ERROR",
 			retryable: false,
+			accountScoped: true,
 			message: `${PROVIDER_NAME$3} reports the account quota is exhausted: ${detail || "no remaining quota"}. Top up or wait for the window to reset.`
 		};
 		return {
 			code: "RATE_LIMIT",
 			retryable: true,
+			accountScoped: false,
 			message: `${PROVIDER_NAME$3} is rate limited or overloaded (429): ${detail || "too many requests"}. Retrying with backoff.`
 		};
 	}
 	if (status >= 500) return {
 		code: "SERVER",
 		retryable: true,
+		accountScoped: false,
 		message: `${PROVIDER_NAME$3} upstream server error (${status}): ${detail || "the model provider is temporarily unavailable"}. Retrying with backoff.`
 	};
 	if (status === 400) return {
 		code: "PROVIDER_ERROR",
 		retryable: false,
+		accountScoped: false,
 		message: `${PROVIDER_NAME$3} rejected the request (400): ${detail || "the request was not accepted"}`
 	};
 	return {
 		code: "PROVIDER_ERROR",
 		retryable: false,
+		accountScoped: false,
 		message: `${PROVIDER_NAME$3} API error (${status}): ${detail || "No response"}`
 	};
 }
+/** Numeric reset the service may ship as a field rather than in prose. */
+const RESET_FIELD_PATTERN = /"?reset[A-Za-z_]*"?\s*[:=]\s*"?(\d{9,16})"?/i;
+/** ISO instant quoted in prose, trusted only when it names its own zone. */
+const RESET_INSTANT_PATTERN = /\d{4}-\d{2}-\d{2}[T ]\d{2}:\d{2}(?::\d{2})?(?:\.\d+)?(?:Z|[+-]\d{2}:?\d{2})/;
+/** The window a billing-cycle refusal names; nothing shorter describes it. */
+const BILLING_CYCLE_PATTERN = /usage limit for this billing cycle/i;
+/**
+* How long one account stays out of rotation after it reports a spent limit.
+*
+* The reset instant is read from the body whenever the service states one — as a
+* numeric field, or as an instant carrying its own zone — and the length of the
+* window the service names is used otherwise. The zone is required for that
+* instant because an unqualified one is ambiguous, and a cooldown computed from
+* the wrong zone is either hours too long or already over; where the body is
+* ambiguous the window length is the honest answer.
+*
+* Both ends of the range are load-bearing. A reset instant barely in the future
+* would otherwise read as no cooldown at all and put the account straight back
+* into rotation to be refused again; and a reset instant already in the past
+* (a stale cache, a clock skew) falls back to the window length above rather
+* than to no cooldown at all.
+*/
+function accountLimitCooldownMs(bodyText, now = Date.now()) {
+	const field = bodyText.match(RESET_FIELD_PATTERN);
+	const instant = bodyText.match(RESET_INSTANT_PATTERN);
+	const resetsAt = field !== null ? parseTimestamp(Number(field[1])) : instant !== null ? Date.parse(instant[0].replace(" ", "T")) : null;
+	if (resetsAt !== null && Number.isFinite(resetsAt) && resetsAt > now) return Math.min(Math.max(resetsAt - now, POOL_COOLDOWN_MS$2), MAX_ACCOUNT_LIMIT_COOLDOWN_MS);
+	return BILLING_CYCLE_PATTERN.test(bodyText) ? BILLING_CYCLE_COOLDOWN_MS : ACCOUNT_LIMIT_COOLDOWN_MS;
+}
 /** Cooldown one rate-limited account takes when the provider states no delay. */
 const POOL_COOLDOWN_MS$2 = 15 * 6e4;
+/**
+* Fallback cooldown for an account whose usage window is spent.
+*
+* The service states the reset time only sometimes, so the fallback is the
+* window's own length. The 15 minutes a 429 gets would put the account back
+* into rotation four times inside the very window that just refused it.
+*/
+const ACCOUNT_LIMIT_COOLDOWN_MS = 300 * 6e4;
+/**
+* A billing cycle is not a window: a 5-hour cooldown would spend a request every
+* five hours to learn an answer that changes at most once a month.
+*/
+const BILLING_CYCLE_COOLDOWN_MS = 720 * 60 * 6e4;
+/** Ceiling, so a malformed reset time cannot park an account for years. */
+const MAX_ACCOUNT_LIMIT_COOLDOWN_MS = BILLING_CYCLE_COOLDOWN_MS;
 var KimiCodeAdapter = class extends LlmAdapter {
 	store;
 	modelSettings;
@@ -21936,11 +22371,15 @@ var KimiCodeAdapter = class extends LlmAdapter {
 			if (pool !== null && accountId !== void 0) {
 				const planScoped = response.status === 429 && matchesAny$1(detail, ENTITLEMENT_PATTERNS);
 				if (response.status === 429 && !planScoped) {
-					await pool.markCooldown(accountId, after ?? POOL_COOLDOWN_MS$2, `${PROVIDER_NAME$3} 429`).catch(() => void 0);
+					const cooldownMs = failure.accountScoped ? accountLimitCooldownMs(detail) : after ?? POOL_COOLDOWN_MS$2;
+					await pool.markCooldown(accountId, cooldownMs, `${PROVIDER_NAME$3} 429`).catch(() => void 0);
 					if (await pool.hasAnotherAvailableAccount(tried)) continue;
 				} else if (failure.code === "INVALID_CREDENTIAL") {
 					await pool.markAuthFailed(accountId, failure.message).catch(() => void 0);
 					if (await pool.hasAnotherAvailableAccount(tried)) continue;
+				} else if (failure.accountScoped) {
+					await pool.markCooldown(accountId, accountLimitCooldownMs(detail), `${PROVIDER_NAME$3} 403`).catch(() => void 0);
+					if (await pool.hasAnotherAvailableAccount(tried)) continue;
 				}
 			}
 			throw new LlmError(failure.message, failure.code, {
@@ -21954,27 +22393,98 @@ var KimiCodeAdapter = class extends LlmAdapter {
 		const state = createStreamState$3(wire, requestOptions.sessionId);
 		markSessionActive(requestOptions.sessionId);
 		let buffer = "";
+		/**
+		* Set the moment ANY chunk reaches the caller, and never cleared.
+		*
+		* Deliberately broader than "text arrived": a block-start the caller has
+		* already seen repeats just as badly as text it has already read, so the
+		* conservative reading is the one that cannot be wrong. It is what keeps the
+		* in-band reclassification below pre-output only.
+		*/
+		let outputStarted = false;
 		try {
-			while (true) {
-				const { done, value } = await reader.read();
-				if (done) break;
-				buffer += decoder.decode(value, { stream: true });
-				const lines = buffer.split("\n");
-				buffer = lines.pop() ?? "";
-				for (const line of lines) {
-					for (const chunk of processLine(line, state, wire)) yield chunk;
-					if (state.finished) return;
+			try {
+				while (true) {
+					const { done, value } = await reader.read();
+					if (done) break;
+					buffer += decoder.decode(value, { stream: true });
+					const lines = buffer.split("\n");
+					buffer = lines.pop() ?? "";
+					for (const line of lines) {
+						for (const chunk of processLine(line, state, wire)) {
+							outputStarted = true;
+							yield chunk;
+						}
+						if (state.finished) return;
+					}
+				}
+				buffer += decoder.decode();
+				if (buffer.trim() !== "") for (const line of buffer.split("\n")) for (const chunk of processLine(line, state, wire)) {
+					outputStarted = true;
+					yield chunk;
+				}
+				if (state.finished) return;
+				assertStreamComplete$3(state);
+				for (const chunk of closeStream$2(state, servedByAccountId)) {
+					outputStarted = true;
+					yield chunk;
 				}
+			} catch (error) {
+				const inBand = state.streamError;
+				if (error instanceof LlmError && !outputStarted && error.code === "PROVIDER_ERROR" && inBand !== void 0) throw await this.inBandStreamFailure(error, inBand, servedByAccountId);
+				throw error;
 			}
-			buffer += decoder.decode();
-			if (buffer.trim() !== "") for (const line of buffer.split("\n")) for (const chunk of processLine(line, state, wire)) yield chunk;
-			if (state.finished) return;
-			assertStreamComplete$3(state);
-			for (const chunk of closeStream$2(state, servedByAccountId)) yield chunk;
 		} finally {
 			reader.cancel().catch(() => void 0);
 		}
 	}
+	/**
+	* What one in-band failure becomes once the pool has been told what the
+	* verdict means for it.
+	*
+	* Reclassification and the pool reaction are separate because they answer
+	* different questions about the same classification: the code decides whether
+	* the harness retries, `accountScoped` decides whether a retry could land
+	* somewhere that can answer. The mapper can supply only the first, which is
+	* why the wire object is recorded there and read here.
+	*
+	* The body is the one a non-2xx response would have carried, so every pattern
+	* below reads exactly what the HTTP branch reads and a limit cannot mean one
+	* thing as a status and another as an event. The request itself is NOT
+	* re-issued: its body is already open and part-read, and the harness retry
+	* policy owns the repeat.
+	*/
+	async inBandStreamFailure(thrown, inBand, accountId) {
+		const bodyText = JSON.stringify({ error: inBand.error });
+		const reclassified = inBand.vocabulary === "anthropic" ? reclassifyInBandError(thrown, inBand.error, classifyKimiFailure) : reclassifyInBandResponsesError(thrown, inBand.error, inBand.message, classifyKimiFailure);
+		const pool = this.accountPool;
+		if (pool === null || accountId === void 0) return reclassified;
+		if (inBand.vocabulary === "anthropic") {
+			const status = inBandAnthropicStatus(inBand.error);
+			if (status === null) return reclassified;
+			const failure = classifyKimiFailure(status, bodyText);
+			const planScoped = status === 429 && matchesAny$1(bodyText, ENTITLEMENT_PATTERNS);
+			if (status === 429 && !planScoped) await this.coolInBandRateLimit(pool, accountId, failure.accountScoped, bodyText);
+			else if (failure.code === "INVALID_CREDENTIAL") await pool.markAuthFailed(accountId, failure.message).catch(() => void 0);
+			else if (failure.accountScoped) await pool.markCooldown(accountId, accountLimitCooldownMs(bodyText), `${PROVIDER_NAME$3} in-stream 403`).catch(() => void 0);
+			return reclassified;
+		}
+		if (reclassified.code === "RATE_LIMIT" && !matchesAny$1(bodyText, ENTITLEMENT_PATTERNS)) await this.coolInBandRateLimit(pool, accountId, classifyKimiFailure(429, bodyText).accountScoped, bodyText);
+		return reclassified;
+	}
+	/**
+	* Cool the account whose own rate limit the stream reported.
+	*
+	* The window this route already defaults to, never a Retry-After: there is no
+	* 429 response to read one from, and a delay stated on the 200 belongs to a
+	* different message. An account-scoped verdict takes its own window instead,
+	* because a spent balance is not back-pressure and 15 minutes would put the
+	* account back into rotation four times inside the window that refused it.
+	*/
+	async coolInBandRateLimit(pool, accountId, accountScoped, bodyText) {
+		const cooldownMs = accountScoped ? accountLimitCooldownMs(bodyText) : POOL_COOLDOWN_MS$2;
+		await pool.markCooldown(accountId, cooldownMs, `${PROVIDER_NAME$3} in-stream 429`).catch(() => void 0);
+	}
 };
 function processLine(line, state, wire) {
 	return wire === "anthropic" ? processAnthropicStreamLine(line, state) : processOpenAIStreamLine(line, state);
@@ -23937,6 +24447,26 @@ function outputConfigFor(modelId, requestedEffort) {
 	const effort = effortForModel(modelId, requestedEffort ?? null);
 	return effort === "default" ? void 0 : { effort };
 }
+/**
+* Raise a cap that is too small for this line's forced thinking.
+*
+* Left alone: a model that can disable thinking (its caller may have asked for
+* thinking off on purpose, and raising the cap would not enable it but a caller
+* reading the number back would be misled), and an absent cap (the caller stated
+* none, so \`maxOutputTokensFor\` owns the ceiling).
+*
+* @param modelId - the routed model id.
+* @param requested - the caller's cap, undefined when it stated none.
+* @returns the cap to request, raised only when that is the difference between an
+* empty truncated answer and a usable one.
+*/
+function floorForcedThinkingTokens(modelId, requested) {
+	if (requested === void 0) return void 0;
+	const model = catalogEntry(modelId);
+	const thinking = model?.thinking;
+	if (thinking !== "always-on" && thinking !== "forced-effort") return requested;
+	return Math.min(Math.max(requested, 512), Math.max(requested, model?.maxTokens ?? requested));
+}
 /** Output cap one request asks for, tracked against the model's declared ceiling. */
 function maxOutputTokensFor$1(modelId, contextWindow) {
 	const model = catalogEntry(modelId);
@@ -24455,6 +24985,7 @@ function processMinimaxStreamLine(line, state) {
 	}
 	if (type === "error") {
 		const error = isRecord$16(event.error) ? event.error : {};
+		state.streamError = 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;
@@ -27115,7 +27646,7 @@ var MinimaxCodeAdapter = class extends LlmAdapter {
 		}
 		let authOwner = computeMinimaxAuthOwner(credentials, accountId);
 		const withUploads = await uploadOversizedMedia(requestOptions, images, videos, credentials, accountId, this.options.fetchFn ?? fetch, signal);
-		const effectiveMax = clampOutputToContext(requestOptions.maxTokens ?? maxOutputTokensFor$1(options.model, contextWindow), contextWindow, estimatedInputTokens(requestOptions));
+		const effectiveMax = clampOutputToContext(floorForcedThinkingTokens(options.model, requestOptions.maxTokens) ?? maxOutputTokensFor$1(options.model, contextWindow), contextWindow, estimatedInputTokens(requestOptions));
 		const buildBodyForOwner = (owner) => {
 			return assertRequestBodyFits(buildMinimaxRequest({
 				...withUploads,
@@ -27175,23 +27706,46 @@ var MinimaxCodeAdapter = class extends LlmAdapter {
 			provider: PROVIDER_ID$2
 		});
 		let buffer = "";
+		/**
+		* Set the moment ANY chunk reaches the caller, and never cleared.
+		*
+		* Deliberately broader than "a content delta or a tool block": a retry would
+		* duplicate a block-start the caller has already seen just as surely, so the
+		* conservative reading is the one that cannot be wrong. It is what keeps the
+		* in-band reclassification below pre-output only.
+		*/
+		let outputStarted = false;
 		try {
-			while (true) {
-				const { done, value } = await reader.read();
-				if (done) break;
-				buffer += decoder.decode(value, { stream: true });
-				const lines = buffer.split("\n");
-				buffer = lines.pop() ?? "";
-				for (const line of lines) {
-					for (const chunk of processMinimaxStreamLine(line, state)) yield chunk;
-					if (state.finished) return;
+			try {
+				while (true) {
+					const { done, value } = await reader.read();
+					if (done) break;
+					buffer += decoder.decode(value, { stream: true });
+					const lines = buffer.split("\n");
+					buffer = lines.pop() ?? "";
+					for (const line of lines) {
+						for (const chunk of processMinimaxStreamLine(line, state)) {
+							outputStarted = true;
+							yield chunk;
+						}
+						if (state.finished) return;
+					}
+				}
+				buffer += decoder.decode();
+				if (buffer.trim() !== "") for (const line of buffer.split("\n")) for (const chunk of processMinimaxStreamLine(line, state)) {
+					outputStarted = true;
+					yield chunk;
+				}
+				if (state.finished) return;
+				assertStreamComplete(state);
+				for (const chunk of closeMinimaxStream(state)) {
+					outputStarted = true;
+					yield chunk;
 				}
+			} catch (error) {
+				if (error instanceof LlmError && !outputStarted && error.code === "PROVIDER_ERROR" && state.streamError !== void 0) throw reclassifyInBandError(error, state.streamError, classifyMinimaxFailure);
+				throw error;
 			}
-			buffer += decoder.decode();
-			if (buffer.trim() !== "") for (const line of buffer.split("\n")) for (const chunk of processMinimaxStreamLine(line, state)) yield chunk;
-			if (state.finished) return;
-			assertStreamComplete(state);
-			for (const chunk of closeMinimaxStream(state)) yield chunk;
 		} finally {
 			reader.cancel().catch(() => void 0);
 		}
@@ -36581,6 +37135,7 @@ 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);
+		state.streamError = error;
 		throw new LlmError$1("Claude stream error" + (kind === void 0 ? "" : " (" + kind + ")") + ": " + message, isContextOverflow(error) ? CONTEXT_OVERFLOW_CODE : "PROVIDER_ERROR");
 	}
 	return out;
@@ -37754,6 +38309,13 @@ function cancelLogin() {
 * - `request` -> PROVIDER_ERROR;
 * - `network` -> TRANSPORT.
 *
+* An in-band `error` event inside a 200 stream carries the same envelope and is
+* classified the same way while nothing has reached the caller; after the first
+* chunk it is surfaced as PROVIDER_ERROR, for the reason in note 1
+* ({@link inBandStreamVerdict}). A reclassified in-band failure is weighed by
+* {@link shouldRotateAccount} exactly as a non-2xx one is, so an account-scoped
+* limit reported inside the stream still takes that account out of rotation.
+*
 * A reported-client-version rejection is a REQUEST problem even though it
 * arrives as a 400 that mentions the client: it must NOT sign the user out, and
 * it has a real remedy (raise the reported version), so it is handled
@@ -38153,8 +38715,19 @@ var ClaudeAdapter = class ClaudeAdapter extends LlmAdapter {
 		* a later reordering of this method cannot silently drop it.
 		*/
 		let outputStarted = false;
+		/**
+		* The account the response now being streamed was issued with.
+		*
+		* Hoisted out of the loop because the in-stream failure path needs it: a
+		* rate limit reported inside a 200 stream belongs to the account that was
+		* asked, and that account is what the failure takes out of rotation — the
+		* harness is about to repeat this request, and repeating it against the same
+		* exhausted account fails identically.
+		*/
+		let usedAccountId;
 		while (true) {
 			const { credentials, accountId } = await this.resolveCredential(pool, tried, fetchFn, signal);
+			usedAccountId = accountId;
 			response = await this.attemptRequest(credentials, requestOptions, images, toolNames, thinking, settings, signal, fetchFn);
 			if (response.ok) break;
 			const detail = (await response.text().catch(() => "")).slice(0, 2e3);
@@ -38200,7 +38773,12 @@ var ClaudeAdapter = class ClaudeAdapter extends LlmAdapter {
 					yield chunk;
 				}
 			} catch (error) {
-				if (error instanceof LlmError) throw error;
+				if (error instanceof LlmError) {
+					if (outputStarted || error.code !== "PROVIDER_ERROR" || state.streamError === void 0) throw error;
+					const verdict = inBandStreamVerdict(error, state.streamError, response.headers);
+					if (verdict.failure !== null && shouldRotateAccount(verdict.failure, outputStarted) && usedAccountId !== void 0) await pool?.markCooldown(usedAccountId, cooldownMsFor(verdict.failure), "Claude（订阅） in-stream 429").catch(() => void 0);
+					throw verdict.error;
+				}
 				if (signal.aborted) throw new LlmError("Claude request aborted", "ABORTED", { cause: error });
 				throw new LlmError("Claude stream failed: " + (error instanceof Error ? error.message : String(error)), outputStarted ? "PROVIDER_ERROR" : "TRANSPORT", { cause: error });
 			}
@@ -38290,6 +38868,49 @@ function codeForFailure(failure) {
 	return "PROVIDER_ERROR";
 }
 /**
+* The DSH error an in-band `error` event becomes when it arrived BEFORE any
+* chunk reached the caller.
+*
+* Anthropic can answer 200 and then report the failure inside the stream, in
+* the same `{ error: { type, message } }` envelope a non-2xx body carries. The
+* mapper types every such event PROVIDER_ERROR, because it cannot know whether
+* output has started — and PROVIDER_ERROR is outside the retry set, so an
+* `overloaded_error` delivered this way ended the turn on the first try while
+* the identical overload delivered as a 529 is retried. Before any output a
+* fresh request is still free (module note 1), so the envelope is classified
+* exactly as a response body would be, and a transient verdict (overloaded,
+* server, rate limit) takes its retryable code.
+*
+* Everything else keeps the mapper's verdict: a request problem, a context
+* overflow, a credential refusal, and a type this line does not recognize. The
+* last matters: classifyFailure falls back to the status line for an unknown
+* type, and no status here describes the failure — the response itself was a
+* 200 — so the status passed is 0 and never consulted for a recognized type.
+*
+* The message stays the mapper's, which names the wire type the user saw.
+*
+* The classified failure rides along beside the error so the caller can weigh it
+* with {@link shouldRotateAccount} instead of re-deriving what it says. It is
+* null whenever the mapper's verdict stands, so a caller never has to guess
+* whether `accountScoped` is meaningful.
+*
+* @param thrown - the mapper's PROVIDER_ERROR (or context-overflow) verdict.
+* @param streamError - the event's wire `error` object, from the stream state.
+* @param headers - the 200 response's headers, read for the rate-limit verdict.
+*/
+function inBandStreamVerdict(thrown, streamError, headers) {
+	const failure = classifyFailure(0, JSON.stringify({ error: streamError }), headers);
+	if (failure.type === null || !failure.retryable) return {
+		error: thrown,
+		failure: null
+	};
+	const retryAfter = failure.retryAfterMs !== null && failure.retryAfterMs > 0 ? failure.retryAfterMs : void 0;
+	return {
+		error: new LlmError(thrown.message, codeForFailure(failure), retryAfter === void 0 ? {} : { providerRetryAfterMs: retryAfter }),
+		failure
+	};
+}
+/**
 * Turn one classified failure into the DSH error the caller sees.
 *
 * The facts the harness needs ride along: the HTTP status, and the delay the