@eddyskywalker/dsh-chatgpt-subscription 0.65.0 → 0.66.0

package/lib/index.js

@@ -809,7 +809,7 @@ const ROUTE_PREFIX$5 = "/api/dsh-chatgpt-subscription";
 * `test/responses-client.test.ts` asserts the two agree, so a release that
 * forgets this constant fails the suite instead of shipping quietly.
 */
-const PLUGIN_VERSION = "0.65.0";
+const PLUGIN_VERSION = "0.66.0";
 const CODEX_CHATGPT_PROVIDER_ID = "codex-chatgpt";
 const CODEX_API_BASE = "https://chatgpt.com/backend-api/codex";
 const CODEX_RESPONSES_URL = `${CODEX_API_BASE}/responses`;
@@ -3656,10 +3656,18 @@ var AccountPoolCore = class {
 		return true;
 	}
 	/** The account a rotation strategy picks out of an already-eligible set. */
-	selectAccount(eligible, data) {
+	selectAccount(eligible, data, hint) {
 		const pinned = this.hooks.preferAccountId?.() ?? null;
 		const pinnedAccount = pinned === null ? void 0 : eligible.find((account) => account.id === pinned);
 		if (pinnedAccount !== void 0) return pinnedAccount;
+		const preferred = hint?.preferAccountIds;
+		if (preferred !== void 0 && preferred.length > 0) {
+			const eligibleById = new Map(eligible.map((account) => [account.id, account]));
+			for (const id of preferred) {
+				const account = eligibleById.get(id);
+				if (account !== void 0) return account;
+			}
+		}
 		if (data.rotationStrategy === "round-robin") return [...eligible].sort((a, b) => (a.lastUsedAt || 0) - (b.lastUsedAt || 0))[0];
 		if (data.rotationStrategy === "sticky") return eligible.find((account) => account.id === data.activeAccountId) || eligible.find((account) => account.isPrimary) || eligible[0];
 		return eligible.find((account) => account.isPrimary) || eligible[0];
@@ -3712,15 +3720,18 @@ var AccountPoolCore = class {
 	* @param forceRefresh - renew even when the stored token still looks usable.
 	* @param options - what the refresh may spend, passed down to the provider's
 	*   refresh hook. Absent means the provider's own default budget.
+	* @param selection - this request's own account preference, when the caller
+	*   has one (a warm prefix cache, for instance). Absent means the rotation
+	*   strategy decides alone, exactly as before.
 	* @throws LlmError with `RATE_LIMIT` when every account is cooling down.
 	*/
-	async getEffectiveAccount(excludeIds, fetchFn = fetch, forceRefresh = false, options) {
+	async getEffectiveAccount(excludeIds, fetchFn = fetch, forceRefresh = false, options, selection) {
 		const data = await this.read();
 		if (data.accounts.length === 0) throw new Error(this.hooks.emptyMessage ?? t("pool.notSignedIn", { provider: this.hooks.displayName }));
 		const now = Date.now();
 		const eligible = data.accounts.filter((account) => this.isEligible(account, now, excludeIds));
 		if (eligible.length === 0) throw this.allUnavailableError(data, now);
-		const selected = this.selectAccount(eligible, data);
+		const selected = this.selectAccount(eligible, data, selection);
 		let refreshed;
 		try {
 			refreshed = await this.freshAccount(selected, fetchFn, forceRefresh, options);
@@ -3728,7 +3739,7 @@ var AccountPoolCore = class {
 			if (this.hooks.refreshFailureStatus?.(error) === void 0 && !(error instanceof PoolAccountUnavailableError)) throw error;
 			const nextExclude = new Set(excludeIds ?? []);
 			nextExclude.add(selected.id);
-			if (await this.hasAnotherAvailableAccount(nextExclude)) return this.getEffectiveAccount(nextExclude, fetchFn, forceRefresh, options);
+			if (await this.hasAnotherAvailableAccount(nextExclude)) return this.getEffectiveAccount(nextExclude, fetchFn, forceRefresh, options, selection);
 			throw error;
 		}
 		const account = await this.updatePool((current) => {
@@ -4477,7 +4488,7 @@ var OAuthService = class {
 			throw new OAuthServiceError("oauth-token-exchange-failed", `The ChatGPT device code could not be issued (${response.status}${detail === null ? "" : `, ${detail}`}).`, response.status);
 		}
 		const payload = await response.json().catch(() => null);
-		const record = isRecord$67(payload) ? payload : {};
+		const record = isRecord$70(payload) ? payload : {};
 		const authId = nonEmptyString$11(record.device_auth_id);
 		const userCode = nonEmptyString$11(record.user_code) ?? nonEmptyString$11(record.usercode);
 		if (authId === void 0 || userCode === void 0) throw new OAuthServiceError("oauth-token-exchange-failed", "ChatGPT returned no device code.");
@@ -4544,7 +4555,7 @@ var OAuthService = class {
 			const detail = deviceErrorIdentifier(payload);
 			throw new OAuthServiceError("oauth-token-exchange-failed", `The ChatGPT device confirmation was refused (${response.status}${detail === void 0 ? "" : `, ${detail}`}).`, response.status);
 		}
-		const record = isRecord$67(payload) ? payload : {};
+		const record = isRecord$70(payload) ? payload : {};
 		const authorizationCode = nonEmptyString$11(record.authorization_code);
 		const codeVerifier = nonEmptyString$11(record.code_verifier);
 		if (authorizationCode === void 0 || codeVerifier === void 0) throw new OAuthServiceError("oauth-token-exchange-failed", "ChatGPT confirmed the device code but returned no authorization code.");
@@ -4877,11 +4888,11 @@ const DEVICE_PENDING_STATUSES = /* @__PURE__ */ new Set([
 function isPendingDeviceAnswer(response, payload) {
 	if (response.status === 403 || response.status === 404) return true;
 	if (!response.ok && response.status !== 400) return false;
-	if (!isRecord$67(payload)) return false;
+	if (!isRecord$70(payload)) return false;
 	const status = payload.status;
 	if (typeof status === "string" && DEVICE_PENDING_STATUSES.has(status.toLowerCase())) return true;
 	const error = payload.error;
-	if (isRecord$67(error)) {
+	if (isRecord$70(error)) {
 		const code = error.code;
 		if (typeof code === "string" && DEVICE_PENDING_STATUSES.has(code.toLowerCase())) return true;
 	}
@@ -4895,18 +4906,18 @@ function isPendingDeviceAnswer(response, payload) {
 * wrong.
 */
 function deviceErrorIdentifier(payload) {
-	if (!isRecord$67(payload)) return void 0;
+	if (!isRecord$70(payload)) return void 0;
 	const candidates = [payload.code];
 	const error = payload.error;
 	if (typeof error === "string") candidates.push(error);
-	else if (isRecord$67(error)) candidates.push(error.code, error.type);
+	else if (isRecord$70(error)) candidates.push(error.code, error.type);
 	candidates.push(payload.status);
 	for (const candidate of candidates) if (typeof candidate === "string" && /^[a-z0-9_.-]{1,64}$/i.test(candidate)) return candidate;
 }
 function nonEmptyString$11(value) {
 	return typeof value === "string" && value !== "" ? value : void 0;
 }
-function isRecord$67(value) {
+function isRecord$70(value) {
 	return typeof value === "object" && value !== null && !Array.isArray(value);
 }
 async function oauthErrorIdentifier(response) {
@@ -5087,7 +5098,7 @@ const CODEX_CLI_CREDENTIAL_SOURCE = "codex";
 * here, and the import has to be repeated afterwards.
 */
 const ADOPTED_CODEX_CREDENTIAL_EXPIRED_HINT = "The sign-in imported from the Codex CLI has expired. Sign in again in the Codex CLI, then import it again from this settings card.";
-function isRecord$66(value) {
+function isRecord$69(value) {
 	return typeof value === "object" && value !== null && !Array.isArray(value);
 }
 /** A trimmed, non-empty string, or nothing. */
@@ -5189,7 +5200,7 @@ function decodeJwtClaims$1(token) {
 	if (parts.length !== 3) return void 0;
 	try {
 		const value = JSON.parse(Buffer.from(parts[1], "base64url").toString("utf8"));
-		return isRecord$66(value) ? value : void 0;
+		return isRecord$69(value) ? value : void 0;
 	} catch {
 		return;
 	}
@@ -5234,10 +5245,10 @@ function expiryFromAccessToken(accessToken) {
 * the file carries cannot reach the result.
 */
 function parseAdoptedCodexCredential(value, sourcePath) {
-	if (!isRecord$66(value)) return void 0;
+	if (!isRecord$69(value)) return void 0;
 	if (value.auth_mode !== "chatgpt") return void 0;
 	const tokens = value[CODEX_CLI_TOKENS_KEY];
-	if (!isRecord$66(tokens)) return void 0;
+	if (!isRecord$69(tokens)) return void 0;
 	const accessToken = nonEmptyString$10(tokens.access_token);
 	if (accessToken === void 0) return void 0;
 	const refreshToken = nonEmptyString$10(tokens.refresh_token);
@@ -5256,7 +5267,7 @@ function parseAdoptedCodexCredential(value, sourcePath) {
 		const email = claims === void 0 ? void 0 : nonEmptyString$10(claims.email);
 		if (email !== void 0) credentials.email = email;
 		const authClaims = claims === void 0 ? void 0 : claims[CODEX_AUTH_CLAIMS_KEY];
-		const planType = isRecord$66(authClaims) ? nonEmptyString$10(authClaims.chatgpt_plan_type) : void 0;
+		const planType = isRecord$69(authClaims) ? nonEmptyString$10(authClaims.chatgpt_plan_type) : void 0;
 		if (planType !== void 0) credentials.planType = planType;
 	}
 	const accountId = nonEmptyString$10(tokens.account_id);
@@ -8149,7 +8160,7 @@ const schemaFields = /* @__PURE__ */ new Set([
 	"required",
 	"propertyOrdering"
 ]);
-function isRecord$65(value) {
+function isRecord$68(value) {
 	return typeof value === "object" && value !== null && !Array.isArray(value);
 }
 function mergeSchemas(left, right) {
@@ -8157,7 +8168,7 @@ function mergeSchemas(left, right) {
 		...left,
 		...right
 	};
-	if (isRecord$65(left.properties) && isRecord$65(right.properties)) merged.properties = {
+	if (isRecord$68(left.properties) && isRecord$68(right.properties)) merged.properties = {
 		...left.properties,
 		...right.properties
 	};
@@ -8193,9 +8204,9 @@ 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$65(branch.properties) ? "object" : isRecord$65(branch.items) ? "array" : "value";
+	const type = typeof branch.type === "string" ? branch.type : isRecord$68(branch.properties) ? "object" : isRecord$68(branch.items) ? "array" : "value";
 	const details = [];
-	if (isRecord$65(branch.properties)) {
+	if (isRecord$68(branch.properties)) {
 		const names = Object.keys(branch.properties);
 		if (names.length > 0) details.push(`{${names.join(",")}}`);
 	}
@@ -8209,7 +8220,7 @@ function describeBranch(branch) {
 }
 /** How well one branch can stand in for a whole union; a named object shape wins. */
 function branchRank(branch) {
-	if (isRecord$65(branch.properties)) return 3;
+	if (isRecord$68(branch.properties)) return 3;
 	if (branch.type === "object") return 2;
 	if (branch.type === "array") return 1;
 	return 0;
@@ -8238,7 +8249,7 @@ function unionShape(branches) {
 		shape: branches[0],
 		folded: false
 	};
-	if (new Set(branches.map((branch) => branch.type)).size === 1 && branches.every((branch) => !isRecord$65(branch.properties) && !isRecord$65(branch.items))) {
+	if (new Set(branches.map((branch) => branch.type)).size === 1 && branches.every((branch) => !isRecord$68(branch.properties) && !isRecord$68(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))];
@@ -8275,7 +8286,7 @@ 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$65(value)) {
+		if (key === "properties" && isRecord$68(value)) {
 			out.properties = Object.fromEntries(Object.entries(value).map(([name, child]) => [name, foldSchemaUnions(child)]));
 			continue;
 		}
@@ -8304,15 +8315,15 @@ function foldOwnFields(node) {
 * argument the model sends anyway fails there, naming the actual problem.
 */
 function foldSchemaUnions(node) {
-	if (!isRecord$65(node)) return node;
+	if (!isRecord$68(node)) return node;
 	let own = foldOwnFields(node);
 	if (Array.isArray(node.allOf)) for (const branch of node.allOf) {
 		const folded = foldSchemaUnions(branch);
-		if (isRecord$65(folded)) own = mergeSchemas(own, folded);
+		if (isRecord$68(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$65);
+	const branches = alternatives.map((branch) => foldSchemaUnions(branch)).filter(isRecord$68);
 	if (branches.length === 0) return own;
 	const shapes = branches.filter((branch) => branch.type !== "null");
 	const acceptsNull = branches.some((branch) => branch.type === "null");
@@ -8328,7 +8339,7 @@ function foldSchemaUnions(node) {
 	return merged;
 }
 function toAntigravityToolSchema(schema, options = {}) {
-	if (!isRecord$65(schema)) return schema;
+	if (!isRecord$68(schema)) return schema;
 	const root = schema;
 	function resolveReference(reference) {
 		let target = root;
@@ -8336,13 +8347,13 @@ 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$65(target) && Object.hasOwn(target, key) ? target[key] : void 0;
+			target = isRecord$68(target) && Object.hasOwn(target, key) ? target[key] : void 0;
 		}
-		if (!isRecord$65(target)) throw new LlmError(`Antigravity tool schema reference was not found: ${reference}`, "PROVIDER_ERROR");
+		if (!isRecord$68(target)) throw new LlmError(`Antigravity tool schema reference was not found: ${reference}`, "PROVIDER_ERROR");
 		return target;
 	}
 	function convert(node, references) {
-		if (!isRecord$65(node)) return {};
+		if (!isRecord$68(node)) return {};
 		let inherited = {};
 		if (typeof node.$ref === "string") {
 			if (references.has(node.$ref)) throw new LlmError(`Antigravity tool schema contains a recursive reference: ${node.$ref}`, "PROVIDER_ERROR");
@@ -8351,8 +8362,8 @@ 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$65(node.properties)) out.properties = Object.fromEntries(Object.entries(node.properties).map(([name, child]) => [name, convert(child, references)]));
-		if (isRecord$65(node.items)) out.items = convert(node.items, references);
+		if (isRecord$68(node.properties)) out.properties = Object.fromEntries(Object.entries(node.properties).map(([name, child]) => [name, convert(child, references)]));
+		if (isRecord$68(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)) {
@@ -8388,7 +8399,7 @@ let toolCallCounter = 0;
 function sanitizeText$9(text) {
 	return text.replace(/\0/g, "");
 }
-function isRecord$64(value) {
+function isRecord$67(value) {
 	return typeof value === "object" && value !== null && !Array.isArray(value);
 }
 function asString$30(value) {
@@ -8408,25 +8419,25 @@ 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$64(raw)) return raw;
+	if (isRecord$67(raw)) return raw;
 	if (raw === void 0 || raw === null || raw === "") return {};
 	const parsed = typeof raw === "string" ? safeJsonParse$9(raw) : raw;
-	return isRecord$64(parsed) ? parsed : {};
+	return isRecord$67(parsed) ? parsed : {};
 }
 const NO_RESOLVED_IMAGES$5 = /* @__PURE__ */ new Map();
 function attachmentOf$5(block) {
 	const attachment = block.attachment;
-	if (!isRecord$64(attachment)) return void 0;
+	if (!isRecord$67(attachment)) return void 0;
 	return typeof attachment.attachmentId === "string" ? attachment : void 0;
 }
 function attachmentLabel$6(block) {
-	const attachment = isRecord$64(block.attachment) ? block.attachment : void 0;
+	const attachment = isRecord$67(block.attachment) ? block.attachment : void 0;
 	return asString$30(attachment?.name) || asString$30(attachment?.attachmentId);
 }
 function collectImageRefs$3(content, refs) {
 	if (!Array.isArray(content)) return;
 	for (const block of content) {
-		if (!isRecord$64(block)) continue;
+		if (!isRecord$67(block)) continue;
 		if (block.type === "image") {
 			const attachment = attachmentOf$5(block);
 			if (attachment) refs.set(attachment.attachmentId, attachment);
@@ -8559,7 +8570,7 @@ function unavailableImageText$6(block) {
 }
 function imageBlockToPart(block, images) {
 	let data = asString$30(block.data) || asString$30(block.base64);
-	const source = isRecord$64(block.source) ? block.source : void 0;
+	const source = isRecord$67(block.source) ? block.source : void 0;
 	if (!data && source) data = asString$30(source.data) || asString$30(source.base64);
 	let mimeType = asString$30(block.mimeType) || asString$30(block.mediaType) || (source ? asString$30(source.mimeType) || asString$30(source.mediaType) : void 0) || "image/png";
 	if (data?.startsWith("data:")) {
@@ -8584,8 +8595,8 @@ function contentToUserParts(content, images) {
 	if (typeof content === "string") return [{ text: sanitizeText$9(content) }];
 	if (!Array.isArray(content)) return [];
 	const parts = [];
-	for (const block of content) if (isRecord$64(block) && block.type === "text" && typeof block.text === "string") parts.push({ text: sanitizeText$9(block.text) });
-	else if (isRecord$64(block) && block.type === "image") {
+	for (const block of content) if (isRecord$67(block) && block.type === "text" && typeof block.text === "string") parts.push({ text: sanitizeText$9(block.text) });
+	else if (isRecord$67(block) && block.type === "image") {
 		const img = imageBlockToPart(block, images);
 		parts.push(img ?? { text: unavailableImageText$6(block) });
 	}
@@ -8600,7 +8611,7 @@ function toolResultImageParts(blocks, images) {
 	if (!Array.isArray(blocks)) return [];
 	const parts = [];
 	for (const block of blocks) {
-		if (!isRecord$64(block)) continue;
+		if (!isRecord$67(block)) continue;
 		if (block.type === "image") {
 			parts.push(imageBlockToPart(block, images) ?? { text: unavailableImageText$6(block) });
 			continue;
@@ -8612,7 +8623,7 @@ function toolResultImageParts(blocks, images) {
 function toolResultText$9(blocks) {
 	if (!Array.isArray(blocks)) return "";
 	return blocks.map((block) => {
-		if (!isRecord$64(block)) return "";
+		if (!isRecord$67(block)) return "";
 		if (block.type === "text" && typeof block.text === "string") return sanitizeText$9(block.text);
 		if (block.type === "tool-result") return toolResultText$9(block.content);
 		if (block.type === "image") {
@@ -8626,9 +8637,9 @@ 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$64(state)) return void 0;
+	if (!isRecord$67(state)) return void 0;
 	if (Array.isArray(state.blocks)) return state.blocks[index];
-	const resp = isRecord$64(state.response) ? state.response : void 0;
+	const resp = isRecord$67(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];
@@ -8730,10 +8741,10 @@ 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$64(block)) continue;
+		if (!isRecord$67(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$64) : [];
+		const originalParts = Array.isArray(replay?.parts) ? replay.parts.filter(isRecord$67) : [];
 		if ((block.type === "text" || block.type === "reasoning") && originalParts.length > 0 && originalParts.every((part) => !part.functionCall) && originalParts.map((part) => asString$30(part.text) || "").join("") === sanitizeText$9(String(block.text || ""))) {
 			parts.push(...originalParts.map(replayPart));
 			continue;
@@ -8754,8 +8765,8 @@ 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$64(part.functionCall));
-			const wireId = asString$30((isRecord$64(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$67(part.functionCall));
+			const wireId = asString$30((isRecord$67(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
@@ -8807,7 +8818,7 @@ 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$64(b) || b.type !== "tool-result"), images);
+		const userParts = contentToUserParts(content.filter((b) => !isRecord$67(b) || b.type !== "tool-result"), images);
 		if (role === "system") {
 			if (userParts.length) contents.push({
 				role: GEMINI_ROLE.user,
@@ -8819,7 +8830,7 @@ function convertMessages(options, model, runtimeModel, images = NO_RESOLVED_IMAG
 			role: GEMINI_ROLE.user,
 			parts: userParts
 		});
-		for (const b of content) if (isRecord$64(b) && b.type === "tool-result") pushToolResult(contents, b, toolCalls, model, runtimeModel, images);
+		for (const b of content) if (isRecord$67(b) && b.type === "tool-result") pushToolResult(contents, b, toolCalls, model, runtimeModel, images);
 	}
 	return contents;
 }
@@ -8962,7 +8973,7 @@ const USAGE_FIELDS = [
 	"totalTokenCount"
 ];
 function collectUsage(value, state) {
-	if (!isRecord$64(value)) return;
+	if (!isRecord$67(value)) return;
 	for (const key of USAGE_FIELDS) {
 		const count = value[key];
 		if (typeof count !== "number" || !Number.isSafeInteger(count) || count < 0) continue;
@@ -8992,17 +9003,17 @@ function processStreamLine$4(line, state) {
 	}
 	if (!json) return [];
 	const chunk = safeJsonParse$9(json);
-	if (!isRecord$64(chunk)) return [];
-	const responseData = isRecord$64(chunk.response) ? chunk.response : chunk;
-	const error = isRecord$64(chunk.error) ? chunk.error : isRecord$64(responseData.error) ? responseData.error : void 0;
+	if (!isRecord$67(chunk)) return [];
+	const responseData = isRecord$67(chunk.response) ? chunk.response : chunk;
+	const error = isRecord$67(chunk.error) ? chunk.error : isRecord$67(responseData.error) ? responseData.error : void 0;
 	if (error !== void 0) throw new LlmError$1(`Antigravity stream error: ${asString$30(error.message) ?? "unknown error"}`, isContextOverflow(error) ? CONTEXT_OVERFLOW_CODE : "PROVIDER_ERROR");
 	const candidates = Array.isArray(responseData.candidates) ? responseData.candidates : [];
-	const candidate = isRecord$64(candidates[0]) ? candidates[0] : void 0;
-	const content = isRecord$64(candidate?.content) ? candidate.content : void 0;
+	const candidate = isRecord$67(candidates[0]) ? candidates[0] : void 0;
+	const content = isRecord$67(candidate?.content) ? candidate.content : void 0;
 	const parts = Array.isArray(content?.parts) ? content.parts : [];
 	const out = [];
 	for (const part of parts) {
-		if (!isRecord$64(part)) continue;
+		if (!isRecord$67(part)) continue;
 		if (typeof part.text === "string" && part.text !== "") {
 			const isThinking = Boolean(part.thought);
 			const blockType = isThinking ? "reasoning" : "text";
@@ -9034,7 +9045,7 @@ function processStreamLine$4(line, state) {
 				index: state.currentBlock.index,
 				text: delta
 			});
-		} else if (!isRecord$64(part.functionCall) && thoughtSignature(part)) {
+		} else if (!isRecord$67(part.functionCall) && thoughtSignature(part)) {
 			if (state.replayBlocks.length === 0) {
 				const type = part.thought ? "reasoning" : "text";
 				state.blocks.push({
@@ -9058,12 +9069,12 @@ function processStreamLine$4(line, state) {
 			}
 			state.replayBlocks[state.replayBlocks.length - 1].parts.push(replayPart(part));
 		}
-		if (isRecord$64(part.functionCall)) {
+		if (isRecord$67(part.functionCall)) {
 			out.push(...closeCurrentBlock(state));
 			const fc = part.functionCall;
 			const toolName = asString$30(fc.name) || "";
 			const toolId = asString$30(fc.id) || sanitizeToolCallId("", toolName);
-			const argsText = JSON.stringify(isRecord$64(fc.args) ? fc.args : {});
+			const argsText = JSON.stringify(isRecord$67(fc.args) ? fc.args : {});
 			const index = state.blocks.length;
 			const block = {
 				type: "tool-call",
@@ -9102,7 +9113,7 @@ function processStreamLine$4(line, state) {
 	if (responseData !== chunk) collectUsage(responseData.usageMetadata, state);
 	const finishMessage = asString$30(candidate?.finishMessage);
 	if (finishMessage) state.finishMessage = finishMessage;
-	const blockReason = asString$30((isRecord$64(responseData.promptFeedback) ? responseData.promptFeedback : void 0)?.blockReason);
+	const blockReason = asString$30((isRecord$67(responseData.promptFeedback) ? responseData.promptFeedback : void 0)?.blockReason);
 	const finishReason = asString$30(candidate?.finishReason) || asString$30(responseData.finishReason) || (blockReason ? `PROMPT_BLOCKED:${blockReason}` : void 0);
 	if (finishReason) {
 		state.finishReason = finishReason;
@@ -9175,7 +9186,7 @@ function closeStream$7(state) {
 *
 * @module dsh-chatgpt-subscription/request-images
 */
-function isRecord$63(value) {
+function isRecord$66(value) {
 	return typeof value === "object" && value !== null && !Array.isArray(value);
 }
 function asString$29(value) {
@@ -9234,17 +9245,17 @@ const SUPPORTED_IMAGE_MEDIA_TYPES$4 = /* @__PURE__ */ new Set([
 ]);
 function attachmentOf$4(block) {
 	const attachment = block.attachment;
-	if (!isRecord$63(attachment)) return void 0;
+	if (!isRecord$66(attachment)) return void 0;
 	return typeof attachment.attachmentId === "string" ? attachment : void 0;
 }
 function attachmentLabel$5(block) {
-	const attachment = isRecord$63(block.attachment) ? block.attachment : void 0;
+	const attachment = isRecord$66(block.attachment) ? block.attachment : void 0;
 	return asString$29(attachment?.name) || asString$29(attachment?.attachmentId);
 }
 function collectImageRefs$2(content, refs) {
 	if (!Array.isArray(content)) return;
 	for (const block of content) {
-		if (!isRecord$63(block)) continue;
+		if (!isRecord$66(block)) continue;
 		if (block.type === "image") {
 			const attachment = attachmentOf$4(block);
 			if (attachment) refs.set(attachment.attachmentId, attachment);
@@ -9274,7 +9285,7 @@ function base64Length$2(bytes) {
 */
 function inlinePayloadOf$1(block) {
 	let data = asString$29(block.data) || asString$29(block.base64);
-	const source = isRecord$63(block.source) ? block.source : void 0;
+	const source = isRecord$66(block.source) ? block.source : void 0;
 	if (!data && source) data = asString$29(source.data) || asString$29(source.base64);
 	if (data?.startsWith("data:")) {
 		const matched = data.match(/^data:([^;,]+);base64,(.*)$/s);
@@ -9304,7 +9315,7 @@ function requestImagePayloadLength(block, images) {
 function collectRequestImagePayloads(content, images, lengths) {
 	if (!Array.isArray(content)) return;
 	for (const block of content) {
-		if (!isRecord$63(block)) continue;
+		if (!isRecord$66(block)) continue;
 		if (block.type === "image") {
 			const bytes = requestImagePayloadLength(block, images);
 			if (bytes !== void 0) lengths.push(bytes);
@@ -9358,7 +9369,7 @@ function replaceOldestRequestImages(options, count, images) {
 		let replaced = false;
 		return {
 			content: content.map((block) => {
-				if (remaining.count === 0 || !isRecord$63(block)) return block;
+				if (remaining.count === 0 || !isRecord$66(block)) return block;
 				if (block.type === "image") {
 					if (requestImagePayloadLength(block, images) === void 0) return block;
 					remaining.count -= 1;
@@ -9500,7 +9511,7 @@ function collectImageOccurrences(options, images) {
 	const walk = (content) => {
 		if (!Array.isArray(content)) return;
 		for (const block of content) {
-			if (!isRecord$63(block)) continue;
+			if (!isRecord$66(block)) continue;
 			if (block.type === "image") {
 				const bytes = requestImagePayloadLength(block, images);
 				const key = requestImageKey(block, images);
@@ -9554,7 +9565,7 @@ function replaceDroppedRequestImages(options, droppedKeys, images) {
 		let replaced = false;
 		return {
 			content: content.map((block) => {
-				if (!isRecord$63(block)) return block;
+				if (!isRecord$66(block)) return block;
 				if (block.type === "image") {
 					const key = requestImageKey(block, images);
 					if (key === void 0 || !droppedKeys.has(key)) return block;
@@ -9739,7 +9750,7 @@ function unavailableImageText$5(block) {
 }
 function imageBlockToInline$3(block, images) {
 	let data = asString$29(block.data) || asString$29(block.base64);
-	const source = isRecord$63(block.source) ? block.source : void 0;
+	const source = isRecord$66(block.source) ? block.source : void 0;
 	if (!data && source) data = asString$29(source.data) || asString$29(source.base64);
 	let mediaType = asString$29(block.mimeType) || asString$29(block.mediaType) || (source ? asString$29(source.mimeType) || asString$29(source.mediaType) : void 0) || "image/png";
 	if (data?.startsWith("data:")) {
@@ -9761,6 +9772,549 @@ function imageBlockToInline$3(block, images) {
 	} : void 0;
 }
 //#endregion
+//#region src/host/common/prompt-cache-policy.ts
+const MINUTE_MS = 6e4;
+const MAX_CACHE_TTL_MS = 720 * 60 * MINUTE_MS;
+const DISABLED = Object.freeze({
+	mode: "none",
+	scope: "unknown",
+	staticPrefix: false,
+	coldStart: "off",
+	stickyImages: false,
+	affinity: false,
+	waitWarmup: false,
+	hint: false
+});
+function policy(input) {
+	return Object.freeze(input);
+}
+/** Built-in facts already represented by a line's mapper, adapter, or official client. */
+const PROMPT_CACHE_POLICIES = Object.freeze({
+	"kimi-code": policy({
+		mode: "automatic",
+		scope: "account",
+		ttlMs: 25 * MINUTE_MS,
+		staticPrefix: true,
+		coldStart: "warm",
+		stickyImages: true,
+		affinity: true,
+		waitWarmup: true,
+		hint: true
+	}),
+	"codex-chatgpt": policy({
+		mode: "automatic",
+		scope: "unknown",
+		staticPrefix: true,
+		coldStart: "in-flight",
+		stickyImages: false,
+		affinity: false,
+		waitWarmup: false,
+		hint: false
+	}),
+	antigravity: policy({
+		mode: "automatic",
+		scope: "unknown",
+		staticPrefix: true,
+		coldStart: "in-flight",
+		stickyImages: false,
+		affinity: false,
+		waitWarmup: false,
+		hint: false
+	}),
+	"step-code": policy({
+		mode: "automatic",
+		scope: "unknown",
+		ttlMs: 5 * MINUTE_MS,
+		staticPrefix: true,
+		coldStart: "in-flight",
+		stickyImages: false,
+		affinity: false,
+		waitWarmup: false,
+		hint: true
+	}),
+	claude: policy({
+		mode: "explicit",
+		scope: "unknown",
+		staticPrefix: true,
+		coldStart: "in-flight",
+		stickyImages: true,
+		affinity: false,
+		waitWarmup: false,
+		hint: false
+	}),
+	"minimax-code": policy({
+		mode: "explicit",
+		scope: "unknown",
+		staticPrefix: true,
+		coldStart: "in-flight",
+		stickyImages: true,
+		affinity: false,
+		waitWarmup: false,
+		hint: false
+	}),
+	"glm-coding-plan": policy({
+		...DISABLED,
+		mode: "automatic"
+	}),
+	"command-code": policy({
+		...DISABLED,
+		mode: "automatic"
+	}),
+	"cline-pass": policy({
+		...DISABLED,
+		mode: "automatic"
+	}),
+	qoder: policy({
+		...DISABLED,
+		mode: "automatic"
+	}),
+	opendesign: policy({
+		...DISABLED,
+		mode: "automatic"
+	}),
+	"workbuddy-subscription": policy({
+		...DISABLED,
+		mode: "automatic"
+	}),
+	ollama: DISABLED
+});
+const OVERRIDE_KEYS = /* @__PURE__ */ new Set([
+	"mode",
+	"scope",
+	"ttlMs",
+	"staticPrefix",
+	"coldStart",
+	"stickyImages",
+	"affinity",
+	"waitWarmup",
+	"hint"
+]);
+const BOOLEAN_KEYS = [
+	"staticPrefix",
+	"stickyImages",
+	"affinity",
+	"waitWarmup",
+	"hint"
+];
+function isRecord$65(value) {
+	return typeof value === "object" && value !== null && !Array.isArray(value);
+}
+/** Validate one provider override without silently dropping unknown keys. */
+function parsePromptCacheOverride(provider, value) {
+	if (!isRecord$65(value)) throw new Error(`prompt cache policy for "${provider}" must be an object`);
+	const out = {};
+	for (const [key, entry] of Object.entries(value)) {
+		if (!OVERRIDE_KEYS.has(key)) throw new Error(`prompt cache policy for "${provider}" has unknown field "${key}"`);
+		if (entry === void 0) continue;
+		if (key === "mode") {
+			if (entry !== "automatic" && entry !== "explicit" && entry !== "none") throw new Error(`invalid prompt cache mode for "${provider}"`);
+		} else if (key === "scope") {
+			if (entry !== "account" && entry !== "shared" && entry !== "unknown") throw new Error(`invalid prompt cache scope for "${provider}"`);
+		} else if (key === "coldStart") {
+			if (entry !== "off" && entry !== "in-flight" && entry !== "warm") throw new Error(`invalid cold-start mode for "${provider}"`);
+		} else if (key === "ttlMs") {
+			if (!Number.isSafeInteger(entry) || entry <= 0 || entry > MAX_CACHE_TTL_MS) throw new Error(`prompt cache ttlMs for "${provider}" must be a positive safe integer no greater than ${MAX_CACHE_TTL_MS}`);
+		} else if (BOOLEAN_KEYS.includes(key) && typeof entry !== "boolean") throw new Error(`prompt cache ${key} for "${provider}" must be boolean`);
+		out[key] = entry;
+	}
+	return out;
+}
+/** Validate a whole config-owned map. Undefined means no overrides. */
+function parsePromptCacheOverrides(value) {
+	if (value === void 0) return {};
+	if (!isRecord$65(value)) throw new Error("promptCacheOverrides must be an object keyed by provider id");
+	const out = {};
+	for (const [provider, override] of Object.entries(value)) {
+		if (!/^[a-z0-9][a-z0-9-]*$/.test(provider)) throw new Error(`invalid prompt cache provider id "${provider}"`);
+		out[provider] = parsePromptCacheOverride(provider, override);
+	}
+	return Object.freeze(out);
+}
+function validateResolved(provider, value) {
+	if (value.mode === "none") return policy({ ...DISABLED });
+	if ((value.coldStart === "warm" || value.waitWarmup || value.hint) && value.ttlMs === void 0) throw new Error(`prompt cache policy for "${provider}" needs a measured ttlMs for warm/waitWarmup/hint`);
+	if (value.affinity && value.scope !== "account") throw new Error(`prompt cache affinity for "${provider}" needs scope:"account"`);
+	return policy(value);
+}
+/** Resolve one line, applying an already-validated or config-shaped override. */
+function promptCachePolicyOf(provider, override) {
+	const base = PROMPT_CACHE_POLICIES[provider] ?? DISABLED;
+	const parsed = override === void 0 ? void 0 : parsePromptCacheOverride(provider, override);
+	return validateResolved(provider, {
+		...base,
+		...parsed
+	});
+}
+//#endregion
+//#region src/host/common/prefix-request-coordinator.ts
+const NOOP_LEASE = Object.freeze({
+	leader: false,
+	release: () => void 0,
+	markWarm: () => void 0
+});
+var PrefixRequestCoordinator = class {
+	leaders = /* @__PURE__ */ new Map();
+	bursts = /* @__PURE__ */ new Map();
+	mainHeaders = /* @__PURE__ */ new Map();
+	warm = /* @__PURE__ */ new Map();
+	waitMs;
+	maxGroups;
+	now;
+	constructor(options = {}) {
+		this.waitMs = options.waitMs ?? 2500;
+		this.maxGroups = options.maxGroups ?? 256;
+		this.now = options.now ?? Date.now;
+	}
+	/** Join or lead one cold prefix group. */
+	async acquire(groupKey, signal, warmTtlMs) {
+		if (groupKey === void 0 || groupKey === "") return NOOP_LEASE;
+		if (warmTtlMs !== void 0 && this.isWarm(groupKey, warmTtlMs)) return NOOP_LEASE;
+		const existing = this.leaders.get(groupKey);
+		if (existing !== void 0 && this.now() - existing.startedAt < this.waitMs) {
+			await this.race(existing.promise, signal);
+			return NOOP_LEASE;
+		}
+		if (existing !== void 0) existing.release();
+		return this.open(this.leaders, groupKey, warmTtlMs);
+	}
+	/** Serialize first-turn bursts on one account/model without a warm cache claim. */
+	async acquireBurst(key, signal) {
+		if (key === "") return NOOP_LEASE;
+		while (true) {
+			const existing = this.bursts.get(key);
+			if (existing === void 0) break;
+			await this.race(existing.promise, signal);
+			if (signal.aborted) return NOOP_LEASE;
+			if (this.bursts.get(key) === existing) existing.release();
+		}
+		return this.open(this.bursts, key);
+	}
+	/** A main request owns the prefill slot until its response headers arrive. */
+	registerMainSession(sessionId) {
+		if (sessionId.trim() === "") return NOOP_LEASE;
+		const key = sessionId.trim();
+		this.mainHeaders.get(key)?.release();
+		return this.open(this.mainHeaders, key);
+	}
+	/** An auxiliary title request waits briefly for the main request's headers. */
+	async waitForMainSession(sessionId, signal) {
+		const existing = this.mainHeaders.get(sessionId.trim());
+		if (existing !== void 0) await this.race(existing.promise, signal);
+	}
+	isWarm(groupKey, ttlMs, at = this.now()) {
+		const warmedAt = this.warm.get(groupKey);
+		return warmedAt !== void 0 && Number.isFinite(ttlMs) && ttlMs > 0 && at - warmedAt < ttlMs;
+	}
+	/** Release every waiter and clear all process-local evidence. */
+	reset() {
+		for (const store of [
+			this.leaders,
+			this.bursts,
+			this.mainHeaders
+		]) {
+			for (const entry of [...store.values()]) entry.release();
+			store.clear();
+		}
+		this.warm.clear();
+	}
+	open(store, key, warmTtlMs) {
+		let resolve;
+		const promise = new Promise((done) => {
+			resolve = done;
+		});
+		let settled = false;
+		const release = () => {
+			if (settled) return;
+			settled = true;
+			resolve();
+			if (store.get(key) === entry) store.delete(key);
+		};
+		const entry = {
+			promise,
+			startedAt: this.now(),
+			release
+		};
+		store.set(key, entry);
+		return {
+			leader: true,
+			release,
+			markWarm: () => {
+				if (warmTtlMs === void 0 || !Number.isFinite(warmTtlMs) || warmTtlMs <= 0) return;
+				this.warm.delete(key);
+				this.warm.set(key, this.now());
+				while (this.warm.size > this.maxGroups) {
+					const oldest = this.warm.keys().next();
+					if (oldest.done !== true) this.warm.delete(oldest.value);
+				}
+			}
+		};
+	}
+	/** Wait for a leader, the bounded budget, or caller cancellation; clean every listener. */
+	async race(leader, signal) {
+		const cleanups = [];
+		const budget = new Promise((resolve) => {
+			const timer = setTimeout(resolve, this.waitMs);
+			timer.unref?.();
+			cleanups.push(() => clearTimeout(timer));
+		});
+		const aborted = new Promise((resolve) => {
+			if (signal.aborted) {
+				resolve();
+				return;
+			}
+			const onAbort = () => resolve();
+			signal.addEventListener("abort", onAbort, { once: true });
+			cleanups.push(() => signal.removeEventListener("abort", onAbort));
+		});
+		await Promise.race([
+			leader,
+			budget,
+			aborted
+		]).finally(() => {
+			for (const cleanup of cleanups) cleanup();
+		});
+	}
+};
+/** Cheap structural gate for auxiliary calls; provider-specific legacy probes may extend it. */
+function isAuxiliaryRequest(options) {
+	return options.purpose !== void 0;
+}
+/** Coordinate one fetch through response headers, releasing on every outcome. */
+async function fetchWithPrefixCoordination(fetchFn, coordinator, groupKey, warmTtlMs, input, init) {
+	const signal = init?.signal;
+	if (signal == null) {
+		const controller = new AbortController();
+		const lease = await coordinator.acquire(groupKey, controller.signal, warmTtlMs);
+		try {
+			const response = await fetchFn(input, init);
+			if (response.ok) lease.markWarm();
+			return response;
+		} finally {
+			lease.release();
+		}
+	}
+	const lease = await coordinator.acquire(groupKey, signal, warmTtlMs);
+	try {
+		const response = await fetchFn(input, init);
+		if (response.ok) lease.markWarm();
+		return response;
+	} finally {
+		lease.release();
+	}
+}
+//#endregion
+//#region src/host/common/prefix-stability.ts
+function isRecord$64(value) {
+	return typeof value === "object" && value !== null && !Array.isArray(value);
+}
+function firstTextOf$1(content) {
+	if (typeof content === "string") return content.trim() === "" ? void 0 : content;
+	if (!Array.isArray(content)) return void 0;
+	for (const block of content) {
+		if (!isRecord$64(block)) continue;
+		if (block.type === "text" && typeof block.text === "string" && block.text !== "") return block.text;
+		if (block.type === "tool-result") {
+			const nested = firstTextOf$1(block.content);
+			if (nested !== void 0) return nested;
+		}
+	}
+}
+function isToolResultMessage$9(message) {
+	return message?.source?.kind === "tool";
+}
+/** Stability tier for an initial bootstrap reminder; null means dynamic/history-bound. */
+function classifyBootstrapUserMessage$1(message) {
+	if (message.role !== "user" || isToolResultMessage$9(message)) return null;
+	const source = isRecord$64(message.source) ? message.source : void 0;
+	const kind = source !== void 0 && typeof source.kind === "string" ? source.kind : void 0;
+	const text = firstTextOf$1(message.content)?.trimStart() ?? "";
+	if (kind === "progress-updates-language" || text.startsWith("Progress update output language override:")) return 1;
+	if (text.startsWith("<system-reminder>\nThe following workspace instructions") || kind === "agent-instructions" && !text.startsWith("<system-reminder>\nUpdated instructions") && !text.startsWith("<system-reminder>\nAdditional instructions") && !text.startsWith("<system-reminder>\nInstructions removed")) return 2;
+	if (text.startsWith("<system-reminder>\nA skill is a reusable set of task-specific instructions.") || kind === "skill-catalog" && !text.startsWith("<system-reminder>\nThe available skill catalog changed.")) return 3;
+	if (kind === "runtime-context" || text.startsWith("Current runtime context. This snapshot supersedes")) return 4;
+	return null;
+}
+function isCompactCheckpointMessage(message) {
+	if (message === void 0 || message.role !== "user" || isToolResultMessage$9(message)) return false;
+	const source = isRecord$64(message.source) ? message.source : void 0;
+	if (source !== void 0 && source.kind === "compact-checkpoint") return true;
+	return firstTextOf$1(message.content)?.trimStart().startsWith("This is an automatically generated checkpoint condensing") === true;
+}
+/** Global instructions precede workspace instructions without reordering equals. */
+function normalizeAgentInstructionsMessage(message) {
+	if (!Array.isArray(message.content) || message.content.length !== 1) return message;
+	const first = message.content[0];
+	if (!isRecord$64(first) || first.type !== "text" || typeof first.text !== "string") return message;
+	const raw = first.text;
+	const firstIdx = raw.indexOf("\n\nInstructions from: ");
+	if (firstIdx < 0) return message;
+	const closeIdx = raw.lastIndexOf("\n</system-reminder>");
+	if (closeIdx <= firstIdx) return message;
+	const header = raw.slice(0, firstIdx);
+	const body = raw.slice(firstIdx + 2, closeIdx);
+	const tail = raw.slice(closeIdx);
+	const sections = body.split("\n\nInstructions from: ");
+	if (sections.length < 2) return message;
+	const tagged = sections.map((sec, idx) => {
+		const firstLineEnd = sec.indexOf("\n");
+		const sourcePath = (firstLineEnd < 0 ? sec : sec.slice(0, firstLineEnd)).trim();
+		return {
+			sec,
+			idx,
+			rank: sourcePath.startsWith("~/") || sourcePath.startsWith("~\\") || sourcePath.includes("/.dsh/") || sourcePath.includes("\\.dsh\\") ? 0 : 1
+		};
+	});
+	if (!tagged.some((item, i) => i > 0 && item.rank < tagged[i - 1].rank)) return message;
+	tagged.sort((a, b) => a.rank - b.rank || a.idx - b.idx);
+	const rebuilt = `${header}\n\nInstructions from: ${tagged.map((item) => item.sec).join("\n\nInstructions from: ")}${tail}`;
+	return {
+		...message,
+		content: [{
+			...first,
+			text: rebuilt
+		}]
+	};
+}
+/** Normalise the initial bootstrap run, including the post-compaction shape. */
+function orderStaticPrefixMessages$1(messages) {
+	try {
+		if (messages.length <= 1) return [...messages];
+		let leadEnd = 0;
+		while (leadEnd < messages.length && messages[leadEnd].role === "user" && !isToolResultMessage$9(messages[leadEnd])) leadEnd += 1;
+		if (leadEnd > 0 && isCompactCheckpointMessage(messages[0])) {
+			const seenInLead = /* @__PURE__ */ new Set();
+			for (let i = 0; i < leadEnd; i++) {
+				const tier = classifyBootstrapUserMessage$1(messages[i]);
+				if (tier !== null) seenInLead.add(tier);
+			}
+			if (!seenInLead.has(1) && !seenInLead.has(2)) {
+				const hoistedByTier = /* @__PURE__ */ new Map();
+				for (let i = leadEnd; i < messages.length; i++) {
+					const msg = messages[i];
+					const tier = classifyBootstrapUserMessage$1(msg);
+					if ((tier === 1 || tier === 2 || tier === 3) && !hoistedByTier.has(tier)) hoistedByTier.set(tier, {
+						index: i,
+						message: tier === 2 ? normalizeAgentInstructionsMessage(msg) : msg
+					});
+				}
+				if (hoistedByTier.size > 0) {
+					const indices = /* @__PURE__ */ new Set();
+					const prefix = [];
+					for (const tier of [
+						1,
+						2,
+						3
+					]) {
+						const entry = hoistedByTier.get(tier);
+						if (entry !== void 0) {
+							indices.add(entry.index);
+							prefix.push(entry.message);
+						}
+					}
+					return [...prefix, ...messages.filter((_, index) => !indices.has(index))];
+				}
+			}
+		}
+		if (leadEnd <= 1) return [...messages];
+		const seen = /* @__PURE__ */ new Set();
+		const tagged = [];
+		let hasBootstrap = false;
+		for (let i = 0; i < leadEnd; i++) {
+			const raw = messages[i];
+			const tier = classifyBootstrapUserMessage$1(raw);
+			if (tier !== null && !seen.has(tier)) {
+				seen.add(tier);
+				hasBootstrap = true;
+				tagged.push({
+					message: tier === 2 ? normalizeAgentInstructionsMessage(raw) : raw,
+					rank: tier,
+					index: i
+				});
+			} else tagged.push({
+				message: raw,
+				rank: 5,
+				index: i
+			});
+		}
+		if (!hasBootstrap) return [...messages];
+		if (!tagged.some((item, i) => i > 0 && item.rank < tagged[i - 1].rank || item.message !== messages[i])) return [...messages];
+		tagged.sort((a, b) => a.rank - b.rank || a.index - b.index);
+		return [...tagged.map((item) => item.message), ...messages.slice(leadEnd)];
+	} catch {
+		return [...messages];
+	}
+}
+/** Apply the policy at an adapter boundary; disabled providers keep object identity. */
+function withStaticPrefixOrdering(provider, options, override) {
+	if (!promptCachePolicyOf(provider, override).staticPrefix) return options;
+	const messages = orderStaticPrefixMessages$1(options.messages);
+	return messages.some((message, index) => message !== options.messages[index]) ? {
+		...options,
+		messages
+	} : options;
+}
+function allTextOf(content) {
+	if (typeof content === "string") return content;
+	if (!Array.isArray(content)) return "";
+	return content.map((block) => {
+		if (!isRecord$64(block)) return "";
+		if (block.type === "text" && typeof block.text === "string") return block.text;
+		if (block.type === "tool-result") return allTextOf(block.content);
+		return "";
+	}).join("");
+}
+function prefixFingerprint(text) {
+	let hash = 2166136261;
+	for (let index = 0; index < text.length; index++) {
+		hash ^= text.charCodeAt(index);
+		hash = Math.imul(hash, 16777619);
+	}
+	return (hash >>> 0).toString(16);
+}
+/** Provider-neutral identity for requests that may share one cold prefix. */
+function sharedPrefixGroupKey(provider, options, namespace, accountScope) {
+	if (isAuxiliaryRequest(options)) return void 0;
+	const systemParts = [];
+	if (typeof options.system === "string" && options.system.trim() !== "") systemParts.push(options.system);
+	for (const message of options.messages) if (message.role === "system") {
+		const text = allTextOf(message.content);
+		if (text !== "") systemParts.push(text);
+	}
+	if (systemParts.length === 0) return void 0;
+	const systemHash = prefixFingerprint(systemParts.join("\n\n"));
+	const tools = options.tools ?? [];
+	const toolsHash = tools.length === 0 ? "none" : prefixFingerprint(JSON.stringify(tools.map((tool) => [
+		tool.name,
+		tool.description ?? "",
+		tool.parameters ?? {}
+	])));
+	const staticParts = [];
+	for (const message of orderStaticPrefixMessages$1(options.messages)) {
+		const tier = classifyBootstrapUserMessage$1(message);
+		if (tier !== 1 && tier !== 2 && tier !== 3) break;
+		staticParts.push(allTextOf(message.content));
+	}
+	const staticHash = staticParts.length === 0 ? "none" : prefixFingerprint(staticParts.join("\n\n"));
+	return [
+		provider,
+		namespace,
+		options.model,
+		systemHash,
+		toolsHash,
+		staticHash,
+		accountScope
+	].join(":");
+}
+/** Pure prefix-diff policy shared by provider-specific telemetry stores. */
+function classifyPrefixDrift(previous, current) {
+	if (previous === void 0 || previous.systemPromptHash === null && previous.toolsHash === null && previous.cacheKey === null) return "first-request";
+	if (current.cacheKey === null) return "cold-key";
+	if (previous.cacheKey !== current.cacheKey) return "cache-key";
+	if (previous.systemPromptHash !== current.systemPromptHash) return "system-prompt";
+	if (previous.toolsHash !== current.toolsHash) return "tools";
+	return "stable";
+}
+//#endregion
 //#region src/host/antigravity/cache-stats.ts
 const EMPTY_TOTALS = {
 	requests: 0,
@@ -10492,6 +11046,7 @@ function streamFailure$2(error, signal) {
 	const chain = errorChain(error).trim();
 	return new LlmError(`Antigravity stream failed: ${chain === "" ? "the response stream ended unexpectedly" : chain.slice(0, 500)}`, "TRANSPORT", { cause: error });
 }
+const antigravityPrefixRequests = new PrefixRequestCoordinator();
 var AntigravityAdapter = class extends LlmAdapter {
 	store;
 	modelSettings;
@@ -10614,7 +11169,7 @@ var AntigravityAdapter = class extends LlmAdapter {
 		const runtimeModel = routing?.routing[effort] || routing?.defaultRequestId || model.id;
 		const family = quotaFamilyForModel(model.id);
 		const runtimeMaxOutput = catalogMaxOutputFor(runtimeModel, catalog);
-		const normalizedOptions = normalizeGenerateOptions(options);
+		const normalizedOptions = withStaticPrefixOrdering("antigravity", normalizeGenerateOptions(options), this.options.promptCache);
 		const images = await resolveRequestImages$3(normalizedOptions, this.options.attachments, signal);
 		const requestOptions = offloadOldestRequestImages$2(normalizedOptions, MAX_REQUEST_IMAGE_BYTES$3, images);
 		let recoveredMissingThinkingSignature = false;
@@ -10656,7 +11211,9 @@ var AntigravityAdapter = class extends LlmAdapter {
 			let firstResponse;
 			for (const endpoint of endpointCandidates()) try {
 				let sentBody = body;
-				response = await fetchFn(`${endpoint}/v1internal:streamGenerateContent?alt=sse`, {
+				const cachePolicy = promptCachePolicyOf("antigravity", this.options.promptCache);
+				const groupKey = cachePolicy.coldStart === "off" ? void 0 : sharedPrefixGroupKey("antigravity", requestOptions, endpoint, account.id);
+				response = await fetchWithPrefixCoordination(fetchFn, antigravityPrefixRequests, groupKey, cachePolicy.coldStart === "warm" ? cachePolicy.ttlMs : void 0, `${endpoint}/v1internal:streamGenerateContent?alt=sse`, {
 					method: "POST",
 					headers,
 					body,
@@ -12173,6 +12730,135 @@ function createControlledModelFetch(baseFetch, options) {
 	};
 }
 //#endregion
+//#region src/host/common/tool-runaway.ts
+/**
+* Provider-neutral tool-call repetition primitives.
+*
+* The detector knows nothing about model wire formats. Providers supply reminder
+* text and may opt into a hard handoff threshold; coordination tools are exempt
+* because they are expected to recur while real work continues.
+*/
+function isRecord$63(value) {
+	return typeof value === "object" && value !== null;
+}
+function isPlainRecord$2(value) {
+	if (!isRecord$63(value) || Array.isArray(value)) return false;
+	const prototype = Object.getPrototypeOf(value);
+	return prototype === Object.prototype || prototype === null;
+}
+function sortJsonValue(value) {
+	if (Array.isArray(value)) return value.map(sortJsonValue);
+	if (!isPlainRecord$2(value)) return value;
+	const out = {};
+	for (const key of Object.keys(value).sort()) out[key] = sortJsonValue(value[key]);
+	return out;
+}
+/** Stable identity for one tool call, independent of object key order. */
+function canonicalToolCallKey(toolName, args) {
+	let json;
+	try {
+		json = JSON.stringify(sortJsonValue(args));
+	} catch {
+		json = void 0;
+	}
+	return toolName + " " + (json ?? String(args));
+}
+/** One immediately-repeated action streak. A different action starts at one. */
+function createExactRepeatTracker() {
+	let current;
+	return {
+		observe(toolName, args) {
+			const key = canonicalToolCallKey(toolName, args);
+			current = {
+				key,
+				count: current?.key === key ? current.count + 1 : 1
+			};
+			return current;
+		},
+		current() {
+			return current;
+		},
+		reset() {
+			current = void 0;
+		}
+	};
+}
+/** Coordination calls that must not break or inflate an action streak. */
+const TOOL_RUNAWAY_EXEMPT_TOOLS = Object.freeze([
+	"progress_update",
+	"todo_write",
+	"ask_user_question",
+	"enter_plan_mode",
+	"exit_plan_mode",
+	"leave_plan_mode",
+	"create_goal",
+	"update_goal",
+	"get_goal",
+	"dispatch_ledger",
+	"present"
+]);
+/** Read the live global switch; legacy Kimi-only installs without it stay on. */
+function resolveToolRunawayGuardEnabled(config) {
+	const field = config?.toolRunawayGuard;
+	if (field !== null && typeof field === "object") try {
+		return field.get() !== false;
+	} catch {
+		return true;
+	}
+	return field !== false;
+}
+/**
+* Exact-repeat state machine shared by provider profiles. A missing forceStopAt
+* gives reminders only; no provider inherits Kimi's hard stop accidentally.
+*/
+function createExactRepeatBreaker(options) {
+	const sessions = /* @__PURE__ */ new Map();
+	const maxSessions = options.maxSessions ?? 256;
+	const exempt = new Set(options.exemptTools ?? TOOL_RUNAWAY_EXEMPT_TOOLS);
+	const stateOf = (sessionId) => {
+		let state = sessions.get(sessionId);
+		if (state === void 0) state = {
+			tracker: createExactRepeatTracker(),
+			count: 0,
+			handoffActive: false
+		};
+		sessions.delete(sessionId);
+		sessions.set(sessionId, state);
+		if (sessions.size > maxSessions) {
+			const oldest = sessions.keys().next();
+			if (oldest.done !== true) sessions.delete(oldest.value);
+		}
+		return state;
+	};
+	return {
+		checkBeforeExecute(sessionId, toolName, args) {
+			if (options.forceStopAt === void 0) return void 0;
+			const state = stateOf(sessionId);
+			if (state.handoffActive) return options.handoffVeto ?? options.handoffReminder;
+			const key = canonicalToolCallKey(toolName, args);
+			if (state.tracker.current()?.key === key && state.count + 1 >= options.forceStopAt) {
+				state.count += 1;
+				state.handoffActive = true;
+				return options.handoffVeto ?? options.handoffReminder;
+			}
+		},
+		recordAfterExecute(sessionId, toolName, args) {
+			if (exempt.has(toolName)) return void 0;
+			const state = stateOf(sessionId);
+			if (state.handoffActive) return options.handoffReminder ?? options.reminder(state.count);
+			state.count = state.tracker.observe(toolName, args).count;
+			if (options.forceStopAt !== void 0 && state.count >= options.forceStopAt) {
+				state.handoffActive = true;
+				return options.handoffReminder ?? options.reminder(state.count);
+			}
+			return options.reminder(state.count);
+		},
+		resetSession(sessionId) {
+			sessions.delete(sessionId);
+		}
+	};
+}
+//#endregion
 //#region src/host/common/provider-route.ts
 /**
 * Claim one provider route for this plugin and keep claiming it.
@@ -12728,7 +13414,7 @@ function preferencesPath() {
 	return path.join(dshHomeDir(), "storages", "dsh-chatgpt-subscription-preferences.json");
 }
 /** Distinguishes concurrent temporary files written by one process. */
-let writeSequence$1 = 0;
+let writeSequence = 0;
 /** One JSON document, validated and shaped by a schemastery schema. */
 var FilePreferencesStore = class {
 	schema;
@@ -12830,8 +13516,8 @@ var FilePreferencesStore = class {
 	}
 	async write(value) {
 		await fs.mkdir(path.dirname(this.filePath), { recursive: true });
-		writeSequence$1 += 1;
-		const tmp = `${this.filePath}.tmp.${process.pid}.${writeSequence$1}`;
+		writeSequence += 1;
+		const tmp = `${this.filePath}.tmp.${process.pid}.${writeSequence}`;
 		try {
 			await fs.writeFile(tmp, JSON.stringify(value, null, 2), {
 				encoding: "utf8",
@@ -14049,9 +14735,17 @@ const REASONING_DELTA_FLUSH_CHARS = 768;
 const REASONING_TRUNCATED_NOTICE = "\n\n[Reasoning summary truncated to keep the DSH web UI responsive.]";
 /** Cap on live turns whose routing state is remembered at once. */
 const MAX_TRACKED_TURN_STATES = 200;
+const codexPrefixRequests = new PrefixRequestCoordinator();
+function codexColdGroupKey(payload, credentials) {
+	const instructions = typeof payload.instructions === "string" ? payload.instructions : "";
+	if (instructions === "") return void 0;
+	const digest = createHash("sha256").update(instructions).update("\0").update(JSON.stringify(Array.isArray(payload.tools) ? payload.tools : [])).digest("hex");
+	return `${String(payload.model ?? "")}:${digest}:${authOwnerKey(credentials)}`;
+}
 var ResponsesClient = class {
 	oauth;
 	attachments;
+	options;
 	fetchFn;
 	onGenerationFinished;
 	outputVerbosity;
@@ -14080,6 +14774,7 @@ var ResponsesClient = class {
 	constructor(oauth, attachments, options = {}) {
 		this.oauth = oauth;
 		this.attachments = attachments;
+		this.options = options;
 		this.fetchFn = options.fetchFn ?? fetch;
 		this.localRawImages = options.localRawImages ?? {};
 		this.onGenerationFinished = options.onGenerationFinished ?? (() => void 0);
@@ -14092,7 +14787,7 @@ var ResponsesClient = class {
 	}
 	localRawImages;
 	async *stream(rawOptions, capabilities) {
-		const options = normalizeGenerateOptions(rawOptions);
+		const options = withStaticPrefixOrdering("codex-chatgpt", normalizeGenerateOptions(rawOptions), this.options.promptCache);
 		const sessionId = stableSessionId(options.sessionId);
 		const resolveCapabilities = capabilities ?? this.capabilities ?? defaultWireCapabilities;
 		const turnKey = turnKeyFor(sessionId);
@@ -14275,7 +14970,9 @@ var ResponsesClient = class {
 	}
 	async request(payload, credentials, sessionId, turnKey, signal, useResponsesLite = false) {
 		for (let attempt = 0;; attempt += 1) try {
-			return await this.fetchFn(CODEX_RESPONSES_URL, {
+			const cachePolicy = promptCachePolicyOf("codex-chatgpt", this.options.promptCache);
+			const groupKey = cachePolicy.coldStart === "off" ? void 0 : codexColdGroupKey(payload, credentials);
+			return await fetchWithPrefixCoordination(this.fetchFn, codexPrefixRequests, groupKey, cachePolicy.coldStart === "warm" ? cachePolicy.ttlMs : void 0, CODEX_RESPONSES_URL, {
 				method: "POST",
 				headers: {
 					...codexHeaders(credentials, sessionId, { turnState: this.turnStateFor(turnKey, credentials) }),
@@ -28497,6 +29194,10 @@ var FileResultReferenceStore = class {
 */
 /** 有界登记表：Map 的插入顺序同时是 LRU 顺序（命中即重新插入到队尾）。 */
 const registry = /* @__PURE__ */ new Map();
+const DEFAULT_USAGE_PROVIDER = "kimi-code";
+function usageScopeKey(sessionId, provider) {
+	return (provider?.trim() || DEFAULT_USAGE_PROVIDER) + ":" + sessionId;
+}
 /** 一个中性返回值，异常路径复用（不含任何容器操作）。 */
 function neutralState(sessionId) {
 	return {
@@ -28601,6 +29302,7 @@ function normalizeUsage(input) {
 	const at = typeof input.at === "number" && Number.isFinite(input.at) ? input.at : Date.now();
 	const normalized = {
 		sessionId,
+		provider: typeof input.provider === "string" && input.provider.trim() !== "" ? input.provider.trim() : DEFAULT_USAGE_PROVIDER,
 		inputTokens: finiteCount(input.inputTokens),
 		cacheReadTokens: finiteCount(input.cacheReadTokens),
 		outputTokens: finiteCount(input.outputTokens),
@@ -28631,16 +29333,19 @@ function stateOf(sessionId, reports) {
 */
 function recordCacheWarmupUsage(input) {
 	let sessionId = "";
+	let provider;
 	try {
 		const usage = normalizeUsage(input);
 		if (usage === void 0) return neutralState("");
 		sessionId = usage.sessionId;
-		const previous = registry.get(usage.sessionId);
+		provider = usage.provider;
+		const scope = usageScopeKey(usage.sessionId, usage.provider);
+		const previous = registry.get(scope);
 		const kept = previous === void 0 ? [] : [...previous];
 		kept.push(Object.freeze({ ...usage }));
 		while (kept.length > CACHE_WARMUP_DEFAULTS.maxUsageReportsPerSession) kept.shift();
-		if (previous !== void 0) registry.delete(usage.sessionId);
-		registry.set(usage.sessionId, Object.freeze(kept));
+		if (previous !== void 0) registry.delete(scope);
+		registry.set(scope, Object.freeze(kept));
 		while (registry.size > CACHE_WARMUP_DEFAULTS.maxUsageSessions) {
 			const oldest = registry.keys().next();
 			if (oldest.done === true) break;
@@ -28649,7 +29354,7 @@ function recordCacheWarmupUsage(input) {
 		return stateOf(usage.sessionId, kept);
 	} catch {
 		try {
-			if (sessionId !== "") registry.delete(sessionId);
+			if (sessionId !== "") registry.delete(usageScopeKey(sessionId, provider));
 		} catch {}
 		return neutralState(sessionId);
 	}
@@ -28659,10 +29364,10 @@ function recordCacheWarmupUsage(input) {
 * @param sessionId - 会话 id。
 * @returns 报告副本；没有或读不到时为空数组。
 */
-function cacheWarmupUsageReports(sessionId) {
+function cacheWarmupUsageReports(sessionId, provider) {
 	try {
 		if (typeof sessionId !== "string" || sessionId === "") return [];
-		const reports = registry.get(sessionId);
+		const reports = registry.get(usageScopeKey(sessionId, provider));
 		return reports === void 0 ? [] : [...reports];
 	} catch {
 		return [];
@@ -28675,11 +29380,12 @@ function cacheWarmupUsageReports(sessionId) {
 function cacheWarmupUsageSessions() {
 	try {
 		const summaries = [];
-		for (const [sessionId, reports] of registry.entries()) {
-			if (reports.length === 0) continue;
+		for (const reports of registry.values()) {
+			const sample = reports[0];
+			if (sample === void 0) continue;
 			const last = reports[reports.length - 1];
 			summaries.push({
-				sessionId,
+				sessionId: sample.sessionId,
 				reports: reports.length,
 				lastAt: last.at,
 				hitRatio: cacheWarmupHitRatio(last)
@@ -28726,10 +29432,10 @@ const WARMUP_BUDGET = Object.freeze({
 * @param durable - 进程重启后尚无内存报告时的持久化回退摘要（可选）。
 * @returns 本次投影可用的预算。
 */
-function cacheWarmupBudget(sessionId, now = Date.now(), durable) {
+function cacheWarmupBudget(sessionId, now = Date.now(), durable, provider) {
 	try {
 		if (typeof sessionId !== "string" || sessionId.trim() === "") return { ...WARMUP_BUDGET };
-		const reports = cacheWarmupUsageReports(sessionId);
+		const reports = cacheWarmupUsageReports(sessionId, provider);
 		if (reports.length === 0) {
 			if (durable !== void 0 && durable !== null && typeof durable === "object") {
 				const lastAt = durable.lastActiveAt;
@@ -28765,7 +29471,7 @@ function cacheWarmupBudget(sessionId, now = Date.now(), durable) {
 * @param now - 当前时间戳（毫秒），默认 `Date.now()`。
 * @returns 快照。
 */
-function cacheWarmupSnapshot(sessionId, now = Date.now()) {
+function cacheWarmupSnapshot(sessionId, now = Date.now(), provider) {
 	const fallback = {
 		budget: { ...WARMUP_BUDGET },
 		defaults: CACHE_WARMUP_DEFAULTS,
@@ -28783,7 +29489,7 @@ function cacheWarmupSnapshot(sessionId, now = Date.now()) {
 		sessions: 0
 	};
 	try {
-		const budget = cacheWarmupBudget(sessionId, now);
+		const budget = cacheWarmupBudget(sessionId, now, void 0, provider);
 		const reports = typeof sessionId === "string" && sessionId.trim() !== "" ? cacheWarmupUsageReports(sessionId) : [];
 		const totals = {
 			inputTokens: 0,
@@ -29378,7 +30084,7 @@ function planRetrievalOutline(text, toolName) {
 * @returns 读到的元数据；两者都没有时 undefined。
 */
 function shellMetadataOf(holder) {
-	const job = isPlainRecord$2(holder.job) ? holder.job : void 0;
+	const job = isPlainRecord$1(holder.job) ? holder.job : void 0;
 	let exitCode;
 	if (typeof holder.exitCode === "number" && Number.isFinite(holder.exitCode)) exitCode = holder.exitCode;
 	if (exitCode === void 0 && typeof job?.exitCode === "number" && Number.isFinite(job.exitCode)) exitCode = job.exitCode;
@@ -29509,7 +30215,7 @@ const JSON_ENVELOPE_LOG_KEYS = [
 /** JSON envelope 其余字段允许占用的最大序列化字符数（防止把含批量源码的混合 envelope 压掉）。 */
 const JSON_ENVELOPE_MAX_OTHER_CHARS = 800;
 /** 判断一个值是不是普通 JSON 对象（不是数组，也不是 null）。 */
-function isPlainRecord$2(value) {
+function isPlainRecord$1(value) {
 	return typeof value === "object" && value !== null && !Array.isArray(value);
 }
 /** 重建信封：只把预览写回被替换的键，其余字段原样保留。 */
@@ -29611,7 +30317,7 @@ function extractJsonEnvelopeLogField(text, toolName) {
 		});
 	}
 	for (const [outerKey, outerValue] of Object.entries(record)) {
-		if (!isPlainRecord$2(outerValue)) continue;
+		if (!isPlainRecord$1(outerValue)) continue;
 		for (const key of JSON_ENVELOPE_LOG_KEYS) {
 			const value = outerValue[key];
 			if (typeof value === "string" && value.length > 0) candidates.push({
@@ -29701,7 +30407,7 @@ function canonicalGrepMatches(value) {
 	if (!Array.isArray(value) || value.length === 0) return void 0;
 	const matches = [];
 	for (const entry of value) {
-		if (!isPlainRecord$2(entry)) return void 0;
+		if (!isPlainRecord$1(entry)) return void 0;
 		if (typeof entry.path !== "string") return void 0;
 		if (!Number.isSafeInteger(entry.lineNumber)) return void 0;
 		if (typeof entry.line !== "string") return void 0;
@@ -29715,12 +30421,12 @@ function canonicalGrepMatches(value) {
 }
 /** 规范读取结果。 */
 function canonicalReadResult(value) {
-	if (!isPlainRecord$2(value)) return void 0;
+	if (!isPlainRecord$1(value)) return void 0;
 	if (typeof value.path !== "string" || value.path.trim() === "") return void 0;
 	if (!Array.isArray(value.lines) || value.lines.length === 0) return void 0;
 	const lines = [];
 	for (const entry of value.lines) {
-		if (!isPlainRecord$2(entry)) return void 0;
+		if (!isPlainRecord$1(entry)) return void 0;
 		if (!Number.isSafeInteger(entry.number)) return void 0;
 		if (typeof entry.text !== "string") return void 0;
 		lines.push({
@@ -29745,7 +30451,7 @@ function findGrepBinding(record) {
 		matches: top
 	};
 	for (const [outerKey, outerValue] of Object.entries(record)) {
-		if (!isPlainRecord$2(outerValue)) continue;
+		if (!isPlainRecord$1(outerValue)) continue;
 		const nested = canonicalGrepMatches(outerValue.matches);
 		if (nested !== void 0) return {
 			binding: {
@@ -29772,7 +30478,7 @@ function findGlobBinding(record) {
 		...typeof record.root === "string" ? { root: record.root } : {}
 	};
 	for (const [outerKey, outerValue] of Object.entries(record)) {
-		if (!isPlainRecord$2(outerValue)) continue;
+		if (!isPlainRecord$1(outerValue)) continue;
 		const nested = canonicalGlobPaths(outerValue.paths);
 		if (nested !== void 0) return {
 			binding: {
@@ -29950,13 +30656,13 @@ function collectCompositeEntries(record) {
 	for (const [key, value] of Object.entries(record)) {
 		if (Array.isArray(value)) {
 			value.forEach((item, index) => {
-				if (isPlainRecord$2(item)) push(`${key}[${index}]`, item);
+				if (isPlainRecord$1(item)) push(`${key}[${index}]`, item);
 			});
 			continue;
 		}
-		if (!isPlainRecord$2(value)) continue;
+		if (!isPlainRecord$1(value)) continue;
 		if (push(key, value)) continue;
-		for (const [subKey, subValue] of Object.entries(value)) if (isPlainRecord$2(subValue)) push(`${key}.${subKey}`, subValue);
+		for (const [subKey, subValue] of Object.entries(value)) if (isPlainRecord$1(subValue)) push(`${key}.${subKey}`, subValue);
 	}
 	return entries;
 }
@@ -30380,6 +31086,17 @@ async function projectToolResult(input) {
 	};
 }
 //#endregion
+//#region src/host/common/failure-classification.ts
+/** Clamp a measured reset delay without parking an account forever or reopening at once. */
+function cooldownFromReset(input) {
+	const now = input.now ?? Date.now();
+	const minimum = input.minimumMs ?? 0;
+	const maximum = input.maximumMs ?? Number.MAX_SAFE_INTEGER;
+	if (input.resetAtMs === void 0 || input.resetAtMs === null || !Number.isFinite(input.resetAtMs)) return Math.min(maximum, Math.max(minimum, input.fallbackMs));
+	if (input.resetAtMs <= now) return Math.min(maximum, Math.max(minimum, input.fallbackMs));
+	return Math.min(maximum, Math.max(minimum, input.resetAtMs - now));
+}
+//#endregion
 //#region src/host/kimi-code/model-catalog.ts
 /**
 * Request-body ceiling the 1M-context tier defaults to.
@@ -31363,6 +32080,67 @@ async function persistRegion(region) {
 		await fs.writeFile(file, region, "utf8");
 	} catch {}
 }
+//#endregion
+//#region src/host/common/cache-hint.ts
+/** Missing data always skips; a provider must supply its own measured TTL. */
+function evaluateCacheHint$1(input) {
+	const { now, lastActiveAt, totalTokens } = input;
+	if (lastActiveAt === void 0 || totalTokens === void 0) return { kind: "skip" };
+	if (!Number.isFinite(lastActiveAt) || !Number.isFinite(totalTokens)) return { kind: "skip" };
+	const ttlMs = input.ttlMs;
+	const minTokens = input.minTokensToHint ?? 1e5;
+	if (ttlMs === void 0 || !Number.isFinite(ttlMs) || ttlMs <= 0) return { kind: "skip" };
+	if (!Number.isFinite(minTokens) || minTokens < 0 || totalTokens < minTokens) return { kind: "skip" };
+	const idleMs = now - lastActiveAt;
+	if (idleMs <= ttlMs) return { kind: "skip" };
+	return {
+		kind: "hint",
+		idleMs,
+		totalTokens
+	};
+}
+//#endregion
+//#region src/host/kimi-code/cache-hint.ts
+/**
+* When it is worth telling the user their prompt cache has gone cold.
+*
+* Kimi's cache is automatic and content-hash based, so a client gets no marker
+* that it is no longer being reused — the only evidence is the read ratio in
+* the usage stream. A conversation that sat idle past the cache lifetime pays
+* the full prompt again on its next turn, and it is large enough on a real
+* codebase to be worth a sentence.
+*
+* The trigger mirrors the official client's: idle beyond the cache lifetime AND
+* a context large enough that reprocessing it actually costs something. The
+* upstream rules arrive from a remote config that a private deployment cannot
+* reach, so the lifetime here is the measured one; the decision function stays
+* a pure function of its inputs so swapping the source later changes only where
+* the number comes from.
+*/
+/**
+* How long an entry survives, in milliseconds.
+*
+* MEASURED, not documented. The coding endpoint publishes no cache lifetime and
+* its `5m`/`1h` write tiers cannot be selected from any official client, so the
+* only figure available is the survival curve: 30 minutes, derived from the
+* endpoint's own measured behaviour.
+*
+* The measurement (2026-10-02/03, three replicas per age, one fresh prefix per
+* age so the probe does not renew what it is timing) put Kimi's coding plan at
+* 3/3 warm through 25 minutes, 1/3 at 30 minutes, and 0/3 at 35 minutes — a
+* 25–35 minute bracket around a fuzzy edge.
+*
+* 25 minutes is the last age at which every replica was still warm, and that is
+* what this notice needs: it must arrive BEFORE the entry is surely gone. The
+* 30-minute point where most replicas had gone cold is too late — by then the
+* user has already paid for the reprocessed prefix, so a notice there describes
+* the bill instead of preventing it.
+*
+* The old value was 5 minutes, taken from the `5m` write tier's NAME. On this
+* endpoint the name is not the lifetime: that estimate warned from minute 6
+* onward while the entry was still warm for another 19.
+*/
+const CACHE_TTL_MS = 1500 * 1e3;
 /**
 * Last time each session did something, epoch ms.
 *
@@ -31394,26 +32172,13 @@ function sessionLastActiveAt(sessionId) {
 	if (sessionId === void 0 || sessionId.trim() === "") return void 0;
 	return lastActiveBySession.get(sessionId.trim());
 }
-/**
-* Whether one conversation should be warned about a cold cache.
-*
-* Every missing-data branch skips. That asymmetry is deliberate and matches
-* the official client: a false negative costs one turn nobody noticed, while a
-* false positive teaches the user the notice is noise and stops them reading
-* the one that mattered.
-*/
+/** Kimi supplies its measured defaults; other providers call the common pure function directly. */
 function evaluateCacheHint(input) {
-	const { now, lastActiveAt, totalTokens } = input;
-	if (lastActiveAt === void 0 || totalTokens === void 0) return { kind: "skip" };
-	if (!Number.isFinite(lastActiveAt) || !Number.isFinite(totalTokens)) return { kind: "skip" };
-	if (totalTokens < 1e5) return { kind: "skip" };
-	const idleMs = now - lastActiveAt;
-	if (idleMs <= 15e5) return { kind: "skip" };
-	return {
-		kind: "hint",
-		idleMs,
-		totalTokens
-	};
+	return evaluateCacheHint$1({
+		...input,
+		ttlMs: input.ttlMs ?? 15e5,
+		minTokensToHint: input.minTokensToHint ?? 1e5
+	});
 }
 //#endregion
 //#region src/host/kimi-code/oauth.ts
@@ -33414,6 +34179,48 @@ function estimatedInputTokens(options) {
 	return Math.ceil(characters / 4);
 }
 //#endregion
+//#region src/host/common/reasoning-dialect.ts
+/** Learn the reasoning field a concrete OpenAI-compatible endpoint speaks. */
+const KNOWN_REASONING_KEYS$1 = [
+	"reasoning_content",
+	"reasoning_details",
+	"reasoning"
+];
+var ReasoningKeyDialect$1 = class {
+	defaultKey;
+	keys;
+	detected;
+	constructor(defaultKey = KNOWN_REASONING_KEYS$1[0], keys = KNOWN_REASONING_KEYS$1) {
+		this.defaultKey = defaultKey;
+		this.keys = keys;
+	}
+	/** Record an observed field and return its text. Empty strings still teach the key. */
+	observe(source) {
+		if (typeof source !== "object" || source === null) return void 0;
+		const record = source;
+		for (const key of this.keys) {
+			const value = record[key];
+			if (typeof value === "string") {
+				this.learn(key);
+				return value;
+			}
+		}
+	}
+	/** Record a previously parsed fragment without re-reading the wire object. */
+	observeKey(key) {
+		if (this.keys.includes(key)) this.learn(key);
+	}
+	learn(key) {
+		this.detected = key;
+	}
+	outboundKey() {
+		return this.detected ?? this.defaultKey;
+	}
+	reset() {
+		this.detected = void 0;
+	}
+};
+//#endregion
 //#region src/host/common/video.ts
 /**
 * Video input, shared by every line in this plugin that accepts it.
@@ -33717,13 +34524,8 @@ function sanitizeText$7(text) {
 * `reasoning_details`. The first entry doubles as the default for a request
 * made before any response has been seen.
 */
-const KNOWN_REASONING_KEYS = [
-	"reasoning_content",
-	"reasoning_details",
-	"reasoning"
-];
-/** Default outbound reasoning key, used until a response proves otherwise. */
-const DEFAULT_REASONING_KEY = KNOWN_REASONING_KEYS[0];
+const KNOWN_REASONING_KEYS = KNOWN_REASONING_KEYS$1;
+KNOWN_REASONING_KEYS[0];
 /**
 * One `reasoning_details` element, in an official or a compatibility shape.
 *
@@ -33846,8 +34648,7 @@ function reasoningFragmentsOf(source) {
 * turn. Detection never clears: a response with no reasoning keeps the last
 * known dialect, and an endpoint that switches is adapted on its next reply.
 */
-var ReasoningKeyDialect = class {
-	detected;
+var ReasoningKeyDialect = class extends ReasoningKeyDialect$1 {
 	/**
 	* Every reasoning fragment one inbound delta carries, remembering the dialect.
 	*
@@ -33867,13 +34668,13 @@ var ReasoningKeyDialect = class {
 		const fragments = reasoningFragmentsOf(source);
 		const teaching = fragments.find((fragment) => fragment.form === void 0 || fragment.form === "string");
 		if (teaching !== void 0) {
-			this.detected = teaching.key;
+			this.learn(teaching.key);
 			return fragments;
 		}
 		if (typeof source === "object" && source !== null) {
 			const record = source;
 			for (const key of KNOWN_REASONING_KEYS) if (typeof record[key] === "string") {
-				this.detected = key;
+				this.learn(key);
 				break;
 			}
 		}
@@ -33883,14 +34684,6 @@ var ReasoningKeyDialect = class {
 	observe(source) {
 		return this.observeAll(source)[0]?.text;
 	}
-	/** The key to serialize thinking into on an outbound assistant message. */
-	outboundKey() {
-		return this.detected ?? DEFAULT_REASONING_KEY;
-	}
-	/** Forget the detected key, returning to the default. Test seam only. */
-	reset() {
-		this.detected = void 0;
-	}
 };
 /**
 * Process-wide reasoning dialect for this route.
@@ -34236,59 +35029,7 @@ function nonSystemMessages$6(options) {
 * arrived (append-only in-history).
 */
 function classifyBootstrapUserMessage(message) {
-	if (message.role !== "user" || isToolResultMessage$7(message)) return null;
-	const source = isRecord$53(message.source) ? message.source : void 0;
-	const kind = source !== void 0 ? asString$21(source.kind) : void 0;
-	const text = firstTextOf(message.content)?.trimStart() ?? "";
-	if (kind === "progress-updates-language" || text.startsWith("Progress update output language override:")) return 1;
-	if (text.startsWith("<system-reminder>\nThe following workspace instructions") || kind === "agent-instructions" && !text.startsWith("<system-reminder>\nUpdated instructions") && !text.startsWith("<system-reminder>\nAdditional instructions") && !text.startsWith("<system-reminder>\nInstructions removed")) return 2;
-	if (text.startsWith("<system-reminder>\nA skill is a reusable set of task-specific instructions.") || kind === "skill-catalog" && !text.startsWith("<system-reminder>\nThe available skill catalog changed.")) return 3;
-	if (kind === "runtime-context" || text.startsWith("Current runtime context. This snapshot supersedes")) return 4;
-	return null;
-}
-function isCompactCheckpointMessage(message) {
-	if (message === void 0 || message.role !== "user" || isToolResultMessage$7(message)) return false;
-	const source = isRecord$53(message.source) ? message.source : void 0;
-	if (source !== void 0 && asString$21(source.kind) === "compact-checkpoint") return true;
-	return (firstTextOf(message.content)?.trimStart() ?? "").startsWith("This is an automatically generated checkpoint condensing");
-}
-/**
-* Normalize `agent-instructions` text so global `~/.dsh/` instructions precede
-* workspace-local `AGENTS.md` instructions when both appear in the same block.
-*/
-function normalizeAgentInstructionsMessage(message) {
-	if (!Array.isArray(message.content) || message.content.length !== 1) return message;
-	const first = message.content[0];
-	if (!isRecord$53(first) || first.type !== "text" || typeof first.text !== "string") return message;
-	const raw = first.text;
-	const firstIdx = raw.indexOf("\n\nInstructions from: ");
-	if (firstIdx < 0) return message;
-	const closeIdx = raw.lastIndexOf("\n</system-reminder>");
-	if (closeIdx <= firstIdx) return message;
-	const header = raw.slice(0, firstIdx);
-	const body = raw.slice(firstIdx + 2, closeIdx);
-	const tail = raw.slice(closeIdx);
-	const sections = body.split("\n\nInstructions from: ");
-	if (sections.length < 2) return message;
-	const tagged = sections.map((sec, idx) => {
-		const firstLineEnd = sec.indexOf("\n");
-		const sourcePath = (firstLineEnd < 0 ? sec : sec.slice(0, firstLineEnd)).trim();
-		return {
-			sec,
-			idx,
-			rank: sourcePath.startsWith("~/") || sourcePath.startsWith("~\\") || sourcePath.includes("/.dsh/") || sourcePath.includes("\\.dsh\\") ? 0 : 1
-		};
-	});
-	if (!tagged.some((item, i) => i > 0 && item.rank < tagged[i - 1].rank)) return message;
-	tagged.sort((a, b) => a.rank - b.rank || a.idx - b.idx);
-	const rebuilt = `${header}\n\nInstructions from: ${tagged.map((t) => t.sec).join("\n\nInstructions from: ")}${tail}`;
-	return {
-		...message,
-		content: [{
-			...first,
-			text: rebuilt
-		}]
-	};
+	return classifyBootstrapUserMessage$1(message);
 }
 /**
 * Deterministic entropy-sorted ordering for leading bootstrap reminders.
@@ -34304,73 +35045,7 @@ function normalizeAgentInstructionsMessage(message) {
 *    compaction boundaries and every step after compaction.
 */
 function orderStaticPrefixMessages(messages) {
-	try {
-		if (messages.length <= 1) return [...messages];
-		let leadEnd = 0;
-		while (leadEnd < messages.length && messages[leadEnd].role === "user" && !isToolResultMessage$7(messages[leadEnd])) leadEnd += 1;
-		if (leadEnd > 0 && isCompactCheckpointMessage(messages[0])) {
-			const seenInLead = /* @__PURE__ */ new Set();
-			for (let i = 0; i < leadEnd; i++) {
-				const tier = classifyBootstrapUserMessage(messages[i]);
-				if (tier !== null) seenInLead.add(tier);
-			}
-			if (!seenInLead.has(1) && !seenInLead.has(2)) {
-				const hoistedByTier = /* @__PURE__ */ new Map();
-				for (let i = leadEnd; i < messages.length; i++) {
-					const msg = messages[i];
-					const tier = classifyBootstrapUserMessage(msg);
-					if ((tier === 1 || tier === 2 || tier === 3) && !hoistedByTier.has(tier)) hoistedByTier.set(tier, {
-						index: i,
-						message: tier === 2 ? normalizeAgentInstructionsMessage(msg) : msg
-					});
-				}
-				if (hoistedByTier.size > 0) {
-					const hoistedIndices = /* @__PURE__ */ new Set();
-					const prefix = [];
-					for (const tier of [
-						1,
-						2,
-						3
-					]) {
-						const entry = hoistedByTier.get(tier);
-						if (entry !== void 0) {
-							hoistedIndices.add(entry.index);
-							prefix.push(entry.message);
-						}
-					}
-					const rest = messages.filter((_, idx) => !hoistedIndices.has(idx));
-					return [...prefix, ...rest];
-				}
-			}
-		}
-		if (leadEnd <= 1) return [...messages];
-		const seenTiers = /* @__PURE__ */ new Set();
-		const leadTagged = [];
-		let hasBootstrap = false;
-		for (let i = 0; i < leadEnd; i++) {
-			const rawMsg = messages[i];
-			const tier = classifyBootstrapUserMessage(rawMsg);
-			if (tier !== null && !seenTiers.has(tier)) {
-				seenTiers.add(tier);
-				hasBootstrap = true;
-				leadTagged.push({
-					message: tier === 2 ? normalizeAgentInstructionsMessage(rawMsg) : rawMsg,
-					rank: tier,
-					index: i
-				});
-			} else leadTagged.push({
-				message: rawMsg,
-				rank: 5,
-				index: i
-			});
-		}
-		if (!hasBootstrap) return [...messages];
-		if (!leadTagged.some((item, i) => i > 0 && item.rank < leadTagged[i - 1].rank || item.message !== messages[i])) return [...messages];
-		leadTagged.sort((a, b) => a.rank - b.rank || a.index - b.index);
-		return [...leadTagged.map((item) => item.message), ...messages.slice(leadEnd)];
-	} catch {
-		return [...messages];
-	}
+	return orderStaticPrefixMessages$1(messages);
 }
 const NO_REASONING = {
 	fields: [],
@@ -35992,12 +36667,11 @@ function trackPrefixStability(options) {
 		toolsHash,
 		cacheKey
 	});
-	if (previous === void 0 || previous.systemPromptHash === null && previous.toolsHash === null && previous.cacheKey === null) return "first-request";
-	if (cacheKey === null) return "cold-key";
-	if (previous.cacheKey !== cacheKey) return "cache-key";
-	if (previous.systemPromptHash !== systemPromptHash) return "system-prompt";
-	if (previous.toolsHash !== toolsHash) return "tools";
-	return "stable";
+	return classifyPrefixDrift(previous, {
+		systemPromptHash,
+		toolsHash,
+		cacheKey
+	});
 }
 /** Drift cause recorded for each session. */
 const driftCauses = /* @__PURE__ */ new Map();
@@ -36315,61 +36989,26 @@ function closeStream$4(state, accountId) {
 function assertStreamComplete$5(state) {
 	if (!state.done && state.finishReason === null) throw new LlmError$1("Kimi Code stream ended before its terminal event", "TRANSPORT");
 }
-/** Where the document lives; beside the other plugin-owned JSON. */
-function sessionCacheStatePath() {
-	return path.join(dshHomeDir(), "storages", "dsh-chatgpt-subscription-kimi-cache.json");
-}
-/** Distinguishes concurrent temporary files written by one process. */
-let writeSequence = 0;
-/**
-* Whether a scheduled write should reach the disk.
-*
-* A unit test that exercises the request path must not overwrite the state file
-* of the machine running it; tests build their own store against a temporary
-* path (which always writes) and call {@link SessionCacheStateStore.flush}.
-*/
-function persistenceEnabled() {
-	return process.env.VITEST !== "true" && process.env.NODE_ENV !== "test";
-}
+//#endregion
+//#region src/host/common/session-cache-state.ts
 /**
-* Per-session cache state held in one JSON document.
-*
-* Reads are synchronous because every caller is answering a status request or
-* closing a turn, both of which are allowed to block on a file this size. Writes
-* are debounced: a turn reports once, and several turns in flight must not
-* rewrite the document several times.
+* Generic bounded, durable session state for cache diagnostics and sticky media.
+* One provider owns one document; records remain keyed by session inside it.
 */
-var SessionCacheStateStore = class {
+var CommonSessionCacheStateStore = class CommonSessionCacheStateStore {
 	document;
 	pending;
-	/**
-	* @param filePath - Document location; defaults to {@link sessionCacheStatePath}.
-	* @param maxEntries - Conversations kept; the oldest are dropped first.
-	*/
 	filePath;
 	maxEntries;
-	/**
-	* Whether this store may write at all.
-	*
-	* A store built on an explicit path is someone's own file - a test's temporary
-	* document - and always writes. The default-path store belongs to the user's
-	* harness home, and a test that never named a path must not overwrite it.
-	*/
 	allowWrite;
-	constructor(filePath = sessionCacheStatePath(), maxEntries = 64) {
-		this.filePath = filePath;
-		this.maxEntries = maxEntries;
-		this.allowWrite = filePath !== sessionCacheStatePath() || persistenceEnabled();
+	normalizeRecord;
+	static writeSequence = 0;
+	constructor(options) {
+		this.filePath = options.filePath;
+		this.maxEntries = options.maxEntries ?? 64;
+		this.allowWrite = options.allowWrite ?? true;
+		this.normalizeRecord = options.normalizeRecord;
 	}
-	/**
-	* The record for one conversation, or the most recently active one.
-	*
-	* The fallback is what makes the status card useful without a session named:
-	* the card describes the line, and the only conversation it can honestly
-	* describe in that case is the one that ran last.
-	* @param sessionId - Conversation to read; omit for the most recent one.
-	* @returns the record, or undefined when nothing has ever been recorded.
-	*/
 	read(sessionId) {
 		const sessions = this.load().sessions;
 		const named = sessionId?.trim();
@@ -36378,80 +37017,59 @@ var SessionCacheStateStore = class {
 		for (const record of Object.values(sessions)) if (newest === void 0 || record.updatedAt > newest.updatedAt) newest = record;
 		return newest;
 	}
-	/**
-	* File one conversation's state and schedule the document write.
-	* @param sessionId - Conversation the record belongs to; ignored when blank.
-	* @param record - Complete record for that conversation.
-	*/
 	write(sessionId, record) {
 		const key = sessionId.trim();
 		if (key === "") return;
 		const document = this.load();
 		document.sessions[key] = record;
-		this.prune(document);
+		const entries = Object.entries(document.sessions);
+		if (entries.length > this.maxEntries) {
+			entries.sort((a, b) => b[1].updatedAt - a[1].updatedAt);
+			document.sessions = Object.fromEntries(entries.slice(0, this.maxEntries));
+		}
 		this.schedule();
 	}
-	/**
-	* Update the warmup consecutive-hit streak on an already-written session record
-	* in place, without issuing a second `write()` call.
-	*/
-	noteWarmupStreak(sessionId, consecutiveHits) {
+	/** Update an existing record; absence is not created implicitly. */
+	update(sessionId, update) {
 		const key = sessionId.trim();
-		if (key === "") return;
+		if (key === "") return false;
 		const document = this.load();
 		const existing = document.sessions[key];
-		if (existing === void 0) return;
-		const normalized = count$1(consecutiveHits);
-		if (normalized > 0) document.sessions[key] = {
-			...existing,
-			warmupConsecutiveHits: normalized
-		};
-		else if (existing.warmupConsecutiveHits !== void 0) {
-			const { warmupConsecutiveHits: _dropped, ...rest } = existing;
-			document.sessions[key] = rest;
-		} else return;
+		if (existing === void 0) return false;
+		const next = update(existing);
+		if (next === existing) return true;
+		document.sessions[key] = next;
 		this.schedule();
+		return true;
 	}
-	/**
-	* Write the document now, synchronously.
-	*
-	* Exposed for tests and for a caller that is about to lose the process; the
-	* debounced path is what normal turns use.
-	*/
 	flush() {
 		this.cancelPending();
-		const document = this.document;
-		if (document === void 0) return;
-		this.persist(document);
+		if (this.document !== void 0) this.persist(this.document);
 	}
-	/** Drop everything older than the newest maxEntries conversations. */
-	prune(document) {
-		const entries = Object.entries(document.sessions);
-		if (entries.length <= this.maxEntries) return;
-		entries.sort((a, b) => b[1].updatedAt - a[1].updatedAt);
-		document.sessions = Object.fromEntries(entries.slice(0, this.maxEntries));
-	}
-	/** Read the document once; a missing or corrupt file reads as empty. */
 	load() {
 		if (this.document !== void 0) return this.document;
+		const sessions = {};
 		try {
 			const parsed = JSON.parse(fs$1.readFileSync(this.filePath, "utf8"));
-			this.document = normalizeDocument(parsed);
-		} catch {
-			this.document = {
-				version: 1,
-				sessions: {}
-			};
-		}
+			if (typeof parsed === "object" && parsed !== null && !Array.isArray(parsed)) {
+				const rawSessions = parsed.sessions;
+				if (typeof rawSessions === "object" && rawSessions !== null && !Array.isArray(rawSessions)) for (const [key, entry] of Object.entries(rawSessions)) {
+					const record = this.normalizeRecord(entry);
+					if (record !== void 0) sessions[key] = record;
+				}
+			}
+		} catch {}
+		this.document = {
+			version: 1,
+			sessions
+		};
 		return this.document;
 	}
-	/** Coalesce the writes one turn's several fields would otherwise cause. */
 	schedule() {
 		if (!this.allowWrite || this.pending !== void 0) return;
 		this.pending = setTimeout(() => {
 			this.pending = void 0;
-			const document = this.document;
-			if (document !== void 0) this.persist(document);
+			if (this.document !== void 0) this.persist(this.document);
 		}, 250);
 		this.pending.unref?.();
 	}
@@ -36460,35 +37078,69 @@ var SessionCacheStateStore = class {
 		clearTimeout(this.pending);
 		this.pending = void 0;
 	}
-	/** Write through a temporary file so a crash cannot leave half a document. */
 	persist(document) {
 		if (!this.allowWrite) return;
 		try {
 			fs$1.mkdirSync(path.dirname(this.filePath), { recursive: true });
-			const temporary = this.filePath + "." + process.pid + "." + ++writeSequence + ".tmp";
+			const temporary = this.filePath + "." + process.pid + "." + ++CommonSessionCacheStateStore.writeSequence + ".tmp";
 			fs$1.writeFileSync(temporary, JSON.stringify(document), "utf8");
 			fs$1.renameSync(temporary, this.filePath);
 		} catch {}
 	}
 };
-/** Validate one parsed document, dropping entries that do not describe a record. */
-function normalizeDocument(value) {
-	const sessions = {};
-	if (isRecord$52(value) && isRecord$52(value.sessions)) for (const [key, entry] of Object.entries(value.sessions)) {
-		const record = normalizeRecord(entry);
-		if (record !== void 0) sessions[key] = record;
-	}
-	return {
-		version: 1,
-		sessions
-	};
+/** Where the document lives; beside the other plugin-owned JSON. */
+function sessionCacheStatePath() {
+	return path.join(dshHomeDir(), "storages", "dsh-chatgpt-subscription-kimi-cache.json");
 }
-function normalizeRecord(value) {
+/**
+* Whether a scheduled write should reach the disk.
+*
+* A unit test that exercises the request path must not overwrite the state file
+* of the machine running it; tests build their own store against a temporary
+* path (which always writes) and call {@link SessionCacheStateStore.flush}.
+*/
+function persistenceEnabled$2() {
+	return process.env.VITEST !== "true" && process.env.NODE_ENV !== "test";
+}
+/**
+* Per-session cache state held in one JSON document.
+*
+* Reads are synchronous because every caller is answering a status request or
+* closing a turn, both of which are allowed to block on a file this size. Writes
+* are debounced: a turn reports once, and several turns in flight must not
+* rewrite the document several times.
+*/
+var SessionCacheStateStore = class extends CommonSessionCacheStateStore {
+	constructor(filePath = sessionCacheStatePath(), maxEntries = 64) {
+		super({
+			filePath,
+			maxEntries,
+			allowWrite: filePath !== sessionCacheStatePath() || persistenceEnabled$2(),
+			normalizeRecord: normalizeRecord$2
+		});
+	}
+	/** Update the warmup streak without creating a record for an unknown session. */
+	noteWarmupStreak(sessionId, consecutiveHits) {
+		this.update(sessionId, (existing) => {
+			const normalized = count$1(consecutiveHits);
+			if (normalized > 0) return {
+				...existing,
+				warmupConsecutiveHits: normalized
+			};
+			if (existing.warmupConsecutiveHits === void 0) return existing;
+			const { warmupConsecutiveHits: _dropped, ...rest } = existing;
+			return rest;
+		});
+	}
+};
+function normalizeRecord$2(value) {
 	if (!isRecord$52(value)) return void 0;
 	const lastActiveAt = finite(value.lastActiveAt);
 	const updatedAt = finite(value.updatedAt);
 	if (lastActiveAt === void 0 || updatedAt === void 0) return void 0;
 	const drift = typeof value.lastDrift === "string" && value.lastDrift !== "" ? value.lastDrift : void 0;
+	const firstRequestCachedTokens = nonNegative(value.firstRequestCachedTokens);
+	const firstRequestPromptTokens = nonNegative(value.firstRequestPromptTokens);
 	const warmupStreak = count$1(value.warmupConsecutiveHits);
 	const budgetDropped = droppedImageKeys(value.budgetDropped);
 	const lastRequestImageBytes = nonNegative(value.lastRequestImageBytes);
@@ -36505,6 +37157,8 @@ function normalizeRecord(value) {
 		cacheWrite5mTokens: count$1(value.cacheWrite5mTokens),
 		cacheWrite1hTokens: count$1(value.cacheWrite1hTokens),
 		...drift === void 0 ? {} : { lastDrift: drift },
+		...firstRequestCachedTokens === void 0 ? {} : { firstRequestCachedTokens },
+		...firstRequestPromptTokens === void 0 ? {} : { firstRequestPromptTokens },
 		...warmupStreak > 0 ? { warmupConsecutiveHits: warmupStreak } : {},
 		...budgetDropped.length > 0 ? { budgetDropped } : {},
 		...lastRequestImageBytes === void 0 ? {} : { lastRequestImageBytes },
@@ -36583,6 +37237,119 @@ function count$1(value) {
 	return number === void 0 || number < 0 ? 0 : Math.floor(number);
 }
 //#endregion
+//#region src/host/common/hosted-media.ts
+/** Default bounded LRU; providers with durable caches can supply their own adapter. */
+var MemoryHostedMediaCache = class {
+	maxEntries;
+	entries = /* @__PURE__ */ new Map();
+	constructor(maxEntries = 256) {
+		this.maxEntries = maxEntries;
+	}
+	get size() {
+		return this.entries.size;
+	}
+	get(key) {
+		const value = this.entries.get(key);
+		if (value === void 0) return void 0;
+		this.entries.delete(key);
+		this.entries.set(key, value);
+		return value;
+	}
+	set(key, value) {
+		this.entries.delete(key);
+		this.entries.set(key, value);
+		while (this.entries.size > this.maxEntries) {
+			const oldest = this.entries.keys().next();
+			if (oldest.done !== true) this.entries.delete(oldest.value);
+		}
+	}
+	delete(key) {
+		return this.entries.delete(key);
+	}
+	clear() {
+		this.entries.clear();
+	}
+};
+var HostedMediaUploadCoordinator = class {
+	cache;
+	flights = /* @__PURE__ */ new Map();
+	constructor(options = {}) {
+		this.cache = options.cache ?? new MemoryHostedMediaCache(options.maxCacheEntries);
+	}
+	get size() {
+		return this.cache.size ?? 0;
+	}
+	/** Upload once per key; one caller aborts only its own wait. */
+	async upload(key, callerSignal, start) {
+		if (callerSignal?.aborted === true) throw abortWithReason(callerSignal);
+		const cached = await this.cache.get(key);
+		if (cached !== void 0) return cached;
+		let flight = this.flights.get(key);
+		if (flight === void 0) {
+			const controller = new AbortController();
+			const promise = start(controller).then(async (value) => {
+				await this.cache.set(key, value);
+				return value;
+			});
+			flight = {
+				promise,
+				controller,
+				waiters: 0
+			};
+			this.flights.set(key, flight);
+			promise.catch(() => void 0);
+			promise.finally(() => {
+				if (this.flights.get(key) === flight) this.flights.delete(key);
+			}).catch(() => void 0);
+		}
+		return this.join(flight, callerSignal);
+	}
+	forget(key) {
+		return this.cache.delete(key);
+	}
+	clear() {
+		return this.cache.clear();
+	}
+	async join(flight, signal) {
+		if (signal?.aborted === true) throw abortWithReason(signal);
+		flight.waiters += 1;
+		let released = false;
+		const release = () => {
+			if (released) return;
+			released = true;
+			flight.waiters -= 1;
+			if (flight.waiters <= 0) {
+				for (const [key, candidate] of this.flights) if (candidate === flight) this.flights.delete(key);
+				flight.controller.abort();
+			}
+		};
+		if (signal === void 0) try {
+			return await flight.promise;
+		} finally {
+			release();
+		}
+		let detach;
+		const cancelled = new Promise((_resolve, reject) => {
+			const onAbort = () => reject(abortWithReason(signal));
+			detach = () => signal.removeEventListener("abort", onAbort);
+			signal.addEventListener("abort", onAbort, { once: true });
+		});
+		try {
+			return await Promise.race([flight.promise, cancelled]);
+		} finally {
+			detach?.();
+			release();
+		}
+	}
+};
+function abortWithReason(signal) {
+	const reason = signal.reason;
+	if (reason instanceof Error) return reason;
+	const error = /* @__PURE__ */ new Error("the media upload was aborted");
+	error.name = "AbortError";
+	return error;
+}
+//#endregion
 //#region src/host/kimi-code/media-files.ts
 /**
 * Kimi Code hosted-file (media upload) client.
@@ -36863,60 +37630,38 @@ function abortError(signal) {
 */
 var KimiMediaFilesUploader = class {
 	fetchFn;
-	maxCacheEntries;
 	timeoutMs;
-	/** key -> `ms://<id>`, in least-recently-used order. */
-	cache = /* @__PURE__ */ new Map();
-	flights = /* @__PURE__ */ new Map();
+	coordinator;
 	constructor(options = {}) {
 		this.fetchFn = options.fetchFn ?? fetch;
-		const cap = options.maxCacheEntries;
-		this.maxCacheEntries = cap !== void 0 && Number.isFinite(cap) && cap > 0 ? Math.floor(cap) : 256;
+		this.coordinator = new HostedMediaUploadCoordinator({ maxCacheEntries: options.maxCacheEntries });
 		const timeout = options.timeoutMs;
 		this.timeoutMs = timeout !== void 0 && Number.isFinite(timeout) ? timeout : MEDIA_UPLOAD_TIMEOUT_MS;
 	}
-	/** Cached references currently held, for a settings surface or a test. */
 	get size() {
-		return this.cache.size;
+		return this.coordinator.size;
 	}
-	/** Uploads one file and returns the `ms://<id>` reference a request cites. */
 	async upload(request) {
 		const resolved = this.resolve(request);
-		if (resolved.signal?.aborted === true) throw abortError(resolved.signal);
-		const cached = this.readCache(resolved.key);
-		if (cached !== void 0) return cached;
-		let flight = this.flights.get(resolved.key);
-		if (flight === void 0) {
-			flight = this.startFlight(resolved);
-			this.flights.set(resolved.key, flight);
-		}
-		return this.join(flight, resolved);
+		return this.coordinator.upload(resolved.key, resolved.signal, (controller) => this.performUpload(controller, { timedOut: false }, resolved));
 	}
-	/**
-	* Drop the reference one upload would use.
-	*
-	* The caller that sees a stored id REJECTED on the wire calls this so the next
-	* attempt uploads again instead of replaying the same dead reference.
-	*/
+	/** Drop a dead reference so the next attempt uploads again. */
 	forget(request) {
-		let key;
 		try {
-			key = mediaFileCacheKey({
+			const key = mediaFileCacheKey({
 				accountScope: request.accountScope,
 				baseUrl: request.baseUrl,
 				mediaType: request.mediaType,
 				data: request.data
 			});
+			return this.coordinator.forget(key);
 		} catch {
 			return false;
 		}
-		return this.cache.delete(key);
 	}
-	/** Forget everything; a sign-out or a test starting over. */
 	clear() {
-		this.cache.clear();
+		this.coordinator.clear();
 	}
-	/** Validate and normalize one request, or throw before anything is sent. */
 	resolve(request) {
 		const accessToken = request.accessToken.trim();
 		if (accessToken === "") throw new KimiMediaUploadError("invalid-request", "the media upload needs an access token");
@@ -36947,77 +37692,6 @@ var KimiMediaFilesUploader = class {
 			})
 		};
 	}
-	readCache(key) {
-		const ref = this.cache.get(key);
-		if (ref === void 0) return void 0;
-		this.cache.delete(key);
-		this.cache.set(key, ref);
-		return ref;
-	}
-	writeCache(key, ref) {
-		this.cache.delete(key);
-		this.cache.set(key, ref);
-		while (this.cache.size > this.maxCacheEntries) {
-			const oldest = this.cache.keys().next();
-			if (oldest.done === true) break;
-			this.cache.delete(oldest.value);
-		}
-	}
-	/** Open one flight and run it; the map entry is dropped as it settles. */
-	startFlight(request) {
-		const controller = new AbortController();
-		const state = { timedOut: false };
-		const promise = this.performUpload(controller, state, request).finally(() => {
-			if (this.flights.get(request.key)?.promise === promise) this.flights.delete(request.key);
-		});
-		promise.catch(() => void 0);
-		return {
-			promise,
-			controller,
-			waiters: 0,
-			state
-		};
-	}
-	/**
-	* Wait for one flight, respecting THIS caller's abort.
-	*
-	* The flight is not bound to the caller's signal: a caller that aborts
-	* detaches, and the transport is cancelled only when the last waiter has gone.
-	*/
-	async join(flight, request) {
-		const signal = request.signal;
-		if (signal?.aborted === true) throw abortError(signal);
-		flight.waiters += 1;
-		let released = false;
-		const release = () => {
-			if (released) return;
-			released = true;
-			flight.waiters -= 1;
-			if (flight.waiters <= 0) {
-				if (this.flights.get(request.key) === flight) this.flights.delete(request.key);
-				flight.controller.abort();
-			}
-		};
-		if (signal === void 0) try {
-			return await flight.promise;
-		} finally {
-			release();
-		}
-		const callerSignal = signal;
-		let detach;
-		const cancelled = new Promise((_resolve, reject) => {
-			const onAbort = () => reject(abortError(callerSignal));
-			detach = () => callerSignal.removeEventListener("abort", onAbort);
-			callerSignal.addEventListener("abort", onAbort, { once: true });
-		});
-		try {
-			return await Promise.race([flight.promise, cancelled]);
-		} finally {
-			detach?.();
-			release();
-		}
-	}
-	/** The one request a flight performs. */
 	async performUpload(controller, state, request) {
 		const form = new FormData();
 		form.append("file", new Blob([request.data], { type: request.mediaType }), request.filename);
@@ -37054,7 +37728,6 @@ var KimiMediaFilesUploader = class {
 			if (timer !== void 0) clearTimeout(timer);
 		}
 	}
-	/** Turn a transport rejection into the error the caller must see. */
 	transportFailure(error, state, request) {
 		if (looksAborted(error)) {
 			if (state.timedOut) return new KimiMediaUploadError("timeout", "the media upload did not finish within " + request.timeoutMs + " ms");
@@ -37062,7 +37735,6 @@ var KimiMediaFilesUploader = class {
 		}
 		return new KimiMediaUploadError("network", "the media upload did not complete: " + redact(error instanceof Error ? error.message : String(error), request.accessToken));
 	}
-	/** Read one response into a reference, or into a typed failure. */
 	settle(response, raw, request) {
 		const status = response.status;
 		if (response.ok !== true) {
@@ -37075,7 +37747,7 @@ var KimiMediaFilesUploader = class {
 				status,
 				detail
 			});
-			if (status === 413) throw new KimiMediaUploadError("too-large", "the Kimi Code hosted-file endpoint refused the upload as too large (HTTP 413)", {
+			if (status === 413) throw new KimiMediaUploadError("too-large", "the Kimi Code hosted-file endpoint refused the upload as too large (HTTP " + status + ")", {
 				status,
 				detail
 			});
@@ -37083,7 +37755,7 @@ var KimiMediaFilesUploader = class {
 				status,
 				detail
 			});
-			throw new KimiMediaUploadError("rejected", "the Kimi Code hosted-file endpoint refused the upload (HTTP " + status + ")", {
+			throw new KimiMediaUploadError("rejected", "the Kimi Code hosted-file endpoint rejected the upload (HTTP " + status + ")", {
 				status,
 				detail
 			});
@@ -37094,13 +37766,11 @@ var KimiMediaFilesUploader = class {
 			detail: redact(raw, request.accessToken)
 		});
 		const id = readFileId$1(parsed.value);
-		if (id === void 0) throw new KimiMediaUploadError("no-id", "the media upload answered HTTP " + status + " without a usable file id", {
+		if (id === void 0) throw new KimiMediaUploadError("no-id", "the media upload answered HTTP " + status + " without a usable id", {
 			status,
 			detail: redact(raw, request.accessToken)
 		});
-		const ref = fileReference(id);
-		this.writeCache(request.key, ref);
-		return ref;
+		return fileReference(id);
 	}
 };
 /** One uploader with the default transport and cache size. */
@@ -37108,6 +37778,118 @@ function createMediaFilesUploader(options = {}) {
 	return new KimiMediaFilesUploader(options);
 }
 createMediaFilesUploader();
+function clampShare(value) {
+	if (!Number.isFinite(value)) return 0;
+	return Math.min(1, Math.max(0, value));
+}
+/** Tightest measured window; unknown remains null rather than pretending full/empty. */
+function remainingQuotaShare$1(quota) {
+	if (quota === null || quota === void 0 || quota.windows.length === 0) return null;
+	let tightest = null;
+	for (const window of quota.windows) {
+		const used = typeof window.usedFraction === "number" ? window.usedFraction : typeof window.usedPercent === "number" ? window.usedPercent / 100 : void 0;
+		if (used === void 0 || !Number.isFinite(used)) continue;
+		const remaining = clampShare(1 - used);
+		tightest = tightest === null ? remaining : Math.min(tightest, remaining);
+	}
+	return tightest;
+}
+/** Warm and sufficiently-funded accounts first; no preference is an empty list. */
+function orderAffinityCandidates$1(input) {
+	const floor = input.floorShare ?? .2;
+	const measured = [];
+	const unmeasured = [];
+	for (const candidate of input.candidates) {
+		if (!candidate.warm) continue;
+		const share = input.remainingShareById?.[candidate.id];
+		if (share === null || share === void 0 || !Number.isFinite(share)) {
+			unmeasured.push(candidate.id);
+			continue;
+		}
+		if (share < floor) continue;
+		measured.push({
+			id: candidate.id,
+			share
+		});
+	}
+	measured.sort((left, right) => right.share - left.share);
+	return [...measured.map((entry) => entry.id), ...unmeasured];
+}
+//#endregion
+//#region src/host/kimi-code/cache-affinity.ts
+/**
+* The share of its tightest window an account has left, or null when unknown.
+*
+* The TIGHTEST window is the only honest answer: a pool row can carry several
+* at once, and an account with 90% of a weekly window but 2% of its five-hour
+* one is about to be refused. A window whose consumption was never measured
+* is skipped rather than read as empty or full, and a snapshot with no measured
+* window at all answers null — "unknown", which the caller treats as "not
+* provably short" instead of blocking the account on a number nobody measured.
+*
+* @param quota - The account's newest quota snapshot, when one was ever read.
+* @returns the smallest remaining fraction in `[0, 1]`, or null when unknown.
+*/
+function remainingQuotaShare(quota) {
+	return remainingQuotaShare$1(quota);
+}
+/** Settings-card description for the cache-affinity switch. */
+const CACHE_AFFINITY_SETTING_DESCRIPTION = "多账号轮换时优先选「这条前缀的缓存还热着」且额度充足的账号；默认用于 Kimi。其他线路需在 promptCacheOverrides 中证明 scope:\"account\" 并显式启用；关闭后回到原轮换策略。";
+/**
+* Environment override for the affinity switch.
+*
+* The same contract as the wait-time warmup switch beside it in settings: an
+* explicit value wins over the stored setting, so a deployment can pin the
+* behavioural change without editing preferences on disk.
+*/
+const CACHE_AFFINITY_ENV = "DSH_KIMI_CODE_CACHE_AFFINITY";
+/**
+* Resolve whether cache-affinity routing is enabled.
+*
+* Enabled by default. `0`/`false`/`no`/`off` and `1`/`true`/`yes`/`on` override
+* the stored setting; otherwise the live `cacheAffinity` setting (default
+* `true`) governs. Anything unreadable falls back to enabled rather than
+* throwing inside a request path.
+*
+* @param config - Plugin config carrying the live setting.
+* @param env - Environment to consult; `process.env` in production.
+* @returns whether a request may state an account preference.
+*/
+function resolveCacheAffinityEnabled(config, env = process.env) {
+	const raw = env[CACHE_AFFINITY_ENV]?.trim().toLowerCase();
+	if (raw === "0" || raw === "false" || raw === "no" || raw === "off") return false;
+	if (raw === "1" || raw === "true" || raw === "yes" || raw === "on") return true;
+	const field = config?.cacheAffinity;
+	if (field === void 0 || field === null) return true;
+	if (typeof field === "object" && typeof field.get === "function") try {
+		return field.get() !== false;
+	} catch {
+		return true;
+	}
+	return field !== false;
+}
+/**
+* Order the accounts this request would rather use, warmest and safest first.
+*
+* Only accounts whose prefix is warm AND whose remaining share is either
+* unknown or at least the floor are returned: a warm account that is nearly
+* spent is deliberately absent, because the pool then falls back to its own
+* rotation instead of spending the last of a window on a discount. An empty
+* result means "no preference" — the caller passes no hint at all and the
+* strategy behaves exactly as it would have without this module.
+*
+* Order: a warm account with a known-and-sufficient share first (largest
+* remaining first, so two warm accounts spread the load instead of always
+* spending one), then a warm account whose share nobody measured. Unknown last
+* because a measured account is a better bet than an unmeasured one, but still
+* listed: an unmeasured warm account is not evidence of anything bad.
+*
+* @param input - candidates plus the quota evidence behind them.
+* @returns account ids in preference order; empty when nothing is preferred.
+*/
+function orderAffinityCandidates(input) {
+	return orderAffinityCandidates$1(input);
+}
 //#endregion
 //#region src/host/kimi-code/adapter.ts
 /**
@@ -37592,9 +38374,13 @@ const BILLING_CYCLE_PATTERN = /usage limit for this billing cycle/i;
 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$5), MAX_ACCOUNT_LIMIT_COOLDOWN_MS);
-	return BILLING_CYCLE_PATTERN.test(bodyText) ? BILLING_CYCLE_COOLDOWN_MS : ACCOUNT_LIMIT_COOLDOWN_MS;
+	return cooldownFromReset({
+		resetAtMs: field !== null ? parseTimestamp(Number(field[1])) : instant !== null ? Date.parse(instant[0].replace(" ", "T")) : null,
+		now,
+		fallbackMs: BILLING_CYCLE_PATTERN.test(bodyText) ? BILLING_CYCLE_COOLDOWN_MS : ACCOUNT_LIMIT_COOLDOWN_MS,
+		minimumMs: POOL_COOLDOWN_MS$5,
+		maximumMs: MAX_ACCOUNT_LIMIT_COOLDOWN_MS
+	});
 }
 /**
 * Cooldown one rate-limited account takes when the provider states no delay.
@@ -37673,122 +38459,30 @@ function toolUpdateMode$2(modelId, catalog) {
 	].includes(raw)) return void 0;
 	return "in-history";
 }
-/**
-* How long a cold-start follower waits for the leader warming the same prefix.
-*
-* The leader only has to reach the response headers: past that the service is
-* already processing the prefix, and a follower that waited out the whole
-* budget would have lost more than the shared warm-up was worth.
-*/
-const COLD_START_WARMUP_WAIT_MS = 2500;
-/** How many prefix groups are remembered as warm. */
-const MAX_WARM_PREFIX_GROUPS = 256;
-/**
-* When each prefix group was last warmed.
-*
-* An entry older than the cache lifetime is treated as cold, because the
-* service's own entry is gone by then too. Bounded, so a long-lived process
-* cannot accumulate one entry per distinct prefix forever.
-*/
-const warmPrefixTimestamps = /* @__PURE__ */ new Map();
-/**
-* The request currently warming each prefix group.
-*
-* The promise is resolved when that leader reaches the response headers, which
-* is what releases the followers waiting behind it.
-*/
-const inFlightColdLeaders = /* @__PURE__ */ new Map();
-/** When each leader started, so a hung leader does not own its group forever. */
-const inFlightColdLeaderStartedAt = /* @__PURE__ */ new Map();
-/**
-* The resolver belonging to each in-flight leader.
-*
-* Kept beside the promise it settles, because a promise cannot be resolved
-* from the outside.
-*/
-const coldLeaderReleases = /* @__PURE__ */ new Map();
-/**
-* In-flight main-session requests waiting for their HTTP response headers,
-* keyed by sessionId. Used so a concurrent `session-title-first-prompt-llm`
-* side-channel request yields to the main turn's prefill instead of competing
-* on the same account during cold start.
-*/
-const inFlightMainSessionHeaders = /* @__PURE__ */ new Map();
-const mainSessionHeaderReleases = /* @__PURE__ */ new Map();
-/**
-* Per-account/model first-turn prefill pacing chain for concurrent subagent bursts.
-*/
-const inFlightSubagentBurstLeaders = /* @__PURE__ */ new Map();
-const subagentBurstReleases = /* @__PURE__ */ new Map();
-/** The gate of a request that neither waits for nor leads a cold start. */
-const NO_COLD_START_GATE = {
-	isLeader: false,
-	release: () => void 0,
-	markWarm: () => void 0
-};
-/**
-* The group of requests that address the same cached prefix.
-*
-* The service derives its cache from the request content, so two requests with
-* the same model, the same system prompt, the same tool list and the same
-* leading static bootstrap reminders hit the same entry. The region is part of
-* the key because each region is a separate service with a cache of its own.
-*
-* A request with no system prompt has no prefix worth warming, and gets no
-* group at all.
-*
-* @param region - The region the request will be answered from.
-* @param options - The request about to be sent.
-* @returns The group key, or undefined when the request has no prefix.
-*/
-function coldStartGroupKey(region, options) {
+const prefixRequests = new PrefixRequestCoordinator({ waitMs: 2500 });
+/** The prefix identity shared by requests on one region/model. */
+function prefixGroupKey(region, options) {
 	const { systemPromptHash, toolsHash, staticPrefixHash } = prefixFingerprints(options);
 	if (systemPromptHash === null) return void 0;
 	return `${region}:${options.model}:${systemPromptHash}:${toolsHash ?? "none"}${staticPrefixHash ? ":" + staticPrefixHash : ""}`;
 }
-/**
-* Remember that a prefix group is warm.
-*
-* @param groupKey - The group that was just warmed.
-*/
-function rememberWarmPrefix(groupKey, now = Date.now()) {
-	warmPrefixTimestamps.delete(groupKey);
-	warmPrefixTimestamps.set(groupKey, now);
-	while (warmPrefixTimestamps.size > MAX_WARM_PREFIX_GROUPS) {
-		const oldest = warmPrefixTimestamps.keys().next();
-		if (oldest.done === true) break;
-		warmPrefixTimestamps.delete(oldest.value);
-	}
+const GROUP_ACCOUNT_SEPARATOR = "|";
+/** The cache group on one account; cross-account warmth is never inferred. */
+function coldStartGroupKey(region, options, accountScope) {
+	const prefix = prefixGroupKey(region, options);
+	return prefix === void 0 ? void 0 : `${prefix}${GROUP_ACCOUNT_SEPARATOR}${accountScope}`;
 }
-/**
-* Register an in-flight main-session request until its HTTP response headers arrive.
-*/
+/** Whether this account's prefix is still warm under the measured Kimi TTL. */
+function isWarmPrefixGroup(region, options, accountScope, now = Date.now()) {
+	const groupKey = coldStartGroupKey(region, options, accountScope);
+	return groupKey !== void 0 && prefixRequests.isWarm(groupKey, 15e5, now);
+}
+/** Register a main request until response headers arrive. */
 function registerInFlightMainSessionHeader(sessionId) {
-	const key = sessionId.trim();
-	if (key === "") return () => void 0;
-	mainSessionHeaderReleases.get(key)?.();
-	let resolveHeader;
-	const promise = new Promise((resolve) => {
-		resolveHeader = resolve;
-	});
-	inFlightMainSessionHeaders.set(key, promise);
-	let released = false;
-	const release = () => {
-		if (released) return;
-		released = true;
-		resolveHeader();
-		if (mainSessionHeaderReleases.get(key) === release) {
-			mainSessionHeaderReleases.delete(key);
-			inFlightMainSessionHeaders.delete(key);
-		}
-	};
-	mainSessionHeaderReleases.set(key, release);
-	return release;
+	const lease = prefixRequests.registerMainSession(sessionId);
+	return () => lease.release();
 }
-/**
-* Whether a first-turn request carries a subagent mandate or large uncached
-* dynamic payload that should stagger behind an already in-flight sibling burst.
-*/
+/** Whether a first-turn request carries a mandate or a large dynamic payload. */
 function isSubagentBurstCandidate(options) {
 	let dynamicChars = 0;
 	for (const message of options.messages) {
@@ -37804,113 +38498,14 @@ function isSubagentBurstCandidate(options) {
 	}
 	return dynamicChars >= 1200;
 }
-/**
-* Stagger concurrent subagent first-turn requests on the same account/model
-* so they do not stampede prefill in the same millisecond.
-*/
+/** Serialize concurrent subagent first turns on one account/model. */
 async function acquireSubagentBurstPacing(pacingKey, signal) {
-	while (true) {
-		const existing = inFlightSubagentBurstLeaders.get(pacingKey);
-		if (existing === void 0) break;
-		await raceColdStartLeader(existing, signal);
-		if (signal.aborted) return () => void 0;
-		if (inFlightSubagentBurstLeaders.get(pacingKey) === existing) {
-			subagentBurstReleases.get(pacingKey)?.();
-			break;
-		}
-	}
-	let resolvePacing;
-	const promise = new Promise((resolve) => {
-		resolvePacing = resolve;
-	});
-	inFlightSubagentBurstLeaders.set(pacingKey, promise);
-	let settled = false;
-	const release = () => {
-		if (settled) return;
-		settled = true;
-		resolvePacing();
-		if (subagentBurstReleases.get(pacingKey) === release) {
-			subagentBurstReleases.delete(pacingKey);
-			inFlightSubagentBurstLeaders.delete(pacingKey);
-		}
-	};
-	subagentBurstReleases.set(pacingKey, release);
-	return release;
-}
-/**
-* Wait for the leader of a prefix group, for the wait budget, or for an abort.
-*
-* The timer and the abort listener are removed whichever way the race ends, so
-* a follower that proceeds leaves neither behind.
-*
-* @param leader - The promise the in-flight leader settles.
-* @param signal - The caller's abort signal; an abort ends the wait.
-*/
-function raceColdStartLeader(leader, signal) {
-	const cleanups = [];
-	const budget = new Promise((resolve) => {
-		const timer = setTimeout(() => resolve(), COLD_START_WARMUP_WAIT_MS);
-		cleanups.push(() => clearTimeout(timer));
-	});
-	const aborted = new Promise((resolve) => {
-		if (signal.aborted) {
-			resolve();
-			return;
-		}
-		const onAbort = () => resolve();
-		signal.addEventListener("abort", onAbort, { once: true });
-		cleanups.push(() => signal.removeEventListener("abort", onAbort));
-	});
-	return Promise.race([
-		leader,
-		budget,
-		aborted
-	]).finally(() => {
-		for (const cleanup of cleanups) cleanup();
-	});
+	const lease = await prefixRequests.acquireBurst(pacingKey, signal);
+	return () => lease.release();
 }
-/**
-* Join the cold-start singleflight for one prefix group.
-*
-* A warm group returns at once. A cold group that already has a leader waits
-* for that leader — but never longer than {@link COLD_START_WARMUP_WAIT_MS}.
-* A cold group with no leader becomes the leader itself.
-*
-* @param groupKey - The prefix group this request belongs to.
-* @param signal - The caller's abort signal; an abort ends the wait.
-* @returns The gate this request holds.
-*/
+/** Join the shared cold-start coordinator with Kimi's measured warm TTL. */
 async function acquireColdStartWarmup(groupKey, signal) {
-	const warmedAt = warmPrefixTimestamps.get(groupKey);
-	if (warmedAt !== void 0 && Date.now() - warmedAt < 15e5) return NO_COLD_START_GATE;
-	if (inFlightColdLeaders.get(groupKey) !== void 0) {
-		const startedAt = inFlightColdLeaderStartedAt.get(groupKey);
-		if (startedAt !== void 0 && Date.now() - startedAt >= 2500) coldLeaderReleases.get(groupKey)?.();
-	}
-	const leader = inFlightColdLeaders.get(groupKey);
-	if (leader !== void 0) {
-		await raceColdStartLeader(leader, signal);
-		return NO_COLD_START_GATE;
-	}
-	let resolveLeader;
-	const promise = new Promise((resolve) => {
-		resolveLeader = resolve;
-	});
-	inFlightColdLeaders.set(groupKey, promise);
-	inFlightColdLeaderStartedAt.set(groupKey, Date.now());
-	const release = () => {
-		resolveLeader();
-		if (coldLeaderReleases.get(groupKey) !== release) return;
-		coldLeaderReleases.delete(groupKey);
-		inFlightColdLeaders.delete(groupKey);
-		inFlightColdLeaderStartedAt.delete(groupKey);
-	};
-	coldLeaderReleases.set(groupKey, release);
-	return {
-		isLeader: true,
-		release,
-		markWarm: () => rememberWarmPrefix(groupKey)
-	};
+	return prefixRequests.acquire(groupKey, signal, CACHE_TTL_MS);
 }
 /** Sessions already told that their prefix broke, so the log stays readable. */
 const driftWarned = /* @__PURE__ */ new Set();
@@ -37944,6 +38539,13 @@ function noteSessionCacheState(sessionId, state) {
 	const liveDropped = imageBudget?.droppedImages ?? 0;
 	const lastRequestImageBytes = liveBytes > 0 ? liveBytes : priorRecord?.lastRequestImageBytes;
 	const lastDroppedImages = liveDropped > 0 ? liveDropped : priorRecord?.lastDroppedImages;
+	const firstRequest = stats.requests === 1 ? {
+		firstRequestCachedTokens: state.cacheReadTokens,
+		firstRequestPromptTokens: state.cacheReadTokens + state.inputTokens
+	} : {
+		...priorRecord?.firstRequestCachedTokens === void 0 ? {} : { firstRequestCachedTokens: priorRecord.firstRequestCachedTokens },
+		...priorRecord?.firstRequestPromptTokens === void 0 ? {} : { firstRequestPromptTokens: priorRecord.firstRequestPromptTokens }
+	};
 	sessionCacheState.write(sessionId, {
 		lastActiveAt: now,
 		totalTokens: state.cacheReadTokens + state.inputTokens,
@@ -37958,10 +38560,12 @@ function noteSessionCacheState(sessionId, state) {
 		...budgetDropped === void 0 ? {} : { budgetDropped },
 		...lastRequestImageBytes === void 0 ? {} : { lastRequestImageBytes },
 		...lastDroppedImages === void 0 ? {} : { lastDroppedImages },
+		...firstRequest,
 		updatedAt: now
 	});
 	const warmupState = recordCacheWarmupUsage({
 		sessionId,
+		provider: PROVIDER_ID$9,
 		inputTokens: state.inputTokens,
 		cacheReadTokens: state.cacheReadTokens,
 		outputTokens: state.outputTokens,
@@ -38078,6 +38682,70 @@ var KimiCodeAdapter = class extends LlmAdapter {
 	settings() {
 		return this.preferences ? Promise.resolve(this.preferences.status()) : this.modelSettings.read();
 	}
+	/** Whether this deployment routes by warm prefix cache; on unless switched off. */
+	cacheAffinityEnabled() {
+		try {
+			return this.options.cacheAffinity?.() !== false;
+		} catch {
+			return true;
+		}
+	}
+	/**
+	* Which accounts this request would rather be served by, when the prefix cache
+	* can say something useful about it.
+	*
+	* The service keeps one prefix cache per account, so rotation — which knows
+	* cooldowns, tokens and last use, never prompt content — can spend a request on
+	* an account that has never seen this prompt while another one has it hot. Both
+	* facts needed to do better are already in this process: the cold-start gate
+	* records which (prefix, account) pairs are warm, and the quota cache says how
+	* much each account has left.
+	*
+	* Returns undefined — no preference at all — wherever a preference would be a
+	* guess: the switch is off, there is no pool or only one account, the pool
+	* document cannot be read, a region cannot be resolved, or no account has this
+	* prefix warm. The pool then applies its rotation strategy exactly as it would
+	* have without this method.
+	*
+	* @param pool - The account pool serving this route, or null when none is installed.
+	* @param options - The request about to be sent; its system prompt, tools and
+	*   static bootstrap are what decide the prefix identity.
+	* @returns a preference for the pool, or undefined for "strategy decides".
+	*/
+	async cacheAffinityHint(pool, options) {
+		if (pool === null || !this.cacheAffinityEnabled()) return void 0;
+		let scopes;
+		try {
+			scopes = await pool.accountScopes();
+		} catch {
+			return;
+		}
+		if (scopes.length < 2) return void 0;
+		let fallbackRegion = null;
+		if (scopes.some((scope) => scope.region === null)) try {
+			fallbackRegion = await resolveRegion();
+		} catch {
+			fallbackRegion = null;
+		}
+		const candidates = [];
+		for (const scope of scopes) {
+			const region = scope.region ?? fallbackRegion;
+			if (region === null) continue;
+			if (!isWarmPrefixGroup(region, options, scope.id)) continue;
+			candidates.push({
+				id: scope.id,
+				warm: true
+			});
+		}
+		if (candidates.length === 0) return void 0;
+		const remainingShareById = {};
+		for (const candidate of candidates) remainingShareById[candidate.id] = remainingQuotaShare(cachedQuotaFor$1(candidate.id));
+		const preferAccountIds = orderAffinityCandidates({
+			candidates,
+			remainingShareById
+		});
+		return preferAccountIds.length === 0 ? void 0 : { preferAccountIds };
+	}
 	/**
 	* Catalog for the picker: the live managed listing when reachable, the
 	* shipped fallback otherwise, narrowed by the user's enabled selection.
@@ -38182,7 +38850,7 @@ var KimiCodeAdapter = class extends LlmAdapter {
 	async *requestStream(options, signal, activity) {
 		const fetchFn = this.options.fetchFn ?? fetch;
 		const normalized = normalizeGenerateOptions(options);
-		const sideChannel = isSideChannelTitleRequest(normalized);
+		const sideChannel = isAuxiliaryRequest(normalized) || isSideChannelTitleRequest(normalized);
 		const sessionKey = typeof options.sessionId === "string" ? options.sessionId.trim() : "";
 		const mainHeaderRelease = !sideChannel && sessionKey !== "" ? registerInFlightMainSessionHeader(sessionKey) : () => void 0;
 		try {
@@ -38299,6 +38967,7 @@ var KimiCodeAdapter = class extends LlmAdapter {
 			let sentBuild;
 			const pool = this.accountPool;
 			const tried = /* @__PURE__ */ new Set();
+			const affinity = sideChannel ? void 0 : await this.cacheAffinityHint(pool, requestOptions);
 			let servedByAccountId;
 			let response;
 			while (true) {
@@ -38316,7 +38985,7 @@ var KimiCodeAdapter = class extends LlmAdapter {
 				else {
 					let effective;
 					try {
-						effective = await pool.getEffectiveCredential(tried, fetchFn, { maxAttempts: MODEL_REFRESH_MAX_ATTEMPTS });
+						effective = await pool.getEffectiveCredential(tried, fetchFn, { maxAttempts: MODEL_REFRESH_MAX_ATTEMPTS }, affinity);
 					} catch (error) {
 						throw credentialFailure(error, signal);
 					}
@@ -38374,13 +39043,10 @@ var KimiCodeAdapter = class extends LlmAdapter {
 					}
 				};
 				if (sideChannel && sessionKey !== "") {
-					const pendingMain = inFlightMainSessionHeaders.get(sessionKey);
-					if (pendingMain !== void 0) {
-						await raceColdStartLeader(pendingMain, signal);
-						if (signal.aborted) throw new LlmError("Kimi Code request aborted", "ABORTED");
-					}
+					await prefixRequests.waitForMainSession(sessionKey, signal);
+					if (signal.aborted) throw new LlmError("Kimi Code request aborted", "ABORTED");
 				}
-				const coldGroupKey = sideChannel ? void 0 : coldStartGroupKey(region, requestOptions);
+				const coldGroupKey = sideChannel ? void 0 : coldStartGroupKey(region, requestOptions, accountScope);
 				const coldGate = coldGroupKey === void 0 ? void 0 : await acquireColdStartWarmup(coldGroupKey, signal);
 				const burstPacingRelease = !sideChannel && sessionKey !== "" && getCacheStats(sessionKey).requests === 0 && isSubagentBurstCandidate(requestOptions) ? await acquireSubagentBurstPacing(`${accountScope}:${options.model}`, signal) : () => void 0;
 				let coldGateSettled = false;
@@ -38857,9 +39523,28 @@ var KimiCodeAccountPool = class extends AccountPoolCore {
 	*
 	* The returned credential carries its own region, OAuth host and base URL, so
 	* the caller must use those rather than the process-wide defaults.
+	*
+	* @param selection - this request's own preference, used by the cache-affinity
+	*   routing to put an account whose prefix cache is still warm first. Absent
+	*   means the rotation strategy decides alone.
+	*/
+	async getEffectiveCredential(excludeIds, fetchFn = fetch, options, selection) {
+		return this.getEffectiveAccount(excludeIds, fetchFn, false, options, selection);
+	}
+	/**
+	* Every pooled account's id with the region that decides its prefix group.
+	*
+	* The cache-affinity routing needs exactly two facts per account — which one
+	* it is, and which regional service answers it, because each region keeps its
+	* own prefix cache — and nothing else. Deliberately a narrow projection over
+	* the pool document rather than a credential read: tokens never leave the
+	* pool, and nothing here can be mistaken for a credential accessor.
 	*/
-	async getEffectiveCredential(excludeIds, fetchFn = fetch, options) {
-		return this.getEffectiveAccount(excludeIds, fetchFn, false, options);
+	async accountScopes() {
+		return (await this.read()).accounts.map((account) => ({
+			id: account.id,
+			region: account.region ?? account.credentials.region ?? null
+		}));
 	}
 	/** The pre-pool credential file this pool mirrors its primary account into. */
 	mirrorStore() {
@@ -39051,15 +39736,37 @@ function warmupStatusOf(sessionId) {
 		requiredHits: CACHE_WARMUP_DEFAULTS.consecutiveHitsRequired
 	};
 }
+/**
+* How much of this conversation's first prompt was served from cache.
+*
+* Read from the durable record on BOTH paths, because the in-memory totals carry
+* rolling sums only: the first request's own numbers exist exactly once, at the
+* turn that closed it, and the record is where the adapter filed them. A record
+* written before this field existed simply has nothing to say, which is the
+* honest answer for a conversation that started under an older build.
+*/
+function openingInheritanceOf(sessionId) {
+	const durable = sessionCacheState.read(sessionId);
+	const cachedTokens = durable?.firstRequestCachedTokens;
+	const promptTokens = durable?.firstRequestPromptTokens;
+	if (cachedTokens === void 0 || promptTokens === void 0 || promptTokens <= 0) return void 0;
+	return {
+		cachedTokens,
+		promptTokens,
+		hitRatio: Math.min(1, Math.max(0, cachedTokens / promptTokens))
+	};
+}
 function cacheStatsOrNull(sessionId) {
 	const stats = getCacheStats(sessionId);
 	if (stats.requests === 0) return durableCacheStats(sessionId);
 	const lastDrift = getLastDriftCause(sessionId);
 	const warmup = warmupStatusOf(sessionId);
+	const opening = openingInheritanceOf(sessionId);
 	return {
 		...stats,
 		...lastDrift === void 0 ? {} : { lastDrift },
 		...warmup === void 0 ? {} : { warmup },
+		...opening === void 0 ? {} : { opening },
 		imageBudget: imageBudgetOf(sessionId),
 		cacheWriteTokens: stats.cacheWriteTokens,
 		cacheWrite5mTokens: stats.cacheWrite5mTokens,
@@ -39103,6 +39810,7 @@ function durableCacheStats(sessionId) {
 	if (durable === void 0 || durable.requests === 0) return null;
 	const lastDrift = driftCauseOf(durable.lastDrift);
 	const warmup = warmupStatusOf(sessionId);
+	const opening = openingInheritanceOf(sessionId);
 	return {
 		requests: durable.requests,
 		cachedTokens: durable.cachedTokens,
@@ -39114,6 +39822,7 @@ function durableCacheStats(sessionId) {
 		cacheWrite1hTokens: durable.cacheWrite1hTokens,
 		...lastDrift === void 0 ? {} : { lastDrift },
 		...warmup === void 0 ? {} : { warmup },
+		...opening === void 0 ? {} : { opening },
 		imageBudget: imageBudgetOf(sessionId)
 	};
 }
@@ -40391,7 +41100,7 @@ const MINIMAX_CODE_PROVIDER$1 = "minimax-code";
 * @param value - 任意值。
 * @returns 是普通对象时为 true。
 */
-function isPlainRecord$1(value) {
+function isPlainRecord(value) {
 	if (value === null || typeof value !== "object" || Array.isArray(value)) return false;
 	const prototype = Object.getPrototypeOf(value);
 	return prototype === Object.prototype || prototype === null;
@@ -40511,7 +41220,7 @@ function coerceBySchema(value, schema, fixes) {
 	const type = schema["type"];
 	if (type === "array") {
 		let current = value;
-		if (isPlainRecord$1(current) && Object.keys(current).length === 1 && "item" in current) {
+		if (isPlainRecord(current) && Object.keys(current).length === 1 && "item" in current) {
 			const inner = current["item"];
 			current = Array.isArray(inner) ? [...inner] : [inner];
 			fixes.push("item_wrapper");
@@ -40543,7 +41252,7 @@ function coerceBySchema(value, schema, fixes) {
 	}
 	if (type === "object") {
 		const properties = schema["properties"];
-		if (!isPlainRecord$1(value) || !isRecord$50(properties)) return value;
+		if (!isPlainRecord(value) || !isRecord$50(properties)) return value;
 		const next = { ...value };
 		for (const [key, subSchema] of Object.entries(properties)) {
 			if (!(key in next)) continue;
@@ -40552,7 +41261,7 @@ function coerceBySchema(value, schema, fixes) {
 		return next;
 	}
 	const properties = schema["properties"];
-	if (isPlainRecord$1(value) && isRecord$50(properties)) {
+	if (isPlainRecord(value) && isRecord$50(properties)) {
 		const next = { ...value };
 		for (const [key, subSchema] of Object.entries(properties)) {
 			if (!(key in next)) continue;
@@ -40612,7 +41321,7 @@ function coerceByKnownFields(args, fixes) {
 * @returns 修复后的参数与修复项列表。
 */
 function repairMinimaxToolArguments(toolName, args, schema) {
-	if (!isPlainRecord$1(args)) return {
+	if (!isPlainRecord(args)) return {
 		value: args,
 		fixes: []
 	};
@@ -40627,7 +41336,7 @@ function repairMinimaxToolArguments(toolName, args, schema) {
 	}
 	if (schema !== void 0) {
 		const coerced = coerceBySchema(working, schema, fixes);
-		if (isPlainRecord$1(coerced)) Object.assign(working, coerced);
+		if (isPlainRecord(coerced)) Object.assign(working, coerced);
 	} else coerceByKnownFields(working, fixes);
 	return {
 		value: working,
@@ -40773,28 +41482,6 @@ function errorFamilyOf(text) {
 	return (text.split(/\r?\n/, 1)[0] ?? "").replace(/\d+/g, "#").slice(0, 200);
 }
 /**
-* 把参数规范化成可比较的字符串（键排序）。
-* @param value - 参数值。
-* @returns 规范化后的 JSON 文本。
-*/
-function canonicalize(value) {
-	return JSON.stringify(sortJsonValue$1(value));
-}
-/**
-* 深拷贝一份键排序后的 JSON 值。
-* @param value - 任意值。
-* @returns 排序后的副本。
-*/
-function sortJsonValue$1(value) {
-	if (Array.isArray(value)) return value.map(sortJsonValue$1);
-	if (isPlainRecord$1(value)) {
-		const sorted = {};
-		for (const key of Object.keys(value).sort()) sorted[key] = sortJsonValue$1(value[key]);
-		return sorted;
-	}
-	return value;
-}
-/**
 * 造一台工具级失控守卫。
 *
 * 三种可提醒信号（顺序即官方提醒优先级）：
@@ -40813,8 +41500,8 @@ function sortJsonValue$1(value) {
 */
 function createMinimaxRunawayGuard(options = {}) {
 	const threshold = options.remindAfterOccurrences ?? 3;
-	/** 同一次动作的连击。 */
-	let actionStreak;
+	/** 同一次动作的连击，由所有线路共享的规范键与跟踪器计算。 */
+	const exactActionRepeats = createExactRepeatTracker();
 	/** 同一族错误的连击（键是工具名 + 错误族）。 */
 	const errorStreaks = /* @__PURE__ */ new Map();
 	/** 轮询连击（键是工具名）。 */
@@ -40847,23 +41534,16 @@ function createMinimaxRunawayGuard(options = {}) {
 	*/
 	const isPolling = (observation) => {
 		if (!POLLING_TOOL_NAMES.includes(observation.toolName)) return false;
-		return (isPlainRecord$1(observation.arguments) ? observation.arguments : void 0)?.["wait"] !== true;
+		return (isPlainRecord(observation.arguments) ? observation.arguments : void 0)?.["wait"] !== true;
 	};
 	return {
 		observe(observation) {
 			const toolName = observation.toolName;
-			const actionKey = `${toolName}\u0000${canonicalize(observation.arguments)}`;
+			const actionKey = canonicalToolCallKey(toolName, observation.arguments);
 			const countsAsAction = !POLLING_TOOL_NAMES.includes(toolName);
 			if (countsAsAction) {
-				if (actionStreak?.key === actionKey) actionStreak = {
-					key: actionKey,
-					count: actionStreak.count + 1
-				};
-				else actionStreak = {
-					key: actionKey,
-					count: 1
-				};
-				record("exact_action_repeat", actionStreak.count, toolName);
+				const action = exactActionRepeats.observe(toolName, observation.arguments);
+				record("exact_action_repeat", action.count, toolName);
 			}
 			if (countsAsAction) {
 				recentActions.push(actionKey);
@@ -40890,7 +41570,7 @@ function createMinimaxRunawayGuard(options = {}) {
 				record("same_error_family", count, toolName);
 			} else for (const [key, entry] of [...errorStreaks]) if (entry.toolName === toolName) errorStreaks.delete(key);
 			if (isPolling(observation)) {
-				const key = `${toolName}\u0000${canonicalize(observation.arguments)}\u0000${observation.resultText ?? ""}`;
+				const key = `${canonicalToolCallKey(toolName, observation.arguments)}\u0000${observation.resultText ?? ""}`;
 				const previous = pollingStreaks.get(toolName);
 				const count = previous !== void 0 && previous.key === key ? previous.count + 1 : 1;
 				pollingStreaks.set(toolName, {
@@ -40909,6 +41589,7 @@ function createMinimaxRunawayGuard(options = {}) {
 					text: renderSameErrorFamilyReminder(errorCandidate.count)
 				};
 			}
+			const actionStreak = exactActionRepeats.current();
 			if (countsAsAction && actionStreak !== void 0 && actionStreak.count >= threshold) {
 				reminderAttempted = true;
 				return {
@@ -40928,7 +41609,7 @@ function createMinimaxRunawayGuard(options = {}) {
 			}
 		},
 		reset() {
-			actionStreak = void 0;
+			exactActionRepeats.reset();
 			errorStreaks.clear();
 			pollingStreaks.clear();
 			recentActions.length = 0;
@@ -41198,14 +41879,14 @@ function schemaOf(tools, name, agent) {
 function repairExecutionArguments(execution, tools) {
 	const record = isRecord$50(execution) ? execution : void 0;
 	const args = record?.["arguments"];
-	if (!isPlainRecord$1(args)) return;
+	if (!isPlainRecord(args)) return;
 	const name = stringOf$4(record?.["name"]);
 	if (name === void 0) return;
 	const agent = record?.["agent"];
 	const repaired = repairMinimaxToolArguments(name, args, schemaOf(tools, name, agent));
 	if (repaired.fixes.length === 0) return;
 	const target = repaired.value;
-	if (!isPlainRecord$1(target)) return;
+	if (!isPlainRecord(target)) return;
 	try {
 		for (const key of Object.keys(args)) if (!(key in target)) delete args[key];
 		Object.assign(args, target);
@@ -44699,6 +45380,62 @@ var FileIdStore = class FileIdStore {
 	/** Shared by every in-memory instance, so they see one cache. */
 	static memory;
 };
+/** Durable MiniMax cache adapted to the shared upload singleflight core. */
+var MiniMaxHostedFileCache = class {
+	store;
+	now;
+	constructor(options) {
+		this.store = new FileIdStore(options.storePath);
+		this.now = options.nowMs ?? Date.now;
+	}
+	async get(key) {
+		const cached = (await this.store.read().catch(() => ({})))[key];
+		return cached !== void 0 && cached.expiresAtMs > this.now() ? cached.fileId : void 0;
+	}
+	async set(key, fileId) {
+		const cache = await this.store.read().catch(() => ({}));
+		await this.store.write({
+			...cache,
+			[key]: {
+				fileId,
+				expiresAtMs: this.now() + DEFAULT_FILE_ID_TTL_SEC * 1e3
+			}
+		}).catch(() => void 0);
+	}
+	async delete(key) {
+		const cache = await this.store.read().catch(() => ({}));
+		const existed = cache[key] !== void 0;
+		if (existed) {
+			delete cache[key];
+			await this.store.write(cache).catch(() => void 0);
+		}
+		return existed;
+	}
+	async clear() {
+		await this.store.write({}).catch(() => void 0);
+	}
+};
+const UPLOAD_COORDINATORS = /* @__PURE__ */ new Map();
+const CLOCK_IDS = /* @__PURE__ */ new WeakMap();
+let clockSequence = 0;
+function uploadCoordinator(options) {
+	let clockKey = "";
+	if (options.nowMs !== void 0) {
+		let id = CLOCK_IDS.get(options.nowMs);
+		if (id === void 0) {
+			id = ++clockSequence;
+			CLOCK_IDS.set(options.nowMs, id);
+		}
+		clockKey = ":clock:" + String(id);
+	}
+	const key = (options.storePath ?? "<memory>") + clockKey;
+	let coordinator = UPLOAD_COORDINATORS.get(key);
+	if (coordinator === void 0) {
+		coordinator = new HostedMediaUploadCoordinator({ cache: new MiniMaxHostedFileCache(options) });
+		UPLOAD_COORDINATORS.set(key, coordinator);
+	}
+	return coordinator;
+}
 function isRecord$47(value) {
 	return typeof value === "object" && value !== null && !Array.isArray(value);
 }
@@ -44759,25 +45496,19 @@ function filesPurposeFor(mediaType) {
 *   inline instead (too large, unconfigured, or the upload failed).
 */
 async function uploadMediaFile(request, accountKey, options = {}) {
-	const purpose = filesPurposeFor(request.mediaType);
-	if (purpose === void 0) return void 0;
+	if (filesPurposeFor(request.mediaType) === void 0) return void 0;
 	if (request.bytes > 52428800) throw new FilesApiError("too-large", "the file is " + formatBytes(request.bytes) + ", above the " + formatBytes(MAX_UPLOAD_BYTES) + " upload ceiling");
 	const key = cacheKey(request.base64, accountKey);
-	const nowMs = options.nowMs ?? Date.now;
-	const store = new FileIdStore(options.storePath);
-	const cache = await store.read().catch(() => ({}));
-	const cached = cache[key];
-	if (cached !== void 0 && cached.expiresAtMs > nowMs()) return cached.fileId;
+	return uploadCoordinator(options).upload(key, request.signal, (controller) => performMediaUpload(request, options, controller));
+}
+/** One shared transport flight; cache reads/writes live in MiniMaxHostedFileCache. */
+async function performMediaUpload(request, options, controller) {
+	const purpose = filesPurposeFor(request.mediaType);
 	const fetchFn = options.fetchFn ?? fetch;
 	const form = new FormData();
 	form.append("purpose", purpose);
 	form.append("file", new Blob([Buffer.from(request.base64, "base64")], { type: request.mediaType }), filenameForMediaType(request.mediaType));
-	const controller = new AbortController();
 	const timer = setTimeout(() => controller.abort(), UPLOAD_TIMEOUT_MS);
-	const abortForTurn = () => controller.abort(request.signal?.reason);
-	if (request.signal?.aborted === true) abortForTurn();
-	else request.signal?.addEventListener("abort", abortForTurn, { once: true });
-	let fileId;
 	try {
 		const response = await fetchFn(filesUploadUrl(request.baseUrl), {
 			method: "POST",
@@ -44790,24 +45521,16 @@ async function uploadMediaFile(request, accountKey, options = {}) {
 		if (!response.ok) throw new FilesApiError("rejected", "the Files API upload failed with HTTP " + response.status);
 		const raw = await response.text().catch(() => "");
 		const payload = safeJson(raw);
-		fileId = readFileId(payload);
+		const fileId = readFileId(payload);
 		if (fileId === void 0) throw new FilesApiError("rejected", describeRejection(payload, raw));
+		return fileId;
 	} catch (error) {
 		if (error instanceof FilesApiError) throw error;
-		if (isAbort$1(error, request.signal)) throw error;
+		if (isAbort$1(error, controller.signal)) throw error;
 		throw new FilesApiError("network", "the Files API upload did not complete: " + (error instanceof Error ? error.message : String(error)));
 	} finally {
 		clearTimeout(timer);
-		request.signal?.removeEventListener("abort", abortForTurn);
 	}
-	await store.write({
-		...cache,
-		[key]: {
-			fileId,
-			expiresAtMs: nowMs() + DEFAULT_FILE_ID_TTL_SEC * 1e3
-		}
-	}).catch(() => void 0);
-	return fileId;
 }
 /** Whether one failure is a cancellation rather than a real upload failure. */
 function isAbort$1(error, signal) {
@@ -44848,6 +45571,50 @@ function formatBytes(bytes) {
 	return (bytes / (1024 * 1024)).toFixed(1) + " MB";
 }
 //#endregion
+//#region src/host/common/sticky-image-budget.ts
+/** Durable dropped-image identities for a provider's sticky budget. */
+function normalizeRecord$1(value) {
+	if (typeof value !== "object" || value === null || Array.isArray(value)) return void 0;
+	const record = value;
+	const updatedAt = record.updatedAt;
+	if (typeof updatedAt !== "number" || !Number.isFinite(updatedAt)) return void 0;
+	const keys = Array.isArray(record.budgetDropped) ? record.budgetDropped.filter((key) => typeof key === "string" && key !== "").slice(0, 512) : [];
+	return {
+		updatedAt,
+		budgetDropped: [...new Set(keys)]
+	};
+}
+function persistenceEnabled$1() {
+	return process.env.VITEST !== "true" && process.env.NODE_ENV !== "test";
+}
+var StickyImageBudgetStore = class {
+	store;
+	constructor(provider, filePath) {
+		const target = filePath ?? path.join(dshHomeDir(), "storages", `dsh-chatgpt-subscription-${provider}-cache.json`);
+		this.store = new CommonSessionCacheStateStore({
+			filePath: target,
+			allowWrite: filePath !== void 0 || persistenceEnabled$1(),
+			normalizeRecord: normalizeRecord$1
+		});
+	}
+	/** Identities already omitted in this session, including before a restart. */
+	priorDroppedKeys(sessionId) {
+		if (sessionId === void 0 || sessionId.trim() === "") return /* @__PURE__ */ new Set();
+		return new Set(this.store.read(sessionId)?.budgetDropped ?? []);
+	}
+	/** Persist the complete set, not just the newest omissions. */
+	record(sessionId, keys) {
+		if (sessionId === void 0 || sessionId.trim() === "") return;
+		this.store.write(sessionId, {
+			updatedAt: Date.now(),
+			budgetDropped: [...keys].slice(0, 512)
+		});
+	}
+	flush() {
+		this.store.flush();
+	}
+};
+//#endregion
 //#region src/host/minimax-code/adapter.ts
 /**
 * The DSH adapter for the MiniMax Code subscription.
@@ -45150,6 +45917,8 @@ function classifyMinimaxFailure(status, bodyText) {
 		message: "MiniMax Code（编程订阅） API error (" + status + "): " + (detail === "" ? "No response" : detail)
 	};
 }
+const minimaxPrefixRequests = new PrefixRequestCoordinator();
+const minimaxStickyImages = new StickyImageBudgetStore("minimax-code");
 /** The reasoning levels one model advertises, from the hardcoded directory. */
 function effortsForModel(modelId) {
 	return [...minimaxCodeModelDef(modelId)?.reasoningEfforts ?? []];
@@ -45517,7 +46286,11 @@ var MinimaxCodeAdapter = class extends LlmAdapter {
 		const settings = await this.settings();
 		minimaxCodeModelDef(options.model);
 		const contextWindow = contextWindowForModel(options.model, settings.contextWindowOverrides);
-		const requestOptions = offloadOldestRequestVideos(offloadOldestRequestImages$2(normalizeGenerateOptions(options), maxRequestImageBytes()), MAX_REQUEST_VIDEO_BYTES);
+		const normalized = withStaticPrefixOrdering("minimax-code", normalizeGenerateOptions(options), this.options.promptCache);
+		const imageBudget = maxRequestImageBytes();
+		const sticky = promptCachePolicyOf("minimax-code", this.options.promptCache).stickyImages ? offloadRequestImagesWithStickyBudget(normalized, imageBudget, Math.max(1, Math.floor(imageBudget / 2)), void 0, minimaxStickyImages.priorDroppedKeys(normalized.sessionId)) : void 0;
+		const requestOptions = offloadOldestRequestVideos(sticky?.options ?? offloadOldestRequestImages$2(normalized, imageBudget), MAX_REQUEST_VIDEO_BYTES);
+		if (sticky !== void 0 && sticky.newlyDroppedKeys.length > 0) minimaxStickyImages.record(normalized.sessionId, sticky.allDroppedKeys);
 		const [images, videos] = await Promise.all([resolveRequestImages$2(requestOptions, this.options.attachments, signal), resolveRequestVideos(requestOptions, this.options.videos, signal)]);
 		const effectiveMax = clampOutputToContext(floorForcedThinkingTokens(options.model, requestOptions.maxTokens) ?? maxOutputTokensFor$1(options.model, contextWindow), contextWindow, estimatedInputTokens(requestOptions));
 		const tried = /* @__PURE__ */ new Set();
@@ -45590,7 +46363,9 @@ var MinimaxCodeAdapter = class extends LlmAdapter {
 			let refusedBody;
 			for (let attempt = 0;; attempt += 1) {
 				try {
-					response = await fetchWithContextBudgetRecovery(fetchFn)(messagesUrl(credentials.region), {
+					const cachePolicy = promptCachePolicyOf("minimax-code", this.options.promptCache);
+					const groupKey = cachePolicy.coldStart === "off" ? void 0 : sharedPrefixGroupKey("minimax-code", requestOptions, credentials.region, accountId ?? "legacy");
+					response = await fetchWithPrefixCoordination(fetchWithContextBudgetRecovery(fetchFn), minimaxPrefixRequests, groupKey, cachePolicy.coldStart === "warm" ? cachePolicy.ttlMs : void 0, messagesUrl(credentials.region), {
 						method: "POST",
 						headers: modelRequestHeaders(credentials.accessToken, requestContext),
 						body,
@@ -48681,34 +49456,6 @@ function makeReminderText2(repeatCount) {
 const REMINDER_TEXT_3 = "<system-reminder>\nWrite your final response now, without any further tool calls. Cover: the current blocker, each approach you have tried and what it established, and the specific information or decision you need from the user to unblock progress. Text only.\n</system-reminder>";
 /** 强停那一轮之后，守卫用来拒绝后续工具调用的文案。 */
 const HANDOFF_VETO_TEXT = "This turn was ended by the repeat breaker after the same tool call was issued " + String(12) + " times in a row. This step accepts a text response only, so the tool call was not executed. Reply in text: the current blocker, what you tried, and what you need next.";
-/**
-* 不计入连击的协调类工具。
-*
-* 这些工具**本来就会**被反复调用：报一次进展、更新一次待办、问一次用户。把它们
-* 算进连击等于让刹车对着正常的心跳乱响。
-*/
-const REPEAT_BREAKER_EXEMPT_TOOLS = Object.freeze([
-	"progress_update",
-	"todo_write",
-	"ask_user_question",
-	"enter_plan_mode",
-	"exit_plan_mode",
-	"leave_plan_mode",
-	"create_goal",
-	"update_goal",
-	"get_goal",
-	"dispatch_ledger",
-	"present"
-]);
-/**
-* 判断一个工具是不是不计入连击的协调类工具。
-*
-* @param toolName - 工具名。
-* @returns 豁免时为 true。
-*/
-function isRepeatBreakerExemptTool(toolName) {
-	return REPEAT_BREAKER_EXEMPT_TOOLS.includes(toolName);
-}
 /** 明确表示「关掉」的环境变量取值。 */
 const DISABLED_ENV_VALUES = /* @__PURE__ */ new Set([
 	"0",
@@ -48731,17 +49478,6 @@ function isKimiRepeatBreakerEnabled(env = process.env) {
 	return !DISABLED_ENV_VALUES.has(raw.trim().toLowerCase());
 }
 /**
-* 判断一个值是不是普通对象（原型是 Object.prototype 或 null）。
-*
-* @param value - 任意值。
-* @returns 是普通对象时为 true。
-*/
-function isPlainRecord(value) {
-	if (value === null || typeof value !== "object" || Array.isArray(value)) return false;
-	const prototype = Object.getPrototypeOf(value);
-	return prototype === Object.prototype || prototype === null;
-}
-/**
 * 判断一个值是不是对象（数组也算）。
 *
 * @param value - 任意值。
@@ -48796,42 +49532,6 @@ function kimiSessionRouteOf(agent) {
 	};
 }
 /**
-* 递归把一个值重排成「键名有序」的形状。
-*
-* 数组**保持原序**：数组元素的顺序本身就是语义（[1, 2] 与 [2, 1] 不是同一件事），
-* 只有对象的键名顺序没有语义。
-*
-* @param value - 当前值。
-* @returns 重排后的值。
-*/
-function sortJsonValue(value) {
-	if (Array.isArray(value)) return value.map(sortJsonValue);
-	if (!isPlainRecord(value)) return value;
-	const out = {};
-	for (const key of Object.keys(value).sort()) out[key] = sortJsonValue(value[key]);
-	return out;
-}
-/**
-* 造一次工具调用的**规范键**。
-*
-* 键 = 工具名 + 参数的规范 JSON。参数先递归重排键名再序列化，于是
-* { a: 1, b: 2 } 与 { b: 2, a: 1 } 得到**完全相同**的键 —— 否则模型只是把参数
-* 的书写顺序换了一下，刹车就看不见这是一个连击。
-*
-* @param toolName - 工具名。
-* @param args - 工具参数。
-* @returns 规范键。
-*/
-function canonicalKimiToolCallKey(toolName, args) {
-	let json;
-	try {
-		json = JSON.stringify(sortJsonValue(args));
-	} catch {
-		json = void 0;
-	}
-	return toolName + " " + (json ?? String(args));
-}
-/**
 * 造一台重复刹车。
 *
 * 状态**按会话**分开：连击是「这场对话里的同一件事重复了多少次」，两个会话
@@ -48839,53 +49539,22 @@ function canonicalKimiToolCallKey(toolName, args) {
 *
 * @returns 一台重复刹车。
 */
+function repeatReminder(count) {
+	if (count >= 8) return REMINDER_TEXT_3;
+	if (count >= 5) return makeReminderText2(count);
+	if (count >= 3) return REMINDER_TEXT_1;
+}
 function createKimiRepeatBreaker() {
-	const sessions = /* @__PURE__ */ new Map();
-	/** 取一个会话的状态；没有就新建一个。 */
-	const stateOf = (sessionId) => {
-		const existing = sessions.get(sessionId);
-		if (existing !== void 0) return existing;
-		const created = {
-			consecutiveKey: null,
-			consecutiveCount: 0,
-			handoffActive: false
-		};
-		sessions.set(sessionId, created);
-		return created;
-	};
-	return {
-		checkBeforeExecute(sessionId, toolName, args) {
-			const state = stateOf(sessionId);
-			if (state.handoffActive) return HANDOFF_VETO_TEXT;
-			const key = canonicalKimiToolCallKey(toolName, args);
-			if (state.consecutiveKey === key && state.consecutiveCount + 1 >= 12) {
-				state.consecutiveCount += 1;
-				state.handoffActive = true;
-				return HANDOFF_VETO_TEXT;
-			}
-		},
-		recordAfterExecute(sessionId, toolName, args) {
-			if (isRepeatBreakerExemptTool(toolName)) return void 0;
-			const state = stateOf(sessionId);
-			if (state.handoffActive) return REMINDER_TEXT_3;
-			const key = canonicalKimiToolCallKey(toolName, args);
-			if (state.consecutiveKey === key) state.consecutiveCount += 1;
-			else {
-				state.consecutiveKey = key;
-				state.consecutiveCount = 1;
-			}
-			if (state.consecutiveCount >= 12) {
-				state.handoffActive = true;
-				return REMINDER_TEXT_3;
-			}
-			if (state.consecutiveCount >= 8) return REMINDER_TEXT_3;
-			if (state.consecutiveCount >= 5) return makeReminderText2(state.consecutiveCount);
-			if (state.consecutiveCount >= 3) return REMINDER_TEXT_1;
-		},
-		resetSession(sessionId) {
-			sessions.delete(sessionId);
-		}
-	};
+	return createExactRepeatBreaker({
+		reminder: repeatReminder,
+		forceStopAt: 12,
+		handoffReminder: REMINDER_TEXT_3,
+		handoffVeto: HANDOFF_VETO_TEXT
+	});
+}
+/** Other providers receive the graduated reminders, never Kimi's hard stop. */
+function createReminderOnlyRepeatBreaker() {
+	return createExactRepeatBreaker({ reminder: repeatReminder });
 }
 /**
 * 读宿主上下文上的工具注册面。
@@ -49007,37 +49676,43 @@ function isRealUserTurn(message) {
 function installKimiRepeatBreaker(ctx, options = {}) {
 	const releaseMount = claimMount$1(ctx);
 	if (releaseMount === void 0) return () => void 0;
-	const breaker = createKimiRepeatBreaker();
+	const kimiBreaker = createKimiRepeatBreaker();
+	const reminderBreaker = createReminderOnlyRepeatBreaker();
 	const enabled = options.enabled ?? (() => isKimiRepeatBreakerEnabled());
-	/** 读一次开关。读失败按**开启**处理：保护不该因为一次读失败而失效。 */
-	const isEnabled = () => {
+	const globalEnabled = options.globalEnabled ?? (() => true);
+	const providers = new Set(options.providers ?? ["kimi-code"]);
+	const readSwitch = (read) => {
 		try {
-			return enabled() !== false;
+			return read() !== false;
 		} catch {
 			return true;
 		}
 	};
-	/**
-	* 取一次执行所属的会话；不是 Kimi Code 这条线路时是 undefined。
-	*
-	* @param agent - 调用方 agent。
-	* @returns 会话标识；不该管时是 undefined。
-	*/
-	const kimiSessionIdOf = (agent) => {
+	const targetOf = (agent) => {
 		const route = kimiSessionRouteOf(agent);
-		if (route.provider !== "kimi-code") return void 0;
-		return route.sessionId;
+		const provider = route.provider;
+		if (provider === void 0 || route.sessionId === void 0 || !providers.has(provider)) return void 0;
+		return {
+			provider,
+			key: provider + ":" + route.sessionId,
+			breaker: provider === "kimi-code" ? kimiBreaker : reminderBreaker
+		};
+	};
+	const activeTargetOf = (agent) => {
+		if (!readSwitch(globalEnabled)) return void 0;
+		const target = targetOf(agent);
+		if (target === void 0) return void 0;
+		if (target.provider === "kimi-code" && !readSwitch(enabled)) return void 0;
+		return target;
 	};
 	const disposeGuard = toolsServiceOf(ctx)?.guard?.((execution) => {
 		try {
-			if (!isEnabled()) return void 0;
 			const record = isRecord$45(execution) ? execution : void 0;
-			const agent = record?.["agent"];
-			const sessionId = kimiSessionIdOf(agent);
-			if (sessionId === void 0) return void 0;
+			const target = activeTargetOf(record?.["agent"]);
+			if (target === void 0) return void 0;
 			const toolName = stringOf$3(record?.["name"]);
 			if (toolName === void 0) return void 0;
-			const veto = breaker.checkBeforeExecute(sessionId, toolName, record?.["arguments"]);
+			const veto = target.breaker.checkBeforeExecute(target.key, toolName, record?.["arguments"]);
 			if (veto === HANDOFF_VETO_TEXT) return HANDOFF_VETO_TEXT;
 			return veto;
 		} catch {
@@ -49047,13 +49722,11 @@ function installKimiRepeatBreaker(ctx, options = {}) {
 	const disposePostExecute = ctx.on?.("tools/post-execute", (async (exec, result, next) => {
 		const decision = await next();
 		try {
-			if (!isEnabled()) return decision;
-			const agent = exec.agent;
-			const sessionId = kimiSessionIdOf(agent);
-			if (sessionId === void 0) return decision;
+			const target = activeTargetOf(exec.agent);
+			if (target === void 0) return decision;
 			const toolName = stringOf$3(exec.name);
 			if (toolName === void 0) return decision;
-			const reminder = breaker.recordAfterExecute(sessionId, toolName, exec.arguments);
+			const reminder = target.breaker.recordAfterExecute(target.key, toolName, exec.arguments);
 			if (reminder === void 0) return decision;
 			return withAdditionalContexts(decision, [createRepeatBreakerReminder(reminder)]);
 		} catch {
@@ -49064,9 +49737,9 @@ function installKimiRepeatBreaker(ctx, options = {}) {
 		const decision = await next();
 		try {
 			if (!isRealUserTurn((Array.isArray(payload?.messages) ? payload.messages : []).at(-1))) return decision;
-			const sessionId = kimiSessionIdOf(payload?.agent);
-			if (sessionId === void 0) return decision;
-			breaker.resetSession(sessionId);
+			const target = activeTargetOf(payload?.agent);
+			if (target === void 0) return decision;
+			target.breaker.resetSession(target.key);
 		} catch {
 			return decision;
 		}
@@ -49074,9 +49747,9 @@ function installKimiRepeatBreaker(ctx, options = {}) {
 	}));
 	const disposeDisposed = ctx.on?.("agent/disposed", ((payload) => {
 		try {
-			const sessionId = kimiSessionIdOf(payload?.agent);
-			if (sessionId === void 0) return;
-			breaker.resetSession(sessionId);
+			const target = targetOf(payload?.agent);
+			if (target === void 0) return;
+			target.breaker.resetSession(target.key);
 		} catch {}
 	}));
 	return () => {
@@ -56502,6 +57175,8 @@ function assertClaudeCliVersionMeetsFloor(modelId, catalog = FALLBACK_MODELS) {
 	if (meetsDottedVersionFloor(effective, floor)) return;
 	throw new LlmError("Claude（订阅） will not serve " + modelId + " to a client reporting version " + effective + ": upstream requires " + floor + " or newer for this model, and a lower claim is refused with claude_code_version_too_old. This was caught locally, before the request was sent. Raise the reported version with the DSH_CLAUDE_CLI_VERSION environment variable, or with a setClaudeCliVersion pin, then retry. The stored sign-in is still valid — this is not a credential problem and signing in again will not change it.", "PROVIDER_ERROR");
 }
+const claudePrefixRequests = new PrefixRequestCoordinator();
+const claudeStickyImages = new StickyImageBudgetStore("claude");
 /**
 * The immutable id of the account a token came from, presented as the two-method
 * store oauth.ts wants. See module note 3.
@@ -56755,7 +57430,9 @@ var ClaudeAdapter = class ClaudeAdapter extends LlmAdapter {
 			cacheTtl
 		});
 		try {
-			return await fetchWithContextBudgetRecovery(fetchFn)("https://api.anthropic.com/v1/messages?beta=true", {
+			const cachePolicy = promptCachePolicyOf("claude", this.options.promptCache);
+			const groupKey = cachePolicy.coldStart === "off" ? void 0 : sharedPrefixGroupKey("claude", requestOptions, "subscription", credentials.account?.uuid ?? "legacy");
+			return await fetchWithPrefixCoordination(fetchWithContextBudgetRecovery(fetchFn), claudePrefixRequests, groupKey, cachePolicy.coldStart === "warm" ? cachePolicy.ttlMs : void 0, "https://api.anthropic.com/v1/messages?beta=true", {
 				method: "POST",
 				headers: buildClaudeHeaders(credentials.accessToken, {
 					model: requestOptions.model,
@@ -56775,9 +57452,12 @@ var ClaudeAdapter = class ClaudeAdapter extends LlmAdapter {
 	}
 	async *requestStream(options, signal, activity) {
 		const fetchFn = this.options.fetchFn ?? fetch;
-		const normalized = normalizeGenerateOptions(options);
+		const normalized = withStaticPrefixOrdering("claude", normalizeGenerateOptions(options), this.options.promptCache);
 		const images = await resolveRequestImages(normalized, this.options.attachments, signal);
-		const requestOptions = offloadOldestRequestImages(normalized, images);
+		const stickyImages = new Map([...images].map(([id, image]) => [id, image.kind === "oversized" ? { kind: "unavailable" } : image]));
+		const sticky = promptCachePolicyOf("claude", this.options.promptCache).stickyImages ? offloadRequestImagesWithStickyBudget(normalized, MAX_REQUEST_IMAGE_BYTES$2, Math.max(1, Math.floor(MAX_REQUEST_IMAGE_BYTES$2 / 2)), stickyImages, claudeStickyImages.priorDroppedKeys(normalized.sessionId)) : void 0;
+		const requestOptions = sticky?.options ?? offloadOldestRequestImages(normalized, images);
+		if (sticky !== void 0 && sticky.newlyDroppedKeys.length > 0) claudeStickyImages.record(normalized.sessionId, sticky.allDroppedKeys);
 		const toolNames = claudeToolNames(requestOptions.tools);
 		assertClaudeCliVersionMeetsFloor(requestOptions.model, await this.catalogForRequest());
 		const effort = requestOptions.reasoningEffort === void 0 || requestOptions.reasoningEffort === null ? void 0 : String(requestOptions.reasoningEffort);
@@ -66765,11 +67445,77 @@ function classifyStepCodeFailure(status, detail) {
 }
 //#endregion
 //#region src/host/step-code/cache-stats.ts
+/**
+* Misses at or below this are cache-breakpoint granularity, not waste.
+*
+* A prefix cache breaks at block boundaries, so a few hundred tokens of drift
+* between two turns is normal re-shuffling. Reporting it teaches the user to
+* ignore the notice, which costs the one notice that mattered.
+*/
+const NOISE_FLOOR_TOKENS = 1024;
 /** Aggregate prompt count of one turn: uncached input plus both cache buckets. */
 function promptTokensOf(usage) {
 	return usage.inputTokens + (usage.cacheReadTokens ?? 0) + (usage.cacheWriteTokens ?? 0);
 }
 //#endregion
+//#region src/host/common/cache-activity.ts
+/** Last request activity used by cache-expiry hints and wait-time keepalive. */
+function normalizeRecord(value) {
+	if (typeof value !== "object" || value === null || Array.isArray(value)) return void 0;
+	const record = value;
+	const lastActiveAt = record.lastActiveAt;
+	const updatedAt = record.updatedAt;
+	if (typeof lastActiveAt !== "number" || !Number.isFinite(lastActiveAt)) return void 0;
+	if (typeof updatedAt !== "number" || !Number.isFinite(updatedAt)) return void 0;
+	return {
+		lastActiveAt,
+		totalTokens: typeof record.totalTokens === "number" && Number.isFinite(record.totalTokens) ? Math.max(0, Math.floor(record.totalTokens)) : 0,
+		updatedAt
+	};
+}
+function persistenceEnabled() {
+	return process.env.VITEST !== "true" && process.env.NODE_ENV !== "test";
+}
+var PromptCacheActivityStore = class {
+	store;
+	constructor(provider, filePath) {
+		const target = filePath ?? path.join(dshHomeDir(), "storages", `dsh-chatgpt-subscription-${provider}-activity.json`);
+		this.store = new CommonSessionCacheStateStore({
+			filePath: target,
+			allowWrite: filePath !== void 0 || persistenceEnabled(),
+			normalizeRecord
+		});
+	}
+	read(sessionId) {
+		const record = this.store.read(sessionId);
+		return record === void 0 ? void 0 : {
+			lastActiveAt: record.lastActiveAt,
+			totalTokens: record.totalTokens
+		};
+	}
+	record(sessionId, totalTokens, at = Date.now()) {
+		if (sessionId === void 0 || sessionId.trim() === "") return;
+		this.store.write(sessionId, {
+			lastActiveAt: at,
+			totalTokens: Number.isFinite(totalTokens) ? Math.max(0, Math.floor(totalTokens)) : 0,
+			updatedAt: at
+		});
+	}
+	flush() {
+		this.store.flush();
+	}
+};
+const SHARED_ACTIVITY_STORES = /* @__PURE__ */ new Map();
+/** One live store per provider; tests can still construct isolated instances by path. */
+function promptCacheActivityStore(provider) {
+	let store = SHARED_ACTIVITY_STORES.get(provider);
+	if (store === void 0) {
+		store = new PromptCacheActivityStore(provider);
+		SHARED_ACTIVITY_STORES.set(provider, store);
+	}
+	return store;
+}
+//#endregion
 //#region src/host/step-code/client.ts
 /**
 * Step Code 线路的上游 I/O：模型目录 + 开发者中心额度。
@@ -68446,6 +69192,8 @@ function streamHeadersFor(credentials) {
 		accept: "text/event-stream"
 	};
 }
+const stepPrefixRequests = new PrefixRequestCoordinator();
+const stepCacheActivity = promptCacheActivityStore("step-code");
 /**
 * Step Code（Step Plan 订阅）线路的 LlmAdapter 实现。
 *
@@ -68494,10 +69242,10 @@ var StepCodeAdapter = class extends LlmAdapter {
 	* 两条路径都**不刷新**：Step 的 api_key 永不过期，号池的 refresh 钩子恒抛
 	* `StepCodeCredentialRefreshError`，所以这里不存在「刷新失败后再试一次」的分支。
 	*/
-	async credentialFor(excludeIds) {
+	async credentialFor(excludeIds, selection) {
 		const fetchFn = this.options.fetchFn ?? fetch;
 		if (this.accountPool !== null) {
-			const effective = await this.accountPool.getEffectiveCredential(excludeIds, fetchFn);
+			const effective = await this.accountPool.getEffectiveCredential(excludeIds, fetchFn, selection);
 			const credentials = effective.credentials;
 			return {
 				accountId: effective.account.id,
@@ -68519,6 +69267,31 @@ var StepCodeAdapter = class extends LlmAdapter {
 			}
 		};
 	}
+	/** Warm-prefix account preference, inert unless a deployment proves account scope. */
+	async cacheAffinityHint(options) {
+		const pool = this.accountPool;
+		if (pool === null) return void 0;
+		const policy = promptCachePolicyOf("step-code", this.options.promptCache);
+		if (!policy.affinity || policy.ttlMs === void 0) return void 0;
+		const data = await pool.read().catch(() => void 0);
+		if (data === void 0 || data.accounts.length < 2) return void 0;
+		const candidates = [];
+		const remainingShareById = {};
+		for (const account of data.accounts) {
+			const key = sharedPrefixGroupKey("step-code", options, account.credentials.profile, account.id);
+			if (key === void 0 || !stepPrefixRequests.isWarm(key, policy.ttlMs)) continue;
+			candidates.push({
+				id: account.id,
+				warm: true
+			});
+			remainingShareById[account.id] = remainingQuotaShare$1(getCachedStepQuotaFor(account.id));
+		}
+		const preferAccountIds = orderAffinityCandidates$1({
+			candidates,
+			remainingShareById
+		});
+		return preferAccountIds.length === 0 ? void 0 : { preferAccountIds };
+	}
 	async listModels(provider) {
 		const prov = provider || PROVIDER_ID$6;
 		const settings = await this.settings();
@@ -68581,13 +69354,14 @@ var StepCodeAdapter = class extends LlmAdapter {
 	async *requestStream(options, signal, activity) {
 		const fetchFn = this.options.fetchFn ?? fetch;
 		const settings = await this.settings();
-		const normalized = normalizeGenerateOptions(options);
+		const normalized = withStaticPrefixOrdering("step-code", normalizeGenerateOptions(options), this.options.promptCache);
 		const images = await resolveRequestImages$2(normalized, this.options.attachments, signal);
 		const requestOptions = offloadOldestStepCodeRequestImages(normalized, images);
 		const tried = /* @__PURE__ */ new Set();
 		let response;
 		let url = "";
-		let picked = await this.credentialFor(tried);
+		const affinity = await this.cacheAffinityHint(requestOptions);
+		let picked = await this.credentialFor(tried, affinity);
 		while (true) {
 			if (picked.accountId !== null) tried.add(picked.accountId);
 			const catalog = this.catalog(picked.credentials.profile);
@@ -68595,7 +69369,9 @@ var StepCodeAdapter = class extends LlmAdapter {
 			const body = JSON.stringify(buildChatRequest$1(requestOptions, images, stepCodeRequestOptionsFor(entry, { cacheRequestParams: settings.cacheRequestParams })));
 			url = stepChatCompletionsUrl(picked.credentials.profile);
 			try {
-				response = await fetchWithContextBudgetRecovery(fetchFn)(url, {
+				const cachePolicy = promptCachePolicyOf("step-code", this.options.promptCache);
+				const groupKey = cachePolicy.coldStart === "off" ? void 0 : sharedPrefixGroupKey("step-code", requestOptions, picked.credentials.profile, picked.accountId ?? "legacy");
+				response = await fetchWithPrefixCoordination(fetchWithContextBudgetRecovery(fetchFn), stepPrefixRequests, groupKey, cachePolicy.coldStart === "warm" ? cachePolicy.ttlMs : void 0, url, {
 					method: "POST",
 					headers: streamHeadersFor(picked.credentials),
 					body,
@@ -68617,13 +69393,13 @@ var StepCodeAdapter = class extends LlmAdapter {
 					const after = retryAfterMs$1(response.headers);
 					await this.accountPool.markCooldown(picked.accountId, after ?? 3e5, "Step Code（Step Plan 订阅） 429：上游限流或额度用尽").catch(() => void 0);
 					if (await this.accountPool.hasAnotherAvailableAccount(tried)) {
-						picked = await this.credentialFor(tried);
+						picked = await this.credentialFor(tried, affinity);
 						continue;
 					}
 				} else if (failure.code === "AUTH") {
 					await this.accountPool.markAuthFailed(picked.accountId, failure.message).catch(() => void 0);
 					if (await this.accountPool.hasAnotherAvailableAccount(tried)) {
-						picked = await this.credentialFor(tried);
+						picked = await this.credentialFor(tried, affinity);
 						continue;
 					}
 				}
@@ -68656,16 +69432,42 @@ var StepCodeAdapter = class extends LlmAdapter {
 			assertStreamComplete$2(state);
 			for (const chunk of closeStream$1(state)) yield chunk;
 		} finally {
-			if (state.sawUsage) rememberStepCodeCacheUsage({
-				inputTokens: state.inputTokens,
-				outputTokens: state.outputTokens,
-				cacheReadTokens: state.cacheReadTokens,
-				cacheWriteTokens: state.cacheWriteTokens
-			});
+			if (state.sawUsage) {
+				rememberStepCodeCacheUsage({
+					inputTokens: state.inputTokens,
+					outputTokens: state.outputTokens,
+					cacheReadTokens: state.cacheReadTokens,
+					cacheWriteTokens: state.cacheWriteTokens
+				});
+				stepCacheActivity.record(options.sessionId, promptTokensOf({
+					inputTokens: state.inputTokens,
+					outputTokens: state.outputTokens,
+					cacheReadTokens: state.cacheReadTokens,
+					cacheWriteTokens: state.cacheWriteTokens
+				}));
+			}
 			reader.cancel().catch(() => void 0);
 		}
 	}
 };
+/** Cache-expiry hint for one session, or the latest recorded session when omitted. */
+function readStepCodeCacheHint(sessionId, override) {
+	const policy = promptCachePolicyOf("step-code", override);
+	if (!policy.hint) return null;
+	const activity = stepCacheActivity.read(sessionId);
+	if (activity === void 0) return null;
+	const decision = evaluateCacheHint$1({
+		now: Date.now(),
+		lastActiveAt: activity.lastActiveAt,
+		totalTokens: activity.totalTokens,
+		ttlMs: policy.ttlMs,
+		minTokensToHint: NOISE_FLOOR_TOKENS
+	});
+	return decision.kind === "hint" ? {
+		idleMs: decision.idleMs,
+		totalTokens: decision.totalTokens
+	} : null;
+}
 //#endregion
 //#region src/host/step-code/oauth.ts
 /**
@@ -69457,8 +70259,8 @@ var StepCodeAccountPool = class extends AccountPoolCore {
 	* 核心在这里做轮转、冷却、可用性判断与（被拒绝的）刷新；本方法只多给适配器一件事：
 	* 一次请求内换过账号时，把试过的账号排除掉。
 	*/
-	async getEffectiveCredential(excludeIds, fetchFn = fetch) {
-		return this.getEffectiveAccount(excludeIds, fetchFn);
+	async getEffectiveCredential(excludeIds, fetchFn = fetch, selection) {
+		return this.getEffectiveAccount(excludeIds, fetchFn, false, void 0, selection);
 	}
 	/**
 	* 一次额度读属于哪个账号，连同读它用的凭据。
@@ -69949,6 +70751,7 @@ async function getStepCodeStatus(sources) {
 	const settingsValue = await settings.status();
 	const quota = sources.quota;
 	const cacheStats = sources.cacheStats;
+	const cacheHint = readStepCodeCacheHint(void 0, sources.options.promptCache);
 	const accountCount = sources.pool.accounts.length;
 	return {
 		...sources.pool,
@@ -69973,6 +70776,7 @@ async function getStepCodeStatus(sources) {
 		quotaCredential: sources.quotaCredential,
 		cacheRequestParams: settingsValue.cacheRequestParams === true,
 		...cacheStats === void 0 ? {} : { cacheStats },
+		...cacheHint === null ? {} : { cacheHint },
 		serving: readOption$2(options.serving, true),
 		conflict: readOption$2(options.conflict, null)
 	};
@@ -72296,6 +73100,7 @@ function isRecord$26(value) {
 *
 * @module dsh-chatgpt-subscription/cline-pass-mapper
 */
+const clineReasoningDialect = new ReasoningKeyDialect$1();
 /**
 * The one wording the shared detector does not know.
 *
@@ -72329,7 +73134,7 @@ function serializeAssistant$1(message) {
 	return {
 		role: "assistant",
 		content: text,
-		...reasoning.length > 0 ? { reasoning_content: reasoning } : {},
+		...reasoning.length > 0 ? { [clineReasoningDialect.outboundKey()]: reasoning } : {},
 		...toolCalls.length > 0 ? { tool_calls: toolCalls } : {}
 	};
 }
@@ -72624,6 +73429,7 @@ function acceptIdentity$1(current, incoming) {
 * parts, which is where a gateway that omits the flat field puts the same text.
 */
 function reasoningOf$1(delta) {
+	clineReasoningDialect.observe(delta);
 	for (const field of ["reasoning", "reasoning_content"]) {
 		const value = delta[field];
 		if (typeof value === "string" && value.length > 0) return value;
@@ -79028,6 +79834,7 @@ async function loadQoderQuota(credentials, options = {}) {
 *
 * @module dsh-chatgpt-subscription/qoder-mapper
 */
+const qoderReasoningDialect = new ReasoningKeyDialect$1();
 function isRecord$21(value) {
 	return typeof value === "object" && value !== null && !Array.isArray(value);
 }
@@ -79374,7 +80181,7 @@ function buildQoderChatRequest(options, images, request, context) {
 				role: "assistant",
 				content: content !== "" ? content : toolCalls.length > 0 || reasoning !== "" ? " " : null
 			};
-			if (reasoning !== "") entry.reasoning_content = reasoning;
+			if (reasoning !== "") Object.assign(entry, { [qoderReasoningDialect.outboundKey()]: reasoning });
 			if (toolCalls.length > 0) entry.tool_calls = toolCalls;
 			if (entry.content !== null) messages.push(entry);
 			continue;
@@ -79723,6 +80530,7 @@ function processQoderStreamLine(line, state) {
 	const choice = isRecord$21(choices[0]) ? choices[0] : void 0;
 	const delta = choice !== void 0 && isRecord$21(choice.delta) ? choice.delta : void 0;
 	if (delta !== void 0) {
+		qoderReasoningDialect.observe(delta);
 		const reasoning = asString$7(delta.reasoning_content) ?? asString$7(delta.reasoning);
 		if (reasoning !== void 0 && reasoning !== "") out.push(...appendTextDelta(state, "reasoning", sanitizeText$1(reasoning)));
 		const content = asString$7(delta.content);
@@ -84252,6 +85060,7 @@ async function fetchCodingPlanUsage({ fetchFn, controlKey, runtimeKey, workspace
 *
 * @module dsh-chatgpt-subscription/opendesign-mapper
 */
+const openDesignReasoningDialect = new ReasoningKeyDialect$1();
 /** The sentence a no-content finish carries, shared with the sibling lines. */
 const NO_ANSWER_PHRASE = "ended the turn without an answer or a tool call";
 /**
@@ -84287,7 +85096,7 @@ function serializeAssistant(message) {
 	return {
 		role: "assistant",
 		content: text,
-		...reasoning.length > 0 ? { reasoning_content: reasoning } : {},
+		...reasoning.length > 0 ? { [openDesignReasoningDialect.outboundKey()]: reasoning } : {},
 		...toolCalls.length > 0 ? { tool_calls: toolCalls } : {}
 	};
 }
@@ -84630,6 +85439,7 @@ function acceptIdentity(current, incoming) {
 * parts, which is where a gateway that omits the flat field puts the same text.
 */
 function reasoningOf(delta) {
+	openDesignReasoningDialect.observe(delta);
 	for (const field of ["reasoning", "reasoning_content"]) {
 		const value = delta[field];
 		if (typeof value === "string" && value.length > 0) return value;
@@ -96110,7 +96920,7 @@ function createCacheWarmupResultTool(deps) {
 			const route = deps.routeOf(exec);
 			const refId = args.refId;
 			const action = args.action;
-			if (route.provider !== deps.provider) return failure(action, refId, route.sessionId ?? "", "WRONG_PROVIDER", "ref handles belong to the " + deps.provider + " line; this session is routed to \"" + (route.provider ?? "unknown") + "\"");
+			if (!(route.provider !== void 0 && (deps.allowProvider?.(route.provider) ?? route.provider === deps.provider))) return failure(action, refId, route.sessionId ?? "", "WRONG_PROVIDER", "ref handles are not available on this session route: \"" + (route.provider ?? "unknown") + "\"");
 			const sessionId = route.sessionId;
 			if (sessionId === void 0) return failure(action, refId, "", "NO_SESSION", "the current session could not be resolved, so the handle cannot be checked");
 			const store = deps.store();
@@ -96252,8 +97062,8 @@ function cacheWarmupNavigationIntentOf(agent) {
 /**
 * 本模块服务的线路。
 *
-* 结果投影的正确性依赖这条线路的缓存语义（内容前缀命中），别的线路上把结果换成
-* 预览只有坏处没有好处，所以这里写死一个 provider 名而不是做成开关。
+* Kimi 是缺少显式 provider 列表时的兼容回退；宿主正常接线会注入受控线路表。
+* 是否投影由当前设置决定，是否补读已有 ref 则由会话所有权与受支持路由决定。
 */
 const CACHE_WARMUP_PROVIDER = "kimi-code";
 /**
@@ -96385,9 +97195,9 @@ function claimMount(ctx) {
 * @param sessionId - 会话标识。
 * @returns 快照，或读不到时的 undefined。
 */
-function readCacheWarmupSnapshot(sessionId) {
+function readCacheWarmupSnapshot(sessionId, provider) {
 	try {
-		const raw = cacheWarmupSnapshot(sessionId);
+		const raw = cacheWarmupSnapshot(sessionId, void 0, provider);
 		return typeof raw === "object" && raw !== null ? raw : void 0;
 	} catch {
 		return;
@@ -96453,6 +97263,7 @@ function installCacheWarmup(ctx, options) {
 		if (tools === void 0) return () => void 0;
 		return tools.register(createCacheWarmupResultTool({
 			provider: CACHE_WARMUP_PROVIDER,
+			allowProvider: (provider) => CONTROLLED_PROVIDERS.includes(provider),
 			routeOf: (exec) => cacheWarmupRouteOf(exec?.agent),
 			store: storeOf,
 			log
@@ -96474,12 +97285,13 @@ function installCacheWarmup(ctx, options) {
 	* @param sessionId - 会话标识。
 	* @param budget - 本轮读到的预算对象。
 	*/
-	const trackWarmup = (sessionId, budget) => {
-		const snapshot = readCacheWarmupSnapshot(sessionId);
+	const trackWarmup = (sessionId, budget, provider) => {
+		const snapshot = readCacheWarmupSnapshot(sessionId, provider);
 		const warmup = booleanOf(snapshot?.warmup) ?? booleanOf(budget.warmup);
 		const consecutiveHits = numberOf$1(snapshot?.consecutiveHits) ?? numberOf$1(budget.consecutiveHits);
 		const hitRatio = numberOf$1(snapshot?.hitRatio) ?? numberOf$1(budget.hitRatio);
-		if (warmupState.get(sessionId) === true && warmup === false) {
+		const scope = (provider ?? "kimi-code") + ":" + sessionId;
+		if (warmupState.get(scope) === true && warmup === false) {
 			const reason = exitReasonOf(consecutiveHits);
 			log({
 				event: "warmup-exit",
@@ -96489,7 +97301,7 @@ function installCacheWarmup(ctx, options) {
 				...hitRatio === void 0 ? {} : { hitRatio }
 			});
 		}
-		if (warmup !== void 0) warmupState.set(sessionId, warmup);
+		if (warmup !== void 0) warmupState.set(scope, warmup);
 		return {
 			...warmup === void 0 ? {} : { warmup },
 			...consecutiveHits === void 0 ? {} : { consecutiveHits },
@@ -96533,7 +97345,8 @@ function installCacheWarmup(ctx, options) {
 		if (retrievalTool && navigation !== true) return decision;
 		if (sourceTool && !settings.enabled && navigation !== true) return decision;
 		const route = cacheWarmupRouteOf(exec.agent);
-		if (route.provider !== "kimi-code") return decision;
+		const admittedProviders = settings.providers ?? ["kimi-code"];
+		if (route.provider === void 0 || !admittedProviders.includes(route.provider)) return decision;
 		const sessionId = route.sessionId;
 		if (sessionId === void 0) return decision;
 		const store = storeOf();
@@ -96541,11 +97354,11 @@ function installCacheWarmup(ctx, options) {
 		const content = blocks;
 		let durable;
 		try {
-			durable = options.durableHint?.(sessionId);
+			durable = options.durableHint?.(sessionId, route.provider);
 		} catch {
 			durable = void 0;
 		}
-		const budget = durable === void 0 ? cacheWarmupBudget(sessionId) : cacheWarmupBudget(sessionId, Date.now(), durable);
+		const budget = durable === void 0 ? cacheWarmupBudget(sessionId, Date.now(), void 0, route.provider) : cacheWarmupBudget(sessionId, Date.now(), durable, route.provider);
 		const projected = await projectToolResult({
 			toolName: exec.name,
 			provider: route.provider,
@@ -96560,7 +97373,7 @@ function installCacheWarmup(ctx, options) {
 			loggerOf()?.warn?.("[dsh-chatgpt-subscription] cache warmup projection skipped: " + cacheWarmupErrorLabel(error));
 		});
 		if (projected === void 0) return decision;
-		const metrics = trackWarmup(sessionId, budget);
+		const metrics = trackWarmup(sessionId, budget, route.provider);
 		log({
 			event: "projection",
 			session: sessionId,
@@ -96607,7 +97420,7 @@ function blocksOf(content) {
 /** Environment override for wait-time cache keepalive warmup. */
 const CACHE_WAIT_WARMUP_ENV = "DSH_KIMI_CODE_CACHE_WAIT_WARMUP";
 /** Settings description for the progress-updates / Kimi cache-wait switch. */
-const CACHE_WAIT_WARMUP_SETTING_DESCRIPTION = "Kimi Code 专用：等待子代理完成期间，主会话每空闲满 20 分钟自动触发一次轻量缓存续热，只要子代理仍在运行就持续续热以防 25 分钟缓存过期重建；非 Kimi 线路开关无影响。";
+const CACHE_WAIT_WARMUP_SETTING_DESCRIPTION = "等待子代理完成期间按线路缓存 TTL 轻量续热；默认用于 Kimi（20/25 分钟），其他线路需明确 TTL 并在 promptCacheOverrides 中显式启用。";
 const MAX_TRACKED_SESSIONS = 256;
 /** Pure cache-wait policy; no child deadline or forced progress is involved. */
 function evaluateCacheWait(input) {
@@ -96678,7 +97491,8 @@ function defaultTimer() {
 		}
 	};
 }
-function defaultActivity(sessionId) {
+function defaultActivity(sessionId, provider = "kimi-code") {
+	if (provider !== "kimi-code") return promptCacheActivityStore(provider).read(sessionId);
 	const durable = sessionCacheState.read(sessionId);
 	const lastActiveAt = sessionLastActiveAt(sessionId) ?? durable?.lastActiveAt;
 	if (lastActiveAt === void 0) return void 0;
@@ -96707,7 +97521,6 @@ var KimiCacheWaitController = class {
 	warmup;
 	logger;
 	decisionMs;
-	ttlMs;
 	minWarmupTokens;
 	disposed = false;
 	constructor(options = {}) {
@@ -96721,7 +97534,6 @@ var KimiCacheWaitController = class {
 		this.warmup = typeof warmupOption === "function" ? warmupOption : () => warmupOption ?? true;
 		this.logger = options.logger;
 		this.decisionMs = options.decisionMs ?? 12e5;
-		this.ttlMs = options.ttlMs ?? 15e5;
 		this.minWarmupTokens = options.minWarmupTokens ?? 0;
 	}
 	/** Observe one durable Session event. */
@@ -96854,15 +97666,15 @@ var KimiCacheWaitController = class {
 		this.timer.clear(state.timer);
 		state.timer = void 0;
 	}
-	isKimiSession(session) {
-		if (this.kimiSession !== void 0) return this.kimiSession(session);
+	providerForSession(session) {
+		if (this.kimiSession !== void 0) return this.kimiSession(session) ? "kimi-code" : void 0;
 		try {
 			const current = text(session.requestHeader?.()?.config?.provider);
-			if (current !== void 0) return current === "kimi-code";
+			if (current !== void 0) return current;
 		} catch {}
 		const latest = this.providerBySession.get(session.id);
-		if (latest !== void 0) return latest === "kimi-code";
-		return this.activity(session.id)?.lastActiveAt !== void 0;
+		if (latest !== void 0) return latest;
+		return this.activity(session.id, "kimi-code")?.lastActiveAt !== void 0 ? "kimi-code" : void 0;
 	}
 	isWarmupEnabled() {
 		try {
@@ -96874,8 +97686,11 @@ var KimiCacheWaitController = class {
 	reconcile(session, state) {
 		this.clearTimer(state);
 		if (this.disposed || !state.waiting || state.handled || state.children.size === 0) return;
-		if (!this.isKimiSession(session)) return;
-		const activity = this.activity(session.id);
+		const provider = this.providerForSession(session);
+		if (provider === void 0) return;
+		const cachePolicy = promptCachePolicyOf(provider, this.options.overrides?.[provider]);
+		if (!cachePolicy.waitWarmup || cachePolicy.ttlMs === void 0) return;
+		const activity = this.activity(session.id, provider);
 		if (activity === void 0 && state.lastWarmupAt === void 0) return;
 		const effectiveLastActiveAt = Math.max(activity?.lastActiveAt ?? -Infinity, state.lastWarmupAt ?? -Infinity);
 		if (!Number.isFinite(effectiveLastActiveAt)) return;
@@ -96888,8 +97703,8 @@ var KimiCacheWaitController = class {
 			totalTokens: activity?.totalTokens,
 			warmupEnabled: this.isWarmupEnabled(),
 			warmupFailed: state.warmupFailed,
-			decisionMs: this.decisionMs,
-			ttlMs: this.ttlMs,
+			decisionMs: provider === "kimi-code" ? this.decisionMs : Math.max(1, Math.floor(cachePolicy.ttlMs * .8)),
+			ttlMs: cachePolicy.ttlMs,
 			minWarmupTokens: this.minWarmupTokens
 		});
 		if (action.kind === "none") return;
@@ -96944,8 +97759,14 @@ function installKimiCacheWait(ctx, options = {}) {
 * schema（`src/index.ts` 的 `Config`）与设置卡共用同一份文案：两处各写一份，
 * 迟早变成「卡片说只影响新结果、schema 说立刻生效」这种自相矛盾。
 */
-const CACHE_WARMUP_SETTING_DESCRIPTION = "Kimi Code 专用：启用缓存预热的结果投影——Kimi 线路上可以可靠结构化提取的超大工具结果会被替换成带 ref:// 引用的预览，需要原文时由 cache_warmup_result 精确补读。关闭只影响之后产生的新结果，已进入历史的内容不变。本开关只对 Kimi Code 线路生效，其他线路开着关着都没有区别。";
-const CACHE_WARMUP_RETRIEVAL_SETTING_DESCRIPTION = "Kimi Code 专用：启用检索型结果投影——明确导航任务中的大文件、网页和搜索结果可显示为大纲或分布预览，原文由 cache_warmup_result 补读。编辑、审查和意图不明的读取保持原文。默认开启，只影响之后的新结果。本开关只对 Kimi Code 线路生效，其他线路开着关着都没有区别。";
+const CACHE_WARMUP_SETTING_DESCRIPTION = "按线路配置结果投影：可以可靠结构化提取的超大工具结果会被替换成带 ref:// 引用的预览，需要原文时由 cache_warmup_result 精确补读。关闭只影响之后产生的新结果，已进入历史的内容不变。";
+const CACHE_WARMUP_RETRIEVAL_SETTING_DESCRIPTION = "按线路配置检索型结果投影：明确导航任务中的大文件、网页和搜索结果可显示为大纲或分布预览，原文由 cache_warmup_result 补读。编辑、审查和意图不明的读取保持原文。默认开启，只影响之后的新结果。";
+/** 缺省设置由宿主注入 controlled provider 列表；独立使用时保持 Kimi 兼容。 */
+const DEFAULT_CACHE_WARMUP_SETTINGS = {
+	enabled: true,
+	retrievalEnabled: true,
+	providers: ["kimi-code"]
+};
 /**
 * 读取一个配置字段，沿着 volatile 活引用取当前值。
 * @param field - 字段本身。
@@ -96967,9 +97788,12 @@ function readCacheWarmupConfigField(field) {
 function resolveCacheWarmupSettings(config) {
 	const enabled = readCacheWarmupConfigField(config?.cacheWarmup);
 	const retrievalEnabled = readCacheWarmupConfigField(config?.cacheWarmupRetrieval);
+	const configuredProviders = readCacheWarmupConfigField(config?.cacheWarmupProviders);
+	const providers = configuredProviders === void 0 ? DEFAULT_CACHE_WARMUP_SETTINGS.providers : [...new Set(configuredProviders.map((provider) => provider.trim()).filter((provider) => provider !== ""))];
 	return {
 		enabled: enabled !== false,
-		retrievalEnabled: retrievalEnabled !== false
+		retrievalEnabled: retrievalEnabled !== false,
+		providers
 	};
 }
 //#endregion
@@ -105052,10 +105876,14 @@ const Config = z$1.object({
 	]).default("auto").volatile().description("进展更新的输出语言：auto 跟随对话语言，zh 中文，en 英文。"),
 	planModePipeline: z$1.boolean().default(true).volatile().description("启用计划模式管线：允许调度模式在澄清访谈收敛后用 enter_plan_mode 进入计划模式，放弃规划时用 leave_plan_mode 退出。"),
 	dispatchExecutionGuard: z$1.boolean().default(true).volatile().description("启用调度执行守卫：在执行期拦截子代理越权调用主控专用工具，并拦截等待子代理期间的纯 sleep 空等命令；不修改工具声明面，零损耗保护 Kimi 缓存。"),
+	toolRunawayGuard: z$1.boolean().default(true).volatile().description("启用工具失控治理：重复调用同一工具与参数时分级提醒；Kimi 保留第 12 次强停，其他线路只提醒。"),
 	teamPolicyCacheGuard: z$1.boolean().default(true).volatile().description("把内置 Agent Teams 的 team:policy 段从 system prompt 摘掉，改为开关切换时以尾部用户消息投递一次，避免中途开团队协作时刷掉整段前缀缓存；默认开启。"),
 	cacheWarmup: z$1.boolean().default(true).volatile().description(CACHE_WARMUP_SETTING_DESCRIPTION),
+	cacheWarmupProviders: z$1.array(z$1.string()).default(CONTROLLED_PROVIDERS).volatile().description("允许产生新结果预览的线路；默认覆盖全部受控线路，空数组表示关闭所有新投影。"),
 	cacheWarmupRetrieval: z$1.boolean().default(true).volatile().description(CACHE_WARMUP_RETRIEVAL_SETTING_DESCRIPTION),
 	cacheWaitWarmup: z$1.boolean().default(true).volatile().description(CACHE_WAIT_WARMUP_SETTING_DESCRIPTION),
+	cacheAffinity: z$1.boolean().default(true).volatile().description(CACHE_AFFINITY_SETTING_DESCRIPTION),
+	promptCacheOverrides: z$1.object({}).default({}).description("按线路覆盖缓存语义（ttlMs、scope、coldStart、affinity、waitWarmup 等）；仅用于已有证据的部署，修改后重载插件。"),
 	codexStreamIdleTimeoutMs: z$1.number().min(1).max(MAX_TIMER_DELAY_MS).step(1).default(DEFAULT_STREAM_IDLE_TIMEOUT_MS$2),
 	poolStorage: z$1.union([z$1.const("keyring"), z$1.const("credentials")]).default("keyring").description("号池存储后端：keyring（默认，桌面平台的系统密钥环）或 credentials（DSH 凭据库，服务器推荐）。")
 });
@@ -105068,6 +105896,7 @@ const inject = [
 	"loader"
 ];
 function apply(ctx, pluginConfig = {}) {
+	const promptCacheOverrides = parsePromptCacheOverrides(pluginConfig.promptCacheOverrides);
 	installHostLocale(new HostLocaleRuntime());
 	ctx.effect(() => () => installHostLocale(void 0), "dsh-chatgpt-subscription: host locale");
 	const fetchConfiguration = resolveFetchConfiguration(pluginConfig);
@@ -105134,7 +105963,10 @@ function apply(ctx, pluginConfig = {}) {
 		return () => guard.dispose();
 	}, "dsh-chatgpt-subscription: compaction policy guard");
 	ctx.effect(() => installMinimaxRuntimeEnhancements(ctx), "dsh-chatgpt-subscription: MiniMax runtime enhancements");
-	ctx.effect(() => installKimiRepeatBreaker(ctx), "dsh-chatgpt-subscription: Kimi repeat breaker");
+	ctx.effect(() => installKimiRepeatBreaker(ctx, {
+		providers: CONTROLLED_PROVIDERS.filter((provider) => provider !== "minimax-code"),
+		globalEnabled: () => resolveToolRunawayGuardEnabled(pluginConfig)
+	}), "dsh-chatgpt-subscription: shared tool runaway guard");
 	ctx.effect(() => installDispatchExecutionGuard(ctx, { enabled: () => resolveDispatchExecutionGuardEnabled(pluginConfig) }), "dsh-chatgpt-subscription: dispatch execution guard");
 	ctx.effect(() => installDispatchLedger(ctx, {
 		enabled: () => resolveDispatchExecutionGuardEnabled(pluginConfig),
@@ -105247,7 +106079,8 @@ function apply(ctx, pluginConfig = {}) {
 	}
 	ctx.effect(() => installCacheWarmup(ctx, {
 		settings: () => resolveCacheWarmupSettings(pluginConfig),
-		durableHint: (sessionId) => {
+		durableHint: (sessionId, provider) => {
+			if (provider !== void 0 && provider !== "kimi-code") return void 0;
 			const record = sessionCacheState.read(sessionId);
 			if (record === void 0) return void 0;
 			const total = record.cachedTokens + record.freshTokens;
@@ -105259,7 +106092,10 @@ function apply(ctx, pluginConfig = {}) {
 		}
 	}), "dsh-chatgpt-subscription: cache warmup");
 	ctx.effect(() => {
-		const dispose = installKimiCacheWait(ctx, { warmup: () => resolveCacheWaitWarmupEnabled(pluginConfig, process.env) });
+		const dispose = installKimiCacheWait(ctx, {
+			warmup: () => resolveCacheWaitWarmupEnabled(pluginConfig, process.env),
+			overrides: promptCacheOverrides
+		});
 		if (dispose === void 0) {
 			ctx.logger.warn("[dsh-chatgpt-subscription] Kimi cache wait skipped: this harness exposes no event bus");
 			return () => void 0;
@@ -105304,7 +106140,8 @@ function apply(ctx, pluginConfig = {}) {
 			label: PROVIDER_NAME$11,
 			adapter: new AntigravityAdapter(antigravityStore, antigravityModelSettings, antigravityPreferences, {
 				fetchFn: modelFetch.antigravity,
-				attachments: ctx.attachments
+				attachments: ctx.attachments,
+				promptCache: promptCacheOverrides.antigravity
 			}, antigravityAccountPool)
 		});
 		const disposeAntigravityRoutes = registerAntigravityRoutes(ctx, antigravityStore, antigravityModelSettings, antigravityPreferences, antigravityProxyFetch, antigravityAccountPool);
@@ -105333,7 +106170,8 @@ function apply(ctx, pluginConfig = {}) {
 				videos: { readVideo: async (ref) => ({
 					data: await readVideoBytes(ref),
 					mediaType: ref.mediaType
-				}) }
+				}) },
+				cacheAffinity: () => resolveCacheAffinityEnabled(pluginConfig, process.env)
 			}, kimiCodeAccountPool)
 		});
 		const ollamaClaim = claimProviderRoute(ctx, {
@@ -105378,7 +106216,8 @@ function apply(ctx, pluginConfig = {}) {
 			adapter: new MinimaxCodeAdapter(minimaxCodeStore, {
 				fetchFn: modelFetch["minimax-code"],
 				attachments: ctx.attachments,
-				accountPool: minimaxCodeAccountPool
+				accountPool: minimaxCodeAccountPool,
+				promptCache: promptCacheOverrides["minimax-code"]
 			}, minimaxCodeModelSettings, minimaxCodePreferences)
 		});
 		let minimaxCodeAppVersion;
@@ -105415,7 +106254,8 @@ function apply(ctx, pluginConfig = {}) {
 			adapter: new ClaudeAdapter(claudeStore, claudeModelSettings, claudePreferences, {
 				fetchFn: modelFetch.claude,
 				attachments: ctx.attachments,
-				accountPool: claudeAccountPool
+				accountPool: claudeAccountPool,
+				promptCache: promptCacheOverrides.claude
 			})
 		});
 		const disposeClaudeRoutes = registerClaudeRoutes(ctx, claudeStore, claudeModelSettings, claudePreferences, {
@@ -105430,14 +106270,16 @@ function apply(ctx, pluginConfig = {}) {
 			adapter: new StepCodeAdapter(stepCodeStore, stepCodeModelSettings, stepCodePreferences, {
 				fetchFn: modelFetch[PROVIDER_ID$6],
 				attachments: ctx.attachments,
-				accountPool: stepCodeAccountPool
+				accountPool: stepCodeAccountPool,
+				promptCache: promptCacheOverrides["step-code"]
 			})
 		});
 		const disposeStepCodeRoutes = registerStepCodeRoutes(ctx, stepCodeSettingsSource(stepCodePreferences, stepCodeModelSettings), {
 			fetchFn: proxyFetch,
 			serving: () => stepCodeClaim.serving(),
 			conflict: () => stepCodeClaim.conflict(),
-			accountPool: stepCodeAccountPool
+			accountPool: stepCodeAccountPool,
+			promptCache: promptCacheOverrides["step-code"]
 		});
 		const clinePassLine = createClinePassLine(ctx, {
 			fetchFn: proxyFetch,
@@ -105551,7 +106393,8 @@ function apply(ctx, pluginConfig = {}) {
 			outputVerbosity: () => preferences.status().outputVerbosity,
 			fastMode: () => preferences.status().fastMode,
 			reasoningSummary: () => preferences.status().reasoningSummary,
-			codexStreamIdleTimeoutMs: pluginConfig.codexStreamIdleTimeoutMs
+			codexStreamIdleTimeoutMs: pluginConfig.codexStreamIdleTimeoutMs,
+			promptCache: promptCacheOverrides["codex-chatgpt"]
 		}), preferences, (options) => oauth.credentials(void 0, { purpose: "tool" }).then((credentials) => loadCodexCatalog(credentials, {
 			fetchFn: proxyFetch,
 			signal: options?.signal