@eddyskywalker/dsh-chatgpt-subscription 0.66.0 → 0.66.1

package/lib/index.js

@@ -809,7 +809,7 @@ const ROUTE_PREFIX$5 = "/api/dsh-chatgpt-subscription";
 * `test/responses-client.test.ts` asserts the two agree, so a release that
 * forgets this constant fails the suite instead of shipping quietly.
 */
-const PLUGIN_VERSION = "0.66.0";
+const PLUGIN_VERSION = "0.66.1";
 const CODEX_CHATGPT_PROVIDER_ID = "codex-chatgpt";
 const CODEX_API_BASE = "https://chatgpt.com/backend-api/codex";
 const CODEX_RESPONSES_URL = `${CODEX_API_BASE}/responses`;
@@ -15092,7 +15092,7 @@ async function* parseResponsesStream(response, signal, _legacyHiddenSandboxContr
 	const tools = /* @__PURE__ */ new Map();
 	const toolFor = (event, item) => {
 		const itemId = string(event.item_id) ?? string(item?.id);
-		const outputIndex = number(event.output_index);
+		const outputIndex = number$1(event.output_index);
 		const key = itemId ?? (outputIndex === void 0 ? `tool-${tools.size}` : `index-${outputIndex}`);
 		let tool = tools.get(key);
 		if (tool === void 0) {
@@ -15421,12 +15421,12 @@ function visibleReasoningDelta(delta, currentVisibleChars, alreadyTruncated) {
 */
 function mapUsage$2(value) {
 	if (value === null) return null;
-	const totalInput = number(value.input_tokens) ?? 0;
-	const outputTokens = number(value.output_tokens) ?? 0;
+	const totalInput = number$1(value.input_tokens) ?? 0;
+	const outputTokens = number$1(value.output_tokens) ?? 0;
 	const details = record$3(value.input_tokens_details);
-	const cached = number(details?.cached_tokens);
-	const written = number(details?.cache_write_tokens);
-	const reasoning = number(record$3(value.output_tokens_details)?.reasoning_tokens);
+	const cached = number$1(details?.cached_tokens);
+	const written = number$1(details?.cache_write_tokens);
+	const reasoning = number$1(record$3(value.output_tokens_details)?.reasoning_tokens);
 	return {
 		inputTokens: Math.max(0, totalInput - (cached ?? 0) - (written ?? 0)),
 		outputTokens,
@@ -15481,7 +15481,7 @@ function record$3(value) {
 function string(value) {
 	return typeof value === "string" ? value : void 0;
 }
-function number(value) {
+function number$1(value) {
 	const parsed = Number(value);
 	return Number.isFinite(parsed) ? parsed : void 0;
 }
@@ -16579,8 +16579,8 @@ function parseCodexUsage(value) {
 	for (const [index, value] of additional.entries()) {
 		const limit = record$2(value);
 		if (limit === null) continue;
-		const idSource = text$4(limit.limit_name) ?? text$4(limit.metered_feature) ?? `additional-${index + 1}`;
-		addBucket(buckets, usedIds, uniqueId(slug(idSource), usedIds), readableLimitName(text$4(limit.limit_name) ?? text$4(limit.metered_feature) ?? idSource), planType, limit.rate_limit);
+		const idSource = text$5(limit.limit_name) ?? text$5(limit.metered_feature) ?? `additional-${index + 1}`;
+		addBucket(buckets, usedIds, uniqueId(slug(idSource), usedIds), readableLimitName(text$5(limit.limit_name) ?? text$5(limit.metered_feature) ?? idSource), planType, limit.rate_limit);
 	}
 	return {
 		buckets,
@@ -16665,7 +16665,7 @@ function parseResetCredits(value) {
 		availableCreditIds: []
 	};
 	const available = (Array.isArray(data.credits) ? data.credits : []).map(record$2).filter((credit) => credit !== null && credit.status === "available").map((credit) => ({
-		id: text$4(credit.id),
+		id: text$5(credit.id),
 		expiresAt: timestamp(credit.expires_at)
 	})).filter((credit) => credit.id !== null).sort((left, right) => (left.expiresAt ?? Number.MAX_SAFE_INTEGER) - (right.expiresAt ?? Number.MAX_SAFE_INTEGER));
 	const reported = numeric(data.available_count);
@@ -16699,7 +16699,7 @@ function decimalText(value) {
 	const trimmed = value.trim();
 	return trimmed !== "" ? trimmed : null;
 }
-function text$4(value) {
+function text$5(value) {
 	return typeof value === "string" && value.trim() !== "" ? value.trim() : null;
 }
 function slug(value) {
@@ -24922,7 +24922,7 @@ function classifyCommandCodeFailure(status, detail) {
 	};
 }
 /** Cooldown one rate-limited key takes when the provider states no delay. */
-const POOL_COOLDOWN_MS$7 = 15 * 6e4;
+const POOL_COOLDOWN_MS$6 = 15 * 6e4;
 var CommandCodeAdapter = class extends LlmAdapter {
 	store;
 	modelSettings;
@@ -25110,7 +25110,7 @@ var CommandCodeAdapter = class extends LlmAdapter {
 			}
 			if (response.status === 429) {
 				const after = retryAfterMs$1(response.headers);
-				await pool.markCooldown(accountId, after ?? POOL_COOLDOWN_MS$7, `${PROVIDER_NAME$10} 429`).catch(() => void 0);
+				await pool.markCooldown(accountId, after ?? POOL_COOLDOWN_MS$6, `${PROVIDER_NAME$10} 429`).catch(() => void 0);
 				if (await pool.hasAnotherAvailableAccount(tried)) continue;
 				break;
 			}
@@ -25185,7 +25185,7 @@ var CommandCodeAdapter = class extends LlmAdapter {
 					const reclassified = reclassifyInBandStreamError(error, inBand);
 					if (reclassified.code === "RATE_LIMIT" && pool !== null && accountId !== void 0) {
 						const after = retryAfterMs$1(response.headers);
-						await pool.markCooldown(accountId, after ?? POOL_COOLDOWN_MS$7, `${PROVIDER_NAME$10} in-stream 429`).catch(() => void 0);
+						await pool.markCooldown(accountId, after ?? POOL_COOLDOWN_MS$6, `${PROVIDER_NAME$10} in-stream 429`).catch(() => void 0);
 					}
 					throw reclassified;
 				}
@@ -26371,7 +26371,7 @@ const STREAM_IDLE_TIMEOUT_MS$5 = 3e5;
 * A 429 here is a real quota or rate limit, not a blip, so the account is parked
 * for long enough that the next request goes to a different one.
 */
-const POOL_COOLDOWN_MS$6 = 15 * 6e4;
+const POOL_COOLDOWN_MS$5 = 15 * 6e4;
 /**
 * NO fallback model names.
 *
@@ -27766,7 +27766,7 @@ var OllamaAdapter = class extends LlmAdapter {
 				break;
 			}
 			if (lastStatus === 429) {
-				await pool.markCooldown(accountId, POOL_COOLDOWN_MS$6, `${PROVIDER_NAME$9} 429`).catch(() => void 0);
+				await pool.markCooldown(accountId, POOL_COOLDOWN_MS$5, `${PROVIDER_NAME$9} 429`).catch(() => void 0);
 				if (await pool.hasAnotherAvailableAccount(tried)) continue;
 				break;
 			}
@@ -29294,7 +29294,7 @@ function consecutiveWarmupHits(reports) {
 	}
 }
 /** 把任意输入归一成一条用量报告；形状不对返回 undefined（绝不抛）。 */
-function normalizeUsage(input) {
+function normalizeUsage$1(input) {
 	if (input === null || input === void 0 || typeof input !== "object") return void 0;
 	const sessionId = input.sessionId;
 	if (typeof sessionId !== "string" || sessionId.trim() === "") return void 0;
@@ -29335,7 +29335,7 @@ function recordCacheWarmupUsage(input) {
 	let sessionId = "";
 	let provider;
 	try {
-		const usage = normalizeUsage(input);
+		const usage = normalizeUsage$1(input);
 		if (usage === void 0) return neutralState("");
 		sessionId = usage.sessionId;
 		provider = usage.provider;
@@ -30799,6 +30799,57 @@ function planMultiSectionProjection(text) {
 		sections
 	};
 }
+/**
+* PTC 工具目录：保留完整的 namespace/name 索引，重 schema 由原文 ref 补读。
+*
+* 只认 code_tools_list / code_tools_search 的根对象形状。任意字段缺失、额外根成员
+* 或非 PTC 调用都回退原文，避免把业务 JSON 里的 tools 数组误当成工具发现结果。
+*/
+function planToolCatalogProjection(text, toolName) {
+	if (!isPtcTransportName(toolName)) return void 0;
+	const candidate = parseJsonCandidate(text);
+	if (candidate?.kind !== "object") return void 0;
+	const record = candidate.record;
+	if (Object.keys(record).some((key) => key !== "tools" && key !== "fallback")) return void 0;
+	if (!Array.isArray(record.tools) || record.tools.length === 0) return void 0;
+	const lines = [];
+	for (const entry of record.tools) {
+		if (!isPlainRecord$1(entry)) return void 0;
+		if (typeof entry.name !== "string" || entry.name.trim() === "") return void 0;
+		if (typeof entry.namespace !== "string" || entry.namespace.trim() === "") return void 0;
+		if (typeof entry.description !== "string") return void 0;
+		const name = entry.name.trim();
+		const namespace = entry.namespace.trim();
+		const description = entry.description.replace(/\s+/g, " ").trim();
+		lines.push(`  ${namespace}/${name}: ${clip(description, 180)}`);
+	}
+	const header = [`tools: ${record.tools.length} (complete name index; descriptions clipped)`];
+	if (typeof record.fallback === "string" && record.fallback.trim() !== "") header.push("fallback: " + clip(record.fallback.replace(/\s+/g, " ").trim(), 240));
+	return {
+		kind: "preview",
+		strategy: "tool-catalog",
+		sections: [{ lines: header }, { lines }]
+	};
+}
+/** 成功测试日志必须有明确通过汇总，且没有任何可识别的失败证据。 */
+function isSuccessfulTestEvidence(text, evidence) {
+	if (evidence.failures.length > 0 || evidence.summary.length === 0) return false;
+	if (!evidence.summary.some((line) => /\b\d+\s+passed\b/i.test(line))) return false;
+	if (/(?:^|\s)[1-9]\d*\s+failed\b/i.test(text)) return false;
+	const exitMatch = /(?:exit code|exited with code|process exited with code)[:=\s]+(-?\d+)/i.exec(text) ?? /\[exit code: (-?\d+)\]/i.exec(text);
+	if (exitMatch !== null && Number(exitMatch[1]) !== 0) return false;
+	for (const raw of text.split("\n")) {
+		const trimmed = raw.trim();
+		if (trimmed === "" || /^npm notice\b/i.test(trimmed)) continue;
+		if (/^(?:PASS\b|[✓✔]|ok\s+\d+)/i.test(trimmed)) continue;
+		if (/^(?:stdout|stderr)\s*\|/i.test(trimmed)) continue;
+		if (/^\(node:\d+\).*Warning:/i.test(trimmed)) continue;
+		if (/^\s*(?:0|no)\s+(?:errors?|failures?|failed)\b/i.test(trimmed)) continue;
+		if (/\b\d+\s+passed\b/i.test(trimmed)) continue;
+		if (SHELL_ERROR_LINE.test(trimmed)) return false;
+	}
+	return true;
+}
 /** 对一段纯文本（非 JSON 文档）做投影计划判定。 */
 function planPlainTextProjection(rawText, toolName) {
 	const text = stripAnsi(rawText);
@@ -30816,6 +30867,11 @@ function planPlainTextProjection(rawText, toolName) {
 		strategy: "test-failure",
 		sections: buildTestSections(test)
 	};
+	if (test !== void 0 && isSuccessfulTestEvidence(text, test)) return {
+		kind: "preview",
+		strategy: "test-success",
+		sections: buildTestSections(test)
+	};
 	const grep = extractGrepEvidence(text);
 	if (grep !== void 0) return {
 		kind: "preview",
@@ -30886,6 +30942,10 @@ function planProjection(text, toolName, navigation, retrievalOnly, spilled = fal
 		const multiSection = planMultiSectionProjection(text);
 		if (multiSection !== void 0) return multiSection;
 	}
+	if (isJsonDocument(text)) {
+		const toolCatalog = planToolCatalogProjection(text, toolName);
+		if (toolCatalog !== void 0) return toolCatalog;
+	}
 	if (looksLikeContract(text)) return {
 		kind: "verbatim",
 		strategy: "verbatim-contract"
@@ -34470,6 +34530,150 @@ function videoBlockToInline(block, videos, videoAccepted, omit) {
 	} } : { omission: omit("unsupported-container", label) };
 }
 //#endregion
+//#region src/host/kimi-code/tool-argument-guard.ts
+const TOOL_ARGUMENT_BYTES_EXCEEDED_CODE = "KIMI_TOOL_ARGUMENT_BYTES_EXCEEDED";
+const TOOL_ARGUMENT_REPEATED_KEY_CODE = "KIMI_TOOL_ARGUMENT_REPEATED_KEY";
+/** Whether one adapter failure belongs to the recoverable argument guard. */
+function isKimiToolArgumentGuardCode(code) {
+	return code === "KIMI_TOOL_ARGUMENT_BYTES_EXCEEDED" || code === "KIMI_TOOL_ARGUMENT_REPEATED_KEY";
+}
+const ASK_USER_ARGUMENT_LIMIT = 32 * 1024;
+const LARGE_PROGRAM_ARGUMENT_LIMIT = 1024 * 1024;
+const DEFAULT_ARGUMENT_LIMIT = 256 * 1024;
+const REPEATED_KEY_LIMIT = 8;
+const MAX_TRACKED_KEY_GROUPS = 64;
+const MAX_SCANNED_KEY_CHARS = 1024;
+/** Large source/file payloads need a different ceiling from coordination cards. */
+function toolArgumentLimit(toolName) {
+	if (toolName === "ask_user_question") return ASK_USER_ARGUMENT_LIMIT;
+	if (toolName === "write" || toolName === "edit" || toolName === "code_run" || toolName === "run_code") return LARGE_PROGRAM_ARGUMENT_LIMIT;
+	return DEFAULT_ARGUMENT_LIMIT;
+}
+/** Split `header_comment`, `header_unused`, and `header___` into one family. */
+function keyFamily(key) {
+	const suffix = /_(?:comment|unused)$|_+$/.exec(key);
+	if (suffix === null || suffix.index === 0) return {
+		stem: key,
+		variant: false
+	};
+	return {
+		stem: key.slice(0, suffix.index),
+		variant: true
+	};
+}
+var KimiToolArgumentGuard = class {
+	toolName;
+	bytes = 0;
+	inString = false;
+	escaped = false;
+	afterString = false;
+	stringRaw = "";
+	stringTooLong = false;
+	highSurrogate = "";
+	keyGroups = /* @__PURE__ */ new Map();
+	violation;
+	constructor(toolName) {
+		this.toolName = toolName;
+	}
+	observe(delta) {
+		if (this.violation !== void 0) return this.violation;
+		const text = this.highSurrogate + delta;
+		this.highSurrogate = "";
+		if (text === "") return void 0;
+		const last = text.charCodeAt(text.length - 1);
+		const complete = last >= 55296 && last <= 56319 ? text.slice(0, -1) : text;
+		if (complete.length !== text.length) this.highSurrogate = text.at(-1) ?? "";
+		this.bytes += Buffer.byteLength(complete, "utf8");
+		if (this.bytes > toolArgumentLimit(this.toolName)) return this.fail("argument-bytes");
+		for (const char of complete) {
+			const violation = this.observeChar(char);
+			if (violation !== void 0) return violation;
+		}
+	}
+	/** Account for a trailing unmatched surrogate without leaving it unbudgeted. */
+	finish() {
+		if (this.highSurrogate !== "") {
+			this.highSurrogate = "";
+			this.bytes += Buffer.byteLength("\ud800", "utf8");
+			if (this.bytes > toolArgumentLimit(this.toolName)) return this.fail("argument-bytes");
+		}
+		return this.violation;
+	}
+	fail(rule) {
+		this.violation = {
+			toolName: this.toolName || "unknown",
+			rule,
+			bytes: this.bytes
+		};
+		return this.violation;
+	}
+	observeChar(char) {
+		if (this.inString) {
+			if (this.escaped) {
+				this.escaped = false;
+				if (!this.stringTooLong) this.stringRaw += char;
+				return;
+			}
+			if (char === "\\") {
+				this.escaped = true;
+				if (!this.stringTooLong) this.stringRaw += char;
+				return;
+			}
+			if (char === "\"") {
+				this.inString = false;
+				this.afterString = true;
+				return;
+			}
+			if (!this.stringTooLong) {
+				this.stringRaw += char;
+				if (this.stringRaw.length > MAX_SCANNED_KEY_CHARS) {
+					this.stringTooLong = true;
+					this.stringRaw = "";
+				}
+			}
+			return;
+		}
+		if (this.afterString) {
+			if (/\s/.test(char)) return void 0;
+			this.afterString = false;
+			if (char === ":") {
+				this.observeKey();
+				return this.violation;
+			}
+			this.stringRaw = "";
+			this.stringTooLong = false;
+		}
+		if (char === "\"") {
+			this.inString = true;
+			this.escaped = false;
+			this.stringRaw = "";
+			this.stringTooLong = false;
+		}
+	}
+	observeKey() {
+		if (this.stringTooLong || this.stringRaw === "") return;
+		try {
+			const key = JSON.parse("\"" + this.stringRaw + "\"");
+			if (typeof key !== "string" || key === "") return;
+			const family = keyFamily(key);
+			if (family.stem === "") return;
+			const group = this.keyGroups.get(family.stem) ?? {
+				count: 0,
+				sawVariant: false
+			};
+			group.count += 1;
+			group.sawVariant ||= family.variant;
+			this.keyGroups.delete(family.stem);
+			this.keyGroups.set(family.stem, group);
+			while (this.keyGroups.size > MAX_TRACKED_KEY_GROUPS) {
+				const oldest = this.keyGroups.keys().next();
+				if (oldest.done !== true) this.keyGroups.delete(oldest.value);
+			}
+			if (group.sawVariant && group.count >= REPEATED_KEY_LIMIT) this.fail("repeated-key");
+		} catch {}
+	}
+};
+//#endregion
 //#region src/host/kimi-code/mapper.ts
 /**
 * Provider-wire mapping for the Kimi Code coding endpoints.
@@ -36154,6 +36358,9 @@ function createStreamState$5(wire, sessionId, sideChannel = false) {
 		cacheWrite5mTokens: 0,
 		cacheWrite1hTokens: 0,
 		reasoningTokens: 0,
+		cacheReadReported: false,
+		cacheWriteReported: false,
+		reasoningReported: false,
 		sawUsage: false,
 		toolCallIds: new ToolCallIdNormalizer()
 	};
@@ -36176,9 +36383,22 @@ function closeCurrent$5(state) {
 		block
 	}];
 }
+function throwToolGuardViolation(state, call, violation) {
+	if (violation === void 0) return;
+	state.toolGuard = violation;
+	const reason = violation.rule === "argument-bytes" ? `arguments exceeded the ${violation.rule} safety budget` : "arguments contained a runaway repeated-key family";
+	throw new LlmError$1(`Kimi Code stopped tool call "${violation.toolName}": ${reason} at ${violation.bytes} bytes. The tool was not executed.`, violation.rule === "argument-bytes" ? TOOL_ARGUMENT_BYTES_EXCEEDED_CODE : TOOL_ARGUMENT_REPEATED_KEY_CODE);
+}
+function appendGuardedArguments(state, call, delta) {
+	if (delta === "") return;
+	call.arguments += delta;
+	call.guard ??= new KimiToolArgumentGuard(call.name);
+	throwToolGuardViolation(state, call, call.guard.observe(delta));
+}
 function closeToolCalls$5(state) {
 	const out = [];
 	for (const [wireIndex, call] of [...state.toolCalls.entries()].sort((a, b) => a[0] - b[0])) {
+		throwToolGuardViolation(state, call, call.guard?.finish());
 		const block = {
 			type: "tool-call",
 			id: toToolCallId(state.toolCallIds.normalize(call.id)),
@@ -36247,8 +36467,14 @@ function processOpenAIStreamLine(line, state) {
 		state.inputTokens = Math.max(0, prompt - cached - written);
 		state.cacheReadTokens = cached;
 		state.cacheWriteTokens = written;
+		state.cacheReadReported = details !== void 0 && typeof details.cached_tokens === "number" && Number.isFinite(details.cached_tokens);
+		state.cacheWriteReported = details !== void 0 && typeof details.cache_write_tokens === "number" && Number.isFinite(details.cache_write_tokens);
 		state.outputTokens = numberOr$8(usage.completion_tokens, state.outputTokens);
-		if (isRecord$53(usage.completion_tokens_details)) state.reasoningTokens = numberOr$8(usage.completion_tokens_details.reasoning_tokens, state.reasoningTokens);
+		if (isRecord$53(usage.completion_tokens_details)) {
+			const reasoning = usage.completion_tokens_details.reasoning_tokens;
+			state.reasoningReported = typeof reasoning === "number" && Number.isFinite(reasoning);
+			state.reasoningTokens = numberOr$8(reasoning, state.reasoningTokens);
+		} else state.reasoningReported = false;
 	}
 	const choices = Array.isArray(chunk.choices) ? chunk.choices : [];
 	const choice = isRecord$53(choices[0]) ? choices[0] : void 0;
@@ -36352,7 +36578,7 @@ function applyOpenAIToolDelta(entry, state) {
 		if (name !== void 0 && name !== "") call.name = name;
 	}
 	const argsDelta = asString$21(fn.arguments) ?? "";
-	if (argsDelta !== "") call.arguments += argsDelta;
+	appendGuardedArguments(state, call, argsDelta);
 	if (!call.started) {
 		call.started = true;
 		state.hasToolCall = true;
@@ -36390,7 +36616,13 @@ function processAnthropicStreamLine(line, state) {
 			state.inputTokens = numberOr$8(usage.input_tokens, 0);
 			state.cacheReadTokens = numberOr$8(usage.cache_read_input_tokens, 0);
 			state.cacheWriteTokens = numberOr$8(usage.cache_creation_input_tokens, 0);
+			state.cacheReadReported = typeof usage.cache_read_input_tokens === "number" && Number.isFinite(usage.cache_read_input_tokens);
+			state.cacheWriteReported = typeof usage.cache_creation_input_tokens === "number" && Number.isFinite(usage.cache_creation_input_tokens);
 			state.outputTokens = numberOr$8(usage.output_tokens, 0);
+			const reasoningDetails = isRecord$53(usage.output_tokens_details) ? usage.output_tokens_details : void 0;
+			const reasoning = usage.reasoning_tokens ?? reasoningDetails?.reasoning_tokens;
+			state.reasoningReported = typeof reasoning === "number" && Number.isFinite(reasoning);
+			state.reasoningTokens = numberOr$8(reasoning, state.reasoningTokens);
 		}
 		const stop = message ? asString$21(message.stop_reason) : void 0;
 		if (stop !== void 0) state.finishReason = stop;
@@ -36454,7 +36686,7 @@ function processAnthropicStreamLine(line, state) {
 			const pending = state.toolCalls.get(contentIndex);
 			const partial = asString$21(delta.partial_json) ?? "";
 			if (pending !== void 0) {
-				pending.arguments += partial;
+				appendGuardedArguments(state, pending, partial);
 				out.push({
 					type: "tool-call-delta",
 					index: pending.blockIndex,
@@ -36502,6 +36734,7 @@ function processAnthropicStreamLine(line, state) {
 		const contentIndex = numberOr$8(event.index, 0);
 		const pending = state.toolCalls.get(contentIndex);
 		if (pending !== void 0) {
+			throwToolGuardViolation(state, pending, pending.guard?.finish());
 			state.toolCalls.delete(contentIndex);
 			const block = {
 				type: "tool-call",
@@ -36528,6 +36761,12 @@ function processAnthropicStreamLine(line, state) {
 		if (usage) {
 			state.sawUsage = true;
 			state.outputTokens = numberOr$8(usage.output_tokens, state.outputTokens);
+			const reasoningDetails = isRecord$53(usage.output_tokens_details) ? usage.output_tokens_details : void 0;
+			const reasoning = usage.reasoning_tokens ?? reasoningDetails?.reasoning_tokens;
+			if (typeof reasoning === "number" && Number.isFinite(reasoning)) {
+				state.reasoningReported = true;
+				state.reasoningTokens = reasoning;
+			}
 		}
 		return out;
 	}
@@ -36971,7 +37210,7 @@ function closeStream$4(state, accountId) {
 	state.finished = true;
 	const out = [...closeCurrent$5(state), ...closeToolCalls$5(state)];
 	if (state.sawUsage) {
-		recordCacheStats(state, state.sessionId, accountId);
+		recordCacheStats(state, state.sessionId, accountId ?? state.accountId);
 		out.push({
 			type: "usage",
 			usage: tokenUsage$5(state)
@@ -37119,6 +37358,27 @@ var SessionCacheStateStore = class extends CommonSessionCacheStateStore {
 			normalizeRecord: normalizeRecord$2
 		});
 	}
+	/** Append bounded request evidence, including a first request that failed before usage. */
+	noteRequest(sessionId, request) {
+		const base = this.read(sessionId) ?? {
+			lastActiveAt: request.startedAt,
+			totalTokens: 0,
+			requests: 0,
+			cachedTokens: 0,
+			freshTokens: 0,
+			outputTokens: 0,
+			cacheWriteTokens: 0,
+			cacheWrite5mTokens: 0,
+			cacheWrite1hTokens: 0,
+			updatedAt: request.endedAt
+		};
+		this.write(sessionId, {
+			...base,
+			lastActiveAt: Math.max(base.lastActiveAt, request.endedAt),
+			recentRequests: [...base.recentRequests ?? [], request].slice(-20),
+			updatedAt: request.endedAt
+		});
+	}
 	/** Update the warmup streak without creating a record for an unknown session. */
 	noteWarmupStreak(sessionId, consecutiveHits) {
 		this.update(sessionId, (existing) => {
@@ -37133,6 +37393,94 @@ var SessionCacheStateStore = class extends CommonSessionCacheStateStore {
 		});
 	}
 };
+function optionalText(value, maxLength = 256) {
+	return typeof value === "string" && value.trim() !== "" ? value.trim().slice(0, maxLength) : void 0;
+}
+function nullableCount(value) {
+	return value === null ? null : nonNegative(value);
+}
+function normalizeAttempt(value) {
+	if (!isRecord$52(value)) return void 0;
+	const startedAt = finite(value.startedAt);
+	if (startedAt === void 0) return void 0;
+	const poolAction = value.poolAction === "retry-same-account" || value.poolAction === "cooldown-and-rotate" || value.poolAction === "mark-auth-failed" || value.poolAction === "none" ? value.poolAction : void 0;
+	const status = nonNegative(value.status);
+	const endedAt = finite(value.endedAt);
+	const retryAfterMs = nonNegative(value.retryAfterMs);
+	const accountId = optionalText(value.accountId);
+	const code = optionalText(value.code);
+	const providerRequestId = optionalText(value.providerRequestId, 128);
+	return {
+		startedAt,
+		...accountId === void 0 ? {} : { accountId },
+		...endedAt === void 0 ? {} : { endedAt },
+		...status === void 0 ? {} : { status },
+		...code === void 0 ? {} : { code },
+		...poolAction === void 0 ? {} : { poolAction },
+		...retryAfterMs === void 0 ? {} : { retryAfterMs },
+		...providerRequestId === void 0 ? {} : { providerRequestId }
+	};
+}
+function normalizeUsage(value) {
+	if (!isRecord$52(value)) return void 0;
+	const freshTokens = nonNegative(value.freshTokens);
+	const outputTokens = nonNegative(value.outputTokens);
+	const cacheReadTokens = nullableCount(value.cacheReadTokens);
+	const cacheWriteTokens = nullableCount(value.cacheWriteTokens);
+	const reasoningTokens = nullableCount(value.reasoningTokens);
+	if (freshTokens === void 0 || outputTokens === void 0 || cacheReadTokens === void 0 || cacheWriteTokens === void 0 || reasoningTokens === void 0) return void 0;
+	return {
+		freshTokens,
+		cacheReadTokens,
+		cacheWriteTokens,
+		outputTokens,
+		reasoningTokens
+	};
+}
+function normalizeRequestDiagnostic(value) {
+	if (!isRecord$52(value)) return void 0;
+	const id = optionalText(value.id, 128);
+	const model = optionalText(value.model, 256);
+	const startedAt = finite(value.startedAt);
+	const endedAt = finite(value.endedAt);
+	const wire = value.wire === "openai" || value.wire === "anthropic" ? value.wire : void 0;
+	const outcome = value.outcome === "success" || value.outcome === "error" || value.outcome === "aborted" ? value.outcome : void 0;
+	if (id === void 0 || model === void 0 || startedAt === void 0 || endedAt === void 0 || wire === void 0 || outcome === void 0) return void 0;
+	const usage = normalizeUsage(value.usage);
+	const attempts = (Array.isArray(value.attempts) ? value.attempts : []).map(normalizeAttempt).filter((attempt) => attempt !== void 0);
+	const accountId = optionalText(value.accountId);
+	const previousAccountId = optionalText(value.previousAccountId);
+	const code = optionalText(value.code);
+	const finishReason = optionalText(value.finishReason);
+	const guard = isRecord$52(value.toolGuard) ? value.toolGuard : void 0;
+	const guardName = optionalText(guard?.toolName);
+	const guardBytes = nonNegative(guard?.bytes);
+	const guardRule = guard?.rule === "argument-bytes" || guard?.rule === "repeated-key" ? guard.rule : void 0;
+	return {
+		id,
+		startedAt,
+		endedAt,
+		model,
+		wire,
+		accountChanged: value.accountChanged === true,
+		outcome,
+		attempts,
+		...accountId === void 0 ? {} : { accountId },
+		...previousAccountId === void 0 ? {} : { previousAccountId },
+		...code === void 0 ? {} : { code },
+		...finishReason === void 0 ? {} : { finishReason },
+		...usage === void 0 ? {} : { usage },
+		...guardName === void 0 || guardBytes === void 0 || guardRule === void 0 ? {} : { toolGuard: {
+			toolName: guardName,
+			bytes: guardBytes,
+			rule: guardRule
+		} }
+	};
+}
+function normalizeRecentRequests(value) {
+	if (!Array.isArray(value)) return [];
+	return value.map(normalizeRequestDiagnostic).filter((request) => request !== void 0).slice(-20);
+}
 function normalizeRecord$2(value) {
 	if (!isRecord$52(value)) return void 0;
 	const lastActiveAt = finite(value.lastActiveAt);
@@ -37141,10 +37489,16 @@ function normalizeRecord$2(value) {
 	const drift = typeof value.lastDrift === "string" && value.lastDrift !== "" ? value.lastDrift : void 0;
 	const firstRequestCachedTokens = nonNegative(value.firstRequestCachedTokens);
 	const firstRequestPromptTokens = nonNegative(value.firstRequestPromptTokens);
+	const lastRequestCachedTokens = nonNegative(value.lastRequestCachedTokens);
+	const lastRequestFreshTokens = nonNegative(value.lastRequestFreshTokens);
+	const previousRequestPromptTokens = nonNegative(value.previousRequestPromptTokens);
+	const lastAccountId = typeof value.lastAccountId === "string" && value.lastAccountId.trim() !== "" ? value.lastAccountId.trim() : void 0;
+	const lastAccountAt = nonNegative(value.lastAccountAt);
 	const warmupStreak = count$1(value.warmupConsecutiveHits);
 	const budgetDropped = droppedImageKeys(value.budgetDropped);
 	const lastRequestImageBytes = nonNegative(value.lastRequestImageBytes);
 	const lastDroppedImages = nonNegative(value.lastDroppedImages);
+	const recentRequests = normalizeRecentRequests(value.recentRequests);
 	return {
 		lastActiveAt,
 		updatedAt,
@@ -37159,6 +37513,12 @@ function normalizeRecord$2(value) {
 		...drift === void 0 ? {} : { lastDrift: drift },
 		...firstRequestCachedTokens === void 0 ? {} : { firstRequestCachedTokens },
 		...firstRequestPromptTokens === void 0 ? {} : { firstRequestPromptTokens },
+		...lastRequestCachedTokens === void 0 ? {} : { lastRequestCachedTokens },
+		...lastRequestFreshTokens === void 0 ? {} : { lastRequestFreshTokens },
+		...previousRequestPromptTokens === void 0 ? {} : { previousRequestPromptTokens },
+		...lastAccountId === void 0 ? {} : { lastAccountId },
+		...lastAccountAt === void 0 ? {} : { lastAccountAt },
+		...recentRequests.length > 0 ? { recentRequests } : {},
 		...warmupStreak > 0 ? { warmupConsecutiveHits: warmupStreak } : {},
 		...budgetDropped.length > 0 ? { budgetDropped } : {},
 		...lastRequestImageBytes === void 0 ? {} : { lastRequestImageBytes },
@@ -38259,6 +38619,7 @@ function classifyKimiFailure(status, bodyText) {
 	const detail = summarizeFailureBody$1(bodyText);
 	if (isHttpContextOverflow(status, bodyText)) return {
 		code: CONTEXT_OVERFLOW_CODE,
+		poolAction: "none",
 		retryable: false,
 		accountScoped: false,
 		upstreamOverload: false,
@@ -38266,6 +38627,7 @@ function classifyKimiFailure(status, bodyText) {
 	};
 	if (status === 400 && isKimiBodySizeRefusal(detail)) return {
 		code: CONTEXT_OVERFLOW_CODE,
+		poolAction: "none",
 		retryable: false,
 		accountScoped: false,
 		upstreamOverload: false,
@@ -38273,6 +38635,7 @@ function classifyKimiFailure(status, bodyText) {
 	};
 	if (status === 402) return {
 		code: "SERVER",
+		poolAction: "retry-same-account",
 		retryable: true,
 		accountScoped: false,
 		upstreamOverload: false,
@@ -38281,6 +38644,7 @@ function classifyKimiFailure(status, bodyText) {
 	if (status === 401 || status === 403) {
 		if (matchesAny$1(detail, ENTITLEMENT_PATTERNS)) return {
 			code: "PROVIDER_ERROR",
+			poolAction: "none",
 			retryable: false,
 			accountScoped: false,
 			upstreamOverload: false,
@@ -38290,6 +38654,7 @@ function classifyKimiFailure(status, bodyText) {
 			const limitReached = matchesAny$1(detail, ACCOUNT_LIMIT_PATTERNS);
 			return {
 				code: "PROVIDER_ERROR",
+				poolAction: "cooldown-and-rotate",
 				retryable: false,
 				accountScoped: true,
 				upstreamOverload: false,
@@ -38298,6 +38663,7 @@ function classifyKimiFailure(status, bodyText) {
 		}
 		return {
 			code: "INVALID_CREDENTIAL",
+			poolAction: "mark-auth-failed",
 			retryable: false,
 			accountScoped: false,
 			upstreamOverload: false,
@@ -38307,6 +38673,7 @@ function classifyKimiFailure(status, bodyText) {
 	if (status === 429) {
 		if (matchesAny$1(detail, QUOTA_EXHAUSTED_PATTERNS$1)) return {
 			code: "PROVIDER_ERROR",
+			poolAction: "cooldown-and-rotate",
 			retryable: false,
 			accountScoped: true,
 			upstreamOverload: false,
@@ -38314,6 +38681,7 @@ function classifyKimiFailure(status, bodyText) {
 		};
 		if (matchesAny$1(detail, UPSTREAM_OVERLOAD_PATTERNS)) return {
 			code: "SERVER",
+			poolAction: "retry-same-account",
 			retryable: true,
 			accountScoped: false,
 			upstreamOverload: true,
@@ -38321,6 +38689,7 @@ function classifyKimiFailure(status, bodyText) {
 		};
 		return {
 			code: "RATE_LIMIT",
+			poolAction: "retry-same-account",
 			retryable: true,
 			accountScoped: false,
 			upstreamOverload: false,
@@ -38329,6 +38698,7 @@ function classifyKimiFailure(status, bodyText) {
 	}
 	if (status >= 500) return {
 		code: "SERVER",
+		poolAction: "retry-same-account",
 		retryable: true,
 		accountScoped: false,
 		upstreamOverload: matchesAny$1(detail, UPSTREAM_OVERLOAD_PATTERNS),
@@ -38336,6 +38706,7 @@ function classifyKimiFailure(status, bodyText) {
 	};
 	if (status === 400) return {
 		code: "PROVIDER_ERROR",
+		poolAction: "none",
 		retryable: false,
 		accountScoped: false,
 		upstreamOverload: false,
@@ -38343,6 +38714,7 @@ function classifyKimiFailure(status, bodyText) {
 	};
 	return {
 		code: "PROVIDER_ERROR",
+		poolAction: "none",
 		retryable: false,
 		accountScoped: false,
 		upstreamOverload: false,
@@ -38378,18 +38750,33 @@ function accountLimitCooldownMs(bodyText, now = Date.now()) {
 		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,
+		minimumMs: MIN_ACCOUNT_LIMIT_COOLDOWN_MS,
 		maximumMs: MAX_ACCOUNT_LIMIT_COOLDOWN_MS
 	});
 }
+/** Minimum cooldown for a proven account-owned usage limit near its reset. */
+const MIN_ACCOUNT_LIMIT_COOLDOWN_MS = 15 * 6e4;
 /**
-* Cooldown one rate-limited account takes when the provider states no delay.
+* The account a shared/transient failure should use on the harness's next attempt.
 *
-* 只在**真正的**限流上使用：上游引擎过载走 SERVER，既不落这条窗口，也不消耗轮转
-* 名额。它是 15 分钟而不是服务自己说的时长，正是因为拿不准——拿不准的那次猜测
-* 落在一个只有单个账号的号池上，代价是整个号池在 15 分钟里彻底不可用。
+* DSH rebuilds GenerateOptions for a retry, but keeps the step's original abort
+* signal. Keying by that signal avoids a process-global session latch: concurrent
+* sessions cannot affect one another, and an abandoned retry chain disappears by GC.
 */
-const POOL_COOLDOWN_MS$5 = 15 * 6e4;
+const sameAccountRetryBySignal = /* @__PURE__ */ new WeakMap();
+function rememberSameAccountRetry(signal, accountId) {
+	if (signal !== void 0 && accountId !== void 0 && !signal.aborted) sameAccountRetryBySignal.set(signal, accountId);
+}
+function clearSameAccountRetry(signal) {
+	if (signal !== void 0) sameAccountRetryBySignal.delete(signal);
+}
+/** Only these non-secret, bounded upstream identities may enter durable diagnostics. */
+function upstreamRequestId(headers) {
+	for (const name of ["x-request-id", "msh-request-id"]) {
+		const value = headers.get(name)?.trim() ?? "";
+		if (/^[A-Za-z0-9._:-]{1,128}$/.test(value)) return value;
+	}
+}
 /**
 * Fallback cooldown for an account whose usage window is spent.
 *
@@ -38525,30 +38912,39 @@ function reportDriftCause(sessionId, cause) {
 	if (driftWarned.size > 256) driftWarned.clear();
 	console.warn("[kimi-code] prompt cache rebuilt for session " + sessionId + ": the " + cause + " changed mid-conversation, so the whole prompt was reprocessed once.");
 }
-function noteSessionCacheState(sessionId, state) {
+function noteSessionCacheState(sessionId, state, successful, requestDiagnostic) {
 	if (sessionId === void 0 || sessionId.trim() === "") return;
 	const stats = getCacheStats(sessionId);
-	if (stats.requests === 0) return;
+	if (stats.requests === 0 && requestDiagnostic === void 0) return;
 	const lastDrift = getLastDriftCause(sessionId);
 	reportDriftCause(sessionId, lastDrift);
 	const now = Date.now();
 	const imageBudget = getSessionImageBudgetStats(sessionId);
 	const priorRecord = sessionCacheState.read(sessionId);
+	const previousRequestPromptTokens = priorRecord?.totalTokens;
 	const budgetDropped = imageBudget !== void 0 && imageBudget.droppedKeys.length > 0 ? imageBudget.droppedKeys : priorRecord?.budgetDropped;
 	const liveBytes = imageBudget?.requestImageBytes ?? 0;
 	const liveDropped = imageBudget?.droppedImages ?? 0;
 	const lastRequestImageBytes = liveBytes > 0 ? liveBytes : priorRecord?.lastRequestImageBytes;
 	const lastDroppedImages = liveDropped > 0 ? liveDropped : priorRecord?.lastDroppedImages;
+	const recentRequests = requestDiagnostic === void 0 ? priorRecord?.recentRequests : [...priorRecord?.recentRequests ?? [], requestDiagnostic].slice(-20);
+	const accountObservation = successful && state.accountId !== void 0 ? {
+		lastAccountId: state.accountId,
+		lastAccountAt: now
+	} : {
+		...priorRecord?.lastAccountId === void 0 ? {} : { lastAccountId: priorRecord.lastAccountId },
+		...priorRecord?.lastAccountAt === void 0 ? {} : { lastAccountAt: priorRecord.lastAccountAt }
+	};
 	const firstRequest = stats.requests === 1 ? {
 		firstRequestCachedTokens: state.cacheReadTokens,
-		firstRequestPromptTokens: state.cacheReadTokens + state.inputTokens
+		firstRequestPromptTokens: state.cacheReadTokens + state.inputTokens + state.cacheWriteTokens
 	} : {
 		...priorRecord?.firstRequestCachedTokens === void 0 ? {} : { firstRequestCachedTokens: priorRecord.firstRequestCachedTokens },
 		...priorRecord?.firstRequestPromptTokens === void 0 ? {} : { firstRequestPromptTokens: priorRecord.firstRequestPromptTokens }
 	};
 	sessionCacheState.write(sessionId, {
 		lastActiveAt: now,
-		totalTokens: state.cacheReadTokens + state.inputTokens,
+		totalTokens: state.cacheReadTokens + state.inputTokens + state.cacheWriteTokens,
 		requests: stats.requests,
 		cachedTokens: stats.cachedTokens,
 		freshTokens: stats.freshTokens,
@@ -38556,6 +38952,11 @@ function noteSessionCacheState(sessionId, state) {
 		cacheWriteTokens: stats.cacheWriteTokens,
 		cacheWrite5mTokens: stats.cacheWrite5mTokens,
 		cacheWrite1hTokens: stats.cacheWrite1hTokens,
+		lastRequestCachedTokens: state.cacheReadTokens,
+		lastRequestFreshTokens: state.inputTokens,
+		...accountObservation,
+		...recentRequests === void 0 ? {} : { recentRequests },
+		...previousRequestPromptTokens === void 0 ? {} : { previousRequestPromptTokens },
 		...lastDrift === void 0 ? {} : { lastDrift },
 		...budgetDropped === void 0 ? {} : { budgetDropped },
 		...lastRequestImageBytes === void 0 ? {} : { lastRequestImageBytes },
@@ -38569,7 +38970,7 @@ function noteSessionCacheState(sessionId, state) {
 		inputTokens: state.inputTokens,
 		cacheReadTokens: state.cacheReadTokens,
 		outputTokens: state.outputTokens,
-		totalTokens: state.cacheReadTokens + state.inputTokens,
+		totalTokens: state.cacheReadTokens + state.inputTokens + state.cacheWriteTokens,
 		at: now,
 		...lastDrift === void 0 ? {} : { driftCause: lastDrift }
 	});
@@ -38737,13 +39138,18 @@ var KimiCodeAdapter = class extends LlmAdapter {
 				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({
+		const warmAccountIds = orderAffinityCandidates({
 			candidates,
 			remainingShareById
 		});
+		const sessionId = typeof options.sessionId === "string" ? options.sessionId.trim() : "";
+		const durable = sessionId === "" ? void 0 : sessionCacheState.read(sessionId);
+		const previousAccountId = durable?.lastAccountId;
+		const previousAge = durable?.lastAccountAt === void 0 ? Infinity : Date.now() - durable.lastAccountAt;
+		const previousIsWarm = previousAccountId !== void 0 && scopes.some((scope) => scope.id === previousAccountId) && previousAge >= 0 && previousAge < 15e5;
+		const preferAccountIds = [.../* @__PURE__ */ new Set([...previousIsWarm ? [previousAccountId] : [], ...warmAccountIds])];
 		return preferAccountIds.length === 0 ? void 0 : { preferAccountIds };
 	}
 	/**
@@ -38852,10 +39258,27 @@ var KimiCodeAdapter = class extends LlmAdapter {
 		const normalized = normalizeGenerateOptions(options);
 		const sideChannel = isAuxiliaryRequest(normalized) || isSideChannelTitleRequest(normalized);
 		const sessionKey = typeof options.sessionId === "string" ? options.sessionId.trim() : "";
+		const retrySignal = options.signal;
+		const sameAccountRetryId = sideChannel || retrySignal === void 0 ? void 0 : sameAccountRetryBySignal.get(retrySignal);
+		const previousAccountId = sideChannel || sessionKey === "" ? void 0 : sessionCacheState.read(sessionKey)?.lastAccountId;
+		const requestDiagnostic = {
+			id: randomUUID(),
+			startedAt: Date.now(),
+			endedAt: Date.now(),
+			model: options.model,
+			wire: "openai",
+			...previousAccountId === void 0 ? {} : { previousAccountId },
+			accountChanged: false,
+			outcome: "error",
+			attempts: []
+		};
+		let diagnosticState;
+		let requestError;
 		const mainHeaderRelease = !sideChannel && sessionKey !== "" ? registerInFlightMainSessionHeader(sessionKey) : () => void 0;
 		try {
 			const catalog = await this.catalog().catch(() => []);
 			const wire = wireForCatalogEntry(options.model, catalog);
+			requestDiagnostic.wire = wire;
 			const entry = catalog.find((model) => model.id === options.model);
 			const images = await resolveRequestImages$2(normalized, this.options.attachments, signal, KIMI_CODE_REQUEST_IMAGE_MAX_EDGE);
 			const priorDroppedKeys = getSessionDroppedImageKeys(options.sessionId, options.sessionId === void 0 ? void 0 : sessionCacheState.read(options.sessionId)?.budgetDropped);
@@ -38967,7 +39390,8 @@ var KimiCodeAdapter = class extends LlmAdapter {
 			let sentBuild;
 			const pool = this.accountPool;
 			const tried = /* @__PURE__ */ new Set();
-			const affinity = sideChannel ? void 0 : await this.cacheAffinityHint(pool, requestOptions);
+			const cacheAffinity = sideChannel ? void 0 : await this.cacheAffinityHint(pool, requestOptions);
+			const affinity = sameAccountRetryId === void 0 ? cacheAffinity : { preferAccountIds: [.../* @__PURE__ */ new Set([sameAccountRetryId, ...cacheAffinity?.preferAccountIds ?? []])] };
 			let servedByAccountId;
 			let response;
 			while (true) {
@@ -38994,6 +39418,11 @@ var KimiCodeAdapter = class extends LlmAdapter {
 					tried.add(accountId);
 					apiCredentials = effective.credentials;
 				}
+				const attemptDiagnostic = {
+					startedAt: Date.now(),
+					...accountId === void 0 ? {} : { accountId }
+				};
+				if (!sideChannel) requestDiagnostic.attempts.push(attemptDiagnostic);
 				const region = apiCredentials.region ?? await resolveRegion();
 				const base = (apiCredentials.baseUrl ?? codingBaseUrl(region)).replace(/\/+$/, "");
 				const endpoint = wire === "anthropic" ? `${base}/v1/messages?beta=true` : `${base}/v1/chat/completions`;
@@ -39088,21 +39517,30 @@ var KimiCodeAdapter = class extends LlmAdapter {
 						detail = response.ok ? "" : (await response.text().catch(() => "")).slice(0, 2e3);
 					}
 				}
-				if (response.ok) break;
+				attemptDiagnostic.status = response.status;
+				const providerRequestId = upstreamRequestId(response.headers);
+				if (providerRequestId !== void 0) attemptDiagnostic.providerRequestId = providerRequestId;
+				if (response.ok) {
+					attemptDiagnostic.endedAt = Date.now();
+					if (accountId !== void 0) requestDiagnostic.accountId = accountId;
+					clearSameAccountRetry(retrySignal);
+					break;
+				}
 				const failure = classifyKimiFailure(response.status, detail);
+				attemptDiagnostic.code = failure.code;
+				attemptDiagnostic.poolAction = failure.poolAction;
+				attemptDiagnostic.endedAt = Date.now();
 				const after = response.status === 429 ? retryAfterMs$1(response.headers) : void 0;
+				if (after !== void 0) attemptDiagnostic.retryAfterMs = after;
+				if (failure.poolAction === "retry-same-account") rememberSameAccountRetry(retrySignal, accountId);
+				else clearSameAccountRetry(retrySignal);
 				if (pool !== null && accountId !== void 0) {
-					const planScoped = response.status === 429 && matchesAny$1(detail, ENTITLEMENT_PATTERNS);
-					if (response.status === 429 && !planScoped && !failure.upstreamOverload) {
-						const cooldownMs = failure.accountScoped ? accountLimitCooldownMs(detail) : after ?? POOL_COOLDOWN_MS$5;
-						await pool.markCooldown(accountId, cooldownMs, `${PROVIDER_NAME$8} 429`).catch(() => void 0);
+					if (failure.poolAction === "cooldown-and-rotate") {
+						await pool.markCooldown(accountId, accountLimitCooldownMs(detail), `${PROVIDER_NAME$8} ${response.status}`).catch(() => void 0);
 						if (await pool.hasAnotherAvailableAccount(tried)) continue;
-					} else if (failure.code === "INVALID_CREDENTIAL") {
+					} else if (failure.poolAction === "mark-auth-failed") {
 						await pool.markAuthFailed(accountId, failure.message).catch(() => void 0);
 						if (await pool.hasAnotherAvailableAccount(tried)) continue;
-					} else if (failure.accountScoped) {
-						await pool.markCooldown(accountId, accountLimitCooldownMs(detail), `${PROVIDER_NAME$8} 403`).catch(() => void 0);
-						if (await pool.hasAnotherAvailableAccount(tried)) continue;
 					}
 				}
 				throw new LlmError(failure.message, failure.code, {
@@ -39127,6 +39565,8 @@ var KimiCodeAdapter = class extends LlmAdapter {
 			const reader = response.body.getReader();
 			const decoder = new TextDecoder();
 			const state = createStreamState$5(wire, requestOptions.sessionId, sideChannel);
+			state.accountId = servedByAccountId;
+			diagnosticState = state;
 			applyCacheWriteTiers(state, response.headers);
 			if (!sideChannel) markSessionActive(requestOptions.sessionId);
 			let buffer = "";
@@ -39170,14 +39610,42 @@ var KimiCodeAdapter = class extends LlmAdapter {
 					}
 				} catch (error) {
 					const inBand = state.streamError;
-					if (error instanceof LlmError && !outputStarted && error.code === "PROVIDER_ERROR" && inBand !== void 0) throw await this.inBandStreamFailure(error, inBand, servedByAccountId);
+					if (error instanceof LlmError && !outputStarted && error.code === "PROVIDER_ERROR" && inBand !== void 0) throw await this.inBandStreamFailure(error, inBand, servedByAccountId, retrySignal, requestDiagnostic.attempts.at(-1));
 					throw streamFailure(error, signal);
 				}
 			} finally {
-				if (state.sawUsage && !sideChannel) noteSessionCacheState(requestOptions.sessionId ?? state.sessionId, state);
 				reader.cancel().catch(() => void 0);
 			}
+		} catch (error) {
+			if (error instanceof LlmError && isKimiToolArgumentGuardCode(error.code)) rememberSameAccountRetry(retrySignal, diagnosticState?.accountId);
+			requestError = error;
+			throw error;
 		} finally {
+			const endedAt = Date.now();
+			const finalAccountId = diagnosticState?.accountId ?? requestDiagnostic.accountId ?? requestDiagnostic.attempts.at(-1)?.accountId;
+			const errorCode = requestError instanceof LlmError ? requestError.code : requestError === void 0 ? void 0 : "UNKNOWN";
+			const lastAttempt = requestDiagnostic.attempts.at(-1);
+			if (lastAttempt !== void 0 && errorCode !== void 0) {
+				lastAttempt.code = errorCode;
+				lastAttempt.endedAt ??= endedAt;
+			}
+			requestDiagnostic.endedAt = endedAt;
+			if (finalAccountId !== void 0) requestDiagnostic.accountId = finalAccountId;
+			requestDiagnostic.accountChanged = previousAccountId !== void 0 && finalAccountId !== void 0 && previousAccountId !== finalAccountId;
+			requestDiagnostic.outcome = retrySignal?.aborted === true || errorCode === "ABORTED" ? "aborted" : requestError === void 0 && diagnosticState?.finished === true ? "success" : "error";
+			if (errorCode !== void 0) requestDiagnostic.code = errorCode;
+			else if (requestDiagnostic.outcome === "error") requestDiagnostic.code = "INCOMPLETE_STREAM";
+			if (diagnosticState?.finishReason !== null && diagnosticState?.finishReason !== void 0) requestDiagnostic.finishReason = diagnosticState.finishReason;
+			if (diagnosticState?.sawUsage === true) requestDiagnostic.usage = {
+				freshTokens: diagnosticState.inputTokens,
+				cacheReadTokens: diagnosticState.cacheReadReported ? diagnosticState.cacheReadTokens : null,
+				cacheWriteTokens: diagnosticState.cacheWriteReported ? diagnosticState.cacheWriteTokens : null,
+				outputTokens: diagnosticState.outputTokens,
+				reasoningTokens: diagnosticState.reasoningReported ? diagnosticState.reasoningTokens : null
+			};
+			if (diagnosticState?.toolGuard !== void 0) requestDiagnostic.toolGuard = diagnosticState.toolGuard;
+			if (!sideChannel && sessionKey !== "") if (diagnosticState?.sawUsage === true) noteSessionCacheState(diagnosticState.sessionId, diagnosticState, requestError === void 0 && diagnosticState.finished, requestDiagnostic);
+			else sessionCacheState.noteRequest(sessionKey, requestDiagnostic);
 			mainHeaderRelease();
 		}
 	}
@@ -39197,37 +39665,32 @@ var KimiCodeAdapter = class extends LlmAdapter {
 	* re-issued: its body is already open and part-read, and the harness retry
 	* policy owns the repeat.
 	*/
-	async inBandStreamFailure(thrown, inBand, accountId) {
+	async inBandStreamFailure(thrown, inBand, accountId, retrySignal, attemptDiagnostic) {
 		const bodyText = JSON.stringify({ error: inBand.error });
 		const reclassified = inBand.vocabulary === "anthropic" ? reclassifyInBandError(thrown, inBand.error, classifyKimiFailure) : reclassifyInBandResponsesError(thrown, inBand.error, inBand.message, classifyKimiFailure);
 		const pool = this.accountPool;
 		if (pool === null || accountId === void 0) return reclassified;
+		let action = "none";
 		if (inBand.vocabulary === "anthropic") {
 			const status = inBandAnthropicStatus(inBand.error);
 			if (status === null) return reclassified;
-			const failure = classifyKimiFailure(status, bodyText);
-			const planScoped = status === 429 && matchesAny$1(bodyText, ENTITLEMENT_PATTERNS);
-			if (status === 429 && !planScoped && !failure.upstreamOverload) await this.coolInBandRateLimit(pool, accountId, failure.accountScoped, bodyText);
-			else if (failure.code === "INVALID_CREDENTIAL") await pool.markAuthFailed(accountId, failure.message).catch(() => void 0);
-			else if (failure.accountScoped) await pool.markCooldown(accountId, accountLimitCooldownMs(bodyText), `${PROVIDER_NAME$8} in-stream 403`).catch(() => void 0);
+			action = classifyKimiFailure(status, bodyText).poolAction;
+		} else if (reclassified.code === "RATE_LIMIT" || matchesAny$1(bodyText, QUOTA_EXHAUSTED_PATTERNS$1)) action = classifyKimiFailure(429, bodyText).poolAction;
+		else if (reclassified.code === "SERVER") action = "retry-same-account";
+		if (attemptDiagnostic !== void 0) {
+			attemptDiagnostic.code = reclassified.code;
+			attemptDiagnostic.poolAction = action;
+			attemptDiagnostic.endedAt = Date.now();
+		}
+		if (action === "retry-same-account") {
+			rememberSameAccountRetry(retrySignal, accountId);
 			return reclassified;
 		}
-		if (reclassified.code === "RATE_LIMIT" && !matchesAny$1(bodyText, ENTITLEMENT_PATTERNS)) await this.coolInBandRateLimit(pool, accountId, classifyKimiFailure(429, bodyText).accountScoped, bodyText);
+		clearSameAccountRetry(retrySignal);
+		if (action === "cooldown-and-rotate") await pool.markCooldown(accountId, accountLimitCooldownMs(bodyText), `${PROVIDER_NAME$8} in-stream limit`).catch(() => void 0);
+		else if (action === "mark-auth-failed") await pool.markAuthFailed(accountId, reclassified.message).catch(() => void 0);
 		return reclassified;
 	}
-	/**
-	* Cool the account whose own rate limit the stream reported.
-	*
-	* The window this route already defaults to, never a Retry-After: there is no
-	* 429 response to read one from, and a delay stated on the 200 belongs to a
-	* different message. An account-scoped verdict takes its own window instead,
-	* because a spent balance is not back-pressure and 15 minutes would put the
-	* account back into rotation four times inside the window that refused it.
-	*/
-	async coolInBandRateLimit(pool, accountId, accountScoped, bodyText) {
-		const cooldownMs = accountScoped ? accountLimitCooldownMs(bodyText) : POOL_COOLDOWN_MS$5;
-		await pool.markCooldown(accountId, cooldownMs, `${PROVIDER_NAME$8} in-stream 429`).catch(() => void 0);
-	}
 };
 function isRefreshTimeout(error) {
 	const seen = /* @__PURE__ */ new Set();
@@ -39756,17 +40219,45 @@ function openingInheritanceOf(sessionId) {
 		hitRatio: Math.min(1, Math.max(0, cachedTokens / promptTokens))
 	};
 }
+/**
+* Latest request's continuity against the previous prompt's observed 256-token
+* alignment baseline. Raw counts travel with the ratio so an above-baseline
+* reading stays visible instead of being hidden by clamping.
+*/
+function continuationOf(sessionId) {
+	const durable = sessionCacheState.read(sessionId);
+	const cachedTokens = durable?.lastRequestCachedTokens;
+	const freshTokens = durable?.lastRequestFreshTokens;
+	const previousPromptTokens = durable?.previousRequestPromptTokens;
+	if (cachedTokens === void 0 || freshTokens === void 0 || previousPromptTokens === void 0) return void 0;
+	const expectedCachedTokens = Math.floor(previousPromptTokens / 256) * 256;
+	if (expectedCachedTokens <= 0) return void 0;
+	return {
+		cachedTokens,
+		freshTokens,
+		previousPromptTokens,
+		expectedCachedTokens,
+		retentionRatio: Math.min(1, Math.max(0, cachedTokens / expectedCachedTokens))
+	};
+}
+function lastRequestDiagnosticOf(sessionId) {
+	return sessionCacheState.read(sessionId)?.recentRequests?.at(-1);
+}
 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);
+	const continuation = continuationOf(sessionId);
+	const lastRequest = lastRequestDiagnosticOf(sessionId);
 	return {
 		...stats,
 		...lastDrift === void 0 ? {} : { lastDrift },
 		...warmup === void 0 ? {} : { warmup },
 		...opening === void 0 ? {} : { opening },
+		...continuation === void 0 ? {} : { continuation },
+		...lastRequest === void 0 ? {} : { lastRequest },
 		imageBudget: imageBudgetOf(sessionId),
 		cacheWriteTokens: stats.cacheWriteTokens,
 		cacheWrite5mTokens: stats.cacheWrite5mTokens,
@@ -39807,10 +40298,12 @@ function cacheHintOrNull(sessionId) {
 */
 function durableCacheStats(sessionId) {
 	const durable = sessionCacheState.read(sessionId);
-	if (durable === void 0 || durable.requests === 0) return null;
+	if (durable === void 0 || durable.requests === 0 && durable.recentRequests === void 0) return null;
 	const lastDrift = driftCauseOf(durable.lastDrift);
 	const warmup = warmupStatusOf(sessionId);
 	const opening = openingInheritanceOf(sessionId);
+	const continuation = continuationOf(sessionId);
+	const lastRequest = lastRequestDiagnosticOf(sessionId);
 	return {
 		requests: durable.requests,
 		cachedTokens: durable.cachedTokens,
@@ -39823,6 +40316,8 @@ function durableCacheStats(sessionId) {
 		...lastDrift === void 0 ? {} : { lastDrift },
 		...warmup === void 0 ? {} : { warmup },
 		...opening === void 0 ? {} : { opening },
+		...continuation === void 0 ? {} : { continuation },
+		...lastRequest === void 0 ? {} : { lastRequest },
 		imageBudget: imageBudgetOf(sessionId)
 	};
 }
@@ -49761,6 +50256,83 @@ function installKimiRepeatBreaker(ctx, options = {}) {
 	};
 }
 //#endregion
+//#region src/host/kimi-code/tool-argument-recovery.ts
+/**
+* One guided retry after a Kimi tool-argument stream enters a runaway shape.
+*
+* The failed attempt is already settled before `agent/request-error` runs. The
+* recovery notice is appended AFTER it, so the next same-step deriveMessages()
+* sees the instruction without changing the system prompt or tool declarations.
+* `agent.steer()` is deliberately not used: it targets a later step boundary.
+*/
+const MAX_RECOVERY_STEPS = 256;
+function text$4(value) {
+	return typeof value === "string" && value.trim() !== "" ? value.trim() : void 0;
+}
+function number(value) {
+	return typeof value === "number" && Number.isSafeInteger(value) && value >= 0 ? value : void 0;
+}
+/** The model-facing recovery instruction; raw tool arguments never enter it. */
+function createKimiToolArgumentRecoveryMessage(code) {
+	const repeated = code === TOOL_ARGUMENT_REPEATED_KEY_CODE;
+	return createUserMessage$1({
+		content: [{
+			type: "text",
+			text: repeated ? "The previous tool call was stopped because its JSON arguments contained a runaway repeated-key family. Do not resume, reuse, or repair that payload. If you still need a tool, make one fresh call with minimal valid JSON and only the exact property names in its schema. For ask_user_question, use only questions, id, question, header, options, label, description, and multi_select as applicable; never append documentation annotations or underscores to property names. If no tool call is necessary, continue with the evidence already available or explain the blocker in text." : "The previous tool call was stopped because its arguments exceeded the route byte budget. Do not resend the same payload. If the existing evidence is sufficient, finish in text. Otherwise make one fresh call with the smallest sufficient arguments; for a large file or program, use a legal smaller/chunked approach instead of repeating the oversized content."
+		}],
+		source: {
+			kind: PLUGIN_MESSAGE_SOURCE_KIND,
+			form: "notice",
+			summary: boundContextSummary(repeated ? "Runaway tool arguments: retry once with fresh minimal JSON." : "Oversized tool arguments: retry once with a smaller payload.")
+		}
+	});
+}
+/** Install the one-per-step request recovery hook. */
+function installKimiToolArgumentRecovery(ctx) {
+	const attempted = /* @__PURE__ */ new Set();
+	const clearAgent = (raw) => {
+		const agent = raw;
+		const sessionId = text$4(agent?.session?.id) ?? text$4(agent?.id);
+		if (sessionId === void 0) return;
+		const prefix = sessionId + ":";
+		for (const key of [...attempted]) if (key.startsWith(prefix)) attempted.delete(key);
+	};
+	const disposeRequestError = ctx.on?.("agent/request-error", (async (payload, next) => {
+		const code = payload.failure?.code;
+		if (payload.provider !== "kimi-code" || !isKimiToolArgumentGuardCode(code)) return await next();
+		if (payload.signal?.aborted === true) return await next();
+		const agent = payload.agent;
+		const sessionId = text$4(agent?.session?.id) ?? text$4(agent?.id);
+		const turn = number(payload.turn);
+		const step = number(payload.step);
+		const session = agent?.session;
+		const append = session?.append;
+		if (sessionId === void 0 || turn === void 0 || step === void 0 || session === void 0 || append === void 0) return await next();
+		const key = `${sessionId}:${turn}:${step}`;
+		if (attempted.has(key)) return await next();
+		const message = createKimiToolArgumentRecoveryMessage(code);
+		try {
+			append.call(session, "user/message", message, { surfaceOp: "append" });
+		} catch {
+			return await next();
+		}
+		attempted.add(key);
+		while (attempted.size > MAX_RECOVERY_STEPS) {
+			const oldest = attempted.keys().next();
+			if (oldest.done !== true) attempted.delete(oldest.value);
+		}
+		return { kind: "retry" };
+	}));
+	const disposeDisposed = ctx.on?.("agent/disposed", ((payload) => {
+		clearAgent(payload?.agent);
+	}));
+	return () => {
+		disposeRequestError?.();
+		disposeDisposed?.();
+		attempted.clear();
+	};
+}
+//#endregion
 //#region src/host/minimax-code/account-pool.ts
 /**
 * MiniMax Code 号池：the account pool for the MiniMax Code (编程订阅) line.
@@ -64533,6 +65105,7 @@ var GlmCodingPlanAdapter = class extends LlmAdapter {
 	preferences;
 	accountPool;
 	options;
+	catalogRegionSnapshot;
 	constructor(store = new GlmCodingPlanCredentialStore(), modelSettings = new GlmCodingPlanModelSettingsStore(), preferences, options = {}) {
 		super();
 		this.store = store;
@@ -64565,6 +65138,38 @@ var GlmCodingPlanAdapter = class extends LlmAdapter {
 		return cached.length > 0 ? cached : modelsForRegion(region, GLM_MODEL_CATALOG);
 	}
 	/**
+	* 模型目录只需要区域，不需要一把可发请求的凭据。
+	*
+	* 这里刻意不调用 credentialFor：那是真实请求的选号入口，会重复解密号池并回写
+	* lastUsedAt。目录构建会为每个模型并发调用 resolveModel，把它放在这里会让一次
+	* 选择器刷新放大成一串 DPAPI 读写。区域快照按号池 revision 与指定账号失效，
+	* 同一代的并发调用共享同一次只读；读不到或尚未登录时不缓存，登录后即可重试。
+	*/
+	catalogRegion(settings) {
+		const pool = this.accountPool;
+		if (pool === null) return this.readCatalogRegion(settings).catch(() => null);
+		const revision = pool.currentIdentityRevision();
+		const selectedAccountId = settings.selectedAccountId;
+		const cached = this.catalogRegionSnapshot;
+		if (cached !== void 0 && cached.revision === revision && cached.selectedAccountId === selectedAccountId) return cached.region;
+		const snapshot = {
+			revision,
+			selectedAccountId,
+			region: this.readCatalogRegion(settings).catch(() => null)
+		};
+		this.catalogRegionSnapshot = snapshot;
+		snapshot.region.then((region) => {
+			if (region === null && this.catalogRegionSnapshot === snapshot) this.catalogRegionSnapshot = void 0;
+		});
+		return snapshot.region;
+	}
+	async readCatalogRegion(settings) {
+		if (this.accountPool === null) return (await this.store.read())?.codingPlan.region ?? null;
+		const data = await this.accountPool.read();
+		const usable = data.accounts.filter((account) => account.authStatus === void 0 || account.authStatus === "ok");
+		return (usable.find((account) => account.id === settings.selectedAccountId) ?? usable.find((account) => account.id === data.activeAccountId) ?? usable.find((account) => account.isPrimary) ?? usable[0])?.credentials.codingPlan.region ?? null;
+	}
+	/**
 	* 这一行请求用哪份凭据。
 	*
 	* 有池时轮转；没有池时读单凭据存储（无头 host 与测试的形态）。
@@ -64589,10 +65194,9 @@ var GlmCodingPlanAdapter = class extends LlmAdapter {
 	async listModels(provider) {
 		const prov = provider || PROVIDER_ID$1;
 		const settings = await this.settings();
-		if (settings.enabled === false) return [];
-		const pick = await this.credentialFor().catch(() => null);
-		if (pick === null) return [];
-		const region = pick.credentials.codingPlan.region;
+		if (settings.enabled === false || settings.enabledModelIds.length === 0) return [];
+		const region = await this.catalogRegion(settings);
+		if (region === null) return [];
 		const available = this.catalog(region);
 		const enabled = new Set(settings.enabledModelIds);
 		return available.filter((model) => model.wire !== null && enabled.has(model.id)).map((model) => ({
@@ -64605,7 +65209,7 @@ var GlmCodingPlanAdapter = class extends LlmAdapter {
 	async resolveModel(provider, model, signal) {
 		if (signal?.aborted) throw new LlmError(t("glm.modelResolveAborted", { provider: PROVIDER_NAME$1 }), "ABORTED");
 		const settings = await this.settings();
-		const region = (await this.credentialFor().catch(() => null))?.credentials.codingPlan.region ?? "intl";
+		const region = await this.catalogRegion(settings) ?? "intl";
 		const entry = resolveGlmModel(model, this.catalog(region));
 		const override = settings.contextWindowOverrides[model];
 		const contextWindow = typeof override === "number" && Number.isFinite(override) && override > 0 ? override : entry.defaultContextWindow;
@@ -97338,10 +97942,10 @@ function installCacheWarmup(ctx, options) {
 		const blocks = decision.content ?? result.content;
 		if (blocks.length === 0) return decision;
 		const baseNavigation = retrievalTool || ptcRetrieval ? cacheWarmupNavigationIntentOf(exec.agent) : void 0;
-		const isOversizedSearchDump = (exec.name === "grep" || exec.name === "glob") && blocks.reduce((sum, block) => {
+		const isSearchPreviewCandidate = (exec.name === "grep" || exec.name === "glob") && blocks.reduce((sum, block) => {
 			return sum + (typeof block === "object" && block !== null && typeof block.text === "string" ? block.text : "").length;
-		}, 0) >= 2e4;
-		const navigation = baseNavigation === true || isOversizedSearchDump ? true : baseNavigation;
+		}, 0) >= CACHE_WARMUP_DEFAULTS.warmupMaxChars;
+		const navigation = baseNavigation === true || isSearchPreviewCandidate ? true : baseNavigation;
 		if (retrievalTool && navigation !== true) return decision;
 		if (sourceTool && !settings.enabled && navigation !== true) return decision;
 		const route = cacheWarmupRouteOf(exec.agent);
@@ -105967,6 +106571,7 @@ function apply(ctx, pluginConfig = {}) {
 		providers: CONTROLLED_PROVIDERS.filter((provider) => provider !== "minimax-code"),
 		globalEnabled: () => resolveToolRunawayGuardEnabled(pluginConfig)
 	}), "dsh-chatgpt-subscription: shared tool runaway guard");
+	ctx.effect(() => installKimiToolArgumentRecovery(ctx), "dsh-chatgpt-subscription: Kimi tool argument recovery");
 	ctx.effect(() => installDispatchExecutionGuard(ctx, { enabled: () => resolveDispatchExecutionGuardEnabled(pluginConfig) }), "dsh-chatgpt-subscription: dispatch execution guard");
 	ctx.effect(() => installDispatchLedger(ctx, {
 		enabled: () => resolveDispatchExecutionGuardEnabled(pluginConfig),