@eddyskywalker/dsh-chatgpt-subscription 0.45.0 → 0.46.0

package/lib/index.js

@@ -141,7 +141,7 @@ const ROUTE_PREFIX$4 = "/api/dsh-chatgpt-subscription";
 * `test/responses-client.test.ts` asserts the two agree, so a release that
 * forgets this constant fails the suite instead of shipping quietly.
 */
-const PLUGIN_VERSION = "0.45.0";
+const PLUGIN_VERSION = "0.46.0";
 const CODEX_CHATGPT_PROVIDER_ID = "codex-chatgpt";
 const CODEX_API_BASE = "https://chatgpt.com/backend-api/codex";
 const CODEX_RESPONSES_URL = `${CODEX_API_BASE}/responses`;
@@ -24710,6 +24710,19 @@ const CACHE_WARMUP_DEFAULTS = {
 	hitRatioThreshold: .97,
 	/** 退出预热期需要连续命中的报告条数。 */
 	consecutiveHitsRequired: 2,
+	/**
+	* 稳态第二轨（短会话分母效应兜底）：已命中的缓存前缀至少达到该 token 规模，
+	* 且相比上一条报告一字未掉（`cacheReadTokens >= prev.cacheReadTokens`），
+	* 同时本轮新增未命中部分不超过 `retainedPrefixMaxFreshTokens` 时，同样计为一次有效命中。
+	*/
+	retainedPrefixMinCachedTokens: 1024,
+	/** 稳态第二轨允许的单轮最大新增未命中 token 数（与 Kimi 256-token 缓存块对齐）。 */
+	retainedPrefixMaxFreshTokens: 256,
+	/**
+	* 缓存存活窗口（毫秒，25 分钟，与 Kimi 实测全热上限一致）。
+	* 距离最近一次报告超过该时长时，预算计算自动预判缓存已冷、退回预热期保守预算。
+	*/
+	cacheTtlMs: 1500 * 1e3,
 	/** 单结果原文上限（字节）。 */
 	maxResultBytes: 8 * 1024 * 1024,
 	/** 每 session 原文总量上限（字节）。 */
@@ -24773,12 +24786,13 @@ function cellOf(value) {
 		return "";
 	}
 }
-/** JSON Lines：每一行都是一个对象，且至少两行。 */
+/** JSON Lines：每一行都是一个结构一致的表格式对象，且至少 5 行（防止 2~3 个异构 tool 结果对象被当成一张表压掉）。 */
 function parseJsonLines(lines) {
-	if (lines.length < 2) return void 0;
+	if (lines.length < 5) return void 0;
 	const fields = [];
 	const seen = /* @__PURE__ */ new Set();
 	const rows = [];
+	let firstKeySet;
 	for (const line of lines) {
 		let parsed;
 		try {
@@ -24788,6 +24802,10 @@ function parseJsonLines(lines) {
 		}
 		if (parsed === null || typeof parsed !== "object" || Array.isArray(parsed)) return void 0;
 		const record = parsed;
+		const keys = Object.keys(record);
+		if (keys.length === 0) return void 0;
+		if (firstKeySet === void 0) firstKeySet = new Set(keys);
+		else if (!keys.some((k) => firstKeySet.has(k))) return void 0;
 		const row = {};
 		for (const [key, value] of Object.entries(record)) {
 			if (!seen.has(key)) {
@@ -24804,7 +24822,7 @@ function parseJsonLines(lines) {
 		rows
 	};
 }
-/** 分隔符表格：首行表头 + 至少两行同列数的数据。 */
+/** 分隔符表格：首行合法表头（标识符列名，非 grep 路径行号或代码行）+ 至少两行同列数的数据。 */
 function parseDelimitedRows(lines) {
 	if (lines.length < 3) return void 0;
 	const header = lines[0];
@@ -24812,6 +24830,7 @@ function parseDelimitedRows(lines) {
 	if (delimiter === void 0) return void 0;
 	const fields = splitDelimitedLine(header, delimiter).map((cell) => cell.trim());
 	if (fields.length < 2 || fields.some((field) => field === "")) return void 0;
+	if (fields.some((field) => field.length > 64 || /[:;{}()[\]\\/*'"`]/.test(field) || /^\d+$/.test(field))) return;
 	if (new Set(fields).size !== fields.length) return void 0;
 	const rows = [];
 	for (const line of lines.slice(1)) {
@@ -24916,12 +24935,40 @@ function clampMaxBytes(maxBytes) {
 	if (rounded < 1) return 1;
 	return Math.min(rounded, CACHE_WARMUP_DEFAULTS.pageMaxBytesMax);
 }
-/** 一段文本的行数（末尾没有换行也算一行）。 */
-function countLines(text) {
+/**
+* 判断一段文本是否为不含真实换行（或几乎全在一行内）的 JSON 转义文本。
+*
+* 当 `run_code` 返回 `JSON.stringify({ stdout: "a\nb\nc..." })` 时，原文逐字保存在
+* 存储里（不改任何字节），但整份文本里没有真 `\n`，只有字面量两字符序列 `\n`。
+* 若只数真 `\n`，`limit: 64` 行会退化成一次吐完 `maxBytes`。识别出这种形态后，
+* `pageText` 和 `countLines` 把未转义的 `\n`（即前面有奇数个 `\` 的 `n`）也算作逻辑换行，
+* 既 100% 保持「翻页拼起来逐字节等于原文」，又让 `limit` 行上限对单行 JSON 同样生效。
+*/
+function usesEscapedNewlines(text) {
+	if (text.length < 4) return false;
+	let realNewlines = 0;
+	for (let i = 0; i < text.length; i += 1) if (text.charCodeAt(i) === 10) {
+		realNewlines += 1;
+		if (realNewlines > 2) return false;
+	}
+	return text.includes("\\n");
+}
+/** 判断 `text[index]`（已知为 `'n'`）是否构成一个 JSON 转义换行 `\n`（前面紧邻奇数个 `\`）。 */
+function isEscapedNewlineAt(text, index) {
+	let backslashes = 0;
+	for (let cursor = index - 1; cursor >= 0 && text.charCodeAt(cursor) === 92; cursor -= 1) backslashes += 1;
+	return (backslashes & 1) === 1;
+}
+/** 一段文本的行数（末尾没有换行也算一行；支持单行 JSON 内的转义 `\n`）。 */
+function countLines(text, escapedMode = usesEscapedNewlines(text)) {
 	if (text === "") return 0;
 	let lines = 0;
-	for (const char of text) if (char === "\n") lines += 1;
-	return text.endsWith("\n") ? lines : lines + 1;
+	for (let index = 0; index < text.length; index += 1) {
+		const code = text.charCodeAt(index);
+		if (code === 10) lines += 1;
+		else if (escapedMode && code === 110 && isEscapedNewlineAt(text, index)) lines += 1;
+	}
+	return text.endsWith("\n") || escapedMode && text.endsWith("n") && isEscapedNewlineAt(text, text.length - 1) ? lines : lines + 1;
 }
 /**
 * 按 UTF-8 字节安全分页。
@@ -24937,6 +24984,7 @@ function pageText(args) {
 	if (!Number.isInteger(offset) || offset < 0) throw new CacheWarmupStoreError("offset must be a non-negative integer byte offset", "OFFSET_OUT_OF_RANGE");
 	if (offset > totalBytes) throw new CacheWarmupStoreError(`offset ${offset} is past the ${totalBytes}-byte result`, "OFFSET_OUT_OF_RANGE");
 	if (offset < totalBytes && (buffer[offset] & 192) === 128) throw new CacheWarmupStoreError("offset is not on a UTF-8 character boundary", "OFFSET_INVALID");
+	const escapedMode = usesEscapedNewlines(text);
 	const rest = offset === 0 ? text : new TextDecoder("utf-8").decode(buffer.subarray(offset));
 	let content = "";
 	let contentBytes = 0;
@@ -24960,6 +25008,7 @@ function pageText(args) {
 		content += char;
 		contentBytes += charBytes;
 		if (char === "\n") lines += 1;
+		else if (escapedMode && char === "n" && isEscapedNewlineAt(content, content.length - 1)) lines += 1;
 	}
 	const nextOffset = offset + Buffer.byteLength(content, "utf8");
 	const complete = nextOffset >= totalBytes;
@@ -24975,7 +25024,7 @@ function pageText(args) {
 		omitted: {
 			reason: complete ? "none" : stoppedByBytes ? "bytes" : stoppedByLines ? "lines" : "bytes",
 			chars: remaining.length,
-			lines: countLines(remaining),
+			lines: countLines(remaining, escapedMode),
 			rows: 0,
 			bytes: Buffer.byteLength(remaining, "utf8")
 		},
@@ -25309,6 +25358,28 @@ function cacheWarmupHitRatio(usage) {
 */
 /** Kimi 对每个请求都会报告的生命周期状态；只有其余值才表示真实前缀漂移。 */
 const BENIGN_LIFECYCLE_CAUSES = /* @__PURE__ */ new Set(["stable", "first-request"]);
+/**
+* 判断一条报告是否算作「预热命中」（双轨判据）：
+*
+* 1. **第一轨（比例轨）**：`hitRatio >= 0.97`；
+* 2. **第二轨（前缀完整延续轨）**：解决短会话的分母效应（如 `1792 cached + 84 input = 95.52%`）。
+*    当本轮 `cacheReadTokens >= 1024`（已形成实质缓存前缀）、存在紧邻的上一条无真实漂移报告、
+*    本轮命中缓存没有比上一轮缩水（`cacheReadTokens >= previous.cacheReadTokens`），
+*    且本轮新增未命中部分 `inputTokens <= 256`（一个缓存块的正常尾巴）时，同样视为有效命中。
+*/
+function isWarmupHitReport(report, previous) {
+	if (cacheWarmupHitRatio(report) >= CACHE_WARMUP_DEFAULTS.hitRatioThreshold) return true;
+	if (previous === void 0 || previous === null) return false;
+	const prevDrift = previous.driftCause;
+	if (typeof prevDrift === "string" && prevDrift.trim() !== "" && !BENIGN_LIFECYCLE_CAUSES.has(prevDrift)) return false;
+	const cached = finiteCount(report.cacheReadTokens);
+	const prevCached = finiteCount(previous.cacheReadTokens);
+	const fresh = finiteCount(report.inputTokens);
+	if (cached < CACHE_WARMUP_DEFAULTS.retainedPrefixMinCachedTokens) return false;
+	if (prevCached < CACHE_WARMUP_DEFAULTS.retainedPrefixMinCachedTokens) return false;
+	if (cached < prevCached) return false;
+	return fresh <= CACHE_WARMUP_DEFAULTS.retainedPrefixMaxFreshTokens;
+}
 function consecutiveWarmupHits(reports) {
 	try {
 		if (!Array.isArray(reports)) return 0;
@@ -25318,7 +25389,7 @@ function consecutiveWarmupHits(reports) {
 			if (report === void 0 || report === null) break;
 			const drift = report.driftCause;
 			if (typeof drift === "string" && drift.trim() !== "" && !BENIGN_LIFECYCLE_CAUSES.has(drift)) break;
-			if (cacheWarmupHitRatio(report) < CACHE_WARMUP_DEFAULTS.hitRatioThreshold) break;
+			if (!isWarmupHitReport(report, index > 0 ? reports[index - 1] : void 0)) break;
 			hits += 1;
 		}
 		return hits;
@@ -25447,21 +25518,41 @@ const WARMUP_BUDGET = Object.freeze({
 * 当前会话的预览预算。
 *
 * 规则（约束 4）：预热期 `maxChars = 1200`（约 600-750 tokens）；连续
-* `consecutiveHitsRequired`(2) 次命中率 ≥ 0.97 且期间没有 drift，才进入稳态
-* `maxChars = 8000`。任何一次 drift 都清零连续计数，于是下一条请求立刻回落到预热预算。
+* `consecutiveHitsRequired`(2) 次命中（比例 ≥ 0.97 或前缀完整延续且新增 ≤ 256 tokens）
+* 且期间没有 drift，才进入稳态 `maxChars = 8000`。任何一次 drift 都清零连续计数，
+* 于是下一条请求立刻回落到预热预算。
+*
+* 空闲 TTL 预判：若距离最近一条报告已超过 `cacheTtlMs`（25 分钟，Kimi 实测缓存
+* 全热上限），说明上游缓存大概率已冷，无需等下一轮请求失败，当场退回预热期保守预算。
 *
 * **不抛出**：没有会话、没有报告、登记表异常，一律返回预热期的保守预算。
 * @param sessionId - 会话 id；不传则按最保守的预热预算处理。
+* @param now - 当前时间戳（毫秒），默认 `Date.now()`；测试可注入。
+* @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) {
+function cacheWarmupBudget(sessionId, now = Date.now(), durable) {
 	try {
 		if (typeof sessionId !== "string" || sessionId.trim() === "") return { ...WARMUP_BUDGET };
 		const reports = cacheWarmupUsageReports(sessionId);
-		if (reports.length === 0) return { ...WARMUP_BUDGET };
-		const consecutiveHits = consecutiveWarmupHits(reports);
+		if (reports.length === 0) {
+			if (durable !== void 0 && durable !== null && typeof durable === "object") {
+				const lastAt = durable.lastActiveAt;
+				const hits = typeof lastAt === "number" && Number.isFinite(lastAt) && Number.isFinite(now) && now - lastAt <= CACHE_WARMUP_DEFAULTS.cacheTtlMs && typeof durable.consecutiveHits === "number" && Number.isFinite(durable.consecutiveHits) && durable.consecutiveHits > 0 ? Math.floor(durable.consecutiveHits) : 0;
+				const hitRatio = typeof durable.hitRatio === "number" && Number.isFinite(durable.hitRatio) && durable.hitRatio >= 0 ? Math.min(1, durable.hitRatio) : 0;
+				const warmup = hits < CACHE_WARMUP_DEFAULTS.consecutiveHitsRequired;
+				return {
+					maxChars: warmup ? CACHE_WARMUP_DEFAULTS.warmupMaxChars : CACHE_WARMUP_DEFAULTS.steadyStateMaxChars,
+					warmup,
+					consecutiveHits: hits,
+					hitRatio
+				};
+			}
+			return { ...WARMUP_BUDGET };
+		}
 		const last = reports[reports.length - 1];
 		const hitRatio = last === void 0 ? 0 : cacheWarmupHitRatio(last);
+		const consecutiveHits = last !== void 0 && typeof last.at === "number" && Number.isFinite(last.at) && Number.isFinite(now) && now - last.at > CACHE_WARMUP_DEFAULTS.cacheTtlMs ? 0 : consecutiveWarmupHits(reports);
 		const warmup = consecutiveHits < CACHE_WARMUP_DEFAULTS.consecutiveHitsRequired;
 		return {
 			maxChars: warmup ? CACHE_WARMUP_DEFAULTS.warmupMaxChars : CACHE_WARMUP_DEFAULTS.steadyStateMaxChars,
@@ -25476,9 +25567,10 @@ function cacheWarmupBudget(sessionId) {
 /**
 * 一份可读快照。**不抛出**，且完全无副作用（只读登记表）。
 * @param sessionId - 会话 id；不传则给全局汇总。
+* @param now - 当前时间戳（毫秒），默认 `Date.now()`。
 * @returns 快照。
 */
-function cacheWarmupSnapshot(sessionId) {
+function cacheWarmupSnapshot(sessionId, now = Date.now()) {
 	const fallback = {
 		budget: { ...WARMUP_BUDGET },
 		defaults: CACHE_WARMUP_DEFAULTS,
@@ -25496,7 +25588,7 @@ function cacheWarmupSnapshot(sessionId) {
 		sessions: 0
 	};
 	try {
-		const budget = cacheWarmupBudget(sessionId);
+		const budget = cacheWarmupBudget(sessionId, now);
 		const reports = typeof sessionId === "string" && sessionId.trim() !== "" ? cacheWarmupUsageReports(sessionId) : [];
 		const totals = {
 			inputTokens: 0,
@@ -25559,26 +25651,47 @@ function clip(text, max) {
 	return text.length <= max ? text : `${text.slice(0, max)}…`;
 }
 /**
+* 剥离终端 ANSI/CSI 转义序列（如 `\x1b[31mFAIL\x1b[0m`）与回车符 `\r`。
+*
+* JSON 解码后，原先源码/日志里的 `\u001b[31m` 会还原成控制字符 `\x1b`；
+* 如果不剥离，`^FAIL` 锚点匹配会被颜色码挡住，而再次 `JSON.stringify` 时每个
+* `\x1b` 又会膨胀成 6 个字符 `\u001b`。
+*/
+const ANSI_ESCAPE = /\x1b(?:\[[0-9;?]*[ -/]*[@-~]|\][^\x07\x1b]*(?:\x07|\x1b\\)|[@-Z\\-_])/g;
+function stripAnsi(text) {
+	return text.replace(ANSI_ESCAPE, "").replace(/\r/g, "");
+}
+/**
+* 计算一行文本在最终输出里占用的精确字符宽度：
+* - 普通纯文本模式：`line.length + 1`（含换行符 `\n`）；
+* - JSON 字符串内嵌模式：`JSON.stringify(line).length`（去首尾引号后 `+ 2` 个字符的 `\\n` 换行转义，
+*   恰好等于 `JSON.stringify(line).length`，对 `"`、`\`、`\t`、`\u00xx` 的转义膨胀 100% 精确）。
+*/
+function lineCost(line, jsonEscaped) {
+	return jsonEscaped ? JSON.stringify(line).length : line.length + 1;
+}
+/**
 * 按预算装配预览正文。
 *
 * 必需段落要么整段进来，要么整体装不下就返回 undefined（调用方据此保持原文）；
 * 可选段落逐段尝试，装不下就整段跳过。这样任何输出都不会出现「半截失败清单」。
 */
-function fitSections(sections, maxChars) {
+function fitSections(sections, maxChars, jsonEscaped = false) {
 	const lines = [];
 	let used = 0;
+	const trailingNewlineAllowance = jsonEscaped ? 2 : 1;
 	for (const section of sections) {
 		if (section.optional === true) continue;
 		for (const line of section.lines) {
-			used += line.length + 1;
+			used += lineCost(line, jsonEscaped);
 			lines.push(line);
 		}
 	}
-	if (used > maxChars + 1) return void 0;
+	if (used > maxChars + trailingNewlineAllowance) return void 0;
 	for (const section of sections) {
 		if (section.optional !== true) continue;
-		const size = section.lines.reduce((sum, line) => sum + line.length + 1, 0);
-		if (used + size > maxChars + 1) continue;
+		const size = section.lines.reduce((sum, line) => sum + lineCost(line, jsonEscaped), 0);
+		if (used + size > maxChars + trailingNewlineAllowance) continue;
 		for (const line of section.lines) lines.push(line);
 		used += size;
 	}
@@ -25613,51 +25726,140 @@ function buildTestSections(evidence) {
 }
 /** 一行 shell 错误。用于「完整错误集」，所以匹配宁可宽一点。 */
 const SHELL_ERROR_LINE = /(?:error|fatal|panic|traceback|exception|EACCES|EPERM|ENOENT|EEXIST|ELIFECYCLE|ERR!|FAILED|permission denied|cannot |could not |not found)/i;
-/** 一行 glob 结果的判定：无空白、像路径。 */
+/** TypeScript 编译器报错首行：`path(line,col): error TSxxxx:` 或 `path:line:col - error TSxxxx:`。 */
+const TS_DIAGNOSTIC_HEADER = /^(?:.+?\(\d+,\d+\):\s+error\s+TS\d+:|.+?:\d+:\d+\s+-\s+error\s+TS\d+:)/;
+/**
+* 紧跟在 TypeScript 诊断首行后面的续行（缩进的类型比对、代码高亮行、波浪线指示行），
+* 最多连带收集 6 行，使模型无需补读就能看全 `Type A is not assignable to Type B` 的完整原因。
+*/
+function isTsDiagnosticDetailLine(raw) {
+	if (/^\s{2,}\S/.test(raw)) return true;
+	if (/^\d+\s+\|/.test(raw.trimStart())) return true;
+	if (/^~+$/.test(raw.trim())) return true;
+	return false;
+}
+/** 一行 glob 结果的判定：无空白、像文件路径（至少 80% 含路径分隔符或文件扩展名，不含引号/括号/冒号行号）。 */
 function looksLikeGlobResult(text) {
 	const lines = text.split("\n").filter((line) => line.trim() !== "");
 	if (lines.length < 10) return false;
 	let paths = 0;
+	let pathWithSepOrExt = 0;
 	for (const line of lines) {
 		const trimmed = line.trim();
-		if (trimmed.length === 0 || trimmed.length > 400) continue;
+		if (trimmed.length === 0 || trimmed.length > 400) return false;
+		if (GREP_LINE.test(trimmed)) return false;
+		if (/^===.*===$/.test(trimmed)) return false;
+		if (/['"[\]{}()]/.test(trimmed)) return false;
 		if (/\s/.test(trimmed)) continue;
 		if (/^[-*+]$/.test(trimmed)) continue;
 		paths += 1;
+		if (/[\\/]/.test(trimmed) || /\.[A-Za-z0-9]{1,8}$/.test(trimmed)) pathWithSepOrExt += 1;
+	}
+	return paths / lines.length >= .9 && pathWithSepOrExt / lines.length >= .8;
+}
+/** 判断一行是否只是单纯的文件路径（含可选引号/逗号，如 `"test\\stream-error.test.ts",`），而非报错消息。 */
+function isBareFilePathLine(trimmed) {
+	const unquoted = trimmed.replace(/^['"]|['"],?$/g, "").trim();
+	if (unquoted.length === 0 || unquoted.length > 300 || /\s/.test(unquoted)) return false;
+	return /[\\/]/.test(unquoted) && /\.[A-Za-z0-9]{1,8}$/.test(unquoted) && !/:(\d+)/.test(unquoted);
+}
+/**
+* 判断文本是否为混合证据（如包含已投影的 `[cache-warmup preview]`、内嵌大型 JSON 对象行、
+* 多行 pretty-printed JSON 块、`util.inspect` 字符串拼接对象、分节头等）。
+*/
+function isMixedEvidenceOutput(text) {
+	if (text.includes("[cache-warmup preview]")) return true;
+	if (/^(?:===|---)\s+.+?(?:\s+(?:===|---))?$/m.test(text) || /^###\s+\S/m.test(text) || /^--\s+.+\s+--$/m.test(text)) return true;
+	const lines = text.split("\n");
+	let inspectStringConcat = 0;
+	let jsonPropertyLines = 0;
+	for (const raw of lines.slice(0, 200)) {
+		const trimmed = raw.trim();
+		if (trimmed === "") continue;
+		if (trimmed.length > 80 && (trimmed.startsWith("{\"") && (trimmed.endsWith("}") || trimmed.includes("\":")) || trimmed.startsWith("[{\"") && trimmed.includes("\":"))) return true;
+		if (/^"[A-Za-z0-9_$-]+"\s*:/.test(trimmed)) {
+			jsonPropertyLines += 1;
+			if (jsonPropertyLines >= 2) return true;
+		}
+		if (/^['"].*\\n['"]\s*\+$/.test(trimmed)) {
+			inspectStringConcat += 1;
+			if (inspectStringConcat >= 2) return true;
+		}
 	}
-	return paths / lines.length >= .9;
+	return false;
 }
-/** shell 输出判定：有退出码标记，或至少一行错误。 */
+/**
+* shell 输出判定：
+* - 排除混合证据输出（内嵌 JSON 行、`[cache-warmup preview]`、分节头、`util.inspect` 对象）；
+* - 有明确的退出码标记（`[exit code: N]`、`exit code: N`、`EXIT=N`）；
+* - 有 Python `Traceback (most recent call last):` 或 JS 异常头 + 堆栈行（`^\s+at\s+`）；
+* - 若没有任何退出码或堆栈，只有纯文本行时，要求不存在 grep 行，且错误行不是单纯的文件名。
+*/
 function looksLikeShellOutput(text) {
+	if (isMixedEvidenceOutput(text)) return false;
 	if (/(?:exit code|exited with code|process exited with code)[:=\s]+-?\d+/i.test(text)) return true;
 	if (/\[exit code: -?\d+\]/i.test(text)) return true;
+	if (/^EXIT=-?\d+$/m.test(text)) return true;
 	if (/^Traceback \(most recent call last\):/m.test(text)) return true;
+	if (/^(?:[A-Za-z]*Error|Exception): /m.test(text) && /^\s+at\s+\S/m.test(text)) return true;
 	const lines = text.split("\n");
+	let grepLikes = 0;
 	let errors = 0;
 	for (const line of lines) {
-		if (line.trim() === "") continue;
-		if (/^\s*(?:0|no)\s+(?:errors?|failures?|failed)\b/i.test(line.trim())) continue;
+		const trimmed = line.trim();
+		if (trimmed === "") continue;
+		if (/^(?:PASS\b|[✓✔]|ok\s+\d+\b)/.test(trimmed)) continue;
+		if (/^stdout\s+\|\s+/.test(trimmed)) continue;
+		if (/^\(node:\d+\)\s+\[\w+\]\s+Warning:/.test(trimmed)) continue;
+		if (isBareFilePathLine(trimmed)) continue;
+		if (GREP_LINE.test(trimmed) && !TS_DIAGNOSTIC_HEADER.test(trimmed)) {
+			grepLikes += 1;
+			if (grepLikes >= 2) return false;
+		}
+		if (/^\s*(?:0|no)\s+(?:errors?|failures?|failed)\b/i.test(trimmed)) continue;
 		if (SHELL_ERROR_LINE.test(line)) errors += 1;
-		if (errors >= 1) return true;
 	}
-	return false;
+	return errors >= 1;
 }
-/** shell 预览：退出码 + 完整错误集（必需），首尾片段（可选）。 */
+/** shell 预览：退出码 + 完整错误集（必需，含 TS 多行诊断上下文），首尾片段（可选）。 */
 function buildShellSections(text) {
 	const lines = text.split("\n");
 	const errors = [];
-	for (const raw of lines) {
-		const trimmed = raw.trim();
+	let errorCount = 0;
+	for (let index = 0; index < lines.length; index += 1) {
+		const trimmed = lines[index].trim();
 		if (trimmed === "") continue;
 		if (/^\[exit code: -?\d+\]/i.test(trimmed)) continue;
 		if (/^\s*(?:0|no)\s+(?:errors?|failures?|failed)\b/i.test(trimmed)) continue;
-		if (SHELL_ERROR_LINE.test(trimmed)) errors.push(`  ${clip(trimmed, 240)}`);
+		if (/^(?:PASS\b|[✓✔]|ok\s+\d+\b)/.test(trimmed)) continue;
+		if (/^stdout\s+\|\s+/.test(trimmed)) continue;
+		if (/^\(node:\d+\)\s+\[\w+\]\s+Warning:/.test(trimmed)) continue;
+		if (isBareFilePathLine(trimmed)) continue;
+		if (TS_DIAGNOSTIC_HEADER.test(trimmed)) {
+			errorCount += 1;
+			errors.push(`  ${clip(trimmed, 240)}`);
+			let details = 0;
+			while (index + 1 < lines.length && details < 6) {
+				const nextRaw = lines[index + 1];
+				const nextTrimmed = nextRaw.trim();
+				if (nextTrimmed === "" || TS_DIAGNOSTIC_HEADER.test(nextTrimmed) || /^\[exit code: -?\d+\]/i.test(nextTrimmed)) break;
+				if (!isTsDiagnosticDetailLine(nextRaw)) break;
+				errors.push(`    ${clip(nextTrimmed, 240)}`);
+				index += 1;
+				details += 1;
+			}
+			continue;
+		}
+		if (SHELL_ERROR_LINE.test(trimmed)) {
+			errorCount += 1;
+			errors.push(`  ${clip(trimmed, 360)}`);
+		}
 	}
-	if (errors.length > 300) return void 0;
+	if (errorCount > 300 || errors.length > 600) return void 0;
 	const exitMatch = /(?:exit code|exited with code|process exited with code)[:=\s]+(-?\d+)/i.exec(text) ?? /\[exit code: (-?\d+)\]/i.exec(text);
 	const sections = [];
 	if (exitMatch !== null) sections.push({ lines: [`exit code: ${exitMatch[1]}`] });
-	sections.push({ lines: [`errors: ${errors.length} (complete set)`] });
+	sections.push({ lines: [`errors: ${errorCount} (complete set)`] });
 	if (errors.length > 0) sections.push({ lines: errors });
 	const nonEmpty = lines.map((line) => line.trimEnd()).filter((line) => line.trim() !== "");
 	const head = nonEmpty.slice(0, 5).map((line) => `  ${clip(line.trim(), 160)}`);
@@ -25672,8 +25874,8 @@ function buildShellSections(text) {
 	});
 	return sections;
 }
-/** 一行 grep 结果：`<path>:<line>[:<column>]:<text>`。 */
-const GREP_LINE = /^(.+?):(\d+):(?:\d+:)?(.*)$/;
+/** 一行 grep 结果：`<path>:<line>[:<column>]:<text>` 或 `<path>:<line>(?:\t|\||\s{2,})<text>`。 */
+const GREP_LINE = /^(.+?):(\d+)(?::(?:\d+:)?|\t|\|\s?|\s{2,})(.*)$/;
 /** 抽取 grep 证据；命中线占比不到 80% 就当它不是 grep 输出。 */
 function extractGrepEvidence(text) {
 	const lines = text.split("\n").filter((line) => line.trim() !== "");
@@ -25742,36 +25944,58 @@ function buildStructuredSections(text) {
 		optional: true
 	}];
 }
+/** 剥离 `read` 工具或 `grep -C` 常见的行号前缀（如 `1: `、`1-  `、`1\t`、`1| `、`  12→`）。 */
+const LINE_NUMBER_PREFIX = /^\s*\d+(?::(?:\s|$)|-\s{2,}|\t|\|(?:\s|$)|→\s?)/;
+function stripLineNumberPrefix(line) {
+	return line.replace(LINE_NUMBER_PREFIX, "");
+}
 /** 契约/类型定义：命中两行以上声明就当它是契约，保持原样。 */
 function looksLikeContract(text) {
 	let declarations = 0;
-	for (const line of text.split("\n").slice(0, 800)) {
+	for (const raw of text.split("\n").slice(0, 800)) {
+		const line = stripLineNumberPrefix(raw);
 		if (/^\s*(?:export\s+)?(?:declare\s+)?(?:interface|type)\s+[A-Za-z_$][\w$]*/.test(line)) declarations += 1;
 		if (declarations >= 2) return true;
 	}
 	return false;
 }
-/** 源码判定：代码行占比过半（或含多处 markdown 代码围栏）。 */
+/**
+* 源码或混合证据判定：
+* - 含多处 markdown 代码围栏；
+* - 含 `git diff` 补丁块（`diff --git` / `@@ -x,y +x,y @@`）；
+* - 含至少 5 行带行号的源码切片（`1: ...` 或 `1\t...`），说明是单文件或多文件混合读取输出；
+* - 代码行占比 >= 45%。
+*/
 function looksLikeSourceCode(text) {
+	if (/^diff --git\s+/m.test(text) || /^@@ -\d+(?:,\d+)? \+\d+(?:,\d+)? @@/m.test(text)) return true;
 	let fences = 0;
-	for (const line of text.split("\n").slice(0, 600)) if (line.trim().startsWith("```")) fences += 1;
+	for (const line of text.split("\n").slice(0, 600)) if (stripLineNumberPrefix(line).trim().startsWith("```")) fences += 1;
 	if (fences >= 4) return true;
 	let code = 0;
+	let numbered = 0;
 	let nonEmpty = 0;
 	for (const raw of text.split("\n").slice(0, 600)) {
-		const trimmed = raw.trim();
+		if (raw.trim() === "") continue;
+		if (LINE_NUMBER_PREFIX.test(raw)) numbered += 1;
+		const trimmed = stripLineNumberPrefix(raw).trim();
 		if (trimmed === "") continue;
 		nonEmpty += 1;
-		if (/^(?:import|export|from|const|let|var|function|class|interface|type|enum|async|await|return|if|for|while|switch|case|break|continue|try|catch|finally|throw|new|public|private|protected|static|readonly|def|using|namespace|package|#include|#pragma)\b/.test(trimmed) || /^[})\];]/.test(trimmed) || /^(?:\/\/|\/\*|\*)/.test(trimmed)) code += 1;
+		if (/^(?:import|export|from|const|let|var|function|class|interface|type|enum|async|await|return|if|for|while|switch|case|break|continue|try|catch|finally|throw|new|public|private|protected|static|readonly|def|using|namespace|package|#include|#pragma)\b/.test(trimmed) || /^[})\];]/.test(trimmed) || /^(?:\/\/|\/\*|\*)/.test(trimmed) || /^<\/?[A-Za-z][\w.-]*/.test(trimmed)) code += 1;
 	}
+	if (numbered >= 5) return true;
 	return nonEmpty >= 10 && code / nonEmpty >= .45;
 }
-/** 完整 JSON 文档或被截断的 JSON 对象：可能是批量源码/混合证据，不按裸日志猜测。 */
+/** 完整 JSON 文档（对象、数组、JSON 字符串）或被截断的 JSON 对象：不按裸日志猜测。 */
 function isJsonDocument(text) {
 	const trimmed = text.trim();
 	if (trimmed.length < 2) return false;
 	const first = trimmed[0];
 	const last = trimmed[trimmed.length - 1];
+	if (first === "\"" && last === "\"") try {
+		return typeof JSON.parse(trimmed) === "string";
+	} catch {
+		return false;
+	}
 	if (first !== "{" && first !== "[") return false;
 	if (first === "{" && last === "}" || first === "[" && last === "]") try {
 		const value = JSON.parse(trimmed);
@@ -25784,36 +26008,117 @@ function isJsonDocument(text) {
 	}
 	return false;
 }
+/** PTC `run_code` JSON envelope 中允许独立投影日志的字段白名单。 */
+const JSON_ENVELOPE_LOG_KEYS = [
+	"stdout",
+	"stderr",
+	"output",
+	"logs",
+	"log"
+];
+/** 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;
 /**
-* 决定这一段文本该怎么处理。
+* 解开一层或两层 JSON 包装（支持顶层 JSON 字符串 `"..."` 或双重序列化的 `"{...}"`）。
+*/
+function parseJsonCandidate(text) {
+	const trimmed = text.trim();
+	if (trimmed.length < 2) return void 0;
+	const first = trimmed[0];
+	const last = trimmed[trimmed.length - 1];
+	if (first === "\"" && last === "\"") try {
+		const inner = JSON.parse(trimmed);
+		if (typeof inner !== "string") return void 0;
+		const innerTrimmed = inner.trim();
+		if (innerTrimmed.startsWith("{") && innerTrimmed.endsWith("}")) try {
+			const innerObj = JSON.parse(innerTrimmed);
+			if (innerObj !== null && typeof innerObj === "object" && !Array.isArray(innerObj)) return {
+				kind: "object",
+				record: innerObj,
+				wrappedInString: true
+			};
+		} catch {}
+		return {
+			kind: "string",
+			decoded: inner
+		};
+	} catch {
+		return;
+	}
+	if (first === "{" && last === "}") try {
+		const parsed = JSON.parse(trimmed);
+		if (parsed !== null && typeof parsed === "object" && !Array.isArray(parsed)) return {
+			kind: "object",
+			record: parsed,
+			wrappedInString: false
+		};
+	} catch {
+		return;
+	}
+}
+/**
+* 尝试从完整 JSON（顶层 JSON 字符串或对象 envelope）中解出可结构化投影的日志正文。
 *
-* 顺序就是「可证明性」的顺序：子代理报告/源码工具/契约先被判成原样，
-* 然后是测试失败集、grep、结构化、glob、shell 这些有明确形状的，
-* 最后落到「无法解析 → 原样」。
+* 规则刻意保守：
+* - 截断 JSON、数组一律返回 undefined；
+* - 对象 envelope 除白名单字段外，其余字段序列化后的总长度不超过 `JSON_ENVELOPE_MAX_OTHER_CHARS`；
+* - 解码并剥离 ANSI 后的日志文本自身必须能通过非源码/非契约的结构化日志投影计划；
+* - 返回的 `storeContent` 把解码后的多行日志直接展开（对象 envelope 还会把其余元数据缩进在头部），
+*   彻底避免单行 minified JSON 因没有真实 `\n` 导致 `store.page` 的按行分页失效。
 */
-function planProjection(text, toolName) {
-	if (TOOL_SUBAGENT.test(toolName)) return {
-		kind: "verbatim",
-		strategy: "verbatim-subagent-report"
-	};
-	if (TOOL_SOURCE.test(toolName)) return {
-		kind: "verbatim",
-		strategy: "verbatim-source-code"
+function extractJsonEnvelopeLogField(text, toolName) {
+	const candidate = parseJsonCandidate(text);
+	if (candidate === void 0) return void 0;
+	if (candidate.kind === "string") {
+		const clean = stripAnsi(candidate.decoded);
+		if (isJsonDocument(clean) || isMixedEvidenceOutput(clean)) return void 0;
+		const plan = planPlainTextProjection(clean, toolName);
+		if (plan.kind !== "preview") return void 0;
+		return { plan };
+	}
+	const { record } = candidate;
+	let candidateKey;
+	let candidateText = "";
+	for (const key of JSON_ENVELOPE_LOG_KEYS) {
+		const val = record[key];
+		if (typeof val === "string" && val.length > candidateText.length) {
+			candidateKey = key;
+			candidateText = val;
+		}
+	}
+	if (candidateKey === void 0 || candidateText.length === 0) return void 0;
+	const rest = {};
+	for (const [key, value] of Object.entries(record)) {
+		if (key === candidateKey) continue;
+		if (typeof value === "string" && (value.length > 400 || looksLikeSourceCode(value) || looksLikeContract(value))) return;
+		rest[key] = value;
+	}
+	if (JSON.stringify(rest).length > JSON_ENVELOPE_MAX_OTHER_CHARS) return void 0;
+	const cleanField = stripAnsi(candidateText);
+	if (isJsonDocument(cleanField) || isMixedEvidenceOutput(cleanField)) return void 0;
+	const plan = planPlainTextProjection(cleanField, toolName);
+	if (plan.kind !== "preview") return void 0;
+	return {
+		envelope: {
+			record,
+			key: candidateKey
+		},
+		plan
 	};
+}
+/** 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);
 	if (looksLikeContract(text)) return {
 		kind: "verbatim",
 		strategy: "verbatim-contract"
 	};
-	if (isJsonDocument(text)) return {
-		kind: "verbatim",
-		strategy: "verbatim-unparsed"
-	};
 	if (looksLikeSourceCode(text)) return {
 		kind: "verbatim",
 		strategy: "verbatim-source-code"
 	};
 	const test = extractTestEvidence(text);
-	if (test !== void 0) return {
+	if (test !== void 0 && test.failures.length > 0) return {
 		kind: "preview",
 		strategy: "test-failure",
 		sections: buildTestSections(test)
@@ -25852,6 +26157,42 @@ function planProjection(text, toolName) {
 		strategy: "verbatim-unparsed"
 	};
 }
+/**
+* 决定这一段文本该怎么处理。
+*
+* 顺序就是「可证明性」的顺序：子代理报告/源码工具/契约先被判成原样，
+* 然后是 JSON envelope（只允许纯日志字段投影，混合证据一律原样），
+* 接着是测试失败集、grep、结构化、glob、shell 这些有明确形状的，
+* 最后落到「无法解析 → 原样」。
+*/
+function planProjection(text, toolName) {
+	if (TOOL_SUBAGENT.test(toolName)) return {
+		kind: "verbatim",
+		strategy: "verbatim-subagent-report"
+	};
+	if (TOOL_SOURCE.test(toolName)) return {
+		kind: "verbatim",
+		strategy: "verbatim-source-code"
+	};
+	if (looksLikeContract(text)) return {
+		kind: "verbatim",
+		strategy: "verbatim-contract"
+	};
+	if (isJsonDocument(text)) {
+		const extracted = extractJsonEnvelopeLogField(text, toolName);
+		if (extracted !== void 0) return {
+			kind: "preview",
+			strategy: extracted.plan.strategy,
+			sections: extracted.plan.sections,
+			...extracted.envelope === void 0 ? {} : { envelope: extracted.envelope }
+		};
+		return {
+			kind: "verbatim",
+			strategy: "verbatim-unparsed"
+		};
+	}
+	return planPlainTextProjection(text, toolName);
+}
 /** 预览尾部：把「怎么读全文」写死在预览里，模型不用猜。 */
 function footerFor(refId) {
 	return `[full result: ${refId} · page with offset/limit(default 64 lines)/maxBytes(default 16384 bytes)]`;
@@ -25870,12 +26211,22 @@ async function decide(args) {
 	if (input.isError === true && plan.strategy !== "test-failure" && plan.strategy !== "shell-output") return { reason: "verbatim-unparsed" };
 	const header = `[cache-warmup preview] strategy=${plan.strategy} input=${text.length}ch`;
 	const footerBudget = footerFor(`ref://${input.sessionId}/${"0".repeat(64)}`).length;
-	const bodyBudget = budget.maxChars - header.length - footerBudget - 2;
+	const jsonEscaped = plan.envelope !== void 0;
+	const envelopeShellLength = plan.envelope === void 0 ? 0 : JSON.stringify({
+		...plan.envelope.record,
+		[plan.envelope.key]: ""
+	}).length;
+	const bodyBudget = budget.maxChars - header.length - footerBudget - envelopeShellLength - (jsonEscaped ? 4 : 2);
 	if (bodyBudget <= 0) return { reason: "verbatim-budget" };
-	const body = fitSections(plan.sections, bodyBudget);
+	const body = fitSections(plan.sections, bodyBudget, jsonEscaped);
 	if (body === void 0) return { reason: "verbatim-budget" };
 	const store = input.store;
 	if (store === void 0) return { reason: "store-unavailable" };
+	const placeholderInner = `${header}\n${body}\n${footerFor(`ref://${input.sessionId}/${"0".repeat(64)}`)}`;
+	if ((plan.envelope === void 0 ? placeholderInner : JSON.stringify({
+		...plan.envelope.record,
+		[plan.envelope.key]: placeholderInner
+	})).length > budget.maxChars) return { reason: "verbatim-budget" };
 	let reference;
 	try {
 		reference = await store.put({
@@ -25887,7 +26238,11 @@ async function decide(args) {
 	} catch {
 		return { reason: "store-unavailable" };
 	}
-	const preview = `${header}\n${body}\n${footerFor(reference.refId)}`;
+	const innerPreview = `${header}\n${body}\n${footerFor(reference.refId)}`;
+	const preview = plan.envelope === void 0 ? innerPreview : JSON.stringify({
+		...plan.envelope.record,
+		[plan.envelope.key]: innerPreview
+	});
 	if (preview.length > budget.maxChars) return { reason: "verbatim-budget" };
 	return {
 		preview,
@@ -30195,7 +30550,7 @@ function messageToolsOf(message) {
 * lives in the request's declaration list, which is why the two are resolved
 * together rather than the block carrying a schema of its own.
 */
-function toolAdditionNames(message) {
+function toolAdditionNames$1(message) {
 	const names = [];
 	for (const block of message.content) {
 		if (block.type !== "tool-addition") continue;
@@ -30212,7 +30567,7 @@ function toolAdditionNames(message) {
 * `addition-only` never sees one, because the harness removes the declaration
 * from the list instead.
 */
-function toolRemovalNames(message) {
+function toolRemovalNames$1(message) {
 	const names = [];
 	for (const block of message.content) {
 		if (block.type !== "tool-removal") continue;
@@ -30229,7 +30584,7 @@ function toolRemovalNames(message) {
 * list (that is what keeps the prefix intact), so the only thing standing
 * between the model and a call that would fail is this sentence.
 */
-function removalNotice(names) {
+function removalNotice$1(names) {
 	const list = names.map((name) => "\"" + name + "\"").join(", ");
 	return names.length === 1 ? "[" + list + " is no longer available. Do not call it again: any call will fail.]" : "[" + list + " are no longer available. Do not call them again: any call will fail.]";
 }
@@ -30373,7 +30728,7 @@ function declarationSlots(options, media) {
 	let beforeIndex = 0;
 	for (const message of options.messages) {
 		if (message.role === "developer") {
-			const additions = declarationsFor(toolAdditionNames(message), options.tools);
+			const additions = declarationsFor(toolAdditionNames$1(message), options.tools);
 			const entries = [];
 			if (additions.length > 0) entries.push(media.messageTools === true ? {
 				role: "system",
@@ -30382,10 +30737,10 @@ function declarationSlots(options, media) {
 				role: "system",
 				content: declarationNotice(options.model, additions.length, "capability")
 			});
-			const removals = toolRemovalNames(message);
+			const removals = toolRemovalNames$1(message);
 			if (removals.length > 0) entries.push({
 				role: "system",
-				content: removalNotice(removals)
+				content: removalNotice$1(removals)
 			});
 			if (entries.length > 0) slots.push({
 				beforeIndex,
@@ -31622,6 +31977,27 @@ var SessionCacheStateStore = class {
 		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) {
+		const key = sessionId.trim();
+		if (key === "") return;
+		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;
+		this.schedule();
+	}
+	/**
 	* Write the document now, synchronously.
 	*
 	* Exposed for tests and for a caller that is about to lose the process; the
@@ -31698,6 +32074,7 @@ function normalizeRecord(value) {
 	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 warmupStreak = count$1(value.warmupConsecutiveHits);
 	return {
 		lastActiveAt,
 		updatedAt,
@@ -31709,7 +32086,8 @@ function normalizeRecord(value) {
 		cacheWriteTokens: count$1(value.cacheWriteTokens),
 		cacheWrite5mTokens: count$1(value.cacheWrite5mTokens),
 		cacheWrite1hTokens: count$1(value.cacheWrite1hTokens),
-		...drift === void 0 ? {} : { lastDrift: drift }
+		...drift === void 0 ? {} : { lastDrift: drift },
+		...warmupStreak > 0 ? { warmupConsecutiveHits: warmupStreak } : {}
 	};
 }
 /**
@@ -32809,7 +33187,7 @@ const KIMI_CODE_REQUEST_IMAGE_MAX_EDGE = 2e3;
 * @param catalog - live catalog, possibly empty when offline.
 * @returns the tool-update mode, or undefined when this route must resend everything.
 */
-function toolUpdateMode(modelId, catalog) {
+function toolUpdateMode$1(modelId, catalog) {
 	if (!dynamicToolsForEntry(modelId, catalog)) return void 0;
 	if (wireForCatalogEntry(modelId, catalog) !== "openai") return void 0;
 	const raw = (process.env.DSH_KIMI_CODE_TOOL_UPDATE ?? "").trim().toLowerCase();
@@ -32860,7 +33238,7 @@ function noteSessionCacheState(sessionId, state) {
 		...lastDrift === void 0 ? {} : { lastDrift },
 		updatedAt: now
 	});
-	recordCacheWarmupUsage({
+	const warmupState = recordCacheWarmupUsage({
 		sessionId,
 		inputTokens: state.inputTokens,
 		cacheReadTokens: state.cacheReadTokens,
@@ -32869,6 +33247,7 @@ function noteSessionCacheState(sessionId, state) {
 		at: now,
 		...lastDrift === void 0 ? {} : { driftCause: lastDrift }
 	});
+	sessionCacheState.noteWarmupStreak(sessionId, warmupState.consecutiveHits);
 }
 var KimiCodeAdapter = class extends LlmAdapter {
 	store;
@@ -33052,7 +33431,7 @@ var KimiCodeAdapter = class extends LlmAdapter {
 			id: modelId,
 			name: entry?.name ?? modelId,
 			inputModalities: inputModalitiesForEntry(modelId, catalog),
-			...toolUpdateMode(modelId, catalog) === void 0 ? {} : { toolUpdate: toolUpdateMode(modelId, catalog) },
+			...toolUpdateMode$1(modelId, catalog) === void 0 ? {} : { toolUpdate: toolUpdateMode$1(modelId, catalog) },
 			context: { contextWindow: this.contextWindowFor(modelId, entry, settings.contextWindowOverrides) },
 			...efforts.length === 0 ? {} : { reasoning: {
 				efforts: efforts.map((effort) => ({
@@ -33849,13 +34228,32 @@ async function getKimiCodeWebStatus(store, modelSettings, preferences, options =
 * answers with the process aggregate, which is the right answer for a settings
 * card that is not showing a specific conversation.
 */
+function warmupStatusOf(sessionId) {
+	const targetSession = sessionId?.trim() || cacheWarmupUsageSessions()[0]?.sessionId;
+	const durable = sessionCacheState.read(targetSession);
+	const hint = durable === void 0 ? void 0 : {
+		lastActiveAt: durable.lastActiveAt,
+		consecutiveHits: durable.warmupConsecutiveHits ?? 0,
+		hitRatio: durable.cachedTokens + durable.freshTokens > 0 ? durable.cachedTokens / (durable.cachedTokens + durable.freshTokens) : 0
+	};
+	if (targetSession === void 0 && hint === void 0) return void 0;
+	const budget = cacheWarmupBudget(targetSession ?? "durable-latest", Date.now(), hint);
+	return {
+		warmup: budget.warmup,
+		maxChars: budget.maxChars,
+		consecutiveHits: budget.consecutiveHits,
+		requiredHits: CACHE_WARMUP_DEFAULTS.consecutiveHitsRequired
+	};
+}
 function cacheStatsOrNull(sessionId) {
 	const stats = getCacheStats(sessionId);
 	if (stats.requests === 0) return durableCacheStats(sessionId);
 	const lastDrift = getLastDriftCause(sessionId);
+	const warmup = warmupStatusOf(sessionId);
 	return {
 		...stats,
 		...lastDrift === void 0 ? {} : { lastDrift },
+		...warmup === void 0 ? {} : { warmup },
 		cacheWriteTokens: stats.cacheWriteTokens,
 		cacheWrite5mTokens: stats.cacheWrite5mTokens,
 		cacheWrite1hTokens: stats.cacheWrite1hTokens
@@ -33897,6 +34295,7 @@ function durableCacheStats(sessionId) {
 	const durable = sessionCacheState.read(sessionId);
 	if (durable === void 0 || durable.requests === 0) return null;
 	const lastDrift = driftCauseOf(durable.lastDrift);
+	const warmup = warmupStatusOf(sessionId);
 	return {
 		requests: durable.requests,
 		cachedTokens: durable.cachedTokens,
@@ -33906,7 +34305,8 @@ function durableCacheStats(sessionId) {
 		cacheWriteTokens: durable.cacheWriteTokens,
 		cacheWrite5mTokens: durable.cacheWrite5mTokens,
 		cacheWrite1hTokens: durable.cacheWrite1hTokens,
-		...lastDrift === void 0 ? {} : { lastDrift }
+		...lastDrift === void 0 ? {} : { lastDrift },
+		...warmup === void 0 ? {} : { warmup }
 	};
 }
 /**
@@ -64949,6 +65349,52 @@ function imagePart$1(block, images) {
 		image_url: { url: `data:${inline.mediaType};base64,${inline.data}` }
 	};
 }
+/**
+* Whether one model id names a DeepSeek model.
+*
+* DeepSeek's own Chat Completions contract carries a disabled reasoning level as
+* `thinking: { type: 'disabled' }` rather than as a `reasoning_effort` value, so
+* the translation below is scoped to this model family. Read from the model id's
+* own segment so a provider-prefixed spelling (`cline-pass/...`, or a canonical
+* `deepseek/...` slug) is recognised either way.
+*/
+function isDeepSeekModel$1(modelId) {
+	const bare = modelId.slice(modelId.lastIndexOf("/") + 1).trim();
+	return /^deepseek/i.test(bare);
+}
+/** The names one developer message declares or revokes, in history order. */
+function toolNamesOf(blocks, type) {
+	const names = [];
+	for (const block of blocks ?? []) {
+		const record = block;
+		if (record.type !== type) continue;
+		const name = record.toolName;
+		if (typeof name !== "string" || name.trim().length === 0) continue;
+		names.push(name.trim());
+	}
+	return names;
+}
+/** One quoted, comma-joined list of tool names. */
+function quotedList(names) {
+	return names.map((name) => "\"" + name + "\"").join(", ");
+}
+/**
+* The model-facing text of a mid-conversation tool change, or '' when there is none.
+*
+* `in-history` means the harness stopped rewriting the declaration list: an
+* added tool still travels in `tools`, and a removed one keeps its declaration
+* so the prefix survives. The history note is therefore the only thing that
+* tells the model a removal happened, which is why the removal sentence is
+* imperative rather than descriptive.
+*/
+function toolUpdateNotice(blocks) {
+	const added = toolNamesOf(blocks, "tool-addition");
+	const removed = toolNamesOf(blocks, "tool-removal");
+	const notices = [];
+	if (added.length > 0) notices.push(added.length === 1 ? "[The tool " + quotedList(added) + " is now available.]" : "[The tools " + quotedList(added) + " are now available.]");
+	if (removed.length > 0) notices.push(removed.length === 1 ? "[" + quotedList(removed) + " is no longer available. Do not call it again: any call will fail.]" : "[" + quotedList(removed) + " are no longer available. Do not call them again: any call will fail.]");
+	return notices.join("\n");
+}
 /** Recursively collect image parts from blocks, including nested tool results. */
 function collectImageParts$1(blocks, images, target = []) {
 	if (!Array.isArray(blocks)) return target;
@@ -65001,8 +65447,17 @@ function serializeMessages(messages, images) {
 		pendingToolImages = [];
 		emit(message);
 	};
-	/** A tool answer: it closes its own id, and the group when it was the last. */
+	/**
+	* A tool answer: it closes its own id, and the group when it was the last.
+	*
+	* An answer whose call id has no pending `tool_call` is refused here, before
+	* anything is sent. Forwarding it produces an OpenAI message sequence the
+	* gateway rejects (a `tool` message with no declaration in front of it), and
+	* the gateway's 400 names neither the call nor the turn, so the failure is
+	* raised locally with the id that has no question.
+	*/
 	const answer = (callId, content) => {
+		if (!awaiting.includes(callId)) throw new LlmError(`tool result for call "${callId}" has no tool_call awaiting it: the conversation cannot be expressed as a valid OpenAI message sequence`, "INVALID_REQUEST");
 		wire.push({
 			role: "tool",
 			tool_call_id: callId,
@@ -65015,9 +65470,11 @@ function serializeMessages(messages, images) {
 	for (const message of messages) {
 		if (message.role === "system" || message.role === "developer") {
 			flushToolImages();
+			const content = [toolUpdateNotice(message.content), flattenText$1(message.content)].filter((part) => part.length > 0).join("\n");
+			if (content.length === 0) continue;
 			emit({
 				role: "system",
-				content: flattenText$1(message.content)
+				content
 			});
 			continue;
 		}
@@ -65037,7 +65494,7 @@ function serializeMessages(messages, images) {
 		else if (block.type === "image") parts.push(imagePart$1(block, images));
 		const hasImage = parts.some((part) => part.type === "image_url");
 		const text = parts.filter((part) => part.type === "text").map((part) => String(part.text)).join("");
-		if (hasImage || text.length > 0 || toolResults.length === 0) {
+		if (hasImage || text.length > 0) {
 			flushToolImages();
 			emit({
 				role: "user",
@@ -65072,6 +65529,7 @@ function buildClinePassBody(options, connection, images) {
 		}
 	}));
 	const maxTokens = options.maxTokens ?? connection.maxTokens;
+	const thinkingOff = options.reasoningEffort !== void 0 && String(options.reasoningEffort) === "none" && isDeepSeekModel$1(options.model);
 	return {
 		model: options.model,
 		messages,
@@ -65081,7 +65539,8 @@ function buildClinePassBody(options, connection, images) {
 		...options.temperature === void 0 ? {} : { temperature: options.temperature },
 		...maxTokens === void 0 || maxTokens <= 0 ? {} : { max_tokens: maxTokens },
 		...options.stop === void 0 ? {} : { stop: options.stop },
-		...options.reasoningEffort === void 0 ? {} : { reasoning_effort: String(options.reasoningEffort) }
+		...options.reasoningEffort === void 0 || thinkingOff ? {} : { reasoning_effort: String(options.reasoningEffort) },
+		...thinkingOff ? { thinking: { type: "disabled" } } : {}
 	};
 }
 /** Map an OpenAI `finish_reason` to the harness finish reason. */
@@ -65242,12 +65701,47 @@ function closeBlock$1(block) {
 	}
 }
 /**
+* Whether one failure is the SSE parser's own "the body ended without [DONE]".
+*
+* Structural rather than `instanceof`: what has to match is the code the parser
+* assigns, and a second copy of this module (or a harness generation that ships
+* its own LlmError class) would fail an identity test on a value that is
+* otherwise exactly the signal this decision is about.
+*/
+function isMissingSentinel$1(error) {
+	return typeof error === "object" && error !== null && error.code === "STREAM_CLOSED";
+}
+/**
+* End the payload iteration instead of throwing when a body stops cleanly.
+*
+* The parser can only report "no [DONE] arrived" by throwing, and that is NOT a
+* failure on its own: the gateway sometimes closes the response right after the
+* terminal frame. Whether the response actually completed is decided by the
+* frames that were seen, which is knowledge this generator does not have — so it
+* only removes the transport signal and leaves the verdict to {@link translate}.
+* Every other failure passes through untouched, including the error frames a 200
+* body can carry.
+*/
+async function* tolerateMissingSentinel$1(payloads) {
+	try {
+		for await (const payload of payloads) yield payload;
+	} catch (error) {
+		if (!isMissingSentinel$1(error)) throw error;
+	}
+}
+/**
 * Translate decoded SSE payloads into harness stream chunks.
 *
 * Deltas stream through as they arrive; block ends, usage and the finish reason
 * are all deferred to the `[DONE]` sentinel, so no chunk ever follows `finish`.
 * A completion with no content at all becomes an `EMPTY_RESPONSE` error finish
 * rather than a silent empty message.
+*
+* A missing `[DONE]` is not a failure by itself. Once a valid `finish_reason`
+* has been seen the response IS complete, and it closes exactly as a `[DONE]`
+* one does — block ends, usage and finish all preserved. A body that carried
+* content but never named a finish reason is a truncation, and still fails with
+* `STREAM_CLOSED`.
 */
 async function* translate(payloads) {
 	let nextIndex = 0;
@@ -65274,38 +65768,48 @@ async function* translate(payloads) {
 		order.push(block);
 		return block;
 	};
-	for await (const payload of payloads) {
+	/**
+	* The chunks that close a completed response: every open block, then usage,
+	* then the single finish. Both terminal paths use this one assembly, so a
+	* response read from a clean EOF and one read from `[DONE]` cannot drift.
+	*/
+	const closing = () => {
+		const chunks = [];
+		for (const block of order) chunks.push({
+			type: "block-end",
+			index: block.index,
+			block: closeBlock$1(block)
+		});
+		if (pendingUsage !== void 0) chunks.push({
+			type: "usage",
+			usage: pendingUsage
+		});
+		const reason = pendingFinish ?? {
+			type: "finish",
+			reason: { kind: "stop" }
+		};
+		const hasVisibleAnswer = toolBlocks.size > 0 || textBlock !== void 0 && textBlock.text.trim() !== "";
+		chunks.push({
+			type: "finish",
+			reason: reason.reason.kind === "stop" && order.length === 0 ? {
+				kind: "error",
+				failure: {
+					message: "model returned a completed response with no content",
+					code: EMPTY_RESPONSE_CODE
+				}
+			} : reason.reason.kind === "stop" && !hasVisibleAnswer ? {
+				kind: "error",
+				failure: {
+					message: "Cline Pass ended the turn without an answer or a tool call — the response stopped before it did anything",
+					code: "PROVIDER_ERROR"
+				}
+			} : reason.reason
+		});
+		return chunks;
+	};
+	for await (const payload of tolerateMissingSentinel$1(payloads)) {
 		if (payload === "[DONE]") {
-			for (const block of order) yield {
-				type: "block-end",
-				index: block.index,
-				block: closeBlock$1(block)
-			};
-			if (pendingUsage !== void 0) yield {
-				type: "usage",
-				usage: pendingUsage
-			};
-			const reason = pendingFinish ?? {
-				type: "finish",
-				reason: { kind: "stop" }
-			};
-			const hasVisibleAnswer = toolBlocks.size > 0 || textBlock !== void 0 && textBlock.text.trim() !== "";
-			yield {
-				type: "finish",
-				reason: reason.reason.kind === "stop" && order.length === 0 ? {
-					kind: "error",
-					failure: {
-						message: "model returned a completed response with no content",
-						code: EMPTY_RESPONSE_CODE
-					}
-				} : reason.reason.kind === "stop" && !hasVisibleAnswer ? {
-					kind: "error",
-					failure: {
-						message: "Cline Pass ended the turn without an answer or a tool call — the response stopped before it did anything",
-						code: "PROVIDER_ERROR"
-					}
-				} : reason.reason
-			};
+			for (const chunk of closing()) yield chunk;
 			return;
 		}
 		const chunk = unwrapEnvelope(payload);
@@ -65379,6 +65883,10 @@ async function* translate(payloads) {
 		}
 		if (isRecord$11(chunk.usage)) pendingUsage = mapUsage$1(chunk.usage);
 	}
+	if (pendingFinish !== void 0) {
+		for (const chunk of closing()) yield chunk;
+		return;
+	}
 	throw new LlmError("stream ended without [DONE]", "STREAM_CLOSED");
 }
 //#endregion
@@ -65515,6 +66023,7 @@ var ClinePassAdapter = class extends LlmAdapter {
 			id: modelId,
 			name: info.displayName,
 			inputModalities: info.input,
+			toolUpdate: "in-history",
 			context: { contextWindow: info.contextWindow },
 			...outputReservation(info.contextWindow, info.maxTokens),
 			...info.reasoning ? { reasoning: { efforts: info.efforts.map((effort) => ({
@@ -65535,11 +66044,12 @@ var ClinePassAdapter = class extends LlmAdapter {
 	}
 	async *requestStream(rawOptions, settings, signal, activity) {
 		if (settings.enabled === false) throw new LlmError(`${PROVIDER_NAME$4} is disabled in Settings > Subscription. Enable the line there, then retry.`, "INVALID_REQUEST");
-		const options = normalizeGenerateOptions(rawOptions);
-		const images = await resolveRequestImages$1(options, this.options.attachments, signal).catch((error) => {
+		const normalized = normalizeGenerateOptions(rawOptions);
+		const images = await resolveRequestImages$1(normalized, this.options.attachments, signal).catch((error) => {
 			if (signal.aborted) throw new LlmError("Cline Pass request aborted by caller", "ABORTED", { cause: error });
 			return NO_RESOLVED_IMAGES$3;
 		});
+		const options = offloadOldestRequestImages$1(normalized, void 0, images);
 		const body = buildClinePassBody(options, {
 			baseURL: settings.baseURL,
 			displayName: PROVIDER_NAME$4,
@@ -76386,6 +76896,52 @@ function collectImageParts(blocks, images, target = []) {
 	}
 	return target;
 }
+/** Names of the tools one developer message adds at its own history position. */
+function toolAdditionNames(message) {
+	const names = [];
+	for (const block of message.content) {
+		if (block.type !== "tool-addition") continue;
+		const name = block.toolName;
+		if (typeof name !== "string" || name.trim() === "") continue;
+		names.push(name.trim());
+	}
+	return names;
+}
+/** Names of the tools one developer message revokes at its own history position. */
+function toolRemovalNames(message) {
+	const names = [];
+	for (const block of message.content) {
+		if (block.type !== "tool-removal") continue;
+		const name = block.toolName;
+		if (typeof name !== "string" || name.trim() === "") continue;
+		names.push(name.trim());
+	}
+	return names;
+}
+/**
+* The model-facing note that announces tools a conversation just gained.
+*
+* This route declares `toolUpdate: 'in-history'`, so the harness states an
+* addition as a block at its own history position instead of rewriting the whole
+* request's tool list. OpenAI has no message-level tool declaration, so the
+* change travels as this sentence: it costs one short system turn and leaves the
+* top-level `tools` array — and therefore the prefix cache — intact.
+*/
+function additionNotice(names) {
+	const list = names.map((name) => "\"" + name + "\"").join(", ");
+	return names.length === 1 ? "[" + list + " is now available. Call it when it helps with the task.]" : "[" + list + " are now available. Call them when they help with the task.]";
+}
+/**
+* The model-facing note that revokes one or more tools.
+*
+* Imperative and unambiguous on purpose: the tool is still in the declaration
+* list, so this sentence is the only thing between the model and a call that
+* would fail.
+*/
+function removalNotice(names) {
+	const list = names.map((name) => "\"" + name + "\"").join(", ");
+	return names.length === 1 ? "[" + list + " is no longer available. Do not call it again: any call will fail.]" : "[" + list + " are no longer available. Do not call them again: any call will fail.]";
+}
 /**
 * Convert the harness conversation to wire messages.
 *
@@ -76397,7 +76953,15 @@ function collectImageParts(blocks, images, target = []) {
 * message keeps a parallel group's answers contiguous.
 *
 * `developer` becomes `system`: the newer OpenAI spelling is not accepted by
-* every OpenAI-compatible gateway, while `system` is.
+* every OpenAI-compatible gateway, while `system` is. A tool addition or
+* removal carried by such a message is replayed as its own system sentence —
+* the `in-history` shape this route declares — and ordinary text is kept
+* beside it. A message that contributes neither is dropped rather than sent as
+* an empty system turn.
+*
+* A user message that carries no text and no image contributes nothing to a
+* chat request, so it is dropped too; a pure tool-result message never took that
+* path, because its answers travel as `tool` messages.
 */
 function serializeMessages$1(messages, images) {
 	const wire = [];
@@ -76430,6 +76994,7 @@ function serializeMessages$1(messages, images) {
 	};
 	/** A tool answer: it closes its own id, and the group when it was the last. */
 	const answer = (callId, content) => {
+		if (!awaiting.includes(callId)) throw new LlmError(`OpenDesign request is invalid: the tool result for call id "${callId}" has no pending tool call to answer, so the message list would carry a tool message that no assistant tool_calls message introduced. Nothing was sent.`, "INVALID_REQUEST");
 		wire.push({
 			role: "tool",
 			tool_call_id: callId,
@@ -76442,9 +77007,20 @@ function serializeMessages$1(messages, images) {
 	for (const message of messages) {
 		if (message.role === "system" || message.role === "developer") {
 			flushToolImages();
-			emit({
+			const text = flattenText(message.content);
+			if (text.length > 0) emit({
+				role: "system",
+				content: text
+			});
+			const additions = toolAdditionNames(message);
+			if (additions.length > 0) emit({
+				role: "system",
+				content: additionNotice(additions)
+			});
+			const removals = toolRemovalNames(message);
+			if (removals.length > 0) emit({
 				role: "system",
-				content: flattenText(message.content)
+				content: removalNotice(removals)
 			});
 			continue;
 		}
@@ -76464,7 +77040,7 @@ function serializeMessages$1(messages, images) {
 		else if (block.type === "image") parts.push(imagePart(block, images));
 		const hasImage = parts.some((part) => part.type === "image_url");
 		const text = parts.filter((part) => part.type === "text").map((part) => String(part.text)).join("");
-		if (hasImage || text.length > 0 || toolResults.length === 0) {
+		if (hasImage || text.length > 0) {
 			flushToolImages();
 			emit({
 				role: "user",
@@ -76478,6 +77054,21 @@ function serializeMessages$1(messages, images) {
 	return wire;
 }
 /**
+* Whether one model id is a DeepSeek model.
+*
+* The vendor prefix is stripped before the compare, so the bare wire id this
+* line ships (`deepseek-v4.1-flash`) and a vendor-qualified spelling
+* (`deepseek/deepseek-v4-pro`, as a sibling registry writes it) both match,
+* while an id that merely mentions the word — a gateway alias such as
+* `my-deepseek-proxy` — does not. Only DeepSeek's own chat endpoint documents
+* the `thinking` field, which is why the translation below is scoped to it
+* rather than applied to every model that offers a `none` level.
+*/
+function isDeepSeekModel(model) {
+	const id = String(model ?? "").trim().toLowerCase();
+	return id.slice(Math.max(id.lastIndexOf("/"), id.lastIndexOf(":")) + 1).startsWith("deepseek");
+}
+/**
 * Build the wire request for one harness call.
 * @param options - the harness request (model, system, history, tools, sampling).
 * @param connection - resolved connection facts (default output budget).
@@ -76499,6 +77090,8 @@ function buildOpenDesignBody(options, connection, images) {
 		}
 	}));
 	const maxTokens = options.maxTokens ?? connection.maxTokens;
+	const reasoningEffort = options.reasoningEffort === void 0 ? void 0 : String(options.reasoningEffort);
+	const thinkingOff = reasoningEffort !== void 0 && reasoningEffort.toLowerCase() === "none" && isDeepSeekModel(options.model);
 	return {
 		model: options.model,
 		messages,
@@ -76508,7 +77101,8 @@ function buildOpenDesignBody(options, connection, images) {
 		...options.temperature === void 0 ? {} : { temperature: options.temperature },
 		...maxTokens === void 0 || maxTokens <= 0 ? {} : { max_tokens: maxTokens },
 		...options.stop === void 0 ? {} : { stop: options.stop },
-		...options.reasoningEffort === void 0 ? {} : { reasoning_effort: String(options.reasoningEffort) }
+		...thinkingOff || reasoningEffort === void 0 ? {} : { reasoning_effort: reasoningEffort },
+		...thinkingOff ? { thinking: { type: "disabled" } } : {}
 	};
 }
 /** Map an OpenAI `finish_reason` to the harness finish reason. */
@@ -76627,8 +77221,16 @@ function streamErrorCode(message, status) {
 }
 /**
 * Parse an SSE body into decoded JSON frames.
+*
+* The parser's contract is the sibling line's and stays unsoftened: a body that
+* ends without `[DONE]` is reported by THROWING `STREAM_CLOSED` after its last
+* frame. An SSE parser cannot tell a gateway that closes politely after the
+* terminal frame from a truncated response, because it never interprets
+* `finish_reason` — so it must not decide that question for its callers either.
+* That verdict belongs to {@link translate}, which has the frames, and absorbs
+* this one code from its input.
 * @param body - a web ReadableStream of UTF-8 bytes.
-* @returns every `data:` payload up to `[DONE]`.
+* @returns every `data:` payload up to `[DONE]`, then throws `STREAM_CLOSED`.
 */
 async function* parseServerSentEvents(body) {
 	const decoder = new TextDecoder();
@@ -76674,6 +77276,36 @@ async function* parseServerSentEvents(body) {
 	}
 	throw new LlmError("stream ended without [DONE]", "STREAM_CLOSED");
 }
+/**
+* Whether one failure is the SSE parser's own "the body ended without [DONE]".
+*
+* Structural rather than `instanceof`: what has to match is the CODE the parser
+* assigns, and a second copy of this module (or a harness generation that ships
+* its own LlmError class) would fail an identity test on a value that is
+* otherwise exactly the signal this decision is about.
+*/
+function isMissingSentinel(error) {
+	return typeof error === "object" && error !== null && error.code === "STREAM_CLOSED";
+}
+/**
+* End the payload iteration instead of throwing when a body stops without the
+* sentinel.
+*
+* The parser can only report "no [DONE] arrived" by throwing, and that is NOT a
+* failure on its own: the gateway sometimes closes the response right after the
+* terminal frame. Whether the response actually completed is decided by the
+* frames that were seen, which is knowledge this generator does not have — so it
+* only removes the transport signal and leaves the verdict to {@link translate}.
+* Every other failure passes through untouched, including the error frames a 200
+* body can carry.
+*/
+async function* tolerateMissingSentinel(payloads) {
+	try {
+		for await (const payload of payloads) yield payload;
+	} catch (error) {
+		if (!isMissingSentinel(error)) throw error;
+	}
+}
 /** Assemble the terminal content block for one open block. */
 function closeBlock(block) {
 	switch (block.kind) {
@@ -76697,9 +77329,21 @@ function closeBlock(block) {
 * Translate decoded SSE payloads into harness stream chunks.
 *
 * Deltas stream through as they arrive; block ends, usage and the finish reason
-* are all deferred to the `[DONE]` sentinel, so no chunk ever follows `finish`.
+* are all deferred to the end of the answer, so no chunk ever follows `finish`.
 * A completion with no content at all becomes an `EMPTY_RESPONSE` error finish
 * rather than a silent empty message.
+*
+* The end of the answer is either the `[DONE]` sentinel or the end of the body
+* AFTER a terminal `finish_reason` arrived. Some gateways close the socket
+* instead of writing the sentinel, and failing such an answer threw away content
+* the caller had already received.
+*
+* The parser still raises `STREAM_CLOSED` for a body with no sentinel — that is
+* its contract and it is not widened here. The signal is absorbed where the
+* frames are available instead (see {@link tolerateMissingSentinel}), and only
+* then does the finish reason decide: a body that stated one closes like any
+* other, and a body that never stated one is truncation and still fails with
+* `STREAM_CLOSED`. No other input failure is absorbed.
 */
 async function* translate$1(payloads) {
 	let nextIndex = 0;
@@ -76726,38 +77370,44 @@ async function* translate$1(payloads) {
 		order.push(block);
 		return block;
 	};
-	for await (const payload of payloads) {
+	/** The terminal chunks of one completed answer, in the order the harness accepts. */
+	const terminal = (reason) => {
+		const chunks = [];
+		for (const block of order) chunks.push({
+			type: "block-end",
+			index: block.index,
+			block: closeBlock(block)
+		});
+		if (pendingUsage !== void 0) chunks.push({
+			type: "usage",
+			usage: pendingUsage
+		});
+		const hasVisibleAnswer = toolBlocks.size > 0 || textBlock !== void 0 && textBlock.text.trim() !== "";
+		chunks.push({
+			type: "finish",
+			reason: reason.reason.kind === "stop" && order.length === 0 ? {
+				kind: "error",
+				failure: {
+					message: "model returned a completed response with no content",
+					code: EMPTY_RESPONSE_CODE
+				}
+			} : reason.reason.kind === "stop" && !hasVisibleAnswer ? {
+				kind: "error",
+				failure: {
+					message: "OpenDesign ended the turn without an answer or a tool call — the response stopped before it did anything",
+					code: "PROVIDER_ERROR"
+				}
+			} : reason.reason
+		});
+		return chunks;
+	};
+	for await (const payload of tolerateMissingSentinel(payloads)) {
 		if (payload === "[DONE]") {
-			for (const block of order) yield {
-				type: "block-end",
-				index: block.index,
-				block: closeBlock(block)
-			};
-			if (pendingUsage !== void 0) yield {
-				type: "usage",
-				usage: pendingUsage
-			};
 			const reason = pendingFinish ?? {
 				type: "finish",
 				reason: { kind: "stop" }
 			};
-			const hasVisibleAnswer = toolBlocks.size > 0 || textBlock !== void 0 && textBlock.text.trim() !== "";
-			yield {
-				type: "finish",
-				reason: reason.reason.kind === "stop" && order.length === 0 ? {
-					kind: "error",
-					failure: {
-						message: "model returned a completed response with no content",
-						code: EMPTY_RESPONSE_CODE
-					}
-				} : reason.reason.kind === "stop" && !hasVisibleAnswer ? {
-					kind: "error",
-					failure: {
-						message: "OpenDesign ended the turn without an answer or a tool call — the response stopped before it did anything",
-						code: "PROVIDER_ERROR"
-					}
-				} : reason.reason
-			};
+			for (const chunk of terminal(reason)) yield chunk;
 			return;
 		}
 		if (isErrorPayload(payload)) throw new LlmError(errorText(payload.error), streamErrorCode(errorText(payload.error)));
@@ -76831,6 +77481,10 @@ async function* translate$1(payloads) {
 		}
 		if (isRecord$2(chunk.usage)) pendingUsage = mapUsage(chunk.usage);
 	}
+	if (pendingFinish !== void 0) {
+		for (const chunk of terminal(pendingFinish)) yield chunk;
+		return;
+	}
 	throw new LlmError("stream ended without [DONE]", "STREAM_CLOSED");
 }
 //#endregion
@@ -76891,6 +77545,23 @@ const TRANSPORT_RETRY_BASE_MS = 400;
 * but the ladder still prefers published data for anything the tables know.
 */
 const REASONING_EFFORT_IDS = DEFAULT_EFFORTS;
+/**
+* How this route carries a mid-conversation tool change, or undefined for none.
+*
+* `in-history`, not `addition-only`: both append an addition, but only
+* `in-history` can also express a REMOVAL. Under `addition-only` a revoked
+* tool is deleted from the declaration list, the top-level `tools` array
+* shrinks, and the gateway's prefix match is thrown away for the whole prompt.
+*
+* The field is spread into the descriptor only when this route wants one, so a
+* harness generation that does not know it simply never reads it — no version
+* check, and no change to the message shape either way: mapper.ts renders each
+* change as a system sentence and never introduces a native dynamic-tool
+* declaration, which an OpenAI-compatible gateway has no field for.
+*/
+function toolUpdateMode() {
+	return "in-history";
+}
 var OpenDesignAdapter = class extends LlmAdapter {
 	modelSettings;
 	options;
@@ -76949,6 +77620,7 @@ var OpenDesignAdapter = class extends LlmAdapter {
 			id: modelId,
 			name: info.displayName,
 			inputModalities: info.input,
+			...toolUpdateMode() === void 0 ? {} : { toolUpdate: toolUpdateMode() },
 			context: { contextWindow: info.contextWindow },
 			...outputReservation(info.contextWindow, info.maxTokens),
 			...info.reasoning && info.efforts.length > 0 ? { reasoning: {
@@ -76976,11 +77648,12 @@ var OpenDesignAdapter = class extends LlmAdapter {
 	}
 	async *requestStream(rawOptions, settings, signal, activity) {
 		if (settings.enabled === false) throw new LlmError(`${PROVIDER_NAME$2} is disabled in Settings > Subscription. Enable the line there, then retry.`, "INVALID_REQUEST");
-		const options = normalizeGenerateOptions(rawOptions);
-		const images = await resolveRequestImages$1(options, this.options.attachments, signal).catch((error) => {
+		const normalized = normalizeGenerateOptions(rawOptions);
+		const images = await resolveRequestImages$1(normalized, this.options.attachments, signal).catch((error) => {
 			if (signal.aborted) throw new LlmError("OpenDesign request aborted by caller", "ABORTED", { cause: error });
 			return NO_RESOLVED_IMAGES$3;
 		});
+		const options = offloadOldestRequestImages$1(normalized, MAX_REQUEST_IMAGE_BYTES$1, images);
 		const body = buildOpenDesignBody(options, { maxTokens: (await this.modelInfo(options.model)).maxTokens }, images);
 		const refused = /* @__PURE__ */ new Set();
 		let lastError = null;
@@ -82700,6 +83373,20 @@ function createCacheWarmupResultTool(deps) {
 * 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。
 */
 const CACHE_WARMUP_PROVIDER = "kimi-code";
+/**
+* 允许进入结果投影的工具集合。
+*
+* 除了 PTC 的 `run_code`，宿主的独立 Shell 与后台任务输出工具（`win_bash`、
+* `bash`、`pwsh`、`job_output`）同样经常吐出整屏测试/构建日志。源码读取/编辑类
+* 工具（`read`、`edit`、`grep`、`glob`、CodeGraph 等）不在此列，保持 100% 原样。
+*/
+const CACHE_WARMUP_SOURCE_TOOLS = /* @__PURE__ */ new Set([
+	"run_code",
+	"win_bash",
+	"bash",
+	"pwsh",
+	"job_output"
+]);
 /** 提示词分节名，全局唯一（重复注册会抛错）。 */
 const CACHE_WARMUP_SECTION_NAME = "plugin:cache-warmup";
 /**
@@ -82943,7 +83630,7 @@ function installCacheWarmup(ctx, options) {
 	const projectDecision = async (exec, result, decision) => {
 		if (decision.kind !== "accept") return decision;
 		if (decision.value !== void 0) return decision;
-		if (exec.name !== "run_code") return decision;
+		if (!CACHE_WARMUP_SOURCE_TOOLS.has(exec.name)) return decision;
 		if (!options.settings().enabled) return decision;
 		const route = cacheWarmupRouteOf(exec.agent);
 		if (route.provider !== "kimi-code") return decision;
@@ -82954,7 +83641,13 @@ function installCacheWarmup(ctx, options) {
 		const blocks = decision.content ?? result.content;
 		if (blocks.length === 0) return decision;
 		const content = blocks;
-		const budget = cacheWarmupBudget(sessionId);
+		let durable;
+		try {
+			durable = options.durableHint?.(sessionId);
+		} catch {
+			durable = void 0;
+		}
+		const budget = durable === void 0 ? cacheWarmupBudget(sessionId) : cacheWarmupBudget(sessionId, Date.now(), durable);
 		const projected = await projectToolResult({
 			toolName: exec.name,
 			provider: route.provider,
@@ -83191,7 +83884,19 @@ function apply(ctx, pluginConfig = {}) {
 			return dispose;
 		}, "dsh-chatgpt-subscription: no-answer resume");
 	}
-	ctx.effect(() => installCacheWarmup(ctx, { settings: () => resolveCacheWarmupSettings(pluginConfig) }), "dsh-chatgpt-subscription: cache warmup");
+	ctx.effect(() => installCacheWarmup(ctx, {
+		settings: () => resolveCacheWarmupSettings(pluginConfig),
+		durableHint: (sessionId) => {
+			const record = sessionCacheState.read(sessionId);
+			if (record === void 0) return void 0;
+			const total = record.cachedTokens + record.freshTokens;
+			return {
+				lastActiveAt: record.lastActiveAt,
+				consecutiveHits: record.warmupConsecutiveHits ?? 0,
+				hitRatio: total > 0 ? record.cachedTokens / total : 0
+			};
+		}
+	}), "dsh-chatgpt-subscription: cache warmup");
 	const probeEnvFile = relayProbeEnvFile();
 	if (relayProbeEnabled(process.env, { envFile: probeEnvFile })) {
 		const probePath = relayProbeLogPath(process.env, { envFile: probeEnvFile });