@eddyskywalker/dsh-chatgpt-subscription 0.62.0 → 0.63.1

package/lib/index.js

@@ -14,7 +14,7 @@ import path, { basename, dirname, extname, isAbsolute, join, relative, resolve,
 import os, { homedir, platform, userInfo } from "node:os";
 import { constants, createCipheriv, createDecipheriv, createHash, createHmac, publicEncrypt, randomBytes, randomUUID } from "node:crypto";
 import { isDeepStrictEqual } from "node:util";
-import { execSync, spawn } from "node:child_process";
+import { execFileSync, spawn } from "node:child_process";
 import fs$1, { appendFileSync, constants as constants$1, copyFileSync, existsSync, mkdirSync, readFileSync, readdirSync, realpathSync, rmSync, statSync, utimesSync, writeFileSync } from "node:fs";
 import http, { createServer } from "node:http";
 import { WebError } from "@deepseek-ai/dsh-web";
@@ -808,7 +808,7 @@ const ROUTE_PREFIX$5 = "/api/dsh-chatgpt-subscription";
 * `test/responses-client.test.ts` asserts the two agree, so a release that
 * forgets this constant fails the suite instead of shipping quietly.
 */
-const PLUGIN_VERSION = "0.62.0";
+const PLUGIN_VERSION = "0.63.1";
 const CODEX_CHATGPT_PROVIDER_ID = "codex-chatgpt";
 const CODEX_API_BASE = "https://chatgpt.com/backend-api/codex";
 const CODEX_RESPONSES_URL = `${CODEX_API_BASE}/responses`;
@@ -11378,20 +11378,36 @@ function shouldBypassProxy(target, noProxyList = "") {
 	return noProxyBypasses(url.hostname, url.port || defaultPortFor(url.protocol), noProxyList);
 }
 /**
+* Run one system probe with no console window and no shell layer.
+*
+* `windowsHide` is not cosmetic here. `execSync` runs through `cmd.exe /d /s /c`,
+* and a host that has no console of its own (a detached `dsh web` host on Windows,
+* or any GUI-launched parent) gets a NEW VISIBLE console allocated for that
+* short-lived shell — the black window that flashed on every re-detection, because
+* the proxy cache TTL is only 5 seconds (#19, #91). `execFileSync` additionally
+* drops the shell hop, so one probe is one process instead of two and its argument
+* needs no quoting. `windowsHide` is a no-op on macOS and Linux, but it states the
+* intent once for every platform.
+*/
+function runSystemProbe(file, args) {
+	return execFileSync(file, args, {
+		timeout: 1500,
+		encoding: "utf-8",
+		windowsHide: true,
+		stdio: [
+			"ignore",
+			"pipe",
+			"ignore"
+		]
+	});
+}
+/**
 * Read one GNOME proxy setting; `null` when `gsettings` is absent or the key is
 * unreadable (headless box, no dconf session).
 */
 function readGsettings(key) {
 	try {
-		return execSync(`gsettings get ${key}`, {
-			timeout: 1500,
-			encoding: "utf-8",
-			stdio: [
-				"ignore",
-				"pipe",
-				"ignore"
-			]
-		}).trim();
+		return runSystemProbe("gsettings", ["get", key]).trim();
 	} catch {
 		return null;
 	}
@@ -11421,26 +11437,10 @@ function parseLinuxSystemProxy(read) {
 function detectSystemProxy(platform = process.platform, env = process.env, readSettings = readGsettings) {
 	try {
 		if (platform === "win32") {
-			const detected = parseWindowsProxyRegistry(execSync("reg query \"HKCU\\Software\\Microsoft\\Windows\\CurrentVersion\\Internet Settings\"", {
-				timeout: 1500,
-				encoding: "utf-8",
-				stdio: [
-					"ignore",
-					"pipe",
-					"ignore"
-				]
-			}));
+			const detected = parseWindowsProxyRegistry(runSystemProbe("reg.exe", ["query", "HKCU\\Software\\Microsoft\\Windows\\CurrentVersion\\Internet Settings"]));
 			if (detected) return detected;
 		} else if (platform === "darwin") {
-			const detected = parseMacOsScutilProxy(execSync("scutil --proxy", {
-				timeout: 1500,
-				encoding: "utf-8",
-				stdio: [
-					"ignore",
-					"pipe",
-					"ignore"
-				]
-			}));
+			const detected = parseMacOsScutilProxy(runSystemProbe("scutil", ["--proxy"]));
 			if (detected) return detected;
 		} else if (platform === "linux") {
 			const detected = parseLinuxSystemProxy(readSettings);
@@ -28188,7 +28188,11 @@ function isWarmupHitReport(report, previous) {
 	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;
+	if (fresh <= CACHE_WARMUP_DEFAULTS.retainedPrefixMaxFreshTokens) return true;
+	const prevFresh = finiteCount(previous.inputTokens);
+	const prevAlignedTotal = Math.floor((prevCached + prevFresh) / 256) * 256;
+	if (prevCached >= 4096 && cached > prevCached && cached >= prevAlignedTotal) return true;
+	return false;
 }
 function consecutiveWarmupHits(reports) {
 	try {
@@ -28666,7 +28670,7 @@ function isBareFilePathLine(trimmed) {
 */
 function isMixedEvidenceOutput(text) {
 	if (text.includes("[cache-warmup preview]")) return true;
-	if (/^(?:===|---)\s+.+?(?:\s+(?:===|---))?$/m.test(text) || /^###\s+\S/m.test(text) || /^--\s+.+\s+--$/m.test(text)) return true;
+	if (/^(?:===\s+.+?(?:\s+===)?|---\s+(?!a\/\S).+?(?:\s+---)?)$/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;
@@ -29034,22 +29038,48 @@ function looksLikeContract(text) {
 	return false;
 }
 /**
+* 判断包含 `diff --git` 的文本是否为构建/测试脚本输出中内嵌的局部 diff，而非纯粹的代码补丁。
+*
+* 真实会话中（如 `npm run checks && npm run verify`），脚本前段可能打印一段快照 `diff --git`，
+* 紧接着运行测试或以 `[exit code: 1]` 退出。若一刀切将所有含 `diff --git` 的输出当作源码，
+* 会导致 25KB+ 的构建/测试失败日志无法投影。
+*/
+function isEmbeddedDiffInBuildOrTestLog(text) {
+	const firstNonEmpty = text.split("\n").map((line) => line.trim()).find((line) => line !== "");
+	if (firstNonEmpty === void 0) return false;
+	if (firstNonEmpty.startsWith("diff --git ") || firstNonEmpty.startsWith("--- a/") || firstNonEmpty.startsWith("+++ b/") || firstNonEmpty.startsWith("@@ ") || /^(?:commit\s+)?[0-9a-f]{7,40}\b/.test(firstNonEmpty)) return false;
+	const hasTestSummary = extractTestEvidence(text) !== void 0;
+	const hasExitCode = /\[exit code: -?\d+\]/i.test(text) || /(?:exit code|exited with code|process exited with code)[:=\s]+-?\d+/i.test(text);
+	const hasNpmScriptHeader = /^>\s+\S+@\S+\s+\S+/m.test(text);
+	return hasTestSummary || hasExitCode && hasNpmScriptHeader;
+}
+/**
 * 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 `git diff` 补丁块（`diff --git` / `@@ -x,y +x,y @@`），除非是带测试汇总/脚本退出码的构建日志；
 * - 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;
+	const hasDiff = /^diff --git\s+/m.test(text) || /^@@ -\d+(?:,\d+)? \+\d+(?:,\d+)? @@/m.test(text);
+	if (hasDiff && !isEmbeddedDiffInBuildOrTestLog(text)) return true;
 	let fences = 0;
 	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;
+	let inEmbeddedDiffHunk = false;
 	for (const raw of text.split("\n").slice(0, 600)) {
 		if (raw.trim() === "") continue;
+		if (hasDiff) {
+			if (raw.startsWith("diff --git ") || raw.startsWith("--- a/") || raw.startsWith("+++ b/") || raw.startsWith("@@ ")) {
+				inEmbeddedDiffHunk = true;
+				continue;
+			}
+			if (inEmbeddedDiffHunk && (raw.startsWith("+") || raw.startsWith("-") || raw.startsWith(" "))) continue;
+			inEmbeddedDiffHunk = false;
+		}
 		if (LINE_NUMBER_PREFIX.test(raw)) numbered += 1;
 		const trimmed = stripLineNumberPrefix(raw).trim();
 		if (trimmed === "") continue;
@@ -33807,7 +33837,159 @@ function leadingSystemText$6(options) {
 	return parts.length === 0 ? void 0 : parts.join("\n\n");
 }
 function nonSystemMessages$6(options) {
-	return options.messages.filter((message) => message.role !== "system");
+	return orderStaticPrefixMessages(options.messages.filter((message) => message.role !== "system"));
+}
+/**
+* Stability tier for an initial bootstrap reminder message (`1` = most static, `4` = session-scoped).
+*
+* The harness appends bootstrap reminders (`agent-instructions`, `runtime-context`,
+* `progress-updates-language`, `skill-catalog`) AFTER the first user prompt (or
+* after the retained tail following a compaction checkpoint). Because Kimi's
+* prefix cache matches strictly from byte 0, leaving the dynamic user prompt at
+* index 0 prevents subagents, forked sessions, and post-compaction turns from
+* sharing the ~8,300-token `AGENTS.md` + `skill-catalog` prefix.
+*
+* Mid-conversation delta reminders (`Updated instructions from:`, `Additional
+* instructions from:`, `Instructions removed:`, `The available skill catalog changed.`)
+* return `null` so they stay anchored at the exact history position where they
+* arrived (append-only in-history).
+*/
+function classifyBootstrapUserMessage(message) {
+	if (message.role !== "user" || isToolResultMessage$7(message)) return null;
+	const source = isRecord$53(message.source) ? message.source : void 0;
+	const kind = source !== void 0 ? asString$21(source.kind) : void 0;
+	const text = firstTextOf(message.content)?.trimStart() ?? "";
+	if (kind === "progress-updates-language" || text.startsWith("Progress update output language override:")) return 1;
+	if (text.startsWith("<system-reminder>\nThe following workspace instructions") || kind === "agent-instructions" && !text.startsWith("<system-reminder>\nUpdated instructions") && !text.startsWith("<system-reminder>\nAdditional instructions") && !text.startsWith("<system-reminder>\nInstructions removed")) return 2;
+	if (text.startsWith("<system-reminder>\nA skill is a reusable set of task-specific instructions.") || kind === "skill-catalog" && !text.startsWith("<system-reminder>\nThe available skill catalog changed.")) return 3;
+	if (kind === "runtime-context" || text.startsWith("Current runtime context. This snapshot supersedes")) return 4;
+	return null;
+}
+function isCompactCheckpointMessage(message) {
+	if (message === void 0 || message.role !== "user" || isToolResultMessage$7(message)) return false;
+	const source = isRecord$53(message.source) ? message.source : void 0;
+	if (source !== void 0 && asString$21(source.kind) === "compact-checkpoint") return true;
+	return (firstTextOf(message.content)?.trimStart() ?? "").startsWith("This is an automatically generated checkpoint condensing");
+}
+/**
+* Normalize `agent-instructions` text so global `~/.dsh/` instructions precede
+* workspace-local `AGENTS.md` instructions when both appear in the same block.
+*/
+function normalizeAgentInstructionsMessage(message) {
+	if (!Array.isArray(message.content) || message.content.length !== 1) return message;
+	const first = message.content[0];
+	if (!isRecord$53(first) || first.type !== "text" || typeof first.text !== "string") return message;
+	const raw = first.text;
+	const firstIdx = raw.indexOf("\n\nInstructions from: ");
+	if (firstIdx < 0) return message;
+	const closeIdx = raw.lastIndexOf("\n</system-reminder>");
+	if (closeIdx <= firstIdx) return message;
+	const header = raw.slice(0, firstIdx);
+	const body = raw.slice(firstIdx + 2, closeIdx);
+	const tail = raw.slice(closeIdx);
+	const sections = body.split("\n\nInstructions from: ");
+	if (sections.length < 2) return message;
+	const tagged = sections.map((sec, idx) => {
+		const firstLineEnd = sec.indexOf("\n");
+		const sourcePath = (firstLineEnd < 0 ? sec : sec.slice(0, firstLineEnd)).trim();
+		return {
+			sec,
+			idx,
+			rank: sourcePath.startsWith("~/") || sourcePath.startsWith("~\\") || sourcePath.includes("/.dsh/") || sourcePath.includes("\\.dsh\\") ? 0 : 1
+		};
+	});
+	if (!tagged.some((item, i) => i > 0 && item.rank < tagged[i - 1].rank)) return message;
+	tagged.sort((a, b) => a.rank - b.rank || a.idx - b.idx);
+	const rebuilt = `${header}\n\nInstructions from: ${tagged.map((t) => t.sec).join("\n\nInstructions from: ")}${tail}`;
+	return {
+		...message,
+		content: [{
+			...first,
+			text: rebuilt
+		}]
+	};
+}
+/**
+* Deterministic entropy-sorted ordering for leading bootstrap reminders.
+*
+* 1. Within the leading pre-assistant `user` message run, orders the first
+*    occurrence of each bootstrap tier (`1: progress-updates-language` ->
+*    `2: agent-instructions` -> `3: skill-catalog` -> `4: runtime-context`)
+*    ahead of the dynamic `user` prompt (`5`).
+* 2. When the surface starts with a `compact-checkpoint` that shadowed the
+*    original bootstrap run and the harness re-injected Tier 1/2/3 reminders
+*    after the retained tail, hoists the first occurrence of Tier 1/2/3 ahead
+*    of `compact-checkpoint` so the ~21k static prefix stays intact across
+*    compaction boundaries and every step after compaction.
+*/
+function orderStaticPrefixMessages(messages) {
+	try {
+		if (messages.length <= 1) return [...messages];
+		let leadEnd = 0;
+		while (leadEnd < messages.length && messages[leadEnd].role === "user" && !isToolResultMessage$7(messages[leadEnd])) leadEnd += 1;
+		if (leadEnd > 0 && isCompactCheckpointMessage(messages[0])) {
+			const seenInLead = /* @__PURE__ */ new Set();
+			for (let i = 0; i < leadEnd; i++) {
+				const tier = classifyBootstrapUserMessage(messages[i]);
+				if (tier !== null) seenInLead.add(tier);
+			}
+			if (!seenInLead.has(1) && !seenInLead.has(2)) {
+				const hoistedByTier = /* @__PURE__ */ new Map();
+				for (let i = leadEnd; i < messages.length; i++) {
+					const msg = messages[i];
+					const tier = classifyBootstrapUserMessage(msg);
+					if ((tier === 1 || tier === 2 || tier === 3) && !hoistedByTier.has(tier)) hoistedByTier.set(tier, {
+						index: i,
+						message: tier === 2 ? normalizeAgentInstructionsMessage(msg) : msg
+					});
+				}
+				if (hoistedByTier.size > 0) {
+					const hoistedIndices = /* @__PURE__ */ new Set();
+					const prefix = [];
+					for (const tier of [
+						1,
+						2,
+						3
+					]) {
+						const entry = hoistedByTier.get(tier);
+						if (entry !== void 0) {
+							hoistedIndices.add(entry.index);
+							prefix.push(entry.message);
+						}
+					}
+					const rest = messages.filter((_, idx) => !hoistedIndices.has(idx));
+					return [...prefix, ...rest];
+				}
+			}
+		}
+		if (leadEnd <= 1) return [...messages];
+		const seenTiers = /* @__PURE__ */ new Set();
+		const leadTagged = [];
+		let hasBootstrap = false;
+		for (let i = 0; i < leadEnd; i++) {
+			const rawMsg = messages[i];
+			const tier = classifyBootstrapUserMessage(rawMsg);
+			if (tier !== null && !seenTiers.has(tier)) {
+				seenTiers.add(tier);
+				hasBootstrap = true;
+				leadTagged.push({
+					message: tier === 2 ? normalizeAgentInstructionsMessage(rawMsg) : rawMsg,
+					rank: tier,
+					index: i
+				});
+			} else leadTagged.push({
+				message: rawMsg,
+				rank: 5,
+				index: i
+			});
+		}
+		if (!hasBootstrap) return [...messages];
+		if (!leadTagged.some((item, i) => i > 0 && item.rank < leadTagged[i - 1].rank || item.message !== messages[i])) return [...messages];
+		leadTagged.sort((a, b) => a.rank - b.rank || a.index - b.index);
+		return [...leadTagged.map((item) => item.message), ...messages.slice(leadEnd)];
+	} catch {
+		return [...messages];
+	}
 }
 const NO_REASONING = {
 	fields: [],
@@ -35374,21 +35556,43 @@ function rememberScoped(store, key, value) {
 */
 const prefixSnapshots = /* @__PURE__ */ new Map();
 /**
+* Extract the concatenated text of leading workspace-static bootstrap reminders
+* (Tiers 1–3: `progress-updates-language`, initial `agent-instructions`, initial
+* `skill-catalog`) after static-prefix ordering.
+*/
+function leadingStaticBootstrapText(options) {
+	const ordered = nonSystemMessages$6(options);
+	const parts = [];
+	for (const message of ordered) {
+		const tier = classifyBootstrapUserMessage(message);
+		if (tier === 1 || tier === 2 || tier === 3) {
+			const text = textOf$7(message.content);
+			if (text !== "") parts.push(text);
+			continue;
+		}
+		break;
+	}
+	return parts.join("\n\n");
+}
+/**
 * Hash the prefix-breaking inputs of one request.
 *
 * @param options - The request about to be built.
-* @returns The hashes of its system prompt and its tool list.
+* @returns The hashes of its system prompt, tool list, and optional static bootstrap prefix.
 */
 function prefixFingerprints(options) {
 	const systemText = leadingSystemText$6(options) ?? "";
 	const tools = options.tools ?? [];
+	const staticBootstrap = leadingStaticBootstrapText(options);
+	const staticPrefixHash = staticBootstrap === "" ? null : fingerprint(staticBootstrap);
 	return {
 		systemPromptHash: systemText === "" ? null : fingerprint(systemText),
 		toolsHash: tools.length === 0 ? null : fingerprint(JSON.stringify(tools.map((tool) => [
 			tool.name,
 			tool.description ?? "",
 			tool.parameters ?? {}
-		])))
+		]))),
+		...staticPrefixHash === null ? {} : { staticPrefixHash }
 	};
 }
 /**
@@ -37122,6 +37326,19 @@ const inFlightColdLeaderStartedAt = /* @__PURE__ */ new Map();
 * from the outside.
 */
 const coldLeaderReleases = /* @__PURE__ */ new Map();
+/**
+* In-flight main-session requests waiting for their HTTP response headers,
+* keyed by sessionId. Used so a concurrent `session-title-first-prompt-llm`
+* side-channel request yields to the main turn's prefill instead of competing
+* on the same account during cold start.
+*/
+const inFlightMainSessionHeaders = /* @__PURE__ */ new Map();
+const mainSessionHeaderReleases = /* @__PURE__ */ new Map();
+/**
+* Per-account/model first-turn prefill pacing chain for concurrent subagent bursts.
+*/
+const inFlightSubagentBurstLeaders = /* @__PURE__ */ new Map();
+const subagentBurstReleases = /* @__PURE__ */ new Map();
 /** The gate of a request that neither waits for nor leads a cold start. */
 const NO_COLD_START_GATE = {
 	isLeader: false,
@@ -37132,9 +37349,9 @@ const NO_COLD_START_GATE = {
 * The group of requests that address the same cached prefix.
 *
 * The service derives its cache from the request content, so two requests with
-* the same model, the same system prompt and the same tool list hit the same
-* entry. The region is part of the key because each region is a separate
-* service with a cache of its own.
+* the same model, the same system prompt, the same tool list and the same
+* leading static bootstrap reminders hit the same entry. The region is part of
+* the key because each region is a separate service with a cache of its own.
 *
 * A request with no system prompt has no prefix worth warming, and gets no
 * group at all.
@@ -37144,18 +37361,18 @@ const NO_COLD_START_GATE = {
 * @returns The group key, or undefined when the request has no prefix.
 */
 function coldStartGroupKey(region, options) {
-	const { systemPromptHash, toolsHash } = prefixFingerprints(options);
+	const { systemPromptHash, toolsHash, staticPrefixHash } = prefixFingerprints(options);
 	if (systemPromptHash === null) return void 0;
-	return `${region}:${options.model}:${systemPromptHash}:${toolsHash ?? "none"}`;
+	return `${region}:${options.model}:${systemPromptHash}:${toolsHash ?? "none"}${staticPrefixHash ? ":" + staticPrefixHash : ""}`;
 }
 /**
 * Remember that a prefix group is warm.
 *
 * @param groupKey - The group that was just warmed.
 */
-function rememberWarmPrefix(groupKey) {
+function rememberWarmPrefix(groupKey, now = Date.now()) {
 	warmPrefixTimestamps.delete(groupKey);
-	warmPrefixTimestamps.set(groupKey, Date.now());
+	warmPrefixTimestamps.set(groupKey, now);
 	while (warmPrefixTimestamps.size > MAX_WARM_PREFIX_GROUPS) {
 		const oldest = warmPrefixTimestamps.keys().next();
 		if (oldest.done === true) break;
@@ -37163,6 +37380,83 @@ function rememberWarmPrefix(groupKey) {
 	}
 }
 /**
+* Register an in-flight main-session request until its HTTP response headers arrive.
+*/
+function registerInFlightMainSessionHeader(sessionId) {
+	const key = sessionId.trim();
+	if (key === "") return () => void 0;
+	mainSessionHeaderReleases.get(key)?.();
+	let resolveHeader;
+	const promise = new Promise((resolve) => {
+		resolveHeader = resolve;
+	});
+	inFlightMainSessionHeaders.set(key, promise);
+	let released = false;
+	const release = () => {
+		if (released) return;
+		released = true;
+		resolveHeader();
+		if (mainSessionHeaderReleases.get(key) === release) {
+			mainSessionHeaderReleases.delete(key);
+			inFlightMainSessionHeaders.delete(key);
+		}
+	};
+	mainSessionHeaderReleases.set(key, release);
+	return release;
+}
+/**
+* Whether a first-turn request carries a subagent mandate or large uncached
+* dynamic payload that should stagger behind an already in-flight sibling burst.
+*/
+function isSubagentBurstCandidate(options) {
+	let dynamicChars = 0;
+	for (const message of options.messages) {
+		if (message.role !== "user") continue;
+		const tier = classifyBootstrapUserMessage(message);
+		if (tier === 1 || tier === 2 || tier === 3) continue;
+		if (!Array.isArray(message.content)) continue;
+		for (const block of message.content) if (typeof block === "object" && block !== null && block.type === "text") {
+			const text = typeof block.text === "string" ? block.text : "";
+			if (text.includes("dsh-mandate-v2")) return true;
+			dynamicChars += text.length;
+		}
+	}
+	return dynamicChars >= 1200;
+}
+/**
+* Stagger concurrent subagent first-turn requests on the same account/model
+* so they do not stampede prefill in the same millisecond.
+*/
+async function acquireSubagentBurstPacing(pacingKey, signal) {
+	while (true) {
+		const existing = inFlightSubagentBurstLeaders.get(pacingKey);
+		if (existing === void 0) break;
+		await raceColdStartLeader(existing, signal);
+		if (signal.aborted) return () => void 0;
+		if (inFlightSubagentBurstLeaders.get(pacingKey) === existing) {
+			subagentBurstReleases.get(pacingKey)?.();
+			break;
+		}
+	}
+	let resolvePacing;
+	const promise = new Promise((resolve) => {
+		resolvePacing = resolve;
+	});
+	inFlightSubagentBurstLeaders.set(pacingKey, promise);
+	let settled = false;
+	const release = () => {
+		if (settled) return;
+		settled = true;
+		resolvePacing();
+		if (subagentBurstReleases.get(pacingKey) === release) {
+			subagentBurstReleases.delete(pacingKey);
+			inFlightSubagentBurstLeaders.delete(pacingKey);
+		}
+	};
+	subagentBurstReleases.set(pacingKey, release);
+	return release;
+}
+/**
 * Wait for the leader of a prefix group, for the wait budget, or for an abort.
 *
 * The timer and the abort listener are removed whichever way the race ends, so
@@ -37506,322 +37800,338 @@ var KimiCodeAdapter = class extends LlmAdapter {
 	}
 	async *requestStream(options, signal, activity) {
 		const fetchFn = this.options.fetchFn ?? fetch;
-		const catalog = await this.catalog().catch(() => []);
-		const wire = wireForCatalogEntry(options.model, catalog);
-		const entry = catalog.find((model) => model.id === options.model);
 		const normalized = normalizeGenerateOptions(options);
 		const sideChannel = isSideChannelTitleRequest(normalized);
-		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);
-		const stickyOffload = offloadRequestImagesWithStickyBudget(normalized, maxRequestImageBytes$1(), lowRequestImageBytes(), images, priorDroppedKeys);
-		const imageTrimmed = stickyOffload.options;
-		if (!sideChannel) markPendingImageOffload(options.sessionId, stickyOffload.newlyDroppedKeys.length);
-		const gatedEffort = coerceThinkingEffort(options.model, mapReasoningEffort(options.reasoningEffort === void 0 || options.reasoningEffort === null ? void 0 : String(options.reasoningEffort)));
-		if (gatedEffort.coercedFrom !== null) recordThinkingCoercion({
-			model: options.model,
-			wire,
-			from: gatedEffort.coercedFrom,
-			to: gatedEffort.effort
-		}, options.sessionId);
-		const videoTrimmed = offloadOldestRequestVideos(imageTrimmed, MAX_REQUEST_VIDEO_BYTES$1);
-		const requestOptions = gatedEffort.effort === void 0 ? videoTrimmed : {
-			...videoTrimmed,
-			reasoningEffort: ReasoningEffortId(gatedEffort.effort)
-		};
-		const videos = await resolveRequestVideos(requestOptions, this.options.videos, signal);
-		const settings = await this.settings();
-		const modalities = inputModalitiesForEntry(options.model, catalog);
-		/**
-		* Everything every attempt shares: the media reads and the capability gates.
-		*
-		* Account-specific references are merged on top of this per attempt, so the
-		* gate is applied once and a reference can never be cited for a modality the
-		* model does not declare.
-		*/
-		const baseMedia = {
-			videos,
-			videoAccepted: modalities.includes("video"),
-			messageTools: dynamicToolsForEntry(options.model, catalog)
-		};
-		/**
-		* Which readable occurrences the reference path would cover.
-		*
-		* Decided here so the body-fit guard can measure the body that will actually
-		* be sent; nothing is uploaded, and no network is touched, until that guard
-		* has accepted the request (see the rotation loop).
-		*/
-		const planned = planMediaReferences({
-			wire,
-			imageAccepted: modalities.includes("image"),
-			videoAccepted: baseMedia.videoAccepted === true,
-			images,
-			videos
-		});
-		const contextWindow = this.contextWindowFor(options.model, entry, settings.contextWindowOverrides);
-		const requestedMax = options.maxTokens ?? maxOutputTokensFor$4(options.model, contextWindow);
-		const inputCap = maxInputTokensForEntry(options.model, catalog);
-		const promptLimit = inputCap === void 0 ? contextWindow : Math.min(contextWindow, inputCap);
-		const boundedOptions = {
-			...requestOptions,
-			maxTokens: clampOutputToContext$1(requestedMax, promptLimit, estimatedInputTokens(requestOptions))
-		};
-		const carriesVideo = requestHasVideo(requestOptions);
-		const bodyLimit = carriesVideo ? MAX_VIDEO_MESSAGE_BODY_BYTES : maxRequestBodyBytesFor(kimiCodeModelDef(entry?.id ?? options.model), process.env);
-		/**
-		* Run the greedy body-fit loop for ONE set of media references.
-		*
-		* A function rather than a loop in place because the same loop now serves
-		* every build of one request: the pre-upload MEASUREMENT (placeholder
-		* references), the real reference build after the upload, and the all-inline
-		* rebuild the missing-resource fallback re-sends.
-		*
-		* Every call restarts from `boundedOptions`. That is load bearing rather than
-		* tidy: the inline fallback has to re-fit the inline body from scratch and
-		* stay inside the same ceiling instead of inheriting an omission the
-		* reference path made to fit a different, much smaller body.
-		*/
-		const buildBodyWithinLimit = (media) => {
-			let buildOptions = boundedOptions;
-			for (let attempt = 0;; attempt++) {
-				const built = buildRequest(buildOptions, wire, images, void 0, media);
-				const body = JSON.stringify(built);
-				const bytes = Buffer.byteLength(body, "utf8");
-				if (bytes <= bodyLimit) return {
-					body,
-					options: buildOptions
-				};
-				const context = () => ({
-					droppedImages: countOmitted(buildOptions, requestOptions, images),
-					sendableImagesRemaining: requestImagePayloadLengths(buildOptions, images).length,
-					excessBytes: bytes - bodyLimit
-				});
-				if (attempt > MAX_BODY_FIT_ATTEMPTS || !requestHasSendableImage(buildOptions, images)) {
-					assertRequestBodyFits$1(built, carriesVideo, context(), bodyLimit);
-					throw new Error("unreachable: assertRequestBodyFits returns or throws");
-				}
-				const drop = dropOldestImagesUntilFits(buildOptions, images, bytes - bodyLimit, requestOptions);
-				if (drop.dropped === 0) {
-					assertRequestBodyFits$1(built, carriesVideo, context(), bodyLimit);
-					throw new Error("unreachable: assertRequestBodyFits returns or throws");
-				}
-				buildOptions = drop.options;
-			}
-		};
-		/**
-		* The measured build for the reference path, done at most once.
-		*
-		* Held across rotation attempts because it is account-independent: it decides
-		* whether this request can be sent at all, and that answer cannot change
-		* because a different account answered.
-		*
-		* The OPTIONS travel with the body because they, not the plan, name the media
-		* this request still carries: see {@link mediaAttachmentIdsIn}.
-		*/
-		let measured;
-		let sentBuild;
-		const pool = this.accountPool;
-		const tried = /* @__PURE__ */ new Set();
-		let servedByAccountId;
-		let response;
-		while (true) {
-			let apiCredentials;
-			let accountId;
-			if (pool === null) try {
-				apiCredentials = await ensureAccessToken$1(this.store, {
-					fetchFn,
-					signal,
-					maxAttempts: MODEL_REFRESH_MAX_ATTEMPTS
-				});
-			} catch (error) {
-				throw credentialFailure(error, signal);
-			}
-			else {
-				let effective;
-				try {
-					effective = await pool.getEffectiveCredential(tried, fetchFn, { maxAttempts: MODEL_REFRESH_MAX_ATTEMPTS });
-				} catch (error) {
-					throw credentialFailure(error, signal);
-				}
-				accountId = effective.account.id;
-				servedByAccountId = accountId;
-				tried.add(accountId);
-				apiCredentials = effective.credentials;
-			}
-			const region = apiCredentials.region ?? await resolveRegion();
-			const base = (apiCredentials.baseUrl ?? codingBaseUrl(region)).replace(/\/+$/, "");
-			const endpoint = wire === "anthropic" ? `${base}/v1/messages?beta=true` : `${base}/v1/chat/completions`;
-			const headers = await modelRequestHeaders$1(apiCredentials.accessToken, wire);
+		const sessionKey = typeof options.sessionId === "string" ? options.sessionId.trim() : "";
+		const mainHeaderRelease = !sideChannel && sessionKey !== "" ? registerInFlightMainSessionHeader(sessionKey) : () => void 0;
+		try {
+			const catalog = await this.catalog().catch(() => []);
+			const wire = wireForCatalogEntry(options.model, catalog);
+			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);
+			const stickyOffload = offloadRequestImagesWithStickyBudget(normalized, maxRequestImageBytes$1(), lowRequestImageBytes(), images, priorDroppedKeys);
+			const imageTrimmed = stickyOffload.options;
+			if (!sideChannel) markPendingImageOffload(options.sessionId, stickyOffload.newlyDroppedKeys.length);
+			const gatedEffort = coerceThinkingEffort(options.model, mapReasoningEffort(options.reasoningEffort === void 0 || options.reasoningEffort === null ? void 0 : String(options.reasoningEffort)));
+			if (gatedEffort.coercedFrom !== null) recordThinkingCoercion({
+				model: options.model,
+				wire,
+				from: gatedEffort.coercedFrom,
+				to: gatedEffort.effort
+			}, options.sessionId);
+			const videoTrimmed = offloadOldestRequestVideos(imageTrimmed, MAX_REQUEST_VIDEO_BYTES$1);
+			const requestOptions = gatedEffort.effort === void 0 ? videoTrimmed : {
+				...videoTrimmed,
+				reasoningEffort: ReasoningEffortId(gatedEffort.effort)
+			};
+			const videos = await resolveRequestVideos(requestOptions, this.options.videos, signal);
+			const settings = await this.settings();
+			const modalities = inputModalitiesForEntry(options.model, catalog);
 			/**
-			* Which account's file store this attempt writes to and cites from.
+			* Everything every attempt shares: the media reads and the capability gates.
 			*
-			* A pooled account's own id, because that is the only identity a reference
-			* may be cached under; a single stored credential uses the user id its
-			* sign-in reported, or a constant when it reported none. The access token is
-			* never used, here or anywhere a cache key is built.
+			* Account-specific references are merged on top of this per attempt, so the
+			* gate is applied once and a reference can never be cited for a modality the
+			* model does not declare.
 			*/
-			const accountScope = accountId ?? legacyAccountScope(apiCredentials);
-			if (measured === void 0) measured = buildBodyWithinLimit(planned.length === 0 ? baseMedia : {
-				...baseMedia,
-				...plannedReferenceMaps(planned)
-			});
-			const stillSent = mediaAttachmentIdsIn(measured.options);
-			const uploadable = planned.filter((item) => stillSent.has(item.attachmentId));
-			const uploads = uploadable.length === 0 ? void 0 : await this.uploadPlannedReferences({
-				planned: uploadable,
-				accessToken: apiCredentials.accessToken,
-				baseUrl: base,
-				accountScope,
-				signal
-			});
-			let finalBuild = uploads === void 0 ? planned.length === 0 ? measured : buildBodyWithinLimit(baseMedia) : buildBodyWithinLimit({
-				...baseMedia,
-				imageReferences: uploads.imageReferences,
-				videoReferences: uploads.videoReferences
+			const baseMedia = {
+				videos,
+				videoAccepted: modalities.includes("video"),
+				messageTools: dynamicToolsForEntry(options.model, catalog)
+			};
+			/**
+			* Which readable occurrences the reference path would cover.
+			*
+			* Decided here so the body-fit guard can measure the body that will actually
+			* be sent; nothing is uploaded, and no network is touched, until that guard
+			* has accepted the request (see the rotation loop).
+			*/
+			const planned = planMediaReferences({
+				wire,
+				imageAccepted: modalities.includes("image"),
+				videoAccepted: baseMedia.videoAccepted === true,
+				images,
+				videos
 			});
-			sentBuild = finalBuild;
-			let body = finalBuild.body;
-			const post = async (payload) => {
-				try {
-					const response = await fetchWithContextBudgetRecovery(fetchFn)(endpoint, {
-						method: "POST",
-						headers,
-						body: payload,
-						signal
+			const contextWindow = this.contextWindowFor(options.model, entry, settings.contextWindowOverrides);
+			const requestedMax = options.maxTokens ?? maxOutputTokensFor$4(options.model, contextWindow);
+			const inputCap = maxInputTokensForEntry(options.model, catalog);
+			const promptLimit = inputCap === void 0 ? contextWindow : Math.min(contextWindow, inputCap);
+			const boundedOptions = {
+				...requestOptions,
+				maxTokens: clampOutputToContext$1(requestedMax, promptLimit, estimatedInputTokens(requestOptions))
+			};
+			const carriesVideo = requestHasVideo(requestOptions);
+			const bodyLimit = carriesVideo ? MAX_VIDEO_MESSAGE_BODY_BYTES : maxRequestBodyBytesFor(kimiCodeModelDef(entry?.id ?? options.model), process.env);
+			/**
+			* Run the greedy body-fit loop for ONE set of media references.
+			*
+			* A function rather than a loop in place because the same loop now serves
+			* every build of one request: the pre-upload MEASUREMENT (placeholder
+			* references), the real reference build after the upload, and the all-inline
+			* rebuild the missing-resource fallback re-sends.
+			*
+			* Every call restarts from `boundedOptions`. That is load bearing rather than
+			* tidy: the inline fallback has to re-fit the inline body from scratch and
+			* stay inside the same ceiling instead of inheriting an omission the
+			* reference path made to fit a different, much smaller body.
+			*/
+			const buildBodyWithinLimit = (media) => {
+				let buildOptions = boundedOptions;
+				for (let attempt = 0;; attempt++) {
+					const built = buildRequest(buildOptions, wire, images, void 0, media);
+					const body = JSON.stringify(built);
+					const bytes = Buffer.byteLength(body, "utf8");
+					if (bytes <= bodyLimit) return {
+						body,
+						options: buildOptions
+					};
+					const context = () => ({
+						droppedImages: countOmitted(buildOptions, requestOptions, images),
+						sendableImagesRemaining: requestImagePayloadLengths(buildOptions, images).length,
+						excessBytes: bytes - bodyLimit
 					});
-					activity?.pulse();
-					return response;
-				} catch (error) {
-					if (signal.aborted) throw new LlmError("Kimi Code request aborted", "ABORTED", { cause: error });
-					throw new LlmError(`Kimi Code request failed: ${error instanceof Error ? error.message : String(error)}`, "TRANSPORT", { cause: error });
+					if (attempt > MAX_BODY_FIT_ATTEMPTS || !requestHasSendableImage(buildOptions, images)) {
+						assertRequestBodyFits$1(built, carriesVideo, context(), bodyLimit);
+						throw new Error("unreachable: assertRequestBodyFits returns or throws");
+					}
+					const drop = dropOldestImagesUntilFits(buildOptions, images, bytes - bodyLimit, requestOptions);
+					if (drop.dropped === 0) {
+						assertRequestBodyFits$1(built, carriesVideo, context(), bodyLimit);
+						throw new Error("unreachable: assertRequestBodyFits returns or throws");
+					}
+					buildOptions = drop.options;
 				}
 			};
-			const coldGroupKey = sideChannel ? void 0 : coldStartGroupKey(region, requestOptions);
-			const coldGate = coldGroupKey === void 0 ? void 0 : await acquireColdStartWarmup(coldGroupKey, signal);
-			let coldGateSettled = false;
-			const settleColdGate = (warm) => {
-				if (coldGateSettled) return;
-				coldGateSettled = true;
-				coldGate?.release();
-				if (warm) coldGate?.markWarm();
-			};
-			response = await post(body).then((value) => {
-				settleColdGate(value.ok);
-				return value;
-			}, (error) => {
-				settleColdGate(false);
-				throw error;
-			});
-			let detail = response.ok ? "" : (await response.text().catch(() => "")).slice(0, 2e3);
-			if (response.status === 404 && uploads !== void 0 && uploads.cited.length > 0 && isMissingResourceBody(detail)) {
-				const uploader = this.mediaFiles();
-				for (const item of uploads.cited) uploader.forget({
+			/**
+			* The measured build for the reference path, done at most once.
+			*
+			* Held across rotation attempts because it is account-independent: it decides
+			* whether this request can be sent at all, and that answer cannot change
+			* because a different account answered.
+			*
+			* The OPTIONS travel with the body because they, not the plan, name the media
+			* this request still carries: see {@link mediaAttachmentIdsIn}.
+			*/
+			let measured;
+			let sentBuild;
+			const pool = this.accountPool;
+			const tried = /* @__PURE__ */ new Set();
+			let servedByAccountId;
+			let response;
+			while (true) {
+				let apiCredentials;
+				let accountId;
+				if (pool === null) try {
+					apiCredentials = await ensureAccessToken$1(this.store, {
+						fetchFn,
+						signal,
+						maxAttempts: MODEL_REFRESH_MAX_ATTEMPTS
+					});
+				} catch (error) {
+					throw credentialFailure(error, signal);
+				}
+				else {
+					let effective;
+					try {
+						effective = await pool.getEffectiveCredential(tried, fetchFn, { maxAttempts: MODEL_REFRESH_MAX_ATTEMPTS });
+					} catch (error) {
+						throw credentialFailure(error, signal);
+					}
+					accountId = effective.account.id;
+					servedByAccountId = accountId;
+					tried.add(accountId);
+					apiCredentials = effective.credentials;
+				}
+				const region = apiCredentials.region ?? await resolveRegion();
+				const base = (apiCredentials.baseUrl ?? codingBaseUrl(region)).replace(/\/+$/, "");
+				const endpoint = wire === "anthropic" ? `${base}/v1/messages?beta=true` : `${base}/v1/chat/completions`;
+				const headers = await modelRequestHeaders$1(apiCredentials.accessToken, wire);
+				/**
+				* Which account's file store this attempt writes to and cites from.
+				*
+				* A pooled account's own id, because that is the only identity a reference
+				* may be cached under; a single stored credential uses the user id its
+				* sign-in reported, or a constant when it reported none. The access token is
+				* never used, here or anywhere a cache key is built.
+				*/
+				const accountScope = accountId ?? legacyAccountScope(apiCredentials);
+				if (measured === void 0) measured = buildBodyWithinLimit(planned.length === 0 ? baseMedia : {
+					...baseMedia,
+					...plannedReferenceMaps(planned)
+				});
+				const stillSent = mediaAttachmentIdsIn(measured.options);
+				const uploadable = planned.filter((item) => stillSent.has(item.attachmentId));
+				const uploads = uploadable.length === 0 ? void 0 : await this.uploadPlannedReferences({
+					planned: uploadable,
+					accessToken: apiCredentials.accessToken,
 					baseUrl: base,
 					accountScope,
-					mediaType: item.mediaType,
-					data: item.data
+					signal
 				});
-				let inlineBuild;
-				try {
-					inlineBuild = buildBodyWithinLimit(baseMedia);
-				} catch {
-					inlineBuild = void 0;
+				let finalBuild = uploads === void 0 ? planned.length === 0 ? measured : buildBodyWithinLimit(baseMedia) : buildBodyWithinLimit({
+					...baseMedia,
+					imageReferences: uploads.imageReferences,
+					videoReferences: uploads.videoReferences
+				});
+				sentBuild = finalBuild;
+				let body = finalBuild.body;
+				const post = async (payload) => {
+					try {
+						const response = await fetchWithContextBudgetRecovery(fetchFn)(endpoint, {
+							method: "POST",
+							headers,
+							body: payload,
+							signal
+						});
+						activity?.pulse();
+						return response;
+					} catch (error) {
+						if (signal.aborted) throw new LlmError("Kimi Code request aborted", "ABORTED", { cause: error });
+						throw new LlmError(`Kimi Code request failed: ${error instanceof Error ? error.message : String(error)}`, "TRANSPORT", { cause: error });
+					}
+				};
+				if (sideChannel && sessionKey !== "") {
+					const pendingMain = inFlightMainSessionHeaders.get(sessionKey);
+					if (pendingMain !== void 0) {
+						await raceColdStartLeader(pendingMain, signal);
+						if (signal.aborted) throw new LlmError("Kimi Code request aborted", "ABORTED");
+					}
 				}
-				if (inlineBuild !== void 0) {
-					finalBuild = inlineBuild;
-					sentBuild = inlineBuild;
-					body = inlineBuild.body;
-					response = await post(body);
-					detail = response.ok ? "" : (await response.text().catch(() => "")).slice(0, 2e3);
+				const coldGroupKey = sideChannel ? void 0 : coldStartGroupKey(region, requestOptions);
+				const coldGate = coldGroupKey === void 0 ? void 0 : await acquireColdStartWarmup(coldGroupKey, signal);
+				const burstPacingRelease = !sideChannel && sessionKey !== "" && getCacheStats(sessionKey).requests === 0 && isSubagentBurstCandidate(requestOptions) ? await acquireSubagentBurstPacing(`${accountScope}:${options.model}`, signal) : () => void 0;
+				let coldGateSettled = false;
+				const settleColdGate = (warm) => {
+					if (coldGateSettled) return;
+					coldGateSettled = true;
+					coldGate?.release();
+					burstPacingRelease();
+					mainHeaderRelease();
+					if (warm) coldGate?.markWarm();
+				};
+				response = await post(body).then((value) => {
+					settleColdGate(value.ok);
+					return value;
+				}, (error) => {
+					settleColdGate(false);
+					throw error;
+				});
+				let detail = response.ok ? "" : (await response.text().catch(() => "")).slice(0, 2e3);
+				if (response.status === 404 && uploads !== void 0 && uploads.cited.length > 0 && isMissingResourceBody(detail)) {
+					const uploader = this.mediaFiles();
+					for (const item of uploads.cited) uploader.forget({
+						baseUrl: base,
+						accountScope,
+						mediaType: item.mediaType,
+						data: item.data
+					});
+					let inlineBuild;
+					try {
+						inlineBuild = buildBodyWithinLimit(baseMedia);
+					} catch {
+						inlineBuild = void 0;
+					}
+					if (inlineBuild !== void 0) {
+						finalBuild = inlineBuild;
+						sentBuild = inlineBuild;
+						body = inlineBuild.body;
+						response = await post(body);
+						detail = response.ok ? "" : (await response.text().catch(() => "")).slice(0, 2e3);
+					}
 				}
-			}
-			if (response.ok) break;
-			const failure = classifyKimiFailure(response.status, detail);
-			const after = response.status === 429 ? retryAfterMs$1(response.headers) : void 0;
-			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$8} 429`).catch(() => void 0);
-					if (await pool.hasAnotherAvailableAccount(tried)) continue;
-				} else if (failure.code === "INVALID_CREDENTIAL") {
-					await pool.markAuthFailed(accountId, failure.message).catch(() => void 0);
-					if (await pool.hasAnotherAvailableAccount(tried)) continue;
-				} else if (failure.accountScoped) {
-					await pool.markCooldown(accountId, accountLimitCooldownMs(detail), `${PROVIDER_NAME$8} 403`).catch(() => void 0);
-					if (await pool.hasAnotherAvailableAccount(tried)) continue;
+				if (response.ok) break;
+				const failure = classifyKimiFailure(response.status, detail);
+				const after = response.status === 429 ? retryAfterMs$1(response.headers) : void 0;
+				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$8} 429`).catch(() => void 0);
+						if (await pool.hasAnotherAvailableAccount(tried)) continue;
+					} else if (failure.code === "INVALID_CREDENTIAL") {
+						await pool.markAuthFailed(accountId, failure.message).catch(() => void 0);
+						if (await pool.hasAnotherAvailableAccount(tried)) continue;
+					} else if (failure.accountScoped) {
+						await pool.markCooldown(accountId, accountLimitCooldownMs(detail), `${PROVIDER_NAME$8} 403`).catch(() => void 0);
+						if (await pool.hasAnotherAvailableAccount(tried)) continue;
+					}
 				}
+				throw new LlmError(failure.message, failure.code, {
+					status: response.status,
+					...after === void 0 ? {} : { providerRetryAfterMs: after }
+				});
 			}
-			throw new LlmError(failure.message, failure.code, {
-				status: response.status,
-				...after === void 0 ? {} : { providerRetryAfterMs: after }
-			});
-		}
-		if (sentBuild !== void 0 && !sideChannel) {
-			const requestImageBytes = requestImagePayloadLengths(sentBuild.options, images).reduce((sum, bytes) => sum + bytes, 0);
-			const trimmedKeys = new Set(collectRemainingImageKeys(imageTrimmed, images));
-			const sentKeys = new Set(collectRemainingImageKeys(sentBuild.options, images));
-			const bodyFitDroppedKeys = [...trimmedKeys].filter((key) => !sentKeys.has(key));
-			const allDroppedKeys = /* @__PURE__ */ new Set([...stickyOffload.allDroppedKeys, ...bodyFitDroppedKeys]);
-			recordSessionImageBudget(options.sessionId, {
-				allDroppedKeys,
-				newlyDroppedCount: stickyOffload.newlyDroppedKeys.length + bodyFitDroppedKeys.length,
-				requestImageBytes,
-				droppedImages: stickyOffload.droppedCount + bodyFitDroppedKeys.length
-			}, requestOptions);
-		}
-		if (response.body === null) throw new LlmError("Kimi Code returned an empty response body", "PROVIDER_ERROR");
-		const reader = response.body.getReader();
-		const decoder = new TextDecoder();
-		const state = createStreamState$5(wire, requestOptions.sessionId, sideChannel);
-		applyCacheWriteTiers(state, response.headers);
-		if (!sideChannel) markSessionActive(requestOptions.sessionId);
-		let buffer = "";
-		/**
-		* Set the moment ANY chunk reaches the caller, and never cleared.
-		*
-		* Deliberately broader than "text arrived": a block-start the caller has
-		* already seen repeats just as badly as text it has already read, so the
-		* conservative reading is the one that cannot be wrong. It is what keeps the
-		* in-band reclassification below pre-output only.
-		*/
-		let outputStarted = false;
-		try {
+			if (sentBuild !== void 0 && !sideChannel) {
+				const requestImageBytes = requestImagePayloadLengths(sentBuild.options, images).reduce((sum, bytes) => sum + bytes, 0);
+				const trimmedKeys = new Set(collectRemainingImageKeys(imageTrimmed, images));
+				const sentKeys = new Set(collectRemainingImageKeys(sentBuild.options, images));
+				const bodyFitDroppedKeys = [...trimmedKeys].filter((key) => !sentKeys.has(key));
+				const allDroppedKeys = /* @__PURE__ */ new Set([...stickyOffload.allDroppedKeys, ...bodyFitDroppedKeys]);
+				recordSessionImageBudget(options.sessionId, {
+					allDroppedKeys,
+					newlyDroppedCount: stickyOffload.newlyDroppedKeys.length + bodyFitDroppedKeys.length,
+					requestImageBytes,
+					droppedImages: stickyOffload.droppedCount + bodyFitDroppedKeys.length
+				}, requestOptions);
+			}
+			if (response.body === null) throw new LlmError("Kimi Code returned an empty response body", "PROVIDER_ERROR");
+			const reader = response.body.getReader();
+			const decoder = new TextDecoder();
+			const state = createStreamState$5(wire, requestOptions.sessionId, sideChannel);
+			applyCacheWriteTiers(state, response.headers);
+			if (!sideChannel) markSessionActive(requestOptions.sessionId);
+			let buffer = "";
+			/**
+			* Set the moment ANY chunk reaches the caller, and never cleared.
+			*
+			* Deliberately broader than "text arrived": a block-start the caller has
+			* already seen repeats just as badly as text it has already read, so the
+			* conservative reading is the one that cannot be wrong. It is what keeps the
+			* in-band reclassification below pre-output only.
+			*/
+			let outputStarted = false;
 			try {
-				while (true) {
-					const { done, value } = await reader.read();
-					if (done) break;
-					if (value !== void 0 && value.byteLength > 0) activity?.pulse();
-					buffer += decoder.decode(value, { stream: true });
-					const lines = buffer.split("\n");
-					buffer = lines.pop() ?? "";
-					for (const line of lines) {
-						activity?.pulse();
-						for (const chunk of processLine(line, state, wire)) {
-							outputStarted = true;
-							yield chunk;
+				try {
+					while (true) {
+						const { done, value } = await reader.read();
+						if (done) break;
+						if (value !== void 0 && value.byteLength > 0) activity?.pulse();
+						buffer += decoder.decode(value, { stream: true });
+						const lines = buffer.split("\n");
+						buffer = lines.pop() ?? "";
+						for (const line of lines) {
+							activity?.pulse();
+							for (const chunk of processLine(line, state, wire)) {
+								outputStarted = true;
+								yield chunk;
+							}
+							if (state.finished) return;
 						}
-						if (state.finished) return;
 					}
+					buffer += decoder.decode();
+					if (buffer.trim() !== "") for (const line of buffer.split("\n")) for (const chunk of processLine(line, state, wire)) {
+						outputStarted = true;
+						yield chunk;
+					}
+					if (state.finished) return;
+					assertStreamComplete$5(state);
+					for (const chunk of closeStream$4(state, servedByAccountId)) {
+						outputStarted = true;
+						yield chunk;
+					}
+				} catch (error) {
+					const inBand = state.streamError;
+					if (error instanceof LlmError && !outputStarted && error.code === "PROVIDER_ERROR" && inBand !== void 0) throw await this.inBandStreamFailure(error, inBand, servedByAccountId);
+					throw streamFailure(error, signal);
 				}
-				buffer += decoder.decode();
-				if (buffer.trim() !== "") for (const line of buffer.split("\n")) for (const chunk of processLine(line, state, wire)) {
-					outputStarted = true;
-					yield chunk;
-				}
-				if (state.finished) return;
-				assertStreamComplete$5(state);
-				for (const chunk of closeStream$4(state, servedByAccountId)) {
-					outputStarted = true;
-					yield chunk;
-				}
-			} catch (error) {
-				const inBand = state.streamError;
-				if (error instanceof LlmError && !outputStarted && error.code === "PROVIDER_ERROR" && inBand !== void 0) throw await this.inBandStreamFailure(error, inBand, servedByAccountId);
-				throw streamFailure(error, signal);
+			} finally {
+				if (state.sawUsage && !sideChannel) noteSessionCacheState(requestOptions.sessionId ?? state.sessionId, state);
+				reader.cancel().catch(() => void 0);
 			}
 		} finally {
-			if (state.sawUsage && !sideChannel) noteSessionCacheState(requestOptions.sessionId ?? state.sessionId, state);
-			reader.cancel().catch(() => void 0);
+			mainHeaderRelease();
 		}
 	}
 	/**
@@ -95427,12 +95737,19 @@ function messageText(data) {
 	return "";
 }
 /**
-* 最新用户消息是否明确属于导航任务。
+* 最新用户消息是否明确属于导航任务，或当前会话正处于只读计划模式（`plan/mode` `active: true`）。
 * @param agent - 工具执行事件里的 agent。
-* @returns 只有明确导航且无源码工作信号时为 true。
+* @returns 明确导航（无源码工作信号）或处于只读计划模式时为 true。
 */
 function cacheWarmupNavigationIntentOf(agent) {
 	const events = eventsOf(agent);
+	for (let index = events.length - 1; index >= 0; index -= 1) {
+		const event = events[index];
+		if (event.type === "plan/mode") {
+			if ((typeof event.data === "object" && event.data !== null ? event.data : void 0)?.active === true) return true;
+			break;
+		}
+	}
 	for (let index = events.length - 1; index >= 0; index -= 1) {
 		const event = events[index];
 		if (event.type !== "user/message") continue;
@@ -95721,7 +96038,13 @@ function installCacheWarmup(ctx, options) {
 		const ptcRetrieval = isPtcTransportName(exec.name) && settings.retrievalEnabled === true;
 		if (sourceTool && !settings.enabled && !ptcRetrieval) return decision;
 		if (retrievalTool && settings.retrievalEnabled !== true) return decision;
-		const navigation = retrievalTool || ptcRetrieval ? cacheWarmupNavigationIntentOf(exec.agent) : void 0;
+		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) => {
+			return sum + (typeof block === "object" && block !== null && typeof block.text === "string" ? block.text : "").length;
+		}, 0) >= 2e4;
+		const navigation = baseNavigation === true || isOversizedSearchDump ? true : baseNavigation;
 		if (retrievalTool && navigation !== true) return decision;
 		if (sourceTool && !settings.enabled && navigation !== true) return decision;
 		const route = cacheWarmupRouteOf(exec.agent);
@@ -95730,8 +96053,6 @@ function installCacheWarmup(ctx, options) {
 		if (sessionId === void 0) return decision;
 		const store = storeOf();
 		if (store === void 0) return decision;
-		const blocks = decision.content ?? result.content;
-		if (blocks.length === 0) return decision;
 		const content = blocks;
 		let durable;
 		try {