@shuind/dsh-codex-harness 0.1.24 → 0.1.26

package/README.md

@@ -12,13 +12,15 @@
 - **旧版 Web 兼容**：在未提供新版模型设置/上下文设置 slot 的 DSH Web 中，Fast 和上下文容量会自动回退到旧版 composer slot；client 注入不依赖纯类型 slot 包。
 - **Responses 原生 apply_patch**：在 GPT Responses 请求的插件传输边界，把普通 JSON function tool 改写为 `type: "custom"` + OpenAI `lark` grammar；下一轮历史同步改写为 `custom_tool_call`。如果中转站拒绝 custom tool，则自动回退到普通 function tool，不改变 DSH 工具执行器。
 - **远程搜索与压缩**：OpenAI Responses 请求默认优先使用 hosted `web_search` 和 `/responses/compact`；失败时回退到 DSH 的本地实现。
+- **远程压缩用量**：`/responses/compact` 返回标准 `usage` 时，插件会把 `input_tokens`、缓存读写和 `output_tokens` 原样拆分为 DSH 的用量字段；不再把远程压缩伪造成 `0`。中转站缺失或返回不一致的 usage 时，插件不会编造数字。
 - **上下文容量**：在 Web 中设置 `1K`–`1000K` tokens，设置值作用于下一次请求。
-- **活动状态**：Codex 等待模型响应或进行上下文压缩时，在 `Deep diving...` 右侧显示对应状态和耗时，让长时间 Deep Diving 不再像卡住。
+- **自动压缩阈值**：Codex preset 的自动压缩会读取最新请求持久化的有效上下文容量；例如设置 `400K` 时，默认 `80%` pressure threshold 是 `320K`，不会错误地按适配器默认 `262,144` 在约 `210K` 提前触发。
+- **活动状态**：Codex 请求模型、接收模型回复或进行上下文压缩时，在 `Deep diving...` 右侧显示对应阶段和耗时，让长时间 Deep Diving 不再像卡住。
 
 ## 安装
 
 ```sh
-dsh plugin --profile web add @shuind/dsh-codex-harness@0.1.24
+dsh plugin --profile web add @shuind/dsh-codex-harness@0.1.26
 ```
 
 重启 Web，创建新会话，在模式菜单中选择 **Codex 模式**。
@@ -51,13 +53,13 @@ Fast 设置保存在 `codex` 命名空间，只对 Codex preset 生效。
 
 ### 上下文容量
 
-点击 Web 顶部的上下文使用量指示器，可以设置下一次请求的上下文容量。上方 meter 显示当前请求的实际容量；修改设置后，需要发送下一次请求才会更新。
+点击 Web 顶部的上下文使用量指示器，可以设置下一次请求的上下文容量。上方 meter 显示当前请求的实际容量；修改设置后，需要发送下一次请求才会更新。Codex preset 的自动压缩也使用这个已经持久化的有效容量。
 
 上下文容量使用整数 K tokens，范围为 `1K` 到 `1000K`。实际可用上限仍取决于模型和中转站支持情况。
 
 ### 活动状态
 
-Codex 等待模型首个输出时，Web 会在 `Deep diving...` 右侧显示 `正在等待模型响应... · Ns`；开始上下文压缩后会切换为 `正在压缩上下文... · Ns`。收到模型输出、压缩结束或下一轮开始后自动隐藏。该状态由 Host 的 `sessionProjections` 投影驱动，因此本地压缩和远程 `/responses/compact` 都会显示同一种状态。
+Codex 发起模型请求时，Web 会在 `Deep diving...` 右侧显示 `请求模型... · Ns`；收到首个非空模型 token（非流式适配器则收到 `assistant/message`）后切换为 `模型回复... · Ns`。模型回复结束、进入工具调用、压缩结束或下一轮开始后自动隐藏。该状态由 Host 的 `sessionProjections` 投影驱动，因此本地压缩和远程 `/responses/compact` 都会显示同一种状态。
 
 如果当前 DSH 组合没有加载 projection 服务，插件会跳过注册，客户端也会自动隐藏这行，不影响普通会话。
 
@@ -105,7 +107,7 @@ description: 使用 Codex 工具并开启可选协作提示词。
 order: 10
 ```
 
-然后在 Web 的 **Agent 预设** 中选择 `My Codex`，再新建会话。这个 preset 会加载 Codex 工具，并开启协作提示词；网页搜索、远程压缩、后台任务控制和 Skills 等其他功能需要在同一个 `agent.cordis.yml` 中按需添加。
+然后在 Web 的 **Agent 预设** 中选择 `My Codex`，再新建会话。这个 preset 会加载 Codex 工具，并开启协作提示词；网页搜索、远程压缩、后台任务控制和 Skills 等其他功能需要在同一个 `agent.cordis.yml` 中按需添加。若希望自定义 preset 也修正自动压缩阈值，请在 compaction 隔离组中使用 `@shuind/dsh-codex-harness/compaction` 替代 `@deepseek-ai/dsh-compaction-basic`。
 
 ### 可选协作提示词（私货）
 

package/lib/client.js

@@ -21,13 +21,52 @@ window.__ModuleLoader__.load({
 		//#endregion
 		//#region lib/types/client/activity.js
 		/** Client-only inference used while the host activity projection frame is in flight. */
-		/** Infer the active model phase from the existing conversation snapshot. */
+		/** Equality for selector hooks, whose default is reference identity. */
+		function sameModelActivity(left, right) {
+			return left?.activity === right?.activity && left?.startedAt === right?.startedAt;
+		}
+		/** Infer the request-phase clock from the existing conversation snapshot. */
 		function awaitingModelStartedAt(snapshot) {
 			if (!snapshot.running || snapshot.runningCalls.length > 0) return void 0;
 			let latest;
 			for (const timing of snapshot.turnTimings.values()) if (timing.endTime === void 0) latest = timing.startTime;
 			return latest;
 		}
+		/** Infer the first non-empty model token timestamp from the current snapshot. */
+		function modelReplyStartedAt(snapshot) {
+			if (!snapshot.running || snapshot.runningCalls.length > 0) return void 0;
+			let openTurn;
+			for (const [turn, timing] of snapshot.turnTimings) if (timing.endTime === void 0) openTurn = {
+				turn,
+				startTime: timing.startTime
+			};
+			if (openTurn === void 0) return void 0;
+			if (snapshot.partial !== null && snapshot.partial.turn === openTurn.turn) {
+				const partialReplyStartedAt = snapshot.partial.replyStartedAt;
+				if (partialReplyStartedAt !== void 0) return partialReplyStartedAt;
+			}
+			const chat = snapshot.chat;
+			const step = (chat?.timeline.turns.get(openTurn.turn))?.steps.findLast((candidate) => candidate.end === void 0);
+			if (step === void 0) return void 0;
+			for (const node of [...chat?.nodes.values() ?? []].reverse()) {
+				if (node.kind !== "assistant-step") continue;
+				const data = node.data;
+				if (data.turn === openTurn.turn && data.step === step.step && typeof data.replyStartedAt === "number") return data.replyStartedAt;
+			}
+		}
+		/** Infer the two-phase model activity from the current client snapshot. */
+		function modelActivity(snapshot) {
+			const startedAt = awaitingModelStartedAt(snapshot);
+			if (startedAt === void 0) return void 0;
+			const replyStartedAt = modelReplyStartedAt(snapshot);
+			return replyStartedAt === void 0 ? {
+				activity: "requesting-model",
+				startedAt
+			} : {
+				activity: "model-reply",
+				startedAt: replyStartedAt
+			};
+		}
 		//#endregion
 		//#region lib/types/client/index.js
 		/** Browser controls for the Codex request settings mounted by the Host plugin. */
@@ -50,7 +89,8 @@ window.__ModuleLoader__.load({
 			contextSizeDescription: "Next request capacity in K tokens. The meter above is current.",
 			contextRestore: "Restore model default",
 			activityCompacting: "Compacting context...",
-			activityAwaitingModel: "Waiting for model response..."
+			activityRequestingModel: "Requesting model...",
+			activityModelReply: "Model replying..."
 		};
 		const zh = {
 			fast: "Fast",
@@ -62,7 +102,8 @@ window.__ModuleLoader__.load({
 			contextSizeDescription: "设置下次请求的上下文容量，单位为 K tokens；上方是当前请求。",
 			contextRestore: "恢复模型默认值",
 			activityCompacting: "正在压缩上下文...",
-			activityAwaitingModel: "正在等待模型响应..."
+			activityRequestingModel: "请求模型...",
+			activityModelReply: "模型回复..."
 		};
 		function FastModeButton({ sessionId, useSessions, useSettings, setSetting, t }) {
 			const agentPreset = useSessions((state) => state.byId[sessionId]?.agentPreset);
@@ -283,11 +324,12 @@ window.__ModuleLoader__.load({
 		function ActivityLine({ sessionId, useSession, useSessions, useProjection, t }) {
 			const agentPreset = useSessions((state) => state.byId[sessionId]?.agentPreset);
 			const projectedActivity = useProjection("codexActivity");
-			const fallbackStartedAt = useSession((snapshot) => awaitingModelStartedAt(snapshot));
-			const activity = projectedActivity !== void 0 ? projectedActivity : isCodexPresetId(agentPreset) && fallbackStartedAt !== void 0 ? {
-				activity: "awaiting-model",
-				startedAt: fallbackStartedAt
-			} : null;
+			const fallbackActivity = useSession(modelActivity, sameModelActivity);
+			const normalizedProjectedActivity = projectedActivity?.activity === "awaiting-model" ? {
+				activity: fallbackActivity?.activity ?? "requesting-model",
+				startedAt: fallbackActivity?.startedAt ?? projectedActivity.startedAt
+			} : projectedActivity;
+			const activity = projectedActivity !== void 0 ? normalizedProjectedActivity : isCodexPresetId(agentPreset) && fallbackActivity !== void 0 ? fallbackActivity : null;
 			const [now, setNow] = (0, react.useState)(() => Date.now());
 			(0, react.useEffect)(() => {
 				if (!isCodexPresetId(agentPreset) || activity === void 0 || activity === null) return void 0;
@@ -296,9 +338,13 @@ window.__ModuleLoader__.load({
 					setNow(Date.now());
 				}, 1e3);
 				return () => clearInterval(timer);
-			}, [agentPreset, activity?.startedAt]);
+			}, [
+				agentPreset,
+				activity?.activity,
+				activity?.startedAt
+			]);
 			if (!isCodexPresetId(agentPreset) || activity === void 0 || activity === null) return null;
-			const label = activity.activity === "awaiting-model" ? t("activityAwaitingModel") : t("activityCompacting");
+			const label = activity.activity === "compaction" ? t("activityCompacting") : activity.activity === "model-reply" ? t("activityModelReply") : t("activityRequestingModel");
 			return (0, react_jsx_runtime.jsxs)("span", {
 				id: "codex-activity",
 				role: "status",

package/lib/compaction.js

@@ -0,0 +1,45 @@
+import { AsyncLocalStorage } from "node:async_hooks";
+import BasicCompactionEngine from "@deepseek-ai/dsh-compaction-basic";
+//#region lib/types/compaction.js
+/** Codex compaction backend that honors the latest durable request capacity. */
+/** Overlay only the capacity lookup used by pressure compaction. */
+function withDurableContextCapacity(llm, activeSession) {
+	return new Proxy(llm, { get(target, property) {
+		if (property === "resolveModelInfo") return async (provider, model, signal) => {
+			const info = await target.resolveModelInfo(provider, model, signal);
+			const requestContext = activeSession.getStore()?.requestContext();
+			if (requestContext?.provider !== provider || requestContext.model !== model || requestContext.contextWindow === void 0) return info;
+			return {
+				...info,
+				context: {
+					...info.context,
+					contextWindow: requestContext.contextWindow
+				}
+			};
+		};
+		const value = Reflect.get(target, property, target);
+		return typeof value === "function" ? value.bind(target) : value;
+	} });
+}
+/**
+* Keep the upstream compaction policy, retention, overflow recovery, and
+* transaction implementation intact. The scoped LLM seam changes only the
+* context metadata returned during a pressure check.
+*/
+var CodexCompactionEngine = class extends BasicCompactionEngine {
+	static inject = BasicCompactionEngine.inject;
+	static Config = BasicCompactionEngine.Config;
+	activeSession;
+	constructor(ctx, config = {}) {
+		const activeSession = new AsyncLocalStorage();
+		const llm = ctx.get("llm");
+		if (llm === void 0) throw new Error("codex compaction requires ctx.llm");
+		super(ctx.extend({ llm: withDurableContextCapacity(llm, activeSession) }), config);
+		this.activeSession = activeSession;
+	}
+	compactIfNeeded(agent, trigger, signal) {
+		return this.activeSession.run(agent.session, () => super.compactIfNeeded(agent, trigger, signal));
+	}
+};
+//#endregion
+export { CodexCompactionEngine as default };

package/lib/index.js

@@ -5,6 +5,7 @@ import { randomBytes } from "node:crypto";
 import { resolve } from "node:path";
 import { credentialRef } from "@deepseek-ai/dsh-credentials";
 import { AsyncLocalStorage } from "node:async_hooks";
+import { isTokenDelta } from "@deepseek-ai/dsh-llm/message";
 import { z as z$1 } from "zod";
 //#region lib/types/context.js
 const CODEX_CONTEXT_MAX = 1e6;
@@ -13,9 +14,8 @@ const CODEX_CONTEXT_MAX = 1e6;
 /** Parser and line-oriented applicator for Codex's `apply_patch` language. */
 const APPLY_PATCH_DESCRIPTION = "Edits files using Codex patch syntax with Begin/End Patch markers and file update directives. In hunk lines, the first character is the operation marker; repeat a source-leading marker when the source line itself starts with one.";
 /** The grammar sent to providers that support OpenAI custom grammar tools. */
-const APPLY_PATCH_GRAMMAR = String.raw`start: begin_patch environment_id? hunk+ end_patch
+const APPLY_PATCH_GRAMMAR = String.raw`start: begin_patch hunk+ end_patch
 begin_patch: "*** Begin Patch" LF
-environment_id: "*** Environment ID: " filename LF
 end_patch: "*** End Patch" LF?
 hunk: add_hunk | delete_hunk | update_hunk
 add_hunk: "*** Add File: " filename LF add_line+
@@ -37,7 +37,6 @@ function isFileHeader(line) {
 	const trimmed = line.trim();
 	return trimmed.startsWith("*** Add File: ") || trimmed.startsWith("*** Delete File: ") || trimmed.startsWith("*** Update File: ");
 }
-const ENVIRONMENT_ID_MARKER = "*** Environment ID:";
 function unwrapHeredoc(input) {
 	const normalized = input.replaceAll("\r\n", "\n").trim();
 	const lines = normalized.split("\n");
@@ -66,11 +65,7 @@ function parsePatch(input) {
 	const files = [];
 	let index = 1;
 	const end = lines.length - 1;
-	if (lines[index]?.trim().startsWith(ENVIRONMENT_ID_MARKER) === true) {
-		if (lines[index].trim().slice(19).trim().length === 0) invalid("environment id cannot be empty");
-		index++;
-		if (lines[index]?.trim().startsWith(ENVIRONMENT_ID_MARKER) === true) invalid("environment id cannot be specified more than once");
-	}
+	if (lines[index]?.trim().startsWith("*** Environment ID:") === true) invalid("environment selection is unavailable in this single-environment DSH plugin");
 	while (index < end) {
 		const header = lines[index++]?.trim();
 		if (header === void 0) invalid("unexpected end of input");
@@ -218,6 +213,7 @@ function applyPatchHunks(original, hunks, label = "file") {
 /** Codex `exec_command` and `write_stdin` execution over dsh capability seams. */
 const STATES = /* @__PURE__ */ new WeakMap();
 const MIN_YIELD_TIME_MS = 250;
+const WINDOWS_INITIAL_EXEC_YIELD_TIME_MS = 1e4;
 const MIN_EMPTY_YIELD_TIME_MS = 5e3;
 const MAX_YIELD_TIME_MS = 3e4;
 const MAX_EMPTY_YIELD_TIME_MS = 3e5;
@@ -236,8 +232,9 @@ function stateFor(agent) {
 function positiveFinite(name, value) {
 	if (value !== void 0 && (!Number.isFinite(value) || value < 0)) throw new Error(`${name} must be a non-negative finite number`);
 }
-function normalizeWaitMs(value, fallback, empty = false) {
-	return Math.max(empty ? MIN_EMPTY_YIELD_TIME_MS : MIN_YIELD_TIME_MS, Math.min(empty ? MAX_EMPTY_YIELD_TIME_MS : MAX_YIELD_TIME_MS, Math.trunc(value ?? fallback)));
+function normalizeWaitMs(value, fallback, empty = false, initialExec = false) {
+	const lower = empty ? MIN_EMPTY_YIELD_TIME_MS : initialExec && process.platform === "win32" ? WINDOWS_INITIAL_EXEC_YIELD_TIME_MS : MIN_YIELD_TIME_MS;
+	return Math.max(lower, Math.min(empty ? MAX_EMPTY_YIELD_TIME_MS : MAX_YIELD_TIME_MS, Math.trunc(value ?? fallback)));
 }
 function outputLimit(maxOutputBytes, maxOutputTokens) {
 	const tokens = maxOutputTokens ?? DEFAULT_OUTPUT_TOKENS;
@@ -369,7 +366,7 @@ async function runExecCommand(ctx, args, exec, config) {
 			const sendRequest = {
 				text: commandFor(args),
 				submit: true,
-				waitMs: normalizeWaitMs(args.yield_time_ms, config.defaultYieldTimeMs),
+				waitMs: normalizeWaitMs(args.yield_time_ms, config.defaultYieldTimeMs, false, true),
 				signal: exec.signal
 			};
 			const result = await terminals.startSend(agent, spawned.sessionId, sendRequest).done;
@@ -432,7 +429,7 @@ async function runExecCommand(ctx, args, exec, config) {
 		signal: exec.signal
 	}));
 	try {
-		await waitForShell(process, normalizeWaitMs(args.yield_time_ms, config.defaultYieldTimeMs), exec.signal);
+		await waitForShell(process, normalizeWaitMs(args.yield_time_ms, config.defaultYieldTimeMs, false, true), exec.signal);
 		const output = readShellOutput(process.readOutput(), maxBytes);
 		if (process.status === "running") {
 			if (exec.agent === void 0) {
@@ -899,6 +896,29 @@ function compactText(body) {
 	const encrypted = text.join("\n\n").trim();
 	return encrypted.length === 0 ? "" : `${REMOTE_COMPACTION_OPEN}${encrypted}${REMOTE_COMPACTION_CLOSE}`;
 }
+function nonNegativeInteger(value) {
+	return typeof value === "number" && Number.isSafeInteger(value) && value >= 0 ? value : void 0;
+}
+/** Map the Responses compact usage object without inventing missing counts. */
+function compactUsage(body) {
+	if (typeof body.usage !== "object" || body.usage === null) return void 0;
+	const raw = body.usage;
+	const inputTokens = nonNegativeInteger(raw.input_tokens);
+	const outputTokens = nonNegativeInteger(raw.output_tokens);
+	if (inputTokens === void 0 || outputTokens === void 0) return void 0;
+	const details = typeof raw.input_tokens_details === "object" && raw.input_tokens_details !== null ? raw.input_tokens_details : void 0;
+	const cacheReadTokens = nonNegativeInteger(details?.cached_tokens);
+	const cacheWriteTokens = nonNegativeInteger(details?.cache_write_tokens);
+	const cached = cacheReadTokens ?? 0;
+	const writes = cacheWriteTokens ?? 0;
+	if (cached + writes > inputTokens) return void 0;
+	return {
+		inputTokens: inputTokens - cached - writes,
+		outputTokens,
+		...cacheReadTokens !== void 0 && cacheReadTokens > 0 ? { cacheReadTokens } : {},
+		...cacheWriteTokens !== void 0 && cacheWriteTokens > 0 ? { cacheWriteTokens } : {}
+	};
+}
 async function remoteCompact(ctx, options) {
 	const profile = profileOf(ctx, options.provider);
 	if (!supportsResponses(profile, options.provider) || profile?.baseURL === void 0) throw new Error("Codex remote compaction requires an OpenAI Responses provider with baseURL");
@@ -911,15 +931,20 @@ async function remoteCompact(ctx, options) {
 		...options.signal === void 0 ? {} : { signal: options.signal }
 	});
 	if (!response.ok) throw new Error(`remote compaction returned HTTP ${response.status}`);
-	const text = compactText(await response.json());
+	const body = await response.json();
+	const text = compactText(body);
 	if (text.length === 0) throw new Error("remote compaction returned no compaction text");
-	return text;
+	const usage = compactUsage(body);
+	return {
+		text,
+		...usage === void 0 ? {} : { usage }
+	};
 }
 /** Remote-first compaction waterfall with the existing DSH path as fallback. */
 function remoteCompactStream(ctx, options, next) {
 	return (async function* () {
 		try {
-			const text = await remoteCompact(ctx, options);
+			const result = await remoteCompact(ctx, options);
 			yield {
 				type: "block-start",
 				index: 0,
@@ -928,22 +953,19 @@ function remoteCompactStream(ctx, options, next) {
 			yield {
 				type: "text-delta",
 				index: 0,
-				text
+				text: result.text
 			};
 			yield {
 				type: "block-end",
 				index: 0,
 				block: {
 					type: "text",
-					text
+					text: result.text
 				}
 			};
-			yield {
+			if (result.usage !== void 0) yield {
 				type: "usage",
-				usage: {
-					inputTokens: 0,
-					outputTokens: 0
-				}
+				usage: result.usage
 			};
 			yield {
 				type: "finish",
@@ -959,13 +981,28 @@ function remoteCompactStream(ctx, options, next) {
 }
 //#endregion
 //#region lib/types/activity.js
-/** Host-side activity projection for the Codex compaction status line. */
+/** Host-side activity projection for the Codex status line. */
+/** Only a non-empty model delta starts the reply phase; usage and finish frames do not. */
+function startsModelReply(event) {
+	if (event.type === "assistant/message") return true;
+	if (event.type !== "assistant/chunk" || typeof event.data !== "object" || event.data === null) return false;
+	const chunk = event.data.chunk;
+	return chunk !== void 0 && isTokenDelta(chunk);
+}
 /** Persisted and wire shape of the Codex activity value. */
 const codexActivitySchema = z$1.union([
 	z$1.object({
 		activity: z$1.literal("compaction"),
 		startedAt: z$1.number().int().nonnegative()
 	}),
+	z$1.object({
+		activity: z$1.literal("requesting-model"),
+		startedAt: z$1.number().int().nonnegative()
+	}),
+	z$1.object({
+		activity: z$1.literal("model-reply"),
+		startedAt: z$1.number().int().nonnegative()
+	}),
 	z$1.object({
 		activity: z$1.literal("awaiting-model"),
 		startedAt: z$1.number().int().nonnegative()
@@ -979,11 +1016,15 @@ function applyCodexActivity(state, event) {
 		startedAt: event.time
 	};
 	if (event.type === "step/start" || event.type === "llm/retry-started") return {
-		activity: "awaiting-model",
+		activity: "requesting-model",
+		startedAt: event.time
+	};
+	if (startsModelReply(event) && (state?.activity === "requesting-model" || state?.activity === "awaiting-model")) return {
+		activity: "model-reply",
 		startedAt: event.time
 	};
 	if (event.type === "turn/start" || event.type === "turn/end" || event.type === "compaction/end") return null;
-	if (state?.activity === "awaiting-model" && (event.type === "llm/retry" || event.type === "tool/call" || event.type === "step/end")) return null;
+	if ((state?.activity === "requesting-model" || state?.activity === "model-reply" || state?.activity === "awaiting-model") && (event.type === "llm/retry" || event.type === "tool/call" || event.type === "step/end")) return null;
 	return state;
 }
 const codexActivityProjection = {
@@ -997,7 +1038,7 @@ const codexActivityProjection = {
 	},
 	schema: codexActivitySchema,
 	view: (state) => state,
-	stateVersion: 1
+	stateVersion: 2
 };
 /** Register the optional activity unit without requiring the projection seam in every assembly. */
 function registerCodexActivityProjection(ctx) {
@@ -1018,12 +1059,13 @@ const inject = [
 	"credentials",
 	"settings"
 ];
+const DEFAULT_MAX_OUTPUT_BYTES = 1048576;
 /** Runtime configuration schema for the Codex tool bridge. */
 const Config = z.object({
 	defaultYieldTimeMs: z.number().step(1).min(0).default(1e4),
 	pollYieldTimeMs: z.number().step(1).min(0).default(5e3),
 	writeYieldTimeMs: z.number().step(1).min(0).default(250),
-	maxOutputBytes: z.number().step(1).min(1).default(64e3),
+	maxOutputBytes: z.number().step(1).min(1).default(DEFAULT_MAX_OUTPUT_BYTES),
 	hostedWebSearch: z.boolean().default(true),
 	remoteCompact: z.boolean().default(true),
 	collaborationPrompt: z.boolean().default(false)
@@ -1566,7 +1608,7 @@ function apply(ctx, config = {}) {
 		defaultYieldTimeMs: config.defaultYieldTimeMs ?? 1e4,
 		pollYieldTimeMs: config.pollYieldTimeMs ?? 5e3,
 		writeYieldTimeMs: config.writeYieldTimeMs ?? 250,
-		maxOutputBytes: config.maxOutputBytes ?? 64e3,
+		maxOutputBytes: config.maxOutputBytes ?? DEFAULT_MAX_OUTPUT_BYTES,
 		hostedWebSearch: config.hostedWebSearch ?? true,
 		remoteCompact: config.remoteCompact ?? true,
 		collaborationPrompt: config.collaborationPrompt ?? false

package/lib/types/activity-types.d.ts

@@ -1,6 +1,6 @@
 /** Pure client/host-shared type for the Codex activity projection value. */
 export type CodexActivity = {
-    activity: 'compaction' | 'awaiting-model';
+    activity: 'compaction' | 'requesting-model' | 'model-reply' | 'awaiting-model';
     startedAt: number;
 };
 //# sourceMappingURL=activity-types.d.ts.map

package/lib/types/activity.d.ts

@@ -1,14 +1,15 @@
-/** Host-side activity projection for the Codex compaction status line. */
+/** Host-side activity projection for the Codex status line. */
 import type { Context } from '@deepseek-ai/cordis';
 import { z, type ZodType } from 'zod';
 import type { CodexActivity } from './activity-types.ts';
 export type { CodexActivity } from './activity-types.ts';
-/** The only long-running Codex activity currently surfaced to the client. */
+/** Long-running Codex activities currently surfaced to the client. */
 export type CodexActivityState = CodexActivity | null;
 /** Minimal event shape needed by this projection; no compaction package import is required. */
 export interface CodexActivityEvent {
     type: string;
     time: number;
+    data?: unknown;
 }
 /** Persisted and wire shape of the Codex activity value. */
 export declare const codexActivitySchema: ZodType<CodexActivityState>;
@@ -25,7 +26,7 @@ export declare const codexActivityProjection: {
     };
     readonly schema: z.ZodType<CodexActivityState, unknown, z.core.$ZodTypeInternals<CodexActivityState, unknown>>;
     readonly view: (state: CodexActivityState) => CodexActivityState;
-    readonly stateVersion: 1;
+    readonly stateVersion: 2;
 };
 /** Register the optional activity unit without requiring the projection seam in every assembly. */
 export declare function registerCodexActivityProjection(ctx: Context): void;

package/lib/types/activity.js

@@ -1,11 +1,31 @@
-/** Host-side activity projection for the Codex compaction status line. */
+/** Host-side activity projection for the Codex status line. */
+import { isTokenDelta } from '@deepseek-ai/dsh-llm/message';
 import { z } from 'zod';
+/** Only a non-empty model delta starts the reply phase; usage and finish frames do not. */
+function startsModelReply(event) {
+    if (event.type === 'assistant/message')
+        return true;
+    if (event.type !== 'assistant/chunk' || typeof event.data !== 'object' || event.data === null)
+        return false;
+    const chunk = event.data.chunk;
+    return chunk !== undefined && isTokenDelta(chunk);
+}
 /** Persisted and wire shape of the Codex activity value. */
 export const codexActivitySchema = z.union([
     z.object({
         activity: z.literal('compaction'),
         startedAt: z.number().int().nonnegative(),
     }),
+    z.object({
+        activity: z.literal('requesting-model'),
+        startedAt: z.number().int().nonnegative(),
+    }),
+    z.object({
+        activity: z.literal('model-reply'),
+        startedAt: z.number().int().nonnegative(),
+    }),
+    // Accepted for rolling upgrades: an older Codex host may still send this
+    // ambiguous state until its projection definition is refreshed.
     z.object({
         activity: z.literal('awaiting-model'),
         startedAt: z.number().int().nonnegative(),
@@ -17,14 +37,20 @@ export function applyCodexActivity(state, event) {
     if (event.type === 'compaction/start')
         return { activity: 'compaction', startedAt: event.time };
     if (event.type === 'step/start' || event.type === 'llm/retry-started') {
-        return { activity: 'awaiting-model', startedAt: event.time };
+        return { activity: 'requesting-model', startedAt: event.time };
+    }
+    if (startsModelReply(event)
+        && (state?.activity === 'requesting-model' || state?.activity === 'awaiting-model')) {
+        return { activity: 'model-reply', startedAt: event.time };
     }
     if (event.type === 'turn/start' || event.type === 'turn/end' || event.type === 'compaction/end')
         return null;
     // Keep the line through reasoning and streamed output. It is the only
     // visible indication that the model is still progressing; tool cards have
     // their own status, so a tool dispatch retires this ambient line.
-    if (state?.activity === 'awaiting-model'
+    if ((state?.activity === 'requesting-model'
+        || state?.activity === 'model-reply'
+        || state?.activity === 'awaiting-model')
         && (event.type === 'llm/retry'
             || event.type === 'tool/call'
             || event.type === 'step/end'))
@@ -44,7 +70,7 @@ export const codexActivityProjection = {
     // the `stateSchema`/`wire` split.
     schema: codexActivitySchema,
     view: (state) => state,
-    stateVersion: 1,
+    stateVersion: 2,
 };
 /** Register the optional activity unit without requiring the projection seam in every assembly. */
 export function registerCodexActivityProjection(ctx) {

package/lib/types/client/activity.d.ts

@@ -1,6 +1,19 @@
 /** Client-only inference used while the host activity projection frame is in flight. */
 import type { ConversationSnapshot } from '@deepseek-ai/dsh-client-runtime/client';
-export type ActivitySessionSnapshot = Pick<ConversationSnapshot, 'running' | 'partial' | 'runningCalls' | 'turnTimings'>;
-/** Infer the active model phase from the existing conversation snapshot. */
+export type ActivitySessionSnapshot = Pick<ConversationSnapshot, 'running' | 'partial' | 'runningCalls' | 'turnTimings'> & {
+    /** Optional for hosts that predate the assembled Chat snapshot. */
+    readonly chat?: ConversationSnapshot['chat'];
+};
+export type CodexModelActivity = {
+    activity: 'requesting-model' | 'model-reply';
+    startedAt: number;
+};
+/** Equality for selector hooks, whose default is reference identity. */
+export declare function sameModelActivity(left: CodexModelActivity | undefined, right: CodexModelActivity | undefined): boolean;
+/** Infer the request-phase clock from the existing conversation snapshot. */
 export declare function awaitingModelStartedAt(snapshot: ActivitySessionSnapshot): number | undefined;
+/** Infer the first non-empty model token timestamp from the current snapshot. */
+export declare function modelReplyStartedAt(snapshot: ActivitySessionSnapshot): number | undefined;
+/** Infer the two-phase model activity from the current client snapshot. */
+export declare function modelActivity(snapshot: ActivitySessionSnapshot): CodexModelActivity | undefined;
 //# sourceMappingURL=activity.d.ts.map

package/lib/types/client/activity.js

@@ -1,5 +1,9 @@
 /** Client-only inference used while the host activity projection frame is in flight. */
-/** Infer the active model phase from the existing conversation snapshot. */
+/** Equality for selector hooks, whose default is reference identity. */
+export function sameModelActivity(left, right) {
+    return left?.activity === right?.activity && left?.startedAt === right?.startedAt;
+}
+/** Infer the request-phase clock from the existing conversation snapshot. */
 export function awaitingModelStartedAt(snapshot) {
     if (!snapshot.running || snapshot.runningCalls.length > 0)
         return undefined;
@@ -10,4 +14,48 @@ export function awaitingModelStartedAt(snapshot) {
     }
     return latest;
 }
+/** Infer the first non-empty model token timestamp from the current snapshot. */
+export function modelReplyStartedAt(snapshot) {
+    if (!snapshot.running || snapshot.runningCalls.length > 0)
+        return undefined;
+    let openTurn;
+    for (const [turn, timing] of snapshot.turnTimings) {
+        if (timing.endTime === undefined)
+            openTurn = { turn, startTime: timing.startTime };
+    }
+    if (openTurn === undefined)
+        return undefined;
+    if (snapshot.partial !== null && snapshot.partial.turn === openTurn.turn) {
+        const partialReplyStartedAt = snapshot.partial.replyStartedAt;
+        if (partialReplyStartedAt !== undefined)
+            return partialReplyStartedAt;
+    }
+    // A non-streaming adapter has no partial row. The assembled Chat view still
+    // carries the reply boundary, but only accept it for the currently open step
+    // so an earlier tool step in the same turn cannot mask a new request.
+    const chat = snapshot.chat;
+    const turn = chat?.timeline.turns.get(openTurn.turn);
+    const step = turn?.steps.findLast(candidate => candidate.end === undefined);
+    if (step === undefined)
+        return undefined;
+    for (const node of [...(chat?.nodes.values() ?? [])].reverse()) {
+        if (node.kind !== 'assistant-step')
+            continue;
+        const data = node.data;
+        if (data.turn === openTurn.turn && data.step === step.step && typeof data.replyStartedAt === 'number') {
+            return data.replyStartedAt;
+        }
+    }
+    return undefined;
+}
+/** Infer the two-phase model activity from the current client snapshot. */
+export function modelActivity(snapshot) {
+    const startedAt = awaitingModelStartedAt(snapshot);
+    if (startedAt === undefined)
+        return undefined;
+    const replyStartedAt = modelReplyStartedAt(snapshot);
+    return replyStartedAt === undefined
+        ? { activity: 'requesting-model', startedAt }
+        : { activity: 'model-reply', startedAt: replyStartedAt };
+}
 //# sourceMappingURL=activity.js.map

package/lib/types/client/index.d.ts

@@ -11,7 +11,8 @@ declare const en: {
     readonly contextSizeDescription: "Next request capacity in K tokens. The meter above is current.";
     readonly contextRestore: "Restore model default";
     readonly activityCompacting: "Compacting context...";
-    readonly activityAwaitingModel: "Waiting for model response...";
+    readonly activityRequestingModel: "Requesting model...";
+    readonly activityModelReply: "Model replying...";
 };
 type CodexKey = keyof typeof en;
 declare module '@deepseek-ai/dsh-client-ui-slots' {

package/lib/types/client/index.js

@@ -2,7 +2,7 @@ import { jsx as _jsx, jsxs as _jsxs } from "react/jsx-runtime";
 /** Browser controls for the Codex request settings mounted by the Host plugin. */
 import { useEffect, useState } from 'react';
 import { CODEX_CONTEXT_MAX, CODEX_CONTEXT_UNIT, isCodexPresetId } from "../context.js";
-import { awaitingModelStartedAt } from "./activity.js";
+import { modelActivity, sameModelActivity } from "./activity.js";
 const NS = 'codex';
 const SETTINGS_NAMESPACE = 'codex';
 const MODEL_SETTINGS_SLOT = 'conversation.input.model.settings';
@@ -24,7 +24,8 @@ const en = {
     contextSizeDescription: 'Next request capacity in K tokens. The meter above is current.',
     contextRestore: 'Restore model default',
     activityCompacting: 'Compacting context...',
-    activityAwaitingModel: 'Waiting for model response...',
+    activityRequestingModel: 'Requesting model...',
+    activityModelReply: 'Model replying...',
 };
 const zh = {
     fast: 'Fast',
@@ -36,7 +37,8 @@ const zh = {
     contextSizeDescription: '设置下次请求的上下文容量，单位为 K tokens；上方是当前请求。',
     contextRestore: '恢复模型默认值',
     activityCompacting: '正在压缩上下文...',
-    activityAwaitingModel: '正在等待模型响应...',
+    activityRequestingModel: '请求模型...',
+    activityModelReply: '模型回复...',
 };
 function FastModeButton({ sessionId, useSessions, useSettings, setSetting, t }) {
     const agentPreset = useSessions(state => state.byId[sessionId]?.agentPreset);
@@ -143,14 +145,20 @@ export function ActivityLine({ sessionId, useSession, useSessions, useProjection
     // Keep the projection seam optional: this package must still load when the
     // host has no session-projection registry or has not carried this key.
     const projectedActivity = useProjection('codexActivity');
-    // The projection is the authoritative clock. The session snapshot fallback
-    // only covers the short wire gap before the next projection frame arrives;
-    // running tool calls and streamed output intentionally remain silent.
-    const fallbackStartedAt = useSession((snapshot) => awaitingModelStartedAt(snapshot));
+    // The projection is the authoritative phase and clock. The session snapshot
+    // fallback only covers the short wire gap before the next projection frame;
+    // tool calls intentionally remain silent because their cards own the status.
+    const fallbackActivity = useSession(modelActivity, sameModelActivity);
     // `undefined` means the key has not arrived yet; `null` is an authoritative
     // clear and must not be resurrected by the local snapshot fallback.
-    const activity = projectedActivity !== undefined ? projectedActivity : (isCodexPresetId(agentPreset) && fallbackStartedAt !== undefined
-        ? { activity: 'awaiting-model', startedAt: fallbackStartedAt }
+    const normalizedProjectedActivity = projectedActivity?.activity === 'awaiting-model'
+        ? {
+            activity: fallbackActivity?.activity ?? 'requesting-model',
+            startedAt: fallbackActivity?.startedAt ?? projectedActivity.startedAt,
+        }
+        : projectedActivity;
+    const activity = projectedActivity !== undefined ? normalizedProjectedActivity : (isCodexPresetId(agentPreset) && fallbackActivity !== undefined
+        ? fallbackActivity
         : null);
     const [now, setNow] = useState(() => Date.now());
     useEffect(() => {
@@ -159,12 +167,14 @@ export function ActivityLine({ sessionId, useSession, useSessions, useProjection
         setNow(Date.now());
         const timer = setInterval(() => { setNow(Date.now()); }, 1_000);
         return () => clearInterval(timer);
-    }, [agentPreset, activity?.startedAt]);
+    }, [agentPreset, activity?.activity, activity?.startedAt]);
     if (!isCodexPresetId(agentPreset) || activity === undefined || activity === null)
         return null;
-    const label = activity.activity === 'awaiting-model'
-        ? t('activityAwaitingModel')
-        : t('activityCompacting');
+    const label = activity.activity === 'compaction'
+        ? t('activityCompacting')
+        : activity.activity === 'model-reply'
+            ? t('activityModelReply')
+            : t('activityRequestingModel');
     return (_jsxs("span", { id: "codex-activity", role: "status", "aria-live": "polite", style: {
             display: 'inline-flex',
             alignItems: 'center',

package/lib/types/compaction.d.ts

@@ -0,0 +1,18 @@
+/** Codex compaction backend that honors the latest durable request capacity. */
+import { Context } from '@deepseek-ai/cordis';
+import BasicCompactionEngine from '@deepseek-ai/dsh-compaction-basic';
+import type { CompactionResult, CompactionTrigger } from '@deepseek-ai/dsh-compaction';
+import type { Agent } from '@deepseek-ai/dsh-agent';
+/**
+ * Keep the upstream compaction policy, retention, overflow recovery, and
+ * transaction implementation intact. The scoped LLM seam changes only the
+ * context metadata returned during a pressure check.
+ */
+export default class CodexCompactionEngine extends BasicCompactionEngine {
+    static inject: string[];
+    static Config: import("@deepseek-ai/schemastery").default<import("@deepseek-ai/dsh-compaction-basic").BasicCompactionConfig>;
+    private readonly activeSession;
+    constructor(ctx: Context, config?: ConstructorParameters<typeof BasicCompactionEngine>[1]);
+    compactIfNeeded(agent: Agent, trigger: CompactionTrigger, signal: AbortSignal): Promise<CompactionResult | null>;
+}
+//# sourceMappingURL=compaction.d.ts.map

package/lib/types/compaction.js

@@ -0,0 +1,48 @@
+/** Codex compaction backend that honors the latest durable request capacity. */
+import { AsyncLocalStorage } from 'node:async_hooks';
+import BasicCompactionEngine from '@deepseek-ai/dsh-compaction-basic';
+/** Overlay only the capacity lookup used by pressure compaction. */
+function withDurableContextCapacity(llm, activeSession) {
+    return new Proxy(llm, {
+        get(target, property) {
+            if (property === 'resolveModelInfo') {
+                return async (provider, model, signal) => {
+                    const info = await target.resolveModelInfo(provider, model, signal);
+                    const requestContext = activeSession.getStore()?.requestContext();
+                    if (requestContext?.provider !== provider
+                        || requestContext.model !== model
+                        || requestContext.contextWindow === undefined)
+                        return info;
+                    return {
+                        ...info,
+                        context: { ...info.context, contextWindow: requestContext.contextWindow },
+                    };
+                };
+            }
+            const value = Reflect.get(target, property, target);
+            return typeof value === 'function' ? value.bind(target) : value;
+        },
+    });
+}
+/**
+ * Keep the upstream compaction policy, retention, overflow recovery, and
+ * transaction implementation intact. The scoped LLM seam changes only the
+ * context metadata returned during a pressure check.
+ */
+export default class CodexCompactionEngine extends BasicCompactionEngine {
+    static inject = BasicCompactionEngine.inject;
+    static Config = BasicCompactionEngine.Config;
+    activeSession;
+    constructor(ctx, config = {}) {
+        const activeSession = new AsyncLocalStorage();
+        const llm = ctx.get('llm');
+        if (llm === undefined)
+            throw new Error('codex compaction requires ctx.llm');
+        super(ctx.extend({ llm: withDurableContextCapacity(llm, activeSession) }), config);
+        this.activeSession = activeSession;
+    }
+    compactIfNeeded(agent, trigger, signal) {
+        return this.activeSession.run(agent.session, () => super.compactIfNeeded(agent, trigger, signal));
+    }
+}
+//# sourceMappingURL=compaction.js.map

package/lib/types/exec.d.ts

@@ -29,7 +29,7 @@ export interface ExecResult {
     exit_code?: number;
     original_token_count?: number;
 }
-export declare function normalizeWaitMs(value: number | undefined, fallback: number, empty?: boolean): number;
+export declare function normalizeWaitMs(value: number | undefined, fallback: number, empty?: boolean, initialExec?: boolean): number;
 /** Render the ordinary Responses tool result text used by Codex's unified exec tools. */
 export declare function renderExecResult(result: ExecResult): string;
 /** Execute one Codex command through the configured pipe or PTY capability. */

package/lib/types/exec.js

@@ -3,6 +3,7 @@ import { randomBytes } from 'node:crypto';
 import { resolve as resolvePath } from 'node:path';
 const STATES = new WeakMap();
 const MIN_YIELD_TIME_MS = 250;
+const WINDOWS_INITIAL_EXEC_YIELD_TIME_MS = 10_000;
 const MIN_EMPTY_YIELD_TIME_MS = 5_000;
 const MAX_YIELD_TIME_MS = 30_000;
 const MAX_EMPTY_YIELD_TIME_MS = 300_000;
@@ -21,8 +22,12 @@ function positiveFinite(name, value) {
         throw new Error(`${name} must be a non-negative finite number`);
     }
 }
-export function normalizeWaitMs(value, fallback, empty = false) {
-    const lower = empty ? MIN_EMPTY_YIELD_TIME_MS : MIN_YIELD_TIME_MS;
+export function normalizeWaitMs(value, fallback, empty = false, initialExec = false) {
+    const lower = empty
+        ? MIN_EMPTY_YIELD_TIME_MS
+        : initialExec && process.platform === 'win32'
+            ? WINDOWS_INITIAL_EXEC_YIELD_TIME_MS
+            : MIN_YIELD_TIME_MS;
     const upper = empty ? MAX_EMPTY_YIELD_TIME_MS : MAX_YIELD_TIME_MS;
     return Math.max(lower, Math.min(upper, Math.trunc(value ?? fallback)));
 }
@@ -178,7 +183,7 @@ export async function runExecCommand(ctx, args, exec, config) {
             const sendRequest = {
                 text: commandFor(args),
                 submit: true,
-                waitMs: normalizeWaitMs(args.yield_time_ms, config.defaultYieldTimeMs),
+                waitMs: normalizeWaitMs(args.yield_time_ms, config.defaultYieldTimeMs, false, true),
                 signal: exec.signal,
             };
             const operation = terminals.startSend(agent, spawned.sessionId, sendRequest);
@@ -250,7 +255,7 @@ export async function runExecCommand(ctx, args, exec, config) {
         signal: exec.signal,
     }));
     try {
-        await waitForShell(process, normalizeWaitMs(args.yield_time_ms, config.defaultYieldTimeMs), exec.signal);
+        await waitForShell(process, normalizeWaitMs(args.yield_time_ms, config.defaultYieldTimeMs, false, true), exec.signal);
         const output = readShellOutput(process.readOutput(), maxBytes);
         if (process.status === 'running') {
             if (exec.agent === undefined) {

package/lib/types/index.js

@@ -9,12 +9,13 @@ import { hostedWebSearchStream, installHostedWebSearch, remoteCompactStream } fr
 import { registerCodexActivityProjection } from "./activity.js";
 export const name = 'codex';
 export const inject = ['tools', 'systemPrompt', 'shell', 'fs', 'llm', 'credentials', 'settings'];
+const DEFAULT_MAX_OUTPUT_BYTES = 1024 * 1024;
 /** Runtime configuration schema for the Codex tool bridge. */
 export const Config = z.object({
     defaultYieldTimeMs: z.number().step(1).min(0).default(10_000),
     pollYieldTimeMs: z.number().step(1).min(0).default(5_000),
     writeYieldTimeMs: z.number().step(1).min(0).default(250),
-    maxOutputBytes: z.number().step(1).min(1).default(64_000),
+    maxOutputBytes: z.number().step(1).min(1).default(DEFAULT_MAX_OUTPUT_BYTES),
     hostedWebSearch: z.boolean().default(true),
     remoteCompact: z.boolean().default(true),
     collaborationPrompt: z.boolean().default(false),
@@ -451,7 +452,7 @@ export function apply(ctx, config = {}) {
         defaultYieldTimeMs: config.defaultYieldTimeMs ?? 10_000,
         pollYieldTimeMs: config.pollYieldTimeMs ?? 5_000,
         writeYieldTimeMs: config.writeYieldTimeMs ?? 250,
-        maxOutputBytes: config.maxOutputBytes ?? 64_000,
+        maxOutputBytes: config.maxOutputBytes ?? DEFAULT_MAX_OUTPUT_BYTES,
         hostedWebSearch: config.hostedWebSearch ?? true,
         remoteCompact: config.remoteCompact ?? true,
         collaborationPrompt: config.collaborationPrompt ?? false,

package/lib/types/patch.js

@@ -1,9 +1,8 @@
 /** Parser and line-oriented applicator for Codex's `apply_patch` language. */
 export const APPLY_PATCH_DESCRIPTION = 'Edits files using Codex patch syntax with Begin/End Patch markers and file update directives. In hunk lines, the first character is the operation marker; repeat a source-leading marker when the source line itself starts with one.';
 /** The grammar sent to providers that support OpenAI custom grammar tools. */
-export const APPLY_PATCH_GRAMMAR = String.raw `start: begin_patch environment_id? hunk+ end_patch
+export const APPLY_PATCH_GRAMMAR = String.raw `start: begin_patch hunk+ end_patch
 begin_patch: "*** Begin Patch" LF
-environment_id: "*** Environment ID: " filename LF
 end_patch: "*** End Patch" LF?
 hunk: add_hunk | delete_hunk | update_hunk
 add_hunk: "*** Add File: " filename LF add_line+
@@ -27,7 +26,6 @@ function isFileHeader(line) {
         || trimmed.startsWith('*** Delete File: ')
         || trimmed.startsWith('*** Update File: ');
 }
-const ENVIRONMENT_ID_MARKER = '*** Environment ID:';
 function unwrapHeredoc(input) {
     const normalized = input.replaceAll('\r\n', '\n').trim();
     const lines = normalized.split('\n');
@@ -63,14 +61,8 @@ export function parsePatch(input) {
     const files = [];
     let index = 1;
     const end = lines.length - 1;
-    if (lines[index]?.trim().startsWith(ENVIRONMENT_ID_MARKER) === true) {
-        const environmentId = lines[index].trim().slice(ENVIRONMENT_ID_MARKER.length).trim();
-        if (environmentId.length === 0)
-            invalid('environment id cannot be empty');
-        index++;
-        if (lines[index]?.trim().startsWith(ENVIRONMENT_ID_MARKER) === true) {
-            invalid('environment id cannot be specified more than once');
-        }
+    if (lines[index]?.trim().startsWith('*** Environment ID:') === true) {
+        invalid('environment selection is unavailable in this single-environment DSH plugin');
     }
     while (index < end) {
         const header = lines[index++]?.trim();

package/lib/types/remote.js

@@ -437,6 +437,36 @@ function compactText(body) {
     const encrypted = text.join('\n\n').trim();
     return encrypted.length === 0 ? '' : `${REMOTE_COMPACTION_OPEN}${encrypted}${REMOTE_COMPACTION_CLOSE}`;
 }
+function nonNegativeInteger(value) {
+    return typeof value === 'number' && Number.isSafeInteger(value) && value >= 0 ? value : undefined;
+}
+/** Map the Responses compact usage object without inventing missing counts. */
+function compactUsage(body) {
+    if (typeof body.usage !== 'object' || body.usage === null)
+        return undefined;
+    const raw = body.usage;
+    const inputTokens = nonNegativeInteger(raw.input_tokens);
+    const outputTokens = nonNegativeInteger(raw.output_tokens);
+    if (inputTokens === undefined || outputTokens === undefined)
+        return undefined;
+    const details = typeof raw.input_tokens_details === 'object' && raw.input_tokens_details !== null
+        ? raw.input_tokens_details
+        : undefined;
+    const cacheReadTokens = nonNegativeInteger(details?.cached_tokens);
+    const cacheWriteTokens = nonNegativeInteger(details?.cache_write_tokens);
+    const cached = cacheReadTokens ?? 0;
+    const writes = cacheWriteTokens ?? 0;
+    // Responses input_tokens includes the cache buckets. Reject an inconsistent
+    // response instead of clamping it and reporting numbers that no longer sum.
+    if (cached + writes > inputTokens)
+        return undefined;
+    return {
+        inputTokens: inputTokens - cached - writes,
+        outputTokens,
+        ...cacheReadTokens !== undefined && cacheReadTokens > 0 ? { cacheReadTokens } : {},
+        ...cacheWriteTokens !== undefined && cacheWriteTokens > 0 ? { cacheWriteTokens } : {},
+    };
+}
 async function remoteCompact(ctx, options) {
     const profile = profileOf(ctx, options.provider);
     if (!supportsResponses(profile, options.provider) || profile?.baseURL === undefined) {
@@ -457,17 +487,19 @@ async function remoteCompact(ctx, options) {
     const text = compactText(body);
     if (text.length === 0)
         throw new Error('remote compaction returned no compaction text');
-    return text;
+    const usage = compactUsage(body);
+    return { text, ...usage === undefined ? {} : { usage } };
 }
 /** Remote-first compaction waterfall with the existing DSH path as fallback. */
 export function remoteCompactStream(ctx, options, next) {
     return (async function* () {
         try {
-            const text = await remoteCompact(ctx, options);
+            const result = await remoteCompact(ctx, options);
             yield { type: 'block-start', index: 0, blockType: 'text' };
-            yield { type: 'text-delta', index: 0, text };
-            yield { type: 'block-end', index: 0, block: { type: 'text', text } };
-            yield { type: 'usage', usage: { inputTokens: 0, outputTokens: 0 } };
+            yield { type: 'text-delta', index: 0, text: result.text };
+            yield { type: 'block-end', index: 0, block: { type: 'text', text: result.text } };
+            if (result.usage !== undefined)
+                yield { type: 'usage', usage: result.usage };
             yield { type: 'finish', reason: { kind: 'stop' } };
         }
         catch (error) {

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@shuind/dsh-codex-harness",
-  "version": "0.1.24",
+  "version": "0.1.26",
   "description": "A Codex harness with a streamlined system prompt for GPT models that do not fit DSH's native interface, while remaining compatible with the DSH plugin ecosystem.",
   "type": "module",
   "main": "lib/index.js",
@@ -18,6 +18,10 @@
       "types": "./lib/types/installer.d.ts",
       "default": "./lib/installer.js"
     },
+    "./compaction": {
+      "types": "./lib/types/compaction.d.ts",
+      "default": "./lib/compaction.js"
+    },
     "./client": {
       "types": "./lib/types/client/index.d.ts",
       "default": "./lib/client.js"
@@ -30,6 +34,7 @@
     "lib/index.js",
     "lib/invariant.js",
     "lib/installer.js",
+    "lib/compaction.js",
     "lib/types/**/*.js",
     "lib/types/**/*.d.ts",
     "lib/client.js",
@@ -87,6 +92,9 @@
     "@deepseek-ai/dsh-client-ui-conversation": ">=0.1.0-rc.8",
     "@deepseek-ai/dsh-client-ui-settings": ">=0.1.0-rc.8",
     "@deepseek-ai/dsh-client-ui-slots": ">=0.1.0-rc.8",
+    "@deepseek-ai/dsh-compaction": ">=0.1.1-rc.2",
+    "@deepseek-ai/dsh-compaction-basic": ">=0.1.1-rc.2",
+    "@deepseek-ai/dsh-compaction-tool-result-pruner": ">=0.1.1-rc.2",
     "@deepseek-ai/dsh-credentials": ">=0.1.0-rc.8",
     "@deepseek-ai/dsh-fs": ">=0.0.1-rc.1",
     "@deepseek-ai/dsh-invariants": ">=0.0.1-rc.1",
@@ -95,6 +103,7 @@
     "@deepseek-ai/dsh-settings": ">=0.1.0-rc.8",
     "@deepseek-ai/dsh-sandbox": ">=0.0.1-rc.1",
     "@deepseek-ai/dsh-sandbox-policy": ">=0.0.1-rc.1",
+    "@deepseek-ai/dsh-session": ">=0.1.1-rc.2",
     "@deepseek-ai/dsh-shell": ">=0.0.1-rc.5",
     "@deepseek-ai/dsh-shell-env": ">=0.0.1-rc.3",
     "@deepseek-ai/dsh-system-prompt": ">=0.0.1-rc.1",
@@ -102,6 +111,7 @@
     "@deepseek-ai/dsh-tool-todo": ">=0.0.1-rc.1",
     "@deepseek-ai/dsh-terminal": ">=0.0.1-rc.3",
     "@deepseek-ai/dsh-tool-jobs": ">=0.1.0-rc.8",
+    "@deepseek-ai/dsh-token-meter": ">=0.1.1-rc.2",
     "@deepseek-ai/dsh-tools": ">=0.0.1-rc.1"
   },
   "peerDependenciesMeta": {
@@ -126,6 +136,15 @@
     "@deepseek-ai/dsh-client-ui-slots": {
       "optional": true
     },
+    "@deepseek-ai/dsh-compaction": {
+      "optional": true
+    },
+    "@deepseek-ai/dsh-compaction-basic": {
+      "optional": true
+    },
+    "@deepseek-ai/dsh-compaction-tool-result-pruner": {
+      "optional": true
+    },
     "@deepseek-ai/dsh-credentials": {
       "optional": true
     },
@@ -150,6 +169,9 @@
     "@deepseek-ai/dsh-sandbox-policy": {
       "optional": true
     },
+    "@deepseek-ai/dsh-session": {
+      "optional": true
+    },
     "@deepseek-ai/dsh-shell": {
       "optional": true
     },
@@ -171,6 +193,9 @@
     "@deepseek-ai/dsh-tool-jobs": {
       "optional": true
     },
+    "@deepseek-ai/dsh-token-meter": {
+      "optional": true
+    },
     "@deepseek-ai/dsh-tools": {
       "optional": true
     }
@@ -183,6 +208,9 @@
     "@deepseek-ai/dsh-client-ui-conversation": "0.1.0-rc.8",
     "@deepseek-ai/dsh-client-ui-settings": "0.1.0-rc.8",
     "@deepseek-ai/dsh-client-ui-slots": "0.1.0-rc.8",
+    "@deepseek-ai/dsh-compaction": "0.1.1-rc.2",
+    "@deepseek-ai/dsh-compaction-basic": "0.1.1-rc.2",
+    "@deepseek-ai/dsh-compaction-tool-result-pruner": "0.1.1-rc.2",
     "@deepseek-ai/dsh-credentials": "0.1.0-rc.8",
     "@deepseek-ai/dsh-fs": "0.1.0-rc.6",
     "@deepseek-ai/dsh-invariants": "0.1.0-rc.6",
@@ -191,6 +219,7 @@
     "@deepseek-ai/dsh-settings": "0.1.0-rc.8",
     "@deepseek-ai/dsh-sandbox": "0.1.0-rc.6",
     "@deepseek-ai/dsh-sandbox-policy": "0.1.0-rc.6",
+    "@deepseek-ai/dsh-session": "0.1.1-rc.2",
     "@deepseek-ai/dsh-shell": "0.1.0-rc.6",
     "@deepseek-ai/dsh-shell-env": "0.1.0-rc.6",
     "@deepseek-ai/dsh-system-prompt": "0.1.0-rc.6",
@@ -198,6 +227,7 @@
     "@deepseek-ai/dsh-tool-todo": "0.1.0-rc.6",
     "@deepseek-ai/dsh-terminal": "0.1.0-rc.6",
     "@deepseek-ai/dsh-tool-jobs": "0.1.0-rc.8",
+    "@deepseek-ai/dsh-token-meter": "0.1.1-rc.2",
     "@deepseek-ai/dsh-tools": "0.1.0-rc.6",
     "@types/node": "^22.20.0",
     "@types/react": "~18.3.1",

package/presets/codex/agent.cordis.yml

@@ -70,7 +70,7 @@
     toolResultPruner: true
   config:
     - id: compaction-basic
-      name: '@deepseek-ai/dsh-compaction-basic'
+      name: '@shuind/dsh-codex-harness/compaction'
 
     - id: command-compact
       name: '@deepseek-ai/dsh-command-compact'

package/presets/codex-collaboration/agent.cordis.yml

@@ -70,7 +70,7 @@
     toolResultPruner: true
   config:
     - id: compaction-basic
-      name: '@deepseek-ai/dsh-compaction-basic'
+      name: '@shuind/dsh-codex-harness/compaction'
 
     - id: command-compact
       name: '@deepseek-ai/dsh-command-compact'