reshuibuduo · Sep 5, 2026 · Sep 5, 2026
diff --git a/.gitattributes b/.gitattributes
@@ -1,2 +1,4 @@
 * text=auto eol=lf
 *.tgz binary
+*.pt binary
+runtime/** linguist-vendored -whitespace
diff --git a/.github/workflows/release.yml b/.github/workflows/release.yml
@@ -43,4 +43,5 @@ jobs:
             release/dsh-tmcra-memory-*.tgz \
             release/SHA256SUMS \
             --verify-tag \
-            --generate-notes
+            --generate-notes \
+            --prerelease
diff --git a/CHANGELOG.md b/CHANGELOG.md
@@ -1,5 +1,13 @@
 # Changelog
 
+## 1.0.0-rc.1 - 2026-09-06
+
+- Add the redesigned local memory workspace, provider configuration, knowledge/graph views, task continuity, session modes, and recall budgets.
+- Support confirmed targeted feedback and prevent disabled or correction-discussion turns from later queue replay.
+- Add three pinned embedding/reranker profiles and Windows runtime preview controls; explicit local identity bypasses cloud login and disables cloud model task execution.
+- Ship the verified backend with automatic private Python setup and local identity discovery. `local-install` opens setup without a TMCRA account or server; stale cloud requests are blocked after local selection.
+- Ship all generated runtime chunks and workspace assets. Full-local compiler, organizer, and restart acceptance remains incomplete.
+
 ## 0.1.7 - 2026-09-04
 
 - Activate the configured Writer and background-organizer providers through the authenticated TMCRA user-provider task protocol.

diff --git a/README.md b/README.md
@@ -20,6 +20,8 @@ Cross-app continuity applies to memories that have reached TMCRA and requires bo
 
 ## What it does
 
+Development branch adds a local memory control panel: `dsh-tmcra-memory memory --scope EXACT_PROJECT_SCOPE --session EXACT_SESSION_ID`. It offers tasks, source corrections, ignore/restore, normal/recall-only/off modes, and a 12000-character default recall budget. Disabled periods are never automatically backfilled; already submitted remote work is retained. Effective corrections require the corresponding backend update. New features are not yet published.
+
 - Continues one project across TMCRA-connected tools such as Codex and DeepSeek Harness.
 - Recalls user-global and current-project memory before the first model request of each turn.
 - Continues project work across separate Harness conversations.
@@ -61,17 +63,17 @@ During the initial release period, TMCRA accounts are free to use with no memory
 ## Install the preview tarball
 
 ```bash
-dsh plugin --profile web add https://github.com/reshuibuduo/tmcra-plugin-deepseek-harness/releases/download/v0.1.7/dsh-tmcra-memory-0.1.7.tgz
+dsh plugin --profile web add https://github.com/reshuibuduo/tmcra-plugin-deepseek-harness/releases/download/v1.0.0-rc.1/dsh-tmcra-memory-1.0.0-rc.1.tgz
 dsh plugin --profile web exec dsh-tmcra-memory login
 dsh --profile web --dump-config
 dsh web
 ```
 
 Harness serves its Web UI at `http://127.0.0.1:3080` by default.
 
-The package contributes `cordis.patch.yml`, so the install command activates the plugin in the selected profile. Prefer the prebuilt release tarball. A locally downloaded copy can also be installed with `dsh plugin --profile web add ./dsh-tmcra-memory-0.1.7.tgz`. Installing from a Git source may require an explicit `pnpm` build-script allowance.
+The package contributes `cordis.patch.yml`, so the install command activates the plugin in the selected profile. Prefer the prebuilt release tarball. A locally downloaded copy can also be installed with `dsh plugin --profile web add ./dsh-tmcra-memory-1.0.0-rc.1.tgz`. Installing from a Git source may require an explicit `pnpm` build-script allowance.
 
-DSH `0.1.1-rc.2` on Windows can still re-anchor a local package path incorrectly when the path contains spaces or non-ASCII characters. Copy the tarball to a short ASCII path first, such as `D:\\dsh-packages\\dsh-tmcra-memory-0.1.7.tgz`. The release URL above is not affected.
+DSH `0.1.1-rc.2` on Windows can still re-anchor a local package path incorrectly when the path contains spaces or non-ASCII characters. Copy the tarball to a short ASCII path first, such as `D:\\dsh-packages\\dsh-tmcra-memory-1.0.0-rc.1.tgz`. The release URL above is not affected.
 
 ## Connect your TMCRA account
 
@@ -94,6 +96,14 @@ Harness keeps credential values out of settings and model requests. A model-oper
 
 ## Configure local Writer and organizer providers
 
+### Full-local Windows preview
+
+The package includes the full backend and E5/MiniLM, BGE-M3/BGE reranker, and Qwen3 embedding/reranker profiles. After installing the DSH plugin, run `dsh-tmcra-memory local-install` and choose a profile on the installation page. This entry requires no TMCRA login. Python, dependencies and verified model files are prepared automatically, with a private local identity discovered by DSH after restart. Windows x64 is supported; preinstalled Python and TMCRA servers/accounts are unnecessary. First-time downloads require internet. Existing cloud credentials are retained; stale cloud memory connections and background cloud model requests are blocked after local selection. Failed setup retains local selection. Clear an advanced `TMCRA_CONFIG_FILE` override before automatic setup.
+
+CPU ingest/raw recall passed a synthetic test. Complex compilation timed out; organizer and full-service restart acceptance remains pending. Lightweight startup requires approximately 6.3 GiB free memory. See the [acceptance record](https://github.com/reshuibuduo/tmcra/blob/v1.0.0-rc.1/docs/LOCAL_DEPLOYMENT_PREVIEW.zh-CN.md). A host configured with a cloud main model can still send recalled evidence to its own provider.
+
+### Separate provider APIs
+
 Open the local setup page after installing the plugin:
 
 ```bash

diff --git a/README.zh-CN.md b/README.zh-CN.md
@@ -20,6 +20,14 @@ Agent 回答前，TMCRA 会找回与当前问题有关的用户要求、决策
 
 ## 已实现能力
 
+1.0.0-rc.1 新增任务接续、会话模式、完整证据块预算和本机记忆控制台。使用 `dsh-tmcra-memory memory --scope <当前项目的精确scope> --session <当前会话ID>` 打开面板；可管理任务、查看来源、纠错、忽略或恢复。三种模式为 `normal`、`recall_only`、`off`。关闭期间的内容不会在恢复后补写；已提交的远端任务保持原状态。有效纠错需要服务端反馈接口同步更新，线上服务须单独升级验收。
+
+工作台现已集成 Writer / 后台整理 API 配置、个人知识库和知识图谱。模型密钥只在本机表单输入并保存在受保护文件中，读取配置仅返回是否已保存；测试使用虚构 JSON 推理样本。知识与图谱分别读取已有服务接口，支持范围切换、搜索、节点选择及原始证据追溯。
+
+在支持工具与原生 ApprovalService 的 DSH 中，`tmcra_memory_control` 会对真实的记忆纠错请求先暂停该回合的自动写入，然后在聊天内展示来源、拟修改内容和范围，请用户确认。拒绝、取消或无确认渠道均不提交；已标识的其他正常回合不受影响。语义识别由宿主 Agent 按工具说明完成，假设性讨论不构成修改授权。此流程已通过 DSH 0.1.1-rc.2 原生工具/确认服务的隔离测试。
+
+预算默认 12000 字符；仅在证据仍位于实际模型上下文时跨轮去重。任务状态按凭据指纹和项目隔离；多个本机适配器可使用同一个 `TMCRA_MEMORY_STATE_DIR`。不同登录凭据的本机任务独立，已提交的长期记忆通过相同服务端项目 scope 共享。
+
 - 同一个项目可以在 Codex、DeepSeek Harness 等 TMCRA 接入工具之间继续推进。
 - 每轮第一次模型请求前自动召回，无需 Agent 主动调用工具。
 - 新开一个 Harness 对话后，可以继续同一项目的工作进度。
@@ -61,17 +69,17 @@ Agent 回答前，TMCRA 会找回与当前问题有关的用户要求、决策
 ## 安装技术预览包
 
 ```bash
-dsh plugin --profile web add https://github.com/reshuibuduo/tmcra-plugin-deepseek-harness/releases/download/v0.1.7/dsh-tmcra-memory-0.1.7.tgz
+dsh plugin --profile web add https://github.com/reshuibuduo/tmcra-plugin-deepseek-harness/releases/download/v1.0.0-rc.1/dsh-tmcra-memory-1.0.0-rc.1.tgz
 dsh plugin --profile web exec dsh-tmcra-memory login
 dsh --profile web --dump-config
 dsh web
 ```
 
 Harness Web UI 默认地址为 `http://127.0.0.1:3080`。
 
-压缩包内含 `cordis.patch.yml`，安装后会自动加入指定 Profile 的配置层。建议使用 Release 中已经构建好的 `.tgz`。下载到本地后，也可以执行 `dsh plugin --profile web add ./dsh-tmcra-memory-0.1.7.tgz`。从 Git 源码安装可能还需要在 `pnpm` 中单独允许构建脚本。
+压缩包内含 `cordis.patch.yml`，安装后会自动加入指定 Profile 的配置层。建议使用 Release 中已经构建好的 `.tgz`。下载到本地后，也可以执行 `dsh plugin --profile web add ./dsh-tmcra-memory-1.0.0-rc.1.tgz`。从 Git 源码安装可能还需要在 `pnpm` 中单独允许构建脚本。
 
-DSH `0.1.1-rc.2` 的 Windows 版仍可能把包含空格或非 ASCII 字符的本地包路径错误重组。请先把压缩包复制到纯 ASCII 短路径，例如 `D:\\dsh-packages\\dsh-tmcra-memory-0.1.7.tgz`；上面的 Release URL 不受影响。
+DSH `0.1.1-rc.2` 的 Windows 版仍可能把包含空格或非 ASCII 字符的本地包路径错误重组。请先把压缩包复制到纯 ASCII 短路径，例如 `D:\\dsh-packages\\dsh-tmcra-memory-1.0.0-rc.1.tgz`；上面的 Release URL 不受影响。
 
 ## 连接 TMCRA 账号
 
@@ -94,6 +102,14 @@ Harness 不会把凭据值写入普通设置或模型请求。如果模型控制
 
 ## 配置本地 Writer 与后台整理模型
 
+### 完整本地部署预览
+
+插件包内置完整后端，提供 E5/MiniLM、BGE-M3/BGE reranker、Qwen3 embedding/reranker 三档。装入 DSH 后运行 `dsh-tmcra-memory local-install`，免登录打开安装页并选择档位。Python、依赖和校验后的模型自动准备，本机身份自动生成，重启 DSH 后自动发现连接。支持 Windows x64，无需预装 Python、TMCRA 账号或 TMCRA 服务器。首次下载需要联网。原云端凭据保留；本地选择生效后，旧云端连接和后台云模型请求会被拦截，安装失败也保持本地选择。高级用户应先清除显式 `TMCRA_CONFIG_FILE` 覆盖。
+
+CPU 合成写入和原文召回已通过；复杂编译超时，后台整理和完整重启恢复仍待验收。轻量启动需约 6.3GiB 空闲内存。详见[验收说明](https://github.com/reshuibuduo/tmcra/blob/v1.0.0-rc.1/docs/LOCAL_DEPLOYMENT_PREVIEW.zh-CN.md)。宿主主模型在云端时，召回证据仍可能由宿主发往云端。
+
+### 配置独立模型 API
+
 安装后执行：
 
 ```bash

diff --git a/assets/tmcra-logo.png b/assets/tmcra-logo.png
diff --git a/dist/cli.mjs b/dist/cli.mjs
@@ -1,10 +1,12 @@
 #!/usr/bin/env node
+import { a as createMemoryActions, i as FilePendingTurnQueue, n as readLocalProviderConfig, o as startMemoryCenter, t as localProviderStageReady } from "./local-provider-config-CDBjMheh.mjs";
+import { n as readFullLocalConfig } from "./full-local-config-BiZAIc60.mjs";
 import { createHash, randomBytes } from "node:crypto";
-import { existsSync } from "node:fs";
+import { chmod, mkdir, readFile, rm, stat } from "node:fs/promises";
 import { homedir, platform } from "node:os";
 import { dirname, join, resolve } from "node:path";
-import { chmod, mkdir, readFile, rm, stat } from "node:fs/promises";
 import { spawn } from "node:child_process";
+import { existsSync } from "node:fs";
 import { fileURLToPath, pathToFileURL } from "node:url";
 import { withFileLock, writeFileAtomic } from "@deepseek-ai/dsh-atomic-write";
 import { Document, isMap, parseDocument } from "yaml";
@@ -58,7 +60,7 @@ async function postJson(fetchImpl, url, body) {
 			headers: {
 				Accept: "application/json",
 				"Content-Type": "application/json",
-				"User-Agent": "dsh-tmcra-memory/0.1.7"
+				"User-Agent": "dsh-tmcra-memory/1.0.0-rc.1"
 			},
 			body: JSON.stringify(body),
 			signal: controller.signal
@@ -175,6 +177,15 @@ async function updateHarnessCredentials(updates, dshHome) {
 	return path;
 }
 async function readHarnessCredentialStatus(dshHome) {
+	const local = await readFullLocalConfig();
+	if (local) return {
+		configured: true,
+		credentialsPath: "private-local-installation",
+		apiBaseUrl: local.baseUrl,
+		globalScope: local.globalScope,
+		projectScopePrefix: local.projectScopePrefix,
+		deploymentMode: "local"
+	};
 	const path = credentialsPath(dshHome);
 	if (!existsSync(path)) return {
 		configured: false,
@@ -190,6 +201,24 @@ async function readHarnessCredentialStatus(dshHome) {
 		projectScopePrefix: value[CREDENTIAL_KEYS.projectScopePrefix] || null
 	};
 }
+/** Local control-panel transport only. Never print this object or send it to a model. */
+async function readHarnessMemoryConnection(dshHome) {
+	const local = await readFullLocalConfig();
+	if (local) return local;
+	const path = credentialsPath(dshHome);
+	await assertOwnerOnly(path);
+	const values = existsSync(path) ? credentialsDocument(await readFile(path, "utf8")).toJS() : {};
+	const apiKey = process.env.TMCRA_API_KEY || values[CREDENTIAL_KEYS.apiKey];
+	const baseUrl = process.env.TMCRA_API_BASE_URL || values[CREDENTIAL_KEYS.apiBaseUrl] || "https://api.tmcra.com";
+	const url = new URL(baseUrl);
+	if (url.protocol !== "https:" || url.username || url.password || url.search || url.hash) throw new Error("TMCRA memory API must be an HTTPS origin");
+	if (!apiKey) throw new Error("Sign in with dsh-tmcra-memory login first");
+	return {
+		apiKey,
+		baseUrl,
+		globalScope: process.env.TMCRA_GLOBAL_SCOPE || values[CREDENTIAL_KEYS.globalScope]
+	};
+}
 async function atomicPending(path, value) {
 	await mkdir(dirname(path), { recursive: true });
 	await writeFileAtomic(path, `${JSON.stringify(value, null, 2)}\n`, {
@@ -277,7 +306,7 @@ async function authorizeDeepSeekHarness(options = {}) {
 	const started = await postJson(fetchImpl, `${authBaseUrl}/api/device/v1/authorizations`, {
 		clientId: CLIENT_ID,
 		clientName: `DeepSeek Harness (${platform()} ${process.arch})`,
-		clientVersion: "0.1.7",
+		clientVersion: "1.0.0-rc.1",
 		codeChallenge: challenge,
 		codeChallengeMethod: "S256"
 	});
@@ -383,9 +412,9 @@ function option(name) {
 function printJson(value) {
 	process.stdout.write(`${JSON.stringify(value)}\n`);
 }
-async function runProviderSetup() {
+async function runProviderSetup(filename = "provider_setup.mjs") {
 	const cliDirectory = dirname(fileURLToPath(import.meta.url));
-	const script = resolve(cliDirectory, "..", "scripts", "provider_setup.mjs");
+	const script = resolve(cliDirectory, "..", "scripts", filename);
 	const forwarded = process.argv.slice(3);
 	await new Promise((resolvePromise, reject) => {
 		const child = spawn(process.execPath, [script, ...forwarded], {
@@ -403,6 +432,60 @@ async function main() {
 	const command = process.argv[2] || "help";
 	const json = process.argv.includes("--json");
 	const dshHome = option("--dsh-home");
+	if (command === "local-install") {
+		await runProviderSetup("local_setup.mjs");
+		return;
+	}
+	if (command === "memory") {
+		const connection = await readHarnessMemoryConnection(dshHome);
+		const scope = option("--scope");
+		const sessionId = option("--session");
+		if (!scope || !sessionId) throw new Error("Use memory --scope EXACT_PROJECT_SCOPE --session EXACT_SESSION_ID");
+		const queue = new FilePendingTurnQueue(join(resolveDshHome(dshHome), "tmcra", "deepseek-harness-pending-turns.json"));
+		const invoke = createMemoryActions({
+			config: connection,
+			scope,
+			sessionId,
+			globalScope: connection.globalScope,
+			status: async () => ({ pending: (await queue.list()).filter((row) => row.scopeName === scope).map((row) => ({
+				id: row.idempotencyKey,
+				jobId: row.jobId,
+				state: row.observedStatus || "queued"
+			})) }),
+			request: async (path, options) => {
+				const provider = await readLocalProviderConfig().catch(() => null);
+				const localHeaders = {};
+				if (provider && localProviderStageReady(provider, "writer")) localHeaders["X-TMCRA-Writer-Execution"] = "user-provider";
+				if (provider && localProviderStageReady(provider, "organizer")) localHeaders["X-TMCRA-Organizer-Execution"] = "user-provider";
+				const response = await fetch(`${connection.baseUrl.replace(/\/+$/u, "")}${path}`, {
+					method: options.method,
+					headers: {
+						Authorization: `Bearer ${connection.apiKey}`,
+						"Content-Type": "application/json",
+						"X-TMCRA-Client-Platform": "deepseek_harness",
+						"X-TMCRA-Integration-ID": "tmcra-deepseek-harness",
+						...localHeaders,
+						...options.headers
+					},
+					body: JSON.stringify(options.body),
+					signal: AbortSignal.timeout(3e4),
+					redirect: "error"
+				});
+				if (!response.ok) throw new Error(`TMCRA request failed (${response.status}); check authorization and service version`);
+				return response.json();
+			}
+		});
+		const center = await startMemoryCenter({
+			invoke,
+			open: !process.argv.includes("--no-open")
+		});
+		printJson({
+			url: center.url,
+			credentialsLocalOnly: true
+		});
+		await new Promise((done) => center.server.once("close", done));
+		return;
+	}
 	if (command === "setup" || command === "configure-models") {
 		await runProviderSetup();
 		return;
@@ -447,7 +530,7 @@ async function main() {
 		else process.stdout.write(`TMCRA credentials were removed from ${path}. Revoke the connection in your TMCRA account if this device is no longer trusted.\n`);
 		return;
 	}
-	process.stdout.write(`Usage:\n  dsh-tmcra-memory setup [--no-open] [--config-file PATH] [--json]\n  dsh-tmcra-memory login [--no-open] [--auth-base-url URL] [--dsh-home PATH] [--json]\n  dsh-tmcra-memory status [--dsh-home PATH] [--json]\n  dsh-tmcra-memory logout [--dsh-home PATH] [--json]\n`);
+	process.stdout.write(`Usage:\n  dsh-tmcra-memory local-install [--no-open]\n  dsh-tmcra-memory setup [--no-open] [--config-file PATH] [--json]\n  dsh-tmcra-memory memory --scope EXACT_PROJECT_SCOPE --session EXACT_SESSION_ID [--no-open] [--dsh-home PATH]\n  dsh-tmcra-memory login [--no-open] [--auth-base-url URL] [--dsh-home PATH] [--json]\n  dsh-tmcra-memory status [--dsh-home PATH] [--json]\n  dsh-tmcra-memory logout [--dsh-home PATH] [--json]\n`);
 }
 const invokedPath = process.argv[1] ? pathToFileURL(process.argv[1]).href : "";
 if (import.meta.url === invokedPath) await main().catch((error) => {