@co0ontty/wand 4.89.0 → 4.89.1

package/dist/git-quick-commit.d.ts

@@ -1,3 +1,4 @@
+import { type EmployeeTextDeps } from "./employee-text.js";
 import type { AiTextRequest } from "./types.js";
 import { GitStatusResult, PushResult, QuickCommitResult, SessionProvider, SessionSnapshot, TagHeadResult } from "./types.js";
 export type QuickCommitErrorCode = "CWD_MISSING" | "NO_CWD" | "NOT_A_GIT_REPO" | "NO_COMMIT" | "NOTHING_TO_COMMIT" | "NOTHING_TO_PUSH" | "EMPTY_MESSAGE" | "EMPTY_TAG" | "EMPTY_AI_MESSAGE" | "AI_FALLBACK_FAILED" | "INVALID_AI_TAG" | "TAG_EXISTS" | "GIT_ADD_FAILED" | "GIT_DIFF_FAILED" | "GIT_COMMIT_FAILED" | "GIT_TAG_FAILED" | "CLAUDE_CLI_MISSING" | "CLAUDE_CLI_FAILED" | "CLAUDE_TIMEOUT";
@@ -26,12 +27,15 @@ export interface QuickCommitAiOptions {
     inheritEnv?: boolean;
     /** 内置「系统运维」员工的角色设定：作为系统提示前缀注入。 */
     opsPersona?: string;
-    /** CLI 降级链（按顺序）；未设置时只用 provider/model 这一次调用。 */
+    /** 员工候选链（按顺序）；未设置时只用 provider/model 这一次调用。 */
     cliCandidates?: import("./types.js").AiCliCandidate[];
-    /** 系统应用必须经员工渠道；没有可用 CLI 候选时直接失败，不退回默认 provider。 */
+    /** SDK 使用当前实例配置与现有免费分组服务，不创建其他执行渠道。 */
+    employeeText?: EmployeeTextDeps;
+    /** 系统应用必须经员工渠道；没有可用候选时直接失败，不退回默认 provider。 */
     employeeChannelOnly?: boolean;
     /** 整条候选链的时间上限；默认 CLI_CHAIN_BUDGET_MS。第一条消息里的调用要短得多。 */
     budgetMs?: number;
+    signal?: AbortSignal;
 }
 export declare class QuickCommitError extends Error {
     readonly code: QuickCommitErrorCode;
@@ -47,11 +51,16 @@ export interface CommitInputOptions {
     iterationDigest?: string;
     includeDiff?: boolean;
 }
+type CliAiTextOptions = QuickCommitAiOptions & {
+    deadline?: number;
+};
+export declare function assertGeminiTextCapability(): void;
+export declare function callCliAiText(request: AiTextRequest, cwd: string, language: string, opts: CliAiTextOptions): Promise<string>;
 /** 角色设定走系统提示前缀；任务自己的输出格式与约束必须排在后面。 */
 export declare function withOpsPersona(request: AiTextRequest, persona?: string): AiTextRequest;
 /**
- * Run a lightweight AI request through the system employee's CLI chain.
- * 所有 Wand 自有文本调用都走本机已安装的 CLI，没有任何直连 API 分支。
+ * Run a lightweight AI request through the system employee's configured chain.
+ * provider 与引擎由候选决定；SDK 与 CLI 均保留各自适配和身份。
  * parseOutput 在候选链内解析/校验结果；失败后继续下一条，而非提前宣布成功。
  */
 export declare function callConfiguredAiText(request: AiTextRequest, cwd: string, language: string, opts: QuickCommitAiOptions): Promise<string>;

package/dist/git-quick-commit.js

@@ -1,5 +1,9 @@
-import { existsSync } from "node:fs";
+import { existsSync, mkdtempSync, readdirSync, rmSync, writeFileSync } from "node:fs";
+import { randomUUID } from "node:crypto";
+import { tmpdir } from "node:os";
+import { join } from "node:path";
 import { spawn } from "node:child_process";
+import { abortableEmployeeText, prepareEmployeeTextCandidate } from "./employee-text.js";
 import { buildChildEnv } from "./env-utils.js";
 import { buildLanguageDirective } from "./language-prompt.js";
 import { isSessionProvider, providerCliInstalled } from "./session-provider.js";
@@ -306,7 +310,7 @@ async function callClaudeText(request, cwd, language, opts) {
     const stdout = await runCliText("claude", args, request.prompt, {
         cwd,
         timeoutMs: cliTextTimeoutMs(CLAUDE_MESSAGE_TIMEOUT_MS, opts),
-        inheritEnv: opts.inheritEnv,
+        inheritEnv: opts.inheritEnv, signal: opts.signal,
     });
     const text = stdout.trim();
     if (!text)
@@ -538,6 +542,8 @@ function buildGrokTextArgs(request, opts, allowTools = false) {
     args.push(...systemPromptArgsFrom("grok", request.system));
     if (allowTools)
         args.push("--always-approve");
+    else
+        args.push("--tools", "", "--disallowed-tools", "run_terminal_cmd,read_file,search_replace,grep,list_dir,web_search,web_fetch,todo_write,task,Agent", "--deny", "MCPTool", "--disable-web-search", "--no-subagents");
     return args;
 }
 /**
@@ -568,6 +574,8 @@ function buildQoderTextArgs(request, opts, allowTools = false) {
     args.push(...systemPromptArgsFrom("qoder", request.system));
     if (allowTools)
         args.push("--permission-mode", "bypass_permissions");
+    else
+        args.push("--tools", "", "--strict-mcp-config");
     return args;
 }
 /** provider 自己的系统提示开关 + 规则文本；没有开关的 provider 返回空。 */
@@ -583,35 +591,48 @@ function contentWithSystemPrompt(provider, request) {
     return systemPromptFlag(provider) ? request.prompt : composeSystemFallback(request.system, request.prompt);
 }
 function runCliText(command, args, prompt, opts) {
+    opts.signal?.throwIfAborted();
     return new Promise((resolve, reject) => {
+        const baseEnv = buildChildEnv(opts.inheritEnv !== false);
         const child = spawn(command, args, {
             cwd: opts.cwd,
-            env: buildChildEnv(opts.inheritEnv !== false),
+            env: opts.env ? opts.env(baseEnv) : baseEnv,
             stdio: ["pipe", "pipe", "pipe"],
         });
         let stdout = "";
         let stderr = "";
         let settled = false;
+        const cancelled = () => {
+            if (settled)
+                return;
+            settled = true;
+            clearTimeout(timeout);
+            child.kill("SIGTERM");
+            reject(opts.signal?.reason ?? new Error("AI 文本请求已取消。"));
+        };
+        const cleanup = () => { clearTimeout(timeout); opts.signal?.removeEventListener("abort", cancelled); };
         const timeout = setTimeout(() => {
             settled = true;
+            opts.signal?.removeEventListener("abort", cancelled);
             child.kill("SIGTERM");
             reject(new QuickCommitError(`${command} 调用超时。`, "CLAUDE_TIMEOUT"));
         }, opts.timeoutMs);
+        opts.signal?.addEventListener("abort", cancelled, { once: true });
         child.stdout?.on("data", (chunk) => { stdout += chunk.toString(); });
         child.stderr?.on("data", (chunk) => { stderr += chunk.toString(); });
         child.on("error", (error) => {
             if (settled)
                 return;
             settled = true;
-            clearTimeout(timeout);
+            cleanup();
             const code = error.code === "ENOENT" ? "CLAUDE_CLI_MISSING" : "CLAUDE_CLI_FAILED";
             reject(new QuickCommitError(error.code === "ENOENT" ? `未找到 ${command} CLI。` : `${command} CLI 失败：${error.message}`, code));
         });
         child.on("close", (code) => {
+            cleanup();
             if (settled)
                 return;
             settled = true;
-            clearTimeout(timeout);
             if (code === 0) {
                 resolve(stdout);
                 return;
@@ -632,7 +653,7 @@ function runCliText(command, args, prompt, opts) {
             if (settled || !expectsStdin)
                 return;
             settled = true;
-            clearTimeout(timeout);
+            cleanup();
             child.kill("SIGTERM");
             reject(new QuickCommitError(`${command} CLI stdin failed: ${error.code ?? "write error"}`, "CLAUDE_CLI_FAILED"));
         });
@@ -654,7 +675,7 @@ async function callCodexText(request, cwd, opts) {
     const stdout = await runCliText("codex", args, contentWithSystemPrompt("codex", request), {
         cwd,
         timeoutMs: cliTextTimeoutMs(CODEX_MESSAGE_TIMEOUT_MS, opts),
-        inheritEnv: opts.inheritEnv,
+        inheritEnv: opts.inheritEnv, signal: opts.signal,
     });
     const text = extractCodexText(stdout);
     if (!text) {
@@ -663,7 +684,8 @@ async function callCodexText(request, cwd, opts) {
     return text;
 }
 async function callOpenCodeText(request, cwd, opts) {
-    const args = ["run", "--format", "json"];
+    const agent = `wand-text-${randomUUID()}`;
+    const args = ["run", "--format", "json", "--pure", "--agent", agent];
     const model = opts.model?.trim();
     if (model && model !== "default")
         args.push("--model", model);
@@ -673,7 +695,18 @@ async function callOpenCodeText(request, cwd, opts) {
     const stdout = await runCliText("opencode", args, contentWithSystemPrompt("opencode", request), {
         cwd,
         timeoutMs: cliTextTimeoutMs(CODEX_MESSAGE_TIMEOUT_MS, opts),
-        inheritEnv: opts.inheritEnv,
+        inheritEnv: opts.inheritEnv, signal: opts.signal,
+        // A dedicated per-invocation agent and scalar deny replace rather than merge
+        // wildcard/specific allow rules. Preserve inherited provider/model config.
+        env: base => {
+            const configured = base.OPENCODE_CONFIG_CONTENT ? JSON.parse(base.OPENCODE_CONFIG_CONTENT) : {};
+            if (!configured || typeof configured !== "object" || Array.isArray(configured))
+                throw new Error("OpenCode 内联配置无效。");
+            return { ...base, OPENCODE_PERMISSION: JSON.stringify("deny"), OPENCODE_CONFIG_CONTENT: JSON.stringify({
+                    ...configured, permission: "deny", autoupdate: false, share: "disabled",
+                    agent: { ...configured.agent, [agent]: { mode: "primary", permission: "deny", prompt: "Return only the requested text. Do not use tools." } },
+                }) };
+        },
     });
     const text = extractOpenCodeText(stdout);
     if (!text)
@@ -684,7 +717,7 @@ async function callGrokText(request, cwd, opts) {
     const stdout = await runCliText("grok", buildGrokTextArgs(request, opts), "", {
         cwd,
         timeoutMs: cliTextTimeoutMs(CODEX_MESSAGE_TIMEOUT_MS, opts),
-        inheritEnv: opts.inheritEnv,
+        inheritEnv: opts.inheritEnv, signal: opts.signal,
     });
     const text = extractGrokText(stdout);
     if (!text)
@@ -695,7 +728,7 @@ async function callQoderText(request, cwd, opts) {
     const stdout = await runCliText("qodercli", buildQoderTextArgs(request, opts), "", {
         cwd,
         timeoutMs: cliTextTimeoutMs(CODEX_MESSAGE_TIMEOUT_MS, opts),
-        inheritEnv: opts.inheritEnv,
+        inheritEnv: opts.inheritEnv, signal: opts.signal,
     });
     const text = extractQoderText(stdout);
     if (!text)
@@ -704,25 +737,42 @@ async function callQoderText(request, cwd, opts) {
 }
 async function callPiText(request, cwd, opts) {
     const stdout = await runCliText("pi", buildPiTextArgs(request, opts), "", {
-        cwd, timeoutMs: cliTextTimeoutMs(CODEX_MESSAGE_TIMEOUT_MS, opts), inheritEnv: opts.inheritEnv,
+        cwd, timeoutMs: cliTextTimeoutMs(CODEX_MESSAGE_TIMEOUT_MS, opts), inheritEnv: opts.inheritEnv, signal: opts.signal,
     });
     const text = extractPiText(stdout);
     if (!text)
         throw new QuickCommitError("Pi 返回了空的 commit message。", "EMPTY_AI_MESSAGE");
     return text;
 }
+export function assertGeminiTextCapability() {
+    // Gemini ignores supplemental admin policy when a central policy exists.
+    // Refuse that unsupported text-only path rather than launching without deny.
+    const systemPolicies = process.platform === "darwin" ? "/Library/Application Support/GeminiCli/policies"
+        : process.platform === "win32" ? join(process.env.ProgramData ?? "C:\\ProgramData", "gemini-cli", "policies") : "/etc/gemini-cli/policies";
+    if (existsSync(systemPolicies) && readdirSync(systemPolicies).some(file => file.endsWith(".toml"))) {
+        throw new QuickCommitError("Gemini 中央策略阻止注入纯文本工具边界，请配置可用的文本候选。", "AI_FALLBACK_FAILED");
+    }
+}
 async function callGeminiText(request, cwd, opts) {
-    const stdout = await runCliText("gemini", buildGeminiTextArgs(opts), contentWithSystemPrompt("gemini", request), {
-        cwd,
-        timeoutMs: cliTextTimeoutMs(CODEX_MESSAGE_TIMEOUT_MS, opts),
-        inheritEnv: opts.inheritEnv,
-    });
-    const text = extractGeminiText(stdout);
-    if (!text)
-        throw new QuickCommitError("Gemini 返回了空的 commit message。", "EMPTY_AI_MESSAGE");
-    return text;
+    assertGeminiTextCapability();
+    const root = mkdtempSync(join(tmpdir(), "wand-gemini-text-"));
+    try {
+        const policy = join(root, "text-only.toml");
+        writeFileSync(policy, '[[rule]]\ntoolName = "*"\ndecision = "deny"\npriority = 999\n', { mode: 0o600 });
+        const args = [...buildGeminiTextArgs(opts), "--admin-policy", policy, "--extensions", "none", "--allowed-mcp-server-names", ""];
+        const stdout = await runCliText("gemini", args, contentWithSystemPrompt("gemini", request), {
+            cwd, timeoutMs: cliTextTimeoutMs(CODEX_MESSAGE_TIMEOUT_MS, opts), inheritEnv: opts.inheritEnv, signal: opts.signal,
+        });
+        const text = extractGeminiText(stdout);
+        if (!text)
+            throw new QuickCommitError("Gemini 返回了空的 commit message。", "EMPTY_AI_MESSAGE");
+        return text;
+    }
+    finally {
+        rmSync(root, { recursive: true, force: true });
+    }
 }
-async function callCliAiText(request, cwd, language, opts) {
+export async function callCliAiText(request, cwd, language, opts) {
     const provider = defaultProvider(opts.provider);
     if (provider === "codex") {
         return callCodexText(request, cwd, opts);
@@ -774,28 +824,56 @@ function singleCandidate(opts) {
  * 只有一条候选时保持原行为（错误原样抛出，不摘要）。
  * 仅用于一次性文本生成；执行工具的任务不能使用这条无条件重试路径。
  */
-async function callCliCandidates(request, cwd, language, opts, parseOutput) {
+async function callTextCandidates(request, cwd, language, opts, parseOutput) {
+    opts.signal?.throwIfAborted();
     const { resolveModelGroupModels } = await import("./model-groups.js");
+    const modelGroups = opts.modelGroups ?? opts.employeeText?.config.modelGroups;
+    const seen = new Set();
     const chain = (opts.cliCandidates?.length ? opts.cliCandidates : [singleCandidate(opts)])
-        .flatMap((candidate) => resolveModelGroupModels(opts.modelGroups, candidate.provider, candidate.model, { preferDefault: !candidate.model || candidate.model === "default" })
-        .map((model) => ({ ...candidate, model: model || undefined })));
+        .flatMap((candidate) => resolveModelGroupModels(modelGroups, candidate.provider, candidate.model, { preferDefault: !candidate.model || candidate.model === "default" })
+        .map((model) => ({ ...candidate, model: model || undefined })))
+        .filter(candidate => {
+        const key = JSON.stringify([candidate.provider, candidate.engine ?? "cli", candidate.model ?? "", candidate.thinkingEffort ?? opts.thinkingEffort]);
+        if (seen.has(key))
+            return false;
+        seen.add(key);
+        return true;
+    });
     if (opts.employeeChannelOnly && !opts.cliCandidates?.length) {
-        throw new QuickCommitError("系统员工没有可用的 CLI 候选，已跳过 Wand Agent。", "AI_FALLBACK_FAILED");
+        throw new QuickCommitError("系统员工没有可用的执行候选。", "AI_FALLBACK_FAILED");
     }
     const deadline = Date.now() + Math.max(1, opts.budgetMs ?? CLI_CHAIN_BUDGET_MS);
-    if (chain.length === 1) {
-        return parseOutput(await callCliAiText(request, cwd, language, { ...candidateOptions(opts, chain[0]), deadline }));
-    }
-    const installed = chain.filter((candidate) => providerCliInstalled(candidate.provider));
+    const call = async (candidate) => {
+        if (candidate.engine !== "sdk")
+            return callCliAiText(request, cwd, language, { ...candidateOptions(opts, candidate), deadline });
+        const controller = new AbortController();
+        const timer = setTimeout(() => controller.abort(new QuickCommitError("Wand Agent 调用超时。", "CLAUDE_TIMEOUT")), Math.max(1, deadline - Date.now()));
+        const signal = opts.signal ? AbortSignal.any([opts.signal, controller.signal]) : controller.signal;
+        try {
+            const generate = await abortableEmployeeText(prepareEmployeeTextCandidate({
+                ...opts.employeeText,
+                config: { ...opts.employeeText?.config, defaultCwd: cwd, inheritEnv: opts.inheritEnv ?? opts.employeeText?.config.inheritEnv, modelGroups },
+            }, { ...candidate, model: candidate.model ?? "default", thinkingEffort: candidate.thinkingEffort ?? opts.thinkingEffort ?? "off", mode: "default", kind: "structured" }, request, signal), signal);
+            signal.throwIfAborted();
+            return await abortableEmployeeText(generate(), signal);
+        }
+        finally {
+            clearTimeout(timer);
+        }
+    };
+    if (chain.length === 1)
+        return parseOutput(await call(chain[0]));
+    const installed = chain.filter((candidate) => candidate.engine === "sdk" || providerCliInstalled(candidate.provider));
     const attempts = installed.length ? installed : chain;
     const errors = [];
     for (const candidate of attempts) {
+        opts.signal?.throwIfAborted();
         if (Date.now() >= deadline) {
             errors.push("已超过等待上限");
             break;
         }
         try {
-            const text = await callCliAiText(request, cwd, language, { ...candidateOptions(opts, candidate), deadline });
+            const text = await call(candidate);
             const result = await parseOutput(text);
             if (candidate !== attempts[0]) {
                 console.info(`[SystemAi] 候选 ${chain.indexOf(candidate) + 1}/${chain.length} (${candidate.provider}) 成功`);
@@ -803,16 +881,17 @@ async function callCliCandidates(request, cwd, language, opts, parseOutput) {
             return result;
         }
         catch (error) {
+            opts.signal?.throwIfAborted();
             errors.push(`${candidate.provider}: ${getGitErrorMessage(error)}`);
             // 不记录提示词、路径、CLI 原始响应或可能带鉴权信息的错误正文。
             const code = error instanceof QuickCommitError ? error.code : "INVALID_AI_RESULT";
             console.warn(`[SystemAi] 候选 ${chain.indexOf(candidate) + 1}/${chain.length} (${candidate.provider}) 失败: ${code}`);
         }
     }
-    throw new QuickCommitError(`所有 CLI 候选均失败：${errors.join("；")}`, "AI_FALLBACK_FAILED");
+    throw new QuickCommitError(`所有执行候选均失败：${errors.join("；")}`, "AI_FALLBACK_FAILED");
 }
 export async function callConfiguredAiText(request, cwd, language, opts, parseOutput = (raw) => raw) {
-    return callCliCandidates(withOpsPersona(request, opts.opsPersona), cwd, language, opts, parseOutput);
+    return callTextCandidates(withOpsPersona(request, opts.opsPersona), cwd, language, opts, parseOutput);
 }
 /** Read the unstaged + staged tree without touching the index. */
 async function collectWorkingTreeDiff(cwd) {

package/dist/local-model-setup.d.ts

@@ -3,6 +3,7 @@ import type { DecisionService } from "./decision-service.js";
 import { type ModelFile } from "./laya-model-manifest.js";
 import { runModelSetup } from "./model-setup-process.js";
 import type { SpeechService } from "./speech-service.js";
+import { type SpeechSupport } from "./speech-support.js";
 import type { LocalModelKind, LocalModelsStatus } from "./local-model-types.js";
 export interface LocalModelSetupDeps {
     configDir: string;
@@ -17,6 +18,7 @@ export interface LocalModelSetupDeps {
     run?: typeof runModelSetup;
     checkRuntime?: (python: string, signal: AbortSignal) => Promise<boolean>;
     timeoutMs?: number;
+    speechSupport?: (runtimeAvailable: boolean) => Promise<SpeechSupport>;
 }
 /** Deployment jobs only; existing inference owners retain their protocols, rate limits and capabilities. */
 export declare class LocalModelSetupService {
@@ -28,14 +30,18 @@ export declare class LocalModelSetupService {
     private readonly jobs;
     private sequence;
     private disposed;
+    private speechEnableRevision;
+    private supportCache;
     constructor(deps: LocalModelSetupDeps);
     private supported;
     private managedModel;
     private layaModel;
     private python;
     status(): Promise<LocalModelsStatus>;
+    private speechSupport;
+    setSpeechEnabled(enabled: boolean): Promise<void>;
     private parse;
-    start(kind: LocalModelKind, action: "download" | "initialize", raw: unknown): void;
+    start(kind: LocalModelKind, action: "download" | "initialize" | "activate", raw: unknown): void;
     private update;
     private work;
     private checkRuntime;

package/dist/local-model-setup.js

@@ -7,6 +7,7 @@ import { LAYA_FILES, LAYA_REPOSITORY, layaFileUrl } from "./laya-model-manifest.
 import { ModelFileStore } from "./model-file-store.js";
 import { runModelSetup } from "./model-setup-process.js";
 import { speechModel } from "./speech-models.js";
+import { speechSupport } from "./speech-support.js";
 import { systemEnvValue } from "./env-utils.js";
 import { decisionHardware } from "./decision-hardware.js";
 import { DECISION_EXPERT_ID } from "./decision-expert-identity.js";
@@ -20,6 +21,8 @@ export class LocalModelSetupService {
     jobs = new Map();
     sequence = 0;
     disposed = false;
+    speechEnableRevision = 0;
+    supportCache = null;
     constructor(deps) {
         this.deps = deps;
         this.files = new ModelFileStore(deps.fetch);
@@ -55,6 +58,7 @@ export class LocalModelSetupService {
         const hardware = decisionHardware();
         const operation = this.operations.get("laya") ?? null;
         const selected = speech.models.find((value) => value.id === speech.settings.model);
+        const support = await this.speechSupport(speech.runtime.available);
         return {
             laya: { kind: "laya", label: "LAYA 本地决策", supported,
                 reason: !supported ? hardware.message
@@ -62,12 +66,38 @@ export class LocalModelSetupService {
                 enabled: decision.enabled, model: LAYA_REPOSITORY, modelSize: this.manifest.reduce((sum, file) => sum + file.size, 0),
                 downloaded: model.downloaded, runtimeAvailable: !!python, initialized: decision.state === "ready", busy: this.jobs.has("laya") || decision.queued > 0 || decision.state === "loading",
                 operation: operation ? { ...operation } : null, hardware, decisionEmployeeId: DECISION_EXPERT_ID },
-            speech: { kind: "speech", label: "服务端语音识别", supported: ["darwin", "linux", "win32"].includes(process.platform), reason: speech.reason,
+            speech: { kind: "speech", label: "服务端语音识别", supported: support.supported, supportReason: support.reason, reason: speech.reason,
                 enabled: speech.settings.enabled, model: speech.settings.model, modelSize: selected?.size ?? 0,
                 downloaded: selected?.downloaded ?? false, runtimeAvailable: speech.runtime.available, initialized: speech.initialized === true,
                 busy: speech.busy || this.jobs.has("speech"), operation: this.operations.has("speech") ? { ...this.operations.get("speech") } : null },
         };
     }
+    speechSupport(runtimeAvailable) {
+        if (!this.supportCache || this.supportCache.expires < Date.now() || this.supportCache.runtimeAvailable !== runtimeAvailable) {
+            this.supportCache = { expires: Date.now() + 30_000, runtimeAvailable, value: (this.deps.speechSupport ?? speechSupport)(runtimeAvailable) };
+        }
+        return this.supportCache.value;
+    }
+    async setSpeechEnabled(enabled) {
+        const revision = ++this.speechEnableRevision;
+        const settingsRevision = this.deps.speech.settingsRevision();
+        if (this.disposed)
+            throw new DecisionError("UNAVAILABLE", "模型管理已关闭。", 503);
+        if (!enabled) {
+            this.cancel("speech");
+            this.deps.speech.configure({ enabled: false });
+            return;
+        }
+        const status = await this.deps.speech.status();
+        const support = await this.speechSupport(status.runtime.available);
+        if (revision !== this.speechEnableRevision || this.disposed || settingsRevision !== this.deps.speech.settingsRevision())
+            return;
+        if (!support.supported)
+            throw new DecisionError("UNSUPPORTED", support.reason || "本机不支持语音识别。", 409);
+        if (status.busy || status.download?.phase === "downloading")
+            throw new DecisionError("BUSY", "语音识别或模型下载正在进行，请稍后启用。", 409);
+        this.start("speech", "activate", {});
+    }
     parse(kind, raw) {
         if (!raw || typeof raw !== "object" || Array.isArray(raw))
             throw new DecisionError("INVALID_REQUEST", "模型操作参数无效。");
@@ -90,6 +120,8 @@ export class LocalModelSetupService {
             throw new DecisionError("UNAVAILABLE", "模型管理已关闭。", 503);
         if (kind === "laya" && !this.supported())
             throw new DecisionError("UNSUPPORTED", `LAYA-MLX 需要 Apple Silicon macOS 与 Metal；${decisionHardware().message}`, 409);
+        if (kind === "laya" && action === "activate")
+            throw new DecisionError("INVALID_REQUEST", "一键启用仅用于语音识别。");
         const input = this.parse(kind, raw);
         if (this.jobs.has(kind))
             throw new DecisionError("BUSY", "该模型已有安装任务，请等待或取消。", 409);
@@ -119,6 +151,8 @@ export class LocalModelSetupService {
         timer.unref();
         const signal = abort.signal;
         const phase = (value, message) => this.update(kind, id, { phase: value, message });
+        let activationMaintenance = false;
+        const settingsRevision = this.deps.speech.settingsRevision();
         try {
             if (kind === "laya") {
                 let model = await this.layaModel();
@@ -170,7 +204,7 @@ export class LocalModelSetupService {
             }
             else {
                 const selected = input.model ?? this.deps.speech.settings().model;
-                if (action === "download") {
+                if (action === "download" || action === "activate") {
                     phase("downloading", "下载并校验语音模型");
                     await this.deps.speech.startDownload(selected);
                     while (true) {
@@ -188,7 +222,12 @@ export class LocalModelSetupService {
                         });
                     }
                 }
-                else {
+                if (action !== "download") {
+                    if (action === "activate") {
+                        signal.throwIfAborted();
+                        this.deps.speech.acquireMaintenance();
+                        activationMaintenance = true;
+                    }
                     const status = await this.deps.speech.status();
                     signal.throwIfAborted();
                     const backend = input.backend === "auto" || !input.backend ? process.platform === "darwin" ? "metal" : "cpu" : input.backend;
@@ -207,22 +246,30 @@ export class LocalModelSetupService {
                             await rm(work, { recursive: true, force: true }).catch(() => { });
                         }
                     }
-                    phase("initializing", "加载所选语音模型并用合成静音检查，未启用识别或重启服务");
-                    await this.deps.speech.initializeModel(selected, signal);
+                    phase("initializing", action === "activate" ? "正在检查语音模型，完成后自动启用。" : "加载所选语音模型并用合成静音检查，未启用识别或重启服务");
+                    await this.deps.speech.initializeModel(selected, signal, action === "activate" ? "auto" : undefined);
+                    if (action === "activate") {
+                        signal.throwIfAborted();
+                        if (settingsRevision !== this.deps.speech.settingsRevision())
+                            throw new DecisionError("SETTINGS_CHANGED", "语音设置已由其他操作修改，本次启用已取消。");
+                        this.deps.speech.configure({ enabled: true, acceleration: "auto" });
+                    }
                 }
             }
             signal.throwIfAborted();
-            phase("completed", action === "download" ? "模型下载及完整性校验完成；可继续初始化" : "运行时和模型初始化检查完成；启用状态保持不变");
+            phase("completed", action === "activate" ? "服务端语音识别已启用。" : action === "download" ? "模型下载及完整性校验完成；可继续初始化" : "运行时和模型初始化检查完成；启用状态保持不变");
         }
         catch (error) {
             const cancelled = signal.aborted;
             const message = cancelled ? "安装任务已取消，完整文件保留，可重试" : error instanceof DecisionError || (error instanceof Error && /请先|校验|自定义|独立/.test(error.message))
                 ? error.message : "安装或初始化失败，请检查网络、编译工具/Python版本、磁盘空间和平台能力，再重试。";
             this.update(kind, id, { phase: cancelled ? "cancelled" : "failed", message, ...(cancelled ? {} : { error: message }) });
-            if (kind === "speech" && action === "download" && cancelled)
+            if (kind === "speech" && (action === "download" || action === "activate") && cancelled)
                 this.deps.speech.cancelDownload();
         }
         finally {
+            if (activationMaintenance)
+                this.deps.speech.releaseMaintenance();
             clearTimeout(timer);
         }
     }

package/dist/local-model-types.d.ts

@@ -3,7 +3,7 @@ export type LocalModelKind = "laya" | "speech";
 export type ModelSetupPhase = "downloading" | "verifying" | "runtime" | "initializing" | "completed" | "failed" | "cancelled";
 export interface ModelSetupOperation {
     id: number;
-    action: "download" | "initialize";
+    action: "download" | "initialize" | "activate";
     phase: ModelSetupPhase;
     message: string;
     received: number;
@@ -14,6 +14,7 @@ export interface LocalModelStatus {
     kind: LocalModelKind;
     label: string;
     supported: boolean;
+    supportReason?: string | null;
     reason: string | null;
     enabled: boolean;
     model: string;

package/dist/process-manager.d.ts

@@ -3,6 +3,7 @@ import { EventEmitter } from "node:events";
 import { WandStorage } from "./storage.js";
 import { ExecutionMode, ProcessEventHandler, SessionProvider, SessionSnapshot, SessionSource, WandConfig } from "./types.js";
 import { type PermissionResolution } from "./claude-pty-bridge.js";
+import type { EmployeeTextDeps } from "./employee-text.js";
 import { type PtyTerminalSnapshot, type PtyHistoryPage } from "./pty-terminal-state.js";
 import { samplePtyForegrounds } from "./pty-foreground.js";
 import { type TerminalHost } from "./terminal-host.js";
@@ -34,6 +35,7 @@ export declare const PTY_TURN_IDLE_MS = 3000;
  */
 export declare const PTY_FOREGROUND_SAMPLE_MS = 1000;
 export interface ProcessManagerOptions {
+    free?: EmployeeTextDeps["free"];
     /** Overrides the non-Claude PTY turn quiet window (tests use a short one). */
     ptyTurnIdleMs?: number;
     /** Overrides the foreground-process-group sampling interval (tests use a short one). */
@@ -55,6 +57,7 @@ export declare class ProcessManager extends EventEmitter {
     private readonly ptyForegroundSampleMs;
     /** Foreground probe; injectable so tests can drive it without a real process table. */
     private readonly probePtyForegrounds;
+    private readonly free;
     private readonly resolveDecisionEvaluate;
     /** 24h archive scan timer */
     private archiveTimer;
@@ -177,7 +180,7 @@ export declare class ProcessManager extends EventEmitter {
      */
     setSessionModel(id: string, model: string | null): SessionSnapshot;
     /**
-     * 归档会先停掉还在跑的终端，但保留会话 ID 和 provider session id，恢复时用它们续上。
+     * 归档只写标记，保留正在运行的终端、会话 ID 和 provider session id。
      * 取消归档只清标记；重新拉起由恢复入口负责。
      */
     setSessionArchived(id: string, archived: boolean): SessionSnapshot;

package/dist/process-manager.js

@@ -563,6 +563,7 @@ export class ProcessManager extends EventEmitter {
     ptyForegroundSampleMs;
     /** Foreground probe; injectable so tests can drive it without a real process table. */
     probePtyForegrounds;
+    free;
     resolveDecisionEvaluate;
     /** 24h archive scan timer */
     archiveTimer = null;
@@ -595,6 +596,7 @@ export class ProcessManager extends EventEmitter {
         this.ptyTurnIdleMs = options.ptyTurnIdleMs ?? PTY_TURN_IDLE_MS;
         this.ptyForegroundSampleMs = options.ptyForegroundSampleMs ?? PTY_FOREGROUND_SAMPLE_MS;
         this.probePtyForegrounds = options.samplePtyForegrounds ?? samplePtyForegrounds;
+        this.free = options.free;
         this.resolveDecisionEvaluate = options.resolveDecisionEvaluate ?? (() => undefined);
         this.terminalHost = terminalHost ?? new InProcessTerminalHost();
         this.logger = new SessionLogger(configDir || path.join(process.env.HOME || process.cwd(), ".wand"), config.shortcutLogMaxBytes);
@@ -1229,7 +1231,7 @@ export class ProcessManager extends EventEmitter {
                     ai: resolveSystemAiContext({
                         provider: groupProvider, runner: undefined, command: providerCliCommand(groupProvider),
                         structuredState: undefined, selectedModel: null, thinkingEffort: opts?.thinkingEffort,
-                    }, this.config, this.storage.getSystemSiliconEmployee()),
+                    }, this.config, this.storage.getSystemSiliconEmployee(), this.free),
                     cwd: resolveSessionCwd(cwd, this.config.defaultCwd),
                     language: this.config.language,
                 });
@@ -1775,13 +1777,10 @@ export class ProcessManager extends EventEmitter {
         return this.snapshot(record);
     }
     /**
-     * 归档会先停掉还在跑的终端，但保留会话 ID 和 provider session id，恢复时用它们续上。
+     * 归档只写标记，保留正在运行的终端、会话 ID 和 provider session id。
      * 取消归档只清标记；重新拉起由恢复入口负责。
      */
     setSessionArchived(id, archived) {
-        if (archived && this.mustGet(id).status === "running") {
-            this.stop(id);
-        }
         const record = this.mustGet(id);
         record.archived = archived;
         record.archivedAt = archived ? (record.archivedAt ?? new Date().toISOString()) : null;
@@ -2354,7 +2353,7 @@ export class ProcessManager extends EventEmitter {
             input: prompt,
             cwd: record.cwd,
             language: this.config.language,
-            ai: resolveSystemAiContext(record, this.config, this.storage.getSystemSiliconEmployee()),
+            ai: resolveSystemAiContext(record, this.config, this.storage.getSystemSiliconEmployee(), this.free),
             readNativeTitle: () => readNativeSessionTitle(resolveSessionProvider(record), record.claudeSessionId, record.cwd),
             hasNativeTitle,
             onGenerating: (generating) => {

package/dist/server-employee-routes.d.ts

@@ -1,5 +1,6 @@
 import type { Express } from "express";
 import { type SiliconEmployee, type SiliconEmployeeDraft } from "./ai-team-types.js";
+import type { OpenRouterFreeModelsService } from "./openrouter-free-models.js";
 import type { WandStorage } from "./storage.js";
 import type { WandTaskAgent } from "./task-types.js";
 import type { WandConfig } from "./types.js";
@@ -17,6 +18,7 @@ export declare function registerSiliconEmployeeRoutes(app: Express, deps: {
     storage: WandStorage;
     notifyEmployeeChanged?: (employeeId: string) => void;
     config?: WandConfig;
+    free?: Pick<OpenRouterFreeModelsService, "resolveForCall">;
     /** 测试注入点：默认按系统 AI 配置真实调用模型。 */
     generateDraft?: (request: EmployeeDraftRequest) => Promise<SiliconEmployeeDraft>;
     /** 测试注入点：CLI 可用性探测。 */