@musnows/scriverse 0.8.8 → 0.9.0

package/dist/ai.js

@@ -1,6 +1,7 @@
 import { ANALYSIS_TASK_TYPES, HISTORICAL_ANALYSIS_TASK_TYPES } from "./domain.js";
 import { buildCompletionRequestBody, AI_THINKING_TYPES, isAiProviderProtocol, normalizeProviderBaseUrl, parseCompletionPayload, parseProviderModelListPage, providerCompletionEndpoint, providerModelListPageEndpoint, providerModelEndpoints, providerProtocolLabelText, providerRequestHeaders } from "./ai-protocol.js";
 import { estimateLiteLlmUsageCost } from "./ai-model-pricing.js";
+import { DEFAULT_AI_ANALYSIS_TIMEOUT_SECONDS, isLongRunningAiAnalysisTaskType, normalizeAiAnalysisTimeoutSeconds } from "./ai-analysis-timeout.js";
 import { AGENT_TOOL_RESULT_MAX_CHARS, DEFAULT_AGENT_TOOL_CALL_GLOBAL_MULTIPLIER, MIN_AGENT_TOOL_CALL_LIMIT, agentToolCallGlobalLimit, agentToolCallQuotaNoticeBudgetChars, agentToolCallQuotaUsedAfterCompact, agentToolCallSoftWarningThreshold, clampAgentToolCallGlobalMultiplier, paginateToolResultRecords, resolveMaxAgentToolCallLimit, shouldRejectAgentToolCalls, shouldRejectGlobalToolCalls, structuralToolResultRecords, withAgentToolCallQuotaNotice } from "./ai-tool-results.js";
 import { AiConnectivityTestGate, hashAiConnectivityConfiguration } from "./ai-connectivity-test.js";
 import { aiHttpRetryCount, aiHttpRetryDelayMs, normalizeAiRetryPolicy } from "./ai-retry.js";
@@ -13,7 +14,7 @@ import { assertOfficialGoogleVertexBaseUrl, fetchGoogleOAuthAccessToken, GoogleV
 import { HYBRID_SEARCH_TYPES, MAXIMUM_WORK_SEARCH_QUERY_LENGTH, buildHybridSearchSnippet, documentParagraphLineRangesFromLines, fuseHybridSearchChannels, normalizeWorkSearchQuery } from "./hybrid-search.js";
 import { logger, sanitizeError } from "./logger.js";
 import { paginated, paginationSql } from "./pagination.js";
-import { currentRequestActor } from "./request-context.js";
+import { currentRequestActor, runWithRequestActor } from "./request-context.js";
 import { aiEndpointUsesPrivateNetwork, fetchSafeAiEndpoint } from "./security.js";
 import { defaultAiConversationTitle, normalizeCharacterName } from "./store.js";
 import { composeRoleplayCurrentUserTurn, formatRoleplayScenePinText, roleplayUserTurnTitleSource } from "./roleplay-turn.js";
@@ -53,7 +54,6 @@ function connectivityTestErrorForLog(error) {
 const AUTO_RUN_MAX_ATTEMPTS = 3;
 const AUTO_RUN_RETRY_DELAYS_MS = [5_000, 30_000];
 const AI_INTERACTIVE_TIMEOUT_MS = 60_000;
-const AI_LONG_RUNNING_TIMEOUT_MS = 300_000;
 const FORCE_CONVERSATION_COMPACTION_USAGE_PERCENT = 95;
 const MIN_OUTPUT_RESERVE_TOKENS = 1_024;
 const MIN_CONTEXT_REMAINING_TOKENS = 5_000;
@@ -204,6 +204,9 @@ export function autoRunFailureDisposition(error, attemptCount) {
         pauseImmediately
     };
 }
+const DESKTOP_LOCAL_AI_RUN_LIMIT = 20;
+const DESKTOP_LOCAL_AI_RUN_RETENTION_MS = 10 * 60_000;
+const DESKTOP_LOCAL_AI_RESPONSE_MAX_BYTES = 4 * 1024 * 1024;
 const allowedParameters = new Set(["temperature", "top_p", "max_tokens", "presence_penalty", "frequency_penalty", "seed"]);
 const DEFAULT_MAX_TOKENS = 32_000;
 const MAX_MODEL_OUTPUT_TOKENS = 2_000_000;
@@ -255,6 +258,9 @@ function providerThinkingType(provider) {
     const value = stringValue(provider, "thinking_type");
     return AI_THINKING_TYPES.includes(value) ? value : "enabled";
 }
+function providerAnalysisTimeoutSeconds(provider) {
+    return normalizeAiAnalysisTimeoutSeconds(numberValue(provider, "analysis_timeout_seconds") || DEFAULT_AI_ANALYSIS_TIMEOUT_SECONDS);
+}
 function supportsMultimodalProviderProtocol(provider) {
     return ["openai-chat-completions", "openai-responses", "anthropic-messages", "google-vertex"].includes(providerProtocol(provider));
 }
@@ -1891,6 +1897,7 @@ export class AiManager {
     relationshipIndexTimer = null;
     relationshipIndexDisposed = false;
     providerSchedules = new Map();
+    desktopLocalAiRuns = new Map();
     vertexTokenCache = new GoogleVertexTokenCache();
     connectivityTestGate;
     allowPrivateAiEndpoints;
@@ -2644,6 +2651,12 @@ export class AiManager {
     }
     dispose() {
         logger.info("ai.manager.disposing", { scheduledWorks: this.autoRunTimers.size, activeTasks: this.taskControllers.size });
+        for (const run of this.desktopLocalAiRuns.values()) {
+            run.pending?.dispose();
+            run.pending?.reject(new Error("AI manager disposed"));
+            run.controller.abort(new Error("AI manager disposed"));
+        }
+        this.desktopLocalAiRuns.clear();
         if (this.autoRunStartupTimer)
             clearTimeout(this.autoRunStartupTimer);
         this.autoRunStartupTimer = null;
@@ -2916,9 +2929,17 @@ export class AiManager {
         if (protocol === "google-vertex")
             assertOfficialGoogleVertexBaseUrl(baseUrl);
         this.store.db.run(`INSERT INTO providers (id, work_id, name, base_url, protocol, encrypted_key, key_iv, key_tag, key_hint, status,
-       connection_status, concurrency_limit, rpm_limit, daily_token_quota, monthly_token_quota, max_tokens_parameter, thinking_type, note, created_at, updated_at)
-       VALUES (?, ?, ?, ?, ?, ?, ?, ?, ?, ?, 'unchecked', ?, ?, ?, ?, ?, ?, ?, ?, ?)`, providerId, PLATFORM_AI_WORK_ID, input.name, baseUrl, protocol, encrypted.encrypted, encrypted.iv, encrypted.tag, providerCredentialHint(protocol, input.apiKey), input.status ?? "disabled", input.concurrencyLimit ?? 10, input.rpmLimit ?? 10, input.dailyTokenQuota ?? null, input.monthlyTokenQuota ?? null, maxTokensParameter, input.thinkingType ?? "enabled", input.note ?? "", timestamp, timestamp);
-        this.store.audit(PLATFORM_AI_WORK_ID, "provider.created", "provider", providerId, { name: input.name, baseUrl, protocol, maxTokensParameter, thinkingType: input.thinkingType ?? "enabled" });
+       connection_status, concurrency_limit, rpm_limit, analysis_timeout_seconds, daily_token_quota, monthly_token_quota,
+       max_tokens_parameter, thinking_type, note, created_at, updated_at)
+       VALUES (?, ?, ?, ?, ?, ?, ?, ?, ?, ?, 'unchecked', ?, ?, ?, ?, ?, ?, ?, ?, ?, ?)`, providerId, PLATFORM_AI_WORK_ID, input.name, baseUrl, protocol, encrypted.encrypted, encrypted.iv, encrypted.tag, providerCredentialHint(protocol, input.apiKey), input.status ?? "disabled", input.concurrencyLimit ?? 10, input.rpmLimit ?? 10, input.analysisTimeoutSeconds ?? DEFAULT_AI_ANALYSIS_TIMEOUT_SECONDS, input.dailyTokenQuota ?? null, input.monthlyTokenQuota ?? null, maxTokensParameter, input.thinkingType ?? "enabled", input.note ?? "", timestamp, timestamp);
+        this.store.audit(PLATFORM_AI_WORK_ID, "provider.created", "provider", providerId, {
+            name: input.name,
+            baseUrl,
+            protocol,
+            maxTokensParameter,
+            thinkingType: input.thinkingType ?? "enabled",
+            analysisTimeoutSeconds: input.analysisTimeoutSeconds ?? DEFAULT_AI_ANALYSIS_TIMEOUT_SECONDS
+        });
         return this.getProvider(providerId);
     }
     listProviders() {
@@ -2978,8 +2999,9 @@ export class AiManager {
             ? nullableNumberValue(row, "monthly_token_quota")
             : input.monthlyTokenQuota;
         this.store.db.run(`UPDATE providers SET name = ?, base_url = ?, protocol = ?, encrypted_key = ?, key_iv = ?, key_tag = ?, key_hint = ?,
-       status = ?, connection_status = ?, concurrency_limit = ?, rpm_limit = ?, daily_token_quota = ?, monthly_token_quota = ?,
-       max_tokens_parameter = ?, thinking_type = ?, note = ?, updated_at = ? WHERE id = ?`, input.name ?? stringValue(row, "name"), nextBaseUrl, nextProtocol, encryptedKey, keyIv, keyTag, keyHint, input.status ?? stringValue(row, "status"), connectionStatus, input.concurrencyLimit ?? numberValue(row, "concurrency_limit"), input.rpmLimit ?? numberValue(row, "rpm_limit"), nextDailyTokenQuota, nextMonthlyTokenQuota, nextMaxTokensParameter, nextThinkingType, input.note ?? stringValue(row, "note"), now(), providerId);
+       status = ?, connection_status = ?, concurrency_limit = ?, rpm_limit = ?, analysis_timeout_seconds = ?,
+       daily_token_quota = ?, monthly_token_quota = ?,
+       max_tokens_parameter = ?, thinking_type = ?, note = ?, updated_at = ? WHERE id = ?`, input.name ?? stringValue(row, "name"), nextBaseUrl, nextProtocol, encryptedKey, keyIv, keyTag, keyHint, input.status ?? stringValue(row, "status"), connectionStatus, input.concurrencyLimit ?? numberValue(row, "concurrency_limit"), input.rpmLimit ?? numberValue(row, "rpm_limit"), input.analysisTimeoutSeconds ?? providerAnalysisTimeoutSeconds(row), nextDailyTokenQuota, nextMonthlyTokenQuota, nextMaxTokensParameter, nextThinkingType, input.note ?? stringValue(row, "note"), now(), providerId);
         this.store.audit(PLATFORM_AI_WORK_ID, "provider.updated", "provider", providerId, {
             fields: Object.keys(input).filter((key) => key !== "apiKey"),
             keyReplaced: Boolean(input.apiKey)
@@ -4228,7 +4250,7 @@ export class AiManager {
         this.store.db.run(`INSERT INTO ai_suggestions (id, call_id, work_id, chapter_id, chapter_version, task_type, instruction,
        source_text, content, action, status, created_at, created_by_user_id) VALUES (?, ?, ?, ?, ?, ?, ?, ?, ?, ?, 'pending', ?, ?)`, suggestionId, generated.callId, input.workId, chapter ? String(chapter.id) : null, chapter ? Number(chapter.versionNo) : null, input.taskType, input.instruction, effectiveInput.scope.selection ?? "", generated.content, action, now(), currentRequestActor()?.userId ?? null);
         if (input.taskType === "continue")
-            await this.runSuggestionGuard(suggestionId);
+            await this.runSuggestionGuardWithRuntime(suggestionId, undefined, effectiveInput.runtime);
         return {
             ...this.getSuggestion(suggestionId),
             outputTokens: generated.outputTokens,
@@ -4356,6 +4378,9 @@ export class AiManager {
         }
     }
     async runSuggestionGuard(suggestionId, candidateContent) {
+        return this.runSuggestionGuardWithRuntime(suggestionId, candidateContent);
+    }
+    async runSuggestionGuardWithRuntime(suggestionId, candidateContent, runtime) {
         const suggestion = this.getSuggestion(suggestionId);
         if (suggestion.taskType !== "continue" || !suggestion.chapterId) {
             throw new AppError(409, "GUARD_NOT_APPLICABLE", "只有续写建议可以运行一致性守卫");
@@ -4387,7 +4412,8 @@ export class AiManager {
                     "续写候选：",
                     content
                 ].join("\n\n"),
-                extraSystemPrompt: "你是续写一致性守卫。必须逐项对照人物状态、地点、时间、世界观硬约束、章节大纲和未回收伏笔。"
+                extraSystemPrompt: "你是续写一致性守卫。必须逐项对照人物状态、地点、时间、世界观硬约束、章节大纲和未回收伏笔。",
+                ...(runtime ? { runtime } : {})
             });
             const issues = parseGuardIssues(generated.content);
             return this.store.createContinuationGuard({
@@ -4501,43 +4527,14 @@ export class AiManager {
     }
     listCalls(workId) {
         this.store.getWork(workId);
-        return this.store.db.all("SELECT * FROM ai_calls WHERE work_id = ? ORDER BY created_at DESC LIMIT 200", workId).map((row) => ({
-            id: stringValue(row, "id"),
-            workId: stringValue(row, "work_id"),
-            taskId: row.task_id === null ? null : stringValue(row, "task_id"),
-            taskType: stringValue(row, "task_type"),
-            provider: this.getProvider(stringValue(row, "provider_id")),
-            model: this.getModel(stringValue(row, "model_id")),
-            contextScope: json(stringValue(row, "context_scope_json"), {}),
-            parameters: json(stringValue(row, "parameters_json"), {}),
-            status: stringValue(row, "status"),
-            failure: row.failure === null ? null : stringValue(row, "failure"),
-            inputChars: numberValue(row, "input_chars"),
-            outputChars: numberValue(row, "output_chars"),
-            createdAt: stringValue(row, "created_at"),
-            completedAt: row.completed_at === null ? null : stringValue(row, "completed_at")
-        }));
+        return this.store.db.all("SELECT * FROM ai_calls WHERE work_id = ? ORDER BY created_at DESC LIMIT 200", workId)
+            .map((row) => this.mapCall(row));
     }
     listCallsPage(workId, pagination) {
         this.store.getWork(workId);
         const page = paginationSql(pagination);
         const rows = this.store.db.all(`SELECT * FROM ai_calls WHERE work_id = ? ORDER BY created_at DESC${page.sql}`, workId, ...page.params);
-        return paginated(rows.map((row) => ({
-            id: stringValue(row, "id"),
-            workId: stringValue(row, "work_id"),
-            taskId: row.task_id === null ? null : stringValue(row, "task_id"),
-            taskType: stringValue(row, "task_type"),
-            provider: this.getProvider(stringValue(row, "provider_id")),
-            model: this.getModel(stringValue(row, "model_id")),
-            contextScope: json(stringValue(row, "context_scope_json"), {}),
-            parameters: json(stringValue(row, "parameters_json"), {}),
-            status: stringValue(row, "status"),
-            failure: row.failure === null ? null : stringValue(row, "failure"),
-            inputChars: numberValue(row, "input_chars"),
-            outputChars: numberValue(row, "output_chars"),
-            createdAt: stringValue(row, "created_at"),
-            completedAt: row.completed_at === null ? null : stringValue(row, "completed_at")
-        })), pagination);
+        return paginated(rows.map((row) => this.mapCall(row)), pagination);
     }
     getTaskTrace(taskId) {
         this.store.getTaskWorkId(taskId);
@@ -4782,6 +4779,9 @@ export class AiManager {
     }
     getContextUsage(input) {
         const { model } = this.resolveModel(input.workId, input.taskType, input.modelId);
+        return this.contextUsageForModel(input, model);
+    }
+    contextUsageForModel(input, model) {
         const budget = this.contextBudget(input, model);
         const conversation = budget.conversation;
         const contextPlan = this.buildContextPlan(input, model, budget);
@@ -4835,8 +4835,235 @@ export class AiManager {
             degradedContextBlocks: contextPlan.degradedBlockIds.length
         };
     }
+    async startDesktopLocalAiRun(input, actorScope, actor, permissions) {
+        this.pruneDesktopLocalAiRuns();
+        if (this.desktopLocalAiRuns.size >= DESKTOP_LOCAL_AI_RUN_LIMIT) {
+            throw new AppError(429, "DESKTOP_LOCAL_AI_RUN_LIMIT", "当前正在处理的 Desktop 本地 AI 请求过多，请稍后再试");
+        }
+        const { provider, model } = this.desktopLocalAiRuntimeRows(input.runtimeModel);
+        const imageAttachments = await this.prepareChatImageAttachmentsForModel(input.workId, model, provider, input.imageAttachmentIds ?? [], permissions);
+        const runId = id("desktop-local-ai-run");
+        const timestamp = Date.now();
+        const run = {
+            id: runId,
+            workId: input.workId,
+            actorScope,
+            actor,
+            status: "running",
+            createdAt: timestamp,
+            updatedAt: timestamp,
+            controller: new AbortController(),
+            contextUsage: null,
+            pending: null,
+            result: null,
+            error: null
+        };
+        const runtime = {
+            provider,
+            model,
+            localModelId: input.runtimeModel.id,
+            completionTransport: (request) => this.awaitDesktopLocalAiCompletion(run, request)
+        };
+        this.desktopLocalAiRuns.set(runId, run);
+        const updateContextUsage = (contextUsage) => {
+            run.contextUsage = contextUsage;
+            run.updatedAt = Date.now();
+        };
+        void Promise.resolve().then(() => runWithRequestActor(actor, () => this.createSuggestion({
+            workId: input.workId,
+            taskType: input.taskType,
+            instruction: input.instruction,
+            scope: input.scope,
+            modelId: input.runtimeModel.id,
+            signal: run.controller.signal,
+            runtime,
+            onPrepared: updateContextUsage,
+            onContextCompacted: (event) => updateContextUsage(event.contextUsage),
+            ...(input.conversationId ? { conversationId: input.conversationId } : {}),
+            ...(input.excludeConversationMessageId ? { excludeConversationMessageId: input.excludeConversationMessageId } : {}),
+            ...(imageAttachments.length > 0 ? { imageAttachments } : {}),
+            ...(input.sceneDirection ? { sceneDirection: input.sceneDirection } : {})
+        }))).then((result) => {
+            if (run.status === "cancelled")
+                return;
+            run.status = "completed";
+            run.result = result;
+            run.contextUsage = result.contextUsage && typeof result.contextUsage === "object" && !Array.isArray(result.contextUsage)
+                ? result.contextUsage
+                : run.contextUsage;
+            run.updatedAt = Date.now();
+        }).catch((error) => {
+            if (run.status === "cancelled")
+                return;
+            const appError = error instanceof AppError ? error : null;
+            run.status = "failed";
+            run.error = {
+                status: appError?.status ?? 502,
+                code: appError?.code ?? "AI_CALL_FAILED",
+                message: appError?.message ?? "AI 调用失败"
+            };
+            run.updatedAt = Date.now();
+        });
+        return this.desktopLocalAiRunStatus(runId, input.workId, actorScope);
+    }
+    desktopLocalAiRunStatus(runId, workId, actorScope) {
+        this.pruneDesktopLocalAiRuns();
+        const run = this.desktopLocalAiRun(runId, workId, actorScope);
+        return {
+            id: run.id,
+            status: run.status,
+            ...(run.contextUsage ? { contextUsage: run.contextUsage } : {}),
+            ...(run.pending ? { completion: run.pending.request } : {}),
+            ...(run.result ? { result: run.result } : {}),
+            ...(run.error ? { error: run.error } : {})
+        };
+    }
+    submitDesktopLocalAiCompletion(runId, workId, actorScope, input) {
+        const run = this.desktopLocalAiRun(runId, workId, actorScope);
+        const pending = run.pending;
+        if (!pending || run.status !== "awaiting-completion") {
+            throw new AppError(409, "DESKTOP_LOCAL_AI_NOT_AWAITING", "当前 Desktop 本地 AI 请求不等待模型响应");
+        }
+        if (pending.request.requestId !== input.requestId) {
+            throw new AppError(409, "DESKTOP_LOCAL_AI_REQUEST_MISMATCH", "Desktop 本地 AI 响应与当前请求不匹配");
+        }
+        if (Buffer.byteLength(input.body, "utf8") > DESKTOP_LOCAL_AI_RESPONSE_MAX_BYTES) {
+            throw new AppError(413, "DESKTOP_LOCAL_AI_RESPONSE_TOO_LARGE", "Desktop 本地 AI 响应过大");
+        }
+        run.pending = null;
+        run.status = "running";
+        run.updatedAt = Date.now();
+        pending.dispose();
+        pending.resolve({
+            status: input.status,
+            body: input.body,
+            retryAfter: input.retryAfter ?? null
+        });
+        return this.desktopLocalAiRunStatus(runId, workId, actorScope);
+    }
+    cancelDesktopLocalAiRun(runId, workId, actorScope) {
+        const run = this.desktopLocalAiRun(runId, workId, actorScope);
+        if (run.status === "completed" || run.status === "failed" || run.status === "cancelled") {
+            return this.desktopLocalAiRunStatus(runId, workId, actorScope);
+        }
+        run.status = "cancelled";
+        run.updatedAt = Date.now();
+        run.pending?.dispose();
+        run.pending?.reject(new Error("Desktop local AI run cancelled"));
+        run.pending = null;
+        run.controller.abort(new Error("Desktop local AI run cancelled"));
+        return this.desktopLocalAiRunStatus(runId, workId, actorScope);
+    }
+    desktopLocalAiRuntimeRows(input) {
+        const timestamp = now();
+        return {
+            provider: {
+                id: input.providerId,
+                work_id: PLATFORM_AI_WORK_ID,
+                name: input.providerName,
+                base_url: "",
+                protocol: input.protocol,
+                encrypted_key: "",
+                key_iv: "",
+                key_tag: "",
+                key_hint: "",
+                status: "enabled",
+                connection_status: "success",
+                max_tokens_parameter: input.maxTokensParameter,
+                thinking_type: input.thinkingType,
+                concurrency_limit: input.concurrencyLimit,
+                rpm_limit: input.rpmLimit,
+                analysis_timeout_seconds: input.analysisTimeoutSeconds,
+                daily_token_quota: null,
+                monthly_token_quota: null,
+                default_model_id: input.id,
+                note: input.note,
+                last_error: null,
+                last_success_at: timestamp,
+                created_at: timestamp,
+                updated_at: timestamp,
+                desktop_local: 1
+            },
+            model: {
+                id: input.id,
+                provider_id: input.providerId,
+                display_name: input.displayName,
+                model_id: input.modelId,
+                enabled: 1,
+                purposes_json: JSON.stringify(input.purposes),
+                context_note: input.contextNote,
+                context_window: input.contextWindow,
+                output_note: input.outputNote,
+                preset_json: JSON.stringify(input.preset),
+                thinking_enabled: input.thinkingEnabled ? 1 : 0,
+                thinking_effort: input.thinkingEffort,
+                multimodal_enabled: input.multimodalEnabled ? 1 : 0,
+                note: input.note,
+                created_at: timestamp,
+                updated_at: timestamp,
+                desktop_local: 1
+            }
+        };
+    }
+    desktopLocalAiRun(runId, workId, actorScope) {
+        const run = this.desktopLocalAiRuns.get(runId);
+        if (!run || run.workId !== workId || run.actorScope !== actorScope)
+            throw notFound("Desktop 本地 AI 请求");
+        return run;
+    }
+    awaitDesktopLocalAiCompletion(run, request) {
+        if (run.controller.signal.aborted)
+            return Promise.reject(new Error("Desktop local AI run cancelled"));
+        if (run.pending)
+            return Promise.reject(new Error("Desktop local AI run already has a pending completion"));
+        return new Promise((resolve, reject) => {
+            const timeout = setTimeout(() => {
+                if (run.pending?.request.requestId !== request.requestId)
+                    return;
+                run.pending = null;
+                run.status = "running";
+                run.updatedAt = Date.now();
+                run.controller.signal.removeEventListener("abort", onAbort);
+                reject(new Error(`AI 请求超时（${Math.round(request.timeoutMs / 1_000)} 秒）`));
+            }, request.timeoutMs);
+            const onAbort = () => {
+                if (run.pending?.request.requestId !== request.requestId)
+                    return;
+                run.pending = null;
+                clearTimeout(timeout);
+                reject(new Error("Desktop local AI run cancelled"));
+            };
+            const dispose = () => {
+                clearTimeout(timeout);
+                run.controller.signal.removeEventListener("abort", onAbort);
+            };
+            run.controller.signal.addEventListener("abort", onAbort, { once: true });
+            run.pending = {
+                request,
+                resolve: (response) => {
+                    dispose();
+                    resolve(response);
+                },
+                reject: (error) => {
+                    dispose();
+                    reject(error);
+                },
+                dispose
+            };
+            run.status = "awaiting-completion";
+            run.updatedAt = Date.now();
+        });
+    }
+    pruneDesktopLocalAiRuns() {
+        const cutoff = Date.now() - DESKTOP_LOCAL_AI_RUN_RETENTION_MS;
+        for (const [runId, run] of this.desktopLocalAiRuns) {
+            if (run.updatedAt >= cutoff || run.status === "running" || run.status === "awaiting-completion")
+                continue;
+            this.desktopLocalAiRuns.delete(runId);
+        }
+    }
     completionContextUsage(input, model, messages, tools, reportedUsage) {
-        const baseUsage = this.getContextUsage(input);
+        const baseUsage = this.contextUsageForModel(input, model);
         const contextWindow = numberValue(model, "context_window") || DEFAULT_CONTEXT_WINDOW;
         const serializedMessageTokens = estimateCompletionMessageTokens(messages);
         const systemPromptTokens = messages
@@ -4958,12 +5185,15 @@ export class AiManager {
         return this.mergeInstructionEntityMatches(input.scope, matches);
     }
     async prepareChatImageAttachments(workId, modelId, attachmentIds, permissions) {
+        const { model, provider } = this.resolveModel(workId, "chat", modelId);
+        return this.prepareChatImageAttachmentsForModel(workId, model, provider, attachmentIds, permissions);
+    }
+    async prepareChatImageAttachmentsForModel(workId, model, provider, attachmentIds, permissions) {
         const ids = [...new Set(attachmentIds.map((attachmentId) => String(attachmentId).trim()).filter(Boolean))];
         if (ids.length === 0)
             return [];
         if (ids.length > 4)
             throw new AppError(400, "AI_CHAT_IMAGE_LIMIT", "一次最多添加 4 张图片附件");
-        const { model, provider } = this.resolveModel(workId, "chat", modelId);
         if (!boolValue(model, "multimodal_enabled")) {
             throw new AppError(400, "MODEL_NOT_MULTIMODAL", "当前选择的模型不是多模态模型，无法处理图片附件");
         }
@@ -5006,11 +5236,10 @@ export class AiManager {
         }
         return prepared;
     }
-    async prepareConversationImageAttachments(workId, modelId, conversation) {
+    async prepareConversationImageAttachments(workId, model, provider, conversation) {
         const preparedByMessage = new Map();
         if (!conversation)
             return preparedByMessage;
-        const { model, provider } = this.resolveModel(workId, "chat", modelId);
         if (!boolValue(model, "multimodal_enabled") || !supportsMultimodalProviderProtocol(provider)) {
             return preparedByMessage;
         }
@@ -5023,7 +5252,7 @@ export class AiManager {
                 : [];
             if (ids.length === 0)
                 continue;
-            preparedByMessage.set(message.id, await this.prepareChatImageAttachments(workId, modelId, ids, permissions));
+            preparedByMessage.set(message.id, await this.prepareChatImageAttachmentsForModel(workId, model, provider, ids, permissions));
         }
         return preparedByMessage;
     }
@@ -5486,8 +5715,10 @@ export class AiManager {
             ? this.roleplayCharacterId(workId, conversationId)
             : roleplayCharacterIdOverride;
         const permissions = this.store.getWork(workId).modulePermissions;
+        const requested = requestedToolIds ? new Set(requestedToolIds) : null;
+        if (requested?.size === 0)
+            return [];
         if (roleplayCharacterId) {
-            const requested = requestedToolIds ? new Set(requestedToolIds) : null;
             if (!canReadWorkModule(permissions, "characters"))
                 return [];
             const roleplayTools = [];
@@ -5518,7 +5749,6 @@ export class AiManager {
             : this.store.getWorkAiSettings(workId).agentTools;
         const enabled = new Set(sourceTools
             .filter((item) => typeof item === "string" && CONFIGURED_AGENT_TOOL_IDS.includes(item)));
-        const requested = requestedToolIds ? new Set(requestedToolIds) : null;
         return CONFIGURED_AGENT_TOOL_IDS.filter((toolId) => enabled.has(toolId)
             && (!requested || requested.has(toolId))
             && this.canReadWithAgentTool(permissions, toolId));
@@ -6722,13 +6952,13 @@ export class AiManager {
             max_tokens: Math.min(Number(parameters.max_tokens) || DEFAULT_MAX_TOKENS, contextWindow - inputTokens)
         };
     }
-    constrainParametersForTokenQuota(workId, provider, messages, parameters, tools = [], additionalUsedTokens = 0) {
+    constrainParametersForTokenQuota(workId, provider, messages, parameters, tools = [], additionalUsedTokens = 0, includeProviderQuota = true) {
         const workStatus = this.getWorkTokenQuotaStatus(workId);
-        const providerStatus = this.getProviderTokenQuotaStatus(stringValue(provider, "id"));
+        const providerStatus = includeProviderQuota ? this.getProviderTokenQuotaStatus(stringValue(provider, "id")) : null;
         const dailyTokenQuota = workStatus.dailyTokenQuota === null ? null : Number(workStatus.dailyTokenQuota);
         const monthlyTokenQuota = workStatus.monthlyTokenQuota === null ? null : Number(workStatus.monthlyTokenQuota);
-        const providerDailyTokenQuota = providerStatus.dailyTokenQuota === null ? null : Number(providerStatus.dailyTokenQuota);
-        const providerMonthlyTokenQuota = providerStatus.monthlyTokenQuota === null ? null : Number(providerStatus.monthlyTokenQuota);
+        const providerDailyTokenQuota = providerStatus?.dailyTokenQuota === null || !providerStatus ? null : Number(providerStatus.dailyTokenQuota);
+        const providerMonthlyTokenQuota = providerStatus?.monthlyTokenQuota === null || !providerStatus ? null : Number(providerStatus.monthlyTokenQuota);
         if (dailyTokenQuota === null && monthlyTokenQuota === null && providerDailyTokenQuota === null && providerMonthlyTokenQuota === null)
             return parameters;
         const additionalTokens = Math.max(0, additionalUsedTokens);
@@ -6753,8 +6983,10 @@ export class AiManager {
                 resetsAt: String(workStatus.monthlyResetsAt),
                 startedAt: String(workStatus.monthStartedAt),
                 timezone: String(workStatus.timezone)
-            },
-            {
+            }
+        ];
+        if (providerStatus) {
+            quotas.push({
                 scope: "provider",
                 period: "daily",
                 quota: providerDailyTokenQuota,
@@ -6764,8 +6996,7 @@ export class AiManager {
                 timezone: String(providerStatus.timezone),
                 providerId: stringValue(provider, "id"),
                 providerName: stringValue(provider, "name")
-            },
-            {
+            }, {
                 scope: "provider",
                 period: "monthly",
                 quota: providerMonthlyTokenQuota,
@@ -6775,8 +7006,8 @@ export class AiManager {
                 timezone: String(providerStatus.timezone),
                 providerId: stringValue(provider, "id"),
                 providerName: stringValue(provider, "name")
-            }
-        ];
+            });
+        }
         for (const item of quotas) {
             if (item.quota === null)
                 continue;
@@ -6832,8 +7063,8 @@ export class AiManager {
             ? this.store.getAiConversationContext(input.conversationId, input.workId, input.excludeConversationMessageId)
             : null;
         const generationRoleplayCharacterId = this.roleplayCharacterIdFromConversation(input.workId, conversation);
-        const { model, provider } = this.resolveModel(input.workId, input.taskType, input.modelId);
-        const conversationImageAttachments = await this.prepareConversationImageAttachments(input.workId, String(model.id), conversation);
+        const { model, provider } = input.runtime ?? this.resolveModel(input.workId, input.taskType, input.modelId);
+        const conversationImageAttachments = await this.prepareConversationImageAttachments(input.workId, model, provider, conversation);
         const preset = safeJsonObject(stringValue(model, "preset_json"));
         const requestedParameters = {
             ...this.sanitizeParameters({ ...preset, ...(input.parameters ?? {}) }, stringValue(model, "model_id")),
@@ -6880,14 +7111,24 @@ export class AiManager {
                 modelId: stringValue(model, "id")
             });
         }
-        parameters = this.constrainParametersForTokenQuota(input.workId, provider, messages, parameters, tools);
+        parameters = this.constrainParametersForTokenQuota(input.workId, provider, messages, parameters, tools, 0, input.runtime === undefined);
+        input.onPrepared?.(this.completionContextUsage(effectiveInput, model, messages, tools));
         const completionMessages = [...messages];
         const callId = id("call");
         const timestamp = now();
         const traceRounds = [];
+        const storedParameters = input.runtime
+            ? {
+                ...parameters,
+                __desktopLocalAi: {
+                    provider: this.mapProvider(provider),
+                    model: this.mapModel(model)
+                }
+            }
+            : parameters;
         this.store.db.transaction(() => {
             this.store.db.run(`INSERT INTO ai_calls (id, work_id, task_id, task_type, provider_id, model_id, context_scope_json, parameters_json,
-         status, input_chars, created_at, created_by_user_id) VALUES (?, ?, ?, ?, ?, ?, ?, ?, 'running', ?, ?, ?)`, callId, input.workId, input.taskId ?? null, input.taskType, stringValue(provider, "id"), stringValue(model, "id"), JSON.stringify(input.scope), JSON.stringify(parameters), context.length + input.instruction.length, timestamp, currentRequestActor()?.userId ?? null);
+         status, input_chars, created_at, created_by_user_id) VALUES (?, ?, ?, ?, ?, ?, ?, ?, 'running', ?, ?, ?)`, callId, input.workId, input.taskId ?? null, input.taskType, stringValue(provider, "id"), stringValue(model, "id"), JSON.stringify(input.scope), JSON.stringify(storedParameters), context.length + input.instruction.length, timestamp, currentRequestActor()?.userId ?? null);
             if (input.taskId) {
                 this.store.db.run(`INSERT INTO ai_call_traces (call_id, task_id, initial_messages_json, rounds_json, source_refs_json, created_at, updated_at)
            VALUES (?, ?, ?, '[]', ?, ?, ?)`, callId, input.taskId, JSON.stringify(sanitizeCompletionTraceMessages(messages)), JSON.stringify(taskTraceSourceRefs(messages, [])), timestamp, timestamp);
@@ -6937,11 +7178,16 @@ export class AiManager {
             return "mixed";
         };
         try {
-            const { accessToken, credentialSecret } = await this.resolveProviderAccessToken(provider);
-            activeSecrets = [credentialSecret, accessToken];
-            const endpoint = providerCompletionEndpoint(stringValue(provider, "base_url"), protocol);
-            const timeoutMs = input.taskType === "book-analysis" || input.taskType === "relationship-analysis"
-                ? AI_LONG_RUNNING_TIMEOUT_MS
+            let accessToken = "";
+            let endpoint = "";
+            if (!input.runtime) {
+                const credential = await this.resolveProviderAccessToken(provider);
+                accessToken = credential.accessToken;
+                activeSecrets = [credential.credentialSecret, credential.accessToken];
+                endpoint = providerCompletionEndpoint(stringValue(provider, "base_url"), protocol);
+            }
+            const timeoutMs = isLongRunningAiAnalysisTaskType(input.taskType)
+                ? providerAnalysisTimeoutSeconds(provider) * 1_000
                 : AI_INTERACTIVE_TIMEOUT_MS;
             const legacyMaximumAttempts = Math.round(clamp(input.maxAttempts ?? 3, 1, 5));
             const maximumAttempts = Math.max(legacyMaximumAttempts, this.retryPolicy.retryCount + 1, this.retryPolicy.backoffRetryCount + 1);
@@ -6957,11 +7203,11 @@ export class AiManager {
                 const requestParameters = options.parameters ?? parameters;
                 const purpose = options.purpose ?? "generation";
                 const requestTools = toolChoice === "auto" ? tools : [];
-                const streamResponse = Boolean(onDelta) && purpose === "generation";
+                const streamResponse = !input.runtime && Boolean(onDelta) && purpose === "generation";
                 const processRound = streamResponse ? streamingGenerationRound + 1 : 0;
                 if (streamResponse)
                     streamingGenerationRound = processRound;
-                const roundParameters = this.constrainParametersForTokenQuota(input.workId, provider, requestMessages, this.constrainParametersForContext(model, requestMessages, requestParameters, requestTools), requestTools, trackedInputTokens + trackedOutputTokens);
+                const roundParameters = this.constrainParametersForTokenQuota(input.workId, provider, requestMessages, this.constrainParametersForContext(model, requestMessages, requestParameters, requestTools), requestTools, trackedInputTokens + trackedOutputTokens, input.runtime === undefined);
                 const traceRound = {
                     round: traceRounds.length + 1,
                     requestedAt: now(),
@@ -6976,6 +7222,16 @@ export class AiManager {
                     attempts: [],
                     toolExecutions: []
                 };
+                const completionRequestBody = buildCompletionRequestBody({
+                    protocol,
+                    model: stringValue(model, "model_id"),
+                    messages: requestMessages,
+                    parameters: roundParameters,
+                    maxTokensParameter: providerMaxTokensParameter(provider),
+                    tools: requestTools,
+                    toolChoice,
+                    ...(streamResponse ? { stream: true } : {})
+                });
                 traceRounds.push(traceRound);
                 saveTrace();
                 let streamedThinkingStep = null;
@@ -6997,6 +7253,31 @@ export class AiManager {
                     let streamedRoundContent = "";
                     try {
                         const candidate = await this.scheduleProviderRequest(provider, input.signal, async () => {
+                            if (input.runtime) {
+                                const response = await input.runtime.completionTransport({
+                                    requestId: id("desktop-local-ai-completion"),
+                                    localModelId: input.runtime.localModelId,
+                                    taskType: input.taskType,
+                                    purpose,
+                                    body: completionRequestBody,
+                                    timeoutMs
+                                });
+                                if (response.status < 200 || response.status >= 300) {
+                                    return {
+                                        ok: false,
+                                        status: response.status,
+                                        body: response.body,
+                                        retryAfter: response.retryAfter
+                                    };
+                                }
+                                try {
+                                    const payload = parseCompletionPayload(protocol, JSON.parse(response.body));
+                                    return { ok: true, status: response.status, payload, delivery: "json" };
+                                }
+                                catch {
+                                    throw new Error(`${providerProtocolLabelText(protocol)} returned invalid JSON: ${response.body.slice(0, 500)}`);
+                                }
+                            }
                             const controller = new AbortController();
                             const forwardAbort = () => controller.abort(input.signal?.reason);
                             if (input.signal?.aborted)
@@ -7015,16 +7296,7 @@ export class AiManager {
                                 const response = await this.outboundFetch(endpoint, {
                                     method: "POST",
                                     headers: providerRequestHeaders(protocol, accessToken, streamResponse ? "text/event-stream" : "application/json"),
-                                    body: JSON.stringify(buildCompletionRequestBody({
-                                        protocol,
-                                        model: stringValue(model, "model_id"),
-                                        messages: requestMessages,
-                                        parameters: roundParameters,
-                                        maxTokensParameter: providerMaxTokensParameter(provider),
-                                        tools: requestTools,
-                                        toolChoice,
-                                        ...(streamResponse ? { stream: true } : {})
-                                    })),
+                                    body: JSON.stringify(completionRequestBody),
                                     signal: controller.signal
                                 });
                                 responseReceived = true;
@@ -11677,19 +11949,22 @@ export class AiManager {
         return row;
     }
     mapProvider(row) {
+        const desktopLocal = boolValue(row, "desktop_local");
         let apiKeyHint = stringValue(row, "key_hint");
-        try {
-            const secret = this.decryptKey(row);
-            apiKeyHint = providerCredentialHint(providerProtocol(row), secret);
-        }
-        catch {
-            // 凭据无法解密时保留数据库中的旧掩码，避免影响供应商列表展示。
+        if (!desktopLocal) {
+            try {
+                const secret = this.decryptKey(row);
+                apiKeyHint = providerCredentialHint(providerProtocol(row), secret);
+            }
+            catch {
+                // 凭据无法解密时保留数据库中的旧掩码，避免影响供应商列表展示。
+            }
         }
         return {
             id: stringValue(row, "id"),
-            scope: "platform",
+            scope: desktopLocal ? "local" : "platform",
             name: stringValue(row, "name"),
-            baseUrl: stringValue(row, "base_url"),
+            baseUrl: desktopLocal ? "" : stringValue(row, "base_url"),
             protocol: providerProtocol(row),
             maxTokensParameter: providerMaxTokensParameter(row),
             thinkingType: providerThinkingType(row),
@@ -11698,6 +11973,7 @@ export class AiManager {
             connectionStatus: stringValue(row, "connection_status"),
             concurrencyLimit: numberValue(row, "concurrency_limit") || 10,
             rpmLimit: numberValue(row, "rpm_limit") || 10,
+            analysisTimeoutSeconds: providerAnalysisTimeoutSeconds(row),
             dailyTokenQuota: nullableNumberValue(row, "daily_token_quota"),
             monthlyTokenQuota: nullableNumberValue(row, "monthly_token_quota"),
             defaultModelId: row.default_model_id === null ? null : stringValue(row, "default_model_id"),
@@ -11709,8 +11985,10 @@ export class AiManager {
         };
     }
     mapModel(row) {
+        const desktopLocal = boolValue(row, "desktop_local");
         return {
             id: stringValue(row, "id"),
+            ...(desktopLocal ? { scope: "local" } : {}),
             providerId: stringValue(row, "provider_id"),
             displayName: stringValue(row, "display_name"),
             modelId: stringValue(row, "model_id"),
@@ -11722,15 +12000,57 @@ export class AiManager {
             thinkingEnabled: boolValue(row, "thinking_enabled"),
             thinkingEffort: stringValue(row, "thinking_effort") || "default",
             multimodalEnabled: boolValue(row, "multimodal_enabled"),
-            imageToolDefault: String(this.store.getPlatformAiSettings().imageToolModelId ?? "") === stringValue(row, "id"),
+            imageToolDefault: !desktopLocal && String(this.store.getPlatformAiSettings().imageToolModelId ?? "") === stringValue(row, "id"),
             enabled: boolValue(row, "enabled"),
             note: stringValue(row, "note"),
             createdAt: stringValue(row, "created_at"),
             updatedAt: stringValue(row, "updated_at")
         };
     }
+    aiCallTarget(row) {
+        const parameters = safeJsonObject(stringValue(row, "parameters_json"));
+        const desktopLocal = parameters.__desktopLocalAi;
+        if (desktopLocal && typeof desktopLocal === "object" && !Array.isArray(desktopLocal)) {
+            const snapshot = desktopLocal;
+            if (snapshot.provider && typeof snapshot.provider === "object" && !Array.isArray(snapshot.provider)
+                && snapshot.model && typeof snapshot.model === "object" && !Array.isArray(snapshot.model)) {
+                return {
+                    provider: structuredClone(snapshot.provider),
+                    model: structuredClone(snapshot.model)
+                };
+            }
+        }
+        return {
+            provider: this.getProvider(stringValue(row, "provider_id")),
+            model: this.getModel(stringValue(row, "model_id"))
+        };
+    }
+    publicAiCallParameters(row) {
+        const { __desktopLocalAi: _desktopLocalAi, ...parameters } = safeJsonObject(stringValue(row, "parameters_json"));
+        return parameters;
+    }
+    mapCall(row) {
+        const target = this.aiCallTarget(row);
+        return {
+            id: stringValue(row, "id"),
+            workId: stringValue(row, "work_id"),
+            taskId: row.task_id === null ? null : stringValue(row, "task_id"),
+            taskType: stringValue(row, "task_type"),
+            provider: target.provider,
+            model: target.model,
+            contextScope: json(stringValue(row, "context_scope_json"), {}),
+            parameters: this.publicAiCallParameters(row),
+            status: stringValue(row, "status"),
+            failure: row.failure === null ? null : stringValue(row, "failure"),
+            inputChars: numberValue(row, "input_chars"),
+            outputChars: numberValue(row, "output_chars"),
+            createdAt: stringValue(row, "created_at"),
+            completedAt: row.completed_at === null ? null : stringValue(row, "completed_at")
+        };
+    }
     mapSuggestion(row) {
-        const call = this.store.db.get("SELECT provider_id, model_id FROM ai_calls WHERE id = ?", stringValue(row, "call_id"));
+        const call = this.store.db.get("SELECT provider_id, model_id, parameters_json FROM ai_calls WHERE id = ?", stringValue(row, "call_id"));
+        const target = call ? this.aiCallTarget(call) : null;
         const guard = this.store.getLatestContinuationGuard(stringValue(row, "id"));
         return {
             id: stringValue(row, "id"),
@@ -11746,8 +12066,8 @@ export class AiManager {
             status: stringValue(row, "status"),
             outputTokens: estimateAiTokens(stringValue(row, "content")),
             guard,
-            provider: call ? this.getProvider(stringValue(call, "provider_id")) : null,
-            model: call ? this.getModel(stringValue(call, "model_id")) : null,
+            provider: target?.provider ?? null,
+            model: target?.model ?? null,
             createdAt: stringValue(row, "created_at"),
             decidedAt: row.decided_at === null ? null : stringValue(row, "decided_at")
         };