@musnows/scriverse 0.7.6 → 0.7.7

package/dist/ai.js

@@ -51,6 +51,8 @@ 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;
 const analysisTaskTypes = new Set(ANALYSIS_TASK_TYPES);
 const interactiveStreamErrorCodes = new Set([
     "AI_STREAM_IDLE_TIMEOUT",
@@ -173,6 +175,7 @@ export function autoRunFailureDisposition(error, attemptCount) {
 }
 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;
 const DEFAULT_CONTEXT_WINDOW = 128_000;
 const RELATIONSHIP_MAX_FUZZY_REFERENCES = 32;
 const RELATIONSHIP_MAX_FUZZY_SOURCES = 200;
@@ -405,7 +408,7 @@ function sanitizeCompletionTraceResponse(value) {
 }
 const MAX_AGENT_TOOL_CALLS = 12;
 const TOOL_CONTEXT_COMPACT_MAX_TOKENS = 1_024;
-const TOOL_CONTEXT_RESPONSE_RESERVE_TOKENS = 512;
+const TOOL_CONTEXT_RESPONSE_RESERVE_TOKENS = MIN_OUTPUT_RESERVE_TOKENS;
 const IMAGE_TOOL_MAX_BYTES = 30 * 1024 * 1024;
 const IMAGE_TOOL_MAX_OUTPUT_TOKENS = 8_192;
 const agentToolCursor = z.number().int().min(0).max(100_000).default(0);
@@ -738,7 +741,7 @@ function completionPayloadOutputText(payload) {
 }
 function normalizeModelPreset(input, modelId = "") {
     const maxTokens = typeof input.max_tokens === "number" && Number.isFinite(input.max_tokens)
-        ? Math.round(clamp(input.max_tokens, 1, 32_768))
+        ? Math.round(clamp(input.max_tokens, 1, MAX_MODEL_OUTPUT_TOKENS))
         : DEFAULT_MAX_TOKENS;
     const temperature = input.temperature;
     const defaultTemperature = isKimiModelId(modelId) && !(typeof temperature === "number" && Number.isFinite(temperature))
@@ -3571,7 +3574,7 @@ export class AiManager {
         const contextWindow = numberValue(model, "context_window") || DEFAULT_CONTEXT_WINDOW;
         const preset = safeJsonObject(stringValue(model, "preset_json"));
         const configuredOutputTokens = typeof preset.max_tokens === "number" ? preset.max_tokens : DEFAULT_MAX_TOKENS;
-        const outputReserveTokens = Math.max(512, Math.min(configuredOutputTokens, Math.floor(contextWindow * 0.25), contextWindow - 512));
+        const outputReserveTokens = Math.max(MIN_OUTPUT_RESERVE_TOKENS, Math.min(configuredOutputTokens, Math.floor(contextWindow * 0.25), contextWindow - MIN_OUTPUT_RESERVE_TOKENS));
         const availableInputTokens = Math.max(256, contextWindow - outputReserveTokens - 512);
         const conversation = input.conversationId
             ? this.store.getAiConversationContext(input.conversationId, input.workId, input.excludeConversationMessageId)
@@ -3590,6 +3593,7 @@ export class AiManager {
             - functionTokens);
         return {
             contextWindow,
+            configuredOutputTokens,
             outputReserveTokens,
             availableInputTokens,
             conversation,
@@ -3619,7 +3623,10 @@ export class AiManager {
         const threshold = Math.min(90, Math.max(50, Number(this.store.getWorkAiSettings(input.workId).contextCompactThreshold) || 85));
         const conversation = budget.conversation;
         const conversationUsagePercent = Number(budget.conversationUsagePercent) || 0;
+        const configuredOutputTokens = Number(budget.configuredOutputTokens) || DEFAULT_MAX_TOKENS;
+        const maxOutputUsagePercent = Math.min(100, Math.round(configuredOutputTokens / contextWindow * 100));
         const compactableMessageCount = Math.max(0, (conversation?.messages.length ?? 0) - 2);
+        const contextFallbackReached = remainingTokens <= MIN_CONTEXT_REMAINING_TOKENS;
         return {
             modelId: stringValue(model, "id"),
             contextWindow,
@@ -3628,8 +3635,12 @@ export class AiManager {
             conversationTokens: Number(budget.conversationTokens),
             conversationBudgetTokens: Number(budget.conversationBudgetTokens),
             conversationUsagePercent,
+            maxOutputTokens: configuredOutputTokens,
+            maxOutputUsagePercent,
+            maxOutputThresholdReached: maxOutputUsagePercent >= threshold,
             outputReserveTokens: Number(budget.outputReserveTokens),
             remainingTokens,
+            contextFallbackReached,
             usagePercent: Math.min(100, Math.round(inputTokens / contextWindow * 100)),
             tokenDistribution: {
                 systemPromptTokens,
@@ -3640,7 +3651,7 @@ export class AiManager {
             },
             compactThreshold: threshold,
             compactableMessageCount,
-            compactRecommended: compactableMessageCount > 0 && conversationUsagePercent >= threshold,
+            compactRecommended: compactableMessageCount > 0 && (conversationUsagePercent >= threshold || contextFallbackReached),
             contextWarningPending: conversation?.warningPending ?? false,
             compactedMessageCount: conversation?.compactedMessageCount ?? 0,
             includedContextBlocks: contextPlan.includedBlockIds.length,
@@ -3665,6 +3676,7 @@ export class AiManager {
             contextWindow,
             inputTokens,
             remainingTokens,
+            contextFallbackReached: remainingTokens <= MIN_CONTEXT_REMAINING_TOKENS,
             usagePercent: Math.min(100, Math.round(inputTokens / contextWindow * 100)),
             tokenDistribution: {
                 systemPromptTokens,
@@ -3678,8 +3690,11 @@ export class AiManager {
     inspectConversationContext(input) {
         const usage = this.getContextUsage({ ...input, taskType: "chat" });
         const usagePercent = Number(usage.usagePercent) || 0;
+        const maxOutputThresholdReached = usage.maxOutputThresholdReached === true;
+        const contextFallbackReached = usage.contextFallbackReached === true;
         const compactableMessageCount = Number(usage.compactableMessageCount) || 0;
-        if (usagePercent >= FORCE_CONVERSATION_COMPACTION_USAGE_PERCENT) {
+        const outputThresholdNeedsCompaction = (maxOutputThresholdReached || contextFallbackReached) && compactableMessageCount > 0;
+        if (usagePercent >= FORCE_CONVERSATION_COMPACTION_USAGE_PERCENT || outputThresholdNeedsCompaction) {
             if (compactableMessageCount <= 0) {
                 throw new AppError(409, "AI_CONTEXT_COMPACTION_UNAVAILABLE", "当前请求已占满模型上下文，但没有可压缩的较早对话；请缩短问题、减少引用或新开对话");
             }
@@ -3712,10 +3727,19 @@ export class AiManager {
         }
         const compaction = await this.compactConversation(input);
         if (compaction.changed !== true) {
+            const inputBelowForcedThreshold = (Number(inspection.usage.usagePercent) || 0) < FORCE_CONVERSATION_COMPACTION_USAGE_PERCENT;
+            if ((inspection.usage.maxOutputThresholdReached === true || inspection.usage.contextFallbackReached === true) && inputBelowForcedThreshold) {
+                return {
+                    action: "ready",
+                    reason: "output_budget_already_fits",
+                    usage: { ...inspection.usage, contextWarningPending: false }
+                };
+            }
             throw new AppError(409, "AI_CONTEXT_COMPACTION_UNAVAILABLE", "当前请求已占满模型上下文，但没有可压缩的较早对话；请缩短问题、减少引用或新开对话");
         }
         const compactedUsage = this.getContextUsage({ ...input, taskType: "chat" });
-        if ((Number(compactedUsage.usagePercent) || 0) >= FORCE_CONVERSATION_COMPACTION_USAGE_PERCENT) {
+        if ((Number(compactedUsage.usagePercent) || 0) >= FORCE_CONVERSATION_COMPACTION_USAGE_PERCENT
+            || compactedUsage.contextFallbackReached === true) {
             throw new AppError(413, "AI_CONTEXT_STILL_OVER_LIMIT", "自动压缩后当前请求仍占满模型上下文；请缩短问题、减少引用或新开对话");
         }
         return {
@@ -4751,6 +4775,8 @@ export class AiManager {
             ...this.sanitizeParameters({ ...preset, ...(input.parameters ?? {}) }, stringValue(model, "model_id")),
             ...thinkingParameters(provider, model)
         };
+        const configuredOutputTokens = Number(requestedParameters.max_tokens) || DEFAULT_MAX_TOKENS;
+        const contextCompactThreshold = Math.min(90, Math.max(50, Number(this.store.getWorkAiSettings(input.workId).contextCompactThreshold) || 85));
         let effectiveInput = input;
         let context = this.buildContext(effectiveInput, model);
         let messages = this.buildMessages(effectiveInput, context);
@@ -5167,7 +5193,14 @@ export class AiManager {
                 // 新工具结果可能附带 toolCallQuotaNotice，预估体积时一并计入，避免低估后触发上下文溢出。
                 const noticeBudgetChars = Math.max(agentToolCallQuotaNoticeBudgetChars(1, agentToolCallLimit), agentToolCallQuotaNoticeBudgetChars(agentToolCallSoftWarningThreshold(agentToolCallLimit), agentToolCallLimit));
                 const maximumNewToolTokens = Math.ceil((AGENT_TOOL_RESULT_MAX_CHARS + noticeBudgetChars) * 1.1) * Math.max(1, toolCallCount);
-                return currentTokens + maximumNewToolTokens + TOOL_CONTEXT_RESPONSE_RESERVE_TOKENS >= contextWindow;
+                // 这里只按工具结果写入后的 context 剩余判断；输出 max_tokens 由下方独立判断。
+                const projectedContextTokens = currentTokens + maximumNewToolTokens;
+                const projectedUsagePercent = Math.round(projectedContextTokens / contextWindow * 100);
+                const projectedContextRemainingTokens = Math.max(0, contextWindow - projectedContextTokens);
+                const maxOutputThresholdReached = configuredOutputTokens >= contextWindow * contextCompactThreshold / 100;
+                return projectedUsagePercent >= contextCompactThreshold
+                    || maxOutputThresholdReached
+                    || projectedContextRemainingTokens <= MIN_CONTEXT_REMAINING_TOKENS;
             };
             let payload = await requestCompletion("auto");
             let choice = payload.choices?.[0];
@@ -9187,7 +9220,7 @@ export class AiManager {
         if (typeof output.top_p === "number")
             output.top_p = clamp(output.top_p, 0, 1);
         output.max_tokens = typeof output.max_tokens === "number"
-            ? Math.round(clamp(output.max_tokens, 1, 32_768))
+            ? Math.round(clamp(output.max_tokens, 1, MAX_MODEL_OUTPUT_TOKENS))
             : DEFAULT_MAX_TOKENS;
         return output;
     }