@zhin.js/agent 1.1.39 → 1.1.41

package/lib/core/agent-core-run.js

@@ -5,7 +5,7 @@ import { decodeStructuredReply, structuredReplyInstruction, structuredReplySchem
 import { renderAgentOutput } from '../outbound/agent-output.js';
 import { mergeToolOutboundElements } from '../media/media-tool-bridge.js';
 import { formatCompact, truncatePreview, getLogger } from '@zhin.js/logger';
-import { parseOutput, agentLoop, agentContextFrom, assistantText, createUserMessage, agentToolsToLlmTools, createResilientCompletionPort, isProviderFailureMessage, resolveProviderConfigResilience, resolveProviderResilienceConfig } from '@zhin.js/ai';
+import { createPromptContextBlock, projectPromptContext, estimateContextFixedTokens, parseOutput, agentLoop, agentContextFrom, assistantText, createUserMessage, agentToolsToLlmTools, createResilientCompletionPort, isProviderFailureMessage, resolveProviderConfigResilience, resolveProviderResilienceConfig } from '@zhin.js/ai';
 import { tokenUsageToLegacy } from './agent-run-shared.js';
 import { applyExecPolicyToTools } from '../security/exec-policy.js';
 import { resolveModelHarness } from '../config/model-harness-runtime.js';
@@ -15,7 +15,8 @@ import { planToolRun } from '../tool/runtime.js';
 import { sanitizeAssistantReply, unwrapJsonStringLayers } from './text-sanitize.js';
 import { formatToolCallsForUser } from './tool-calls-user-format.js';
 import { summarizeExhaustedTurn } from './exhausted-turn-reply.js';
-import { decodeLenientStructuredReply, recoverStructuredReply, STRUCTURED_REPLY_FAILURE } from './structured-reply-recovery.js';
+import { decodeLenientStructuredReply, recoverStructuredReply } from './structured-reply-recovery.js';
+import { chatFailureKindOf, chatFailureLine } from '../outbound/chat-failure.js';
 import { trackTurnCompletionUsage } from './turn-completion-usage.js';
 import { shouldSuppressReplyForSpawnDelegation } from './spawn-delegation.js';
 import { logPhase, tokenUsageLogFields, logAgentLoopIterationEnd } from '../internal/phase-trace.js';
@@ -24,7 +25,7 @@ import { buildLlmToolsForProvider } from '../tool/deferred-resolution.js';
 import { applyToolToModelOutput } from '../tool/tool-model-output.js';
 import { resolveAlwaysLoadedSet } from '../tool-catalog/resolve-config.js';
 import { resolveDeferredApiTools, residentToolNames } from '../tool-catalog/tool-catalog.js';
-import { createTurnEventMapperState, mapAgentEventToTurnEvents, turnUsageEvent, } from './turn-event-mapper.js';
+import { createTurnEventMapperState, mapAgentEventToTurnEvents, } from './turn-event-mapper.js';
 import { TurnJournalCommitError } from '../turn/journal-integrity.js';
 import { ExecutableToolRegistry } from '../tool/executable-tool-registry.js';
 import { withoutForbiddenShell } from '../plugin-runtime/native-bash-tool.js';
@@ -33,7 +34,6 @@ import { createMissingToolResolver } from './missing-tool-resolver.js';
 import { DEFAULT_EXPLORATION_BUDGET, EXPLORATION_GUIDANCE, ExplorationBudget, } from './exploration-budget.js';
 import { TurnProgressTracker, autoContinueAllowed, buildContinuationInstruction, isProgress, resolveToolBudget, } from './turn-progress.js';
 import { TurnProgressReporter, summarizeProgress } from './progress-reporter.js';
-import { estimatePromptTokens } from '../prompt/estimate-tokens.js';
 import { noteTurnTrace } from '../trace/turn-trace-registry.js';
 import { buildContextBudgetReport } from '../trace/context-budget.js';
 import { renderDeferredToolIndex } from '../skill/capability-catalog.js';
@@ -46,11 +46,14 @@ function normalizeAssistantMessageText(assistant) {
         }
     }
 }
-function resolveAssistantReplyText(assistant) {
+function resolveAssistantReplyText(assistant, chat) {
     normalizeAssistantMessageText(assistant);
     const text = assistantText(assistant);
     if (text.trim())
         return text;
+    // In a chat the failure is said like a member would say it; operators read the logs.
+    if (assistant.errorMessage && chat)
+        return chatFailureLine(chatFailureKindOf(assistant.errorMessage), chat);
     if (assistant.errorMessage && isProviderFailureMessage(assistant.errorMessage)) {
         return `${assistant.errorMessage}。`;
     }
@@ -184,7 +187,7 @@ async function* runAgentLoopVisionTurnOnceRun(input) {
             modelId,
             label: 'vision',
         }),
-        convertToLlm: (messages) => messages,
+        convertToLlm: projectPromptContext,
         getSteeringMessages: promptHooks?.getSteeringMessages,
         getFollowUpMessages: promptHooks?.getFollowUpMessages,
     };
@@ -320,7 +323,7 @@ export async function* runAgentLoopTextTurnRun(input) {
     const loaded = persistentConversation
         ? await repo.loadContext(sessionId)
         : { messages: [] };
-    const promptMessages = input.initialMessages?.length
+    let promptMessages = input.initialMessages?.length
         ? input.initialMessages
         : [createUserMessage(input.rawContent)];
     const directTools = input.toolLoading === 'direct';
@@ -336,6 +339,9 @@ export async function* runAgentLoopTextTurnRun(input) {
     const messagePreset = configuredSchema === true || configuredSchema === 'segments';
     const imReply = messagePreset && input.turnContext.origin.kind === 'im'
         && input.conversationPersistence !== 'none';
+    const chatFailure = input.turnContext.origin.kind === 'im'
+        ? { master: input.turnContext.principal.roles.includes('master') }
+        : undefined;
     const outputSchema = imReply ? structuredReplySchema(input.replySegments)
         : messagePreset || !configuredSchema ? undefined : configuredSchema;
     const turnShellPolicy = input.turnContext.policy?.shell;
@@ -433,8 +439,15 @@ export async function* runAgentLoopTextTurnRun(input) {
     if (hasTools && !directTools) {
         const toolIndex = renderDeferredToolIndex(deferredCatalogTools(catalog, agentTools)
             .map((item) => ({ name: item.name, description: item.brief })));
-        if (toolIndex)
-            systemPrompt = `${systemPrompt}\n\n${toolIndex}`;
+        if (toolIndex) {
+            const block = createPromptContextBlock('deferred-tools', toolIndex);
+            const userIndex = promptMessages.map(message => message.role === 'user' && Array.isArray(message.content)).lastIndexOf(true);
+            promptMessages = userIndex < 0
+                ? [...promptMessages, { ...createUserMessage(''), content: [block] }]
+                : promptMessages.map((message, index) => index === userIndex && Array.isArray(message.content)
+                    ? { ...message, content: [...message.content, block] }
+                    : message);
+        }
     }
     if (networkWithheld)
         systemPrompt = `${systemPrompt}\n\n${NETWORK_DISABLED_PROMPT_NOTE}`;
@@ -447,6 +460,7 @@ export async function* runAgentLoopTextTurnRun(input) {
             kind: 'context_budget',
             report: buildContextBudgetReport({
                 systemPrompt,
+                outputSchema,
                 sections: promptSections,
                 bootstrapContext: input.promptRuntime?.bootstrapContext,
                 activeSkillsContext: input.promptRuntime?.activeSkillsContext,
@@ -490,8 +504,16 @@ export async function* runAgentLoopTextTurnRun(input) {
     const contextWindow = llmModel.contextWindow ?? host.config.contextTokens;
     // System prompt and tool schemas count against the window too, so the
     // compaction threshold is measured on what is left for messages (#730).
-    const fixedPromptTokens = () => estimatePromptTokens(systemPrompt)
-        + llmTools.reduce((sum, tool) => sum + estimatePromptTokens(`${tool.name} ${tool.description}`), 0);
+    let budgetTools;
+    let fixedBudget;
+    const fixedPromptTokens = () => {
+        // 本回合系统提示不变；工具渐进披露替换数组后重新计算完整 Schema。
+        if (!fixedBudget || budgetTools !== llmTools) {
+            budgetTools = llmTools;
+            fixedBudget = estimateContextFixedTokens({ systemPrompt, tools: llmTools, model: llmModel, outputSchema });
+        }
+        return fixedBudget;
+    };
     const loopHooks = input.loopHooks;
     const completionPort = createTurnCompletionPort({
         host,
@@ -544,7 +566,7 @@ export async function* runAgentLoopTextTurnRun(input) {
             }),
             ...(outputSchema ? { outputSchema } : {}),
         },
-        convertToLlm: (messages) => messages,
+        convertToLlm: projectPromptContext,
         transformContext: async (messages, ctxSignal) => persistentConversation ? host.compactionRuntime.transformContext(messages, ctxSignal, {
             host,
             sessionId,
@@ -552,7 +574,7 @@ export async function* runAgentLoopTextTurnRun(input) {
             model: llmModel,
             compactionConfig: host.config.compaction,
             contextWindow,
-            reservedTokens: fixedPromptTokens(),
+            reservedTokens: await fixedPromptTokens(),
             toldLedger: host.toldLedger?.render(sessionId),
             mode: 'text',
             loopHooks,
@@ -564,7 +586,7 @@ export async function* runAgentLoopTextTurnRun(input) {
             model: llmModel,
             compactionConfig: host.config.compaction,
             contextWindow,
-            reservedTokens: fixedPromptTokens(),
+            reservedTokens: await fixedPromptTokens(),
             toldLedger: host.toldLedger?.render(sessionId),
             mode: 'text',
             force: true,
@@ -644,7 +666,9 @@ export async function* runAgentLoopTextTurnRun(input) {
                 const budgetNote = explorationBudget?.record(resolvedName, isProgress(progressKind));
                 const finalText = budgetNote ? `${resultText}\n\n${budgetNote}` : resultText;
                 toolCalls.push({ tool: resolvedName, args: effectiveArgs, result: finalText });
-                return toolResultToAgentMessage(toolCall, finalText, false);
+                const result = toolResultToAgentMessage(toolCall, finalText, false);
+                const blocks = deferredController?.modelOutputBlocks?.(resolvedName, effectiveArgs, finalText);
+                return blocks ? { ...result, content: blocks } : result;
             }
             catch (err) {
                 if (err instanceof TurnJournalCommitError)
@@ -719,8 +743,7 @@ export async function* runAgentLoopTextTurnRun(input) {
     // Messages of the running segment, for progress summaries before agent_end.
     let segmentMessages = [];
     function* drainCompletionUsage() {
-        for (const usage of completionUsage.drainUsage()) {
-            const event = turnUsageEvent(usage);
+        for (const event of completionUsage.drainEvents()) {
             emitTurnEvent(event);
             yield event;
         }
@@ -773,7 +796,7 @@ export async function* runAgentLoopTextTurnRun(input) {
                 lastAssistantNeedsHostReply = Boolean(assistant.errorMessage)
                     || (!lastAssistantHadToolCalls && !answerText.trim());
                 lastAssistantText = imReply && !lastAssistantNeedsHostReply ? answerText
-                    : resolveAssistantReplyText(imReply ? { ...assistant, content: [] } : assistant);
+                    : resolveAssistantReplyText(imReply ? { ...assistant, content: [] } : assistant, chatFailure);
                 logAssistantIterationFailure(assistant, modelId, sessionId);
                 const toolNames = assistant.content
                     .filter((b) => b.type === 'toolCall')
@@ -904,7 +927,10 @@ export async function* runAgentLoopTextTurnRun(input) {
     const continuationPrefix = continuations > 0 && input.turnContext.origin.kind !== 'im'
         ? `（任务步骤较多，已自动续跑 ${continuations} 次）\n\n` : '';
     const proseReply = structuredReply ? ''
-        : sanitizeAssistantReply(rawReply, { toolSummary: formatToolCallsForUser(toolCalls) });
+        : sanitizeAssistantReply(rawReply, {
+            toolSummary: formatToolCallsForUser(toolCalls),
+            ...(chatFailure ? { emptyMessage: chatFailureLine('malformed_tool_call', chatFailure) } : {}),
+        });
     const textReply = delegatedOnly ? '' : proseReply.trim() ? continuationPrefix + proseReply : proseReply;
     let output = delegatedOnly || structuredReply ? [] : parseOutput(textReply);
     if (!delegatedOnly && structuredReply) {
@@ -926,13 +952,14 @@ export async function* runAgentLoopTextTurnRun(input) {
                     transport: completionUsage.transport, model: llmModel, raw: rawReply,
                     error: error instanceof Error ? error : new Error('Invalid structured reply'), signal,
                     segments: input.replySegments,
+                    failureText: chatFailureLine('format', { ...chatFailure, random: 0 }),
                 });
                 output = recovery.output;
                 // History must match what the user saw: the corrected reply, or the failure text. Never the
                 // correction prompt (it embeds the rejected raw text) and never the rejected raw reply.
                 if (messages && rejected?.role === 'assistant') {
                     messages[messages.length - 1] = recovery.assistant
-                        ?? { ...rejected, content: [{ type: 'text', text: STRUCTURED_REPLY_FAILURE }] };
+                        ?? { ...rejected, content: [{ type: 'text', text: renderAgentOutput(recovery.output) }] };
                 }
             }
         }