micro-models-agent 0.7.9 → 0.8.0

package/dist/core/agent.js

@@ -0,0 +1,626 @@
+import { t } from "../i18n/index";
+import { pc } from "../ui/colors";
+import { PromptBuilder } from "./prompt-builder";
+import { OrchestratorClient } from "../llm/orchestrator";
+import { validatePlan, applyAutoFixes, } from "../modules/execution/plan-validator";
+import { MoEExecutor } from "../modules/execution/moe-executor";
+import { StepVerifier } from "../modules/execution/verifier";
+const TOOL_RESULT_MAX_TOKENS_RATIO = 0.3;
+const TOOL_RESULT_ABSOLUTE_MAX_CHARS = 15000;
+export class Agent {
+    deps;
+    systemPromptAdded = false;
+    constructor(deps) {
+        this.deps = deps;
+    }
+    setScope() {
+        const executor = this.deps.toolExecutor;
+        if (executor && executor.ctx && this.deps.scope) {
+            executor.ctx.scope = this.deps.scope;
+        }
+    }
+    buildSystemPrompt() {
+        const systemBudget = Math.floor(this.deps.config.contextWindow *
+            this.deps.config.contextBudget.systemPrompt);
+        const builder = new PromptBuilder(systemBudget);
+        builder.addBlocks(this.deps.promptBlocks);
+        const dynamic = this.deps.getDynamicPromptBlocks?.() ?? [];
+        if (dynamic.length > 0) {
+            builder.addBlocks(dynamic);
+        }
+        const pluginBlocks = (this.deps.pluginManager.runOnBuildPrompt?.() ?? [])
+            .filter((s) => Boolean(s && s.trim() !== ""))
+            .map((content) => ({
+            content,
+            priority: "low",
+            essential: false,
+            estimatedTokens: this.deps.llmProvider.countTokens(content),
+        }));
+        if (pluginBlocks.length > 0) {
+            builder.addBlocks(pluginBlocks);
+        }
+        return builder.build();
+    }
+    /** Refresh the system prompt in context if dynamic blocks changed. */
+    refreshSystemPrompt() {
+        const { prompt } = this.buildSystemPrompt();
+        const current = this.deps.contextManager
+            .getActiveHistory()
+            .find((m) => m.role === "system");
+        if (!current || current.content !== prompt) {
+            this.deps.contextManager.updateSystemPrompt?.(prompt);
+        }
+    }
+    truncateToolOutput(output, budget, currentTokens) {
+        const remainingBudget = budget.history - currentTokens;
+        const maxCharsByBudget = Math.floor(remainingBudget * 0.5 * 2);
+        const maxCharsByRatio = Math.floor(budget.history * TOOL_RESULT_MAX_TOKENS_RATIO * 2);
+        const maxChars = Math.min(maxCharsByBudget, maxCharsByRatio, TOOL_RESULT_ABSOLUTE_MAX_CHARS);
+        if (output.length <= maxChars)
+            return output;
+        const truncated = output.slice(0, maxChars);
+        const removedChars = output.length - maxChars;
+        return (truncated +
+            `\n\n${t("tool.truncated", { tokens: Math.ceil(removedChars / 2) })}`);
+    }
+    emitPhase(iteration, phase, onPhase) {
+        this.deps.pluginManager.runOnPhase?.({ iteration, logger: this.deps.logger }, phase);
+        onPhase?.(phase);
+    }
+    async run(input, onChunk, onMeta, onTool, onPhase) {
+        this.setScope();
+        const { config, llmProvider, toolExecutor, pluginManager, contextManager, hallucinationDetector, logger, sessionManager, } = this.deps;
+        if (sessionManager && !sessionManager.getActive()) {
+            sessionManager.create();
+            logger.debug(`Session started: ${sessionManager.getActive()}`);
+        }
+        if (!this.systemPromptAdded &&
+            !contextManager.getActiveHistory().some((m) => m.role === "system")) {
+            const { prompt: systemPrompt, excluded } = this.buildSystemPrompt();
+            contextManager.addMessage({ role: "system", content: systemPrompt });
+            this.systemPromptAdded = true;
+            if (sessionManager) {
+                sessionManager.appendLog({
+                    ts: new Date().toISOString(),
+                    type: "system",
+                    content: systemPrompt.slice(0, 2000),
+                });
+                if (excluded.length > 0) {
+                    sessionManager.appendLog({
+                        ts: new Date().toISOString(),
+                        type: "system",
+                        content: `[Excluded prompt blocks: ${excluded.length}]`,
+                    });
+                }
+            }
+            pluginManager.runOnSessionStart({
+                logger,
+                sessionManager: sessionManager?.getActiveMeta(),
+            });
+        }
+        contextManager.addMessage({ role: "user", content: input });
+        if (sessionManager) {
+            sessionManager.appendMessage({
+                role: "user",
+                content: input,
+                timestamp: new Date().toISOString(),
+            });
+            sessionManager.appendLog({
+                ts: new Date().toISOString(),
+                type: "user",
+                content: input,
+            });
+        }
+        if (config.moe?.enabled) {
+            return this.runWithMoE(input, onChunk, onMeta, onTool, onPhase);
+        }
+        return this.executeSingleAgentLoop(input, onChunk, onMeta, onTool, onPhase);
+    }
+    async executeSingleAgentLoop(input, onChunk, onMeta, onTool, onPhase) {
+        const { config, llmProvider, toolExecutor, pluginManager, contextManager, hallucinationDetector, logger, sessionManager, } = this.deps;
+        let iteration = 0;
+        let lastText = "";
+        let hallucinationRetries = 0;
+        let lastToolSignature = "";
+        const MAX_HALLUCINATION_RETRIES = 3;
+        while (iteration < config.maxToolIterations) {
+            iteration++;
+            pluginManager.runOnBeforeThink({
+                iteration,
+                logger,
+                lastUserMessage: input,
+                contextManager,
+                onMeta,
+            });
+            if (contextManager.needsCompaction()) {
+                contextManager.compact();
+                logger.debug("Context compacted");
+                if (sessionManager) {
+                    sessionManager.appendLog({
+                        ts: new Date().toISOString(),
+                        type: "compaction",
+                        content: `regular compaction, iteration ${iteration}`,
+                        iteration,
+                    });
+                }
+            }
+            const currentTokens = contextManager.getEstimatedTokens();
+            const budget = contextManager.getBudget();
+            if (currentTokens > budget.history) {
+                contextManager.compact();
+                logger.warn(`Context overflow (${currentTokens} > ${budget.history}), forced compaction`);
+                if (sessionManager) {
+                    sessionManager.appendLog({
+                        ts: new Date().toISOString(),
+                        type: "compaction",
+                        content: `forced compaction (${currentTokens} > ${budget.history}), iteration ${iteration}`,
+                        iteration,
+                        contextTokens: currentTokens,
+                        contextLimit: budget.history,
+                    });
+                }
+            }
+            this.refreshSystemPrompt();
+            const history = contextManager.getActiveHistory();
+            const allTools = toolExecutor.getToolDefinitions(this.deps.toolTags);
+            if (sessionManager) {
+                sessionManager.appendLog({
+                    ts: new Date().toISOString(),
+                    type: "tool_defs",
+                    toolCount: allTools.length,
+                    toolNames: allTools.map((t) => t.name),
+                    iteration,
+                });
+            }
+            let textContent = "";
+            let reasoningContent = "";
+            const toolCalls = [];
+            let sawToolCall = false;
+            let emittedReasoning = false;
+            this.emitPhase(iteration, "thinking", onPhase);
+            try {
+                for await (const chunk of llmProvider.chat(history, allTools)) {
+                    if (chunk.type === "text" && chunk.content) {
+                        if (emittedReasoning && !textContent) {
+                            onMeta?.("\n\n");
+                        }
+                        textContent += chunk.content;
+                        const textOut = pluginManager.runOnText({ iteration, logger }, chunk.content);
+                        onChunk?.(textOut);
+                    }
+                    if (chunk.type === "reasoning" && chunk.content) {
+                        reasoningContent += chunk.content;
+                        if (config.showReasoning) {
+                            const metaOut = pluginManager.runOnMeta({ iteration, logger }, chunk.content);
+                            if (metaOut) {
+                                onMeta?.(pc.dim(metaOut));
+                            }
+                            emittedReasoning = true;
+                        }
+                    }
+                    if (chunk.type === "tool_call" && chunk.toolCall) {
+                        sawToolCall = true;
+                        let parsedArgs;
+                        try {
+                            parsedArgs = JSON.parse(chunk.toolCall.arguments);
+                        }
+                        catch {
+                            parsedArgs = {};
+                        }
+                        toolCalls.push({
+                            id: chunk.toolCall.id,
+                            name: chunk.toolCall.name,
+                            arguments: parsedArgs,
+                        });
+                    }
+                }
+            }
+            catch (err) {
+                logger.error(`LLM call failed: ${err.message}`);
+                if (sessionManager) {
+                    sessionManager.appendLog({
+                        ts: new Date().toISOString(),
+                        type: "error",
+                        content: err.message,
+                    });
+                }
+                pluginManager.runOnError({ iteration, logger }, err);
+                return {
+                    success: false,
+                    text: lastText,
+                    error: t("error.llm", { message: err.message }),
+                    iterationCount: iteration,
+                };
+            }
+            finally {
+                this.emitPhase(iteration, "done", onPhase);
+            }
+            let llmResponse = null;
+            if (sawToolCall) {
+                llmResponse = { type: "tool_call", calls: toolCalls };
+            }
+            else if (textContent) {
+                llmResponse = { type: "text", content: textContent };
+            }
+            else if (reasoningContent) {
+                llmResponse = { type: "reasoning", content: reasoningContent };
+            }
+            pluginManager.runOnAfterThink({ iteration, logger }, llmResponse);
+            if (this.deps.exitOnComplete && sawToolCall) {
+                const signature = toolCalls
+                    .map((tc) => `${tc.name}:${JSON.stringify(tc.arguments)}`)
+                    .join("|");
+                if (signature && signature === lastToolSignature) {
+                    logger.debug("Exit-on-complete: repeated identical tool call, stopping");
+                    break;
+                }
+                lastToolSignature = signature;
+            }
+            if (sessionManager) {
+                if (reasoningContent) {
+                    sessionManager.appendLog({
+                        ts: new Date().toISOString(),
+                        type: "reasoning",
+                        content: reasoningContent,
+                        iteration,
+                    });
+                }
+                if (sawToolCall) {
+                    sessionManager.appendLog({
+                        ts: new Date().toISOString(),
+                        type: "assistant",
+                        content: textContent || "",
+                        tool_calls: toolCalls.map((tc) => ({
+                            id: tc.id,
+                            name: tc.name,
+                            arguments: tc.arguments,
+                        })),
+                        iteration,
+                    });
+                }
+            }
+            if (sawToolCall) {
+                contextManager.addMessage({
+                    role: "assistant",
+                    content: textContent || "",
+                    tool_calls: toolCalls.map((tc) => ({
+                        id: tc.id,
+                        type: "function",
+                        function: {
+                            name: tc.name,
+                            arguments: JSON.stringify(tc.arguments),
+                        },
+                    })),
+                });
+                const summaries = [];
+                for (const call of toolCalls) {
+                    this.setScope();
+                    const startTime = Date.now();
+                    pluginManager.runOnToolCall({
+                        toolName: call.name,
+                        args: call.arguments,
+                    });
+                    pluginManager.runOnToolStart({ iteration, logger }, { id: call.id, name: call.name, arguments: call.arguments });
+                    onTool?.({ type: "start", tool: call.name, args: call.arguments });
+                    if (sessionManager) {
+                        sessionManager.appendLog({
+                            ts: new Date().toISOString(),
+                            type: "tool_call",
+                            tool: call.name,
+                            tool_call_id: call.id,
+                            args: call.arguments,
+                            iteration,
+                        });
+                    }
+                    const result = await toolExecutor.execute(call);
+                    const duration = Date.now() - startTime;
+                    pluginManager.runOnToolEnd({ iteration, logger }, { id: call.id, name: call.name, arguments: call.arguments }, result, duration);
+                    if (result.display) {
+                        onMeta?.("\n" + result.display + "\n");
+                    }
+                    const metaOut = pluginManager.runOnMeta({ iteration, logger }, result.output);
+                    onMeta?.("\n" + pc.dim(metaOut) + "\n");
+                    if (result.diff) {
+                        onMeta?.("\n" + result.diff + "\n");
+                    }
+                    onTool?.({
+                        type: "end",
+                        tool: call.name,
+                        args: call.arguments,
+                        duration,
+                        error: !result.success,
+                    });
+                    const currentTokens = contextManager.getEstimatedTokens();
+                    const budget = contextManager.getBudget();
+                    const truncatedOutput = this.truncateToolOutput(result.output, budget, currentTokens);
+                    contextManager.addMessage({
+                        role: "tool",
+                        content: truncatedOutput,
+                        name: call.name,
+                        tool_call_id: call.id,
+                    });
+                    summaries.push(`[Tool: ${call.name} (${JSON.stringify(call.arguments)}) → ${truncatedOutput.slice(0, 200)}]`);
+                    if (sessionManager && config.session.autoSave) {
+                        sessionManager.appendMessage({
+                            role: "tool",
+                            content: truncatedOutput.slice(0, 500),
+                            name: call.name,
+                            timestamp: new Date().toISOString(),
+                        });
+                        sessionManager.appendLog({
+                            ts: new Date().toISOString(),
+                            type: "tool_result",
+                            tool: call.name,
+                            tool_call_id: call.id,
+                            success: result.success,
+                            content: result.output.slice(0, 1000),
+                            diff: result.diff,
+                            duration,
+                            iteration,
+                        });
+                    }
+                    if (contextManager.needsCompaction()) {
+                        contextManager.compact();
+                        logger.debug("Context compacted after tool result");
+                    }
+                }
+                contextManager.addMessage({
+                    role: "user",
+                    content: `<system-summary>${summaries.join("\n")}</system-summary>`,
+                });
+                continue;
+            }
+            const hallucinationResult = hallucinationDetector.validate(textContent);
+            if (hallucinationResult.status === "block") {
+                logger.warn(`Response blocked: ${hallucinationResult.reason}`);
+                return {
+                    success: false,
+                    text: lastText,
+                    error: t("error.response_blocked", {
+                        reason: hallucinationResult.reason || "",
+                    }),
+                    iterationCount: iteration,
+                };
+            }
+            if (hallucinationResult.status === "warn") {
+                logger.warn(`Hallucination warning: ${hallucinationResult.reason}`);
+                const warnPrefix = t("hall.uncertainty_prefix");
+                if (onChunk) {
+                    onChunk(warnPrefix);
+                }
+                lastText = warnPrefix + lastText;
+            }
+            if (hallucinationResult.status === "retry") {
+                if (this.deps.exitOnComplete) {
+                    logger.debug("Exit-on-complete: stopping on first response");
+                    break;
+                }
+                if (hallucinationRetries >= MAX_HALLUCINATION_RETRIES) {
+                    logger.warn(`Hallucination retries exhausted (${MAX_HALLUCINATION_RETRIES}), returning error`);
+                    return {
+                        success: false,
+                        text: lastText,
+                        error: t("error.response_blocked", {
+                            reason: t("hall.max_retries_exhausted"),
+                        }),
+                        iterationCount: iteration,
+                    };
+                }
+                hallucinationRetries++;
+                logger.warn(`Hallucination retry (${hallucinationRetries}/${MAX_HALLUCINATION_RETRIES}): ${hallucinationResult.reason}`);
+                if (textContent) {
+                    contextManager.addMessage({
+                        role: "assistant",
+                        content: textContent,
+                    });
+                }
+                contextManager.addMessage({
+                    role: "user",
+                    content: `<system-summary>[Retry context: ${hallucinationResult.reason}. Original task: "${input}". You must either call a needed tool or provide a substantive response. Empty replies are not allowed.]</system-summary>`,
+                });
+                continue;
+            }
+            if (textContent) {
+                contextManager.addMessage({ role: "assistant", content: textContent });
+                if (sessionManager && config.session.autoSave) {
+                    sessionManager.appendMessage({
+                        role: "assistant",
+                        content: textContent.slice(0, 500),
+                        timestamp: new Date().toISOString(),
+                    });
+                }
+            }
+            if (sessionManager) {
+                if (reasoningContent) {
+                    sessionManager.appendLog({
+                        ts: new Date().toISOString(),
+                        type: "reasoning",
+                        content: reasoningContent,
+                        iteration,
+                    });
+                }
+                if (textContent) {
+                    sessionManager.appendLog({
+                        ts: new Date().toISOString(),
+                        type: "assistant",
+                        content: textContent,
+                        iteration,
+                    });
+                }
+                else if (reasoningContent) {
+                    sessionManager.appendLog({
+                        ts: new Date().toISOString(),
+                        type: "assistant",
+                        content: reasoningContent,
+                        iteration,
+                    });
+                }
+            }
+            lastText = textContent;
+            if (!sawToolCall) {
+                if (this.deps.finalAudit) {
+                    const audit = await this.deps.finalAudit();
+                    if (audit && !audit.passed) {
+                        logger.warn(`Final audit incomplete: ${audit.summary}`);
+                        const steps = audit.pendingSteps.slice(0, 5).join("; ") || "—";
+                        contextManager.addMessage({
+                            role: "user",
+                            content: `<system-summary>${t("exec.audit_incomplete", {
+                                summary: audit.summary,
+                                steps,
+                            })}</system-summary>`,
+                        });
+                        if (sessionManager) {
+                            sessionManager.appendLog({
+                                ts: new Date().toISOString(),
+                                type: "audit",
+                                content: audit.summary,
+                                iteration,
+                            });
+                        }
+                        if (iteration >= config.maxToolIterations - 1) {
+                            break;
+                        }
+                        continue;
+                    }
+                }
+                break;
+            }
+        }
+        const tokensUsed = contextManager.getEstimatedTokens();
+        const budget = contextManager.getBudget();
+        if (iteration >= config.maxToolIterations) {
+            return {
+                success: false,
+                text: lastText,
+                error: t("error.max_iters", { max: config.maxToolIterations }),
+                iterationCount: iteration,
+                contextUsed: tokensUsed,
+                contextLimit: budget.history,
+            };
+        }
+        return {
+            success: true,
+            text: lastText,
+            iterationCount: iteration,
+            contextUsed: tokensUsed,
+            contextLimit: budget.history,
+        };
+    }
+    async runWithMoE(input, _onChunk, onMeta, onTool, onPhase) {
+        const { config, llmProvider, logger } = this.deps;
+        const orchestrator = new OrchestratorClient({
+            model: config.orchestrator.model,
+            provider: config.orchestrator.provider,
+        }, llmProvider);
+        if (!orchestrator.isEnabled()) {
+            logger.debug("MoE enabled but no orchestrator model configured — falling back to single-agent");
+            return this.executeSingleAgentLoop(input, _onChunk, onMeta, onTool, onPhase);
+        }
+        onMeta?.("🤖 Planning with MoE mode...\n");
+        this.emitPhase(0, "thinking", onPhase);
+        const planResult = await orchestrator.plan(input);
+        this.emitPhase(0, "done", onPhase);
+        if ("error" in planResult) {
+            logger.warn(`MoE plan failed: ${planResult.error} — falling back to single-agent`);
+            return this.executeSingleAgentLoop(input, _onChunk, onMeta, onTool);
+        }
+        const { plan } = planResult;
+        onMeta?.(`📋 Plan created: "${plan.title}" (${plan.subtasks.length} subtasks)\n`);
+        const validation = validatePlan(plan, config);
+        if (!validation.valid) {
+            const applied = applyAutoFixes(plan, validation.autoFixes);
+            const retry = validatePlan(applied, config);
+            if (!retry.valid) {
+                logger.warn(`MoE plan validation failed: ${retry.errors.join("; ")} — falling back to single-agent`);
+                onMeta?.(`⚠️ Plan validation failed. Falling back to single-agent mode.\n`);
+                return this.executeSingleAgentLoop(input, _onChunk, onMeta, onTool);
+            }
+            Object.assign(plan, applied);
+            onMeta?.(`🔧 Auto-fixed ${validation.autoFixes.length} plan issues.\n`);
+        }
+        const moeDeps = {
+            config,
+            toolRegistry: this.deps.toolExecutor.registry,
+            toolExecutor: this.deps.toolExecutor,
+            llmProvider,
+            logger,
+            baseDir: this.deps.baseDir,
+        };
+        const executor = new MoEExecutor(moeDeps);
+        onMeta?.(`⚙️ Executing ${plan.subtasks.length} subtasks...\n`);
+        const planResults = await executor.executePlan(plan);
+        onMeta?.(`✅ Execution complete: ${planResults.results.filter((r) => r.success).length}/${planResults.results.length} succeeded\n`);
+        const verifier = new StepVerifier(this.deps.baseDir);
+        const knownTags = [
+            "file",
+            "code",
+            "shell",
+            "research",
+            "browser",
+            "vision",
+            "core",
+        ];
+        const verification = await verifier.verifyMoEManifest(plan, config, knownTags);
+        this.emitPhase(0, "thinking", onPhase);
+        const verifyResult = await orchestrator.verifyAndMerge({
+            plan,
+            results: planResults.results.map((r) => ({
+                subtaskId: r.subtaskId,
+                success: r.success,
+                summary: r.summary,
+                result: r.result,
+                error: r.error,
+            })),
+            verifierErrors: verification.errors,
+            verifierWarnings: verification.warnings,
+        });
+        this.emitPhase(0, "done", onPhase);
+        const outputLines = [`## MoE Execution Results\n`];
+        for (const r of planResults.results) {
+            const icon = r.success ? "✅" : "❌";
+            outputLines.push(`${icon} **${r.subtaskId}**: ${r.summary} (${r.durationMs}ms)`);
+        }
+        outputLines.push("");
+        if (verifyResult.type === "final") {
+            outputLines.push(`**Result:** ${verifyResult.finalAnswer || "Complete"}`);
+        }
+        else {
+            outputLines.push(`**Re-plan requested:** ${verifyResult.explanation || ""}`);
+        }
+        const failedCount = planResults.results.filter((r) => !r.success).length;
+        return {
+            success: failedCount === 0 && verification.success,
+            text: outputLines.join("\n"),
+            iterationCount: planResults.results.length,
+        };
+    }
+    clearContext() {
+        this.deps.contextManager.clear();
+        this.systemPromptAdded = false;
+    }
+    setContext(messages) {
+        const { contextManager } = this.deps;
+        contextManager.clear();
+        const { prompt: systemPrompt } = this.buildSystemPrompt();
+        contextManager.addMessage({ role: "system", content: systemPrompt });
+        this.systemPromptAdded = true;
+        for (const msg of messages) {
+            if (msg.role === "system")
+                continue;
+            contextManager.addMessage({
+                role: msg.role,
+                content: msg.content,
+                name: msg.name,
+            });
+        }
+    }
+    shutdown() {
+        const { pluginManager, logger, sessionManager, contextManager } = this.deps;
+        contextManager.onCompact = null;
+        pluginManager.runOnSessionEnd({
+            logger,
+            sessionManager: sessionManager?.getActiveMeta(),
+        });
+    }
+}