@kolisachint/hoocode-agent 0.4.74 → 0.4.76

package/dist/core/agent-session.js

@@ -14,26 +14,21 @@
  */
 import { existsSync, mkdirSync, readFileSync, writeFileSync } from "node:fs";
 import { basename, dirname, resolve } from "node:path";
-import { clampThinkingLevel, cleanupSessionResources, completeSimple, getSupportedThinkingLevels, isContextOverflow, modelsAreEqual, resetApiProviders, } from "@kolisachint/hoocode-ai";
+import { clampThinkingLevel, cleanupSessionResources, getSupportedThinkingLevels, isContextOverflow, modelsAreEqual, resetApiProviders, } from "@kolisachint/hoocode-ai";
 import { theme } from "../modes/interactive/theme/theme.js";
 import { stripFrontmatter } from "../utils/frontmatter.js";
 import { sleep } from "../utils/sleep.js";
 import { loadAgentRegistry } from "./agent-registry.js";
 import { formatNoApiKeyFoundMessage, formatNoModelSelectedMessage } from "./auth-guidance.js";
 import { executeBashWithOperations } from "./bash-executor.js";
-import { calculateContextTokens, collectEntriesForBranchSummary, compact, estimateContextTokens, generateBranchSummary, prepareCompaction, serializeConversation, shouldCompact, } from "./compaction/index.js";
+import { calculateContextTokens, collectEntriesForBranchSummary, compact, estimateContextTokens, generateBranchSummary, prepareCompaction, shouldCompact, } from "./compaction/index.js";
 import { DEFAULT_THINKING_LEVEL } from "./defaults.js";
 import { exportSessionToHtml } from "./export-html/index.js";
 import { createToolHtmlRenderer } from "./export-html/tool-renderer.js";
 import { ExtensionRunner, wrapRegisteredTools, } from "./extensions/index.js";
 import { emitSessionShutdownEvent } from "./extensions/runner.js";
-import { convertToLlm } from "./messages.js";
 import { expandPromptTemplate, tryExpandPromptTemplate } from "./prompt-templates.js";
 import { clearProviderExhaustion, isProviderQuotaError, markProviderExhausted } from "./provider-health.js";
-import { LocalInferenceRouter, resolveRoutingMode } from "./routing/local-inference.js";
-import { logLocalInferenceFallback, logRoutingMetrics } from "./routing/metrics.js";
-import { MlxServerManager } from "./routing/mlx-server.js";
-import { buildToolResultPrompt, stripThinkTags, TOOL_RESULT_SYSTEM_PROMPT } from "./routing/tool-result-prompts.js";
 import { CURRENT_SESSION_VERSION, getLatestCompactionEntry } from "./session-manager.js";
 import { createSyntheticSourceInfo } from "./source-info.js";
 import { updateSubagentSkillPaths } from "./subagent-pool-instance.js";
@@ -41,6 +36,12 @@ import { buildSystemPrompt } from "./system-prompt.js";
 import { createLocalBashOperations } from "./tools/bash.js";
 import { createAllToolDefinitions } from "./tools/index.js";
 import { createToolDefinitionFromAgentTool } from "./tools/tool-definition-wrapper.js";
+/**
+ * Retryable error signatures (overloaded, rate limit, server/network errors,
+ * transport closes). Compiled once at module load instead of on every assistant
+ * response. Context-overflow errors are handled separately by compaction.
+ */
+const RETRYABLE_ERROR_PATTERN = /overloaded|provider.?returned.?error|rate.?limit|too many requests|429|500|502|503|504|service.?unavailable|server.?error|internal.?error|network.?error|connection.?error|connection.?refused|connection.?lost|websocket.?closed|websocket.?error|other side closed|fetch failed|upstream.?connect|reset before headers|socket hang up|ended without|http2 request did not get a response|timed? out|timeout|terminated|retry delay/i;
 /**
  * Parse a skill block from message text.
  * Returns null if the text doesn't contain a skill block.
@@ -113,12 +114,6 @@ export class AgentSession {
     _extensionErrorUnsubscriber;
     // Model registry for API key resolution
     _modelRegistry;
-    // Local-inference routing (opt-in via enableLocalInference). Lazily built.
-    _enableLocalInference;
-    _localRouter;
-    _localRouterBuilt = false;
-    _mlxServer;
-    _mlxServerBuilt = false;
     // Tool registry for extension getTools/setTools
     _toolRegistry = new Map();
     _toolDefinitions = new Map();
@@ -137,7 +132,6 @@ export class AgentSession {
         this._customTools = config.customTools ?? [];
         this._cwd = config.cwd;
         this._modelRegistry = config.modelRegistry;
-        this._enableLocalInference = config.enableLocalInference ?? false;
         this._extensionRunnerRef = config.extensionRunnerRef;
         this._initialActiveToolNames = config.initialActiveToolNames;
         this._allowedToolNames = config.allowedToolNames ? new Set(config.allowedToolNames) : undefined;
@@ -160,57 +154,6 @@ export class AgentSession {
     get modelRegistry() {
         return this._modelRegistry;
     }
-    /**
-     * Local-inference router, built lazily from the flag + env + models.json
-     * routing config. Returns undefined when routing is not active.
-     */
-    get localRouter() {
-        if (!this._localRouterBuilt) {
-            this._localRouterBuilt = true;
-            const mode = resolveRoutingMode({
-                enableFlag: this._enableLocalInference,
-                envMode: process.env.HOOCODE_ROUTING_MODE,
-                configMode: this._modelRegistry.getRoutingConfig()?.mode,
-            });
-            this._localRouter =
-                mode === "primary-only"
-                    ? undefined
-                    : LocalInferenceRouter.create({
-                        mode,
-                        config: this._modelRegistry.getRoutingConfig(),
-                        registry: this._modelRegistry,
-                    });
-        }
-        return this._localRouter;
-    }
-    /**
-     * Ensure the harness-managed executor server (if configured) is healthy before
-     * the executor is used. Returns false when routing is inactive or the server
-     * could not be started (caller degrades to primary/raw). Reuses an externally
-     * started server when one is already healthy.
-     */
-    async _ensureExecutorServer(signal) {
-        const router = this.localRouter;
-        if (!router?.isExecutorAvailable())
-            return false;
-        const serverConfig = router.getExecutorConfig()?.server;
-        if (!serverConfig)
-            return true; // user-managed server; assume available
-        if (!this._mlxServerBuilt) {
-            this._mlxServerBuilt = true;
-            const executor = router.getExecutorModel();
-            if (executor) {
-                this._mlxServer = new MlxServerManager({
-                    config: serverConfig,
-                    modelId: executor.id,
-                    baseUrl: executor.baseUrl,
-                });
-            }
-        }
-        if (!this._mlxServer)
-            return false;
-        return this._mlxServer.ensureStarted(signal);
-    }
     async _getRequiredRequestAuth(model) {
         const result = await this._modelRegistry.getApiKeyAndHeaders(model);
         if (!result.ok) {
@@ -284,92 +227,11 @@ export class AgentSession {
                     changed = true;
                 }
             }
-            // Optionally compress large bash output before it enters context.
-            // Only when local-inference tool-result routing is active; on any
-            // failure the original (uncompressed) content is kept (fallback to raw).
-            if (!resolvedIsError) {
-                const compressed = await this._maybeCompressToolResult(toolCall.name, content);
-                if (compressed) {
-                    content = compressed;
-                    changed = true;
-                }
-            }
             if (!changed)
                 return undefined;
             return { content, details, isError: resolvedIsError };
         };
     }
-    /**
-     * Compress a tool result via the local executor when routing is active and the
-     * tool/size qualify. Returns the compressed content blocks, or undefined to
-     * keep the original (no routing, not compressible, outside the size band, or
-     * any failure).
-     */
-    async _maybeCompressToolResult(toolName, content) {
-        const router = this.localRouter;
-        if (!router)
-            return undefined;
-        const text = extractTextContent(content);
-        if (text === undefined)
-            return undefined;
-        const bytes = Buffer.byteLength(text, "utf8");
-        if (!router.shouldCompressToolResult(toolName, bytes))
-            return undefined;
-        const executor = router.getExecutorModel();
-        const prompt = buildToolResultPrompt(toolName, text);
-        if (!executor || !prompt)
-            return undefined;
-        const startedAt = Date.now();
-        try {
-            if (!(await this._ensureExecutorServer())) {
-                throw new Error("executor server unavailable");
-            }
-            const { apiKey, headers } = await this._getRequiredRequestAuth(executor);
-            const response = await completeSimple(executor, {
-                systemPrompt: TOOL_RESULT_SYSTEM_PROMPT,
-                messages: [
-                    {
-                        role: "user",
-                        content: [{ type: "text", text: prompt }],
-                        timestamp: Date.now(),
-                    },
-                ],
-            }, { apiKey, headers });
-            if (response.stopReason === "error") {
-                throw new Error(response.errorMessage || "executor error");
-            }
-            const rawSummary = response.content
-                .filter((c) => c.type === "text")
-                .map((c) => c.text)
-                .join("\n");
-            const summary = stripThinkTags(rawSummary);
-            if (summary === "")
-                throw new Error("empty compression");
-            const compressedBytes = Buffer.byteLength(summary, "utf8");
-            // Guard: if the model failed to compress (output not smaller), keep raw.
-            if (compressedBytes >= bytes)
-                return undefined;
-            logRoutingMetrics({
-                turnKind: "tool-result",
-                provider: executor.provider,
-                model: executor.id,
-                latencyMs: Date.now() - startedAt,
-                bytesBefore: bytes,
-                bytesAfter: compressedBytes,
-                fallback: false,
-            });
-            return [
-                {
-                    type: "text",
-                    text: `${summary}\n\n[compressed from ${bytes} to ${compressedBytes} bytes by local executor]`,
-                },
-            ];
-        }
-        catch (error) {
-            logLocalInferenceFallback("tool-result", error);
-            return undefined;
-        }
-    }
     // =========================================================================
     // Event Subscription
     // =========================================================================
@@ -680,7 +542,6 @@ export class AgentSession {
         this._extensionRunner.invalidate("This extension ctx is stale after session replacement or reload. Do not use a captured extension API or command ctx after ctx.newSession(), ctx.fork(), ctx.switchSession(), or ctx.reload(). For newSession, fork, and switchSession, move post-replacement work into withSession and use the ctx passed to withSession. For reload, do not use the old ctx after await ctx.reload().");
         this._disconnectFromAgent();
         this._eventListeners = [];
-        this._mlxServer?.stop();
         cleanupSessionResources(this.sessionId);
     }
     // =========================================================================
@@ -1467,7 +1328,7 @@ export class AgentSession {
             }
         }
         const generated = extensionCompaction ??
-            (await this._compactWithRouting(preparation, model, apiKey, headers, customInstructions, signal));
+            (await compact(preparation, model, apiKey, headers, customInstructions, signal, this.thinkingLevel));
         if (signal.aborted) {
             return { status: "cancelled" };
         }
@@ -1488,41 +1349,6 @@ export class AgentSession {
             result: { summary, firstKeptEntryId, tokensBefore, tokensAfter: tokensAfter ?? tokensBefore, details },
         };
     }
-    /**
-     * Run compaction, routing the summarization to the local executor model when
-     * local-inference routing is active for summarization. Any executor
-     * failure falls back to the primary model so compaction never hard-fails.
-     */
-    async _compactWithRouting(preparation, primaryModel, apiKey, headers, customInstructions, signal) {
-        const router = this.localRouter;
-        const executor = router?.selectModel("summarization", primaryModel);
-        // Size band guards local inference globally: only route conversations within
-        // the configured byte band to the executor. Oversized conversations are slow
-        // locally and can OOM small machines, so they fall back to the primary model.
-        // Only serialize to measure size when routing is actually engaged (avoids
-        // needless work on the common primary-only path).
-        if (router && executor && executor !== primaryModel) {
-            const conversationBytes = Buffer.byteLength(serializeConversation(convertToLlm(preparation.messagesToSummarize)), "utf8");
-            if (router.withinSizeBand(conversationBytes)) {
-                try {
-                    if (!(await this._ensureExecutorServer(signal))) {
-                        throw new Error("executor server unavailable");
-                    }
-                    const { apiKey: exKey, headers: exHeaders } = await this._getRequiredRequestAuth(executor);
-                    return await compact(preparation, executor, exKey, exHeaders, customInstructions, signal, 
-                    // Executor runs without thinking (validated config); summary quality
-                    // relies on /no_think via the model's promptSuffix compat setting.
-                    "off");
-                }
-                catch (error) {
-                    if (signal.aborted)
-                        throw error;
-                    logLocalInferenceFallback("summarization", error);
-                }
-            }
-        }
-        return compact(preparation, primaryModel, apiKey, headers, customInstructions, signal, this.thinkingLevel);
-    }
     /**
      * Manually compact the session context.
      * Aborts current agent operation first.
@@ -2118,7 +1944,7 @@ export class AgentSession {
             return false;
         const err = message.errorMessage;
         // Match: overloaded_error, provider returned error, rate limit, 429, 500, 502, 503, 504, service unavailable, network/connection errors (including connection lost), WebSocket transport closes/errors, fetch failed, request ended without sending chunks, HTTP/2 closed before response, terminated, retry delay exceeded
-        return /overloaded|provider.?returned.?error|rate.?limit|too many requests|429|500|502|503|504|service.?unavailable|server.?error|internal.?error|network.?error|connection.?error|connection.?refused|connection.?lost|websocket.?closed|websocket.?error|other side closed|fetch failed|upstream.?connect|reset before headers|socket hang up|ended without|http2 request did not get a response|timed? out|timeout|terminated|retry delay/i.test(err);
+        return RETRYABLE_ERROR_PATTERN.test(err);
     }
     /**
      * Handle retryable errors with exponential backoff.
@@ -2708,19 +2534,4 @@ export class AgentSession {
         return this._extensionRunner;
     }
 }
-/**
- * Join all text blocks of a tool result. Returns undefined if any non-text
- * (e.g. image) content is present, since those must not be compressed.
- */
-function extractTextContent(content) {
-    if (content.length === 0)
-        return undefined;
-    const texts = [];
-    for (const block of content) {
-        if (block.type !== "text")
-            return undefined;
-        texts.push(block.text);
-    }
-    return texts.join("\n");
-}
 //# sourceMappingURL=agent-session.js.map