open-agents-ai 0.31.2 → 0.31.3

package/dist/index.js

@@ -11012,6 +11012,39 @@ Rules:
       setContextWindowSize(size) {
         this.options.contextWindowSize = size;
       }
+      // -------------------------------------------------------------------------
+      // Context-aware limits — all dynamic values derived from contextWindowSize
+      // and modelTier. Call this instead of using hardcoded magic numbers.
+      // -------------------------------------------------------------------------
+      /**
+       * Compute all context-dependent limits from the current contextWindowSize
+       * and modelTier. Returns sensible defaults when contextWindowSize is 0.
+       */
+      contextLimits() {
+        const ctx = this.options.contextWindowSize;
+        const tier = this.options.modelTier ?? "large";
+        const compactionThreshold = ctx > 0 ? Math.min(this.options.compactionThreshold, Math.floor(ctx * 0.75)) : this.options.compactionThreshold;
+        const keepRecentDivisor = tier === "small" ? 2e3 : tier === "medium" ? 3e3 : 4e3;
+        const keepRecent = ctx > 0 ? Math.max(4, Math.min(12, Math.floor(ctx / keepRecentDivisor))) : 12;
+        const maxOutputTokens = ctx > 0 ? Math.min(this.options.maxTokens, Math.max(2048, Math.floor(ctx * 0.25))) : this.options.maxTokens;
+        const toolOutputMaxChars = ctx > 0 ? Math.max(2e3, Math.min(8e3, Math.floor(ctx * 0.5))) : 8e3;
+        const foldLineThreshold = tier === "small" ? 30 : tier === "medium" ? 35 : 40;
+        const foldHeadLines = tier === "small" ? 12 : tier === "medium" ? 16 : 20;
+        const foldTailLines = tier === "small" ? 5 : tier === "medium" ? 8 : 10;
+        const maxSummaryChars = ctx > 0 ? Math.max(2e3, Math.min(8e3, Math.floor(ctx * 0.2))) : 4e3;
+        const repetitionWindow = tier === "small" ? 6 : tier === "medium" ? 8 : 10;
+        return {
+          compactionThreshold,
+          keepRecent,
+          maxOutputTokens,
+          toolOutputMaxChars,
+          foldLineThreshold,
+          foldHeadLines,
+          foldTailLines,
+          maxSummaryChars,
+          repetitionWindow
+        };
+      }
       /** Register a tool for the agent to use */
       registerTool(tool) {
         this.tools.set(tool.name, tool);
@@ -11093,7 +11126,8 @@ Rules:
       detectRepetition(recentToolCalls) {
         if (recentToolCalls.length < 4)
           return 0;
-        const window = recentToolCalls.slice(-8);
+        const { repetitionWindow } = this.contextLimits();
+        const window = recentToolCalls.slice(-repetitionWindow);
         const uniqueKeys = new Set(window.map((tc) => `${tc.name}:${tc.argsKey}`));
         const ratio = 1 - uniqueKeys.size / window.length;
         return ratio;
@@ -11213,8 +11247,7 @@ Integrate this guidance into your current approach. Continue working on the task
             });
           }
           const compacted = this.compactMessages(messages);
-          const ctxWindow = this.options.contextWindowSize;
-          const effectiveMaxTokens = ctxWindow > 0 ? Math.min(this.options.maxTokens, Math.max(2048, Math.floor(ctxWindow * 0.25))) : this.options.maxTokens;
+          const { maxOutputTokens: effectiveMaxTokens } = this.contextLimits();
           const chatRequest = {
             messages: compacted,
             tools: toolDefs,
@@ -11313,8 +11346,7 @@ Integrate this guidance into your current approach. Continue working on the task
                   }
                 }
               }
-              const ctxW = this.options.contextWindowSize;
-              const maxLen = ctxW > 0 ? Math.max(2e3, Math.min(8e3, Math.floor(ctxW * 0.5))) : 8e3;
+              const { toolOutputMaxChars: maxLen } = this.contextLimits();
               const output = result.success ? result.output.length > maxLen ? this.foldOutput(result.output, maxLen) : result.output : `Error: ${result.error || "unknown error"}
 ${result.output.length > maxLen ? this.foldOutput(result.output, maxLen) : result.output}`;
               this.emit({
@@ -11527,8 +11559,8 @@ Integrate this guidance into your current approach. Continue working on the task
                     }
                   }
                 }
-                const maxLen = 8e3;
-                const output = result.success ? result.output.length > maxLen ? result.output.slice(0, maxLen) + `
+                const { toolOutputMaxChars: maxLen2 } = this.contextLimits();
+                const output = result.success ? result.output.length > maxLen2 ? result.output.slice(0, maxLen2) + `
 ...(truncated)` : result.output : `Error: ${result.error || "unknown error"}
 ${result.output}`;
                 this.emit({ type: "tool_result", toolName: tc.name, content: output.slice(0, 200), success: result.success, turn, timestamp: (/* @__PURE__ */ new Date()).toISOString() });
@@ -11629,11 +11661,12 @@ ${marker}` : marker);
       // -------------------------------------------------------------------------
       foldOutput(output, maxChars) {
         const lines = output.split("\n");
-        if (lines.length <= 40) {
+        const { foldLineThreshold, foldHeadLines, foldTailLines } = this.contextLimits();
+        if (lines.length <= foldLineThreshold) {
           return output.slice(0, maxChars) + "\n...(truncated)";
         }
-        const headLines = 20;
-        const tailLines = 10;
+        const headLines = foldHeadLines;
+        const tailLines = foldTailLines;
         const head = lines.slice(0, headLines).join("\n");
         const tail = lines.slice(-tailLines).join("\n");
         const omitted = lines.length - headLines - tailLines;
@@ -11662,13 +11695,11 @@ ${tail}`;
           return sum;
         }, 0);
         const estimatedTokens = totalChars / 4;
-        const ctxWinThreshold = this.options.contextWindowSize > 0 ? Math.floor(this.options.contextWindowSize * 0.75) : Infinity;
-        const effectiveThreshold = Math.min(this.options.compactionThreshold, ctxWinThreshold);
-        if (estimatedTokens < effectiveThreshold) {
+        const limits = this.contextLimits();
+        if (estimatedTokens < limits.compactionThreshold) {
           return messages;
         }
-        const ctxWin = this.options.contextWindowSize;
-        const keepRecent = ctxWin > 0 ? Math.max(4, Math.min(12, Math.floor(ctxWin / 4e3))) : 12;
+        const keepRecent = limits.keepRecent;
         const head = messages.slice(0, 2);
         if (messages.length <= 2 + keepRecent)
           return messages;
@@ -11722,13 +11753,13 @@ ${combinedSummary}
        * When the combined text exceeds the budget, condense the older summary.
        */
       progressiveSummarize(olderSummary, newerSummary) {
-        const MAX_SUMMARY_CHARS = 4e3;
+        const { maxSummaryChars } = this.contextLimits();
         const combined = `${olderSummary}
 
 ---
 
 ${newerSummary}`;
-        if (combined.length <= MAX_SUMMARY_CHARS) {
+        if (combined.length <= maxSummaryChars) {
           return combined;
         }
         const condensed = this.condenseSummary(olderSummary);
@@ -11737,8 +11768,8 @@ ${newerSummary}`;
 ---
 
 ${newerSummary}`;
-        if (result.length > MAX_SUMMARY_CHARS) {
-          const budget = MAX_SUMMARY_CHARS - newerSummary.length - 60;
+        if (result.length > maxSummaryChars) {
+          const budget = maxSummaryChars - newerSummary.length - 60;
           return budget > 200 ? `[Earlier work, condensed]
 ${olderSummary.slice(0, budget)}...
 
@@ -15599,11 +15630,12 @@ async function doSetup(config, rl) {
     const createModelfile = await ask(rl, `  Create optimized model "${c2.bold(customName)}" with ${ctx.label} context? (Y/n) `);
     if (createModelfile.toLowerCase() !== "n") {
       try {
+        const numPredict = Math.min(16384, Math.max(2048, Math.floor(ctx.numCtx * 0.25)));
         const modelfileContent = [
           `FROM ${selectedVariant.tag}`,
           `PARAMETER num_ctx ${ctx.numCtx}`,
           `PARAMETER temperature 0`,
-          `PARAMETER num_predict 16384`,
+          `PARAMETER num_predict ${numPredict}`,
           `PARAMETER stop "<|endoftext|>"`
         ].join("\n");
         const modelDir2 = join24(homedir9(), ".open-agents", "models");
@@ -16005,11 +16037,12 @@ function createExpandedVariant(baseModel, specs, sizeGB) {
   const customName = expandedModelName(baseModel);
   const ctx = calculateContextWindow(specs, sizeGB);
   try {
+    const numPredict = Math.min(16384, Math.max(2048, Math.floor(ctx.numCtx * 0.25)));
     const modelfileContent = [
       `FROM ${baseModel}`,
       `PARAMETER num_ctx ${ctx.numCtx}`,
       `PARAMETER temperature 0`,
-      `PARAMETER num_predict 16384`,
+      `PARAMETER num_predict ${numPredict}`,
       `PARAMETER stop "<|endoftext|>"`
     ].join("\n");
     const modelDir2 = join24(homedir9(), ".open-agents", "models");
@@ -20983,7 +21016,7 @@ function createTaskCompleteTool() {
     }
   };
 }
-function buildTools(repoRoot, config) {
+function buildTools(repoRoot, config, contextWindowSize) {
   const executionTools = [
     new FileReadTool(repoRoot),
     new FileWriteTool(repoRoot),
@@ -21044,11 +21077,11 @@ function buildTools(repoRoot, config) {
   ];
   return [
     ...executionTools.map(adaptTool2),
-    createSubAgentTool(config, repoRoot),
+    createSubAgentTool(config, repoRoot, contextWindowSize),
     createTaskCompleteTool()
   ];
 }
-function createSubAgentTool(config, repoRoot) {
+function createSubAgentTool(config, repoRoot, ctxWindowSize) {
   return {
     name: "sub_agent",
     description: "Delegate a sub-task to an independent agent with its own context window. Each sub-agent creates an independent backend connection, enabling TRUE PARALLEL inference when the backend supports concurrent requests (Ollama with OLLAMA_NUM_PARALLEL > 1). BEST PRACTICE: Launch multiple sub_agent calls with background=true in ONE response to maximize parallelism. Check results via task_status/task_output.",
@@ -21069,13 +21102,18 @@ function createSubAgentTool(config, repoRoot) {
         return { success: false, output: "", error: "task is required" };
       }
       const backend = new OllamaAgenticBackend(config.backendUrl, config.model, config.apiKey);
+      const subCtxWindow = ctxWindowSize ?? 0;
+      const subTier = getModelTier(config.model);
+      const subCompaction = subTier === "small" ? 12e3 : subTier === "medium" ? 24e3 : 4e4;
       const subRunner = new AgenticRunner(backend, {
         maxTurns,
         maxTokens: 16384,
         temperature: 0,
         requestTimeoutMs: config.timeoutMs,
         taskTimeoutMs: config.timeoutMs * 2,
-        compactionThreshold: 4e4
+        compactionThreshold: subCompaction,
+        contextWindowSize: subCtxWindow,
+        modelTier: subTier
       });
       const subTools = [
         new FileReadTool(repoRoot),
@@ -21145,7 +21183,7 @@ function startTask(task, config, repoRoot, voice, stream, taskStores, bruteForce
     // effectively unlimited — no hard timeout, agent runs until complete or aborted
     contextWindowSize: contextWindowSize ?? 0
   });
-  const tools = buildTools(repoRoot, config);
+  const tools = buildTools(repoRoot, config, contextWindowSize);
   if (contextWindowSize && contextWindowSize > 0) {
     for (const tool of tools) {
       if ("setContextWindowSize" in tool && typeof tool.setContextWindowSize === "function") {

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "open-agents-ai",
-  "version": "0.31.2",
+  "version": "0.31.3",
   "description": "AI coding agent powered by open-source models (Ollama/vLLM) — interactive TUI with agentic tool-calling loop",
   "type": "module",
   "main": "./dist/index.js",