@tjamescouch/gro 1.3.13 → 1.3.15

package/dist/drivers/anthropic.js

@@ -128,6 +128,23 @@ function parseResponseContent(data, onToken) {
     } : undefined;
     return { text, toolCalls, usage };
 }
+/**
+ * Determine if a model supports Anthropic adaptive/extended thinking.
+ * Conservative allowlist approach: if we don't recognize the model,
+ * we omit thinking (safe default — API works fine without it).
+ */
+function supportsAdaptiveThinking(model) {
+    const m = model.toLowerCase();
+    if (/claude-opus-4/.test(m))
+        return true;
+    if (/claude-sonnet-4/.test(m))
+        return true;
+    if (/claude-3[.-]7/.test(m))
+        return true;
+    if (/claude-3[.-]5-sonnet.*20241022/.test(m))
+        return true;
+    return false;
+}
 export function makeAnthropicDriver(cfg) {
     const base = (cfg.baseUrl ?? "https://api.anthropic.com").replace(/\/+$/, "");
     const endpoint = `${base}/v1/messages`;
@@ -141,12 +158,13 @@ export function makeAnthropicDriver(cfg) {
         const { system: systemPrompt, apiMessages } = convertMessages(messages);
         const body = {
             model: resolvedModel,
-            thinking: {
-                type: "adaptive"
-            },
             max_tokens: maxTokens,
             messages: apiMessages,
         };
+        // Only include adaptive thinking for models that support it
+        if (supportsAdaptiveThinking(resolvedModel)) {
+            body.thinking = { type: "adaptive" };
+        }
         if (systemPrompt)
             body.system = systemPrompt;
         // Tools support — convert from OpenAI format to Anthropic format

package/dist/main.js

@@ -487,9 +487,9 @@ function formatOutput(text, format) {
  * the model needing to know the full versioned name.
  */
 const MODEL_ALIASES = {
-    "haiku": "claude-haiku-4-20250514",
-    "sonnet": "claude-sonnet-4-20250514",
-    "opus": "claude-opus-4-20250514",
+    "haiku": "claude-haiku-4-5",
+    "sonnet": "claude-sonnet-4-5",
+    "opus": "claude-opus-4-6",
     "gpt4": "gpt-4o",
     "gpt4o": "gpt-4o",
     "gpt4o-mini": "gpt-4o-mini",

package/dist/memory/advanced-memory.js

@@ -20,7 +20,7 @@ export class AdvancedMemory extends AgentMemory {
         this.highRatio = Math.min(0.95, Math.max(0.55, args.highRatio ?? 0.70));
         this.lowRatio = Math.min(this.highRatio - 0.05, Math.max(0.35, args.lowRatio ?? 0.50));
         this.summaryRatio = Math.min(0.50, Math.max(0.15, args.summaryRatio ?? 0.35));
-        this.avgCharsPerToken = Math.max(1.5, Number(args.avgCharsPerToken ?? 4));
+        this.avgCharsPerToken = Math.max(1.5, Number(args.avgCharsPerToken ?? 2.8));
         this.keepRecentPerLane = Math.max(1, Math.floor(args.keepRecentPerLane ?? 4));
         this.keepRecentTools = Math.max(0, Math.floor(args.keepRecentTools ?? 3));
     }
@@ -42,6 +42,40 @@ export class AdvancedMemory extends AgentMemory {
             createdAt: new Date().toISOString(),
         });
     }
+    /**
+     * Return messages for the API, with hard truncation as a safety net.
+     * Even if background summarization hasn't caught up, this ensures we never
+     * send more than the configured context budget to the driver.
+     */
+    messages() {
+        const budget = this.budgetTokens();
+        const all = [...this.messagesBuffer];
+        const estTok = this.estimateTokens(all);
+        // If under budget, return everything (common case)
+        if (estTok <= budget)
+            return all;
+        // Hard truncation: keep system prompt + most recent messages that fit
+        const result = [];
+        let usedTok = 0;
+        // Always keep the system prompt (first message if system role)
+        if (all.length > 0 && all[0].role === "system") {
+            result.push(all[0]);
+            usedTok = this.estimateTokens(result);
+        }
+        // Walk backwards from the end, adding messages until we hit budget
+        const toAdd = [];
+        for (let i = all.length - 1; i >= (result.length > 0 ? 1 : 0); i--) {
+            const candidate = [all[i], ...toAdd];
+            const candidateTok = this.estimateTokens(candidate);
+            if (usedTok + candidateTok <= budget) {
+                toAdd.unshift(all[i]);
+            }
+            else {
+                break; // No more room
+            }
+        }
+        return [...result, ...toAdd];
+    }
     async onAfterAdd() {
         const budget = this.budgetTokens();
         const estTok = this.estimateTokens(this.messagesBuffer);

package/dist/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@tjamescouch/gro",
-  "version": "1.3.13",
+  "version": "1.3.15",
   "description": "Provider-agnostic LLM runtime with context management",
   "bin": {
     "gro": "./dist/main.js"

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@tjamescouch/gro",
-  "version": "1.3.13",
+  "version": "1.3.15",
   "description": "Provider-agnostic LLM runtime with context management",
   "bin": {
     "gro": "./dist/main.js"