open-agents-ai 0.64.0 → 0.66.0

package/dist/index.js

@@ -13801,7 +13801,7 @@ Rules:
           const keepRecentDivisor = tier === "small" ? 2e3 : tier === "medium" ? 3e3 : 4e3;
           keepRecent = Math.max(4, Math.min(keepRecentMax, Math.floor(ctx / keepRecentDivisor)));
         } else {
-          keepRecent = deep ? 20 : 12;
+          keepRecent = deep ? tier === "small" ? 8 : tier === "medium" ? 14 : 20 : tier === "small" ? 4 : tier === "medium" ? 6 : 12;
         }
         const maxOutputTokens = ctx > 0 ? Math.min(this.options.maxTokens, Math.max(2048, Math.floor(ctx * 0.25))) : this.options.maxTokens;
         const toolOutputCeiling = deep ? 16e3 : 8e3;
@@ -14088,7 +14088,14 @@ Integrate this guidance into your current approach. Continue working on the task
           const choiceContent = response.choices[0]?.message?.content ?? "";
           const choiceArgs = response.choices[0]?.message?.toolCalls?.map((tc) => JSON.stringify(tc.arguments)).join("") ?? "";
           estimatedTokens += Math.ceil((choiceContent.length + choiceArgs.length) / 4);
-          const estimatedContextTokens = Math.ceil(compacted.reduce((sum, m) => sum + (typeof m.content === "string" ? m.content.length : 100), 0) / 4);
+          const estimatedContextTokens = Math.ceil(compacted.reduce((sum, m) => {
+            let chars = typeof m.content === "string" ? m.content.length : 100;
+            if (m.tool_calls) {
+              for (const tc of m.tool_calls)
+                chars += tc.function.arguments?.length ?? 0;
+            }
+            return sum + chars;
+          }, 0) / 4);
           this.emit({
             type: "token_usage",
             timestamp: (/* @__PURE__ */ new Date()).toISOString(),
@@ -14386,7 +14393,14 @@ Integrate this guidance into your current approach. Continue working on the task
             const choiceContent2 = response.choices[0]?.message?.content ?? "";
             const choiceArgs2 = response.choices[0]?.message?.toolCalls?.map((tc) => JSON.stringify(tc.arguments)).join("") ?? "";
             estimatedTokens += Math.ceil((choiceContent2.length + choiceArgs2.length) / 4);
-            const bfEstCtx = Math.ceil(compactedMsgs.reduce((sum, m) => sum + (typeof m.content === "string" ? m.content.length : 100), 0) / 4);
+            const bfEstCtx = Math.ceil(compactedMsgs.reduce((sum, m) => {
+              let chars = typeof m.content === "string" ? m.content.length : 100;
+              if (m.tool_calls) {
+                for (const tc of m.tool_calls)
+                  chars += tc.function.arguments?.length ?? 0;
+              }
+              return sum + chars;
+            }, 0) / 4);
             this.emit({
               type: "token_usage",
               timestamp: (/* @__PURE__ */ new Date()).toISOString(),
@@ -14659,12 +14673,19 @@ ${tail}`;
         if (messages.length < 3)
           return messages;
         const totalChars = messages.reduce((sum, m) => {
+          let chars = 0;
           if (typeof m.content === "string")
-            return sum + m.content.length;
-          if (Array.isArray(m.content)) {
-            return sum + m.content.reduce((s, p) => s + (p.text?.length || 0) + (p.image_url ? 1e3 : 0), 0);
+            chars += m.content.length;
+          else if (Array.isArray(m.content)) {
+            chars += m.content.reduce((s, p) => s + (p.text?.length || 0) + (p.image_url ? 1e3 : 0), 0);
+          }
+          if (m.tool_calls) {
+            for (const tc of m.tool_calls) {
+              chars += tc.function.arguments?.length ?? 0;
+              chars += tc.function.name?.length ?? 0;
+            }
           }
-          return sum;
+          return sum + chars;
         }, 0);
         const estimatedTokens = totalChars / 4;
         const limits = this.contextLimits();
@@ -14758,7 +14779,34 @@ ${fullSummary}
 [Continue from the recent context below. Do not repeat work already completed above.]`
         };
         this.persistCheckpoint(fullSummary);
-        const result = [...head, compactionMsg, ...recent];
+        let result = [...head, compactionMsg, ...recent];
+        const ctxWindow = this.options.contextWindowSize;
+        if (ctxWindow > 0) {
+          const estimateResult = (msgs) => msgs.reduce((sum, m) => {
+            let chars = typeof m.content === "string" ? m.content.length : 100;
+            if (m.tool_calls) {
+              for (const tc of m.tool_calls)
+                chars += tc.function.arguments?.length ?? 0;
+            }
+            return sum + chars;
+          }, 0) / 4;
+          const safetyTarget = Math.floor(ctxWindow * 0.65);
+          let trimmedRecent = [...recent];
+          while (estimateResult(result) > safetyTarget && trimmedRecent.length > 2) {
+            trimmedRecent = trimmedRecent.slice(1);
+            while (trimmedRecent.length > 1 && trimmedRecent[0]?.role === "tool") {
+              trimmedRecent = trimmedRecent.slice(1);
+            }
+            result = [...head, compactionMsg, ...trimmedRecent];
+          }
+          if (trimmedRecent.length < recent.length) {
+            this.emit({
+              type: "status",
+              content: `Post-compaction trim: reduced recent from ${recent.length} to ${trimmedRecent.length} messages to fit ${ctxWindow.toLocaleString()}-token context window`,
+              timestamp: (/* @__PURE__ */ new Date()).toISOString()
+            });
+          }
+        }
         if (result.length < 2)
           return messages;
         return result;
@@ -25423,14 +25471,39 @@ async function handleUpdate(subcommand, ctx) {
 async function switchModel(query, ctx, local = false) {
   try {
     const models = await fetchModels(ctx.config.backendUrl, ctx.config.apiKey);
-    const match = findModel(models, query);
+    let match = findModel(models, query);
     if (!match) {
-      renderError(`Model not found: "${query}"`);
-      renderInfo("Available models:");
-      for (const m of models.slice(0, 10)) {
-        renderInfo(`  ${m.name}`);
+      if (ctx.config.backendType === "ollama") {
+        if (!/^[a-zA-Z0-9][a-zA-Z0-9._:/-]*$/.test(query)) {
+          renderError(`Invalid model name: "${query}"`);
+          return;
+        }
+        renderInfo(`Model "${query}" not found locally. Pulling from Ollama registry...`);
+        try {
+          pullModelWithAutoUpdate(query);
+          const refreshedModels = await fetchModels(ctx.config.backendUrl, ctx.config.apiKey);
+          match = findModel(refreshedModels, query);
+          if (!match) {
+            renderError(`Model "${query}" was pulled but could not be found.`);
+            return;
+          }
+        } catch {
+          renderError(`Model "${query}" could not be pulled from Ollama registry.`);
+          renderInfo("Check that the model name is correct: https://ollama.com/library");
+          renderInfo("Available local models:");
+          for (const m of models.slice(0, 10)) {
+            renderInfo(`  ${m.name}`);
+          }
+          return;
+        }
+      } else {
+        renderError(`Model not found: "${query}"`);
+        renderInfo("Available models:");
+        for (const m of models.slice(0, 10)) {
+          renderInfo(`  ${m.name}`);
+        }
+        return;
       }
-      return;
     }
     let finalModel = match.name;
     if (ctx.config.backendType === "ollama") {
@@ -25464,6 +25537,7 @@ async function switchModel(query, ctx, local = false) {
       const caps = await queryModelCapabilities(ctx.config.backendUrl, finalModel);
       ctx.setCapabilities(caps);
     }
+    ctx.refreshModelCache?.();
   } catch (err) {
     renderError(`Failed to switch model: ${err instanceof Error ? err.message : String(err)}`);
   }
@@ -35012,6 +35086,9 @@ async function startInteractive(config, repoPath) {
     if (ctxSize) {
       resolvedContextWindowSize = ctxSize;
       statusBar.setContextWindowSize(ctxSize);
+      if (activeTask) {
+        activeTask.runner.setContextWindowSize(ctxSize);
+      }
     }
   }).catch(() => {
   });
@@ -35239,9 +35316,26 @@ Rationale: ${proposal.rationale}${provenanceNote}`;
   ];
   const discoveredSkillNames = discoverSkills(repoRoot).map((s) => `/${s.name}`);
   const allCompletions = [.../* @__PURE__ */ new Set([...BUILTIN_COMMANDS, ...discoveredSkillNames])].sort();
+  let cachedModelNames = [];
+  function refreshModelCache() {
+    fetchModels(config.backendUrl, config.apiKey).then((models) => {
+      cachedModelNames = models.map((m) => m.name);
+    }).catch(() => {
+    });
+  }
+  refreshModelCache();
   function completer(line) {
     if (!line.startsWith("/"))
       return [[], line];
+    const modelMatch = line.match(/^\/model\s+(.*)$/i);
+    if (modelMatch) {
+      const partial = modelMatch[1].toLowerCase();
+      const hits2 = cachedModelNames.filter((n) => n.toLowerCase().startsWith(partial));
+      if (hits2.length === 0 && partial.length > 0) {
+        return [[modelMatch[1]], modelMatch[1]];
+      }
+      return [hits2, modelMatch[1]];
+    }
     const lower = line.toLowerCase();
     const hits = allCompletions.filter((c3) => c3.toLowerCase().startsWith(lower));
     return [hits, line];
@@ -35914,6 +36008,7 @@ Respond concisely and safely. Remember: you are talking to the general public.`;
       resolvedCaps = caps;
       statusBar.setCapabilities(caps);
     },
+    refreshModelCache,
     hasActiveTask: () => activeTask !== null,
     requestCompaction(strategy) {
       if (!activeTask)

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "open-agents-ai",
-  "version": "0.64.0",
+  "version": "0.66.0",
   "description": "AI coding agent powered by open-source models (Ollama/vLLM) — interactive TUI with agentic tool-calling loop",
   "type": "module",
   "main": "./dist/index.js",