open-agents-ai 0.187.353 → 0.187.355

package/dist/index.js

@@ -270535,7 +270535,21 @@ ${body}`;
        */
       microcompact(messages2, recentToolResults) {
         const tier = this.options.modelTier ?? "large";
-        const keepResults = tier === "small" ? 6 : tier === "medium" ? 10 : 20;
+        let keepResults = tier === "small" ? 6 : tier === "medium" ? 10 : 20;
+        let lastAssistant;
+        for (let i2 = messages2.length - 1; i2 >= 0; i2--) {
+          if (messages2[i2].role === "assistant") {
+            lastAssistant = messages2[i2];
+            break;
+          }
+        }
+        if (lastAssistant && typeof lastAssistant._timestamp === "number") {
+          const gapMs = Date.now() - lastAssistant._timestamp;
+          const IDLE_THRESHOLD_MS = 5 * 6e4;
+          if (gapMs > IDLE_THRESHOLD_MS) {
+            keepResults = Math.min(keepResults, 3);
+          }
+        }
         const toolResultIndices = [];
         for (let i2 = 0; i2 < messages2.length; i2++) {
           if (messages2[i2].role === "tool") {
@@ -273167,10 +273181,15 @@ ${errOutput}`;
 ${result.output}`, "utf-8");
         } catch {
         }
-        return `[${toolName} succeeded — output externalized: ${result.output.length} chars, ${lineCount} lines]
-Handle: ${handleId}
-Preview: ${preview}...
-Full content available via: repl_exec(code="data = retrieve('${handleId}')")  or  memex_retrieve(id="${handleId}")`;
+        const { join: _pj } = __require("node:path");
+        const savedPath = _pj(process.cwd(), ".oa", "tool-results", `${handleId}.txt`);
+        const folded = this.foldOutput(result.output, maxLen);
+        return `[Tool output truncated — ${result.output.length} chars, ${lineCount} lines]
+Full output saved to: ${savedPath}
+To read the complete output, use file_read with path="${savedPath}".
+
+Truncated preview (beginning + end):
+${folded}`;
       }
       /**
        * WO-INF-02: Build structured error recovery guidance for small/medium models.
@@ -332807,28 +332826,30 @@ ${lines.join("\n")}
     dynamicContext += `
 
 <vision-capabilities>
-You have vision capabilities available.
+You have vision capabilities. Choose the RIGHT tool for each situation:
 
-IMAGE ANALYSIS:
-  image_read(image="path")  — View/read an image file directly
-  vision(image="path", action="caption")  — Describe image contents
-  vision(image="path", action="query", prompt="question")  — Visual QA
+FOR IMAGE FILES (photos, screenshots, diagrams):
+  image_read(image="path")  — Read an image file (base64 + OCR)
+  vision(image="path", action="caption")  — Describe image contents via Moondream
+  vision(image="path", action="query", prompt="question")  — Ask questions about an image
   vision(image="path", action="detect", prompt="object")  — Find objects (bounding boxes)
-  vision(image="path", action="point", prompt="element")  — Find element center (for clicking)
 
-DESKTOP INTERACTION PIPELINE:
-  1. desktop_describe()  — Screenshot + describe what's on screen
-  2. desktop_click({ target: "element" })  — Find + click an element
-  3. desktop_click({ target: "file.pdf", click_type: "double" })  — Double click to open
-  4. desktop_click({ target: "icon", button: "right" })  — Right click for context menu
-  Chain calls iteratively: describe → click → wait → describe → click next item
+FOR CAMERA (webcam, USB camera):
+  camera_capture()  — Take a photo from the webcam
+
+FOR SCREENSHOTS (what's on screen right now):
+  screenshot()  — Take a screenshot of the desktop
+
+FOR DESKTOP UI INTERACTION (clicking, navigating):
+  desktop_describe()  — Screenshot + describe UI elements on screen
+  desktop_click({ target: "element" })  — Click a UI element by description
 
 RULES:
   - Do NOT say you cannot see images — you can.
-  - Use desktop_describe first to survey the screen, then desktop_click to interact.
-  - After clicking, call desktop_describe again to verify the result.
-  - For file operations: double-click to open, right-click for context menus.
-  - Use delay_ms parameter if UI needs time to transition between clicks.
+  - For reading an image FILE → use image_read or vision, NOT desktop_describe.
+  - For taking a PHOTO → use camera_capture, NOT desktop_describe.
+  - For seeing the SCREEN → use screenshot or desktop_describe.
+  - Only use desktop_describe + desktop_click for UI automation tasks.
 </vision-capabilities>`;
   }
   let localFirstOverride = false;

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "open-agents-ai",
-  "version": "0.187.353",
+  "version": "0.187.355",
   "description": "AI coding agent powered by open-source models (Ollama/vLLM) \u2014 interactive TUI with agentic tool-calling loop",
   "type": "module",
   "main": "./dist/index.js",

package/prompts/compaction/context-compaction.md

@@ -1,9 +1,44 @@
-You are a context compaction assistant. Given a transcript of agent-tool interactions, produce a structured markdown summary that preserves:
-1. **Goal**: What the agent was trying to accomplish
-2. **Completed Steps**: Specific actions taken and their outcomes
-3. **Files Modified**: List of files changed with what was done
-4. **Key Findings**: Important information discovered during exploration
-5. **Errors & Fixes**: Any errors encountered and how they were resolved
-6. **Current Status**: Where things stand right now
-
-Be concise. Use bullet points. Preserve file paths and error messages verbatim. Do NOT include the raw transcript — only the structured summary.
+You are the component that summarizes internal chat history into a structured snapshot.
+
+When the conversation history grows too large, you will be invoked to distill the entire history into a concise, structured XML snapshot. This snapshot is CRITICAL — it will become the agent's ONLY memory of the past. The agent will resume its work based solely on this snapshot. All crucial details, plans, errors, and user directives MUST be preserved.
+
+First, review the user's overall goal, the agent's actions, tool outputs, file modifications, and any unresolved questions. Identify every piece of information essential for future actions.
+
+Then generate the final snapshot. Be incredibly dense with information. Omit conversational filler.
+
+The structure MUST be as follows:
+
+<state_snapshot>
+    <overall_goal>
+        <!-- A single, concise sentence describing the user's high-level objective. -->
+    </overall_goal>
+
+    <key_knowledge>
+        <!-- Crucial facts, conventions, and constraints the agent must remember. Use bullet points. -->
+        <!-- Include: file paths, commands, tool names, patterns discovered, user preferences -->
+    </key_knowledge>
+
+    <file_system_state>
+        <!-- List files created, read, modified, or deleted. Note status and critical learnings. -->
+        <!-- Format: - CWD: /path  - READ: file.ts - purpose  - MODIFIED: file.ts - what changed -->
+    </file_system_state>
+
+    <recent_actions>
+        <!-- Summary of the last significant agent actions and their outcomes. Focus on facts. -->
+        <!-- Include tool names, arguments, and whether they succeeded or failed. -->
+    </recent_actions>
+
+    <errors_and_fixes>
+        <!-- Any errors encountered, their root causes, and how they were resolved. -->
+        <!-- Preserve error messages verbatim where possible. -->
+    </errors_and_fixes>
+
+    <current_plan>
+        <!-- The agent's step-by-step plan. Mark completed steps. -->
+        <!-- Format: 1. [DONE] Step description  2. [IN PROGRESS] Step  3. [TODO] Step -->
+    </current_plan>
+
+    <failed_approaches>
+        <!-- Approaches that were tried and failed. The agent must NOT repeat these. -->
+    </failed_approaches>
+</state_snapshot>