FIX_SUMMARY.md

Fix Summary: Bot Repetition and JSON Streaming

Issues Fixed

1. Bot Repetition Issue

Problem: The model was repeating the same tool calls (e.g., creating "orchestrator" agent multiple times) because:

• Context window compaction removed critical "already done" context
• No mechanism to detect repeated agent creation attempts
• Streaming retries could duplicate content

Solution Applied:

• Added repetitionPrevention tracking in run_state.go to count repeated agent.create calls
• After 3 attempts to create the same agent, the system returns an error:

  "repetition detected: agent 'orchestrator' creation was already attempted 3 times"
  

• This breaks the loop and forces the model to check previous results

2. JSON Streaming Display

Problem: Tool arguments weren't visible during streaming - only summary/error/duration were sent

Solution Applied:

• Added "arguments" and "output" fields to the tool_end SSE event in engine.go
• Frontend already has logic to display these (chat.js:892)
• Now JSON tool calls are fully visible during streaming

Using Subagents to Keep Context Fresh

The codebase already has excellent subagent infrastructure. Here's how to use it:

When to Delegate to Subagents

1. Tool iteration limit approaching (120 iterations)
2. Repeated no-progress loops (6+ iterations)
3. Context window approaching limit
4. Cross-domain tasks
5. Long-running analysis

Example: Using agent.run Tool

// Delegate a file search task to a specialized subagent
input := tools.AgentRunInput{
    CallerAgentID: "orchestrator",
    TargetAgentID: "file-searcher",  // Use a specialized agent profile
    Message: `Search for all files matching pattern "config*.go".
              Return only the file paths, one per line.
              Do not perform any other actions.`,
    TaskID:        "search-phase-1",
    Source:        "subagent/orchestrator",
    ThinkingMode:  "off",  // Reduce token usage
}

result, err := subAgentRunner.ExecuteSubAgent(ctx, input)
// result.FinalText contains the structured output

Best Practices

1. Use Task IDs for Traceability

   TaskID: "phase-1-analysis"

2. Configure Agent Profiles

   {
     "agents": {
       "profiles": {
         "file-searcher": {
           "enabled": true,
           "model": {"temperature": 0.1, "max_tokens": 2000}
         }
       }
     }
   }

3. Pass Minimal Context

   • Subagents get fresh context windows
   • Only pass task description + minimal required context
   • Use memory system for cross-agent state

4. Source Tagging

   Source: "subagent/orchestrator"

   Creates clear audit trails in run artifacts

Context Overflow Prevention Strategies

Strategy A: Compaction Before Delegation

// Summarize accumulated tool results
// Extract key findings
// Pass only summary to subagent

Strategy B: Parallel Subagents (Map-Reduce)

// Spawn multiple subagents with independent tasks
// Each gets fresh context window
// Results aggregated by parent

Strategy C: Sequential Pipeline

// Subagent 1: Phase 1 → writes results to file/memory
// Subagent 2: Phase 2 → reads checkpoint, continues
// Each phase starts with fresh context

Key Files Modified

• internal/agent/run_state.go: Added repetition detection
• internal/runtime/engine.go: Added arguments/output to tool_end events

Key Infrastructure Files

• internal/tools/agent_tools.go: Subagent tool implementations
• internal/runtime/engine.go: Subagent runner, context management
• docs/MEMORY_SYSTEM.md: Memory system for context persistence