Background sidechain agents consuming tokens without returning results

Bug Description

When starting a new Claude Code session, background \"sidechain\" agents are spawned that consume input tokens but their results are never returned to the main conversation.

Environment

• Claude Code Version: 2.0.61
• OS: Linux (Ubuntu)
• Model: Claude Opus 4.5

Steps to Reproduce

1. Start a new Claude Code session in any project
2. Monitor the ~/.claude/projects/{project}/*.jsonl files for new agent-*.jsonl files
3. Observe that multiple agent files are created immediately on session start

Observed Behavior

When starting a session, multiple agent-*.jsonl files are created with the following characteristics:

{
  "isSidechain": true,
  "sessionId": "96e08ed5...",  // Same as main session
  "output_tokens": 1,
  "stop_reason": null
}

Example from actual session:

• agent-cb4ad2c6.jsonl: 3,703 input tokens consumed
• agent-9845a4a3.jsonl: 3,831 input tokens consumed

These agents:

1. Are marked as isSidechain: true - NOT part of main conversation thread
2. Have stop_reason: null and only 1 output token - never completed
3. Generate responses mentioning tool calls, but these aren't actual tool_use blocks
4. Results are NEVER fed back to the main conversation

Expected Behavior

Either:

1. Background agents should have their results retrieved via AgentOutputTool and integrated into the main conversation
2. OR these agents should not be spawned if their results won't be used

Impact

• Token waste: ~7,500+ input tokens per session start with zero value
• Cost: Unnecessary API costs for users
• Confusion: Activity appears in JSONL logs but not in the UI

Evidence

Agent content shows responses like:

"I'll start by getting recent context from my memory system..."
"I'm in READ-ONLY mode, so I can..."

But these are just text responses - no actual tool execution occurs, and the "context" gathered is discarded.

Additional Context

Discovered while building a JSONL file watcher to monitor Claude Code activity. The watcher showed multiple NEW SESSION [agent-*] entries being created on every session start, with responses that never appeared in the main conversation UI.