/insights includes Claude Agent SDK sessions, skewing results

Description

/insights analyzes all session JSONL files in ~/.claude/projects/, but this directory also contains sessions created by the Claude Agent SDK (via claude-agent-sdk-python). There's no way to distinguish between interactive Claude Code sessions and programmatic Agent SDK sessions, so /insights mixes them together, producing inaccurate usage reports.

How it happens

The Agent SDK's SubprocessCLITransport spawns the Claude Code CLI as a subprocess, passing cwd from ClaudeAgentOptions. The CLI then stores session files in ~/.claude/projects/{encoded-cwd}/ — the same location it uses for interactive Claude Code sessions.

For example, an Agent SDK app with cwd=/Users/me/.observe-agent/observe-agent-chat-sessions creates sessions at:

~/.claude/projects/-Users-me--observe-agent-observe-agent-chat-sessions/*.jsonl

These are indistinguishable from regular Claude Code sessions.

Impact

In my case, I had 159 Agent SDK sessions mixed in with my Claude Code sessions. /insights reported 197 total sessions, but a significant portion were from a programmatic agent, not interactive use. This skewed:

• Project area analysis — agent runs were categorized as if I was doing that work interactively
• Friction analysis — friction patterns from the agent's behavior were attributed to my interactive workflow
• Interaction style narrative — session counts and patterns reflected agent behavior, not my usage

Expected behavior

One or more of:

1. /insights should distinguish between Claude Code sessions and Agent SDK sessions (the CLI sets CLAUDE_CODE_ENTRYPOINT=sdk-py via env — this could be recorded in session metadata)
2. /insights should have a flag to exclude specific project directories
3. The Agent SDK should store sessions in a separate directory (e.g., ~/.claude/sdk-sessions/) rather than sharing ~/.claude/projects/

Environment

• Claude Code: 2.3.1
• claude-agent-sdk-python: 0.1.19
• OS: macOS (Darwin 25.2.0)