Agent context identification in logs

[djdarcy]

Summary

When a Task subagent (Explore, Plan, etc.) runs tools, those tools should be identifiable in the sesslog as coming from an agent rather than the main session.

Current Behavior

[[2026-02-01 18:56:21]] {Bash: ls -d src/*/ }

No indication whether this Bash command was run by the main session or by a subagent.

Desired Behavior

[[2026-02-01 18:56:21]] {Bash|Explore: ls -d src/*/ }
[[2026-02-01 18:56:31]] {Bash: rm test.txt }  # Main session, no prefix

Format: {Tool|AgentType: content } when running inside a subagent.

Implementation Notes

• Need to investigate PostToolUse JSON structure to determine how agent context propagates
• Agent types include: Explore, Plan, Bash, general-purpose, claude-code-guide
• Model info not needed (user can check transcript if needed)
• Should handle nested agents gracefully

Analysis

See 2026-02-01__19-10-22__session-logger-enhancements-analysis.md for detailed analysis.

[djdarcy]

Progress Update - v0.1.4

✅ Framework Added

• Added format_tool_name(tool_info, tool_input) function for agent-prefixed output
• Added debug logging to investigate available JSON fields in PostToolUse

🔍 Investigation Needed

The PostToolUse hook data doesn't clearly expose agent context. Fields examined:

• tool_info.name - Just the tool name
• tool_input - Tool parameters only
• No obvious agent_type or subagent field

Next Steps

1. Debug logging is enabled - need to trigger Task tool with Explore/Plan agents
2. Check debug log at ~/.claude/logs/hook-debug.log for available fields
3. May need to track Task tool calls ourselves and correlate by timing/session state
4. Alternative: Check if session-env folder contains agent context

The agent prefixing format {Tool|AgentType: ...} is ready once detection is solved.