claude-devtools-inspector

Inspect Claude Code session logs by reading JSONL files from ~/.claude/sessions/ to reconstruct readable transcripts. This recovers hidden execution details like exact tool inputs, thinking data, and nested subagent activity without modifying the original logs.

Analyze token usage and context window compaction by reconstructing per-turn data from JSONL logs. The tool breaks down tokens by category and visualizes context windows to show exactly what content was dropped during compaction.

Debug tool calls and agent behavior by recovering exact tool inputs and outputs from prior runs. Reading local session logs allows you to audit actions like Read, Edit, Bash, and Search to understand why specific execution paths were taken.

The best way to export Claude session transcripts to Markdown or JSON is by reconstructing the session data from local JSONL logs. This process extracts tool calls, token attributions, and nested agent trees into structured formats for external review.

Session logs may be missing if Claude Code was not launched with a named session or if Docker mounts are improperly configured. The tool includes troubleshooting guidance to help locate logs in ~/.claude/ and resolve common environment issues.

Session debugging works for nested subagent activity by displaying nested agent trees with cost and duration metrics. It also browses project memory layers to provide full visibility into how subagents interacted during the run.