extract-conversations

Extract conversation data and save to a temporary file for analysis:

python3 scripts/extract_session_data.py [project_dir] > /tmp/conversations.json

Output Schema

{
  "project": "string — project path, or \"ALL\" when --all is used",
  "generated_at": "string — ISO 8601 UTC timestamp",
  "conversation_count": "number",
  "session_count": "number",
  "conversations": [
    {
      "slug": "string | null — conversation topic, used to group sessions. null for ungrouped sessions",
      "sessions": [
        {
          "session_id": "string — unique session identifier",
          "file": "string — absolute path to the source JSONL file",
          "is_continuation": "boolean — true if this session continues a previous one (compaction restart)",
          "version": "string | null — Claude Code version",
          "git_branch": "string | null — active git branch during the session",
          "start": "string | null — ISO 8601 timestamp of first entry",
          "end": "string | null — ISO 8601 timestamp of last entry",
          "turns": [
            {
              "index": "number — zero-based turn index within the session",
              "uuid": "string | null — unique turn identifier from the log entry",
              "timestamp": "string | null — ISO 8601 timestamp of the user message",
              "duration_seconds": "number | null — wall-clock seconds from user message to end of Claude's response",
              "user_text": "string — the user's message text",
              "interrupted": "boolean — true if the user interrupted Claude's response",
              "post_interrupt_text": "string — (only present if interrupted) the user's message after interrupting",
              "estimated_input_tokens": "number — (only present for interrupted turns with no API response) estimated from nearest turn with token data",
              "model": "string | null — primary model used (e.g. \"claude-sonnet-4-6\")",
              "api_calls": "number — deduplicated API call count for this turn",
              "tokens": {
                "input": "number — uncached input tokens",
                "cache_read": "number — tokens read from cache",
                "cache_write": "number — tokens written to cache",
                "output": "number — output tokens generated"
              },
              "tools": "object — {ToolName: count} map, e.g. {\"Edit\": 3, \"Bash\": 1}. Empty object if no tools used",
              "skills": [
                {
                  "name": "string — skill name",
                  "source": "string — \"cli\" (user typed /skill) or \"model\" (Claude invoked via Skill tool)"
                }
              ],
              "errors": ["string — error messages from bash failures and tool errors"],
              "subagents": [
                {
                  "agent_id": "string | null — agent identifier",
                  "description": "string — short description from the Agent tool call",
                  "completed_in_turn": "number | null — turn index where the agent completed, null if same turn",
                  "file": "string | null — absolute path to the subagent's JSONL log file",
                  "model": "string | null — primary model used by the subagent",
                  "api_calls": "number — subagent's API call count",
                  "duration_seconds": "number | null — subagent wall-clock duration",
                  "tokens": {
                    "input": "number",
                    "cache_read": "number",
                    "cache_write": "number",
                    "output": "number"
                  },
                  "tools": "object — {ToolName: count} map",
                  "skills": [{"name": "string", "source": "string"}],
                  "errors": ["string"]
                }
              ]
            }
          ]
        }
      ]
    }
  ]
}

Conditional fields: post_interrupt_text is only present on interrupted turns. estimated_input_tokens is only present on turns interrupted before any API response. subagents is only present on turns that launched agents. All other fields are always present.

Useful flags:

• --from YYYY-MM-DD / --to YYYY-MM-DD — filter by date range
• --no-subdirs — exclude worktrees
• --all — scan all projects
• --session ID — filter to a specific session

After extracting, use jq to query the data. Examples:

# All interrupted turns with post-interrupt text
jq '[.conversations[].sessions[].turns[] | select(.interrupted) | {user_text, post_interrupt_text, duration_seconds}]' /tmp/conversations.json

# Skill usage frequency
jq '[.conversations[].sessions[].turns[].skills[].name] | group_by(.) | map({skill: .[0], count: length}) | sort_by(-.count)' /tmp/conversations.json

# Turns with errors
jq '[.conversations[].sessions[].turns[] | select(.errors | length > 0) | {user_text: .user_text[:80], errors}]' /tmp/conversations.json

# Subagent summary
jq '[.conversations[].sessions[].turns[].subagents[]? | {description, duration_seconds, api_calls, model}]' /tmp/conversations.json

To dig deeper into a specific turn, use the turn's uuid and the session's file:

# Find the full exchange starting from a turn's UUID
jq -c 'select(.uuid == "TARGET_UUID" or .parentUuid == "TARGET_UUID")' SESSION_FILE.jsonl