A Claude Code session's subagent transcripts live beside the session file, not inside it

Fact

Claude Code writes a session to ~/.claude/projects/<encoded-cwd>/<session-id>.jsonl, but a subagent's or workflow agent's own conversation is not appended to that file. It goes into a sibling directory named after the session:

<encoded-cwd>/<session-id>/subagents/<kind>/<group>/agent-<agentId>.jsonl
<encoded-cwd>/<session-id>/subagents/<kind>/<group>/agent-<agentId>.meta.json
<encoded-cwd>/<session-id>/subagents/<kind>/<group>/journal.jsonl

The agent file holds that agent's full turn stream — prompt, thinking, tool calls, results — with isSidechain: true and its agentId on every record. journal.jsonl holds the started / result lifecycle records. The session file itself keeps only the tool call that spawned the run and whatever summary came back; on a store where every subagent ran this way it contains zero isSidechain lines and never mentions an agentId.

Why it matters

Anything that reads a session — a transcript viewer, a token accounting script, an audit of what an agent actually did — silently under-reports when it parses <session-id>.jsonl alone. The parent looks like a short session that made one tool call, while most of the model's output for that turn sits in files the reader never opened. Nothing errors; the numbers are just wrong.

How to apply

Treat the session directory, not the session file, as the unit:

main   = store / project / f"{session_id}.jsonl"
agents = sorted((store / project / session_id).rglob("*.jsonl"))

To join a subagent back to the parent turn that spawned it, use promptId — the parent's tool call and the subagent's first message share it. Do not rely on the parent referencing agentId; it generally does not. Older runs may instead inline sidechain records in the session file, so handle both. See [[claude-code-transcript-human-prompts]] for the other trap in the same files.