The transcript formats

Neither Claude Code nor OpenAI Codex CLI documents the files it writes, and the details matter if you want to read them correctly. Pi documents its own and is open source, which helps, but it has traps of its own. These are notes from reading real corpora, written while building bough and each traced to code that was getting something wrong.

• The Claude Code JSONL transcript format, the largest of the three: where the files live, what a record holds, and the four things that will catch you out.
• The Codex CLI rollout file format, read the same way, from a smaller corpus.
• The Pi session file format, written against Pi's own source: a tree rather than a list, and forks that copy everything.

One pattern is worth naming before any of them. All three agents write the same work more than once, so all three overcount badly if you take the files at face value. On Claude a naive line count runs about three times too high. On Codex the input token figure already contains the cached figure, so adding the two doubles every total. On Pi a fork copies the whole conversation into a new file, so everything before the fork is counted twice.