Agent Sessions

Every coding agent session (Claude Code or Codex) in one table: what each one took in turns, prompts, tokens, cost and time, and where the agent read and wrote. A row opens the session's turn-by-turn trace.

How to read this table

One row per session, that is one transcript of Claude Code (interactive sessions and claude -p runs alike) or of Codex (its rollout files under ~/.codex/sessions). Click a column heading to sort, type in the search box to filter (prompt, folder, model and session id are searched), pick a folder or a model; the tiles sum what is listed. A row opens the session in the trace page, turn by turn.

Prompts are the messages you typed; turns the API calls they took. Re-sent is the context read from the prompt cache over all turns, the whole conversation re-sent with each call, which is where most of a session's tokens go. New is what the calls added to it (tool results, the agent's own words, the setup and the reminders) plus the few tokens sent uncached; output what the model wrote. Files counts the files read and edited, lines read the lines of those files that tool output returned, edits the edit calls.

Tabs. The sessions table; with several folders, a table per folder (a row lists only that folder's sessions); and the files map.

Cost is the recorded cost when the trace has one, otherwise an estimate at list prices (2026-10-09: cache reads, 1-hour cache writes, uncached input, output), shown in grey. The bar is the share of the most expensive listed session. Active time is the model's time plus its tools' time, without the waits between your prompts; span runs from the first API call to the last.

Built from the transcripts the agents keep (Claude Code under ~/.claude/projects, deleted after 30 days by default; Codex under ~/.codex/sessions) with build_sessions.py from the agent-session-tracer repository; serve.py serves the same table live.