Messages
User and assistant messages stay in chronological order without Codex environment wrappers.
Codex observability from the session record
Codex CLI Monitoring & Session Replay
Follow a Codex run from the initial instruction through assistant messages, reasoning, function calls and command output. tracehouse turns local Codex session JSONL into a searchable trace you can replay, compare and share.
live trace · refactor-auth · codex-cli
Shareuseragentthinktoolresult
What the Codex monitor reads
The session structure, not a pile of console text
Reasoning summaries
Recorded reasoning summaries appear as their own trace events when the session includes them.
Function calls
Function names and parsed arguments remain attached to the run that invoked them.
Function output
Command and function results are paired back to their call ID for readable replay.
Session timing
Timestamps show the order and pace of long-running Codex work.
Project context
Session metadata connects a trace to its working directory and Codex model context.
How Codex monitoring works
Watch the files Codex already creates
Codex CLI writes rollout files under~/.codex/sessions/. The tracehouse agent follows new JSONL records, recognizes session metadata and converts response items into trace events. You keep using the normal codex workflow.
# install once
$ curl -fsSL https://tracehouse.ai/install.sh | bash
# start Codex normally
$ codex
✓ watching ~/.codex/sessions
Replay the Codex session
Switch from the flat rollout JSONL to a readable transcript and timeline. Expand the exact function input or output only when you need it.
Compare Codex traces
Put two attempts side by side to find changed instructions, extra tool work, timing regressions and different outcomes.
Share the complete run
Create a trace link for review or debugging instead of pasting a partial terminal excerpt into an issue.
Codex traces and token usage
Clear about what the transcript contains
Current Codex rollout records do not provide the same per-response usage object that Claude Code transcripts provide. tracehouse estimates token contribution from recorded text for analysis and labels it as estimated; it does not present that number as a bill.
One monitor for both coding agents
Compare workflows without flattening them
Claude Code and Codex use different transcript formats. tracehouse keeps dedicated parsers for each one and normalizes the result into the same trace vocabulary, so a team can review sessions consistently without pretending the sources are identical.
See Claude Code monitoringMake the next Codex run reviewable
Install the watcher, run Codex CLI normally, then replay the entire session in tracehouse.
Start monitoring Codex