Selecting Codex on the front page offers Open Codex example, named
codex-coding. This is one live run recorded on 2026-10-08 (UTC) with Codex 0.161.0,
GPT-6-Luna and low reasoning effort, using the interactive recording plugin
with automatic start and save on exit.
The prompt asks the agent to fix interval union duration and peak concurrency, add four regression tests, and run the tests. The trace includes six measured model responses, five scripts and five nested tool calls. Expand Tools → Scripted tool use to follow file inspection, edits and test runs. The completed files are in result. All 12 tests and 500 independent cases passed. The original tests are unchanged.
coding.pftrace retains the captured timestamps, session identity and reported
usage. Scratch workspace paths were replaced with /workspace/intervals.
Input tokens include cached input; the context limit comes from Codex session
metadata. This is a workflow example, not a speed comparison with the other
harness examples, which use different models and timing boundaries.
Context categories were captured from this run's durable session transcript using character-count estimates. They cover observed history with partial attribution; provider instructions and tool definitions remain unavailable.
recording.json contains checksums and validated measurements.
npm run trace:example rebuilds the UI bundle without running a model.
npm run check:example verifies trace import, identity, usage, context counters,
script relationships and the completed task's tests. Refreshing the example
reloads the bundled recording.
To record the task again, copy the starting files from
../pi-opus-5/task into a scratch directory, install the
Codex recording plugin, and run the
prompt with the recording profile. A new recording does not replace this example.