Record one bounded Codex task around a known repository revision
Use the Codex command you normally run and keep the task, repository SHA, and Codex version available. A bounded task makes process and file attribution tractable and lets the final diff remain the primary code-review surface while the AgentSight run explains the execution history that produced it.
Codex has a first-party OpenTelemetry implementation
The current open-source codex-otel crate provides provider wiring for log, trace, and metric exporters, SessionTelemetry for session-scoped business events, metrics APIs, and W3C trace-context helpers. That native data is the best source for Codex-specific application events because it is emitted where those semantics are known.
AgentSight adds host behavior that native events do not automatically enumerate
A Codex tool or shell step can launch tests, compilers, package managers, Git commands, or other helpers. AgentSight can add the process family, observed path operations, network destinations, and resource behavior around those descendants. Use it to answer “what actually ran on this machine?” rather than duplicating Codex's own semantic fields.
Codex plaintext capture is not an OpenSSL assumption
AgentSight v1.0.3 includes stripped-binary rustls detection for supported Codex builds. The codex_offsets implementation scans for machine-code prefixes associated with rustls plaintext write paths, including vectored writes, and validates surrounding instruction blocks. These fingerprints depend on the compiler and rustls build and should be verified against the exact binary rather than treated as a stable ABI.
Use the run to review validation and retries behind the final patch
Check which test/build commands ran, which failed, what edits followed failures, and whether validation occurred after the last material change. Use file and network context to identify external influences that are not visible in the final diff. Keep conclusions bounded to the recorded task and source revision.
Preserve provenance when exporting into a shared telemetry stack
If Codex native OTel and AgentSight-exported GenAI spans land in the same backend, retain source/service attributes so the two measurement paths remain distinguishable. Codex can provide richer session semantics, while AgentSight can provide independent system context. Merging them conceptually is useful; pretending they were produced by the same instrumentation is not.
$ sudo agentsight record -- codex
$ agentsight report audit --json
$ agentsight report serve