Codex exposes its telemetry implementation in source
The current codex-otel crate documents provider wiring for logs, traces, and metrics, SessionTelemetry for session-scoped events, metrics APIs, and W3C trace-context helpers. That makes the application layer directly inspectable and gives you a strong semantic source without reverse engineering the CLI.
The final patch still needs execution context for some reviews
A diff does not show which tests failed, what commands retried, which external paths were read, or which remote services influenced the task. Record the generating Codex run and preserve the command sequence, validation timing, file scope, and relevant network activity as context for the source review rather than as a replacement for it.
AgentSight adds a separate system and TLS boundary
AgentSight follows the Codex process family and records supported host activity. For the currently supported stripped rustls path, v1.0.3 uses validated machine-code fingerprints instead of searching for OpenSSL functions. That compatibility layer is tied to binary/compiler output and should be treated as version-specific rather than a permanent Codex ABI.
Correlate native and system timelines around the same tool phase
When a native Codex event identifies a tool or model phase, inspect the same interval in the process tree and file/network views. When the system trace shows a surprising process or destination, return to the native session context to understand which tool or user action led to it. This avoids the two most common errors: inferring intent from syscalls and inferring low-level effects from tool labels.
Use aggregate profiling when one run is no longer enough
agentpprof can aggregate supported local Codex sessions into token, time, operation, file, and network flamegraphs. That is useful for questions such as which task categories consume model budget or which paths dominate across many runs. Keep the semantic tag rules and source session set with any published chart so the result is reproducible.
State capture limits next to any negative conclusion
A bounded AgentSight session shows observed activity in its configured scope; it does not prove that unrecorded behavior never occurred. A native OTel event similarly represents what Codex instrumented. If a security or provenance conclusion depends on absence, reproduce with a targeted second source before treating that absence as proof.
$ sudo agentsight record -- codex
$ agentsight report audit --json