Complete agent traces require an explicit observation boundary

Complete agent traces require an explicit observation boundary. For subscription-authenticated coding agents, capture the documented client event stream through a signed local collector and call it complete only when its server-issued run has a contiguous, acknowledged sequence and a terminal close event. This can include prompts, agent messages, tool calls/results, command output, file diffs, model/usage facts, lifecycle, and emitted reasoning summaries. It cannot establish exact provider wire frames, hidden reasoning, offline edits, separately launched sessions, or all activity on a user-controlled device.

Hooks are useful lifecycle signals but are not a completeness boundary: they can be declined, fail independently, and expose an unstable transcript path. A machine-readable event stream is safer. Keep subscription credentials owned by Codex; never copy or upload auth files. Pass the assessment bearer only to the Codex host process, exclude TOKEN/SECRET variables from model-launched commands, and use a one-time bootstrap instead of placing the bearer in a copied shell command. Upload via strict typed schemas with bounded chunks, idempotent sequence ACKs, local retry spooling, and an explicit GAP state when continuity is lost. If trace completeness is mandatory, stop accepting new turns after the bounded offline buffer cannot be delivered.

Privacy limits: collect only assessment-thread events, normalize paths to the assessment workspace, omit account email/auth payloads, disclose exact collection scope and retention before launch, and expose only emitted reasoning summaries rather than attempting to obtain hidden chain-of-thought.