Inspect source and generated programs
The model's narration states intent; the admitted PTC-Lisp program is the executable decision you can inspect.
That program shows the functions, tools, data, and return or failure path that were actually chosen. Reviewing it helps debug a surprising result, check authority use, accept or reject a candidate, and improve a prompt.
Inspection does not mutate the active bundle or promote a replacement. Use an explicit inspect, edit, replay, and compare cycle instead.
How do I choose a retrieval surface?
| Need | Start here | Details |
|---|---|---|
| One attached definition | (source agent.prompt/render) | REPL and function references |
| One attached component and its direct dependencies | (component "agent.prompt") | Source inspection and component references |
| Programs a nested loop admitted before the workflow ends | run-outcome with "retain_programs" | Agent library and source inspection |
| Exact programs and component source from a completed run | analysis/read on private inspection | Debug navigation |
What is public versus private?
Public traces carry identities, hashes, dependency projections, and outcomes, but not source. Host logs and telemetry are not source-export surfaces. Exact historical source requires explicitly enabled private inspection.
Use the source-inspection reference for the full selection matrix, return shapes, limits, and privacy rules.