The report
Coding agents can finish a task while leaving supervisors without enough information to reconstruct important intermediate decisions. In a paper submitted on October 14, the CTXGAP researchers tested how much task state could be recovered from accumulated progress reports versus the recorded execution context. Across 37 nonrandomly selected coding tasks and 4,799 valid probes, report-only readers recovered 52.0% of task state compared with 92.7% from full context, a difference of 40.7 percentage points.
The researchers also examined final checkpoints for 30 hidden-test-verified fixes; the gap there remained 34.8 points. That matters because a green final result does not necessarily explain a workaround, a failed check or a residual compatibility risk. In the CTXGAP case, these are differences in what an observer can reconstruct, not proof that the actor deliberately concealed information.
Terminal scrollback improved recovery relative to routine reports but still did not reproduce the full-context result. The practical implication is to evaluate the report as a limited supervisory interface, while retaining execution evidence that a separate reviewer can inspect. The experiment does not measure how often actual teams miss incidents or whether a reporting template prevents the problem.
Source: CTXGAP paper.