A commit 7a8c975 dated October 28, 2026 landed in the zju-real/ageval repository. It is a documentation-only commit for an evaluation tool that swaps the agent under test as a plugin so that the same dataset can be run from anywhere. It adds 81 lines and deletes 0 lines, and it touches four files.
The centerpiece is the newly created skills/eval-ops-pitfalls/SKILL.md. It contains six items to check before a run and a procedure for reviewing evaluator configuration: environment image resolution failures caused by a dead local HTTP_PROXY, the relationship between AGEVAL_AGENT_INVOKE_TIMEOUT and acp_timeout, accessibility of the package mirror used via AGEVAL_APT_MIRROR, routing errors depending on model ID and provider prefix, Python package installation blocked by PEP668 (externally-managed-environment), and evaluator dependency and file permission problems. It organizes, in a table, what to check first for each symptom. This skill is hooked into the routing table in skills/ageval-platform/SKILL.md as the entry 'Pre-run environment checks and suite troubleshooting', and 5 lines were added to the table in the CLI-side skills/ageval-cli/references/failures.md so that environment_image_unresolved, acp_timeout, API 400, PEP668, and the case where nearly everything FAIL each point to $eval-ops-pitfalls. The Docker plugin README also changed by 1 line.
The direction is clear. It is a written judgment that you have to filter out environment preparation failures before results can be interpreted, rather than model performance. It also specifies the order: when nearly all tasks fail or a patch is empty, find the common cause through a representative result, fix it, verify with a representative run, and only then run a new suite. That said, this change is entirely 81 lines of documentation. The defaults and supported ranges of the environment variables mentioned by this skill, and the results of reproducing each error code, are not in this commit. Whether these checks alone are sufficient in any given environment can only be judged after more README and release history accumulate.