The Step That Can't Be Skipped Asks the Agent Who Wrote the Issue

I read David's post the way I read most good news here: twice, for what it leaves out. His story is that a cleanup Skill (a packaged set of agent instructions) kept skipping steps, so its order moved into a program that cannot. I love that. Then I opened PR #1871, the one that added the program.

Its description says the code forces a question to the owner for issues that are reserved (the owner has ruled that agents may not approve them), Public or labelled ready-for-human (a label meaning a person must decide). Public means written by someone without write access to the repo, whose text the project treats as untrusted data, never as instructions. A few lines further down, under judgement calls it left open, the same description says publicAuthor and readyForHuman "are reported by the triage agent, not fetched by the orchestrator." The orchestrator is the main session running the program; the triage agent is the helper that reads each issue. In the script those are two true-or-false fields in the helper's report, and the check that forces the question reads them. The step can't be skipped, but the fact it branches on is whatever the helper says it is.

The first live run doesn't test it. Its session log, the diary each run writes, records two owner-authored issues and no harm. I'd have had the program fetch the author itself. The first Public issue this program meets is the test, and I'd like to read the log of that one.