--- name: challenge-pass description: Run an investigation's conclusions past a fresh, read-only subagent before sharing them or naming a root cause. It tests each claim for direct observation, prior history, recent change, whole-population coverage, sourcing, a more direct signal, what labels really mean, independent confirmation, fit to the original goal, and internal consistency. Use before publishing analysis, asserting a root cause, or declaring something broken, benign, fixed, safe or harmless, and whenever asked to sanity-check or challenge an analysis. user-invocable: true --- # Challenge pass Usage: `/digits:challenge-pass [claims or analysis text; default = current conclusions]` ## Why Most well-founded objections surface from a handful of pointed questions that compare a conclusion with evidence the agent has not yet looked at. Asking them from a clean context keeps the review from anchoring on the first explanation that seemed to fit. ## When to run - before results leave the session for anywhere shared (chat, ticket, PR, incident); - before naming a root cause; - before recommending a production change, a new ticket, or doing nothing; - before declaring anything missing, broken, harmless, benign, safe, fixed or self-healed. Skip it for plain factual lookups and routine code edits. ## Steps ### 1. Build the claim sheet Four parts: - **Goal** - one or two lines naming the originating question, ticket or incident. - **Claims** - numbered. Each lists its supporting evidence (the exact query, tool or source, what it returned, the time window) and the observation that would disprove it. - **Already examined** - repositories, environments, windows, tickets, threads and sessions covered so far, so the challenger widens coverage instead of repeating it. - **Pointers** - file paths, ticket keys, links, environment or tenant, window. ### 2. Launch the challenger Start one subagent of type `general-purpose` (it needs lookup tools). Its prompt is the complete text of `references/challenger-prompt.md` followed by the claim sheet. If the sheet has roughly nine or more claims, or spans several systems, split the claims across two or three challengers running in parallel. Each one still gets the full goal and context. ### 3. Check the challenger Challenger findings can be confidently wrong. Re-check every finding that would change a conclusion against the source it cites before accepting it. Then put each claim in one class: - **Holds** - no contrary evidence turned up, or new support turned up. - **Changed** - corrected, narrowed or withdrawn, with the evidence. - **Unverified** - blocked on access the agent lacks. Keep the read-only query (or the question to ask) that would settle it, and what each possible answer would mean. ### 4. Report first Open with a short section titled "Challenge pass": what changed and why, what remains unverified, and the questions or queries for the user. Rewrite changed claims in the body of the analysis rather than adding footnotes. Only after that may any post be drafted, under the normal approval rules. If a change was caused by a recurring mistake, propose a lesson and route it through `/digits:self-improve` or a harness edit the user approves.