[leomontigatti]

This changes what every project's agent does: it no longer adds speculative hardening (a lock, a retry, a second validation, a new state, all against a race or an edge case that is unlikely to happen) on its own, and no longer applies a review finding that asks for it. It brings the change to the user as an ask instead, with the failure mode, how likely it is, and what the fix costs. No project has to add or rename anything.

The term is defined in implement step 2. Both triage rubrics (implement step 5, babysit-pr/triage.md) send such findings to ask, the implementer lists them under risks, and the correctness reviewer's warning now says how likely its condition is. A traced critical finding is never speculative. The idea comes from opencode's AGENTS.md. No text is carried over, and opencode is not a watched upstream, so there's no decisions.md row and no NOTICE entry.

Review

Tier: readback, reviewed 627d9b2

• readback (Luna): 2 findings, 2 declined (likelihood feeds the triager, not the severity; the long line was already reflowed)

🤖 Generated with Claude Code