Add: classify-then-act-for-background-agents

[jimy-r]

Adds a third pattern extracted from the same production workspace behind the two already in the catalogue (dead-mans-switch-for-scheduled-agent-jobs, context-budget-as-a-governed-resource).

The pattern: a background agent classifies every queued task into has-default, needs-intent or out-of-scope before acting, builds only the has-default work, and builds it speculatively in a sandbox behind a review queue. The write-up is honest about the operational history. The self-discovery wrapper around this classifier failed in production and was replaced with an explicit delegation queue, while the classifier, the sandbox build and the rejection log carried forward. The Trade-offs section covers what failed and why, with the public retrospective linked. Google's Large-Scale Changes process is cited as the industry precedent for the speculative-build-then-review-queue half.

AI disclosure, per AI_POLICY.md: the draft was written with Claude Code (Fable 5) from the public source repo's PATTERNS.md and CHANGELOG. I reviewed it before submitting, checked every quote and anchor against the source, and approved it personally. The commit carries the Assisted-by trailer. Happy to adjust framing or category if you see it differently.

[vercel]

@jimy-r is attempting to deploy a commit to the nkkko's projects Team on Vercel.

A member of the Team first needs to authorize it.