Part of the experimental Factory suite of builder-skills. Reads .agent-factory/config.yaml (artifact-autonomy-policy) and its own optional skill_prompts overlay, which adds project guidance but cannot grant permission. Experimental even within the Factory.
A read-only decision queue: every configured item the agent-handled flows did not resolve and a person still has to decide on. By default it covers all configured categories for the last 7 days at balanced detail. It can be narrowed in plain language (“PRs only, last 30 days, detailed”), but a filter “cannot expand the repositories, source scopes, or tools in the project config.”
It includes PRs waiting for human review, approval, a merge decision, or product judgment; issues, feedback, or errors held for missing context, unclear intent, risk, or conflicting evidence; new answers to information requests that have not been re-triaged; and recurring incidents or delivery work that stopped with no safe authorized next step. It excludes items with a verified terminal outcome unless they reopened. “A label or agent summary alone does not prove an item is resolved.”
Related items are grouped “by the underlying user impact or failure pattern, not by similar words alone”, with every source link kept and report volume separated from the number of affected people. Output is at one of three levels: brief, balanced, or detailed.
It never acts. “Do not reply, approve, merge, assign, close, change status, or notify people from the digest.” Any decision is routed back through the module that owns that action and its separate policy. This is the factory anatomy’s escalation path built as a capability. Off-stage: it tracks decisions that belong to other stages without advancing any of them.
See Also
- beads —
bd human, the store-layer version of a human-needed queue. - topic-software-factory — escalation in the glossary; the human socket in the anatomy diagram.
- builder-factory — the policy that decides what gets held.