📖About & Usage

Owner agent — accountability this droid serves

Catching stale scope in freshly-authored initiatives before active work begins. INI-300-20's manager-droids test suite was the canonical regression — the Author drafted against a test landscape that no longer existed.

Droid responsibility

Elrond must guarantee that every initiative entering active status is grounded in present-day Carol reality, not in stale parent-chain context. IV-S1 reads the initiative's requirements + criteria + plan steps, loads a snapshot of the live regression runner, current designs, cookbook, registry, requirements, and SST, and asks Haiku to verdict each item as keep / refine / strike. Flagged items are fed back to the Author for up to two refinement rounds.

What the droid actually does

• Loads the freshly-authored initiative (title, description, requirements, success criteria, plan steps).
• Loads a present-day Carol state snapshot via shared.initiative_context.snapshot_current_state.
• Calls Haiku to verdict each item: keep, refine, or strike, with a one-line comment and (for refine) a concrete replacement.
• Returns flagged items to the orchestrator for Author refinement.
• Round-trip is capped at two rounds; anything still flagged after the cap is logged as a decision row noting that operator review is needed.
• Never edits the initiative itself — only produces verdicts. The Author is the only writer.

Boundaries

• Only verdicts items already present on the initiative; never authors new items.
• Cannot override the Author — only flags items for refinement.
• Round-trip cap is two rounds; after that, the initiative is left for operator review with a clear decision row.
• Validator failure (Claude call or JSON parse) does not block the initiative — the Author's draft remains and a warning is logged.

👤Owner

📚Recent initiatives

Initiatives that touched this droid — a short summary each; open one for the full story.

No initiatives recorded for this droid yet.