curate

/curate — confabulation check before you trust a claim

Runs the curator subagent (read-only) over a set of claims and returns a verdict report. This is the in-loop guard against claims that sound done but aren't — fabricated tool results, "prod-ready" work that only passed local smoke, stale audit claims treated as current.

Requires the curator agent (agents/curator.md in this collection) installed to ~/.claude/agents/. The out-of-tree journal lookup in step 1 is optional — it only applies if you also use the branch skill; otherwise that bullet is skipped.

Steps

1. Determine the target (in priority order):

   • An explicit argument — a file path, or a claim/list of claims the user pasted → use it.
   • Else the newest docs/summaries/pause-*.md under the repo root (git rev-parse --show-toplevel).
   • Else the ## Active entry for this session in the shared out-of-tree journal at $(bash ~/.claude/coordination/resolve-coord-dir.sh)/journal.md (the in-tree docs/worktree-journal.md is a tombstone — do not read it).
   • Else this session's own stated accomplishments since the last checkpoint. If you can't find a target, ask the user what to curate — don't curate nothing.

2. Dispatch the curator. Call the Agent tool with subagent_type: curator, passing the target (the file path, or the extracted claims verbatim). The curator reads the repo's invariants doc (e.g. LAWS.md) itself, if one exists. For a deeper-reasoning pass on a high-stakes deploy, you may override its model to opus.

3. Surface the report. Print the curator's report as-is. Lead with CONTRADICTED items — those are the lies. Then the high-impact UNVERIFIED items. Do not bury them under VERIFIED rows.

4. Never auto-fix. The curator is a guard, not a fixer. Present findings; let the human decide what to re-check, correct, or override. If they confirm a recurring confabulation pattern (e.g. "handoffs keep claiming prod-ready off local smoke"), offer to append a one-line OPEN observation to ~/.claude/skill-observations/log.md tagged to the relevant skill, so the pattern accumulates.

Notes

• A state change you can't observe is an invalid transition. Process success is not delivery proof — an exit-0, a green check, or an "ok" message is not evidence that work happened. Every claimed step should point to a concrete observable artifact: a diff, a written file, a passing test's actual output, a live read. If the curator can't locate the artifact behind a claim, the verdict is UNVERIFIED, not done — treat an unobservable transition as a lie until proven.
• Read-only by contract: the curator never edits, deploys, or mutates any store. If you find yourself wanting it to "just fix" something, stop — that's a separate, explicit action the human approves.
• This skill is also invoked automatically by /session-pause (step 4.5) and /session-end (Phase 1.5) so handoffs and session logs are curated before they're trusted.