fix(curator): add post-archive safety guard to prevent unverified skill archival#30948
Conversation
…ll archival When the curator LLM archives skills without verified consolidation evidence (e.g., hallucinated umbrella names, fallback/no-evidence pruning), active workflows break silently. This adds a post-archive guard that: - Detects skills removed by the LLM review pass - Checks for tool-call evidence of real consolidation (write_file/patch/ create/edit via skill_manage on an umbrella skill) - If NO consolidation evidence exists: moves skills back from .archive/ to the active skills directory (fail-closed behavior) - Logs a warning and annotates the run summary with the rollback count Fixes NousResearch#29912
|
Note: This is a clean rebase of the previously closed PR #30041. The old PR was closed because it had diverged significantly from upstream main (781 commits behind). This version is cherry-picked onto latest main with no conflicts — only the relevant safety guard code is included. The competing PR #30108 (friday-james) takes a different approach (cron-job-based protection). My approach is post-archive tool-call verification — it catches ALL cases of unverified archival, not just cron-related ones. The two approaches are complementary, not mutually exclusive. |
|
Note to maintainers: The alt-glitch bot flagged this as a duplicate. This is incorrect. The original PRs were closed because they were stale (far behind upstream main). This is the clean rebased version cherry-picked onto latest main. The old PRs referenced by the bot no longer exist. |
Summary
When the curator LLM archives skills without verified consolidation evidence (e.g., hallucinated umbrella names, fallback/no-evidence pruning), active workflows break silently.
Changes
Adds a post-archive guard that:
Testing
Fixes #29912