hermes - 💡(How to fix) Fix Curator may archive active skills during umbrella pass without verified consolidation (fail-open behavior) [4 pull requests]

Bug description

The curator can archive currently active/operationally critical skills during umbrella consolidation, even when evidence for successful consolidation is missing or inconsistent.

In our run, 10 operational skills were archived in one pass, including:

• facebook-high-signal-routine
• linkedin-high-signal-routine
• google-photos-memory-ops
• cron-run-error-recovery

This breaks active workflows because the original skill names are removed from normal lookup and the replacement umbrella may not be a drop-in equivalent.

Evidence

Curator report:

• ~/.hermes/logs/curator/20260521-132901/REPORT.md
• ~/.hermes/logs/curator/20260521-132901/run.json

Key contradictions from run.json:

• counts.before: 10, counts.after: 0, archived_this_run: 10
• consolidated_this_run: 0
• cron_jobs_rewritten: 0
• pruned[*].source: "fallback (model named missing umbrella, no tool-call evidence)"

So all candidates were archived, but no measured consolidations happened and no cron rewrites were applied.

Reproduction (high-level)

1. Have a set of agent-created skills that are still referenced in real operational routines.
2. Run curator pass with consolidation logic enabled.
3. Observe full-wave archive into ~/.hermes/skills/.archive/.
4. Observe active skill names no longer available under normal skill lookup path.

Actual behavior

• Curator archives a whole cluster of active skills in one run.
• Report summary says umbrella consolidation happened.
• Machine counters show zero consolidations and fallback/no-evidence pruning paths.
• Cron references are not rewritten (cron_jobs_rewritten: 0).

Expected behavior

• Never archive skill names that are currently in active use unless safe migration is verified first.
• If umbrella creation fails or is not evidenced, fail closed: do not archive originals.
• Require hard validation before archive:
  • replacement skill exists and passes lookup,
  • reference rewrites (cron + other references) succeeded,
  • post-check proves no unresolved references.
• If any check fails, keep originals active and emit warning-only report.

Impact

High operational risk: active automations lose known skill names and can silently degrade/fail.

Suggested fix directions

1. Add pre-archive guard: block archive when consolidated_this_run == 0 for planned absorption set.
2. Add mandatory reference rewrite/validation gate before delete/archive.
3. Treat fallback ... no tool-call evidence as non-destructive (no archive), only warning.
4. Add a strict "dry-run + explicit approve" mode for umbrella-wave archival.

Environment

• Hermes profile: default
• Curator status at observation time: enabled
• Run ID: 20260521-132901
• Local path: /home/sancho/.hermes/logs/curator/20260521-132901/