feat(skills): add stats/archive/restore/prune CLI subcommands

[elmatadorgh]

Closes #19384

Summary

Adds the four user-facing CLI verbs requested in #19384, exposing the curator's existing tools/skill_usage.py telemetry through hermes skills:

• hermes skills stats [--days N] — ranked usage table (uses, views, patches, last activity, idle days, state, pinned)
• hermes skills archive — archive a skill; refuses if pinned, hints at hermes curator unpin
• hermes skills restore — restore from .archive/
• hermes skills prune [--days N] [--yes] [--dry-run] — bulk-archive idle skills (default 90 days)

All four handlers are thin wrappers over functions that already exist on main: agent_created_report(), archive_skill(), restore_skill(), get_record(). No new storage, no curator behavior changes, no per-message ranking added to the system prompt.

Notes

• Pinning (PR feat(skills): refuse skill_manage writes on pinned skills #17562) is honored: archive refuses pinned skills with a hint pointing at hermes curator unpin , and prune filters them out.
• prune falls back to created_at when last_activity_at is null, so never-used skills can be pruned (otherwise they'd be immortal).
• --days 0 and negative values rejected with exit code 2 (argparse-style user error).
• archive and restore exit non-zero on failure for shell-script friendliness; prune reports per-skill results in a summary line.
• Rich is imported lazily inside _cmd_stats to keep hermes --help fast, with a plain-text fallback if rich is unavailable.

Relationship to #4406

@fathah's #4406 prototyped the same CLI surface against an older SQLite-based curator design. The CLI shape here mirrors that work; the storage backend uses the merged tools/skill_usage.py (sidecar JSON) instead. The ranking/keyword-relevance pieces from #4406 are intentionally not included, per the issue.

Tests

• 8 argparse smoke tests added to tests/hermes_cli/test_skills_subparser.py
• 22 handler tests in new tests/hermes_cli/test_skills_lifecycle.py covering empty/non-empty reports, --days filtering, sort order, pinned refusal, dry-run, confirmation prompts, exclusion rules, and the created_at fallback for never-used skills

tests/hermes_cli/test_skills_subparser.py: 14 passed
tests/hermes_cli/test_skills_lifecycle.py: 22 passed

No regressions in the broader tests/hermes_cli/ run; pre-existing Windows console / missing-deps failures unaffected.

Manual smoke test

Tested on Windows (WSL2 / PowerShell), Python 3.11. hermes skills stats on a workspace with 23 agent-created skills renders the rich table correctly. hermes skills --help and each subcommand's --help render without traceback.

How to test

Run the new tests:

pytest tests/hermes_cli/test_skills_subparser.py tests/hermes_cli/test_skills_lifecycle.py -v

Smoke test the CLI:

hermes skills --help
hermes skills stats
hermes skills prune --dry-run --days 30

The pinned-skill refusal can be exercised by pinning a skill via hermes curator pin first, then running hermes skills archive — should refuse with a hint pointing at hermes curator unpin.

[teknium1]

Adapted and merged via PR #20200 (#20200), closing #19384.

The new verbs — hermes curator archive and hermes curator prune — live under the existing hermes curator namespace instead of a parallel hermes skills surface. The stats and restore parts of your PR duplicated existing curator status and curator restore commands (which landed after the issue was written), so only the two genuinely new verbs were kept, rewritten to match hermes_cli/curator.py's plain-text handler style.

Your authorship is preserved via Co-authored-by: elmatadorgh <coktinbaran5@gmail.com> on the squash commit (436672d). AUTHOR_MAP entry added. Thanks for driving this to a working implementation!