Fixes #2045.

Summary

Adds a skills.loading config option. When set to lazy, the <available_skills> block is replaced with a one-sentence hint that tells the agent to call skills_list() on demand. Default remains eager (current behavior) for backward compatibility.

Why

With N installed skills, the <available_skills> block costs ~30-50 tokens per skill on every turn — regardless of whether any skill will actually be used. On a 72-skill installation this is ~4,800 tokens per API call, re-sent every message. The cost scales linearly with skill count and becomes a hard ceiling on how many skills a user can reasonably install, especially for messaging-platform users (Telegram / Discord) where prompt caching is unavailable and every token is billed.

Issue #2045 proposed lazy loading: since skill_view() + skills_list() already exist, the index itself can be made on-demand. This PR implements that.

Measured impact (72 skills installed)

Trade-off: the first domain-specific turn in a session spends one extra tool call (skills_list) before the agent picks a skill. Subsequent turns are cached. For chat-heavy workloads (personal assistants, messaging bots) the saving dominates; for one-shot CLI sessions where every installed skill could plausibly be needed, eager stays preferable — which is why it remains the default.

Changes

• hermes_cli/config.py — new skills.loading key, default "eager".
• agent/prompt_builder.py — add _is_skills_lazy() helper; when lazy, build_skills_system_prompt() returns a one-sentence hint instead of the full index.
• website/docs/user-guide/features/skills.md — document the new option with token impact and trade-offs.

No existing behavior changes; the feature is opt-in.

Backward compatibility

• Default is eager — existing installations behave exactly as before.
• skills_list() and skill_view() already exist; no new tools added.
• LRU prompt cache still works — lazy mode has a single cache key.

Test plan

• Syntax check on modified Python files.
• Manual end-to-end on a 72-skill install: system prompt size confirmed to drop from ~4,800 to ~120 tokens.
• Agent correctly invokes skills_list() when asked domain-specific questions; correctly skips when asked casual/emotional questions.
• Add unit test under tests/agent/test_prompt_builder.py to verify both modes — happy to add if maintainers prefer.

Future work

Not in this PR, but natural follow-ups:

• Auto-switch to lazy when installed-skill count crosses a threshold (e.g. > 30).
• platform_loading map so Telegram could be lazy while CLI stays eager for the same user.
• A skills_search(query, k=5) tool that does semantic retrieval over skill descriptions — would further reduce first-turn latency in lazy mode.

Happy to iterate on wording, config key naming, or any of the above.