feat: Lazy skill loading — skills.loading config option

[downwind7clawd-ctrl]

Summary

Implements lazy skill loading as proposed in #2045.

Problem

When many skills are installed, all skill names + descriptions are injected into <available_skills> in the system prompt at every session start. With 87 bundled skills, this block alone consumes ~1,500–2,000 tokens — re-sent with every API call. For messaging platform users (Discord/Telegram) paying per-token through proxies without prompt caching, this adds up fast.

Solution

Add a skills.loading config option:

skills:
  loading: lazy    # "eager" (default) or "lazy"

When lazy:

• The full skill listing is NOT injected into the system prompt
• Instead, a short hint tells the agent to call skills_list() on demand
• Agent calls skill_view(name) as normal after discovering relevant skills
• Saves ~1,500–2,000 system-prompt tokens per turn

When eager (default):

• Current behavior — full listing in system prompt (backward compatible)

Changes (3 files, +21 lines)

File	Change
hermes_cli/config.py	Add skills.loading: "eager" default
agent/prompt_builder.py	Add lazy param + on-demand hint branch
run_agent.py	Read config, pass lazy flag to prompt builder

Token savings estimate

Setup	System prompt tokens	Working headroom (200k)
87 skills always loaded	~8,500	~191k
Lazy loading	~6,500 (-2,000)	~193k

Closes #2045

[wilsonchengassistant]

Hermes Agent 程式碼審核\n\n### 總結\n此 PR 實施了延遲載入技能的功能，符合 #2045 的提案。程式碼乾淨，邏輯清晰，預期能節省系統提示的 token 消耗。\n\n### 優點\n- ✅ 配置選項設計良好，預設為 'eager' 保持向後相容。\n- ✅ 在 lazy 模式下提供明確的提示，引導代理使用 skills_list()。\n- ✅ 變更最小，只影響提示建構和配置讀取。\n\n### 建議\n- 在 tests/agent/test_prompt_builder.py 中新增單元測試，驗證 lazy 模式下提示字串的生成（例如檢查是否包含 'skills_list()' 提示）。\n- 考慮在 README 或文件更新中提及此新配置選項。\n\n### 結論\nLGTM！批准合併。\n\n---\n由 Hermes Agent 審核

[alt-glitch]

Likely duplicate of #13980 — both implement lazy skill loading (skills.loading: lazy) for the same feature request #2045.

[teknium1]

Thanks for targeting a real prompt-size cost.

Problems

• agent/prompt_builder.py:609 removes the catalog entirely. Current main deliberately preserves every skill name even when it compacts descriptions (agent/prompt_builder.py:1464-1468), and the regression test explains why: models do not reliably rediscover skills through skills_list after the index goes quiet (tests/agent/test_prompt_builder.py:429-455).
• The patch has no regression coverage for either config mode.
• The integration points have moved: current prompt assembly is agent/system_prompt.py:314-318, and skill config initialization is agent/agent_init.py:1423-1428 (introduced by 2d2cd5e904).

Suggested changes

• Please get maintainer direction on the discoverability trade-off before salvage; the analogous proposal #19174 was closed on this design concern.
• If the direction is approved, implement it at the current integration points and add eager/lazy prompt-builder coverage.

Automated hermes-sweeper review.

[teknium1]

This removes every skill name from the initial prompt. Current main intentionally keeps all names visible even in compact mode because tests/agent/test_prompt_builder.py:429-455 documents that models do not reliably rediscover skills after the index goes quiet. Please resolve that discoverability contract before adopting this branch.