What does this PR do?

Adds a public reload path for the in-process skill caches so newly installed (or removed) skills become visible mid-session without a gateway restart. The existing internal API agent.prompt_builder.clear_skills_system_prompt_cache(clear_snapshot=True) is now exposed across three surfaces:

1. /reload-skills slash command — works in the CLI (cli.py) and the gateway (gateway/run.py), mirroring the shape of /reload-mcp. Includes /reload_skills alias for Telegram autocomplete and an explicit Discord registration.
2. skills_reload agent tool — exposed in the skills toolset alongside skills_list / skill_view / skill_manage so an agent/subagent can rescan after dropping a new skill via the shell or skills_hub.
3. agent.skill_commands.reload_skills() — shared helper used by all three callers.

The helper clears every in-process skill cache:

• agent.skill_commands._skill_commands (slash-command map)
• agent.prompt_builder._SKILLS_PROMPT_CACHE (in-process LRU)
• ~/.hermes/.skills_prompt_snapshot.json (cross-process snapshot)

…and returns a structured diff:

{
    "added":      ["new-skill"],   # bare names, no leading slash
    "removed":    [],
    "unchanged":  ["..."],
    "total":      37,              # post-rescan skill count
    "commands":   37,              # /slash-skill count
}

When the diff is non-empty, both the CLI and gateway append a short [IMPORTANT: Skills have been reloaded …] note to the end of the conversation history so the model sees the change on its next turn (same pattern /reload-mcp uses).

Related Issue

Aligns with #15626 (plugin reload — same shape, different subsystem) and addresses the runtime cache-staleness called out in #14535 / #14536. No specific skill-reload issue exists yet; happy to file one and link, or fold it under one of the above if a maintainer prefers.

Type of Change

• ✨ New feature (non-breaking change that adds functionality)

Changes Made

Total: +682 / -4 across 10 files.

How to Test

Manual

# 1. Start a hermes chat session
hermes chat

# 2. In another terminal, drop a new skill into ~/.hermes/skills/
mkdir -p ~/.hermes/skills/mid-session-demo
cat > ~/.hermes/skills/mid-session-demo/SKILL.md <<'SKILL'
---
name: mid-session-demo
description: a skill installed during a session
---
body
SKILL

# 3. Back in the chat, run:
/reload-skills

# Expected output:
#   🔄 Reloading skills...
#     ➕ Added: mid-session-demo
#     📚 N skill(s) available
#     ✅ Skill cache cleared

The same flow works inside the gateway (Telegram, Discord, Slack, etc.). For agentic flows, the model can call the new skills_reload tool directly after installing a skill.

Automated

pytest tests/agent/test_skill_commands_reload.py \
       tests/cli/test_cli_reload_skills.py \
       tests/gateway/test_reload_skills_command.py -q
# 16 passed

Surrounding regressions (skills tooling, prompt builder, command registry, gateway slash dispatch) all green:

pytest tests/agent/test_skill_commands_reload.py \
       tests/agent/test_skill_commands.py \
       tests/agent/test_prompt_builder.py \
       tests/cli/test_cli_reload_skills.py \
       tests/gateway/test_reload_skills_command.py \
       tests/tools/test_skills_tool.py \
       tests/skills/ \
       tests/hermes_cli/test_commands.py \
       tests/gateway/test_unknown_command.py -q
# 527 passed, 1 skipped

Checklist

Code

• I've read the Contributing Guide
• Commit follows Conventional Commits
• No duplicate PRs
• PR contains only related changes
• I've run pytest tests/ -q for the touched areas — green (527 passed across surrounding files)
• I've added tests for my changes (16 new cases)
• Tested on my platform: macOS 15 Apple Silicon, Python 3.13.3

Documentation & Housekeeping

• Updated relevant documentation — N/A (slash command is auto-listed in /help via the existing CommandDef registry)
• Updated cli-config.yaml.example if config keys changed — N/A
• Updated CONTRIBUTING.md or AGENTS.md if architecture changed — N/A
• Considered cross-platform impact — works the same on Linux/macOS/Windows; Telegram alias and Discord registration covered
• Updated tool descriptions/schemas if tool behavior changed — skills_reload schema added; existing tools unchanged

Background — why this matters

We're integrating Hermes into a sandboxed orchestrator (NemoClaw-style). When the orchestrator drops a freshly-installed skill into ~/.hermes/skills/ mid-session, Hermes can't see it until restart. The internal clear_skills_system_prompt_cache(clear_snapshot=True) already does the right thing — this PR just wires it to a slash command (so users can invoke it) and an agent tool (so agents that just installed a skill via shell can pick it up themselves, closing the agentic-install catch-22 also raised in #15626).