feat(agent): skill-level model routing via SKILL.md metadata.hermes.model

[davidgut1982]

What

Skills can now declare a preferred model in their SKILL.md frontmatter. When the skill is invoked, the gateway swaps the agent to that model for the duration of the skill turn and restores the original model afterward, so no model preference leaks across turns.

Example SKILL.md frontmatter:

---
metadata:
  hermes:
    model: deepseek/deepseek-v4-flash
---

Resolution order (highest wins):

1. skills.model_overrides["skill-name"] in config.yaml — operator-level override
2. metadata.hermes.model in the skill's SKILL.md frontmatter
3. No swap (session model used as-is)

New utilities (agent/skill_utils.py):

• extract_skill_model_override(frontmatter) — extracts metadata.hermes.model from parsed frontmatter dict
• skill_model_swap(agent, new_model) — lightweight swap: sets agent.model = new_model and clears agent._cached_system_prompt to force system-prompt rebuild on next turn. Does NOT call switch_model() to avoid provider round-trips or cache thrash.
• skill_model_restore(agent, snapshot) — restores the pre-swap model and prompt cache

New lookup (agent/skill_commands.py):

• get_skill_model_for_command(cmd_key, config_overrides=None) — resolves the preferred model for a slash command key (must be slash-prefixed, e.g. "/web-search"). Returns None if no preference is declared.

CLI wiring (cli.py):

• Calls get_skill_model_for_command when a skill command is dispatched from the CLI path

Gateway wiring (gateway/run.py):

• _pending_skill_model_swap dict on GatewayRunner.__init__ keyed by session_key
• _handle_message: resolves and records the pending swap when a skill command is dispatched (before falling through to normal message processing)
• _run_agent: applies the swap just before the agent run; reverts in the finally block

Config (hermes_cli/config.py):

• skills.model_overrides: {} — operator-level map {"skill-name": "model-slug"} added to DEFAULT_CONFIG

Why

Different skills have different complexity profiles. A web-search skill needs a fast cheap model for JSON extraction; a code-review skill may want the strongest available reasoning model. Per-skill model preferences let operators and skill authors express this once in the skill definition rather than requiring users to /model back and forth.

Tests

pytest tests/agent/test_skill_model_routing.py -v

14 tests covering: frontmatter extraction, config override precedence, swap/restore cycle, missing frontmatter passthrough, slash-prefix requirement, and gateway turn-isolation (swap applied, reverted after turn). All pass.

Platforms tested

Linux (CT/LXC environment, Python 3.13).

[davidgut1982]

Merge ordering note: This PR adds skills.model_overrides to DEFAULT_CONFIG but does not bump _config_version. The version bump (24 → 25) lives in #37444 (smart_model_routing), which also touches hermes_cli/config.py. If this PR merges before #37444, a reviewer should add the version bump here as well, or it can be squashed in at merge time. No functional impact either way — just keeping the version consistent.

[teknium1]

This would cause too much confusion to users, feel free to build a standalone plugin that makes this possible if you'd like