feat: add LM Studio JIT load mode

[marcelohildebrand]

Summary

• add model.lmstudio_load_mode with a backwards-compatible default of explicit
• support model.lmstudio_load_mode: jit to skip Hermes' explicit LM Studio preload call and let LM Studio JIT/Auto-Evict manage model loading
• document the new LM Studio config option
• add regression coverage for JIT skip behavior and explicit/default preload behavior

Motivation

LM Studio's Hermes integration docs describe JIT loading support, but Hermes currently preloads LM Studio models through the LM Studio management API before the first chat request. That path is useful for enforcing Hermes' minimum context, but it bypasses LM Studio's normal JIT-loading / Auto-Evict behavior. Users who rely on LM Studio Auto-Evict can hit VRAM pressure when the explicit preload path tries to load a model before LM Studio has a chance to evict the previous one.

This change keeps the current explicit preload behavior as the default while allowing users to opt into LM Studio-managed JIT loading:

hermes config set model.lmstudio_load_mode jit

To restore the existing behavior:

hermes config set model.lmstudio_load_mode explicit

Testing

• uv run --with pytest python -m pytest tests/run_agent/test_lmstudio_load_mode.py tests/run_agent/test_switch_model_context.py -q -o 'addopts='
  • 5 passed
• python -m py_compile run_agent.py agent/agent_init.py tests/run_agent/test_lmstudio_load_mode.py
• git diff --cached --check

Notes

The default remains explicit, so existing LM Studio users should keep the same behavior unless they opt into jit.

[teknium1]

Thanks for the focused opt-in path. The current behavior is still present: run_agent.py:761-775 unconditionally routes LM Studio through ensure_lmstudio_model_loaded(), and the shared helper is reached during initialization (agent/agent_init.py:1751), model switches (agent/agent_runtime_helpers.py:2029-2030), and fallback activation (agent/chat_completion_helpers.py:1606-1607).

Problems

• model.lmstudio_load_mode is documented and read by the PR, but it is not added to the config schema. hermes_cli/config.py:976-980 has no default for it, so schema-driven config consumers cannot discover the setting.
• tests/run_agent/test_lmstudio_load_mode.py only constructs a SimpleNamespace and tests the helper guard. It does not cover config loading through init_agent() into the first preload path.

Suggested changes

• Add the explicit default through the existing model-config normalization path, preserving scalar legacy model config compatibility.
• Add a temp-HERMES_HOME config-propagation regression test for jit and the default/explicit case.

Automated hermes-sweeper review.

[teknium1]

Please add model.lmstudio_load_mode: explicit through the config schema/normalization path as well. The runtime fallback is safe, but leaving the new documented setting out of DEFAULT_CONFIG means schema-driven config consumers cannot discover its default.