fix(lmstudio): respect applied runtime context

[lunarnexus]

What does this PR do?

Fixes LM Studio context handling so Hermes stops treating 64_000 as both a preload default and a runtime-state sentinel.

The replacement keeps LM Studio runtime state and allocation intent separate:

• already-loaded LM Studio instances keep their reported loaded_instances[].config.context_length, including 64K;
• unloaded explicit-mode loads omit context_length when Hermes has no explicit override;
• explicit Hermes overrides are sent exactly and rejected if they exceed a known LM Studio maximum;
• Hermes requests echo_load_config and only adopts a context length verified by LM Studio's response or a refreshed loaded-instance query;
• lmstudio_load_mode: jit still skips the management preload endpoint.

Related Issue

Fixes #30178
Related to #25989

#25989 asks for JIT lifecycle behavior. This PR preserves existing JIT no-preload behavior but does not redesign post-JIT reconciliation, fallback lifecycle, or multiple-instance routing.

Type of Change

• Bug fix
• New feature
• Breaking change
• Documentation update
• Refactor / cleanup
• Test-only change

Changes Made

• hermes_cli/models.py

  • Preserves active loaded-instance context exactly.
  • Omits context_length for unloaded/no-override explicit loads.
  • Sends exact explicit override values.
  • Rejects explicit overrides above known max_context_length before POST.
  • Uses echo_load_config or one refreshed catalog query before adopting runtime context.

• run_agent.py

  • Removes the forced max(config_context_length or 0, 64_000) LM Studio preload target.
  • Preserves JIT preload bypass.
  • Prevents rejected or unverifiable management loads from falling back to guessed context budgets.

• agent/agent_init.py

  • Passes verified LM Studio startup runtime into the context engine.
  • Limits sub-64K startup allowance to explicit LM Studio context configuration.

• agent/agent_runtime_helpers.py

  • Uses the destination model's configured LM Studio context during direct model switching.
  • Passes verified runtime into the switched model's context engine.

• tests/...

  • Adds focused regressions for loaded 64K authority, omitted cold-load context, exact overrides, above-max rejection, verification-only adoption, JIT no-preload, and direct switch context propagation.

• website/docs/integrations/providers.md

  • Updates LM Studio context behavior docs.

• contributors/emails/git@lunarnexus.com

  • Adds required contributor attribution mapping.

How to Test

Targeted local verification:

scripts/run_tests.sh --file-retries 0 \
  tests/hermes_cli/test_lmstudio_context_policy.py \
  tests/hermes_cli/test_urllib_security.py \
  tests/run_agent/test_lmstudio_load_mode.py \
  tests/run_agent/test_switch_model_context.py \
  -q

Result:

34 tests passed, 0 failed

Additional checks run:

ruff check hermes_cli/models.py run_agent.py agent/agent_init.py agent/agent_runtime_helpers.py tests/hermes_cli/test_lmstudio_context_policy.py tests/hermes_cli/test_urllib_security.py tests/run_agent/test_lmstudio_load_mode.py tests/run_agent/test_switch_model_context.py
python scripts/check-windows-footguns.py --all
uv lock --check
git diff --check HEAD^ --

Results:

Ruff: passed
Windows footguns: passed
uv lock --check: passed
git diff --check: passed

Manual smoke validation performed earlier on Linux with LM Studio: unloaded explicit-mode cold load no longer forced 64K when no Hermes context override was configured.

Checklist

• I have read the contribution guide
• My commit follows Conventional Commit format
• I searched for related issues/PRs before submitting
• This PR is focused and avoids unrelated changes
• Tests pass locally
• Bug-fix tests were added/updated
• Tested on Linux
• Documentation updated where applicable
• No new dependencies
• No secrets, credentials, or environment values added

Security impact

No new credential sources or auth behavior. Existing credential-safe urllib redirect handling remains covered by tests/hermes_cli/test_urllib_security.py.

Scope intentionally deferred

• Post-JIT context reconciliation after inference-triggered loads
• TTL / Auto-Evict lifecycle handling
• Fallback provider activation transactions
• Multiple loaded instances of the same LM Studio model with different contexts

[lunarnexus]

Additional verification / process notes:

• This PR intentionally consolidates fix: use LM Studio max_context_length when model is loaded at default 64K #45037 and fix(lmstudio): use detected context for runtime preload when no override is set #48884 into one reviewable LM Studio context fix.

• The diff is limited to the two LM Studio context paths and their regression tests:

  • agent/model_metadata.py
  • run_agent.py
  • tests/agent/test_model_metadata_local_ctx.py
  • tests/agent/test_lmstudio_runtime_load.py

• Targeted tests pass locally:

  python -m pytest tests/agent/test_lmstudio_runtime_load.py tests/agent/test_model_metadata_local_ctx.py -q -o 'addopts='
  # 26 passed

• Syntax check passes locally:

  python -m py_compile run_agent.py agent/model_metadata.py tests/agent/test_lmstudio_runtime_load.py tests/agent/test_model_metadata_local_ctx.py

• I also ran the preferred wrapper:

  scripts/run_tests.sh

  It timed out after 600s after running 14,333 tests. The observed failures were outside this PR's touched area; one reproduced on untouched origin/main:

  python -m pytest tests/agent/test_copilot_acp_client.py::test_run_prompt_preserves_real_home_when_profile_home_available -q -o 'addopts='

  Untouched origin/main fails with the same HOME assertion, so that failure is not introduced by this PR.

Maintainers can modify the branch if you'd prefer small wording/code adjustments before merge.

[lunarnexus]

Update: the PR branch has now been rebased onto current origin/main and force-pushed cleanly after updating the fork PAT permissions.

Post-rebase local verification:

python -m pytest tests/agent/test_lmstudio_runtime_load.py tests/agent/test_model_metadata_local_ctx.py -q -o 'addopts='
# 26 passed

python -m py_compile run_agent.py agent/model_metadata.py tests/agent/test_lmstudio_runtime_load.py tests/agent/test_model_metadata_local_ctx.py
# passed

[lunarnexus]

Quick maintainer check: should this be reviewed independently from #36852, or folded somehow? This PR covers the runtime preload + loaded-instance fallback path, not just config lookup: run_agent.py::_ensure_lmstudio_runtime_loaded(...) now uses detected context when no explicit override is set, and agent/model_metadata.py::_query_local_context_length(...) falls back to LM Studio max_context_length when the loaded instance is only at the default/minimum 64K.

Focused verification is already in the PR body (26 passed across the two LM Studio regression suites plus py_compile). If there’s a preferred way to reconcile this with #36852, I’m happy to rework it — I mainly want maintainer guidance on whether this should be evaluated as the runtime half of the fix or folded into a different path.

[teknium1]

Thanks for consolidating the two LM Studio paths. The 64K preload behavior is still present on current main (run_agent.py:772), but the proposed runtime-detection heuristic needs rework.

Problems

• Current main deliberately treats loaded_instances[].config.context_length as the actual runtime constraint (agent/model_metadata.py:1751-1757). Commit c030ac1d8520fec3088f10134a82b9560ea712af added that behavior, and tests/agent/test_model_metadata_local_ctx.py:407-442 protects it.
• The PR changes any loaded value <= 64K to the advertised maximum. A loaded 64K instance with a larger max_context_length is therefore indistinguishable from the claimed default case and can be reloaded at the larger value through the new preload target.

Suggested changes

• Keep every loaded-instance context authoritative; split unloaded-model capability discovery from active-runtime detection using an explicit server-state signal.
• Add a regression case for loaded_instances.context_length == 64_000 plus a larger advertised maximum, preserving the active 64K limit.

Automated hermes-sweeper review.

[lunarnexus]

Correction: I narrowed this branch after the replacement grew beyond the reviewer request.

This revision removes the fallback lifecycle, cache invalidation, and broad transaction-snapshot work from the PR. Those are useful follow-ups, but not needed for the core LM Studio context bug.

Current scope is limited to:

• preserving active loaded_instances[].config.context_length, including 64K;
• removing the forced 64K explicit preload target;
• omitting context_length on unloaded/no-override explicit loads;
• sending exact explicit overrides and rejecting above-known-maximum values;
• adopting only LM Studio-verified applied runtime;
• preserving lmstudio_load_mode: jit no-preload behavior;
• carrying the verified runtime through startup and direct /model switching.

The earlier broad rework comment is obsolete and should be ignored in favor of this narrower revision.