PR #75925: fix(lmstudio): skip manual model load when skipModelLoad is enabled [AI-assisted]

• Repository: openclaw/openclaw
• Author: luyao618
• State: closed | merged: False
• Link: https://github.com/openclaw/openclaw/pull/75925

Description (problem / solution / changelog)

🤖 AI-assisted (built with Hermes orchestration). Test level: fully tested. Prompt summary available on request.

Summary

• Problem: When LM Studio JIT model loading is enabled, OpenClaw's call to /api/v1/models/load marks models as manually loaded, preventing LM Studio's JIT auto-unload from functioning.
• Why it matters: Users who rely on JIT to swap between large models on memory-constrained hardware cannot benefit from automatic model lifecycle management.
• What changed: Added a skipModelLoad boolean option to the LM Studio provider config (models.providers.lmstudio.skipModelLoad). When set to true, both inference preload and embedding warmup paths return early without calling the load API. Registered the new field in ModelProviderConfig, Zod schema, generated JSON schema, labels, and help text.
• What did NOT change (scope boundary): Default behavior (false) is fully preserved. No changes to LM Studio model discovery, inference, or embedding logic outside the preload guard.

Change Type (select all)

• Bug fix

Scope (select all touched areas)

• Extensions / Plugins

Linked Issue/PR

• Closes #75921
• This PR fixes a bug or regression

Root Cause

• Root cause: ensureLmstudioModelLoaded unconditionally calls /api/v1/models/load, which marks models as manually loaded in LM Studio, overriding JIT TTL-based lifecycle management.
• Missing detection / guardrail: No config opt-out existed for the preload call.
• Contributing context: The preload was added for reliability but conflicts with LM Studio's newer JIT auto-load/unload feature.

Regression Test Plan

• Coverage level that should have caught this:
  • Unit test
• Target test or file: extensions/lmstudio/src/stream.test.ts
• Scenario the test should lock in: When skipModelLoad: true is set in provider config, ensureLmstudioModelLoaded must not be called.
• Why this is the smallest reliable guardrail: The preload mock assertion directly verifies the skip path without requiring a live LM Studio instance.
• Existing test that already covers this (if any): N/A (new behavior)
• If no new test is added, why not: N/A (test added)

User-visible / Behavior Changes

New optional config key models.providers.lmstudio.skipModelLoad (boolean, default false). When true, OpenClaw does not call /api/v1/models/load before inference or embedding requests.

Diagram (if applicable)

N/A

Security Impact (required)

• New permissions/capabilities? No
• This change only adds an opt-in config flag that skips an outbound HTTP call to the user's own local LM Studio instance. No new secrets, dependencies, or broader permissions.

Changed files

• extensions/lmstudio/src/embedding-provider.ts (modified, +19/-15)
• extensions/lmstudio/src/stream.test.ts (modified, +24/-0)
• extensions/lmstudio/src/stream.ts (modified, +6/-0)
• src/config/schema.base.generated.ts (modified, +11/-0)
• src/config/schema.help.ts (modified, +2/-0)
• src/config/schema.labels.ts (modified, +1/-0)
• src/config/types.models.ts (modified, +2/-0)
• src/config/zod-schema.core.ts (modified, +1/-0)

PR #75939: fix: add sendLoadApi option for LM Studio provider

• Repository: openclaw/openclaw
• Author: 1yihui
• State: closed | merged: False
• Link: https://github.com/openclaw/openclaw/pull/75939

Description (problem / solution / changelog)

Summary

Adds a new provider config option sendLoadApi (default: true) that controls whether OpenClaw sends the /api/v1/models/load request to LM Studio before inference.

When set to false, OpenClaw skips the explicit model loading call, allowing LM Studio's Just-In-Time (JIT) model loading feature to work as intended. This lets LM Studio handle automatic model loading and unloading based on its own TTL and unload settings.

Changes

• Added sendLoadApi?: boolean to ModelProviderConfig type
• Added sendLoadApi: z.boolean().optional() to Zod schema
• Added help text and label for models.providers.*.sendLoadApi
• Modified ensureLmstudioModelLoadedBestEffort in stream.ts to skip load call when sendLoadApi === false
• Modified createLmstudioEmbeddingProvider in embedding-provider.ts to also respect sendLoadApi === false

Testing

• All 68 existing LM Studio extension tests pass
• The fix was verified by checking that ensureLmstudioModelLoaded is not called when sendLoadApi is false

Fixes openclaw/openclaw#75921

Changed files

• extensions/lmstudio/src/embedding-provider.ts (modified, +21/-15)
• extensions/lmstudio/src/stream.ts (modified, +7/-0)
• src/config/schema.help.ts (modified, +2/-0)
• src/config/schema.labels.ts (modified, +1/-0)
• src/config/types.models.ts (modified, +1/-0)
• src/config/zod-schema.core.ts (modified, +1/-0)