#4837: fix: strip tools for Venice models without function calling support

## Problem Venice AI models that don't support function calling (`hermes-3-llama-3.1-405b`, `venice-uncensored`, `deepseek-v3.2`) receive tools from OpenClaw unconditionally, causing Venice to reject the request with validation errors and return empty responses. Venice's `/v1/models` endpoint correctly reports `supportsFunctionCalling: false` for these 3 models. ## Solution 1. **`src/config/types.models.ts`**: Added optional `supportsFunctionCalling?: boolean` to `ModelDefinitionConfig` 2. **`src/agents/venice-models.ts`**: - Marked 3 models with `supportsFunctionCalling: false` in the static catalog - Updated `buildVeniceModelDefinition()` to propagate the field - Updated `discoverVeniceModels()` to read `supportsFunctionCalling` from the Venice API for newly discovered models - Exported `veniceModelSupportsFunctionCalling()` helper for runtime lookups 3. **`src/agents/pi-embedded-runner/run/attempt.ts`**: When provider is `venice`, checks function calling support before creating tools — strips tools for unsupported models (same behavior as `disableTools`) ## Notes - The pi-ai `Model` interface doesn't support custom fields, so the runtime check uses a direct catalog lookup rather than threading through the model object - Models default to supporting function calling; only explicitly marked models are affected - Build passes cleanly (`pnpm build`) Co-authored-by: jonisjongithub <jonisjongithub@users.noreply.github.com> Co-authored-by: Clawdbot <bot@clawd.bot> <!-- greptile_comment --> <h2>Greptile Overview</h2> <h3>Greptile Summary</h3> This PR adds a `supportsFunctionCalling?: boolean` capability flag to `ModelDefinitionConfig`, annotates known Venice models that lack function calling support, propagates the flag through Venice model discovery, and uses a new `veniceModelSupportsFunctionCalling()` helper to strip tools at runtime for Venice models that can’t handle tool schemas. This prevents Venice request validation errors and empty responses caused by sending tools unconditionally. <h3>Confidence Score: 4/5</h3> - This PR is generally safe to merge and addresses a real provider capability mismatch, with one potential provider-id mismatch to verify. - Changes are localized (Venice model catalog/discovery and tool creation gating) and align with Venice’s reported capabilities. Main remaining risk is whether `params.provider` vs `params.model.provider` can diverge, which could cause tools to be stripped (or not) unexpectedly. - src/agents/pi-embedded-runner/run/attempt.ts (provider check logic) <!-- greptile_other_comments_section --> <!-- /greptile_comment -->