#9212: fix: ensure model.input is always an array for custom providers

## Summary Fixes #8372 - `Cannot read properties of undefined (reading 'includes')` error when using custom providers with custom models. ## Problem When resolving models from custom providers via `modelRegistry.find()`, the returned model may be missing the `input` field. This causes pi-ai to crash when it tries to call `model.input.includes("image")` in `openai-completions.ts`. ## Root Cause In `src/agents/pi-embedded-runner/model.ts`, the `resolveModel()` function returns models from `modelRegistry.find()` without ensuring the `input` field exists. While the fallback paths (inline models and fallback models) correctly set `input: ["text"]`, models found in the registry may not have this field defined. ## Solution Added a safety check after `modelRegistry.find()` to ensure `model.input` is always an array before returning: ```typescript const normalized = normalizeModelCompat(model); // Ensure input field is always present to prevent pi-ai crashes when // model.input.includes() is called on models from custom providers. // See: https://github.com/openclaw/openclaw/issues/8372 if (!normalized.input || !Array.isArray(normalized.input)) { normalized.input = ["text"]; } return { model: normalized, authStorage, modelRegistry }; ``` ## Testing - Tested with custom provider configuration connecting to a local Ollama instance - Previously crashed immediately with `TypeError: Cannot read properties of undefined (reading 'includes')` - After fix: Successfully processes messages without error ## Configuration Example (that was failing before) ```yaml models: providers: local-llm: baseUrl: "http://127.0.0.1:4002/v1" apiKey: "local-llm-key" api: "openai-chat" models: - id: "deepseek-local" name: "DeepSeek R1 8B" contextWindow: 8192 maxTokens: 4096 ``` <!-- greptile_comment --> <h2>Greptile Overview</h2> <h3>Greptile Summary</h3> This change updates `resolveModel()` (`src/agents/pi-embedded-runner/model.ts`) so models returned from `modelRegistry.find()` are normalized and always have a valid `input` array. If a custom provider’s model definition omits `input` (or provides a non-array), the resolver now defaults it to `["text"]`, preventing downstream crashes where code calls `model.input.includes(...)`. This fits the existing resolver behavior: the inline-model and fallback-model paths already supply `input: ["text"]`; this patch brings registry-resolved models in line with those expectations. <h3>Confidence Score: 5/5</h3> - This PR is safe to merge with minimal risk. - The change is narrowly scoped to normalizing a missing/invalid `model.input` field for registry-resolved models, aligning behavior with existing fallback paths and preventing a known runtime crash. No API surface changes and no broader control-flow impact were found in review. - src/agents/pi-embedded-runner/model.ts <!-- greptile_other_comments_section --> _{(2/5) Greptile learns from your feedback when you react with thumbs up/down!} <!-- /greptile_comment -->