Description

In the latest stable version (v2026.3.2, build 85377a2) running on Windows 11 (10.0.26200, x64), images received from Telegram are silently discarded when routed to a model configured with the native google-generative-ai API adapter.

The same images work flawlessly when using the openai-completions adapter pointing to the exact same backend model and endpoint.

Steps to Reproduce

1. Configure two providers in openclaw.json:
   • Provider A: api: "google-generative-ai" (native Gemini SDK, /v1beta)
   • Provider B: api: "openai-completions" (OpenAI-compatible, /v1)
   • Both pointing to the same backend model (e.g. gemini-3.1-pro)
2. Send an image via Telegram to the agent
3. Switch the agent's model between Provider A and Provider B

Expected Behavior

Both adapters should include the image payload in the API request, and the model should describe/process the image.

Actual Behavior

• Provider B (openai-completions): Image is included correctly. Model sees and describes the image. ✅
• Provider A (google-generative-ai): Image payload is silently dropped. Model responds as if no image was sent, or says it cannot see any image. ❌

No error is logged. The image simply disappears from the request payload.

Likely Cause

This closely resembles a regression of the previously resolved #25878 (SafeOpenError — stat.dev vs lstat.dev returning 0 on Windows in fs-safe.ts / openVerifiedLocalFile()).

The native google-generative-ai adapter likely reads the file from disk via openVerifiedLocalFile(), which fails the path verification on Windows. The openai-completions adapter presumably uses a different code path (base64 inline or URL reference) that bypasses this check entirely.

Relevant source files:

• src/fs-safe.ts (openVerifiedLocalFile)
• src/agents/pi-embedded-helpers/images.ts (sanitizeSessionMessagesImages)
• src/agents/tool-images.ts (sanitizeContentBlocksImages)

Environment

• OS: Windows 11 (10.0.26200, x64)
• OpenClaw: v2026.3.2 (build 85377a2)
• Node: v22.16.0
• Shell: PowerShell (pwsh)
• Channel: Telegram
• Related closed issue: Bug on Windows: async SafeOpen validation blocks all local media files for reasoning models due to stat.dev !== lstat.dev mismatch #25878

Workaround

Use openai-completions adapter instead of google-generative-ai for image-related tasks on Windows.