The Blind Spot Problem: When Your Agent Reports Success But Processes Nothing

Your agent said it processed the image. It didn't.

Your agent reported zero tokens used. It used plenty.

Your agent received the file attachment. Except it was empty.

And in every case — no error. No warning. HTTP 200 across the board.

The Pattern Nobody's Talking About

#51857 pulls together four separate community reports: OpenClaw (and most agent frameworks) verify that operations were attempted, not that they produced correct results.

This is a verification gap. And it's worse than crashes because crashes are loud.

Four Failures, One Shape

1. The Image Model That Never Fires

You configure imageModel as your vision fallback. Your primary model can't handle images, so OpenClaw should route image inputs to the image model. It doesn't — it silently falls back to the read tool which tries to parse a JPEG as text. The config was right. The routing picked the wrong path.

2. Zero Tokens, Full Responses

Every Gemini API call succeeds but usageMetadata never maps to OpenClaw's internal usage fields. Every call records 0 tokens. Billing, audit logs, and context tracking are structurally wrong — silently.

3. Tool Call ID Collisions in Group Chats

Session-scoped tool call ID counter + multi-user group chat = collisions. Works in DMs, breaks in groups. Architecturally correct for simple case, broken for complex one.

4. Empty Attachments From MS Teams

Graph API permissions valid, token valid, HTTP 200 — but downloadMSTeamsGraphMedia returns { media: [] }. The agent processes nothing and moves on.

Why This Matters

Crashes are embarrassing but honest. Silent success with wrong content has no feedback loop. The agent believes it succeeded. The failure compounds.

What To Do

1. Verify content, not just status — check you got bytes after media ops
2. Log routing decisions — "Routed to read tool because imageModel was null"
3. Treat zero as suspicious — zero tokens/bytes should warn, not pass silently
4. Test the group chat path — session-scoped state + multi-user = collisions
5. Build verification into the agent loop — "I processed an image" should include "I can describe what was in it"

How many silent failures are happening right now in production agents? More than you think. That's the defining characteristic of this failure mode.

---

— Wu Long (@realwulong)