From 81ce512705ef27f9b54840e9c175dbf8fffe16c0 Mon Sep 17 00:00:00 2001
From: MoerAI <friendnt@g.skku.edu>
Date: Tue, 19 May 2026 10:13:52 +0900
Subject: [PATCH] fix(agents): declare multimodal-looker tool allowlist in
 prompt to prevent death loop on small VL models (fixes #4116)

The multimodal-looker prompt described what to do but never told the model which tools are available. Smaller VL models (e.g. Qwen3-VL-8B) would try to call non-existent tools and enter an infinite loop emitting:

  Model tried to call unavailable tool 'invalid'. Available tools: call_omo_agent, read.

Add a single sentence at the top of the prompt that explicitly enumerates the only allowed tools ('read' and 'call_omo_agent') and forbids calling any other tool. This matches the runtime allowlist enforced by createAgentToolAllowlist(["read"]).

Regression test asserts the prompt contains the available-tools enumeration so future prompt rewrites don't regress.
---
 src/agents/multimodal-looker.test.ts | 27 +++++++++++++++++++++++++++
 src/agents/multimodal-looker.ts      |  2 ++
 2 files changed, 29 insertions(+)
 create mode 100644 src/agents/multimodal-looker.test.ts

diff --git a/src/agents/multimodal-looker.test.ts b/src/agents/multimodal-looker.test.ts
new file mode 100644
index 000000000..f2f282644
--- /dev/null
+++ b/src/agents/multimodal-looker.test.ts
@@ -0,0 +1,27 @@
+import { describe, test, expect } from "bun:test"
+import { createMultimodalLookerAgent } from "./multimodal-looker"
+
+describe("createMultimodalLookerAgent", () => {
+  test("prompt explicitly enumerates the agent's available tools to prevent death loop on small VL models", () => {
+    // given
+    const agent = createMultimodalLookerAgent("openai/gpt-5-nano")
+
+    // when
+    const prompt = typeof agent.prompt === "string" ? agent.prompt : ""
+
+    // then
+    expect(prompt).toMatch(/available tools/i)
+    expect(prompt).toContain("read")
+  })
+
+  test("prompt instructs the agent never to call other tools", () => {
+    // given
+    const agent = createMultimodalLookerAgent("openai/gpt-5-nano")
+
+    // when
+    const prompt = typeof agent.prompt === "string" ? agent.prompt : ""
+
+    // then
+    expect(prompt.toLowerCase()).toContain("never")
+  })
+})
diff --git a/src/agents/multimodal-looker.ts b/src/agents/multimodal-looker.ts
index 2d89e422b..2a4b80431 100644
--- a/src/agents/multimodal-looker.ts
+++ b/src/agents/multimodal-looker.ts
@@ -23,6 +23,8 @@ export function createMultimodalLookerAgent(model: string): AgentConfig {
     ...restrictions,
     prompt: `You interpret media files that cannot be read as plain text.
 
+Your only available tools are 'read' and 'call_omo_agent'. Always use 'read' to load the file first, then analyze the returned content. Never attempt to call any other tool.
+
 Your job: examine the attached file and extract ONLY what was requested.
 
 When to use you: