Mateo/dev 25 write connecting arcade tools to your llm page

[torresmateo]

Preview here: https://docs-git-mateo-dev-25-write-connecting-arcade-a84eaa-arcade-ai.vercel.app/en/home/connect-arcade-to-your-llm

---

Note

Introduces a step-by-step Python guide for wiring Arcade tool-calling into an LLM app.

• New setup-arcade-with-your-llm-python guide: project setup with uv, environment config, initializing Arcade and OpenRouter clients, fetching formatted tool definitions, helper for tool authorization/execution, ReAct-style multi-turn loop with tool calls, interactive chat runner, and example code
• Updates app/en/guides/agent-frameworks/_meta.tsx to include the new guide in navigation
• Refreshes public/llms.txt (git-sha/timestamp) and adds a "Connect Arcade to your LLM" entry for discovery

^{Written by Cursor Bugbot for commit 120198a. This will update automatically on new commits. Configure here.}

[vercel]

The latest updates on your projects. Learn more about Vercel for GitHub.

Project	Deployment	Review	Updated (UTC)
docs	Ready	Preview, Comment	Jan 9, 2026 2:04am

[cursor]

Bug: Missing assistant message before tool results in history

When the LLM returns tool calls, the code appends tool result messages to history but never appends the assistant message that contained the tool_calls. The OpenAI API (and compatible APIs like OpenRouter) requires the assistant message with tool_calls to appear in the conversation history before the corresponding tool result messages. This will cause an API error on the next iteration of the loop when the malformed history is sent back to the model.

Additional Locations (1)

• app/en/home/connect-arcade-to-your-llm/page.mdx#L386-L400

 

[nearestnabors]

I need more openrouter tokens to actually test that it works, but wanted to share what I have so far!

[cursor]

Bug: Tool result shown as assistant response when max_turns exhausted

When invoke_llm exhausts max_turns while the assistant is still making tool calls, the function returns with a tool response as the last history item. The chat() function then accesses history[-1]["content"] and prints it prefixed with "🤖 Assistant:", displaying raw tool output as if it were the assistant's response. This produces confusing output when many consecutive tool calls are needed.

Additional Locations (1)

• app/en/home/connect-arcade-to-your-llm/page.mdx#L443-L446

 

[nearestnabors]

Still getting the same malformed directory structure after running the commands in step 1:

[torresmateo]

holding the merge until nav is merged
FYI @nearestnabors

[cursor]

Max turns exceeded causes tool result printed as response

Medium Severity

The chat() function assumes history[-1] is always an assistant message and prints its content as the assistant's response. However, when invoke_llm exits because max_turns is exceeded while still processing tool calls, history[-1] is a tool message with role: "tool". This causes the raw tool result (likely a JSON string) to be displayed to the user as "🤖 Assistant:" output, creating confusing behavior.

🔬 Verification Test

Why verification test was not possible: This edge case requires the LLM to continuously make tool calls until the max_turns limit (5) is reached without providing a final response. This is difficult to trigger reliably in testing as it depends on LLM behavior and requires actual API credentials. The logic flaw is evident from code inspection: when the while loop exits due to turns >= max_turns during tool processing, no assistant message is appended, yet chat() unconditionally treats history[-1]["content"] as the assistant response.

Additional Locations (1)

• app/en/guides/agent-frameworks/setup-arcade-with-your-llm-python/page.mdx#L460-L463