Strata + Oh my Pi reliably failing MCP tool calls

[tannisroot]

My setup:
Strata 0.2.11
Unsloth Qwen 3.8 Flash Next IQ4_XS
Oh my Pi v18.18.0 (same issue in v18.17.0) deployed via Docker using official Dockerfile
Playwright MCP (mcr.microsoft.com/playwright/mcp:v0.0.83)

When asking the model to perform a very simple task:
Grab a screenshot of phoronix.com using Playwright
the model reliably fails to do so and just stops on a malformed tool call:

The user is asking for a screenshot of phoronix.com using Playwright. Let's use the Playwright MCP tool: navigate to the page and take a screenshot.                                                                         

<tool_call>                                                                                                                                                                                                                  
<function=mcp__playwright_browser_navigate>                                                                                                                                                                                  
<parameter=url>                                                                                                                                                                                                              
https://www.phoronix.com                                                                                                                                                                                                     
</parameter>                                                                                                                                                                                                                 
</function>                                                                                                                                                                                                                  
</tool_call>  

This happens 95% with this exact request, very easily reproducible.
The workaround is to set this in the config:

tools:
 xdev: false

However it balloons the context.
Per ChatGPT's suggestion I've made the following APPEND_SYSTEM.md:

## Mandatory tool routing

When using a tool mounted through xdev:

- Never invoke the original tool name directly.
- Always invoke it through the write tool.
- Set write.path to the exact xd:// URI shown in Tool Routes.
- Set write.content to a JSON-encoded string containing the tool arguments.
- Do not put the tool name in the function field.

Example:
write({
path: "xd://<exact-device-uri>",
"content": "{\"argument_name\":\"argument_value\"}"
})

and this reliably fixes it.
But it seems like a hack I should not have to add to fix basic harness functionality.
Is this a bug in the harness I should open an issue about or is there a better, cleaner solution to this?

[Medamineelkhattabi]

I don't see a cleaner fix in v18.18.0 than the two you found. The call gets lost between Strata and oh-my-pi, not in oh-my-pi's dispatcher. What's missing is a way to pin specific tools, which #9019 already asks for (PR #6864 is open), so I'd add your case there instead of opening a new issue.

I checked the v18.18.0 source. MCP tools don't declare a loadMode, so they default to discoverable (essential-tools.ts, sdk.ts). With tools.xdev on, discoverable tools are dropped from the request's tools array and mounted at paths like xd://mcp__playwright_browser_navigate (xdev.ts, L85-L88). So no Playwright function is declared to the server, and the "MCP Tool Routes" section tells the model to write JSON args to each path (mcp-xdev-guidance.md).

Your model wrote a native call for mcp__playwright_browser_navigate instead. Strata 0.2.11 only turns a <tool_call> into a structured call when the request declared that name. Otherwise the text goes back as written, with finish_reason: stop (frontend.py, DETAILS.md). The Strata console should print "did not read as a call of a declared tool and went back as text" along with the name. oh-my-pi just gets a text answer, so the turn ends.

If the call had arrived as a structured tool call, oh-my-pi would have run it. When a call names a mounted device directly, the agent loop falls back to a resolver that maps the name to the tool (agent-loop.ts, sdk.ts). What it lacks is a parser for Qwen's <function=...> markup in plain text.

There is no include list per tool or per server. tools.essentialOverride was removed with no replacement (settings.ts). tools.xdevDocs and tools.xdevInlineDevices only change how much device documentation goes into the system prompt. The tools still aren't declared, so Strata still won't accept the call.

For now that leaves xdev: false or your APPEND_SYSTEM.md. Adding instructions on top of the default prompt is what that file is for (docs). web_search is already kept top-level because models call it directly (xdev.ts, #5973). A Strata + Qwen report on #9019 supports the tools.xdevTopLevelDevices glob from #6864, which would let you pin mcp__playwright_* and keep xdev on.

[tannisroot]

I actually fixed this by setting:

tools:
  format: qwen3

in config.yml
upd: solved one issue and created another

[tannisroot]

my bad - that solved one issue and created another - calls repeatedly stopping and producing nothing.
I'm falling back to the APPEND_SYSTEM.md for now.