Tool calling canary fails for tool-capable models via OpenAI-compatible API (OpenCode client)

[ewanc26]

Using the OpenAI-compatible endpoint (https://cocore.dev/api/v1) as a custom
provider in OpenCode, tool-calling requests fail against multiple models with:

No connected provider supports tool calling for this model. The provider
must be started with vLLM tool calling enabled and pass the startup
forced-tool canary.

Tried against:

• mlx-community/Qwen3.5-4B-MLX-4bit (6 machines)
• mlx-community/Qwen3.6-35B-A3B-4bit-DWQ

Client config declares "tools": true on the model entry, and plain
(non-tool) chat completions work fine against the same models — this only
surfaces once tools/tool_choice are included in the request.

Since OpenCode's core workflow depends on tool calling, this makes any
co/core-backed model currently unusable as an OpenCode provider, regardless
of how many machines are serving it.

[DGaffney]

Thanks for the detailed report — two things were going on here.

Tool calling on co/core is opt-in per machine: an owner flips a "tool calls" switch, and the agent then verifies each model with a forced-tool canary at startup before the machine advertises tool support. When you tested, no connected machine had verified tool-calling for the models you tried, so the gate (correctly, if unhelpfully) rejected the request. Qwen3.5-4B-MLX-4bit now has verified capacity, so that one is worth a retry.

That said, you exposed two real bugs, both fixed in #172: dispatch wasn't tool-aware (a request that passed the gate could still land on a non-tool-capable machine serving the same model and fail mid-stream), and the -DWQ quant of Qwen3.6-35B was missing its parser pairing, so machines serving it could never verify tool calls at all. The first is live once the advisor deploys; the second lands with the next agent release. Availability for a given model still depends on owners opting in, so you may see the error intermittently on thin models until adoption grows.

[DGaffney]

@ewanc26 lmk if this patch worked when you have a min otherwise I can try another approach!

[ewanc26]

will do, thanks!

[codegod100]

Qwen3.5-4B-MLX-4bit now has verified capacity, so that one is worth a retry.

is there a way to see which models have tool support?