Colibri for Agentic Coding Development

[tim-lab-57]

Has anyone gotten Colibri to work with tool calling in a standard agentic environment? I've tried both opencode and pi, but neither seems to work reliably, it usually "attempts" to call tools but fails, and only calls them successfully some of the time.

Anyone gotten colibri working reliably in opencode?

[basil-k-aji-dev]

"Attempts the call, succeeds only sometimes" is a specific failure with a
specific cause, and the server already tells you when you're hitting it.

Tool calls are not a structured channel here — the model emits them as ordinary
text markers that c/openai_server.py parses back into OpenAI tool_calls. GLM
uses <tool_call>{name}<arg_key>k</arg_key><arg_value>v</arg_value></tool_call>;
parse_arch_tool_calls (:558) dispatches to a different parser per
architecture, with the GLM parser as the fallback. Because the syntax is text
produced by a heavily quantized model, it sometimes comes out mangled — and a
strict parse of mangled markup yields zero calls, which is exactly the
intermittency you're describing.

First, check the server's stderr. When tools were declared and markers were
present but nothing parsed, it prints:

[api] tools declared and tool-call markers present, but no call parsed --
output may be quantization-mangled; try COLI_TOOL_SALVAGE=1

(openai_server.py:377-380). If you see that line on the failing turns, the
question is answered: the model is emitting the call, the parse is rejecting it.

Two recoveries exist:

• COLI_TOOL_SALVAGE=1 (:257-259) — opt-in de-mangler for models that drop
  the <arg_key>/<arg_value> structure. It maps a lone payload onto the tool's
  primary parameter, using the schema you sent for the ordering. Off by default
  because it never rewrites well-formed output.
• Unclosed-box recovery, already on — a <tool_call> the model opened and
  never closed, or closed as </tool_cal, is recovered when unambiguous
  (_unclosed_tail, :317-334). Deliberately narrow so prose that merely
  mentions the tag can't be turned into a call.

So the order I'd work in: run the server with stderr visible, reproduce a failing
tool call in opencode, and see whether that diagnostic fires. If it does, set
COLI_TOOL_SALVAGE=1 and retry the same prompt. If it doesn't, the model isn't
emitting markers at all, which is a different problem — prompt/template rather
than parsing — and worth reporting with the model you're running.

Which model are you on? The parser is architecture-specific (deepseek_v4,
deepseek_v41, kimi, qwen38 each have their own; everything else uses the GLM
path), so "colibri + opencode" behaves differently depending on the answer, and
families whose chat template carries no tool syntax at all are a separate case
again.

Read from c/openai_server.py at f028d26.

[MichaelFomenko]

I used Kimi K3 and Qwen Next Flash in OpenCode and it worked. But the Prefilling was extremely slow, maybe there is a way to use the GPU or something to speed the prefilling.