Module tool_markup 

Fallback parser for tool calls emitted as inline markup instead of the provider’s native function-calling JSON.

Some models — notably DeepSeek served through certain OpenAI-compatible proxies (e.g. opencode-go) — emit tool calls as XML-ish markup inside the assistant content stream rather than as OpenAI tool_calls:

<｜｜DSML｜｜tool_calls>
<｜｜DSML｜｜invoke name="read">
<｜｜DSML｜｜parameter name="file_path" string="true">config.py</｜｜DSML｜｜parameter>
</｜｜DSML｜｜invoke>
</｜｜DSML｜｜tool_calls>

Anthropic-style <invoke name="..."> blocks use the same shape without the ｜｜DSML｜｜ token. When the OpenAI-compatible layer sees this in content (with finish_reason: "stop"), the tool would otherwise leak to the user as raw text and never execute. We detect and normalize both forms into real tool calls.

Structs

ParsedToolCall

Functions

extract_tool_calls

Extract every <invoke name="X"> … <parameter name="Y">Z</parameter> … </invoke> block (tolerating the ｜｜DSML｜｜ token prefix) into structured tool calls.

looks_like_tool_markup

Cheap pre-check: does this text contain inline tool-call markup we can parse? Used to (a) decide whether to suppress the raw text from the user and (b) whether to run the (more expensive) full extraction at stream end.