pi-dsml

A pi package that executes DeepSeek DSML tool calls which arrive as plain assistant text.

The problem

Some gateways serve DeepSeek models without native tool calling. The model falls back to its own DSML grammar and writes the call into the response body:

I'll check the project setup first.

<｜｜DSML｜｜ calls>
<｜｜DSML｜｜ invoke name="bash">
<｜｜DSML｜｜ parameter name="command" string="true">ls -la</｜｜DSML｜｜ parameter>
</｜｜DSML｜｜ invoke>
</｜｜DSML｜｜ calls>

Pi receives a message with text blocks and no toolCall blocks, so it runs nothing — while the model carries on as if the tool had answered. The turn derails from there.

pi-dsml parses that text back into real tool calls at message_end, which is early enough that pi's agent loop still executes them.

Install

pi install npm:pi-dsml

Or, for local development:

pi -e /path/to/pi-dsml/extensions/index.ts

Grammar coverage

Both the documented form and the mangled form real gateways emit:

The mangled form comes from re-decoding the special tokens: bars double up, a space appears where tool_ should be, and the block tag is left as a bare calls. <｜end▁of▁sentence｜> is stripped too.

Per the spec, string="true" means a raw string and string="false" means a JSON value (number, boolean, array, object).

Behaviour

• Unknown tool names are recovered as-is. Models sometimes invent names (read_file, explore, Bash). Pi answers those with its own Tool X not found result, which the model reads and corrects on the next turn — a better feedback loop than silently dropping the call.
• Messages that already have native tool calls only get the leaked text cleaned up; the parsed copies are discarded so nothing runs twice.
• Fenced code is never touched, so documenting the grammar cannot trigger an execution.
• Truncated blocks still yield whatever complete calls arrived before the cut.
• Thinking blocks are cleaned the same way as text.
• While streaming, a markdown transformer hides DSML spans behind ⟨tool call⟩, so the terminal does not fill with raw markers.

Text with no DSML in it short-circuits immediately, so models with working native tool calls pay nothing.

Configuration

Everything has a default; the file is optional. Create ~/.pi/agent/extensions/pi-dsml.json:

{
  "enabled": true,
  "models": ["deepseek"],
  "hideWhileStreaming": true,
  "debug": false
}

Development

npm install
npm run check   # tsc --noEmit + vitest

test/replay.test.ts replays the real leaked messages found in ~/.pi/agent/sessions, and skips itself when that history is not present.

License

MIT