[Bug][Safety] dsh-llm-pi-ai: malformed tool-call arguments become {} and the Tool runs with empty input

[xiangxiaoyong]

Version: 0.2.0-rc.2 (@deepseek-ai/dsh-llm-pi-ai, which resolves @earendil-works/pi-ai 0.87.x)

Problem — a safety issue. When a provider returns malformed or truncated function-call arguments, for example a stream cut mid-JSON or invalid JSON text, pi-ai parses them into {} and DSH logs and executes the Tool with empty input. The model's actual text is lost, so the Session log cannot show what was sent, and a Tool with all-optional input runs with defaults the model never chose.

Repro. Use any route through dsh-llm-pi-ai, with a provider response whose tool-call arguments is invalid JSON, e.g. {"a": [1, 2. The resulting tool/call has arguments: "{}", and the Tool executes.

What we do today. Two small patches:

• pi-ai keeps ToolCall.rawArguments, the provider's exact argument text (Responses output_item.done, Chat Completions finalize), and replays it verbatim.
• dsh-llm-pi-ai emits rawArguments as the DSH tool-call.arguments string and hands it back on replay.

The 0.2 agent loop then parses the arguments itself, and a malformed call fails Tool validation instead of executing.

Ask. Carry the provider's raw argument text through the adapter, and never substitute {} for unparseable arguments.