v0.17.0

v0.17.0 Release Summary

What Changed

The ResponseAdapter trait now supports unified extraction of model name, finish reason, token usage, and tool calls across all LLM provider response types. A new ChatResponse::from_response_value method enables dynamic deserialization of unknown JSON payloads into the correct provider variant.

---

Breaking Changes

The ResponseAdapter trait gains 6 new required methods. Any type implementing ResponseAdapter outside of this crate must add the following:

Method	Return type
model_name()	Option<&str>
finish_reason()	Option<&str>
input_tokens()	Option<i64>
output_tokens()	Option<i64>
total_tokens()	Option<i64>
get_tool_calls()	Vec<ToolCallInfo>

ToolCallInfo is a new public struct in potato_type::tools.

---

Changes

ResponseAdapter trait extensions

Six methods added to ResponseAdapter in potato_type::traits. Each provider implements them against its native response schema:

• OpenAI: Reads from choices[0].finish_reason, usage.prompt_tokens / completion_tokens / total_tokens, and choices[*].message.tool_calls.
• Anthropic: Maps stop_reason enum variants to string slices ("end_turn", "max_tokens", "stop_sequence", "tool_use"). Extracts tool calls from content blocks of type ToolUse.
• Google Gemini / Vertex Generate: Reads candidates[0].finish_reason (via new FinishReason::as_str()), usage_metadata token counts, and functionCall parts. FinishReason also gains a Display impl.
• Google Vertex Predict (embeddings): All new methods return None or empty — embeddings have no finish reason, token counts, or tool calls.

ToolCallInfo

New struct in potato_type::tools:

pub struct ToolCallInfo {
    pub name: String,
    pub arguments: serde_json::Value,
    pub call_id: Option<String>,
    pub result: Option<serde_json::Value>,
}

Provides a provider-agnostic representation of a tool call. Each provider's get_tool_calls() maps its native format (OpenAI's stringified JSON arguments, Anthropic's direct Value input, Google's Map<String, Value> args) into this struct.

ChatResponse::from_response_value

Discriminant-based JSON router on the ChatResponse enum. Inspects top-level keys to select the correct variant:

Key(s) present	Deserializes as
choices	OpenAIChatResponse
predictions	PredictResponse (Vertex)
candidates	GenerateContentResponse (Gemini)
stop_reason + content	AnthropicMessageResponse

Returns ProviderError::DeserializationError (new variant) if no pattern matches or deserialization fails.

ChatResponse trait dispatch

ChatResponse now delegates all 6 new ResponseAdapter methods to the inner variant, matching the existing pattern for the other trait methods.

Test coverage

98 new tests across 5 files covering all ResponseAdapter implementations, ChatResponse dispatch, from_response_value routing, and JSON deserialization from raw strings.

---

Upgrading from v0.16.1

1. If you implement ResponseAdapter on custom types, add the 6 new methods listed in Breaking Changes above.
2. Add use potato_type::tools::ToolCallInfo where needed for get_tool_calls() return values.
3. No schema changes, no config changes, no migration required beyond the trait additions.