openai_compat runs tools on default arguments when the stream truncates; Anthropic refuses to

[yuanhao]

Found while adding SSE tool-call coverage in #166, and pinned there as a test rather than changed — closing it is a behavioural decision, not a test fix.

The divergence

Two providers disagree about what to do when a tool call's arguments cannot be parsed.

anthropic.rs fails the turn, and states why at the sweep:

a tool handed {"__partial_json": ...} runs on its defaults instead of what the model asked for

So the turn returns StopReason::Error, the agent loop returns before extracting tool calls, and the tool never runs.

openai_compat.rs falls back to an empty object:

let args = serde_json::from_str(&buf.arguments).unwrap_or_else(|e| {
    if !buf.arguments.is_empty() {
        warn!(tool = %buf.name, len = buf.arguments.len(),
              "tool-call arguments failed to parse ({e}); using empty object");
    }
    serde_json::Value::Object(Default::default())
});

The call proceeds with {} — which is exactly the hazard Anthropic's path exists to prevent, and it is the default on 15+ providers: OpenAI, Groq, Together, DeepSeek, Fireworks, Mistral, xAI.

Pinned by truncated_tool_arguments_fall_back_to_empty_unlike_anthropic in tests/openai_compat_stream_test.rs.

Why it matters

A truncated stream is not exotic — a dropped connection mid-arguments is the ordinary case. The failure is silent apart from a warn!, and a library cannot assume its host installed a subscriber.

Concretely: delete_files({"paths": ["/tmp/scratch"]}) truncated after {"paths": becomes delete_files({}). What that does depends entirely on the tool's own defaults, and the model's actual intent is gone. The agent then sees a plausible-looking result and continues.

The empty case is not affected and should stay as-is: a zero-argument tool legitimately streams "", and {} is the right answer there. Both are now pinned, so any fix must keep them distinct — which is the same empty-vs-malformed distinction that shipped the Anthropic bug.

Options

1. Match Anthropic — fail the turn on a non-empty, unparseable buffer. Consistent, and the loop already handles StopReason::Error. Behavioural change for 15+ providers.
2. Surface it as an error tool result — let the model see "your arguments did not parse" and retry. Fits the crate's "tools return stderr so the LLM can self-correct" convention, and is less abrupt than failing the turn.
3. Leave it, document it. Defensible if the truncation rate is genuinely negligible, but nothing measures that today.

Option 2 looks best on the crate's own conventions, but it is a real design call rather than a bug fix.

Acceptance

• Decide deliberately between the three.
• Whatever is chosen, the empty-argument case keeps working — that distinction is load-bearing and has already been got wrong once.
• truncated_tool_arguments_fall_back_to_empty_unlike_anthropic is updated rather than deleted, so the change is visible in the diff.
• Check whether google.rs, bedrock.rs, openai_responses.rs and azure_openai.rs have a third behaviour — only Anthropic and openai_compat were examined.