bug: unguarded json.loads on streamed tool-call arguments crashes the request on empty/truncated JSON

bug: unguarded json.loads on streamed tool-call arguments crashes the request on empty/truncated JSON

Bug description

Providers occasionally stream empty or truncated tool-call argument JSON
(e.g. output cut off by token limits or an upstream hiccup). The unguarded
json.loads() on streamed arguments then raises JSONDecodeError, which
surfaces to the user as a raw error and aborts the in-flight step:

Error: Expecting value: line 1 column 1 (char 0)

This is provider-agnostic — any of the streaming providers can emit an empty
or truncated arguments payload.

Expected behavior

A malformed/empty streamed arguments payload should degrade gracefully, not
kill the request.

Suggested approach

Add a shared parse helper that:

1. Parses strictly,
2. Falls back to the existing lenient extract_json() helper,
3. On total failure, raises a descriptive error naming the tool and including a payload snippet (instead of a bare JSONDecodeError).

Apply it to all streaming providers (Anthropic, OpenAI chat completions,
OpenAI Responses) and to Anthropic history replay, where a stale corrupt
payload in history should degrade to {} rather than breaking the current
request.

Environment

• Latest main

Locate the shared extract_json() helper and the streaming argument parsing paths for Anthropic, OpenAI chat completions, OpenAI Responses, plus Anthropic history replay. Trace how empty or truncated payloads currently reach json.loads(), then verify that malformed streaming input produces a descriptive tool-specific error and corrupt history degrades to {} without aborting the request.