~ LOSSY

pydantic-ai 2.61.0

truncation.tool-call-args-stream · anthropic

Requirements

Id	Level	Status	Evidence
no-complete-partial-call	MUST	SATISFIED	outcome=raised; error signalled
cause-is-truncation	MUST	NOT_OBSERVABLE	finish.value="length" (raw: ModelResponse.finish_reason); truth.finish="length"; error.category="validation"; structured.returned_as_success=false

What the provider sent

Finish: length
Raw finish: max_tokens
Text: null
Tool calls: [ { "arguments": null, "arguments_parsed_ok": false, "arguments_raw": "{\"title\": \"Pottery in Málaga", "id": "toolu_01A09q90qw90lq917835lq9", "name": "create_event" } ]
Usage: { "alternatives": {}, "cached_input": null, "input": 25, "output": 20, "raw": { "input_tokens": 25, "output_tokens": 20, "output_tokens_details": { "thinking_tokens": 6 } }, "reasoning": 6, "total": 45 }
Stream completed: true
Error: null

Wire excerpt

status: 200
content-type: text/event-stream
body (sse):
event: message_start
data: {"type": "message_start", "message": {"id": "msg_01Aq9w938a90dw8q", "type": "message", "role": "assistant", "content": [], "model": "claude-sonnet-5-5", "stop_reason": null, "stop_sequence": null, "usage": {"input_tokens": 25, "output_tokens": 1}}}


event: content_block_start
data: {"type": "content_block_start", "index": 0, "content_block": {"type": "thinking", "thinking": "", "signature": ""}}


event: ping
data: {"type": "ping"}


event: content_block_delta
data: {"type": "content_block_delta", "index": 0, "delta": {"type": "thinking_delta", "thinking": ""}}


event: content_block_delta
data: {"type": "content_block_delta", "index": 0, "delta": {"type": "signature_delta", "signature": "EqQBCgIYAhIM1gbcDa9GJwZA2b3hGgxBdjrkzLoky3dl1pkiMOYds..."}}


event: content_block_stop
data: {"type": "content_block_stop", "index": 0}


event: content_block_start
data: {"type": "content_block_start", "index": 1, "content_block": {"type": "tool_use", "id": "toolu_01A09q90qw90lq917835lq9", "name": "create_event", "input": {}}}


event: content_block_delta
data: {"type": "content_block_delta", "index": 1, "delta": {"type": "input_json_delta", "partial_json": "{\"title\": \"Pot"}}


event: content_block_delta
data: {"type": "content_block_delta", "index": 1, "delta": {"type": "input_json_delta", "partial_json": "tery in Málaga"}}


event: content_block_stop
data: {"type": "content_block_stop", "index": 1}


event: message_delta
data: {"type":"message_delta","delta":{"stop_reason":"max_tokens","stop_sequence":null},"usage":{"output_tokens":20,"output_tokens_details":{"thinking_tokens":6}}}


event: message_stop
data: {"type": "message_stop"}

What the library reported

Outcome

raised

Finish

{
  "raw": "max_tokens",
  "source": "ModelResponse.finish_reason",
  "surface": "raw",
  "value": "length"
}

Text

Value: null
Surface: absent

Tool calls

[
  {
    "arguments": {
      "INVALID_JSON": "{\"title\": \"Pottery in Málaga"
    },
    "arguments_parsed_ok": true,
    "arguments_raw": "{\"INVALID_JSON\":\"{\\\"title\\\": \\\"Pottery in Málaga\"}",
    "id": "toolu_01A09q90qw90lq917835lq9",
    "name": "create_event"
  }
]

Usage

{
  "cached_input": 0,
  "input": 25,
  "output": 20,
  "reasoning": 6,
  "surface": {
    "cached_input": "raw",
    "input": "raw",
    "output": "raw",
    "reasoning": "raw",
    "total": "raw"
  },
  "total": 45
}

Stream

{
  "chunks_seen": 10,
  "completed": false,
  "used": true
}

Error

Type: pydantic_ai.exceptions.UnexpectedModelBehavior
Message: Tool 'create_event' exceeded max retries count of 0. Consider raising the retry limit, or see the docs on tool retries: https://pydantic.dev/docs/ai/tools-toolsets/tools-advanced/#tool-retries
Category: validation
Retry after: null
Stack top: pydantic_ai/tool_manager.py:523

Unsupported reason

null

Full observation JSON

{
  "elapsed_ms": 190.30865095555782,
  "error": {
    "category": "validation",
    "message": "Tool 'create_event' exceeded max retries count of 0. Consider raising the retry limit, or see the docs on tool retries: https://pydantic.dev/docs/ai/tools-toolsets/tools-advanced/#tool-retries",
    "retry_after": null,
    "stack_top": "pydantic_ai/tool_manager.py:523",
    "type": "pydantic_ai.exceptions.UnexpectedModelBehavior"
  },
  "finish": {
    "raw": "max_tokens",
    "source": "ModelResponse.finish_reason",
    "surface": "raw",
    "value": "length"
  },
  "format": "wiretruth.observation/v1",
  "library": {
    "api": "Agent.run_stream_events",
    "language": "python",
    "name": "pydantic-ai",
    "version": "2.61.0"
  },
  "notes": [],
  "outcome": "raised",
  "stream": {
    "chunks_seen": 10,
    "completed": false,
    "used": true
  },
  "structured": {
    "attempted": false,
    "returned_as_success": false,
    "value": null
  },
  "text": {
    "surface": "absent",
    "value": null
  },
  "tool_calls": [
    {
      "arguments": {
        "INVALID_JSON": "{\"title\": \"Pottery in Málaga"
      },
      "arguments_parsed_ok": true,
      "arguments_raw": "{\"INVALID_JSON\":\"{\\\"title\\\": \\\"Pottery in Málaga\"}",
      "id": "toolu_01A09q90qw90lq917835lq9",
      "name": "create_event"
    }
  ],
  "unsupported_reason": null,
  "usage": {
    "cached_input": 0,
    "input": 25,
    "output": 20,
    "reasoning": 6,
    "surface": {
      "cached_input": "raw",
      "input": "raw",
      "output": "raw",
      "reasoning": "raw",
      "total": "raw"
    },
    "total": 45
  }
}

Fixture

documented anthropic/truncation.tool-call-args-stream

Provenance	Value
Source	https://platform.claude.com/docs/en/agents-and-tools/tool-use/fine-grained-tool-streaming#handling-invalid-json-in-tool-responses
Notes	Documented shapes, not a recording. Message envelope and stop fields follow handling-stop-reasons; content blocks follow thinking/streaming and handle-tool-calls. Model changed to claude-sonnet-5-5; IDs and scenario text are filled in. Token counts are illustrative, filled in to include thinking within the requested output limit, not measured. eager_input_streaming and partial JSON at max_tokens are documented; the input fragments are original scenario text. ASSUMED composition with an omitted thinking block and a truncated tool block; the complete SSE transcript for this limit is not shown. The initial input: {} is only the documented placeholder, not the tool arguments. SSE names, JSON shapes, field order and spaced text/thinking events follow streaming examples; the compact message_delta follows Streaming request with tool use. Fragment boundaries and counts are filled in. Signature placeholder is copied literally from the signature_delta example; empty thinking_delta follows the documented omitted display flow. The output_tokens_details.thinking_tokens breakdown follows thinking-steering.md (https://platform.claude.com/docs/en/build-with-claude/thinking-steering-and-cost#pricing); the thinking count of 6 is filled in, not measured. It appears only in the final message_delta usage.

Provenance

Value

Source

https://platform.claude.com/docs/en/agents-and-tools/tool-use/fine-grained-tool-streaming#handling-invalid-json-in-tool-responses

Notes

Documented shapes, not a recording. Message envelope and stop fields follow handling-stop-reasons; content blocks follow thinking/streaming and handle-tool-calls. Model changed to claude-sonnet-5-5; IDs and scenario text are filled in. Token counts are illustrative, filled in to include thinking within the requested output limit, not measured. eager_input_streaming and partial JSON at max_tokens are documented; the input fragments are original scenario text. ASSUMED composition with an omitted thinking block and a truncated tool block; the complete SSE transcript for this limit is not shown. The initial input: {} is only the documented placeholder, not the tool arguments. SSE names, JSON shapes, field order and spaced text/thinking events follow streaming examples; the compact message_delta follows Streaming request with tool use. Fragment boundaries and counts are filled in. Signature placeholder is copied literally from the signature_delta example; empty thinking_delta follows the documented omitted display flow. The output_tokens_details.thinking_tokens breakdown follows thinking-steering.md (https://platform.claude.com/docs/en/build-with-claude/thinking-steering-and-cost#pricing); the thinking count of 6 is filled in, not measured. It appears only in the final message_delta usage.

Reproduce

wiretruth serve anthropic/truncation.tool-call-args-stream

Plain-library snippet from the shim manifest.

AnthropicModel(model, provider=AnthropicProvider(base_url=WIRETRUTH_BASE_URL_ANTHROPIC, api_key=WIRETRUTH_API_KEY)); host root

Upstream

reported Issue

Fixed in: not specified

Same report as the entry above, for the streamed Anthropic cell (the non-streamed Anthropic cell passes).

History

Version	Grade
2.61.0	~ LOSSY
2.56.0	~ LOSSY