fix(anthropic): strip FORMAT_TOOL from re-encoded history to fix resume() HTTP 400

[devteamaegis]

Problem

response.resume() fails with HTTP 400 when using Anthropic with structured outputs (format=SomeModel):

BadRequestError: 400 - tool_use ids were found without tool_result blocks immediately after

Fixes #2503

Root cause

Anthropic structured-output mode injects a private FORMAT_TOOL_NAME tool. The model response contains a tool_use block for that tool. When the assistant message appears in the history of a resume() call:

1. _encode_message detects the format tool (raw_message_has_format_tool → True) and falls through to re-encode from content parts instead of using the raw message verbatim.
2. encode_content encodes every tool_call part as a ToolUseBlockParam — including the FORMAT_TOOL_NAME call.
3. Anthropic rejects the request because there is no tool_result in the following user message (there never is — the format tool is a Mirascope detail that is never actually invoked).

Fix

In encode_content, skip any tool_call part whose name starts with FORMAT_TOOL_NAME:

elif part.type == "tool_call":
    if part.name.startswith(FORMAT_TOOL_NAME):
        continue   # ← new guard
    blocks.append(ToolUseBlockParam(...))

This preserves all real tool_use blocks while stripping the implementation-detail format tool from re-encoded history.

Tests

Added to python/tests/llm/providers/anthropic/test_anthropic_encoding.py:

test	what it checks
test_encode_content_skips_format_tool_call	FORMAT_TOOL tool_use absent from re-encoded blocks
test_encode_content_keeps_real_tool_calls	genuine tool_use blocks still emitted

pytest tests/llm/providers/anthropic/test_anthropic_encoding.py -v
3 passed

[willbakst]

Doesn't this mean the content of this final tool call isn't included at all? Shouldn't we be re-encoding it as text in this case?

[devteamaegis]

Good catch @willbakst — you're right, continue dropped the content entirely. The format tool's arguments are the assistant's actual structured response, so instead of skipping I now re-encode them as a text block: the JSON the model produced is preserved in the history, and there's still no orphaned tool_use block to trigger the 400 on resume. Updated the test to assert both text parts survive (["Here is the result.", '{"value": 7}']). Tests pass.