Claude code execution: 400 on multi-turn (code_execution_tool_result.citations: Extra inputs are not permitted)

[vishalr15]

Description

A conversation that uses Claude's code execution tool fails on the second (and later) turn with an Anthropic 400:

400 - messages.N.content.M.code_execution_tool_result.citations:
      Extra inputs are not permitted

The first turn succeeds; the error only appears once the prior assistant response is replayed as input.

Root cause

Anthropic emits a citations field on code_execution_tool_result blocks but rejects it on input (the output and input schemas differ). agno preserves server-tool result blocks for history reconstruction:

# agno/models/anthropic/claude.py
server_tool_blocks.append(block.model_dump())   # model_dump() keeps `citations`
...
model_response.provider_data.setdefault("server_tool_blocks", []).extend(server_tool_blocks)

On the next turn format_messages replays those stored blocks verbatim:

# agno/utils/models/claude.py
for block_dict in message.provider_data["server_tool_blocks"]:
    ...
    content.append(block_dict)   # `citations` echoed back as input -> 400

Because the Messages API is stateless, the whole prior response is re-sent every turn, so the illegal field is rejected before the request reaches the model.

Steps to reproduce

1. Create a Claude agent with skills / code execution enabled.
2. Send a message that triggers code execution (turn 1) — succeeds.
3. Send any follow-up in the same conversation (turn 2) — 400 with the error above.

Expected

Server-tool result blocks should be replayable across turns; output-only fields Anthropic rejects on input should not be echoed back.

Environment

• agno: 2.6.x (reproduced against main)
• Provider: Anthropic (Claude), code execution / skills

Proposed fix

Strip the output-only citations field from code_execution_tool_result / bash_code_execution_tool_result blocks at capture time (so it's never persisted), with a matching strip at the replay point for already-persisted history. The citation data is preserved separately in ModelResponse.citations. PR to follow.