fix(agent): steer on non-string tool content only as anthropic blocks, coerce for chat-completions (#78598)

[webtecnica]

Summary

Fixes /steer on tool results with non-string content (e.g. multimodal block lists). apply_pending_steer_to_tool_results() in agent/agent_runtime_helpers.py unconditionally converted non-string tool content into an Anthropic Messages API block list. For chat-completions providers (OpenRouter, Xiaomi, DeepSeek, DeepInfra — routed via /chat/completions), ChatCompletionToolMessage.content must be a plain string; the block list is rejected with HTTP 400 and the poisoned message stays in the history, causing a 400 loop on every subsequent call including fallbacks.

Root Cause

if not isinstance(existing_content, str):
    blocks = list(existing_content) if existing_content else []
    blocks.append({"type": "text", "text": marker.lstrip()})
    messages[target_idx]["content"] = blocks

The block-list shape is only valid for the Anthropic Messages API wire format. The function never consulted the active wire format (agent.api_mode) before choosing the content shape.

Change

• agent/agent_runtime_helpers.py:
  • apply_pending_steer_to_tool_results() now only builds the block-list shape when agent.api_mode == "anthropic_messages" (previous behavior preserved for Anthropic).
  • For any other wire format, non-string tool content is coerced to a plain string BEFORE appending the steer marker, via a new module-level helper _coerce_tool_content_to_string() (JSON serialization with str() fallback, matching the pattern already used elsewhere in the codebase).
• tests/run_agent/test_steer.py:
  • test_multimodal_content_list_preserved now pins api_mode="anthropic_messages" (block-list behavior is Anthropic-only).
  • New test_chat_completions_coerces_non_string_content_to_plain_string — chat-completions mode + block-list content produces a plain string with the marker appended after the serialized content.
  • New test_string_content_shape_unchanged_in_anthropic_mode — string content stays a string with the marker in both wire formats.

Verification

• python3 -m pytest tests/run_agent/test_steer.py -q — 29 passed.
• python3 -m pytest tests/ -k "steer or tool_result" -q — 154 passed, 3 skipped, 0 failures.

Closes

Closes #78598