feat(core): detect tool calls leaked into reasoning and text

[Grigory-Rylov]

Problem

LLM sometimes emits tool calls as XML-style text in reasoning output instead of using the structured tool_calls API field:

<tool_call>
<function=grep>
<parameter=path>...</parameter>
<parameter=pattern>...</parameter>
</function>
</tool_call>

This causes the agent to hang because the tool call is never executed.

Solution

• Added findToolCallName() to detect XML-style tool calls (<function=name>, [Assistant tool call]: name(, and JSON {"name": "..."} patterns)
• Applied detection to both text and reasoning fragments in publish-llm-event.ts
• When detected: replaces the content with an [ERROR] message and sets textBasedToolCall = true, triggering a retry turn

Changes

• packages/core/src/session/runner/publish-llm-event.ts — findToolCallName() + reasoning fragment check
• packages/core/test/publish-llm-event.test.ts — 12 tests for findToolCallName

[github-actions]

This PR doesn't fully meet our contributing guidelines and PR template.

What needs to be fixed:

• PR description is missing required template sections. Please use the PR template.

Please edit this PR description to address the above within 2 hours, or it will be automatically closed.

If you believe this was flagged incorrectly, please let a maintainer know.