fix: update_instructions() now reflected in tool call response generation

[weiguangli-io]

Summary

• Fixes update_instructions() not reflected in tool call response generation #4242: when update_instructions() is called inside a function tool, the updated instructions are now used for the subsequent tool response generation
• Before this fix, the chat context was snapshotted before tool execution, so any instruction updates made during tool execution were lost for the current turn's response
• The fix refreshes the instructions in chat_ctx from agent._instructions right before generating the tool response

Context

Maintainer @longcw confirmed this is unexpected behavior:

"it makes sense to add a update_instructions for this turn in a function tool"
"no, it's not expected"

Changes

In agent_activity.py, added a call to update_instructions() to refresh the chat context's instructions from the agent's current _instructions before the recursive _pipeline_reply_task call that generates the tool response. This ensures that any update_instructions() calls made inside function tools are reflected in the LLM's response generation for the current turn.

Test plan

• Verify that calling agent.update_instructions() inside a @function_tool updates the instructions used for the current turn's response
• Verify that normal (non-tool) instruction updates continue to work as before
• Verify that instructions persist correctly across multiple tool call steps

🤖 Generated with Claude Code

[devin-ai-integration]

Devin Review found 1 potential issue.

View 3 additional findings in Devin Review.

[devin-ai-integration]

🟡 Missing try/except ValueError around update_instructions unlike all other call sites

The new update_instructions() call is not wrapped in a try/except ValueError, unlike every other call site in the same file (agent_activity.py:600-606 and agent_activity.py:1960-1964). The update_instructions function in generation.py:820-823 explicitly raises ValueError if the existing instructions item is not of type "message". If this exception fires, it will propagate uncaught and crash the _pipeline_reply_task_impl coroutine, potentially leaving the agent in a broken state (e.g., stuck in "thinking" without transitioning back to "listening").

Suggested change

	update_instructions(
	chat_ctx,
	instructions=self._agent._instructions,
	add_if_missing=False,
	)
	try:
	update_instructions(
	chat_ctx,
	instructions=self._agent._instructions,
	add_if_missing=False,
	)
	except ValueError:
	logger.exception("failed to update the instructions")

---

Was this helpful? React with 👍 or 👎 to provide feedback.