[changed]After a tool returns afunctionResponse, the agent must either call another tool or give a user-facing text response — empty replies with no text and no tool calls are no longer allowed.[changed]After edit or write-file tool runs, the agent must summarize what changed and outline verification or next steps (for example running tests), and must not finish the turn with zero text tokens.
Models & prompts
- Built-in tool-usage guidance now includes explicit Tool Execution Response Rules and Post-Edit Response Rules so the model consistently continues the turn after tools and file edits.
Upgrade notes
No special upgrade steps; existing installs pick up the clearer post-tool and post-edit response behaviour automatically.