This commit fixes a bug where model text preceding a tool call would
be displayed after the tool call in the conversation history.
We now explicitly flush any pending agent text to the conversation
history (via 'addItem') immediately before a tool request is tracked.
This ensures the model's intent is rendered above the execution result.
This change ensures that tool execution in the Agent Protocol experiment
reaches visual parity with the legacy implementation by passing rich
metadata and objects (like FileDiff) through the AgentEvent stream.
Key changes:
- Core 'LegacyAgentSession' now builds temporary invocations for tool requests
to capture rich, argument-aware descriptions (e.g. 'Writing to poem.md').
- Core 'LegacyAgentSession' now attaches the raw 'resultDisplay' object and
'outputFile' path to the 'tool_response' event metadata.
- Core 'LegacyAgentSession' now passes the rich description through 'tool_update'
events to ensure dynamic descriptions are updated during execution.
- UI 'useAgentStream' hook now extracts these rich values from event metadata
to populate the local 'trackedTools' state, allowing 'mapToDisplay' to
correctly trigger bespoke rendering components (like the diff viewer).
This commit achieves visual parity for tool execution in the interactive stream when using the experimental 'useAgentProtocol' flag. It removes direct UI dependency on the tool scheduler's internal state.
Key changes:
- Core 'LegacyAgentSession' now attaches display metadata (displayName, description, etc.) to 'tool_request' AgentEvents.
- Core 'LegacyAgentSession' listens to the MessageBus to emit 'tool_update' AgentEvents for live output (e.g., shell commands).
- UI 'useAgentStream' now maintains its own 'trackedTools' local state, constructed entirely from incoming 'tool_request', 'tool_update', and 'tool_response' events.
- The local 'trackedTools' state is mapped to 'pendingToolGroupItems' using the existing 'mapToDisplay' function for seamless visual parity.
This change introduces an experimental 'useAgentProtocol' flag to interactive mode.
When enabled, the UI uses the new 'useAgentStream' hook which leverages
the core 'LegacyAgentSession' (AgentProtocol) instead of the custom
'useGeminiStream' logic.
Key changes:
- Added 'useAgentProtocol' experimental setting to CLI and Core config.
- Implemented 'useAgentStream' hook with basic interaction and thought support.
- Modified 'useToolScheduler' to expose its internal Scheduler instance to ensure implementation parity.
- Updated 'AppContainer' to conditionally branch between implementations via ternary operator.
- Added comprehensive unit tests for the new hook.
Michael Bleigh