- Rename `create_new_topic` to `update_topic` to support continuous narrative heartbeats.
- Introduce `strategic_intent` parameter to capture tactical reasoning within tool-use turns.
- Consolidate `previous_summary` and `current_summary` into a single `summary` parameter to simplify model narrative bridging.
- Implement Smart Narrative Routing in `UpdateTopicTool` to dynamically toggle between full chapter headers and tactical intent-only heartbeats based on topic transitions.
- Update `TopicState` to track both active topic and active intent session-wide.
- Overhaul system prompt snippets to mandate the Heartbeat protocol, enforcing strictly granular, single-idea chapters and forbidding clubbed phases.
- Gate all logic and tool accessibility behind the `topicUpdateNarration` experimental flag.
- Update security policies (read-only and plan) to whitelist the heartbeat tool.
- Refactor scheduler logic to prioritize `update_topic` as the first execution in any batch.
- Synchronize all core and CLI configuration tests with the new tool schema and naming.
- Mandate granular, single-focus topics (e.g., 'Researching', 'Implementing [Idea]') to prevent broad, combined topics.
- Introduce required 'current_topic_summary' and 'previous_topic_summary' (on transition) parameters to the create_new_topic tool.
- Require 5-10 sentence detailed summaries for both parameters to improve context injection and strategic alignment.
- Mandate creating a new topic if work deviates from the current topic's stated goal.
- Relax silent-mode restrictions to allow Strategic Narration at the start of chapters for better transparency.
- Update tool declarations, system prompts, and unit tests to reflect these changes.
- Move topic tracking from global TopicManager to a session-scoped TopicState within Config.
- Implement sanitization to strip newlines and carriage returns from topic titles.
- Reject empty or whitespace-only topic titles for improved robustness.
- Update CreateNewTopicTool and PromptProvider to use the session-scoped state.
- Disable topicUpdateNarration by default in .gemini/settings.json.
- Enhance test coverage for state independence and sanitization.
- Added explicit policy rules to always allow 'create_new_topic' in all modes.
- Updated topic tool output to use 'Current topic' phrasing for clarity.
- Replaced the 'Thinking Protocol' with a mandate for internal reasoning to prevent literal thought block output in the UI.
- Added and updated unit tests to verify policy and tool behavior.
Organize agent work into logical phases (Chapters) using a dedicated tool
instead of prompt-only narration. This improves UI organization and
provides the model with explicit state tracking.
Key changes:
- Introduced 'create_new_topic' tool for semantic phase transitions.
- Added TopicManager singleton to maintain session chapter state.
- Gated the feature behind 'experimental.topicUpdateNarration' flag.
- Limited tool registration to Gemini-3 model set.
- Modified Scheduler to prioritize topic markers at start of batches.
- Injected active topic context into system prompt footers.
- Replaced SI narration in snippets with tool-based mandates.
Christian Gunderman