AI Assistant Updates: News, Tasks, and Prompts

The news tool has been simplified from six parameters/four actions to a two-parameter interface: query and an optional category. If no category is provided, Google News is retrieved using geo derived from browser telemetry instead of hardcoded settings.

The persistent task system was redesigned to use a clean three-phase execution: Plan, Execute (using ACT loop with a sub_agent skill), and Surface. Tasks now stall on failure instead of retrying infinitely.

A new sub_agent innate skill was introduced for spawning focused workers, and tool deny lists were implemented to prevent recursive spawning.

The unnecessary post-loop synthesis LLM call, which fired when the loop exited without a final response, was removed to reduce token cost and prevent misleading outputs.

System prompts now include guidance on when to invoke skills versus responding directly, addressing issues like unnecessary tool firing or failing to search memory.

Explicit feedback signals are now wired into the score calculation to boost the UNIFIED score on acknowledgments, which reduces unwarranted tool dispatch.

Principles such as auto-store and active-recall were added to the frontal cortex documentation, and a proactive information gathering section was added to ACT documentation.

• Replaced news tool with 2-param interface (query + optional category).

• Redesigned persistent tasks for three-phase execution: Plan -> Execute -> Surface.

• Introduced sub_agent innate skill for spawning focused workers in tasks.

• Removed wasteful post-loop synthesis LLM call from task execution.

• Added skill invocation guidance principles to system prompts.

• Wired explicit_feedback signal to reduce unnecessary tool dispatch.