April 21, 2026

Refining Tooling and Enhancing Personalization

Tool descriptions were sharpened to require imperative triggers like "deep dive" or "research" to ensure models dispatch tools before timing out on inline answe

Tool descriptions were sharpened to require imperative triggers like “deep dive” or “research” to ensure models dispatch tools before timing out on inline answers.

Document skill tool descriptions were rewritten to be unambiguous, explicitly stating when to use the tool to create, save, write, or delete a document.

Backfilling of data graph embeddings and FTS was implemented, calling a routine to generate and store embeddings for raw-SQL-inserted rows lacking them.

Code evaluation output retrieval was fixed by passing locals to exec(), allowing the captured print output to be correctly read and used by the LLM.

The goal pursuit mechanism now persists the goal to the goals table before spawning the background thread, ensuring ledger accuracy.

The news tool was updated to correctly accept the topic keyword argument as used by the internal tool registry.

A comprehensive personality feature was introduced, driven by a 5-slider voice corpus, which prepends a voice line to the system prompt.

The web interface was updated to document the new Personality subtab, detailing the slider UI, live preview, and save functionality.

  • Sharpened tool description to fire before inline answer using imperative triggers.

  • Rewrote document tool description to explicitly signal when to use create/save/etc.

  • Implemented lazy backfill for missing data graph embeddings and FTS.

  • Fixed code evaluation by capturing PrintCollector output via exec locals.

  • Launched personality feature using 5-slider voice corpus driving system prompt.

  • Updated news tool to accept the topic keyword for registry convention.