April 16, 2026

Major Refactor: Thinking Gate, Data Graph & Classifier

The thinking-gate was redesigned to utilize provider-native deliberation flags instead of a classifier-driven approach, moving exploration to a single, same-job

The thinking-gate was redesigned to utilize provider-native deliberation flags instead of a classifier-driven approach, moving exploration to a single, same-job pass on high-mode only. This single pass now persists the exploration as a tool call row named “thinking” for later re-injection into the ACT loop via a wrapper function.

The architecture now decouples the thinking time budget (THINKING_TIMEOUT=600s) from the main ACT loop’s MAX_TIMEOUT envelope, allowing the exploration LLM call to run independently.

The data graph was overhauled by replacing the contradiction classifier with a deterministic Concept LUT system, which enforces temporal and coexistence rules for 27 high-level concepts.

This LUT system includes a new forget action with rule-aware hard-delete semantics and updates the data graph service to return structured results.

Episode extraction was completely redesigned to handle transcript linkage, triggering per-channel extraction every 35 turns and handling overlaps by creating super episodes.

The testing philosophy shifted dramatically to feature tests only, eliminating 1548 lines of mock theater across six classifier test files. New tests focus on real-world inputs and observable outputs.

Fixes were implemented to ensure robust system state, including exposing degraded status if ONNX classifier heads fail to boot, and ensuring existing trait metadata is correctly forwarded to the ONNX input.

Debugging support was improved by logging Ollama HTTP 4xx response bodies before re-raising errors, aiding in diagnosing upstream rejection payloads.

  • Replaced contradiction classifier with deterministic Concept LUT system with 27 concepts.

  • Thinking mode now uses provider-native flags (e.g., ollama think=True) for deliberation.

  • Exploration LLM call runs once per turn, with its time budget set at 600s.

  • Removed anti-demotion guard, making the classifier the sole decision maker.

  • Episode extraction pipeline redesigned with per-channel triggers and super episode creation.

  • Switched classifier tests to zero-mock feature tests, focusing on real-world input/output validation.