March 16, 2026

AI Core Enhancements: Mode-Gate, Context, and Stability

The mode-gate ONNX classifier has been deployed, a binary RESPOND/ACT classifier achieving 97

The mode-gate ONNX classifier has been deployed, a binary RESPOND/ACT classifier achieving 97.9% accuracy, saving LLM tokens on ~48% of messages by triggering a RESPOND fast-path when confidence is $\ge 0.85$.

Chat responses now emit a present_response intent alongside sse pub/sub, and digest worker emissions of show_narration intent are routed for ACT narrations.

Chat history persistence is improved: after a server restart, the system now correctly retrieves the still-active thread instead of only checking for expired threads in SQLite.

A Stage 8 cognitive contract introduced wrapper API enhancements, Signal ingestion, and improved trait extraction safety by adding quoted-content rules and skipping extraction for long messages.

The SituationModelService and ConversationPhaseService were added to aggregate ambient context and track conversation lifecycle phases, injecting this data into prompts via the $`` placeholder.

To handle GPT-5.4 returning multi-object responses, the ACT loop JSON parser was fixed using $json.JSONDecoder().raw_decode()$ as an early recovery layer to prevent silent loss of tool invocations.

The tool registry HTML detection logic was tightened from $"<" in text$ to $re.search(r'<[a-zA-Z/]', text)$ to correctly process plain-text tool output containing comparison operators.

Build log regeneration now scans commit counts per date across the last 30 days to ensure stale entries are regenerated if commit counts have changed.

  • Mode-gate ONNX classifier deployed with 97.9% accuracy to accelerate RESPOND fast-path.

  • SituationModelService tracks ambient context and conversation phases, injecting data into prompts.

  • ACT loop JSON parser fixed to handle multi-object responses from GPT-5.4.

  • Trait extraction is safer, adding quoted-content rules and skipping extraction for messages over 300 characters.

  • Chat history now correctly loads active threads following a server restart.

  • HTML detection logic tightened to avoid stripping plain-text output containing comparison operators.