Observability And Logging

Last updated: April 11, 2026

Assistant Observability and Logging

The assistant module includes a robust observability stack for tracking performance, costs, tool health, and system audits.

Data Persistence

The observability system (in assistant/observability.go) uses a SQLite database with WAL (Write-Ahead Logging) for high-performance telemetry storage.

Key Metrics Tracked

  • Request Logs: Every interaction includes model name, tokens (prompt/completion), cost, duration, and status.
  • Tool Execution: Tracks average duration, success rates, and common error patterns for every tool.
  • Provider Analytics: Compares latency and cost across different LLM providers (Anthropic, OpenAI, Ollama, etc.).

Logging Strategy

The system employs a dual logging strategy:

  1. JSONL Session Logs: Every step of the ReAct loop is logged in real-time to a .jsonl file in the chats/logs/ directory. This allows for live debugging and auditing via Mission Control.
  2. HTTP Request Logs: Standard HTTP telemetry is captured in logs/http.log via the LoggingMiddleware.

Mission Control Integration

The MissionControl dashboard (in assistant/missioncontrol.go) provides a real-time UI for these metrics:

  • Real-time Reasoning: SSE-powered "Chain of Thought" logs.
  • Tool Health Grid: Visualizes failing tools and sample error messages.
  • Cost Monitoring: Real-time spending analysis per user and per provider.

Component Diagram: Observability Flow

graph TD
    P[Planner] --> L[JSONL Session Logs]
    P --> O[Observability DB]
    O --> S[SQLite / WAL]
    S --> MC[Mission Control Dashboard]
    L --> MC
    
    subgraph "Audit Layer"
        P --> A[Architect Audit]
        A --> SN[Sovereign/Infra Alerts]
    end

Key Files & Functions

  • assistant/observability.go: Core ObservabilityDB implementation and metric recording.
  • assistant/session_logs.go: Logic for immediate, non-buffered JSONL log commitment.
  • assistant/missioncontrol.go: SSE and WebSocket broadcasting of telemetry data.

Guidance for AI Agents

  • Detailed Tool Responses: When a tool fails, provide descriptive error messages; these are captured in the health grid and help developers fix issues.
  • Audit Awareness: Be aware that your reasoning (thoughts) and actions are logged and audited by the Architect agent for logic and security compliance.

Cross-References