Skip to main content

Changelog

All notable changes to TraceAgent are documented here.

tip

For the most up-to-date release information, see the GitHub Releases page.

Unreleased​

Added​

  • Pydantic AI integration (trace-agent-pydantic-ai): a capability that records model calls and tools of Pydantic AI 2.x agents without blocking them, available as run.as_pydantic_ai_capability()
  • OpenTelemetry ingestion: OTLP/HTTP receiver at POST /v1/traces (protobuf or JSON, gzip). One run per trace; attach to an SDK run with the trace_agent.run_id resource attribute. Settings TRACE_AGENT_OTLP_ENABLED and TRACE_AGENT_OTLP_MAX_BODY_BYTES
  • Examples for both routes (pydantic_ai_agent.py, pydantic_ai_otel_agent.py)
  • Plain-language names in the timeline and graphs, and a readable inspector for events and decisions. A decision shows the events it is based on
  • parent_step_id in POST /api/runs/{id}/tool-results, to hang a tool from its parent step
  • Documentation site with full SDK reference, LangChain guide, and examples
  • Local search integration
  • Command palette for quick navigation (Ctrl/Cmd + K)
  • Showcase page for community use cases

Changed​

  • Runs show provider, model and token totals for framework integrations (previously only for runs created through the server's own model proxy). The Compare view counts those tokens too
  • The timeline nests model calls and tools under the agent step that ran them
  • The inspector hides identifiers, coordinates and raw JSON under "Technical details"
  • The server now accepts openai>=1.52,<4 (it was <2), so it installs next to pydantic-ai-slim[openai], and depends on opentelemetry-proto
  • Updated branding with new detective cat logo

Fixed​

  • LangChain: a tool call was recorded twice and counted twice as an error. It now appears once, with its real name, provider and model
  • Live updates (SSE) did not reach the UI because of an event type mismatch
  • Events of nested steps were hidden in the timeline
  • Inspecting an event or graph node of one run no longer shows it when you switch to another run

v0.1.0 — Initial Release​

Added​

  • TraceAgent SDK — Python client for recording agent actions
    • Start/finish runs
    • Record file writes, reads, commands, and tool calls
    • Attach artifacts and metadata
  • TraceAgent LangChain — Callback handler for automatic LangChain tracing
    • Chain start/end events
    • LLM call/response capture
    • Tool usage tracking
    • Retriever activity logging
  • TraceAgent Server — FastAPI backend
    • SQLite storage (default)
    • PostgreSQL support
    • Audit metrics collection
    • REST API
  • TraceAgent UI — Interactive dashboard
    • Execution timeline visualization
    • Run listing and search
    • File diff viewer
    • Artifact browser
  • Docker images for Server and UI
  • Example scripts:
    • sample_agent.py — Minimal SDK smoke test
    • langchain_agent.py — LangChain callback integration
    • Local model agent — Any OpenAI-compatible model with the SDK
    • Planner agent — Multi-tool planning
    • coding_agent_debugger_demo.py — Synthetic debug scenarios
    • rich_demo_agent.py — Full-fidelity trace demo

💬 Comments