AgentPulse v0.1.0

First release of AgentPulse — open-source drift investigation for multi-agent systems: which agent drifted, where it originated, and what to check next.

Highlights

Capture

• One-line instrument() for AutoGen / LangChain runs; Anthropic and OpenAI SDKs patched transparently (sync, async, and streaming)
• Per-run metrics: per-agent tokens, latency, cost, tool calls, retries, handoffs, termination
• Full replay capture: verbatim LLM request/response payloads, tool arguments and results, and handoff payload text, stored in SQLite with idempotent in-place migrations

Investigate

• Web dashboard (light theme): project overview, run history, drift investigation, metrics explorer, and event timeline
• Version snapshots with baseline comparison — drift is measured against the version you pick
• Config-change tracking with before/after prompt diffs

Replay

• python storage/otel_export.py <run_id|--last> --submit exports any captured run as an OpenTelemetry gen_ai.* trace and submits it to a replay server's /api/ingest
• Exports round-trip the replay harness's request matching: messages verbatim, system prompt as a request param, content blocks preserved

MCP

• agentpulse_mcp.py exposes findings, next check steps, and version comparisons to Claude