gtracer 0.1.3

Lightweight span-based tracing for LangChain/LangGraph agents

 ██████╗ ████████╗██████╗  █████╗  ██████╗███████╗██████╗
██╔════╝ ╚══██╔══╝██╔══██╗██╔══██╗██╔════╝██╔════╝██╔══██╗
██║  ███╗   ██║   ██████╔╝███████║██║     █████╗  ██████╔╝
██║   ██║   ██║   ██╔══██╗██╔══██║██║     ██╔══╝  ██╔══██╗
╚██████╔╝   ██║   ██║  ██║██║  ██║╚██████╗███████╗██║  ██║
 ╚═════╝    ╚═╝   ╚═╝  ╚═╝╚═╝  ╚═╝ ╚═════╝╚══════╝╚═╝  ╚═╝

Lightweight span-based tracing for LangChain and LangGraph agents.

Emits structured JSONL spans via Python's standard logging — no new infrastructure, no agents, no dashboards required.

---

What it does

Every time an LLM call happens inside your agent, gtracer captures it as a structured span and writes it to stdout as JSON:

run
└── agent "main"
    ├── llm_call seq:1  ← tokens, model, message delta, latency
    │   └── tool_call search_database  ← input, result, duration
    ├── llm_call seq:2
    │   └── tool_call calculator
    └── llm_call seq:3  ← final answer

Works with CloudWatch, Datadog, or any stdout log consumer. Zero configuration — spans are live the moment you import the package.

---

Install

pip install gtracer
# or
uv add gtracer

---

Quick Start

1. Import and go

import gtracer  # spans are live immediately — nothing else needed

gtracer auto-configures at import time. It attaches its own JSON handler with propagate=False — it never touches your app's root logger, no double-emission, no interference.

2. Wrap your agent

from gtracer import tracer, tracing_handler

async def run(session_id: str, user_input: str):
    tracer.start_trace(session_id)

    with tracer.span("run", tags={"session_id": session_id}):
        with tracer.span("agent", attrs={"agent": "main"}) as agent_span:
            result = await my_agent.ainvoke(
                {"messages": [{"role": "user", "content": user_input}]},
                config={"callbacks": [tracing_handler]},
            )
            agent_span.set("output_type", type(result).__name__)
            return result

Every LLM call is now automatically captured — tokens, model, latency, message deltas.

3. Instrument your tools

from langchain_core.tools import tool
from gtracer import tracer, tracing_handler, _span_id

@tool
async def search_database(query: str) -> str:
    """Run a database query."""
    llm_parent = tracing_handler.last_llm_span(_span_id.get())
    with tracer.span("tool_call",
                     attrs={"tool": "search_database", "input": {"query": query}},
                     parent_span_id=llm_parent) as span:
        result = await execute_query(query)
        span.set("result", result)
        return result

---

Silence in Production

Set the env var — no code change needed:

GTRACER_ENABLED=false

Tracing mechanics (spans, callbacks, token counts) stay fully active. Only stdout output is suppressed.

---

Span Schema

Every span event is a flat JSON object on a single line:

{
  "ts": "2026-03-30T10:00:00",
  "level": "TRACE",
  "event": "span.end",
  "span_name": "llm_call",
  "trace_id": "abc123",
  "span_id": "a1b2c3d4",
  "parent_span_id": "e5f6a7b8",
  "status": "ok",
  "duration_ms": 1823,
  "attrs": {
    "agent": "main",
    "model": "claude-sonnet-4-6",
    "seq": 2,
    "tokens": {
      "input": 461,
      "output": 277,
      "total": 738,
      "input_cache_read": 15541
    },
    "stop_reason": "tool_use"
  }
}

Each llm_call span captures:

Field	Description
attrs.tokens	input, output, total, cache_read, cache_creation
attrs.model	exact model ID from the provider response
attrs.delta	new messages added since the previous LLM call
duration_ms	wall-clock latency in milliseconds
attrs.stop_reason	tool_use, end_turn, etc.

---

Configuration

from gtracer import configure

configure(truncation_limit=50_000)  # max chars for message content fields (default)

---

Supported Patterns

Pattern	Description
create_agent	ReAct loop with tool use and structured output
StateGraph	LangGraph graphs with custom nodes
Nested agents	Agent-as-a-tool with causal span parenting
Deep Agents	LangChain create_deep_agent with sub-agents
Parallel tools	Concurrent tool calls under the same llm_call parent

See docs/documentation.md for full integration patterns, API reference, and gotchas.

---

Requirements