continuum-observability

Continuum Observability Skill

Authoritative source: docs/observability.md.

---

Quick start

Tracing is on by default if Langfuse is configured.

docker compose --profile observability up -d              # brings up Langfuse on :3000

LANGFUSE_ENABLED=true
LANGFUSE_HOST=http://localhost:3000
LANGFUSE_PUBLIC_KEY=…                                      # from Langfuse UI
LANGFUSE_SECRET_KEY=…

Open http://localhost:3000 — every agent run, LLM call, and tool call shows up automatically.

To disable: LANGFUSE_ENABLED=false.

---

Trace your own functions

from continuum.observability import observe, SpanLevel

@observe(name="my-pipeline", capture_input=True, capture_output=True,
         metadata={"version": "v2"}, level=SpanLevel.DEFAULT)
async def run_pipeline(data):
    ...

Specialized variants:

from continuum.observability import trace_tool, trace_agent

@trace_tool(name="search_db", tool_type="database")
def search_db(query): ...

@trace_agent
async def my_custom_agent_runner(...): ...

---

Manual spans

from continuum.core.container import get_container

mgr = get_container().tracing_manager

with mgr.trace(name="onboarding", user_id="u1", session_id="s1",
               metadata={"tier": "pro"}, tags=["beta"]) as trace:
    with mgr.span(name="step-1", input={"foo": 1}) as span:
        span.event(name="cache-hit")
        span.score(name="quality", value=0.92)

Generation spans (LLM calls):

with mgr.span(name="step") as span:
    gen = span.generation(name="llm-call", model="gpt-4o-mini",
                          input=messages)
    # … call the LLM …
    gen.end(output=response_text,
            usage_prompt_tokens=120, usage_completion_tokens=80, usage_total_tokens=200)

---

Async-safe trace context

from continuum.observability import (
    set_trace_context, restore_trace_context,
    get_current_trace_id, get_current_session_id, get_current_user_id,
)

token = set_trace_context(trace_id="abc", user_id="u1", session_id="s1")
try:
    await client.chat(messages)                # inherits the trace context
finally:
    restore_trace_context(token)