retrace-sdk 0.16.3

Record, replay, fork & share AI agent executions

Record every LLM call, tool invocation, and error your agent makes. Replay it step-by-step like a video. Fork from any point, change the input, and watch the whole agent re-execute down a new path. Share any run as an interactive, public link.

Quick Start · How It Works · Recipes · Enforcement · Docs

---

Install

pip install retrace-sdk

Requires Python 3.10+.

---

Quick Start

import retrace

retrace.configure(api_key="rt_...")  # get a key in the dashboard

@retrace.record(name="research-agent", resumable=True)
def run_agent(prompt: str):
    plan = call_planner(prompt)        # captured automatically
    results = call_tools(plan)         # captured automatically
    return summarize(results)          # captured automatically

run_agent("What changed in vector databases this year?")

That's it. The run appears in the dashboard as an interactive timeline you can scrub, replay, and fork.

---

How It Works

The SDK is a thin capture layer. It wraps your function, auto-instruments provider calls, and streams spans to Retrace over a resilient transport — never blocking or crashing your agent.

flowchart LR
    subgraph yourproc["Your process"]
        fn["@record"] --> cap["auto-captured spans<br/>(LLM · tool · error)"]
        cap --> buf["offline buffer<br/>(bounded, flush on reconnect)"]
    end

    buf == "WebSocket (primary)" ==> api["Retrace"]
    buf -. "HTTP fallback" .-> api
    api == "resume: re-execute from fork point" ==> fn

    classDef p fill:#0f3d2e,stroke:#10b981,color:#d1fae5;
    class api p;

• Capture — provider calls (OpenAI, Anthropic, Gemini) are intercepted automatically; you can also emit manual spans.
• Transport — spans stream over a WebSocket; if it drops, the SDK falls back to HTTP and replays a bounded offline buffer on reconnect, so nothing is lost. A flush-on-exit hook captures runs that end abruptly.
• Resumable — with resumable=True, a dedicated listener handles a resume command and re-executes your function from a fork point with modified input, powering cascade replay from the dashboard.
• Safe by default — failures in the SDK never raise into your agent; typed errors surface real problems explicitly.

---

Auto-Instrumentation

LLM calls from major providers are captured with no extra code — just install the provider SDK alongside retrace-sdk:

Provider	Captured automatically
OpenAI	chat completions
Anthropic	messages
Google Gemini	generate content

---

Capabilities

Capability	What it does
Record	One @record decorator captures the full execution tree.
Cascade replay	resumable=True lets a dashboard fork re-execute the whole function from any step.
Enforcement	Local budget/step/loop ceilings stop a runaway agent before the next call.
Multi-agent	Tag spans with an agent id/role for topology + inter-agent detectors.
Golden cassettes	Record a run as a CI regression fixture and gate on it offline.
Sampling	Record a fraction of traffic in production.
Sessions	Group multi-turn conversations under one session.

---

Usage Recipes

Configure

import retrace

retrace.configure(
    api_key="rt_...",                  # or RETRACE_API_KEY
    base_url="https://api.retraceai.tech",
    project_id="...",                  # or RETRACE_PROJECT_ID
)

Set RETRACE_ENABLED=false to disable recording without changing code.

Resumable execution (cascade replay)

@retrace.record(name="my-agent", resumable=True)
def run_agent(prompt: str):
    plan = call_planner(prompt)
    result = call_executor(plan)
    return summarize(result)

When you fork at any span in the dashboard, the SDK re-executes the entire function with the modified input — not just one call. Every downstream step that depends on the change diverges.

Enforcement (circuit breakers)

import retrace
from retrace import RetraceEnforcementError

retrace.configure(
    api_key="rt_...",
    max_steps_per_run=50,
    max_usd_per_run=2.0,
    server_enforcement=True,  # optional: also consult centrally-managed server policies
)

try:
    run_agent("...")
except RetraceEnforcementError as e:
    print(e.verdict, e.reason)  # e.g. "block", "Local USD ceiling reached: ..."

Local ceilings are enforced offline (zero network). Precedence: explicit arg > env var (RETRACE_MAX_STEPS_PER_RUN, RETRACE_MAX_TOKENS_PER_RUN, RETRACE_MAX_USD_PER_RUN, RETRACE_SERVER_ENFORCEMENT) > unset. If the server check is unreachable, local limits still apply.

Multi-agent context

import retrace

with retrace.agent("planner", role="planner"):
    plan = call_planner(prompt)
with retrace.agent("executor", role="executor"):
    result = call_executor(plan)

Tags spans so the dashboard can draw the agent topology and run inter-agent detectors (ping-pong, reasoning–action mismatch, task derailment).

Golden cassettes (CI regression gates)

from retrace import write_golden_cassette

write_golden_cassette("golden.json", recorder=rec)

Gate on it offline in CI with retrace ci replay.

Sampling

retrace.configure(api_key="rt_...", sample_rate=0.1)  # record 10% of traces

Error handling

from retrace import (
    RetraceError,
    RetraceAuthError,
    RetraceCreditsExhaustedError,
    RetraceRateLimitError,
    RetraceEnforcementError,
)

Typed errors for auth failures, credit exhaustion, rate limiting, and enforcement blocks. Transient transport problems never crash your agent.

---

Links

• Documentation
• GitHub
• PyPI

License

MIT