rewind-agent 0.14.0

Chrome DevTools for AI agents — record, inspect, fork, replay, diff.

rewind-agent

Python SDK for Rewind — the time-travel debugger for AI agents.

Record every LLM call. See the exact context window. Fork, fix, replay from failure — without re-running.

Install

pip install rewind-agent

This installs both the Python SDK and the rewind CLI. The native binary is auto-downloaded on first use.

Quick Start

One line to start recording — no proxy, no setup:

import rewind_agent
import openai

rewind_agent.init()  # patches OpenAI + Anthropic automatically

client = openai.OpenAI()
client.chat.completions.create(model="gpt-4o", messages=[...])
# Recorded to ~/.rewind/ — inspect with: rewind show latest

Or as a scoped session:

with rewind_agent.session("my-agent"):
    client = openai.OpenAI()
    client.chat.completions.create(model="gpt-4o", messages=[...])

Two Recording Modes

	Direct mode (default)	Proxy mode
Setup	rewind_agent.init()	rewind record in a second terminal
How	Monkey-patches SDK clients in-process	HTTP proxy intercepts LLM traffic
Best for	Python agents, quick iteration	Any language, polyglot teams

# Direct mode (default — no proxy needed)
rewind_agent.init(mode="direct")

# Proxy mode (requires `rewind record` running)
rewind_agent.init(mode="proxy", proxy_url="http://127.0.0.1:8443")

Replay from Failure

Agent failed at step 5? Fix your code, then replay — steps 1-4 are cached (instant, free), step 5 re-runs live:

with rewind_agent.replay("latest", from_step=4):
    result = my_agent.run("Research Tokyo population")
    # Steps 1-4: instant cached responses (0ms, 0 tokens)
    # Step 5+: live LLM calls, recorded to a new forked timeline

After the replay, diff the timelines: rewind diff <session> main replayed

Regression Testing

Turn any session into a baseline. After code changes, check for regressions:

from rewind_agent import Assertions

# Check the latest session against a known-good baseline
result = Assertions().check("booking-happy-path", "latest")
assert result.passed, f"Regression: {result.failed_checks} checks failed"

Checks step types, models, tool calls, error status, and token usage. Supports configurable tolerance:

result = Assertions().check("my-baseline", "latest", token_tolerance=0.15)
print(f"Passed: {result.passed_checks}/{result.total_checks}")

Agent Hooks

Enrich recordings with semantic labels:

@rewind_agent.step("search")
def search(query: str) -> str:
    return client.chat.completions.create(...)

@rewind_agent.tool("calculator")
def calculate(a: float, b: float) -> float:
    return a + b

with rewind_agent.trace("analysis"):
    rewind_agent.annotate("confidence", 0.92)
    result = search("Tokyo population")

Framework Support

Native (auto-detected on init()): OpenAI Agents SDK, Pydantic AI — zero config.

Wrapper (manual setup):

# LangGraph
graph = rewind_agent.wrap_langgraph(compiled_graph)

# CrewAI
crew = rewind_agent.wrap_crew(crew)

Any other framework works via the HTTP proxy — point OPENAI_BASE_URL at rewind record.

Learn More

• GitHub
• Changelog