Rewind is a time-travel debugger for AI agents -- a single binary with zero dependencies. Below is the complete command reference.
| Command | Description |
|---|---|
rewind record [--replay] |
Start the recording proxy. --replay enables instant replay caching. |
rewind sessions |
List all recorded sessions |
rewind show <id|latest> |
Print a session's step-by-step trace |
rewind inspect <id|latest> |
Open the interactive TUI |
rewind replay <id> --from <step> [--fork-id <id>] |
Replay from a fork point -- cached steps instant, live from fork onward. --fork-id reuses an existing fork (set by MCP / web UI to avoid double-forking) |
rewind fork <id> --at <step> |
Create a timeline branch at a specific step |
rewind diff <id> <left> <right> |
Compare two timelines side by side |
rewind snapshot [dir] --label <name> |
Capture workspace state as a checkpoint |
rewind restore <id|label> |
Restore workspace from a snapshot |
rewind snapshots |
List all snapshots |
rewind cache |
Show instant replay cache statistics |
rewind assert baseline <id> --name <name> |
Create a regression baseline from a session |
rewind assert check <id> --against <name> |
Check a session against a baseline |
rewind assert list |
List all baselines |
rewind assert show <name> |
Show baseline step signatures |
rewind assert delete <name> |
Delete a baseline |
rewind eval dataset create <name> |
Create a new evaluation dataset |
rewind eval dataset import <name> <file.jsonl> |
Import test cases from JSONL |
rewind eval dataset show <name> |
Show dataset with example previews |
rewind eval evaluator create <name> -t <type> |
Create an evaluator (exact_match, contains, regex, json_schema, tool_use_match, llm_judge, custom) |
rewind eval run <dataset> -c <cmd> -e <evaluator> |
Run experiment -- execute command per example, score, aggregate |
rewind eval compare <left> <right> |
Compare two experiments side-by-side |
rewind eval show <experiment> |
Show detailed experiment results |
rewind eval experiments |
List all experiments |
rewind fix <session> |
AI-powered diagnosis -- find root cause, suggest fix, optionally fork + replay with patch (guide) |
rewind fix <session> --apply |
Apply the suggested fix: fork, start proxy with rewrites, wait for re-run |
rewind fix <session> --apply -c <cmd> |
Fully automated: diagnose, fork, run agent, report savings |
rewind fix <session> --hypothesis <fix> |
Skip diagnosis, test a fix directly (e.g., swap_model:gpt-4o) |
rewind eval score <session> -e <evaluator> |
Score a session's timeline with LLM-as-judge or other evaluators |
rewind share <session> |
Generate a self-contained HTML file for sharing (no install needed) |
rewind import otel --json-file <path> |
Import OTLP traces from JSON or protobuf files |
rewind import from-langfuse --trace <id> |
Import a trace from Langfuse by ID |
rewind hooks install |
Install Claude Code hooks for session observation |
rewind export otel <session> |
Export a session as OpenTelemetry traces via OTLP (guide) |
rewind query "SQL" |
Run a read-only SQL query against the Rewind database |
rewind query --tables |
Show all tables and their column schemas |
rewind web [--port 4800] |
Start the web dashboard (flight recorder + live) |
rewind record --web |
Start recording with live web dashboard |
rewind demo |
Seed demo data to explore without API keys |