@agentrewind/cli

Command-line inspection and packaging tools for AgentRewind sessions.

The umbrella package @agentrewind/sdk installs the same binaries, so most users do not need to install this package directly.

Commands

agentrewind quickstart [openai|openai-compatible|openrouter|anthropic]
agentrewind quickstart openai --format ts
agentrewind quickstart openai --out agentrewind-openai.ts
agentrewind list [.rewind]
agentrewind list [.rewind] --json
agentrewind doctor <session>
agentrewind doctor <session-id> --store .rewind
agentrewind doctor latest --store .rewind
agentrewind inspect <session>
agentrewind timeline <session> --kind model_call --site <name>
agentrewind inspect <session> --json
agentrewind context <session>
agentrewind prompt <session> --site <name>
agentrewind context <session> --step <n>
agentrewind context <session> --site <name>
agentrewind context <session> --json
agentrewind diff <session>
agentrewind diff <session> --from <a> --to <b>
agentrewind diff <session> --from-site <a> --to-site <b>
agentrewind fork <session> --site <name> --system "Updated system prompt"
agentrewind fork latest --store .rewind --step <n> --model <model-id>
agentrewind tool <session>
agentrewind tool <session> --name <tool>
agentrewind tool <session> --step <n>
agentrewind tool <session> --json
agentrewind entropy <session>
agentrewind entropy <session> --source uuid
agentrewind entropy <session> --step <n>
agentrewind pack <session> <out.rewind>
agentrewind unpack <out.rewind> <dir>

Alias:

arw inspect <session>

Notes

• quickstart prints install commands and a minimal TypeScript record/replay starter for the selected provider. Use --manager pnpm, --manager yarn, or --manager bun for a different install command. Use --format ts to print raw TypeScript, or --out <file> to write a starter file. Starters use defineHarness() so TypeScript infers the harness result without a manual generic annotation, and they validate required environment variables before recording. They also format validation, record, and replay failures with explainRewindError() and a known session path. --out refuses to overwrite existing files unless --force is passed.
• list shows the sessions inside a store such as .rewind, including each session id, provider, origin, model/tool/entropy counts, token usage, and full session path. Use it when you know the store directory but not the exact .rewind/<session-id> path. ls is an alias, and --json is available for scripts.
• Single-session commands (doctor, inspect, context, diff, and pack) accept a full session path, a session id with --store, the special selector latest --store .rewind, or a store directory when it contains exactly one session.
• doctor validates that the session can be read and prints a plain-language summary with provider, event counts, model/tool steps, redaction status, token usage, and suggested next commands. Use --json for automation.
• inspect prints a labeled timeline table with step, kind, lane, site, a short fingerprint, request detail, token usage, and flags such as stream, error, or provenance=live. timeline is an alias. Use --kind, --site, --errors, --live, --from, and --to to focus large sessions. Use --full-fingerprint when comparing exact request/tool hashes, --json for automation, or --no-header for compact TSV output.
• context and diff operate on model-call steps. By default, context prints a readable view of the first model call and diff compares the first two model calls. prompt is an alias for context; add --json for the raw recorded message array. Pass --site, --from-site, or --to-site when you know the stable site name from your harness. Pass --step, --from, or --to when you need a specific numeric step from inspect.
• fork starts at a recorded model-call step, reuses the recorded prefix, and sends the tail to a live provider client. Pick the fork point with --site or --step, change the live tail with --system and/or --model, and add --dry-run to verify the plan without spending provider tokens. Dry runs do not check credentials unless you add --check-provider. Built-in providers are openai, openai-compatible, openrouter, and anthropic. API keys are read from OPENAI_API_KEY, COMPATIBLE_API_KEY, OPENROUTER_API_KEY, or ANTHROPIC_API_KEY; generic compatible endpoints also need COMPATIBLE_BASE_URL or --base-url. Successful forks write a complete child session: recorded prefix boundaries keep provenance=recorded, live tail model calls get provenance=live, and the child can be inspected or replayed as one complete trajectory.
• tool prints a readable recorded tool call by default. Add --json for args, result, error, stream chunks, latency, and provenance as JSON. Use --name when the tool appears once, or --step after inspect when a tool appears more than once.
• entropy prints a recorded ctx.clock(), ctx.random(), ctx.uuid(), or ctx.env() draw as JSON. Use --source when that source appears once, or --step after inspect when the same source appears more than once.
• pack excludes the local vault and prints a redaction summary before writing the .rewind bundle.
• Session read failures are formatted as AgentRewind errors. If a command cannot identify the session, run agentrewind list .rewind, pass a full session directory such as .rewind/<session-id>, or pass <session-id> --store .rewind.
• Recording and harness-aware replay/fork experiments are programmatic because they need your harness code.