REWINDWatch 3-min demo

OpenAI Build Week 2026/ Developer tools

A green test is not enough.

Rewind records what a coding agent intended to change, which exact Git state passed its checks, what changed afterward, and where you can recover—without disturbing your current work.

  • Local + network-free demo
  • Deterministic findings
  • Branch-only recovery
REWIND / TASK RECEIPT REVIEW

Task outcome

Completed with items to review

task_demo
!Scope2 files outside
!Freshness2 files changed
!ProtectedReview required
Recoverycp_02 ready
01Intent
02Checkpoint
03Check
04Receipt

The test passed. The final tree changed. Rewind shows both.

3-MINUTE JUDGE PATH0:00 Receipt1:30 Forensics2:15 Build story2:50 Limits

01 / INTERACTIVE WALKTHROUGH

Rewind a real task.

A deterministic walkthrough of the generated upload-worker demo. No backend, model call, or simulated live CLI.

Declared intent

Add retry handling to the upload worker

demo_app/upload/**tests/**

03 / RECEIPT

4 ITEMS TO REVIEW

REVIEW

The test stayed green. The final task state tells a more complete story.

ScopeREVIEW

2 files outside scope

EvidenceREVIEW

2 files changed after pass

ProtectedREVIEW

1 dependency · 1 protected

RecoveryREADY

cp_02 is ready

What the green test missed

pyproject.tomloutside scope · dependency
deploy.yamloutside scope · changed after test
demo_app/upload/worker.pychanged after test
01

Did the agent stay inside the scope I declared?

02

Which exact Git tree passed the check?

03

What changed after that check?

04

Where can I recover without disturbing my work?

The terminal and self-contained HTML report render the same rewind.receipt.v1result. There is no second risk engine.

02 / TRUST CHAIN

Evidence bound to exact state.

Rewind stays intentionally narrow: inspectable Python, normal Git, and deterministic results from CLI to report.

  1. 01

    Declare

    Sign the task intent, allowed paths, and protected-path policy.

  2. 02

    Checkpoint

    Create a real Git commit through a temporary index without moving HEAD.

  3. 03

    Verify

    Run argv directly and hash stdout, stderr, exit status, and exact tree.

  4. 04

    Receipt

    Compare the tested tree with the final tree and surface review items.

  5. 05

    Recover

    Create a new branch only after checkpoint and evidence integrity pass.

03 / HISTORICAL REPLAY

Authenticity is not authority.

A signed, authentic record can still prove that an observed action was not authorized. Select a failure case to replay the four levels.

Replay verdict

NOT AUTHORIZED

FAIL CLOSED
L0 RecordPASS
L1 IntegrityPASS
L2 BindingPASS
L3 AuthorityFAIL

The record is authentic and correctly bound. As-of authority fails because the actor's role had already been revoked.

Rewind records deployment_observed. It does not claim to perform or block deployment.

04 / BUILD WEEK PROVENANCE

Built with Codex. Deterministic at runtime.

Rewind began as our team's research idea. Codex with GPT-5.6 Sol helped explore product shapes, implement the Python CLI and MCP adapter, pressure-test failure cases, and harden the final workflow. Humans locked the scope and trust decisions.

After an explicit bootstrap exception, Rewind recorded the rest of its own build. Rewind itself makes no model or OpenAI API calls at runtime; its findings are deterministic.

FAST Prototype directionsULTRA Overnight subagent testingHUMAN Scope + trust decisions
47tests passed
4forensic fixtures
5MCP tools with CLI parity
0runtime model calls

05 / VERIFY IT YOURSELF

One command path. No network.

Generate the developer task receipt and all four signed forensic cases locally. The output is disposable by design—no project key or private recorder state is committed.

JUDGE QUICK START
python3 -m venv venv
venv/bin/pip install -e '.[dev]'
venv/bin/rewind demo --output /tmp/rewind-judge-demo
→ task-report.html→ forensic-report.html

06 / HONEST BOUNDARY

Tamper-evident, not magic.

01One trusted local recorder key—not multi-party non-repudiation.

02Append-only JSONL is not immutable storage without external anchoring.

03Records and detects; it is not a universal filesystem gateway.

04Policy replay intentionally supports only documented v1 and v2.

05Recovery creates a branch; it never rolls back the working tree.

06macOS and Linux are supported. Windows remains untested.