StateHinge

How it works

From agent output to accepted project state—without skipping the decision in between.

StateHinge keeps generated work separate, runs declared checks, shows the exact candidate to a human, and only then allows deliberate promotion. Recovery is verified against the expected baseline.

Plain English

“Promotion” just means moving an approved candidate into the project state your team accepts.

StateHinge does not make the agent a trusted authority. It makes the transition from generated work to accepted work explicit and inspectable.

The five-step path.

1. Bound the task

Define which project and paths the run is allowed to touch.

2. Build separately

The agent produces a candidate away from accepted state.

3. Run checks

Declared deterministic validators must finish in an acceptable state.

4. Review the exact change

A human sees the concrete diff/candidate before approving anything.

5. Promote or recover

Accepted work crosses the boundary deliberately; if recovery is needed, StateHinge verifies the restored baseline.

Without an explicit acceptance layer

  • Agent output and accepted state can blur together.
  • Review reconstructs context after the fact.
  • Interrupted runs can repeat or leave unclear partial state.
  • Rollback status may not prove the baseline actually returned.

With StateHinge

  • Candidate and accepted state stay separate.
  • Checks and exact diff come before approval.
  • Completed checkpoints can be resumed rather than blindly repeated.
  • Recovery evidence checks the expected baseline.

Try the boundary, not the pitch

Use one bounded workflow for 14 days.

No card. No automatic conversion. Keep your current coding-agent workflow.