Skip to content

minimax-ai/verification-loop

v0.1.0Apache-2.0

Enforce a claim-to-proof workflow: define the evidence, run the right verification, and distinguish verified results from assumptions before declaring work complete.

GitHub

Verification Loop

The problem

Agents routinely declare work "done" without evidence: the fix was written but never run, the test suite was never invoked, or a claim of "works now" rests on confidence rather than output. Downstream, a human discovers the gap only when something breaks.

This Plugin installs one Skill that enforces a claim-to-proof loop: state the claim, decide what evidence would prove it, run the smallest meaningful verification, read the actual result, and report verified / not verified / could-not-verify explicitly.

Try it

Use the verification-loop skill, finish the pagination fix, and report completion with evidence.

Expected result: the agent states each claim it is about to make ("page 2 returns 11-20"), runs the smallest check that proves it (the endpoint or the targeted test), quotes the real output, and labels anything it could not check as not verified.

What the Skill does

  • Core rule: a change is not done until the relevant claim has been verified with the smallest meaningful evidence.
  • Six-step loop: enumerate every completion claim, decide evidence for each, run the smallest verification for each, record the command/action and concise observed result, iterate on any failure, and account for every claim before completion.
  • Verification categories: behavior (run the flow or minimal repro), tests (targeted first, then broader), types/build/lint when relevant, review claims tied to concrete output, and separating product failures from environment limits.
  • Completion format: every "Verified" or "Failed" claim includes concrete evidence (command + exit status + observed result, or manual action + outcome). Untested work uses "Not verified"; blocked work uses "Could not verify" with the reason and "Next proof step" explicitly.

Requirements

  • No additional executables, accounts, or paid services. The Skill directs how the agent uses the host's existing run/test tools.
  • Platform-independent; supported wherever MiniMax Code Agent Plugins 1.0 Skills are available.

Data and network

  • This Plugin has no fixed network dependency or destination. Each host-project command or user-facing flow selected for verification retains its own network behavior; even a command run locally may contact registries, test services, or remote endpoints. The Skill requires inspecting that behavior first; unknown behavior is treated as remote and skipped until the destination and data flow are explicitly confirmed.
  • Before any remote action, the Skill requires disclosure of the exact destination, minimal data sent, and whether remote state may change. The destination must be user-supplied or explicitly confirmed by the user. Any potentially mutating remote check requires explicit confirmation even when the destination was already supplied.
  • Credentials, personal data, private endpoints, and unrelated repository content must not be sent or included in completion evidence.
  • If host policy, connectivity, or data-safety constraints block the request, the Skill records the remote behavior as "Could not verify", states why it was blocked, and gives the next safe proof step.
  • This Plugin requires no credentials of its own.

License

Apache-2.0. See LICENSE.