Skip to content
Elenchos
Install
QuickstartVerificationKaneMCPAgentsEvidenceSecurity

We've locked the contract.

The task and Kane test are hashed before the run. Changing either one during implementation, repair, or Kane execution stops the run.

Isolate

01

A detached Git worktree holds the agent edits.

Kane decides pass or fail

Structured Kane events map each acceptance criterion. An unmapped criterion stays UNVERIFIED and cannot produce VERIFIED.

Repair

A confirmed product FAIL comes back to the agent, bounded.

03

Evidence

Each attempt stores Kane events, logs, and a workspace snapshot.

04

A normal run

  • Requires a clean Git worktree.
  • Creates a detached worktree under `.elenchos/workspaces/<run-id>`.
  • Sends the task to the configured agent in that worktree.
  • Starts the application and waits for the configured URL.
  • Runs `kane-cli testmd run <test> --agent`.
  • Compares Git HEAD and working content before and after Kane.
  • Repairs a confirmed product FAIL up to `maxRepairAttempts`.
  • Removes the detached worktree after evidence capture unless you retain it.

Task format

A task needs an id, a title, and at least one acceptance criterion. Explicit criterion IDs map Kane observations back to the contract.

demo/tasks/add-task.json
{
  "id": "add-task",
  "title": "Add a task",
  "description": "A user can add a task from the main screen.",
  "acceptanceCriteria": [
    { "id": "AC-001", "description": "The task form is visible" },
    { "id": "AC-002", "description": "The new task appears in the list" }
  ],
  "verification": {
    "testFile": "tests/add-task_test.md"
  }
}

Run states

StateMeaning
VERIFIEDKane completed the contract and the repository stayed stable.
FAILEDKane confirmed a product failure, or verify mode finished without a pass.
ERRORThe verifier could not establish a product result.

Kane statuses inside an attempt

StatusMeaning
PASSStructured Kane output passed, and every declared criterion mapped to pass.
FAILKane confirmed a product failure.
VERIFIER_ERRORKane itself could not complete a trustworthy result.
INCONCLUSIVEThe response was incomplete, contradictory, or unusable.

Stability checks

  • Task JSON and Kane test hashes must match the pre-run contract.
  • Git HEAD and working content must match across the Kane window.
  • An overall Kane pass is accepted only when every criterion has a mapped pass.