Skip to content

Tests and delivery proof

HelmIt uses your project’s test, build, and lint commands. Preparation records them in .helmit/KEEL.md; if a command does not exist, that is made explicit. The agent chooses a check that matches the promised outcome, runs the applicable commands, and records the result.

MomentWhat is checked
At every logical commitA quick floor checks the staged content and stops the commit if it fails.
When a task or correction is completedThe task’s focused check proves its declared outcome. Intermediate commits do not repeat that final check.
At integration and phase closureComplete test, build, and lint commands run when the work needs to be checked together. At ship, HelmIt also checks requirements, proof, and acceptance criteria. A complete proof that is still valid can be reused.

A CHG follows the same commit floor and has its own focused check, even though it sits outside phases. The optional /helmit:validate command lets you inspect final proof before ship; it adds no mandatory stage.

A check must show the requested behavior, not merely exit successfully. For example, if a task fixes a broken link, a successful site build alone does not show that the link works. For a documentation-only change, checking the published page may be more useful than adding a code test.

  • Reproducible defect: when possible, capture the failure before the fix and confirm that the same check passes afterward. There is no need to manufacture a red test for every change.
  • New behavior: check the outcome someone using the feature can observe, along with any necessary structural checks.
  • Refactor: start with a known result and confirm that behavior is preserved.
  • Human judgment: reserve it for questions requiring human assessment or access. A check the agent can perform remains the agent’s responsibility.

In the plan, verify: identifies the focused check. satisfies: means a task delivers a requirement; supports: means it contributes without completing the requirement on its own. At ship, a requirement is proved only when its mapped evidence passes and its acceptance criteria are met.

A failed commit floor stops the commit. A failed focused check stops the task from being declared complete. Failed complete proof, a missing requirement, or unmet acceptance stops phase closure. The agent fixes the issue and repeats the affected check; --no-verify is not an escape hatch.

Proof runs on a temporary copy of the Git content for the moment being checked. This keeps repository-relative test and build output out of the working tree. It does not isolate commands that access external paths, the network, or services; those remain subject to normal authorization.

See Project artifacts to locate plans, requirements, and results, and The delivery lifecycle to place these checks in the overall flow.