How delivery would work
Check that the plan preserves what matters
The cancellation plan adds validation and persistence. It gives no task for preserving the shipped-order guard. The reviewer sees which criterion has no task and returns the plan for revision.
Requirements traceability from tasks to tests
Where verification sits
See how criteria constrain a coding plan and how the team distinguishes proposed checks from results tied to a source version.
Criteria are written in Define, covered by tasks in Plan and checked in Verify. Each link is a record a reviewer can open.
Understand
Read the twin around the area you want to change: paths, rules, data and gaps.
AI drafts · you read
Define
Write the use case, stories and acceptance criteria against the current model.
AI drafts · lead edits
Approve requirementPlan
Propose files, symbols, tasks and the impact map. Every criterion needs a task.
AI proposes · dev reviews
Approve planImplement
The coding agent gets an approved, version-pinned brief and stops if inputs change.
Agent writes · dev guides
Review and mergeVerify
Test cases trace back to criteria. Results name the build they ran against.
QA reviews · tests run
Accept resultsRefresh
The merge becomes a new commit. The twin rebuilds and old approvals are rechecked.
Automatic
Every push starts the loop again from a fresh twin. Approvals built on the old code are flagged, not silently kept.
Walkthrough
Step by step.
From criterion to task
Map each requirement to the work that implements or preserves it. A criterion with no covering task is a planning gap. A task with weak source support also needs a review decision before it can count as a complete plan.
From proposed check to observed result
A configured test case describes the expected outcome. A result shows what ran, under which inputs, and against which version. Show incomplete runs, skipped cases and stale results by name.
Read approvals and delivery evidence independently
Review of a plan, QA review of cases, execution results, code review and merge answer different questions. Distinct records preserve who decided what and help a later reader see the behaviour a release actually exercised.
What a reviewer can actually conclude
| Artifact | What it supports | Still needed |
|---|---|---|
| Criterion mapped to a task | The plan accounts for the shipped-order rule | The task and check still need execution |
| Passed fixture assertion | The guard resisted that input in the local fixture | Production compatibility and database effects |
| Skipped client check | The required client version was unavailable | A supported older client or explicit rollout decision |