| name | review |
| description | Defines the full-scope review loop until dry for backend, frontend, and overall review objectives. Read for a review objective or when a final objective must complete an omitted review. |
Review
The compiler reports defects in artifacts that exist. It cannot report a required model, operation, test, screen, or journey nobody created. Completeness therefore requires direct review of the entire current scope.
docs/analysis/ is immutable and authoritative. Correct the application, never the requirements. Any application artifact may be wrong, including an upstream schema or DTO that every downstream layer copied consistently.
Read the detailed procedure for the current objective before beginning:
- Backend Review:
backend.md
- Frontend Review:
frontend.md
- Overall Review:
overall.md
Review Loop Until Dry
Review loop until dry means this exact procedure:
- Perform one read-only, literal full reading of the entire current scope and apply every branch in the detailed procedure.
- Keep a current-round finding list while continuing through the final scoped artifact. Do not edit the product or stop the round after finding several problems.
- After the full round ends, fix every finding and its complete consequence surface, regenerate derived artifacts, and settle the required compiler and runtime gates.
- Start another literal full reading at the first requirement. Correction work and reads from earlier rounds count as nothing in the new round.
- Repeat without any round limit until one entire current round reaches the end, finds no problem, and makes no product or generated-file edit.
No exception or discretionary judgment may shorten or replace this procedure.
Scope
The detailed procedure defines the product files, live behavior, journeys, and relationships in scope.
Where a procedure puts generated SDK output in scope, it is a current consumer contract: inspect it, but correct its authored source and regenerate rather than editing the generated file. Dependencies, caches, logs, screenshots, build output, generated Prisma clients, and review bookkeeping are not product review scope.
Within that scope, do not exclude an item because it is large, repetitive, configuration-only, apparently trivial, unchanged, or already examined during an earlier objective, round, or correction.
Findings
Keep the current round's findings in the current objective's working context. Do not create or modify a product file merely to preserve review bookkeeping.
Any required artifact or behavior that is absent is a finding.
For each finding, retain the requirement or upstream contract, the exact conflicting artifact or behavior, why they disagree, and the complete consequence surface. Do not replace this information with transcripts, command output, or guesses.
Literal Full Reading
A file counts as read only when its complete current contents are returned and examined during the current round.
- Build and return one complete sorted current-scope manifest before every round. Reusing or splitting an earlier manifest does not count.
- Start at the first requirement and continue in manifest order through the final scoped artifact.
- Read exactly one manifest file per command. A command that reads multiple files counts as reading none of them.
- Return the file in full without truncation. Consecutive ranges for one large file count only when they cover the first through final line with no gap.
- Searches, match excerpts, summaries, inventories, diffs, Git status, line counts, builds, lint output, test output, and previous reads count as zero.
- Track the current position and completed propagation roots in the current objective. Never call a round full, complete, clean, final, or finished before both reach their end.
- If a product or generated file changes during the round, invalidate the round and restart it from the first requirement against a new manifest.
- After compaction or resume, continue only when the exact manifest, round, next item, and completed propagation roots are known. Otherwise restart the round from the first requirement.
Never partition a round by file, package, layer, requirement subset, journey, review lens, time window, or agent. Never compose partial passes into a round.
Correction And Completion
When a completed round has findings:
- fix every finding at its owning layer and every downstream consequence, including every unimplemented requirement regardless of scale or redesign;
- regenerate every affected derived artifact;
- run the gates named by the instruction and fix every failure;
- reconcile every finding with the actual correction; and
- begin a new full round at the first requirement.
When a completed round has no finding and made no scoped edit, run the instruction's final gates. A failure or resulting change invalidates the round and requires correction followed by another full round.
There is no small-fix exception. Any scoped change caused by a compiler, generator, test, browser, runtime check, or temporary calibration invalidates the round. Restore temporary changes, then start a new full round. Gates must describe the workspace after its last scoped change.
The review Goal is complete only after one dry round and unchanged clean gates.
Mutation Calibration
A round that finds nothing is a clean candidate, not a dry round. Before the round that may qualify as dry, prove the suite can still fail: temporarily remove one material behavior, run its test and require failure, restore the behavior exactly, then require success.
That calibration is itself a scoped change, so it invalidates the round it ran in. Calibrate, then perform a fresh full round, and only that round may qualify as dry.
No Discretionary Stop
Never stop because:
- the review was extensive, productive, expensive, time-consuming, repetitive, or under context pressure;
- many or important defects were found, most defects were probably found, another round seems unlikely to help, or further review seems inefficient;
- a build, test, live journey, search, diff, summary, previous review, or sample passed; or
- a later Final objective might catch or repair an omission.
None of these satisfy review loop until dry. When a literal dry round is not proven, remain active and continue the same review Goal.
