SuperAPI AI Directory

review-conformance

Testing & quality

View repository

Updated 10-17 16:22

About this skill

Review Conformance

Announce at start: "Using the review-conformance skill against ."

Preconditions

  • You were dispatched with exactly one guidelines file, the paths it governs and a diff range. One dispatch reviews one file; a second file is a second dispatch.
  • You have no implementation narrative: no plan, no dispatch reports, no account of why the code is the way it is. The diff and the file are the whole case.
  • Read the guidelines file and the diff, nothing else of the project's guidance: not another guidelines file, not the installed default, not a unit's file or the root's in its place. The file you were handed stands alone.
  • Change nothing: no edit, no commit, no task worktree. Another agent fixes your findings, and the next round reviews the fix.

Steps

  1. Check every changed line under the governed paths against every clause. Restrict the diff to those paths (git diff <range> -- <path>...). A clause with no changed line it applies to needs nothing; a clause that applies is checked on every line it applies to, not on a sample.

  2. Write each finding in one shape:

    **finding-N**: guideline, <low | medium | high> — <clause cited as Heading N, quoted> — <file:line>
    

    Number findings from 1 in the report. Cite the clause by its heading and number and quote the words the diff breaks. Severity is as merge-change's review checklist defines it: high loses work or makes a gate pass that should fail in normal use; medium misleads in normal use or leaves a stated rule unmet; low needs a rare or contrived state, or is a gap where the code is right.

  3. Report a clause that contradicts the floor as a finding against the clause, cited at the clause's own line in the guidelines file. The floor wins, and a change that obeys such a clause against the floor is also a finding. The floor, in full:

    • every test carries verifies:;
    • every new test was watched red;
    • every test fails when the behavior it verifies: breaks, so an assertion only that a double was called meets it only where the call is the requirement;
    • class B and C: abnormal-input tests for every REQ and LLR;
    • class C: tests at every touched SDD item's interface;
    • every SDD item traces to a REQ;
    • class C items have LLRs;
    • a mechanism is stated for segregation between items of different classes;
    • every third-party component is in soup.md;
    • a cross-unit dependency is declared in depends_on:, with segregated_from: where the classes differ.
  4. Report a clause that no longer stands as written, each as a finding against the clause:

    • a clause whose reason no longer fits its rule, because the rule was changed and the reason kept;
    • a clause that refers to another guidelines file or to a document the project does not contain, since the reader cannot apply a rule it cannot open.
  5. What the file does not state is a note:, never a finding. A note may say what you would have flagged and why; it blocks nothing.

  6. End with verdict: <one line>: the count of findings by severity and whether any is above low.

Red flags

Thought Reality
"This is bad but the file doesn't say so" A note, not a finding.
"The clause seems wrong here, I'll let it go" A finding against the clause. Never a silent waiver.
"I'll read the installed default for context" You read one file.
"The floor is someone else's review" A clause that contradicts it is yours to report.
"This finding is obvious, the clause need not be cited" A finding with no clause is a note.
"The fix is one line; I'll make it" You change nothing. Report it; a fixer makes it.

Done when

  • Every clause was checked against the governed diff.
  • Every finding cites a clause, by heading and number, quoted.
  • The report ends with a verdict line.