Checking conformance and fixing differences

Conformance measures an instance — generated code on disk — against the template it came from. A difference is something the template declares that the instance lacks.

The one rule that governs every language

The instance must carry everything its template declares. Anything it adds is never a difference.

Extra declarations, extra keys, extra lines, extra files, reordered class members: all fine. So the fix for a difference is almost always to add what is missing, never to delete what is extra. An agent that starts removing content to make a file conform is working against the mechanism.

Implement freely inside generated files. Function bodies are not compared at all.

Running it

Per project, or across the workspace:

nx run <project>:conformetry-validate
nx run-many --targets=conformetry-validate

Without Nx, the command-line host does the same:

conformetry validate

Both accept a language filter, and the command-line host also accepts a one-off glob override in place of the configured instances and a --templates filter naming which templates to measure against. Narrowing is for iterating; let the full run be the gate.

conformetry validate --templates nestjs-service-module
conformetry validate --templates nestjs-service-module --languages typescript

--templates and --instances narrow opposite halves of the pairing and the run is their intersection — neither overrides the other. Naming a template that does not exist is refused with the real names listed, and a selection matching no instances is reported as "no instances belong to …" rather than as a clean report, so a typo cannot read as a pass.

A template whose instance groups carry tags cannot be located by the command-line host at all. Those globs are read inside each project the tags select, so src/modules/* means nothing until a project root is joined to it. The host leaves such a group out rather than globbing it from the working directory, and a narrowed run tells you so instead of reporting nothing found. Reach for the Nx targets there, or pass --instances with the paths. This is why conformetry validate --templates nestjs-service-module can report nothing to check while nx run <project>:conformetry-validate checks plenty.

An agent shell is not a terminal, so omitting --templates there validates every template exactly as it did before the flag existed — it never waits on the picker a terminal would be offered. --templates all says the same thing explicitly; all is reserved and cannot name a template.

Reading a report

A report opens with a score summary when anything scored below a perfect match, then lists each file, the instance and template it compared, and every difference with a suggested fix:

  1. file: orders.service.ts
     Instance: packages/widgets/src/modules/orders/orders.service.ts
     Template: configuration/conformetry-templates/nestjs-service-module/…

     1. Missing Decorator "Injectable"
        Instance: Line 12, Column 1
        Template: Line 8, Column 1
        Fix     : Add the missing Decorator "Injectable" to the instance file.

Read the Fix line before inventing a remedy. Every difference carries one, and it names the specific construct rather than describing the file in general.

Locations are 1-based. JSON differences report a path such as scripts.build[0] instead of a line, because a key has no meaningful line in a reformatted file.

For what a difference means in a particular file type — why a string literal is required verbatim, why a heading is pinned to its level, why a function body is free — see references/languages.md.

The kinds of difference, and what each asks of you

A missing file or directory. Checked before any content comparison, and a missing directory suppresses reporting for the files inside it. So create the directory, re-run, and expect a new batch naming its contents. That is the check working, not a regression.

A missing declaration or comment. Add it. Comments are compared as an ordered subsequence, so a section marker in the wrong position reads as missing even though the text is present.

An unmatched instance — no template explains this path. Either the path is not generated code and should not be in the instance globs, or it has drifted so far that nothing recognizes it.

An ambiguous instance — two templates fit equally well and only partially. Either give the templates distinguishing files, or narrow the glob so one applies.

Those last two are attribution failures rather than content failures, and they are the hardest to act on from the report alone. Diagnose them by asking which templates explain the path:

conformetry templates --instances packages/widgets/src/modules/orders

  nestjs-service-module (nsm)
    Generate a NestJS service module
    Template: configuration/conformetry-templates/nestjs-service-module
    Instances:
      packages/widgets/src/modules/orders 5/5 files 100%
  nestjs-command-module (ncm)
    Generate a NestJS command module
    Template: configuration/conformetry-templates/nestjs-command-module
    Instances:
      packages/widgets/src/modules/orders 3/5 files 60%

Every template that explains the path is listed, not just the winner — which is the point. A path legitimately belongs to more than one template, so a single verdict would hide the tie that caused the difference. Read it as a ranking: a template you expected near the top sitting low means its files are not where the instance has them.

The complement asks the other direction — what generated code exists, and what each piece answers to:

conformetry instances
conformetry instances --templates nestjs-service-module

Each path it prints is usable as the --instances argument above, so the two commands compose. --templates narrows to the instances a given template explains, which is how you find every instance that would be affected by changing that template.

Both accept --json for parsing, and --config to read a configuration elsewhere. references/template-matching.md explains how attribution works and how to read the ranking.

Three behaviors that look like bugs and are not

A second run can report new differences after you fixed everything. Only the smallest matching template's requirements are reported per file. Once those are satisfied, requirements unique to a larger template surface. Iterate until clean rather than treating the second report as a regression.

A missing Python interpreter is reported as a difference, not a crash. If differences appear against .py files complaining that Python is unavailable, the fix is to install python3, not to change the template.

Nothing is fixed automatically. There is no autofix and no write mode; the only flag that changes the outcome is --threshold, which changes what passes rather than changing any file. Remediation is either editing the instance by hand or regenerating it — and regeneration overwrites unconditionally, so read conformetry-generate before reaching for it.

Deciding between editing and regenerating

Edit the instance when the differences are few and the file holds work you want. That is almost always the answer.

Regenerate only when the instance has drifted so far that reconstructing it by hand is worse, and only after reading what will be destroyed — generation has no existence check, no merge, and no conflict detection.

Scores and thresholds

A run does not only say whether an instance conforms — it scores how much of its template the instance honours, and fails that instance when the score falls below the threshold that applies to it. Differences carry weights, and the score is the share of weighed requirements honoured.

The report prints a Conformance scores: summary above the differences, listing every instance that scored below a perfect match, whether it met its threshold, and a workspace total.

The threshold defaults to 1 — a perfect match — so a fresh workspace is strict. Three levels set it, narrowest winning: an instance group's threshold, then the generator's threshold, then a run-level --threshold.

Two consequences worth holding on to:

• A lowered threshold permits drift, it does not hide it. Differences still print for an instance that cleared its threshold. Reading a report and finding differences does not mean the run failed — check the score summary.
• Lower a threshold to migrate, not to silence. It is the tool for moving existing instances onto a new template gradually. If an instance can conform, make it conform; there is still no per-difference severity to downgrade.

Before claiming the work is done

A clean conformance run is not optional polish — it is the check that generated code still matches its standard. Run it after generating, after editing generated files, and before saying an implementation is finished.

Reproducing a difference in a sandbox

conformetry-examples carries one instance per kind of difference, deliberately broken, so a report can be understood without the surrounding project's conventions muddying it.

• drift-catalogue — a missing file, a missing directory, a deleted export, a dropped section comment, and a renamed class, with the whole report quoted.
• structural-not-textual — a reformatted instance that passes beside one missing a single export that does not.
• scoring-thresholds — differences printing for an instance that cleared a lowered threshold, and a run that still exits zero.
• ambiguous-attribution — an instance two templates explain equally well.
• failure-modes — two things that conform for the wrong reasons.

Its AGENTS.md maps report text to the example that reproduces it, and lists which changes to an instance are differences and which are not. The broken instances there are broken on purpose — do not repair them.