Review Summary

Nice work, @kwin! Replacing the Roaster + ASM dependencies with a standard JDK javadoc doclet is a clean architectural improvement. The intermediate properties file format provides good separation between metadata extraction and rendering, and the fileContains pre-filtering to avoid classpath resolution issues is a smart optimization. The behavioral equivalence with the old code looks correct — e.g., the Long suffix stripping is correctly omitted since VariableElement.getConstantValue() returns a proper Long object.

Two minor observations on the new code:

Low Severity

1. Redundant BufferedWriter wrapping — In ConfigurationCollectorDoclet.java, Files.newBufferedWriter() already returns a BufferedWriter, so the outer new BufferedWriter(...) wrapper is unnecessary. Harmless, but an extra allocation:

   try (Writer writer = Files.newBufferedWriter(output, StandardCharsets.UTF_8)) {
   

2. Broad exception catch in fileContains — In CollectConfiguration.java, the fileContains method catches Exception instead of IOException. Since Files.readString() only throws IOException as a checked exception, catching the broader Exception could mask programming errors (e.g., NullPointerException) by silently returning false. Consider narrowing to IOException.

Notes

• The PR title appears truncated: "Use custom doclet to extract configuration metadata with the help of" — consider completing it.
• CI checks are still pending at time of review.

This review was generated by an AI agent and may contain inaccuracies. Please verify all suggestions before applying.

Claude Code on behalf of Guillaume Nodet

Follow-up Review

Thanks for the quick update, @kwin! Both findings from the initial review have been addressed in the new commit:

• ✅ Redundant BufferedWriter wrapping removed
• ✅ Broad Exception catch narrowed to IOException

The new enum resolution logic is well-structured — the fallback chain from compile-time constant → enum constant kind → initializer expression analysis covers the practical cases nicely. The test fixture exercises both indirect enum references (DEFAULT_ENUM = SampleEnum.VALUE_A) and direct enum constant references (SampleEnum#VALUE_B), which is good coverage.

Three minor observations on the new code (all low severity):

1. Stale assertion message — assertEquals(4, ...) says "expected three configuration keys" (see inline)
2. Redundant docTrees.getTree(field) call — called twice in resolveEnumReference, could use Java 16+ pattern matching instanceof (see inline)
3. Theoretical null from docTrees.getPath() — resolveReferencedField passes the result of docTrees.getPath(contextField) directly to new DocTreePath(...) without a null check. In practice this is very unlikely since the doclet only processes included source elements, but a null guard would be more defensive.

None of these are blocking — the PR looks good overall.

This review was generated by an AI agent and may contain inaccuracies. Please verify all suggestions before applying.

Claude Code on behalf of Guillaume Nodet

Follow-up Review (commits 3 & 4)

Thanks for applying the suggestions from the last review, @kwin — commit 4 (727e16bf) correctly adopts pattern matching instanceof and fixes the assertion message. Both look good.

Regarding commit 3 ("Restore maven mode"), the dual-mode architecture is clean — the split into processResolverField / processMavenField keeps the two code paths well-separated, and the @Config annotation attribute reading is correct. The mode is properly threaded from picocli CLI through to the doclet. A few observations:

Medium

1. No test coverage for Maven mode — The existing test only exercises the resolver path (Javadoc block tags). The Maven mode has meaningfully different logic (annotation-based metadata extraction), and without a test fixture that includes an @Config-annotated field, regressions in processMavenField / getAnnotation would go undetected. Consider adding a second fixture and test method similar to the existing one but with --mode=maven.

Low

2. No mode validation in doclet — The mode field is a raw String set by the --mode option. An invalid value (e.g. --mode=typo) silently falls through to resolver mode since doRun checks "maven".equals(mode) and treats everything else as resolver. Practical risk is low because CollectConfiguration validates via the picocli Mode enum, but if the doclet is ever invoked standalone, this could be surprising.

3. Dead null check on description — In processMavenField, the if (description != null) guard on the replace("*", ...) call is unreachable: when docComment != null, renderContent(getFullBody()) returns a non-null String; when docComment == null, description is "". Harmless, but slightly misleading.

Notes

• The PR description still says "Remove unused 'maven' mode" — consider updating it to reflect that Maven mode is now restored.
• Good use of Locale.ROOT in source.toLowerCase() — an improvement over the default-locale alternative.

This review was generated by an AI agent and may contain inaccuracies. Please verify all suggestions before applying.

Claude Code on behalf of Guillaume Nodet

Follow-up review #4 — commit 154e76df (mode validation & cleanup)

Nice work — this commit addresses two of the three findings from the previous round:

✅ Mode validation (previously flagged): The explicit "resolver".equals(mode) check with else throw IllegalArgumentException correctly prevents unknown modes from silently falling through to resolver processing. The exception is caught by the RuntimeException handler in run() and properly reported.

✅ Dead null check (previously flagged): The if (description != null) guard was correctly removed — the ternary always produces a non-null string.

✅ system_properties source: Good addition completing the enum mapping alongside model and user_properties.

One low-severity observation (not blocking): The mode validation happens inside the field iteration loop (line 157). If scanned sources happened to contain no constant fields, the loop body would never execute and an invalid mode would silently succeed. In practice this is unreachable because CollectConfiguration constrains the mode to valid enum values and pre-filters sources to those containing configuration markers — but moving the validation to the top of doRun() (alongside the existing output == null check) would make the invariant self-documenting.

Still open from prior rounds: No test coverage for the maven mode path or the new IllegalArgumentException path. Not blocking for this commit.

This review was generated by an AI agent and may contain inaccuracies. Please verify all suggestions before applying.

Claude Code on behalf of Guillaume Nodet

Follow-up review #5 — commits 6–8 (pattern matching, mode validation, FQCN enum types)

Great progress — commits 6-7 address our prior review findings (pattern matching, dead null check, mode validation), and commit 8 is a substantial improvement with proper FQCN resolution using DocTreePath/DocTrees.getElement(). The Reporter integration replacing System.err is also a nice touch.

Findings:

1. [Important] Wrong tag name in extractClassLink() error messages

extractClassLink() processes @configurationType content:

String configurationType =
    getConfigurationType(extractClassLink(field, docComment, blockTags.get("configurationType")));

But all three error messages inside the method reference @configurationDefaultValue:

throw new IllegalArgumentException("Missing content for @configurationDefaultValue");
// ...
"Expected a class link in @configurationDefaultValue, but got a member reference: ..."
// ...
"No valid {@link ...} reference found in @configurationDefaultValue"

This would send developers chasing the wrong tag when they hit these errors. Should be @configurationType in all three messages.

2. [Minor] Stale javadoc on resolveEnumReference()

The method javadoc still says it returns "including the enum type in the result (e.g. SomeEnum.VALUE)", but after commit 8 it now returns just the constant name (e.g., VALUE). The example contradicts the behavior.

Non-blocking observations:

• The {@link ...} prefix/suffix stripping in getConfigurationType() is now dead code (since extractClassLink() returns a FQCN directly), but the java.lang. stripping is still needed. Minor cleanup opportunity.
• In lookupConstant(), the else branch uses field.getConstantValue() again instead of the already-extracted value variable.

CI: Still running (all jobs pending at time of review).

Verdict: Solid improvements overall. Finding #1 is the only actionable item — a quick fix to swap the tag name in 3 error messages.

Reviewed with AI on behalf of @gnodet. This review was generated by an AI agent and may contain inaccuracies; please verify all suggestions before applying.