[jgrodziski]

Why

build.sh builds every extractor, and five of them already have a bin/ command: committed wrapper scripts that run their build output, resolved relative to themselves. The Java jar and the C# binary had none, yet docs/metrics.md tells clone users to run ./bin/codegraph-java, which failed with "no such file".

What

• bin/codegraph-java runs extractors/java/target/codegraph-java.jar on the first Java at or above pom.xml's maven.compiler.release (17), looking in JAVA_HOME, on PATH, under ~/.sdkman, then macOS's java_home. A JRE is enough, since ECJ only needs the JVM's class library. Non-interactive shells never source sdkman, so no export is needed. The floor is read from the pom, not hard-coded.
• bin/codegraph-csharp runs extractors/csharp/dist/<host rid>/codegraph-csharp, with the .exe fallback the ruby and metrics wrappers use.
• test.sh: a new gate runs the jar through ./bin/codegraph-java and must reproduce fixtures/java/expected/model.jsonl, with the same root rewrite as the native gate. The published-C# gate now goes through ./bin/codegraph-csharp, as the Elixir and Ruby gates already do with their wrappers.
• ci.yml csharp-smoke downloaded its artifact into bin/. On the four Unix runners that would overwrite the committed bin/codegraph-csharp wrapper. The download now lands in extractors/csharp/dist/<rid>/, where build.sh publishes it, and the job still runs the binary directly.
• Docs: CLAUDE.md, the build.sh header, extractors/java/README.md, docs/cli.md, docs/csharp-extractor.md and the site's install and Java/C# reference pages now name the two commands.

Verification

• Red first: before the wrappers existed, both gate commands exited 127.
• ./test.sh --java: green. That covers 220 Maven tests, the jar via bin/codegraph-java matching the snapshot, and the native binary matching it.
• ./test.sh --csharp: green. That covers 84 tests and the published binary via bin/codegraph-csharp, byte for byte.
• Java selection, by hand:
  • A missing jar or binary prints the ./build.sh --… hint and exits 1.
  • When only a Java 11 is available, the wrapper refuses with a clear message.
  • A Java 11 in JAVA_HOME is skipped in favor of sdkman's Java 25, which extracts the fixture.
  • JAVA_HOME=21 takes precedence.
  • Version strings parse correctly: 1.8.0_392→8, 21→21, 17-ea→17, 21+35→21.
• Not verified locally:
  • The csharp-smoke change runs only in CI on five runners, so this PR is its test.
  • The website pages were not built with Hugo.

Noticed, not changed

scripts/lib.sh's _java_major matches only N.N versions, so a JDK printing a bare openjdk version "21" (an initial GA release) is silently skipped by build.sh and test.sh. The new wrapper's parser handles that case.

🤖 Generated with Claude Code

https://claude.ai/code/session_01PKCVh1pKrt8Z17Ln1pEEkq

[cloudflare-workers-and-pages]

Deploying with    Cloudflare Workers

The latest updates on your project. Learn more about integrating Git with Workers.

Status	Name	Latest Commit	Preview URL	Updated (UTC)
✅ Deployment successful! View logs	codegraph	f32c62d	Commit Preview URL Branch Preview URL	Nov 05 2026, 11:54 AM