Upstream issues

Upstream issues

Known issues in upstream tooling that affect jarvis. These are not jarvis bugs — they are limitations or bugs in the SCIP indexers or the scip CLI.

typeHierarchy empty on upstream scip v0.9.0

• Symptom: typeHierarchy returns an error or empty on real indexes.
• Root cause: upstream scip expt-convert never populates global_symbols.relationships (sourcegraph/scip#464).
• Workaround: setup.sh installs a fork build (phuongddx/scip) carrying the #465 fix. Verify with scip --version.
• Status: awaiting upstream merge.

scip-java + Android/AGP

• Symptom: zero SCIP shards produced for Android/Gradle projects.
• Root cause: AGP replaces source sets (sourcegraph/scip-java#177).
• Workaround: jarvis auto-falls-back to --search-only. Navigation tools won’t work, but searchCode and semanticSearch will.

scip-kotlinc version mismatch

• Symptom: scip-kotlinc fails (compiled against one exact Kotlin version).
• Root cause: the compiler-plugin API is unstable across Kotlin versions.
• Workaround: jarvis auto-falls-back to --search-only.

displayName/kind often null

• Symptom: some symbols show null kind or displayName.
• Root cause: upstream scip indexing gap (sourcegraph/scip#464).
• Workaround: jarvis backfills from the symbol string. Works in practice for most repos.

Swift USR stability across toolchain versions

• Symptom: indexes may not be comparable across Swift versions.
• Root cause: clang USRs (which scip-swift emits as symbol names) may change between Swift toolchain versions.
• Workaround: pin .swift-version (currently 6.2.4); reindex after toolchain upgrades.

--semantic-include has no effect on prose-only repos

• Symptom: passing --semantic-include on a markdown-only or otherwise prose-only repo doesn’t change semanticSearch results.
• Root cause: semantic chunking is code-symbol-scoped — it chunks around SCIP symbols, so a repo with no code symbols has nothing for --semantic-include to add (jarvis-index#9, open).
• This is by design. It is a settled limitation, not a bug: a prose-only repo needs its own chunking strategy, which semanticSearch doesn’t implement.