Angular component sibling files (ts/html/scss) merged into one File node; search_code silently incomplete

[artaommahe]

Version

codebase-memory-mcp 0.8.1

Platform

macOS (Apple Silicon)

Install channel

GitHub release archive / install.sh / install.ps1

Binary variant

standard

What happened, and what did you expect?

When indexing an Angular project, the indexer creates one File node per component and folds the component's sibling files (foo.component.ts / .html / .scss) into it. Only one sibling's path survives as file_path, and which one wins is inconsistent (sometimes .ts, sometimes .html, sometimes .scss). The other siblings disappear from the file inventory entirely.

Consequences:

1. search_code silently returns incomplete results — it only greps files that have their own File node, so matches inside merged-away templates are never found, with no truncation warning. On a real ~10k-file Angular monorepo, a sweep for a CSS class found 95 of 137 grep matches (~70%); ~630 of 1,041 .html templates had no File node.
2. Wrong-facet paths in results — a component can be represented by its .scss path, so query results/callers point at a stylesheet.
3. extension / file_path mismatch — on the large repo, nodes appear with extension: '.html' but a file_path ending in .ts or .scss.

Expected: one File node per file on disk (or, if component+template merging is intentional for call-graph purposes, the merged siblings should still be present in the file inventory that search_code and file-level queries operate on — call extraction itself handles the merge fine, see below).

Reproduction

Public minimal repro (13 source files, dummy code): https://github.com/artaommahe/codebase-memory-mcp-angular-repro @ bbc0779

It contains three tiny components engineered to trigger each merge outcome (add-absence: ts+html; checkbox: ts+html+spec+stories; badge: ts+html+scss), a standalone help.html control with no component sibling, and a FeatureService.isEnabled() called from two components. All four .html files contain the literal marker repro-marker.

1. index_repository {"repo_path":"<repro>","mode":"full"} → 34 nodes / 54 edges

2. query_graph {"query":"MATCH (f:File) RETURN f.file_path, f.extension ORDER BY f.file_path"}

   Result — 7 File nodes instead of 13:

   src/app/add-absence/add-absence.component.ts   .ts     <- add-absence.component.html GONE
   src/app/badge/badge.component.scss             .scss   <- badge .ts AND .html GONE
   src/app/checkbox/checkbox.component.html       .html   <- checkbox.component.ts GONE
   src/app/checkbox/checkbox.component.spec.ts    .ts
   src/app/checkbox/checkbox.component.stories.ts .ts
   src/app/feature.service.ts                     .ts
   src/standalone/help.html                       .html
   

3. search_code {"pattern":"repro-marker","mode":"files"}

   Result — 2 of 4 files (checkbox.component.html, help.html); missing add-absence.component.html and badge.component.html, both found by plain grep -rl repro-marker src/.

4. Control checks: trace_path {"function_name":"isEnabled","direction":"inbound"} correctly returns both callers (AddAbsenceComponent.save, BadgeComponent.isHighlighted) — call extraction is unaffected; the standalone help.html (no component sibling) is indexed correctly.

Full expected-vs-actual write-up in the repro repo's README.

Project scale (if relevant)

Repro: 34 nodes / 54 edges / 13 files. Real-world Angular monorepo where this was found: 60,541 nodes / 124,930 edges / 10,330 files.

Confirmations

• I searched existing issues and this is not a duplicate (closest is task: Parsing & extraction quality — language/format coverage gaps #592, the generic parsing-quality meta-task).
• My reproduction uses shareable code (a dummy snippet or a public OSS repository), not proprietary code.

[artaommahe]

Follow-up on consequence 3 (extension / file_path mismatch) — it now reproduces in the repro repo too, and it turns out to require git history:

• At the initial commit (bbc0779, single commit), extensions are consistent with the surviving path (add-absence.component.ts → .ts, badge.component.scss → .scss).
• After adding commits that touch the merged siblings separately (10f18b4..5d07f94: one edits the .html, one the .scss, one the .ts) and re-running index_repository (mode full), the merged nodes' extension is corrupted:

src/app/add-absence/add-absence.component.ts   ""      <- was ".ts" without history
src/app/badge/badge.component.scss             ""      <- was ".scss"
src/app/checkbox/checkbox.component.html       .html   <- unaffected (template path won the merge)

Stable across repeated re-indexes. Only merged component nodes where a non-template sibling won the merge are affected. On the large real-world repo the same class of nodes instead shows extension: '.html' with a .ts/.scss file_path — so the value is order/environment-dependent: three different outcomes for the same logical situation.

Repro repo updated with the exact steps: https://github.com/artaommahe/codebase-memory-mcp-angular-repro (README, "Follow-up" section, commit 5e39bf9).

Posted by Claude Code on behalf of @artaommahe

[DeusData]

Thanks for the Angular repro. Triage: high-priority parsing/search correctness under #592 and #396.

File identity must remain one File node per actual source/template/style file; sibling Angular component files should not collapse into one graph node. The search_code impact makes this more than a cosmetic graph issue, so the regression should assert separate File nodes and complete text-search coverage for .ts, .html, and .scss siblings.