fast mode silently excludes `scripts/` — first-class Python source dropped from graph (built-in skip list?)

[Limebird01]

Version

codebase-memory-mcp 0.9.0

Platform

Linux (x64)

Install channel

GitHub release archive / install.sh / install.ps1

Binary variant

standard (portable static build, linux-amd64)

What happened, and what did you expect?

index_repository --mode fast silently excluded the entire scripts/ directory of a mid-size Python repo. Only 46 of 156 .py files were indexed, and scripts appeared in the excluded.dirs list of the index output. Re-indexing the same repo with --mode full captured everything (5,286 nodes / 17,328 edges, 6.0 s, ~232 MB max RSS).

No .gitignore or .cbmignore pattern matches scripts/. In this repo (as in many Python projects) scripts/ holds first-class production entry-point modules — it is a common Python layout, not a bin of throwaway files.

Expected: fast mode should reduce analysis depth (cheaper resolution strategies, fewer passes), not silently drop first-class source directories from the graph — or at minimum it should emit a visible warning that a top-level source dir was skipped by a built-in heuristic.

This looks like the same silent-exclusion family as #1184 (built-in skip list drops deploy/deployment(s) dirs, mode full, not overridable via .cbmignore) and #411 (closed; moderate mode silently drops subtrees), and #1219 (default assets exclusion swallows API route files) — but none of those covers scripts/ in fast mode specifically. Note the docs state the built-in skip list is not user-overridable, so there is currently no workaround other than switching to --mode full.

Reproduction

1. Code: any generic Python repo with a top-level scripts/ directory containing real importable modules, e.g.:

   repo/
     src/pkg/core.py
     scripts/deploy_tool/main.py   # first-class entry points, plain .py modules
     scripts/reporting/report.py
   

2. Commands:

   codebase-memory-mcp cli index_repository --repo-path /path/to/repo --mode fast --persistence false
   codebase-memory-mcp cli index_repository --repo-path /path/to/repo --mode full --persistence false
   

3. Result: fast-mode output lists scripts under excluded.dirs; indexed file count drops to 46/156 .py; every symbol under scripts/ is absent from the graph. Full mode indexes all 156 files. Expected: identical file coverage in both modes (only analysis depth should differ), or an explicit warning naming the skipped directory.

Logs

Diagnostics trajectory (memory / performance / leak issues)

Project scale (if relevant)

5,286 nodes / 17,328 edges / 156 .py files (full mode)

Confirmations

• I searched existing issues and this is not a duplicate.
• My reproduction uses shareable code (a dummy snippet or a public OSS repository), not proprietary code.

[AmirF194]

Reproduced on current main (d6be58ef, 2026-07-31), clean container. Minimal repo: src/pkg/core.py plus scripts/deploy_tool/main.py and scripts/reporting/report.py (each one importable function). fast mode:

{"project":"r1406fast","excluded":{"dirs":["scripts"],"count":1,...},"nodes":19,"edges":17,...}

full mode on the same tree: nodes":28,"edges":49 with nothing excluded, i.e. the two scripts/ functions and their edges only show up in full.

Root cause: src/discover/discover.c, FAST_SKIP_DIRS[] (around line 59) hardcodes "scripts" alongside "docs", "examples", "samples", "fixtures", etc, and cbm_should_skip_dir() applies that list whenever mode != CBM_MODE_FULL. There's no separate signal for "looks like CI/deploy tooling" vs "looks like a source package"; scripts/ is skipped by name alone. History-wise, the whole list (including this entry) came in as one bulk add in 94a9a92f ("Add C-native pipeline..."), with nothing that specifically calls out scripts/ versus its neighbors, so this reads as an oversight in the original list rather than an intentional exclusion.

Competing-PR sweep on src/discover/discover.c (87/87 open PRs, uncapped): 6 candidates touch the file. Two touch FAST_SKIP_DIRS itself: #1166 only adds an unrelated ALWAYS_SKIP_DIRS entry (.codebase-memory) and #1245 (a large incremental-indexing rewrite) reorders the same list but keeps "scripts" in it verbatim. Neither removes or gates it. #1203 is unrelated (shebang-based extensionless-script detection).