NeuralMind v4.15.0 — scan-for-secrets scans what you can commit, once: linked worktrees skipped, .gitignore honoured, each finding marked tracked or not

A 12-repository field audit (#635) found neuralmind scan-for-secrets hard to read in the repositories that need it most. Repositories that keep their worktrees inside the checkout (.worktrees/<id>/, .claude/worktrees/<id>/) got every finding once per worktree, which buried the real ones. And none of the HIGH findings across the 12 repositories was a committed secret: they were an untracked .env, a Firebase client config and an environment lookup with a localhost default. build already left those worktrees out (they were gitignored), so the scan and the index disagreed about what the project is.

What changed

• Linked worktrees are skipped. A directory whose .git is a file naming a git worktree add checkout is another copy of a repository, so the scan no longer walks it, whether or not .gitignore covers it. The report lists what it skipped. Submodules, which also have a .git file, and nested repositories are still scanned: they are separate code, and build indexes them.
• .gitignore is honoured by default. Inside a git repository the scan covers what build indexes, from the same git ls-files listing: tracked files and untracked files that .gitignore doesn’t exclude. Ignored files can’t be committed and aren’t indexed. --include-ignored scans them too, for a full sweep. Outside a repository the scan reads everything, as before.
• A clean scan says what it left out. The text output ends with the worktrees and the number of gitignored files it skipped, with the flag that scans them, so a skipped .env is never a silent all-clear.
• Each finding says whether git tracks its file. [tracked] or [untracked] on each line, "tracked": true | false in --json (null outside a repository), and the summary counts high-confidence findings in tracked files. A credential in a tracked file is already in history, or one commit from it: rotate that one first.

Exit codes are unchanged: any high-confidence finding exits 1, so a CI gate behaves as before on what it scans. The legacy .neuralmind/last_output.json cache is still scanned explicitly, and .neuralmindignore still applies only with --use-neuralmindignore.

What the agent actually sees

An agent that runs the scan, or reads its output from a hook or CI log, gets each finding once, labelled:

NeuralMind secret scan — /home/dev/myproject

  [HIGH ] src/config.py:8  aws-access-key-id  (AKIA…(20 chars))  [tracked]
  [HIGH ] scratch.py:1  anthropic-api-key  (sk-a…(108 chars))  [untracked]

  2 high-confidence (1 in tracked files), 0 heuristic.

  ...
  A credential in a tracked file is in git history: rotate it first.

  Skipped 2 linked worktrees (.claude/worktrees/a1, .worktrees/b2): other checkouts of a repository; run the scan inside one to check it.
  Skipped 1 file that .gitignore excludes (never committed or indexed). --include-ignored scans them too.

--json adds tracked to each finding and gitignore_respected, ignored_files_skipped and skipped_worktrees to the report.

Per-agent expectations

Upgrading

• A gitignored file with a credential is no longer reported by default. A CI step that relied on the scan failing on an ignored, uncommitted file (a .env written by an earlier step, say) should add --include-ignored. The report still counts the ignored files it skipped.
• A script that reads --json sees the new fields; none were removed.
• No index change, no rebuild.

Reproduce

pytest tests/test_secret_scan_scope.py tests/test_secret_scan.py -v

tests/test_secret_scan_scope.py drives real git: two linked worktrees inside the checkout are skipped and listed, a submodule and a nested repository are still scanned, a gitignored .env is skipped and counted (and found with --include-ignored), and each finding’s tracked matches what git tracks.

Related

• #635 — the issue this release fixes
• #640 — the 12-repository field audit it came from
• CLI reference: scan-for-secrets