Repository navigation
Description
Summary
Since 836c366 ("prune nested repositories from the parent source walk"), the source walker skips every non-root directory that contains a .git entry. That fixed a real problem (linked worktrees inside a checkout being indexed many times over), but it also drops git submodules from the parent project's index, with no setting to bring them back.
I'd like an opt-in way to index submodules as part of the parent, while still skipping linked worktrees and unrelated nested clones.
Use case
A "super repo" that pulls in the main application as a submodule:
super-repo/ <- where I start every agent session
.gitmodules
dev/ <- scripts that launch and wrap the app
docs/
app/ <- git submodule, ~1,000 tracked files
I always start agent sessions from super-repo/. Today:
- The
super-repoindex covers only its own files; nothing underapp/is indexed, and nothing reports that it was skipped. - To search the app I have to
tracedecay initit as a separate project and passproject_selectoron every call, or usemulti_root_*scope sets. - Even then there are no call or import edges between the two, so a path from
dev/intoapp/never appears as one chain incallers/impact/call_chain.
For this layout the submodule is effectively part of the project, and one index is what I want.
Telling submodules apart from worktrees
With current git, both a linked worktree and a checked-out submodule have a .git file, not a directory, so the presence check can't distinguish them. Some signals that do:
- The
gitdir:target. A linked worktree points to<common-dir>/worktrees/<name>; a submodule points to<parent>/.git/modules/<name>. - The parent's index. A submodule is a gitlink entry (mode
160000) at that path, and is usually listed in.gitmodules. - Linked worktrees also show up in
git worktree listfor the parent.
Possible shape
Something like an index setting (name up to you):
[index]
include_submodules = true # default false keeps today's behaviourWhen enabled, the walker would descend into a directory only if it is a registered submodule of the root (gitlink plus .gitmodules entry), and keep pruning linked worktrees and other nested repositories exactly as now.
Related
- Because grep, ast-grep, and unmounted-files share this walker, they also skip submodule contents today.
- It would help if
statusorunmounted_filesreported "skipped nested repository: app/", so users know why that code is missing.
Environment
- tracedecay 1.0.0-beta.91, macOS, Claude Code plugin install
- git 2.54
Activity
fix(index): withdraw incomplete submodule source admission
ScriptedAlchemy commentedon Nov 8, 2026
Implementation investigation (2026-11-08 UTC), Refs #4593 — issue remains open; no fix is claimed.
A walker-only
include_submodulesswitch cannot satisfy this issue without silently disagreeing with the indexed graph. I checkedmasterat91c72aba7750a398dd28e83d64bf0c6d6fef4172, rootAGENTS.md(no nested AGENTS entries in the complete tree), the issue (zero comments before this investigation), both currently open PRs, and prior closed/unmerged #4719 (which explicitly withdrew a walker-only implementation).The additional integration points that must change together are:
crates/tracedecay-code-index/src/source_walk.rs:167-176excludes every nonroot.gitdirectory; live grep, ast-grep, and unmounted-file auditing reuse this.crates/tracedecay-code-index-runtime/src/code_index_scheduler/classification.rsbuilds candidates from the parent Git index and explicitly selects.index_worktree_submodules(None). The parent indexes the gitlink path, not the child source files.crates/tracedecay-code-index-runtime/src/code_index_scheduler/git_tree_capture.rsexplicitly skipsentry.mode.is_commit(), so the clean-HEAD/exact-tree fast path drops gitlinks.reconcile.rs:4011+uses that fast path to publish clean generations; merely changing the filesystem walker cannot affect the index.freshness_witness.rs::source_candidatesderives its roster from the parent classification, andidentity.rs::GitMetadataFingerprintV1samples only the parent Git metadata. A changed child revision or same-size content rewrite could otherwise remain falsely fresh.An independent disposable native Git probe executed five nonvacuous assertions: the parent records
160000 ... appbut notapp/feature.rs; an initialized child has.git -> ../.git/modules/app; a linked worktree has a different.git -> .../worktrees/...target; and the child file really exists despite absence from the parent index. This is Git/filesystem behavior, not repository Rust/Bazel, CI, or native macOS validation.Acceptance needs a single explicit, restart-applied opt-in from the canonical project configuration; Git-index gitlink plus
.gitmodulesmembership and bound child repository/worktree identity; consistent nested-source admission, sanitized capture, sealed graph/retrieval, and both clean and dirty snapshot authorities; freshness invalidation on child changes and removals; fail-closed linked-worktree/untrusted clone/metadata substitution handling. Tests must exercise default-deny, opt-in parent→child call edges and grep/AST parity, commit/dirty change, restart/reuse, removal, denied/nested clone, and macOS/native behavior. No secondary walker-only authority.No production changes or PR were published: the environment providing this investigation has no repository checkout or Bazel/native macOS toolchain, so it cannot legitimately compile, run nonzero repository tests, or establish a safe completed implementation. The issue-specific owner remains active; this comment is a concrete implementation boundary, not a closure recommendation.
fix(code-index): report skipped nested repositories in unmounted audit
fix(code-index): bound repository audits and report walk failures
style(code-index): format nested repository regressions
fix(code-index): align nested checkout capture with source walks
style(code-index): format nested repository boundary tests
fix(code-index): invalidate cached rosters on nested Git markers
test(code-index): assert roster invalidation across marker removal
fix(code-index): refuse markerless parent gitlink directories
style(code-index): satisfy Bazel formatting for gitlink guard
fix(code-index): deny ignored admission beneath Git gitlinks
fix(code-index): preserve HEAD gitlink boundary on staged removal
fix(code-index): deny staged-removal gitlink in source admission
style(code-index): fix ignored gitlink regression formatting
7 remaining items
fix(code-index): require explicit sealed submodule policy
perf(code-index): restrict scoped Git boundary traversal
style(code-index): apply Bazel rustfmt to scoped Git boundaries
fix(code-index): revalidate parent submodule authority during reads
style(code-index): apply Bazel rustfmt to submodule authority test
fix(code-index): recheck child repository worktree binding
fix(code-index): revalidate bound submodules during capture
style(code-index): apply Bazel rustfmt to child binding tests
fix(submodules): report parent Git failures during source reads
perf(code-index): reuse stable submodule binding witnesses
fix(code-index): witness split-index Git authority
fix(code-index): watch parent worktree config in child bindings
test(code-index): emit actual parent worktree Git configuration
ScriptedAlchemy commentedon Nov 16, 2026
Closing as completed. #4801 added the submodule opt-in, and #4905 adds tests that pin its boundaries.
97caca9f0e53be3649019a469bbb165a1e55ebce.f8878bf7e8f17287e6e00929cf5c61da90e42632(test(code-index): pin submodule opt-in boundaries #4905).How each request is met:
index.include_submodules.v1defaults tofalseand takes effect after a daemon restart. With the default, submodule contents stay out of the index, as before..gitmodulesregisters it. Linked worktrees, unrelated nested clones, deeper gitlinks, and uninitialized children stay excluded. test(code-index): pin submodule opt-in boundaries #4905 adds a test that creates a linked worktree inside the parent with the opt-in on. The worktree stays pruned andapp/first.rsis still walked.parent.rsand nothing underapp/, so the default check can't pass on an empty result. With the opt-in on, each findsapp/first.rs. The MCP grep tool searches the resolved project's root with that project's own path policy.unmounted_filesadds a "Skipped nested repositories" note listing the skipped paths (EcosystemAudit.note). The issue asked for this instatusorunmounted_files. It is inunmounted_files.Unit suites run locally on the #4905 change: code-index 415 passed, configuration 41 passed, code-index-runtime 746 passed with 2 ignored. The three new tests passed at
bec33fbc. The merged headb8212e8differs from that commit only in rustfmt line wrapping, and all required checks passed on it.This issue set no benchmark or remote performance criterion. None was added, and none is claimed here.