[Idea] History-Based Indexing — giving VecGrep storytelling ability

[iamvirul]

This discussion is based on issue #16 opened by @Kavirubc. Moving here for broader community input before we decide on an implementation approach.

---

Idea summary

Give VecGrep the ability to answer temporal questions about code — not just what the code looks like now, but how it evolved and why it changed.

What problem does this solve?

Right now VecGrep is a point-in-time tool. You can ask "how does authentication work?" but not:

• "What was the auth module like before the OAuth refactor?"
• "Why did the developer change this function here?"
• "What did this file look like 3 commits ago?"

This limits the depth of understanding Claude can provide when helping with debugging, code reviews, or onboarding into an unfamiliar codebase.

Proposed solution

Two approaches worth discussing:

Option A — Lightweight git tool (low cost, quick win)
Add a get_file_history(path) MCP tool that wraps git log and git blame and returns structured output. No schema changes, no extra storage. Claude gets the storytelling ability for git repos immediately.

get_file_history("/path/to/project/src/auth.py")
# → last 10 commits touching this file, with messages and authors
# → git blame output mapping line ranges to commits

Option B — Full historical chunk indexing (high cost, high capability)
Store versioned chunk snapshots with commit metadata: commit_hash, commit_message, author, committed_at. Allows semantic search across historical versions — e.g. "show me the auth module before the performance refactor."

Trade-off: index size multiplies by number of significant commits. A repo with 100 commits and 10k chunks could mean up to 1M historical entries.

Alternatives considered

• Using git log -p / git blame directly via Claude shell access — works today but requires Claude to know the right commands and parse raw output
• Storing only the previous version of each chunk (1-deep history) rather than full commit history — much cheaper storage, covers the "what was here before" use case

Would you be willing to contribute this?

Open for community input — what level of history do you actually need day-to-day? Drop your thoughts below.

---

cc @Kavirubc — would love your thoughts on which option is worth pursuing, given you raised the original idea in #16.