Claude Code Provenance: Track Every Agentic Edit

Guide

Claude Code edits files autonomously — full tasks, multi-file refactors, terminal runs. LineageLens captures the complete chain: prompt, model version, tool use, and every insertion, via a transparent local proxy. Here is how.

Agentic edits are invisible to git

When Claude Code completes a task, it may touch a dozen files through dozens of tool calls. Git records one commit by one human. The prompt that drove the change, the model that produced it, and the intermediate steps are gone the moment the session ends.

Proxy capture: the full request chain

LineageLens Lite runs a transparent proxy at localhost:8788. Point Claude Code at it (one environment variable) and every request is recorded: the full prompt, the model version, tool-use chains, and the code each response inserted. The proxy setup guide covers configuration.

Records feed the self-hosted dashboard with per-insertion risk scoring — license similarity, secret leakage, confidence — and export to AI-BOM for audit.

Blame, search, and gates

  • lineagelens blame — per-line model + prompt attribution in the terminal.
  • Semantic search — which changes came from prompts mentioning payments? (pgvector, Plus).
  • CI gate — GitHub Action annotates PRs containing unreviewed high-risk AI code (Plus).

Start in five minutes

  1. Deploy Lite: one Docker command on a $5 VPS.
  2. Open the setup wizard, point Claude Code at the proxy.
  3. Watch the live feed as Claude Code works — every edit attributed.