AI coding agents write files faster than anyone can remember where they came from. Git records versions; DVC and OpenLineage record pipelines you declared up front; nothing records the unplanned output — the CSV an agent wrote mid-task, the PNG it regenerated, the notebook that silently touched a data file.
trace-file-lineage (MIT, ~361★, v0.7.0) is a local CLI that answers "which script, notebook, data, command, or AI agent produced this file?" with five confidence labels — verified, strong-candidate, candidate, weak-signal, insufficient — and it labels guesses as guesses. I have not tested or run this tool; everything below is from reading the README and source files.
Two modes, one honest split. Retrospective: for files you already have, it reconstructs the most likely origins from code, document metadata, and Git history — "ranked guesses with the reasoning attached" (README). Prospective: lineage run --task "..." -- python sweep.py wraps a command, captures the task boundary, and files changed during the run are marked verified — "These answers are proof" (README). It never moves, renames, or deletes files; lineage layout --suggest only proposes a destination.
What verified means in source. In evidence.py, every fact carries a basis (observation / declaration / inference / confirmation), an assurance field (verified when exact, candidate otherwise), the source path and line it came from, and a deterministic id (uuid5(NAMESPACE_URL, canonical_json)), so re-scanning the same fact yields the same id. Evidence about the evidence, with location attached — that design is the part worth copying.
Trust order and privacy. The README ranks sources: your own confirmation → recorded runs → imported provenance → declarations → static code → content → names/timestamps. Extracted text lives in .file-lineage/ (SQLite-backed, git-ignored by default). Privacy claims: nothing is uploaded, no account or API key; scanning never executes your code — only lineage run runs the explicit command after --, and password-looking arguments are stripped from recorded commands.
Deliberate limits. JS/TS get "a cautious static scan, not real language understanding" (README); other languages are searched, not parsed; runtime-built paths cannot be resolved. Performance is claimed and reproducible, not something I ran: tests/benchmark.py on macOS / Python 3.14 reports 1,000 files → 0.5 s cold scan / 0.1 s warm; 10,000 files → 16.5 s / 1.1 s; individual queries in milliseconds.
Bottom line. If you run Python/notebook workflows and a pile of agent-generated files has become unanswerable trivia, this is worth evaluating — lineage enable even writes the tracing rules into CLAUDE.md and AGENTS.md (an instruction, not enforcement, per the README). If you have no agent-generated files, or cannot tolerate a git-ignored text index in your project, it is premature. A comparison with Git/DVC/OpenLineage is in docs/comparison.md.
Top comments (1)
Evidence labels are exactly the missing layer. Agents often treat every file as equally authoritative because the filesystem gives them paths, not confidence. A small provenance label - generated, user-authored, copied, stale, derived - can prevent a lot of bad decisions before the model even starts reasoning.