Docs navigation

Files & line blame

Browse the repo at HEAD with per-line AI attribution, and open the conversation behind any AI-written line.

The Files tab is the per-line view: the repository at the default branch’s HEAD, with an AI-attribution overlay on every file. It answers the question code review always asks — who wrote this line, and why?

Browsing

The left pane is the repo’s file tree at HEAD; select a file to open it with the attribution overlay. Every line is either attributed to a specific model and coding tool, or carries no attribution (human-written, or its traces were never pushed). Attribution is read from the pushed ledgers — deterministic, never heuristic.

The header badge shows exactly which branch and commit you’re looking at.

The conversation behind a line

Click any AI-attributed line to open its context:

  • Model and tool — what wrote the line.
  • The trace — the recorded editing session it came from.
  • The conversation — the prompt and session transcript behind the change, if it was pushed. This is the intent that never survives into the code itself: what was asked for, what constraints were given, what the agent decided.

If a team pushes attribution only (no conversations — see Data & privacy), the line still shows its model and tool; there’s just no transcript to open.

Requirements

  • Traces enabled on the project, with at least one push covering commits reachable from the default branch’s HEAD.
  • GitHub App access to the repository — the file contents are read at HEAD via the App. If the view says the codebase is unavailable, check the installation under Settings → Connected GitHub (Connecting GitHub).

Anyone with read access to the project can browse — reviewers, new teammates, auditors — with no local setup at all.

The same answers, elsewhere

  • Locally, without TracesHub: agent-trace blame src/app.py per file, or agent-trace viewer for a browser view of your local record (blame, viewer).
  • For your coding agents: agent-trace context src/app.py -l 10-50 fetches the prompt and transcript behind a line range from the terminal, and agent-trace rule add context-for-agents teaches Cursor / Claude Code / Codex to look it up on their own before editing (context, rule). After an agent-trace pull, this works against everything teammates pushed, too.