$ agitrack ┌ claude · session-1 ──────────────────────────────────┐ > add a --version flag to the cli · editing agitrack/cli.py … · added --version using importlib.metadata └──────────────────────────────────────────────────────┘ ✓ Created <aGiTrack> commit 4b2f31a. aGiTrack · Ctrl-G $ git log --oneline -2 4b2f31a <aGiTrack> add a --version flag to the cli a91c07e Initial commit $ git show 4b2f31a --stat <aGiTrack> add a --version flag to the cli agitrack/cli.py | 6 ++++++ trace + model + token cost in the body ↓
Agent code without provenance is unreviewable.
A coding agent edits twelve files while you get coffee. A week later nobody can say which prompt produced which change, what the model was told, or what it cost. aGiTrack makes the unit of agent work a commit, so review, blame, bisect, and revert work on agent output the same way they work on yours.
<aGiTrack> add a --version flag to the cli # Interaction Trace ## User add a --version flag to the cli ## Agent Added --version using the package metadata; prints and exits. # aGiTrack Metadata commit_type: agent backend: claude model: claude-opus-4-8 session_name: session-1 context_tokens: 18204 tokens_since_last_commit_input: 411 tokens_since_last_commit_output: 2390
Your branch only ever moves by merge.
Each session runs in its own git worktree on its own branch, so concurrent sessions never stomp each other or your working tree. Finished turns integrate back into the branch you launched from; conflicts are handed to the agent to resolve, and the resolution is itself a traceable <aGiTrack-merge> commit.
Isolated sessions
One worktree per session, per-turn branches, automatic recovery of anything a crash leaves behind. Run several agents side by side.
Backend-native
No model API keys, no proxy of your tokens. aGiTrack drives the real Claude Code / OpenCode TUI and reads the session's own transcript for the facts.
Costs on the record
Input, output, and cache token counts per commit, straight from the backend's session record: a price tag on every change.
A live dashboard of what the agents did.
Run agit -d and aGiTrack serves a live, auto-refreshing dashboard on localhost and opens it in your browser. Every number is computed from commit metadata alone, so it's identical on every clone — no database, no tracking server.
Honest provenance
Lines aGiTrack tracked as the agent's, versus everything else — it never claims a human wrote what the model did, even when a person made the commit.
Filter live
Narrow the whole dashboard to one committer (merged to their GitHub ID), a backend, a model, or a time range — recomputed instantly in the browser.
Cost & drill-down
Token totals and an efficiency ratio, a heads-up when a conversation loops, and a commit log you can click to read the full message and jump to the commit on GitHub.
agitrack -d
agitrack -d text
Three commands to the first traced commit.
pip install agitrack
cd your-repo && agitrack
Ctrl-G
Prompt the agent as usual. When the turn settles, the commit is already there. git log tells the rest.
Put your agents on the record.
Apache-2.0. Issues, ideas, and first contributions welcome. The issue tracker has good first issue tags waiting.