Repo Revival Agent

repo-revival-agent

Autonomous agent that classifies dead Python repos and either opens a modernization PR, files a polite "use this instead" issue, or **refuses to do anything** — whichever the evidence supports.

The most useful thing this agent did during the build was refuse to open a PR. Not because it crashed — because it ran the test suite, compared signatures against the baseline, and correctly concluded its changes would make things worse.

Live artifact

**[rholder/retrying#101](https://github.com/rholder/retrying/issues/101)** — agent opened this issue on a real abandoned repo, pointing to `tenacity` as the maintained successor.

Demo


Agent classifies `rholder/retrying` as `let_rest`, generates a polite issue suggesting `tenacity` as the maintained successor, and stops in dry-run mode.

What it does

**Scan.** Collects GitHub metadata: stars, days inactive, issue counts, license, CI, dependencies, README excerpt. No clone needed at this stage.

**Classify.** A Claude Opus tool-use agent searches GitHub for modern alternatives and emits one of: `revive` (deps need bumping, repo is otherwise fine), `fork` (worth porting, open a modernization PR), `let_rest` (superseded — point users to the successor). Runs three times; only emits a verdict if all three agree. Otherwise: `uncertain`. The agent refuses to act when it doesn't know.

**Act.** `let_rest` opens a polite GitHub issue. `revive` clones the repo into a bot-owned fork, runs the dependency bumper and codemods, runs the test suite in an isolated venv before *and* after changes, compares failure signatures between the two runs, and only opens a PR if no regressions were introduced. All actions run under a separate bot account (`@repo-revival-agent`), not under a personal identity.

What it explicitly won't do

• Open a PR if tests regress against the baseline. Comparison is by signature (file_path::ExceptionType for collection errors, full nodeid for test failu