Show HN: Jeffy Loop, an autonomous code improvement loop for Claude Code

I'll pull the GitHub repo and HN thread so the paragraphs stay grounded in real project details rather than inventing specs.Jeffy Loop, published as a Show HN by lenamonj at https://github.com/lenamonj/jeffy-loop, is an autonomous code-improvement loop built for Claude Code. The author reports running it against 17 open-source projects across 8 languages (Python, JavaScript, Java, C#, C++, Go, Rust, Ruby): 16 runs converged, and 3 fixes merged upstream (fasthttp parseUintBuf overflow, bat bidi sanitizer completeness, chalk ansi256 level downconversion). At posting, the HN thread had 1 point and 0 comments.

Mechanically, a run is `/jeffy [N] [focus...]` with a default budget of 10 iterations. The loop audits the project surface first, builds a backlog where every task has a runnable acceptance check, then executes one verified, checkpointed task per iteration behind a repo-level verify gate that reverts anything that breaks the tree. Convergence requires a clean fresh audit, an empty backlog, an adversarial evaluator countersignature from a fresh-context sub-agent, and a machine-checked Stop hook (`skills/jeffy/hooks/stop-hook.sh`) that re-runs the project's own verify command. State lives in PLAN.md, BACKLOG.md, and JOURNAL.md; checkpoints are local git commits only—nothing is pushed. Enhance mode reuses the same engine for capability work rather than defect hunting. The engine ships no language-specific analyzer; it uses each project's own tests and verify command. CI holds the engine to 125 behavioural checks on Linux, Windows, and macOS.

For engineers, the useful claim is not “an agent rewrote the code,” but that green suites often hide real defects and that a disciplined loop can surface them with red-green proof. Examples from the public receipts: quantstats (7,489 stars) yielded 29 findings behind 125 green tests and ended smaller after fixes; records hid four High data-loss bugs behind a green suite; fasthttp findings were on the v1.73.0 tag that users already ship; bat’s new `--sanitize` flag was inert when output was piped—the exact case sanitization exists for—and that bidi fix merged before first release. Builders get a concrete pattern: audit with evidence, one task per iteration, acceptance checks, checkpoint-and-revert, and refuse “done” until a second agent and a shell gate agree.

Market context is the growing stack of agent loops around coding tools. Jeffy Loop descends from Geoffrey Huntley’s Ralph technique (re-feed one prompt in a loop) and wraps that engine with method drawn from Anthropic’s Claude Code practices and Boris Cherny’s public workflow: runnable checks, small tasks, promote lessons into files, prefer short fresh-context runs over one long session. Against a raw prompt loop, Jeffy adds the verify gate, surface inventory, severity envelope, priced declines for out-of-budget Lows, and the adversarial evaluator. Receipts deliberately include null results (gson: two iterations, one declined Low, zero lines changed; RuboCop: full sweep, zero findings) and hard grinds (python-dotenv: eight runs and 73 iterations before a clean audit), which is stronger evidence than only cherry-picked wins.

Practical takeaway: install needs Claude Code and git; `./install.sh` (or install.ps1) copies skills to `~/.claude/skills` and registers the Stop hook. Prefer several short runs over one large N, restart the session between runs so context resets while state files carry work forward, and never allowlist push or force operations for unattended loops. What to watch next is upstream follow-through—open PRs remain on dayjs, jsoncpp, yfinance, python-dotenv, PyPortfolioOpt, and PapaParse’s four Highs—and whether third parties can reproduce the same convergence discipline on their own trees from the public evals under `evals/`.