A cron job opened its first PR: one helper, 11 copy-pastes gone

Rasputin is a one-person project, so I have a standing interest in code review that happens while I sleep. As of July 20th, 2026 the control-plane repo runs a scheduled AI quality sweep: a Claude agent wakes on a Monday cron, hunts for duplicated and dead code, tries to refute its own findings, and opens a PR if anything survives. The first real PR merged the same day — it consolidated a SQLite open sequence that had been copy-pasted byte-for-byte across all 11 store packages, and deleted three dead symbols. Getting there took four runs and two failure modes worth writing down.

Timeline

• 2026-07-19 — researched the 2025–2026 agent-quality tooling before building anything (24 sources, top claims adversarially verified). The consistent line across every credible source: agents find, fix, and open PRs; a human merges. Nobody has published a guardrail stack they trust for auto-merge. That constraint shaped the whole design.
• 2026-07-19 — quality-sweep.yml lands: anthropics/claude-code-action@v1 in automation mode, Mondays 10:00 UTC. The prompt is find → verify → fix → gate → ship: refute each candidate before touching it, pass the exact checks ci.yml runs (gofmt, go vet, go test -race, go build, UI typecheck + build), push a branch, open a PR it cannot merge.
• 2026-07-20, run 1 — green in 2m40s, no PR, no visible reasoning. The action hides the agent transcript by default, so a no-findings run and a silent failure look identical. show_full_output + display_report fixed that.
• 2026-07-20, run 2 — skipped by the action itself. It refuses to run a workflow file that differs from main’s copy — token-exfiltration protection, and it means every change to the sweep needs a merge before it can be tested.
• 2026-07-20, run 3 — failed at “Reached maximum number of turns (50)”, 8.5 minutes in, four edits made, nothing pushed.
• 2026-07-20, run 4 — 96 turns, 16 minutes, PR #16 opened with CI green. Merged the same day.

The starvation bug

Run 3’s transcript is the instructive one. The agent wanted staticcheck for dead-code analysis — the right call. The Bash allowlist permitted go:*, so go install honnef.co/go/tools/cmd/staticcheck succeeded; executing ~/go/bin/staticcheck matched no allowed prefix, so every attempt bounced. It burned roughly 15 of its 50 turns on GOBIN tricks, symlinks, and find / -name staticcheck before discovering the loophole — go run honnef.co/go/tools/cmd/staticcheck starts with go. By then the budget was gone.

The fix was three lines of YAML: preinstall staticcheck@2025.1.1 in a setup step and put it on PATH, add Bash(staticcheck:*) to the allowlist, and raise --max-turns to 150 — sized from the measured run, with the 60-minute job timeout as the hard stop. The prompt now states the toolbox up front: these tools exist, install nothing else.

What the first PR found

• The open sequence — ?_pragma=journal_mode(WAL)&_pragma=busy_timeout(5000)&_pragma=foreign_keys(on), SetMaxOpenConns(1), then the schema exec — pasted identically in 11 store packages. Now a single dbutil.Open helper, with each package’s error-message prefix preserved exactly. Changing one pragma used to mean editing 11 files.
• It left alerts/store.go alone after confirming its DSN diverges on purpose — the refute-your-own-finding step doing its job.
• Three staticcheck U1000 symbols removed, each with grep evidence in the PR body: an unused msPtr, a function-local type cell, an orphaned errNoGrubenvState.
• +66/−115 overall, every gate green before it pushed.

Cost: the workflow authenticates with an OAuth token from claude setup-token, so runs draw from the same Claude Max subscription I already use for daily development instead of metered per-token API credits — the two billing paths are easy to conflate, and Anthropic’s plan-usage article explains the split better than I can keep current here. 96 turns and 16 minutes on a weekly cadence rounds to nothing against the plan’s limits. The guardrails stay: ~400-line diff cap, no dependency or workflow changes, no touching the /healthz install contract, and the merge button stays human.

Takeaway: the agent was never the hard part. The harness was — transcript visibility, tool access, a measured turn budget, and gates the run cannot skip.

Rasputin is an open-source (AGPL) homelab cluster system — a control plane, node OS, and firewall image that make a few Raspberry Pis or N100 boxes behave like one appliance. It’s pre-alpha, on GitHub, and I’m looking for a handful of design partners to run it and tell me what’s broken.