[hey-research-lab]

Problem

The first scheduled Snapshot run (run 39215293192, 2026-10-19) failed after 26 minutes, 500 of 4,842 detail reads in, on one answer:

hey-data: HEY answered 520. — retry after 60s
Process completed with exit code 6.

520 is what Cloudflare, in front of heyresearch.xyz, answers when the origin errs. politeFetch retried only 429, 500, 502, 503 and 504, so a single transient edge error ended a run of thousands of requests. A thrown network error or timeout was not retried either. The weekly release has never been produced.

What changed

• src/http.ts: 520–524 join the retryable set and are retried like a 502 (the server's retry-after, else 30 s), within --max-retries. A network failure or timeout is retried the same way. An oversized body is still not retried. Every request hey-data makes is a GET, so a retry is safe.
• .github/workflows/snapshot.yml: the run passes --max-retries 3.
• test/http.test.ts: four new tests:
  • 520/522/524 are retried;
  • the last 520 is returned once the retries run out;
  • a network failure is retried, then thrown;
  • an oversized body is not retried.
• test/generate.test.ts: the three resume tests stop their runs with maxRetries: 0. They inject one network failure to produce a stopped run, and with the default single retry that failure would now be absorbed instead of stopping the run.
• README, --help text and CHANGELOG describe the new retry set.

Validation

• pnpm scan, pnpm lint, pnpm typecheck, pnpm build, prettier --check .: pass.
• pnpm test: 42 passed. Three of the new tests fail against main's src/http.ts and pass with this change. The fourth, the oversized-body test, guards behaviour that is already correct.
• A live run against heyresearch.xyz with the built CLI finished with exit 0 in 21 min 50 s: generate --no-details --evidence-limit 0 --max-retries 3 wrote 4,851 projects, 4,353 ships and 26,281 changes, with totals checks passing. That run met no transient error, so it shows the change breaks nothing; the retry paths themselves are covered by the unit tests.

Limitations

• More retries cannot outlast a longer outage. The run still stops with exit 6, and the existing resume_from_run input resumes it from the uploaded partial work.
• The full schedule (details on, about 3 hours) was not re-run here; the next Monday schedule, or a manual dispatch after merge, is the end-to-end check.