lychee.toml

Rama main · réplica de lectura del repositorio original.

# Lychee link checker config — used by .github/workflows/docs-quality.yml
exclude_path = [
  ".claude/worktrees",
  "node_modules",
  ".git",
  "packages/docs/dist",
  "packages/docs/.astro",
]
exclude = [
  # Asciinema casts get uploaded post-merge; skip until the ID exists
  "asciinema\\.org/a/<id>",
  # localhost references in code examples
  "localhost",
  "127\\.0\\.0\\.1",
  # GitHub Releases anchors that are page-internal
  "github\\.com/.*/releases/tag",
  # release-please CHANGELOG compare links — the higher tag is only created
  # by the release job AFTER the release PR merges, so the link 404s while
  # the PR is open. Resolves automatically post-merge.
  "github\\.com/nimbus-agent/Nimbus/compare/v",
  # External hosts that block scripted user-agents or rate-limit aggressively
  # from GitHub-hosted runner egress. Verified live in browsers; the failures
  # are infrastructure noise, not broken links.
  "npmjs\\.com",           # 403 to non-browser UAs
  "canva\\.com",           # 403 to non-browser UAs
  "opengraph\\.xyz",       # aggressive 429 rate limit
  "tauri\\.app",           # intermittent connection failures from runners
  "conventionalcommits\\.org",  # intermittent TCP RST from GH runner egress
  "asciinema\\.org/?$",    # homepage times out from runner egress; verified live (200)
  "asciinema\\.org/a/",    # cast pages also time out under load from runner egress; verified live (200)
  "nimbus-agent\\.dev",    # exclude until PR 1 DNS propagates
  "glama\\.ai",            # 525/403 to scripted UAs due to Cloudflare
  # 403 from GitHub runner egress only. Checked rather than assumed: from a dev
  # machine it returns 200 to BOTH a browser UA and curl's default UA, so this is
  # origin/IP-based blocking, not the scripted-UA kind the entries above describe.
  # Pre-existing link (the Ramp connector CHANGELOG entry), unrelated to the PR
  # that had to exclude it.
  "ramp\\.com",
  # github-action-benchmark dashboard on the perf-data orphan branch — the branch
  # and dev/bench dir are created by the first post-merge push, so the link 404s
  # while the PR documenting it is open. Resolves automatically post-merge.
  "github\\.com/nimbus-agent/Nimbus/tree/perf-data",
  # open-vsx.org's signed-in settings pages (cited in docs/ci-secrets.md and
  # docs/credential-hygiene.md) answer 503 intermittently — measured 2026-09-17 from a
  # dev machine, not only runner egress: 3 of 6 samples 503 on `/user-settings/tokens`
  # while `/extension/...` pages returned 200 on all 6. lychee does not retry a
  # rejected status, so both passes failed in three runs across two PRs that day. Scoped to
  # `/user-settings/` so the public open-vsx.org links stay checked.
  "open-vsx\\.org/user-settings/",
]
# 202 Accepted is a 2xx success — figma.com's homepage returns it to scripted UAs,
# which lychee otherwise rejects (the long-standing intermittent figma.com flake).
accept = ["200", "202", "203", "204", "206", "301", "302", "304"]
max_redirects = 5
timeout = 30
# CORRECTION (supersedes the claim this comment used to make): `max_retries` and
# `retry_wait_time` DO NOT apply to a rejected status code. Measured against a local
# server that counts requests — a 504 and a 503 were each fetched exactly ONCE, while
# a dropped connection was fetched FOUR times (1 + 3 retries). lychee retries
# TRANSPORT errors only; a response that arrives and fails the `accept` list is final.
#
# So these two settings do nothing for the recurring github.com 504, which was the
# reason given for adding `retry_wait_time` in #1200. They are kept because they are
# still correct for genuine transport flakes (connection reset, TLS hiccup), which do
# occur from runner egress — but they are not the 504 mitigation.
#
# The 504 is handled where it can be: `scripts/check-links.sh` (used by both
# `audit:links` and the workflow step) caches successes with `--cache` and retries the
# run once, so the retry re-checks ONLY what failed — verified: 2 OK URLs fetched once
# across two passes, the 504 URL fetched twice. A genuinely dead link still fails both
# passes and the gate still goes red, which accepting 5xx as success would not
# preserve.
max_retries = 3
retry_wait_time = 2
# Per-HOST, not global: almost every link here is github.com, so the run was
# opening far more simultaneous connections to one host than it needs to. This
# caps that without throttling other hosts.
#
# It is not free. Measured locally over the real link set (1,313 links, 845
# unique), one run each: 111s before, 127s after — roughly +15%, some of which
# is network variance. That is judged worth it against a gate that has produced
# two false failures in a day, each costing a re-run and eroding trust in the
# job. Raise this number if the run time starts to matter more than the flake.
host_concurrency = 8