[RUM] npm shorthand package specs (github:owner/repo, bare owner/repo) bypass redactErrorText — a customer repo reached Error Tracking unredacted

[dawsontoth]

The gap

redactErrorText (src/integrations/datadog/redactErrorText.ts) recognises exactly two
shapes of repository reference:

• URL_TOKEN — anything matching scheme://…
• SCP_GIT_REMOTE — git@host:owner/repo

npm package specifiers do not have to be either. github:owner/repo,
gitlab:owner/repo, bitbucket:owner/repo and the bare owner/repo GitHub shorthand
are all valid npm install arguments, they are what Harper passes to git clone /
npm install on a component deploy, and none of them contains :// or a user@host:
prefix. They pass through the redactor untouched, carrying the customer's repository
owner and name into Error Tracking, which is the exact retention-and-search surface the
module's header says it exists to keep them off.

Probing the shipped function directly:

LEAKS   | Failed to clone package github:acme-corp/private-billing: fatal: Too many arguments.
LEAKS   | Failed to download package gitlab:acme-corp/private-billing: npm error
LEAKS   | Failed to download package bitbucket:acme-corp/private-billing: npm error
LEAKS   | Failed to download package acme-corp/private-billing: npm error
redacts | Failed to download package https://github.com/<redacted>: npm error
redacts | Failed to clone package git@github.com:<redacted>: fatal

Production evidence

A SSEOperationError on 2026-08-28 reached RUM with a real third-party GitHub
org and repository name intact, in the github:<owner>/<repo> form, in both
error.message and error.stack. The identifiers are omitted here deliberately — this
repo is public — and are in the Datadog event.

This is after the redactor shipped (a8576681, 2026-07-30) and after the #1609
deploy gap closed, and <redacted> is present in the currently-served entry chunk, so
it is not a deploy-lag artifact: the redactor ran on this message and had no rule that
matched. Sibling events in the same window were redacted
(Failed to clone package https://github.com/<redacted>).html: …), which is what makes
the shorthand form the specific hole rather than the redactor being off.

Volume is low — 23 SSEOperationError events across 3 days in the 30 days to
2026-09-02, of which one carried an unredacted shorthand reference. The severity is not
the volume; it is that the leak is silent and the data is retained.

Why it isn't a one-line regex

github:/gitlab:/bitbucket: are tractable — they are a closed set of npm-recognised
aliases with a <segment>/<segment> payload. The bare owner/repo form is not: no
regex can distinguish it from ordinary prose or a file path without context, and a
greedy rule would shred stack frames and Harper's own paths.

The tractable framing is to redact by position, not by shape — Harper composes these
messages from a known set of prefixes (Failed to clone package <spec>:,
Failed to download package <spec>:, Failed to install dependencies for <spec> using…),
so the token immediately following a recognised prefix can be replaced wholesale
regardless of its form. That also closes the bare-shorthand case, which no shape-based
rule reaches.

That is a design call rather than a mechanical fix, which is why this is an issue and
not a PR.

Related

• [RUM] Secrets operation failures publish Harper's composed error text to Error Tracking #1664 — Harper's composed error text reaching Error Tracking (secrets); same class,
  different message family.
• [RUM] error.causes[] bypasses every redaction layer in beforeSend #1665 — error.causes[] bypasses every redaction layer; independent gap in the same
  module's coverage.
• a8576681 — the commit that introduced redactErrorText and the private-repo drop rule.