chore(repo): tolerate stdin EPIPE races

[7ttp]

TL;DR

fixes a flake where a fast exiting child races the parent's
stdin write and rejects with EPIPE instead of the descriptive exit-code error

ref:

• spotted in: https://github.com/supabase/cli/actions/runs/34261833658/job/102181439072
  & many more runs.....

[github-actions]

🤖 AI Review

Codex reported no findings. Of Claude’s six findings, one error-propagation concern and one test-coverage gap are confirmed; four speculative correctness/maintainability concerns are refuted by the subprocess protocols, downstream validation, and actual implementation differences.

Findings

Severity	Location	Category	Sources	Claim
🟡 MINOR	apps/cli/tests/helpers/cli.ts:397	error-handling	claude	Throwing inside the stdin error-event listener surfaces non-EPIPE failures as uncaught exceptions instead of associating them with the spawned command.
⚪ NIT	apps/cli/tests/helpers/cli.ts:396	test-coverage	claude	No deterministic regression test directly exercises the newly tolerated EPIPE behavior.

Refuted findings (kept for transparency, not posted as review comments)

• packages/config/scripts/semantic-release-path-filter.ts:110 (error-handling): Swallowing EPIPE could allow a partial hash list to be treated as complete, silently dropping release commits.
  Refuted: git diff-tree --stdin reads hashes until EOF and cannot successfully exit early with an arbitrary partial list. An EPIPE caused by git rejecting the invocation produces a nonzero exit, which lines 119-122 surface; an exit of zero means git completed its stdin protocol.
• packages/config/scripts/build.ts:74 (error-handling): A swallowed EPIPE followed by a zero formatter exit could silently write an empty or truncated generated schema.
  Refuted: The fixed oxfmt process must successfully consume and format stdin to exit zero. More importantly, the same build subsequently requires both schema files as JSON at lines 237-240, so empty or malformed output fails the build rather than silently shipping. Trusted package documentation also states that dist is gitignored and rebuilt, contrary to the claim that this corrupts a committed artifact.
• apps/cli/tests/helpers/cli.ts:398 (error-handling): The helper should also swallow ECONNRESET, ERR_STREAM_DESTROYED, ERR_STREAM_WRITE_AFTER_END, and similar codes for the same fast-exit race.
  Refuted: For a child-process stdin pipe whose reader exits during this immediate write, Node reports EPIPE. ERR_STREAM_WRITE_AFTER_END is impossible in this single-write-before-end control flow, ERR_STREAM_DESTROYED applies to later writes to an already destroyed stream, and ECONNRESET describes a different transport failure. Swallowing those codes would hide distinct errors.
• packages/config/scripts/build.ts:77 (maintainability): The same five-line EPIPE guard is duplicated three times without explanatory comments.
  Refuted: The implementations are not identical: the config scripts catch rejected Bun FileSink operations and narrow unknown errors with instanceof Error, whereas the CLI helper handles Node EventEmitter stream errors. Only two short predicates are duplicated within the same package, and the trusted conventions do not require extracting such a two-use predicate.

Stats

Claude findings: 6 · Codex findings: 0 · Confirmed: 2 · Refuted: 4 · Uncertain: 0

---

Models: claude-opus-5 + gpt-5.6-sol · Trigger: auto · Workflow run

This review runs once per PR. A maintainer can request another with a /ai-review comment.