[kixelated]

Problem

The host land gate ejects a whole batch when any module fails, yet just ci keeps every lane running to the end. On 2026-10-31, moq's failed batches ran 20-50 more minutes after their first failure: check quest in 3 s, check rs in about 40 s, test rs in 226 s. No moq PR landed between 11:27 and 16:30 while 60+ waited.

Approach

• CHECK_FAIL_FAST=1 is an opt-in that the land gate sets.
  • A lane returns at its first failed module.
  • The dispatcher cancels the running lanes, using the existing cancel, which records each one as (cancelled).
  • It starts no more lanes.
  • The failed module is still named, and the seconds-per-module report still prints.
• Without the variable, just ci and just check still run every lane to the end, as the dispatcher intends: one unavailable check shouldn't hide unrelated failures.
• Host side: kixelated/nixos's moq check command exports CHECK_FAIL_FAST=1. That lands in kixelated/nixos#16 once this merges.

Impact

• No public API changes.

Alternatives

• The host watches just ci's output and kills the run at the first ci: … failed line. Rejected as a workaround for something the dispatcher should own.
• Dropping the batching gate for per-PR CI and auto-merge. Rejected by the maintainer, because PRs that pass alone but break together would break main.

Follow-ups

• This is the planned quest/m0/ci/ci-fail-fast.md from #7456, which is still queued. That quest is deleted once both land.

Checks

• python3 sh/dispatch-query.test.py: the two new tests pass, and both fail without the change. The serial one fails outright; the parked one hangs until its timeout. The existing failure and continuation tests still pass.
• just check passes.
• just test passes.

Review

quest review --base origin/main (Codex), round 1 on head c367838: no findings.

The maintainer approved landing this past the queue with the admin bypass, since the queue it unblocks is the one it would wait in.

🤖 Generated with Claude Code

(Written by Claude Opus 5.5)

[pullfrog]

✅ No new issues found.

Reviewed changes

• Opt-in fail-fast in the dispatcher — CHECK_FAIL_FAST=1 makes run_lane return at its first failed phase and the main loop set stopping, cancel the running lanes via the existing cancel, and start no further lanes; unset, every lane still runs to the end.
• Cancelled lanes stay distinct from failures — the [[ -z "$stopping" ]] guard keeps a cancelled lane's nonzero status out of failed, so only the first failure is named while cancelled phases show as (cancelled).
• Tests — a parallel test proves parked rs/cpp lanes are cancelled and later lanes never start, and a serial test proves a first failure stops the run with exactly one failure line.
• Comment updates — justfile and the sh/dispatch.sh header document the new host-gate variable.

Verified locally: the full sh/dispatch-query.test.py (31 tests) passes, shellcheck sh/dispatch.sh is clean, and both new tests fail without the change (the parallel one hangs to its 60s timeout; the serial one reports two failures instead of one).

One environment note, not a code change: the two new tests need zizmor on PATH because gh is a repository-wide module selected on every diff and CI_STRICT=1 turns its absence into an error before any lane runs. That tool is present in CI's dev shell, so it only affects bare checkouts.

  ^{｜ View workflow run ｜ Using deepseek-v4.1-flash (free via Pullfrog for OSS) ｜ 𝕏}