The merge queue's gates, by failing test
Every gate the merge queue runs — a batch, a search prefix, a suspect alone —
is kept for 7 days in gate_result (src/checks/gate_result.ts): its pulls,
green, red or cut_off, the reason (tests, typecheck, install, …),
how many tests passed and failed, up to 20 failing tests by name with their
file and the first 5 lines of their error, and how long it took.
The queue's log line names the tests:
gated 8 pulls: red (2 failed: "a > b", "c > d")
Read them:
curl -sH "authorization: Bearer " \
"/api/v1/repos/<owner>/<name>/queue/gates?limit=20"
{ items, next_cursor }, newest first; limit is 1–100 (default 20), and
cursor is the last item's id. An agent asks the MCP tool queue_gates
with { repo, limit? } and gets the same gates.
Each failing test carries flaky: true when the same test (file and name)
failed in two red gates of the week that share no pull — no single pull's
code explains both.
The counts come from the gate's own output (src/checks/gate_tests.ts):
passed sums each bun test process's N pass; failed counts distinct
(fail) names, so the shards' tally — which repeats every failure as
(fail) name [1ms] [shard 3] — no longer counts a test twice.