@feigi/fleet-ctl 3.23.36 → 3.23.37

This diff represents the content of publicly available package versions that have been released to one of the supported registries. The information contained in this diff is provided for informational purposes only and reflects changes between package versions as they appear in their respective public registries.
Files changed (204) hide show
  1. package/package.json +1 -1
  2. package/scripts/main-checkout.mjs +1 -1
  3. package/scripts/member-write-guard.mjs +1 -1
  4. package/scripts/repo-root.mjs +1 -1
  5. package/skills/run-team/references/reaping.md +1 -1
  6. package/scripts/agent-test-dir-prose.test.mjs +0 -59
  7. package/scripts/agent-test.test.mjs +0 -1640
  8. package/scripts/ambient-git-vars-mjs-prose.test.mjs +0 -397
  9. package/scripts/ambient-git-vars-prose.test.mjs +0 -192
  10. package/scripts/ancestry-check-position-prose.test.mjs +0 -175
  11. package/scripts/apply-ruleset.test.mjs +0 -421
  12. package/scripts/arg-header-probes-prose.test.mjs +0 -394
  13. package/scripts/arg.test.mjs +0 -1354
  14. package/scripts/board-cli.test.mjs +0 -710
  15. package/scripts/board-identity.test.mjs +0 -429
  16. package/scripts/board-liveness-banner.test.mjs +0 -119
  17. package/scripts/board-prev-shape.test.mjs +0 -228
  18. package/scripts/board-tryparse-null.test.mjs +0 -119
  19. package/scripts/board.test.mjs +0 -3121
  20. package/scripts/brief-after-cut-prose.test.mjs +0 -53
  21. package/scripts/candidates-exit3-prose.test.mjs +0 -302
  22. package/scripts/candidates.test.mjs +0 -2312
  23. package/scripts/caveat-unification-prose.test.mjs +0 -98
  24. package/scripts/cell-readout.test.mjs +0 -615
  25. package/scripts/ci-comment-rot-prose.test.mjs +0 -2025
  26. package/scripts/ci-completes-premise-prose.test.mjs +0 -257
  27. package/scripts/ci-state-prose.test.mjs +0 -113
  28. package/scripts/ci-state.test.mjs +0 -1933
  29. package/scripts/ci-vacuous-green.test.mjs +0 -368
  30. package/scripts/citation-sweep-prose.test.mjs +0 -1712
  31. package/scripts/claim-lifecycle.test.mjs +0 -187
  32. package/scripts/claim-ticket.test.mjs +0 -1122
  33. package/scripts/closing-keyword-prose.test.mjs +0 -193
  34. package/scripts/cockpit-launch-idempotent-prose.test.mjs +0 -165
  35. package/scripts/cockpit-tick-nonblocking.test.mjs +0 -210
  36. package/scripts/compute-board.test.mjs +0 -1214
  37. package/scripts/compute-spend.test.mjs +0 -517
  38. package/scripts/consumer-technology-prose.test.mjs +0 -126
  39. package/scripts/cross-repo-citation-prose.test.mjs +0 -155
  40. package/scripts/cwd-isolation-pins.mjs +0 -83
  41. package/scripts/dedupe-guard-prose.test.mjs +0 -243
  42. package/scripts/delete-merged-branch.test.mjs +0 -576
  43. package/scripts/derive-testcmd.test.mjs +0 -595
  44. package/scripts/diff-stats.test.mjs +0 -404
  45. package/scripts/dispatch-block-golden-prose.test.mjs +0 -1130
  46. package/scripts/dispatch-block-pins-prose.test.mjs +0 -615
  47. package/scripts/dispositions-check.test.mjs +0 -936
  48. package/scripts/dispositions-gate-prose.test.mjs +0 -129
  49. package/scripts/drop-merged-label.test.mjs +0 -479
  50. package/scripts/exec-stub.mjs +0 -93
  51. package/scripts/exec-stub.test.mjs +0 -69
  52. package/scripts/expected-jobs-prose.test.mjs +0 -80
  53. package/scripts/fan-out-authorization-prose.test.mjs +0 -80
  54. package/scripts/filing-bar-seat-prose.test.mjs +0 -242
  55. package/scripts/filing-claim-bar-prose.test.mjs +0 -166
  56. package/scripts/finisher-died-prose.test.mjs +0 -31
  57. package/scripts/finisher-dispatch-premise-prose.test.mjs +0 -312
  58. package/scripts/finisher-halt-prose.test.mjs +0 -106
  59. package/scripts/finisher-name-prose.test.mjs +0 -134
  60. package/scripts/finisher-own-tree-prose.test.mjs +0 -299
  61. package/scripts/finisher-pin-race-prose.test.mjs +0 -339
  62. package/scripts/finisher-test-verdict-prose.test.mjs +0 -102
  63. package/scripts/fix-applier-correction-rules-prose.test.mjs +0 -202
  64. package/scripts/fix-applier-result-file-prose.test.mjs +0 -149
  65. package/scripts/fleet-bootstrap.test.mjs +0 -75
  66. package/scripts/fleet-heartbeat.test.mjs +0 -450
  67. package/scripts/fleet-provenance.test.mjs +0 -138
  68. package/scripts/fleet-run.test.mjs +0 -90
  69. package/scripts/fleet-state.test.mjs +0 -460
  70. package/scripts/fleet-tick-prose.test.mjs +0 -300
  71. package/scripts/fleet-tick.test.mjs +0 -2444
  72. package/scripts/force-with-lease-reason-prose.test.mjs +0 -184
  73. package/scripts/frontmatter-check.test.mjs +0 -436
  74. package/scripts/git-env.test.mjs +0 -149
  75. package/scripts/git-status-untracked-mode-sweep.test.mjs +0 -216
  76. package/scripts/gone-walk.test.mjs +0 -213
  77. package/scripts/gp-sep-invariant-prose.test.mjs +0 -188
  78. package/scripts/head-moved-after-label-prose.test.mjs +0 -233
  79. package/scripts/immutable-body-claim-prose.test.mjs +0 -204
  80. package/scripts/implementer-model-tier.test.mjs +0 -235
  81. package/scripts/inflight-citation-prose.test.mjs +0 -144
  82. package/scripts/inflight-exit2-prose.test.mjs +0 -81
  83. package/scripts/inflight-probe-guard-prose.test.mjs +0 -109
  84. package/scripts/inflight-scan-scope-prose.test.mjs +0 -57
  85. package/scripts/inflight.test.mjs +0 -2882
  86. package/scripts/injection-control-prose.test.mjs +0 -404
  87. package/scripts/install-root-audit.test.mjs +0 -464
  88. package/scripts/instrument-check-prose.test.mjs +0 -235
  89. package/scripts/instruments.test.mjs +0 -787
  90. package/scripts/is-cli.test.mjs +0 -99
  91. package/scripts/issue-tracker-prose.test.mjs +0 -71
  92. package/scripts/json.test.mjs +0 -364
  93. package/scripts/ledger-dispatch.test.mjs +0 -927
  94. package/scripts/ledger-grammar.test.mjs +0 -227
  95. package/scripts/ledger-read-require-file.test.mjs +0 -197
  96. package/scripts/ledger-rotate.test.mjs +0 -142
  97. package/scripts/ledger.test.mjs +0 -3436
  98. package/scripts/line-distance-prose.test.mjs +0 -245
  99. package/scripts/liveness-rationale-prose.test.mjs +0 -83
  100. package/scripts/locale-pin-prose.test.mjs +0 -189
  101. package/scripts/main-checkout.test.mjs +0 -332
  102. package/scripts/member-outcome-prose.test.mjs +0 -167
  103. package/scripts/member-outcomes-header.test.mjs +0 -129
  104. package/scripts/member-outcomes-prose.test.mjs +0 -63
  105. package/scripts/member-outcomes.test.mjs +0 -638
  106. package/scripts/member-prompt-prose.test.mjs +0 -165
  107. package/scripts/member-record.test.mjs +0 -541
  108. package/scripts/member-write-guard.test.mjs +0 -214
  109. package/scripts/merge-commit-probe-prose.test.mjs +0 -122
  110. package/scripts/merge-gate.test.mjs +0 -613
  111. package/scripts/muted-git-guard-sweep.test.mjs +0 -493
  112. package/scripts/net-ssh-precedence-prose.test.mjs +0 -91
  113. package/scripts/net-watchdog-prose.test.mjs +0 -79
  114. package/scripts/net.test.mjs +0 -211
  115. package/scripts/no-ci-gate-prose.test.mjs +0 -108
  116. package/scripts/no-undo-audit.test.mjs +0 -3961
  117. package/scripts/no-wave.test.mjs +0 -138
  118. package/scripts/node-floor-sweep.test.mjs +0 -611
  119. package/scripts/pointer-target-prose.test.mjs +0 -186
  120. package/scripts/pr-cost.test.mjs +0 -521
  121. package/scripts/pr-overlap.test.mjs +0 -692
  122. package/scripts/printf-die-sweep.test.mjs +0 -384
  123. package/scripts/prompt-renderer.mjs +0 -40
  124. package/scripts/prompt-renderer.test.mjs +0 -68
  125. package/scripts/prose-pin.mjs +0 -807
  126. package/scripts/prose-pin.test.mjs +0 -1034
  127. package/scripts/prove-merge.test.mjs +0 -995
  128. package/scripts/pull-table-relabel-prose.test.mjs +0 -60
  129. package/scripts/quiet-payload-prose.test.mjs +0 -222
  130. package/scripts/reap-cherry-check-prose.test.mjs +0 -80
  131. package/scripts/reap.test.mjs +0 -4432
  132. package/scripts/reaping-prose.test.mjs +0 -239
  133. package/scripts/reaping-reap-authorizers-prose.test.mjs +0 -112
  134. package/scripts/reaping-release-authorizers-prose.test.mjs +0 -85
  135. package/scripts/recipe-derivation-prose.test.mjs +0 -112
  136. package/scripts/recipe-prove.test.mjs +0 -1042
  137. package/scripts/refuter-scratch-prose.test.mjs +0 -135
  138. package/scripts/release-label-required-prose.test.mjs +0 -71
  139. package/scripts/release-label-split-prose.test.mjs +0 -184
  140. package/scripts/release-ticket-reap-pairing-prose.test.mjs +0 -63
  141. package/scripts/release-ticket.test.mjs +0 -5099
  142. package/scripts/renovate-release-contract.test.mjs +0 -379
  143. package/scripts/repo-root.test.mjs +0 -873
  144. package/scripts/reporter-output-rule-prose.test.mjs +0 -208
  145. package/scripts/review-agent-report-only.test.mjs +0 -49
  146. package/scripts/review-core-booking.test.mjs +0 -35
  147. package/scripts/review-core-cwd-isolation.test.mjs +0 -374
  148. package/scripts/review-core-reads.test.mjs +0 -478
  149. package/scripts/review-core-refuter-scratch.test.mjs +0 -148
  150. package/scripts/review-core-scratch-refusal.test.mjs +0 -64
  151. package/scripts/review-core-shared-test-run.test.mjs +0 -275
  152. package/scripts/review-core-snapshot-path.test.mjs +0 -830
  153. package/scripts/review-core-specialist-scratch.test.mjs +0 -191
  154. package/scripts/review-core-testcmd.test.mjs +0 -335
  155. package/scripts/review-core-unrun.test.mjs +0 -566
  156. package/scripts/review-core-unverified-discriminant.test.mjs +0 -208
  157. package/scripts/review-eval-resolver-path.test.mjs +0 -55
  158. package/scripts/review-host-fixture.mjs +0 -89
  159. package/scripts/review-in-run-retry.test.mjs +0 -211
  160. package/scripts/review-no-vendor-dispatch.test.mjs +0 -45
  161. package/scripts/review-path-default.test.mjs +0 -796
  162. package/scripts/review-runner-spawns.test.mjs +0 -95
  163. package/scripts/review-runner.test.mjs +0 -166
  164. package/scripts/review-tier-audit.test.mjs +0 -140
  165. package/scripts/run-artifact-pr-prose.test.mjs +0 -341
  166. package/scripts/run-merge-bot-prose.test.mjs +0 -1413
  167. package/scripts/run-team-local-hook.test.mjs +0 -442
  168. package/scripts/scratch-partition-invariant-prose.test.mjs +0 -74
  169. package/scripts/scripts-path-citation-sweep.test.mjs +0 -73
  170. package/scripts/section-between-guard-prose.test.mjs +0 -379
  171. package/scripts/select-dimensions.test.mjs +0 -674
  172. package/scripts/shared-refusal.test.mjs +0 -388
  173. package/scripts/shell-traps-prose.test.mjs +0 -504
  174. package/scripts/shortlist.test.mjs +0 -380
  175. package/scripts/slow-transport.mjs +0 -72
  176. package/scripts/snapshot-repo.test.mjs +0 -1006
  177. package/scripts/snapshot-runner-audience-prose.test.mjs +0 -130
  178. package/scripts/spend-view.test.mjs +0 -312
  179. package/scripts/staleness-qualifier-prose.test.mjs +0 -251
  180. package/scripts/staleness-verdict-prose.test.mjs +0 -142
  181. package/scripts/staleness.test.mjs +0 -893
  182. package/scripts/stalled-member-liveness-prose.test.mjs +0 -219
  183. package/scripts/stash-prohibition-prose.test.mjs +0 -313
  184. package/scripts/strip-comments.mjs +0 -44
  185. package/scripts/temp-dir.mjs +0 -47
  186. package/scripts/ticket-router.test.mjs +0 -981
  187. package/scripts/tier-check.test.mjs +0 -828
  188. package/scripts/tier-outcomes-header.test.mjs +0 -518
  189. package/scripts/tier-outcomes.test.mjs +0 -572
  190. package/scripts/tier-roles.test.mjs +0 -535
  191. package/scripts/tracker-block-copy-prose.test.mjs +0 -288
  192. package/scripts/triage-labels-gloss-prose.test.mjs +0 -97
  193. package/scripts/unattended-git-sweep.test.mjs +0 -142
  194. package/scripts/verify-sha-exit2-prose.test.mjs +0 -269
  195. package/scripts/verify-sha.test.mjs +0 -692
  196. package/scripts/watcher-degraded-block.test.mjs +0 -369
  197. package/scripts/watcher-key-prose.test.mjs +0 -106
  198. package/scripts/within-run-pair-prose.test.mjs +0 -212
  199. package/scripts/worktree-audit.test.mjs +0 -1408
  200. package/scripts/worktree-head-prose.test.mjs +0 -276
  201. package/scripts/worktree-listing-sweep.test.mjs +0 -217
  202. package/scripts/worktree-model-prose.test.mjs +0 -216
  203. package/scripts/worktree-porcelain.mjs +0 -35
  204. package/scripts/worktree-porcelain.test.mjs +0 -102
@@ -1,2312 +0,0 @@
1
- // Regression gate for the candidate query's two filters — to-spec specs never
2
- // reach the queue, and the list comes back oldest-first — and for its exit
3
- // codes, which are the whole interface for a caller that reads no stderr: 2 is
4
- // "the query broke", 1 is "the queue is empty", and the two must never swap.
5
- //
6
- // candidates.mjs applies its reduction through gh's `--jq`, so the stub below
7
- // runs the same expression gh would have received — under the system jq it
8
- // execs by default, not the gojq gh embeds and applies in its own process.
9
- // Both predicates are under test that way; the ENGINE itself is under test
10
- // only where a real gojq binary is reachable (`findGojq()` below) — the
11
- // dependency scan since #331, the spec predicate since #204. `\s`, `\d` and
12
- // `\b` are Unicode-aware in jq's Oniguruma and ASCII-only in Go's RE2, so the
13
- // same body used to reduce differently depending on which engine ran it, and
14
- // gojq's answer is the one that reaches production: gh applies gojq, not jq,
15
- // so dropSpecs is the only line of defence against what gojq actually
16
- // decides. #383 closed that divergence by spelling an explicit class at every
17
- // position, each chosen from what GFM renders — so the class after a
18
- // heading's `#` marker is deliberately NARROWER than the class at the end of
19
- // that same heading's line. The expression now contains no `\s`, `\d` or
20
- // `\b`, and all 24 measured fixtures agree on both engines.
21
- //
22
- // That agreement is also what removed the old proof of which engine ran.
23
- // Until #383 the gated tests identified their engine by one fixture the two
24
- // answered differently — a trailing U+00A0 that made a heading a spec under
25
- // jq and not under gojq. Those rows are ruled now and agree, so the proof is
26
- // direct instead: the STUB records the `--version` of the binary it executed
27
- // and every gated test asserts that recording is gojq (`assertRanGojq`).
28
- // A pattern gojq rejects outright at least exits non-zero; a class that
29
- // merely matches differently would leave this suite green on jq alone, which
30
- // is what the gated gojq tests below exist to catch. Stubbing gh to return
31
- // already-reduced JSON would leave the jq expression — which is where both
32
- // predicates actually live — completely untested.
33
-
34
- import { test } from "node:test";
35
- import assert from "node:assert/strict";
36
- import { spawnSync } from "node:child_process";
37
- import { mkdtempSync, readFileSync, rmSync, writeFileSync } from "node:fs";
38
- import { tmpdir } from "node:os";
39
- import { join } from "node:path";
40
- import { fileURLToPath } from "node:url";
41
- import { stripComments } from "./strip-comments.mjs";
42
- import { between, phrase } from "./prose-pin.mjs";
43
- import { writeExecStub } from "./exec-stub.mjs";
44
-
45
- const SCRIPT = fileURLToPath(new URL("./candidates.mjs", import.meta.url));
46
- const ARG_MODULE = fileURLToPath(new URL("./arg.mjs", import.meta.url));
47
-
48
- const STUB = `#!/bin/sh
49
- # Stand-in for \`gh issue list … --jq <expr>\`. Applies the expression gh was
50
- # given to the fixture, so the expression is under test rather than assumed.
51
- #
52
- # ONE plumb point for which binary applies it, and it records itself. The
53
- # gated gojq tests below can no longer tell the engines apart by their
54
- # ANSWERS: #383 ruled every divergent position and all 24 fixtures now agree
55
- # on both engines, which is the point of the ticket and also what blinded the
56
- # old proof. So the engine names itself here, by the same --version test
57
- # isGojq applies, and those tests assert the recording. Resolved ONCE and
58
- # reused for the version line and both apply paths below, so the recording
59
- # can never describe a different binary from the one that reduced the
60
- # fixture. Delete this plumb and every gated test reds, because the recording
61
- # then reads jq- and not gojq.
62
- JQ_RAN="\${JQ_BIN:-jq}"
63
- [ -n "$ENGINE_LOG" ] && "$JQ_RAN" --version > "$ENGINE_LOG" 2>&1
64
- expr=""
65
- search=""
66
- fields=""
67
- while [ $# -gt 0 ]; do
68
- case "$1" in
69
- --jq) shift; expr="$1" ;;
70
- --search) shift; search="$1" ;;
71
- --json) shift; fields="$1" ;;
72
- esac
73
- shift
74
- done
75
- # The fallback is the query with no POSITIVE label: term. The leading space is
76
- # the whole test and cannot be dropped: every exclusion is spelled \`-label:\`,
77
- # so a bare \`*label:*\` matches the unfiltered query too and silently serves it
78
- # the labeled fixture — the fallback then looks untaken however well it works.
79
- # \$search is padded so the term still matches when it leads: gh ignores
80
- # qualifier order, so reordering query()'s template must not invert this.
81
- # No second fixture supplied — both queries see the same rows, as before.
82
- # The two fixtures are deliberately DISJOINT, which real gh could never be
83
- # (unfiltered is a superset). A faithful superset makes the stub's ignored
84
- # --limit stop the cap tests firing. So never assert a labeled row is ABSENT
85
- # from the fallback payload — it was never in that fixture, and it passes
86
- # whatever the code does.
87
- # gh applies \`--jq\` itself, inside its own process, so a reduction that did not
88
- # take is a real failure mode — an old gh, an expression it rejects, a proxy
89
- # answering with an error object. Overriding the expression is the only way a
90
- # test reaches the empty and non-array refusals: the real one always emits an
91
- # array. It is NOT the only way to reach the row-shape refusal, which a fixture
92
- # alone trips whenever a field's type is wrong upstream (\`"number":"11"\` reduces
93
- # to \`{"n":"11"}\`) — the expression fixes the key set, never the value types.
94
- [ -n "$JQ_OVERRIDE" ] && expr="$JQ_OVERRIDE"
95
- fixture="$FIXTURE"
96
- case " $search " in
97
- *\\ label:*)
98
- # GitHub ends an UNQUOTED qualifier value at the first space; every word
99
- # after the first becomes a free-text term instead. Measured against the
100
- # live repo 2026-08-17: \`label:ready-for-agent candidates\` returns
101
- # strictly fewer issues than \`label:ready-for-agent\`, while
102
- # \`label:"ready-for-agent"\` returns exactly as many — the relationship and
103
- # not the counts, which drift with the queue from hour to hour.
104
- # Reproduce that split or no test can tell the two forms apart — both
105
- # contain \` label:\`, so the dispatch above serves the labeled fixture
106
- # either way and #175 stays green while broken.
107
- #
108
- # Opt-in through \$LABEL_EXPECT: unset, the term is parsed and ignored, so
109
- # every test written before #175 keeps exactly its old dispatch.
110
- padded=" $search "
111
- term=\${padded#* label:}
112
- case "$term" in
113
- \\"*) label=\${term#\\"}; label=\${label%%\\"*} ;;
114
- *) label=\${term%% *} ;;
115
- esac
116
- ;;
117
- # No positive \`label:\` term at all, so nothing was parsed out. An empty
118
- # \$label matches no \$LABEL_EXPECT, which is what refuses the OTHER
119
- # direction of #175 — a query that dropped the qualifier outright rather
120
- # than misparsing it. Only a check BELOW the case can see that: such a
121
- # search never enters the arm above.
122
- *) label=""; [ -n "$FIXTURE_UNFILTERED" ] && fixture="$FIXTURE_UNFILTERED" ;;
123
- esac
124
- if [ -n "$LABEL_EXPECT" ] && [ "$label" != "$LABEL_EXPECT" ]; then
125
- # An empty array, not a missing one: the real query would have SUCCEEDED
126
- # and matched nothing, which is the whole confusion #175 is about.
127
- echo '[]' | "$JQ_RAN" -c "$expr"
128
- exit
129
- fi
130
- # gh answers \`--json\` with EXACTLY the fields it names, never the whole issue,
131
- # so the fixture is cut down to them before the expression sees it. Uncut, the
132
- # stub would serve fields the query never asked for: drop \`blockedBy\` from
133
- # candidates.mjs's \`--json\` and real gh sends no edge at all, while every
134
- # native-edge fixture below would still reduce to its blockers (#1741). Cut by
135
- # the same binary that applies the expression, so a gated test stays gojq end
136
- # to end.
137
- if [ -n "$fields" ]; then
138
- "$JQ_RAN" -c --arg fields "$fields" '($fields | split(",") | map({(.): true}) | add) as $keep | map(with_entries(select($keep[.key])))' "$fixture" > "$fixture.fields" || exit
139
- fixture="$fixture.fields"
140
- fi
141
- exec "$JQ_RAN" -c "$expr" "$fixture"
142
- `;
143
-
144
- function run(issues, args = ["--require-label", "ready-for-agent"], unfiltered = null, extraEnv = {}) {
145
- const dir = mkdtempSync(join(tmpdir(), "candidates-"));
146
- const fixture = join(dir, "issues.json");
147
- writeFileSync(fixture, JSON.stringify(issues));
148
- const gh = join(dir, "gh");
149
- writeExecStub(gh, STUB);
150
- const engineLog = join(dir, "engine-version");
151
- const env = { ...process.env, PATH: `${dir}:${process.env.PATH}`, FIXTURE: fixture, ENGINE_LOG: engineLog, ...extraEnv };
152
- if (unfiltered) {
153
- const second = join(dir, "unfiltered.json");
154
- writeFileSync(second, JSON.stringify(unfiltered));
155
- env.FIXTURE_UNFILTERED = second;
156
- }
157
- const r = spawnSync(process.execPath, [SCRIPT, ...args], { encoding: "utf8", env });
158
- // Read the engine recording BEFORE the temp dir goes. This is the only
159
- // evidence of which binary actually applied the expression (#383); it
160
- // stays "" when the STUB never ran (ENOENT), which is itself a failure the
161
- // gated tests report rather than skipping past. Only ENOENT is expected
162
- // here — any OTHER fs error (EACCES, EMFILE/ENFILE, ENOTDIR, EISDIR…)
163
- // means the STUB (and the engine it invoked) may well have run fine and
164
- // the log path itself is what's broken; swallowing it too would have
165
- // `assertRanGojq` misreport a real I/O failure as "never reached gojq".
166
- let engine = "";
167
- try {
168
- engine = readFileSync(engineLog, "utf8").trim();
169
- } catch (e) {
170
- if (e.code !== "ENOENT") throw e;
171
- }
172
- rmSync(dir, { recursive: true, force: true });
173
- return { ...r, rows: r.stdout.trim() ? JSON.parse(r.stdout) : [], engine };
174
- }
175
-
176
- // For the failure shapes the fixture stub cannot reach — it always `exec`s jq,
177
- // so it can only ever fail by exit status. A gh that kills itself, or one that
178
- // is absent, is a different `execFileSync` error object entirely.
179
- function runWithGh(script) {
180
- const dir = mkdtempSync(join(tmpdir(), "candidates-gh-"));
181
- const gh = join(dir, "gh");
182
- writeExecStub(gh, script);
183
- const env = { ...process.env, PATH: `${dir}:${process.env.PATH}` };
184
- const r = spawnSync(process.execPath, [SCRIPT, "--require-label", "ready-for-agent"], {
185
- encoding: "utf8",
186
- env,
187
- });
188
- rmSync(dir, { recursive: true, force: true });
189
- return r;
190
- }
191
-
192
- // Labels are settled by the fixture, not by the query: the stub ignores
193
- // `--search`'s label term beyond picking a fixture. A row in the unfiltered
194
- // fixture therefore has to carry a label that would NOT have matched the
195
- // labeled query, or the two fixtures contradict each other.
196
- const ticket = (n, body, labels = ["ready-for-agent"]) => ({
197
- number: n,
198
- title: `ticket ${n}`,
199
- labels: labels.map((name) => ({ name })),
200
- body,
201
- });
202
-
203
- // #1306: EXCLUDE is a single string literal, pinned whole rather than by N
204
- // independent `assert.match` calls for individual terms — a dropped or
205
- // reordered term sits in the gap BETWEEN two such matches and neither would
206
- // see it. One `assert.equal` over the exact captured span covers every term
207
- // and their order together.
208
- //
209
- // Read through stripComments(), not raw source — the same escape :1039's
210
- // `die()` pin already closes. An unanchored match against raw text is
211
- // satisfied by a declaration parked in a `/* */` block above a reverted one:
212
- // measured, with the correct five-term EXCLUDE moved into a block comment
213
- // and the live declaration reverted to the pre-#1306 five terms, a raw-source
214
- // version of this assertion still passed while the unfiltered scan leaked
215
- // all eight wayfinder tickets back in. stripComments() blanks whole-line and
216
- // block comments (its own documented ceiling), so the mutant reds under it
217
- // and the real declaration's capture is unaffected either way.
218
- test("EXCLUDE carries every canonical exclusion term, in order, as one pinned string (#1306)", () => {
219
- const src = stripComments(readFileSync(SCRIPT, "utf8"));
220
- const m = src.match(/const EXCLUDE =\n\s*"([^"]+)";/);
221
- assert.notEqual(m, null, "candidates.mjs's EXCLUDE declaration no longer matches this pin's shape — update this test");
222
- assert.equal(
223
- m[1],
224
- "-label:in-progress -label:onhold -label:wontfix -label:needs-triage -label:needs-info -label:wayfinder:map -label:wayfinder:research -label:wayfinder:prototype -label:wayfinder:grilling -label:wayfinder:task",
225
- );
226
- });
227
-
228
- test("a to-spec spec is dropped — it is to-tickets' input, not a claimable ticket", () => {
229
- const { rows } = run([
230
- ticket(10, "## Problem Statement\n\nx\n\n## User Stories\n\n1. As a user, I want…\n"),
231
- ticket(11, "## What to build\n\nAdd a --json flag.\n"),
232
- ]);
233
- assert.deepEqual(rows.map((r) => r.n), [11]);
234
- });
235
-
236
- test("every spec is dropped and named — a filter that stops at the first leaks the rest", () => {
237
- const { rows, stderr } = run([
238
- ticket(10, "## User Stories\n\n1. As a user, I want…\n"),
239
- ticket(11, "## Problem Statement\n\ny\n\n## User Stories\n\n2. As a user…\n"),
240
- ticket(12, "## What to build\n\nx\n"),
241
- ]);
242
- assert.deepEqual(rows.map((r) => r.n), [12]);
243
- // deepEqual on the parsed lines, not a substring match: `/dropped #10/` is
244
- // satisfied by a hardcoded number and by logging kept rows too. This pins
245
- // both that 10 and 11 are named and that 12 is not.
246
- assert.deepEqual(stderr.match(/dropped #\d+/g), ["dropped #10", "dropped #11"]);
247
- });
248
-
249
- test("a drop from the no-label run is tagged [unfiltered] — the one pass tag no other test reaches", () => {
250
- // Every other run() in this file passes --require-label, so the pass-1
251
- // ternary always takes its `label:` arm and the `unfiltered` arm is
252
- // unpinned: rename it to anything and the whole suite stays green. Both
253
- // documented callers do pass the flag (next-ticket/SKILL.md's `## 1. Candidates`
254
- // step, and shortlist.mjs, which run-team/SKILL.md's **Build the Shortlist** step runs), so this arm is reachable only by a hand-run — but
255
- // it is a supported invocation, and its tag is what tells a reader which
256
- // query dropped what. Two rows, not one: the spec supplies the drop line,
257
- // the ticket keeps the queue non-empty so this exits 0, not the empty-queue 1.
258
- const { status, stderr, rows } = run(
259
- [
260
- ticket(10, "## User Stories\n\n1. As a user…\n", ["ready-for-human"]),
261
- ticket(11, "## What to build\n\nx\n", ["ready-for-human"]),
262
- ],
263
- [],
264
- );
265
- assert.equal(status, 0);
266
- assert.deepEqual(rows.map((r) => r.n), [11]);
267
- // The closing `\]` is load-bearing: without it this is also satisfied by the
268
- // fallback pass's `[unfiltered (fallback)]`, the one tag already pinned above.
269
- assert.match(stderr, /dropped #10 — to-spec spec, not a ticket \(User Stories heading\) \[unfiltered\]/);
270
- });
271
- test("near misses are kept — the predicate's shape is specified, not accidental", () => {
272
- // One fixture per dimension the regex commits to. Without these, every
273
- // loosening of the heading match still passes: the dropped fixture differs
274
- // from a ticket in all of them at once, so it discriminates none.
275
- //
276
- // #65: nesting depth (`### User Stories`) moved OUT of this list — it is no
277
- // longer a near miss, see the depth test below — and the split-heading
278
- // fixture moved IN, replacing it as the dimension this list now pins.
279
- const { rows } = run([
280
- ticket(2, "## user stories\n\nlowercase\n"),
281
- ticket(3, "##User Stories\n\nno separating space\n"),
282
- ticket(4, "## User Stories (draft)\n\ntrailing text\n"),
283
- ticket(5, "Mentions ## User Stories mid-line, not a heading.\n"),
284
- // #65: the marker alone on its line, heading text starting the next —
285
- // `\s` spans the newline so this used to read as the same heading. Kept
286
- // is the fixed behaviour; `dropped #6` in a future run is the regression.
287
- ticket(6, "##\nUser Stories\n\nmarker and text on separate lines\n"),
288
- // #65: both ends of the depth range. `#` is an ordinary ticket's h1 and the
289
- // floor the regex's own name commits to — loosen `#{2,6}` to `#+` and this
290
- // is what goes red. `#######` is past CommonMark's six-`#` cap, so it is not
291
- // a heading at all and must not read as the signature; it pins the ceiling
292
- // the way #7 below pins that the range does not stop at three.
293
- ticket(10, "# User Stories\n\nan h1, not the signature depth\n"),
294
- ticket(11, "####### User Stories\n\nseven hashes, past CommonMark's cap\n"),
295
- ]);
296
- assert.deepEqual(rows.map((r) => r.n), [2, 3, 4, 5, 6, 10, 11]);
297
- });
298
-
299
- test("a spec heading is dropped at depth two through six, however separated and ended — #65", () => {
300
- // Why the predicate widened at all — see candidates.mjs's spec-predicate
301
- // comment. What this test adds is the shape of the range and its edges:
302
- // `####` pins that it does not stop at three (the near-miss list above pins
303
- // both ends), and #12/#9/#13 ride the same run to pin what may separate the
304
- // marker from the text and what may end the line.
305
- const { rows } = run([
306
- ticket(1, "### User Stories\n\nnested one level deeper\n"),
307
- ticket(7, "#### User Stories\n\nnested deeper still\n"),
308
- // #65: CRLF. `$` under `(?m)` matches BEFORE the `\n`, so the line still
309
- // ends in `\r` and a trailing class of `[ \t]*` cannot reach the anchor —
310
- // this body stops matching and the spec leaks, silently, with nothing on
311
- // stderr. GitHub's web textarea submits CRLF, so it is a real body shape,
312
- // and this is the only line of defence. Narrow the trailing `\s*` and
313
- // `dropped #12` is what goes missing.
314
- ticket(12, "## User Stories\r\n\r\n1. As a user…\r\n"),
315
- ticket(9, "## User Stories \n\n1. As a user…\n"),
316
- // #65: a TAB between marker and text. The class is `[ \t]`, but every other
317
- // fixture here uses a space, so narrowing it to `[ ]` passed the whole repo
318
- // suite green — the documented half was asserted in prose and pinned
319
- // nowhere.
320
- ticket(13, "##\tUser Stories\n\na tab between marker and text\n"),
321
- ticket(8, "## What to build\n\nx\n"),
322
- ]);
323
- // #8 is a real kept ticket, so this is a POSITIVE assertion: a query that
324
- // dies outright fails it too, which `deepEqual(…, [])` over a spec-only
325
- // fixture could not. Every other row above must be absent, and the failure
326
- // names whichever one leaked.
327
- assert.deepEqual(rows.map((r) => r.n), [8]);
328
- });
329
-
330
- // Dependency-scan tests (#58). The forms below are the ones to-tickets
331
- // actually publishes — see candidates.mjs's `depnums` comment — not the bare
332
- // inline phrasings the OLD regex covered alone.
333
-
334
- test("a Blocked-by heading with bulleted refs yields every number, not just the first", () => {
335
- const { rows } = run([ticket(9, "## Blocked by\n\n- #12\n- #13\n")]);
336
- assert.deepEqual(rows[0].d, [12, 13]);
337
- });
338
-
339
- test("None — can start immediately under a heading yields no dependencies", () => {
340
- const { rows } = run([ticket(9, "## Blocked by\n\n- None — can start immediately\n")]);
341
- assert.deepEqual(rows[0].d, []);
342
- });
343
-
344
- test("a reference in unrelated prose, not under a blocking declaration, is not a dependency", () => {
345
- const { rows } = run([ticket(9, "- A reference to #77 blocking ticket\n")]);
346
- assert.deepEqual(rows[0].d, []);
347
- });
348
-
349
- test("the bold-label form yields its number — to-tickets' local-file template writes this, not the issue template", () => {
350
- const { rows } = run([ticket(9, "**Blocked by:** #12\n")]);
351
- assert.deepEqual(rows[0].d, [12]);
352
- });
353
-
354
- test("the original bare inline phrasings still work, now reduced to numbers rather than matched phrase text", () => {
355
- const { rows } = run([
356
- ticket(1, "depends on #5\n"),
357
- ticket(2, "blocked by #99 inline form\n"),
358
- ticket(3, "requires #3\n"),
359
- ticket(4, "after #4\n"),
360
- ]);
361
- assert.deepEqual(rows.map((r) => r.d), [[5], [99], [3], [4]]);
362
- });
363
-
364
- test("several references on one inline label line are all found, not just the first", () => {
365
- const { rows } = run([ticket(9, "Blocked by: #12, #13\n")]);
366
- assert.deepEqual(rows[0].d, [12, 13]);
367
- });
368
-
369
- test("a chained ref separated by FORM FEED or a bare CR is still collected, not silently dropped (#383)", () => {
370
- // The inline label/ref separator class must match candidates.mjs's
371
- // end-of-heading-line class ([\t\f\r \p{Zs}]) for the same reasons: \f is
372
- // stripped by GFM and \r keeps CRLF bodies working. A narrower class here
373
- // than there SILENTLY DROPS the ref after the odd separator — the worse
374
- // direction, since a reader still sees `Blocked by #12,\f#13` as two
375
- // dependencies and GitHub still autolinks both.
376
- const { rows } = run([
377
- ticket(1, "Blocked by #12,\f#13\n"),
378
- ticket(2, "Blocked by #12,\r#13\n"),
379
- ]);
380
- assert.deepEqual(rows.map((r) => r.d), [[12, 13], [12, 13]]);
381
- });
382
-
383
- test("a heading section ends at the NEXT heading, blocking or not — a ref past it is not swept in", () => {
384
- const { rows } = run([ticket(9, "## Blocked by\n\n- #12\n\n## Notes\n\nsee #999 for context\n")]);
385
- assert.deepEqual(rows[0].d, [12]);
386
- });
387
-
388
- test("a bare marker-only heading (no title text) still closes a blocking section, on both LF and CRLF bodies (#383)", () => {
389
- // GFM still renders a bare `##` as a real (empty) `<h2>` — the toggle that
390
- // decides whether a line starts a NEW heading must see it too, or the
391
- // still-open `## Blocked by` section wrongly keeps collecting refs past
392
- // it. Over-collection, the direction #383's own gate ruling did not cover:
393
- // a bare `[ \t]` marker class matches only a TITLED heading, so this one
394
- // was invisible to the toggle and the section never closed.
395
- const { rows } = run([
396
- ticket(1, "## Blocked by\n\n- #12\n\n##\n\n- #13\n"),
397
- ticket(2, "## Blocked by\r\n\r\n- #12\r\n\r\n##\r\n\r\n- #13\r\n"),
398
- ]);
399
- assert.deepEqual(rows.map((r) => r.d), [[12], [12]]);
400
- });
401
-
402
- test("d holds plain issue numbers, not matched phrase strings", () => {
403
- const { rows } = run([ticket(9, "depends on #5\n")]);
404
- assert.deepEqual(rows[0].d, [5]);
405
- assert.equal(typeof rows[0].d[0], "number");
406
- });
407
-
408
- test("an open blocking section reads its list items only — a later prose ref is not a phantom blocker", () => {
409
- // The section flag closes on the NEXT heading, so a `## Blocked by` that is
410
- // the last heading in the body stays armed to EOF. Collecting every line
411
- // there turns "None — can start immediately" plus ordinary notes into a
412
- // blocker the body never declared, and the consumer (next-ticket step 2)
413
- // drops the ticket without re-reading the body.
414
- const { rows } = run([
415
- ticket(9, "## Blocked by\n\n- None — can start immediately\n\nImplementation notes: mirror what #300 did.\n"),
416
- ]);
417
- assert.deepEqual(rows[0].d, []);
418
- });
419
-
420
- test("a fenced code sample inside an open blocking section is not swept in", () => {
421
- const { rows } = run([ticket(9, "## Blocked by\n\n- #12\n\n```\ngit log #999\n```\n")]);
422
- assert.deepEqual(rows[0].d, [12]);
423
- });
424
-
425
- test("both passes are line-local — a phrase whose ref opens the next line is not collected", () => {
426
- // The removed regex's `\s+` crossed newlines and did collect these. The
427
- // narrowing is deliberate (see candidates.mjs's `depnums` comment); this
428
- // pins it so it stays a decision rather than drifting back by accident.
429
- const { rows } = run([
430
- ticket(1, "Blocked by:\n#12\n"),
431
- ticket(2, "This depends on\n#5 landing first\n"),
432
- ticket(3, "requires\n#7\n"),
433
- ]);
434
- assert.deepEqual(rows.map((r) => r.d), [[], [], []]);
435
- });
436
-
437
- test("'and' joins refs on an inline label line, the same as a comma", () => {
438
- const { rows } = run([ticket(9, "Blocked by: #12 and #13\n")]);
439
- assert.deepEqual(rows[0].d, [12, 13]);
440
- });
441
-
442
- test("every declarative phrase arms a heading section, not just 'Blocked by'", () => {
443
- // Dropping `requires` or `depends on` from the heading alternation left the
444
- // whole suite green: only `## Blocked by` was ever pinned.
445
- const { rows } = run([
446
- ticket(1, "## Requires\n\n- #8\n"),
447
- ticket(2, "## Depends on\n\n- #9\n"),
448
- ]);
449
- assert.deepEqual(rows.map((r) => r.d), [[8], [9]]);
450
- });
451
-
452
- test("an `after` HEADING arms nothing — it is narrative, not a declaration", () => {
453
- // `## After the migration` / `## Before` + `## After` are ordinary ticket
454
- // prose, unlike the three declarative phrases, which head a section only to
455
- // declare one. A heading that really does name a blocker still resolves,
456
- // through the inline pass — `after` stays in that alternation.
457
- const { rows } = run([
458
- ticket(1, "## After the migration\n\n- see #77 for background\n"),
459
- ticket(2, "## Before\n\n- old path\n\n## After\n\n- new path short-circuits, see #12\n"),
460
- ticket(3, "## After #4 lands\n\n- do the thing\n"),
461
- ]);
462
- assert.deepEqual(rows.map((r) => r.d), [[], [], [4]]);
463
- });
464
-
465
- test("the label's asterisks may close BEFORE the colon as well as after it", () => {
466
- // `**Blocked by:** #12` and `**Blocked by**: #12` are both bold labels a
467
- // human writes; only the first was covered, and the regex group carrying
468
- // the second could be deleted with the whole suite still green.
469
- const { rows } = run([ticket(9, "**Blocked by**: #12\n")]);
470
- assert.deepEqual(rows[0].d, [12]);
471
- });
472
-
473
- test("a bolded REF yields its number — whitespace and asterisks interleave freely between label and ref", () => {
474
- // #439: the asterisk groups used to sit ahead of the separating whitespace,
475
- // so only asterisks flush against the label matched. Markdown puts them on
476
- // the other side of the space, on the ref — which is what #208's brief
477
- // wrote, and the scan then read that whole section as no dependency at all.
478
- const { rows } = run([
479
- ticket(1, "Blocked by **#179** (decides gone shape)\n"),
480
- ticket(2, "blocked by **#178**\n"),
481
- ticket(3, "Depends on **#5**\n"),
482
- ticket(4, "Blocked by: **#12**\n"),
483
- ticket(5, "**Blocked by:** **#12**\n"),
484
- ticket(6, "Blocked by *#12*\n"),
485
- ticket(7, "**Blocked by** : #12\n"),
486
- ]);
487
- assert.deepEqual(rows.map((r) => r.d), [[179], [178], [5], [12], [12], [12], [12]]);
488
- });
489
-
490
- test("a bolded ref AFTER the first one still joins the run — asterisks interleave between refs too", () => {
491
- // #439's residual, found reviewing the fix: the separator between the label
492
- // and the first ref was widened, the continuation between refs was not, so
493
- // the second ref's leading `**` ended the run and the number was dropped —
494
- // no error, no warning, the same wrong admission #439 was filed for, one ref
495
- // to the right. Revert the continuation and every row here reds.
496
- const { rows } = run([
497
- ticket(1, "Blocked by **#12** and **#13**\n"),
498
- ticket(2, "Blocked by: **#12**, **#13**\n"),
499
- ticket(3, "Blocked by #12 and **#13**\n"),
500
- ticket(4, "Depends on **#5**, #6\n"),
501
- ]);
502
- assert.deepEqual(rows.map((r) => r.d), [[12, 13], [12, 13], [12, 13], [5, 6]]);
503
- });
504
-
505
- test("a bold split landing MID-PHRASE misses the label — ACCEPTED LIMITATION, pinned, not desired behavior", () => {
506
- // #1033. READ ROWS 1-4 AS A BOUNDARY, NOT AS A SPECIFICATION of what the
507
- // scan should do. They are shapes an author could plausibly write and the
508
- // scan does NOT read: the separator AROUND the label tolerates asterisks
509
- // freely (`**Blocked by:** #12`, `Blocked by **#12**` — the two tests
510
- // above), but the label PHRASE itself must still appear as one unbroken
511
- // literal, so emphasis closing BETWEEN its two words drops the ref silently
512
- // — exit 0, nothing on stderr, the ticket reading as claimable while its
513
- // blocker is open. Row 3 shows a colon does not rescue it and row 4 that the
514
- // run length does not matter: it is where the seam falls, not the emphasis,
515
- // that decides. Measured at 662004f on both engines, agreeing — `/usr/bin/jq`
516
- // 1.7.1-apple and `gojq` 0.12.19 both return [] for rows 1-4.
517
- //
518
- // Widening the alternation to tolerate emphasis INSIDE the phrase was
519
- // considered at triage and refused, for two reasons worth keeping here:
520
- // - Nothing produces this shape. to-tickets' templates bold the whole
521
- // label (`**Blocked by:**`) or the whole ref (`Blocked by **#12**`),
522
- // never split the two words — the form is derived from the regex, never
523
- // observed in a real body.
524
- // - A label alternation loosened for bold is the exact class of change
525
- // that reopened the phantom-blocker sweep before; "the phantom-blocker
526
- // sweep stays closed against the widened label and heading gates" below
527
- // is what catches that. Widening spends the expensive half of the trade
528
- // on a shape nobody writes.
529
- //
530
- // So, to a future widener: this fixture going red is the PROMPT, not a
531
- // regression. Before flipping rows 1-4 to expect numbers, re-measure the
532
- // phantom-blocker sweep on BOTH engines — system jq (Oniguruma) and gojq
533
- // (RE2, the one gh actually applies) — because a loosened label gate is
534
- // what reopens that sweep, and the two engines need not agree about it.
535
- //
536
- // Rows 5-7 are the controls, and are what keeps this from being a vacuous
537
- // pin: the seam forms — emphasis closing at the phrase boundary, or absent
538
- // altogether — still resolve. Break the scan outright and rows 1-4 stay
539
- // green while 5-7 red.
540
- const { rows } = run([
541
- ticket(1, "**Blocked** by #12\n"),
542
- ticket(2, "**Depends** on #5\n"),
543
- ticket(3, "**Blocked** by: #12\n"),
544
- ticket(4, "*Blocked* by #12\n"),
545
- ticket(5, "**Blocked by** #12\n"),
546
- ticket(6, "**Depends on** #5\n"),
547
- ticket(7, "Blocked by #12\n"),
548
- ]);
549
- assert.deepEqual(rows.map((r) => r.d), [[], [], [], [], [12], [5], [12]]);
550
- });
551
-
552
- test("a noun-form Dependencies heading arms a section, at any heading depth and with a trailing colon", () => {
553
- // #439's second, independent miss: the gate named the verb forms only, so
554
- // the heading #208's brief actually used opened nothing, and the list-item
555
- // branch never engaged either. Both routes into that section failed.
556
- const { rows } = run([
557
- ticket(1, "## Dependencies\n\n- #12\n- #13\n"),
558
- ticket(2, "## Dependency\n\n- #12\n"),
559
- ticket(3, "### Dependencies:\n\n- #12\n"),
560
- ticket(4, "###### Dependencies\n\n- #12\n"),
561
- ]);
562
- assert.deepEqual(rows.map((r) => r.d), [[12, 13], [12], [12], [12]]);
563
- });
564
-
565
- test("a bolded dependency heading arms its section — emphasis is not a different heading", () => {
566
- // #1031: the gate admitted no emphasis at all, so `## **Dependencies**` armed
567
- // NO section and every ref its bullets declared was dropped — exit 0, nothing
568
- // on stderr, the same silent wrong admission #439 was filed for — reached
569
- // through the heading gate rather than through the inline label separator
570
- // #439 widened. Both emphasis runs are load-bearing: markdown
571
- // closes the bold either before the colon (`**Dependencies**:`) or after it
572
- // (`**Dependencies:**`), and the run is `*`, `**` or `***` depending on
573
- // whether the author wanted italic, bold or both. Substitute the
574
- // `\*{0,2}` pair the ticket hypothesised — one leading run, one ahead of the
575
- // colon and none after it — and rows 5 and 8 red, on both engines.
576
- const { rows } = run([
577
- ticket(1, "## **Dependencies**\n\n- #12\n- #13\n"),
578
- ticket(2, "## **Blocked by**\n\n- #12\n"),
579
- ticket(3, "## **Depends on**\n\n- #9\n"),
580
- ticket(4, "## **Requires**\n\n- #8\n"),
581
- ticket(5, "## **Dependencies:**\n\n- #12\n"),
582
- ticket(6, "## **Dependencies**:\n\n- #12\n"),
583
- ticket(7, "## *Dependencies*\n\n- #12\n"),
584
- ticket(8, "## ***Dependencies***\n\n- #12\n"),
585
- ticket(9, "###### **Dependency**\n\n- #12\n"),
586
- ]);
587
- assert.deepEqual(
588
- rows.map((r) => r.d),
589
- [[12, 13], [12], [9], [8], [12], [12], [12], [12], [12]],
590
- );
591
- });
592
-
593
- test("the noun form arms only when it is the WHOLE heading — bolded or not, a Dependency-injection section is prose", () => {
594
- // The verb forms end at `\b` and tolerate trailing text, because a heading
595
- // opening with `Blocked by` declares one whatever follows it. The noun form
596
- // cannot afford that: `## Dependency injection` is an ordinary section title
597
- // in a code repo, and arming on it turns every `#N` in its bullets into a
598
- // blocker the body never declared, starving the ticket out of the queue —
599
- // the same over-fire the list-item restriction exists to prevent. Anchoring
600
- // the noun alternative to end-of-line is what keeps these three closed;
601
- // widen it to `\b` like its neighbours and this test is what reds.
602
- //
603
- // The bolded halves are #1031's other side: emphasis tolerance is added
604
- // AROUND that anchor, never by relaxing it, so the run of asterisks must not
605
- // smuggle in the prefix match the anchor exists to refuse. The runs match
606
- // ASTERISKS only, which is what keeps ` injection` outside them — spell
607
- // either one as a general wildcard (`.*`) and all six rows arm, silently
608
- // inventing #300 as a blocker on every DI section in the queue.
609
- const { rows } = run([
610
- ticket(1, "## Dependency injection\n\n- rework the container, see #300\n"),
611
- ticket(2, "## Dependencies (blocking)\n\n- #300\n"),
612
- ticket(3, "## Dependency injection:\n\n- see #300\n"),
613
- ticket(4, "## **Dependency injection**\n\n- rework the container, see #300\n"),
614
- ticket(5, "## **Dependencies (blocking)**\n\n- #300\n"),
615
- ticket(6, "## **Dependency injection:**\n\n- see #300\n"),
616
- ]);
617
- assert.deepEqual(rows.map((r) => r.d), [[], [], [], [], [], []]);
618
- });
619
-
620
- test("#208's own brief text yields both of the open blockers it declared in prose", () => {
621
- // The live wrong admission #439 was filed from: this section reduced to no
622
- // dependency at all, so a ticket with two open blockers reached the survivor
623
- // read as claimable. What this fixture pins is the noun-form heading half:
624
- // its bolded refs sit on list-item lines inside the armed section, so the
625
- // list-item branch reads them whatever the label separator does — revert
626
- // that separator to its pre-#439 form and this row stays green, while
627
- // reverting the heading gate to the verb forms is what reds it. The
628
- // separator half is pinned by the two bolded-ref tests above instead.
629
- // The `PR #390` bullet is collected too: in-section collection reads every
630
- // `#N` on a list-item line by construction, and a merged ref is inert at the
631
- // consumer, which drops a ticket only for a blocker that is still open.
632
- const { rows } = run([
633
- ticket(208, [
634
- "## Dependencies",
635
- "",
636
- "- Blocked by **#179** (decides `gone()`'s shape, which `Indeterminate` needs), which is itself blocked by **#178**.",
637
- "- Vocabulary from **PR #390**.",
638
- "",
639
- "Recording those as native dependency edges so the fleet does not claim this early.",
640
- "",
641
- ].join("\n")),
642
- ]);
643
- assert.deepEqual(rows[0].d, [178, 179, 390]);
644
- });
645
-
646
- test("the phantom-blocker sweep stays closed against the widened label and heading gates", () => {
647
- // PR #331 closed these by restricting in-section collection to list-item
648
- // lines. #439 loosens the label separator and widens the heading gate,
649
- // neither of which touches that restriction — this pins that they did not,
650
- // because a regex widened for bold is exactly the change that reopens them.
651
- // The noun-form rows carry the same shapes through the newly-armed heading,
652
- // and #1031's bolded rows carry them once more through the emphasis the gate
653
- // tolerates next: a section armed by `## **Dependencies**` reads its LIST ITEMS
654
- // only, exactly as the unbolded one does, so the prose ref in row 8 and the
655
- // fenced one in row 9 stay out of `d`. Widening the gate does not widen what
656
- // an armed section then sweeps, and that is the half this re-run pins.
657
- const { rows } = run([
658
- ticket(1, "## Blocked by\n\nSee #99 for context\n"),
659
- ticket(2, "## Blocked by\n\n- #1\n\nSome later text mentioning #42\n"),
660
- // Row 3 is kept to pin an ACCEPTANCE CRITERION, not to discriminate a
661
- // mutant — #439's own criteria name this shape verbatim among the four the
662
- // phantom-blocker sweep must keep closed: "a fenced code block containing
663
- // `fixes #77` → `[]`". It is the executable record that that criterion
664
- // holds, and that is the whole job it is here for. Under the mutant this
665
- // test's NAME invokes — `$nextsec and $item` → `$nextsec`, dropping #331's
666
- // list-item restriction — every other row in this table moves and row 3
667
- // does not: no heading arms a section here, so that restriction is never
668
- // reached. Both facts measured. Do not delete it on the second one; it has
669
- // already been re-filed on exactly that reading once (#1034).
670
- // Row 6 is the SIBLING that covers the fenced case with real discriminating
671
- // power — a fence inside an armed `## Dependencies` section, which answers
672
- // `[12, 999]` under that same mutant (row 9 is its bolded twin). The two
673
- // are not duplicates and neither subsumes the other: row 6 discriminates
674
- // the mutant, row 3 records the criterion — and row 3 is in fact the only
675
- // row here that moves when the inline verb alternation is widened to admit
676
- // `fixes`, where it answers `[77]` and every other row holds still.
677
- ticket(3, "```\nfixes #77\n```\n"),
678
- ticket(4, "## Dependencies\n\n- #12 and then\n more about #999\n"),
679
- ticket(5, "## Dependencies\n\nSee #99 for context\n"),
680
- ticket(6, "## Dependencies\n\n- #12\n\n```\ngit log #999\n```\n"),
681
- ticket(7, "## **Dependencies**\n\n- #12 and then\n more about #999\n"),
682
- ticket(8, "## **Blocked by**\n\nSee #99 for context\n"),
683
- ticket(9, "## **Dependencies**\n\n- #12\n\n```\ngit log #999\n```\n"),
684
- ticket(10, "## **Dependencies**\n\n- None — can start immediately\n\nMirror what #300 did.\n"),
685
- ]);
686
- assert.deepEqual(
687
- rows.map((r) => r.d),
688
- [[], [1], [], [12], [], [12], [12], [], [12], []],
689
- );
690
- });
691
-
692
- // #1032. The per-candidate row prints `deps:` only when `d` is non-empty, so a
693
- // heading that named a dependency the gate refused printed EXACTLY the row of a
694
- // ticket that declared none — and the refusal is by design (`## Dependencies
695
- // (blocking)`, two tests above), so that shape declared blockers into a void
696
- // with nothing said. The new line is the signal, and matching on its wording is
697
- // the point rather than an accident: #1032's central constraint is that it reads
698
- // as "declared, not parsed" and never as a fault, because the near-miss is
699
- // legitimate usage and a fault-worded diagnostic false-alarms on every ticket
700
- // using it.
701
- const MISS = /not parsed as a dependency section/;
702
- const missLines = (stderr) => stderr.split("\n").filter((l) => MISS.test(l));
703
-
704
- test("a dependency heading that armed no section is named on stderr, and an armed one is not (#1032)", () => {
705
- const { status, stdout, stderr, rows } = run([
706
- // The near-miss the gate refuses on purpose: fires.
707
- ticket(1, "## Dependencies (blocking)\n\n- #300\n"),
708
- // No dependency heading at all — a `#N` in ordinary prose is not one: silent.
709
- ticket(2, "## Notes\n\n- see #300 for context\n"),
710
- // Armed correctly, refs collected: silent. The signal is about the
711
- // near-miss, not about every heading that mentions the concept.
712
- ticket(3, "## Dependencies\n\n- #12\n"),
713
- // The deliberate over-fire. Detection is LOOSER than the gate because the
714
- // headings worth naming are the ones the gate rejected — tighten it to the
715
- // gate and nothing is ever reported, which is what the neutral wording
716
- // below pays for.
717
- ticket(4, "## Dependency injection\n\n- rework the container, see #300\n"),
718
- // One line per CANDIDATE, not per heading.
719
- ticket(5, "## Dependencies (blocking)\n\n- #300\n\n## Dependency injection\n\n- see #42\n"),
720
- ]);
721
- const lines = missLines(stderr);
722
- assert.deepEqual(lines.map((l) => l.match(/#(\d+)/)[1]), ["1", "4", "5"], stderr);
723
- assert.ok(lines[0].includes("'## Dependencies (blocking)'"), lines[0]);
724
- assert.ok(lines[1].includes("'## Dependency injection'"), lines[1]);
725
- assert.ok(
726
- lines[2].includes("'## Dependencies (blocking)'") && lines[2].includes("'## Dependency injection'"),
727
- lines[2],
728
- );
729
- // Points at the heading, and reads as a fact about the scan rather than a
730
- // fault in the ticket. Both halves of that are load-bearing: a line worded
731
- // as an error is the false alarm #1032 refuses, and a line that names no
732
- // heading leaves the reader to guess which one the scan skipped.
733
- assert.ok(lines[0].includes("check the heading"), lines[0]);
734
- assert.doesNotMatch(lines[0], /error|fail|invalid|malformed|broken|warning|must /i);
735
- // The row format is untouched: #1 still prints no `deps:` — the empty list is
736
- // still an empty list — and #3 still prints its refs the same way.
737
- assert.match(stderr, /^ {4}#1 \[ready-for-agent\] ticket 1$/m);
738
- assert.match(stderr, /^ {4}#3 \[ready-for-agent\] ticket 3 {2}deps:12$/m);
739
- // Exit status and payload unchanged. The heading text reaches this process as
740
- // a jq field (the body never does), so the assertion that matters is that the
741
- // field is stripped before stdout: phase 0 and fleet-tick's readSupply() read the
742
- // same four keys they read before — `spec` is the reduction's other
743
- // stderr-only field, and `dropSpecs` has already destructured that one out.
744
- assert.equal(status, 0);
745
- assert.deepEqual(rows.map((r) => r.d), [[], [], [12], [], []]);
746
- for (const r of rows) assert.deepEqual(Object.keys(r).sort(), ["d", "l", "n", "t"]);
747
- assert.doesNotMatch(stdout, /dh/);
748
- });
749
-
750
- // #1741. `d` is the union of the body scan and the issue's native "blocked by"
751
- // edges. Fixtures carry `blockedBy` in the shape gh 2.100.0 emits —
752
- // `{nodes:[{number,state,…}], totalCount}` — and the STUB cuts every fixture
753
- // down to candidates.mjs's own `--json` list the way real gh does, so these
754
- // rows also pin that the query ASKS for the field: take `blockedBy` out of it
755
- // and no edge ever arrives. Every fixture before this block carries no
756
- // `blockedBy` at all, so each earlier `d` pin, unmodified, is also the pin
757
- // that an issue without edges still reduces to the body scan alone.
758
- const edges = (...nodes) => ({
759
- nodes: nodes.map(([number, state]) => ({ number, state })),
760
- totalCount: nodes.length,
761
- });
762
- const withEdges = (t, blockedBy) => ({ ...t, blockedBy });
763
- // #593's blocker line, verbatim from the live issue. Its first parenthetical
764
- // breaks the scan's ref chain, so the body alone yields [239] and never #531 —
765
- // the admission gap #1741 was filed on, left in place (the grammar is out of
766
- // scope), so the native edge is the only thing carrying #531.
767
- const BODY_593 =
768
- "**Blocked by #239** (the new bar must exist before anything is re-labelled against it) and **#531** (the dedupe guard's ledger arm is empty, so an unknown share of the pile is duplicates — promoting duplicates is worse than filing them).\n";
769
- // A blocker list longer than the one page gh fetches (`blockedBy(first:50)`).
770
- const SHORT_50_OF_51 = {
771
- nodes: Array.from({ length: 50 }, (_, i) => ({ number: 100 + i, state: "OPEN" })),
772
- totalCount: 51,
773
- };
774
-
775
- test("#593's shape: the blocker the body scan misses reaches `d` through its native edge, and the shared one lands once (#1741)", () => {
776
- // #592 is the same body with no edge: the scan's own answer, which is the
777
- // gap. #593's edges arrive 531-first; `d` comes out sorted and deduplicated,
778
- // so #239 — named by both the body and an edge — appears exactly once.
779
- const { status, rows, stderr } = run([
780
- ticket(592, BODY_593),
781
- withEdges(ticket(593, BODY_593), edges([531, "OPEN"], [239, "CLOSED"])),
782
- ]);
783
- assert.equal(status, 0, stderr);
784
- assert.deepEqual(rows.map((r) => r.d), [[239], [239, 531]]);
785
- });
786
-
787
- test("edges alone make `d` when the body names no blocker — a closed one kept, as the scan keeps one (#1741)", () => {
788
- // `d` stays RAW: openness is the consumer's call (run-team / next-ticket
789
- // step 2), so an edge's CLOSED state does not filter its number out here.
790
- const { rows } = run([withEdges(ticket(9, "## What to build\n\nx\n"), edges([12, "OPEN"], [7, "CLOSED"]))]);
791
- assert.deepEqual(rows[0].d, [7, 12]);
792
- });
793
-
794
- test("a blocker list gh cut short refuses at exit 2 — `d` never ships with blockers missing (#1741)", () => {
795
- // Unreachable today (GitHub caps an issue at 50 edges per relationship, and
796
- // gh fetches 50), which is exactly why it must be loud if either moves: a
797
- // truncated list reads as a smaller one, and #12's 51st blocker would drop
798
- // out of `d` with nothing said. The whole query refuses, like a capped
799
- // list does, and the message names the issue and the shortfall.
800
- const { status, stderr } = run([
801
- ticket(11, "## What to build\n\nx\n"),
802
- withEdges(ticket(12, "## What to build\n\nx\n"), SHORT_50_OF_51),
803
- ]);
804
- assert.equal(status, 2, stderr);
805
- assert.match(stderr, /#12: blockedBy lists 50 of 51 blockers, so d would miss the rest/);
806
- assert.match(stderr, /^candidates: gh issue list failed/m);
807
- });
808
-
809
- // The gap #63 named: the STUB above execs system jq (Oniguruma), but gh
810
- // applies `--jq` with its embedded gojq (RE2) — a different engine, and every
811
- // other test in this file accepts that gap rather than closing it. This one
812
- // closes it for the dependency scan specifically, since #58 asks for it by
813
- // name: when a real `gojq` binary is reachable (`go install
814
- // github.com/itchyny/gojq/cmd/gojq@v0.12.19`), the STUB runs the fixtures
815
- // through it instead of system jq, so the regex is checked against the exact
816
- // engine gh uses — not merely a same-family stand-in. No `gojq` on this
817
- // machine → skip, loudly, rather than silently passing on the weaker engine.
818
- // Answering `--version` is not the same as BEING gojq: system jq answers it
819
- // too, and resolving to it would run the fixtures on Oniguruma under a name
820
- // claiming RE2 — the silent degrade this whole check exists to refuse. gojq
821
- // prints `gojq 0.12.19 (rev: …)`, jq prints `jq-1.7.1-apple`.
822
- // Single source of truth for "is this gojq" — isGojq (which engine ran the
823
- // --version probe) and assertRanGojq (which engine the STUB recorded) each
824
- // own their own site under #383, and only sharing this pattern by
825
- // construction keeps them from silently desyncing if gojq's version-string
826
- // format ever changes: a future edit to one site would otherwise quietly
827
- // reopen the exact "gated test proved nothing" failure mode #383 is fixing.
828
- const GOJQ_VERSION_RE = /^gojq /;
829
- const isGojq = (bin) =>
830
- GOJQ_VERSION_RE.test(spawnSync(bin, ["--version"], { encoding: "utf8" }).stdout ?? "");
831
-
832
- function findGojq() {
833
- const named = process.env.GOJQ_BIN;
834
- // A caller who names the binary asked for that engine. Falling back past it
835
- // would run the check on something else and still report the gojq test green.
836
- if (named && !isGojq(named)) throw new Error(`GOJQ_BIN=${named} is not gojq`);
837
- if (named) return named;
838
- if (isGojq("gojq")) return "gojq";
839
- // No `go` on PATH → spawnSync fails ENOENT and `stdout` is undefined, not "".
840
- // Without `?.` this throws at import, taking every test in the file with it —
841
- // the opposite of the loud skip the comment above promises.
842
- const gopath = spawnSync("go", ["env", "GOPATH"], { encoding: "utf8" }).stdout?.trim();
843
- if (!gopath) return null;
844
- const candidate = join(gopath, "bin", "gojq");
845
- return isGojq(candidate) ? candidate : null;
846
- }
847
- const GOJQ = findGojq();
848
- const SKIP_WITHOUT_GOJQ = { skip: GOJQ ? false : "no gojq on PATH — go install github.com/itchyny/gojq/cmd/gojq@v0.12.19 to run this check" };
849
-
850
- // What proves a gated test REACHED gojq (#383). Until these rulings the proof
851
- // was indirect: one fixture per test that the two engines ANSWERED
852
- // differently — a U+00A0 that reduced to [12] under Oniguruma and [] under
853
- // RE2. The rulings removed every such row on purpose, so all 24 fixtures now
854
- // agree on both engines, and no assertion about OUTPUT can say which engine
855
- // produced it any more. The STUB therefore records the --version of the
856
- // binary it executed and this reads that recording, applying the very test
857
- // isGojq does. A strengthening of the harness, not a thinning of it
858
- // (.out-of-scope/gojq-parity-harness.md): the old discriminator proved the
859
- // engine only as a side effect of one disagreeing row, so it went blind the
860
- // moment that row was ruled, and it proved nothing about the OTHER assertions
861
- // in its own test — this proves it directly, for every gated assertion, and
862
- // keeps proving it however the fixtures are ruled later. Mutation-checked:
863
- // delete the STUB's JQ_RAN plumb and every gated test reds here.
864
- const assertRanGojq = (r) =>
865
- assert.match(
866
- r.engine,
867
- GOJQ_VERSION_RE,
868
- `the STUB recorded ${JSON.stringify(r.engine)} — this gated test never reached gojq`,
869
- );
870
-
871
- // The pre-#383 program, verbatim (from the commit #383's own diff replaced),
872
- // kept here ONLY so the discriminating test below can run it through
873
- // JQ_OVERRIDE and prove what it actually did on system jq — the engine the
874
- // three gojq-only fixtures below #908 cannot reach, because RE2 was already
875
- // ASCII-only at every one of these positions before #383 too, so gojq never
876
- // had the bug being fixed. Never used to run the real query; the current JQ
877
- // constant in candidates.mjs is the only one this script ever ships with.
878
- const OLD_JQ =
879
- 'def armed: test("(?i)^#{1,6}\\\\s+\\\\**((?:depends on|blocked by|requires)\\\\b|dependenc(?:y|ies)\\\\**:?\\\\**\\\\s*$)");\n' +
880
- '\n' +
881
- 'def depmiss:\n' +
882
- ' [ split("\\n")[]\n' +
883
- ' | select(test("(?i)^#{1,6}\\\\s.*(dependenc|blocked by|depends on|requires)") and (armed | not))\n' +
884
- ' | rtrimstr("\\r")\n' +
885
- ' ];\n' +
886
- '\n' +
887
- 'def depnums:\n' +
888
- ' (reduce (split("\\n"))[] as $line (\n' +
889
- ' {insec: false, nums: []};\n' +
890
- ' ($line | armed) as $bh\n' +
891
- ' | ($line | test("^#{1,6}\\\\s")) as $any\n' +
892
- ' | (if $any then $bh else .insec end) as $nextsec\n' +
893
- ' | ($line | test("^\\\\s*([-*+]|[0-9]+[.)])\\\\s")) as $item\n' +
894
- ' | {\n' +
895
- ' insec: $nextsec,\n' +
896
- ' nums: (\n' +
897
- ' .nums\n' +
898
- ' + (if $nextsec and $item then [$line | scan("#\\\\d+")] else [] end)\n' +
899
- ' + [ $line\n' +
900
- ' | scan("(?i)(?:depends on|blocked by|requires|after)[\\\\s*]*:?[\\\\s*]*(#\\\\d+(?:[\\\\s*]*(?:,|and)?[\\\\s*]*#\\\\d+)*)")\n' +
901
- ' | .[0]\n' +
902
- ' | scan("#\\\\d+")\n' +
903
- ' ]\n' +
904
- ' )\n' +
905
- ' }\n' +
906
- ' )).nums\n' +
907
- ' | map(ltrimstr("#") | tonumber)\n' +
908
- ' | unique;\n' +
909
- '\n' +
910
- '[.[] | {n:.number,t:.title,l:[.labels[].name],\n' +
911
- ' spec:((.body//"")|test("(?m)^#{2,6}[ \\\\t]+User Stories\\\\s*$")),\n' +
912
- ' d:((.body//"")|depnums),\n' +
913
- ' dh:((.body//"")|depmiss)}]\n';
914
-
915
- test(
916
- "dependency forms hold under gojq, the engine gh actually applies — not only system jq",
917
- SKIP_WITHOUT_GOJQ,
918
- () => {
919
- const extraEnv = { JQ_BIN: GOJQ };
920
- // gojq REJECTING the program is the failure this test exists to catch. Read
921
- // the exit status first: `rows[0]` is undefined the moment the child exits
922
- // non-zero, so without this the report is `Cannot read properties of
923
- // undefined (reading 'd')`, naming neither gojq nor the refusal it printed.
924
- const deps = (body) => {
925
- const r = run([ticket(9, body)], undefined, null, extraEnv);
926
- assert.equal(r.status, 0, r.stderr);
927
- // Every form, not just one row: the proof is per-application, so no
928
- // assertion below can pass having silently run on system jq.
929
- assertRanGojq(r);
930
- return r.rows[0].d;
931
- };
932
- assert.deepEqual(deps("## Blocked by\n\n- #12\n- #13\n"), [12, 13]);
933
- assert.deepEqual(deps("## Blocked by\n\n- None — can start immediately\n"), []);
934
- assert.deepEqual(deps("- A reference to #77 blocking ticket\n"), []);
935
- assert.deepEqual(deps("**Blocked by:** #12\n"), [12]);
936
- assert.deepEqual(deps("depends on #5\n"), [5]);
937
- assert.deepEqual(deps("Blocked by: #12, #13\n"), [12, 13]);
938
- assert.deepEqual(deps("## Blocked by\n\n- #12\n\n## Notes\n\nsee #999 for context\n"), [12]);
939
- // #439's two forms, on the engine that produced the live wrong admission.
940
- assert.deepEqual(deps("Blocked by **#179**\n"), [179]);
941
- assert.deepEqual(deps("Blocked by: **#12**\n"), [12]);
942
- assert.deepEqual(deps("Blocked by **#12** and **#13**\n"), [12, 13]);
943
- assert.deepEqual(deps("Blocked by: **#12**, **#13**\n"), [12, 13]);
944
- assert.deepEqual(deps("**Blocked by** : #12\n"), [12]);
945
- assert.deepEqual(deps("## Dependencies\n\n- #12\n- #13\n"), [12, 13]);
946
- assert.deepEqual(deps("## Dependency injection\n\n- see #300\n"), []);
947
- assert.deepEqual(deps("## Blocked by\n\nSee #99 for context\n"), []);
948
- // #1031's widened heading gate, on the engine gh actually applies. The
949
- // emphasis runs are plain `\*`, which Oniguruma and RE2 spell alike — so
950
- // these rows are expected to AGREE with the system-jq fixtures above, and
951
- // exist to confirm that rather than because a divergence was found. The
952
- // negative rows matter most here: RE2 has no backtracking, so a `$` anchor
953
- // reached through a widened alternation is exactly where the two engines
954
- // could have parted, and `## **Dependency injection**` arming under gojq
955
- // alone would be invisible to every other test in this file.
956
- assert.deepEqual(deps("## **Dependencies**\n\n- #12\n- #13\n"), [12, 13]);
957
- assert.deepEqual(deps("## **Blocked by**\n\n- #12\n"), [12]);
958
- assert.deepEqual(deps("## **Dependencies:**\n\n- #12\n"), [12]);
959
- assert.deepEqual(deps("## ***Dependencies***\n\n- #12\n"), [12]);
960
- assert.deepEqual(deps("## **Dependency injection**\n\n- see #300\n"), []);
961
- assert.deepEqual(deps("## **Dependency injection:**\n\n- see #300\n"), []);
962
- assert.deepEqual(deps("## **Dependencies (blocking)**\n\n- #300\n"), []);
963
- assert.deepEqual(deps("## **Dependencies**\n\nSee #99 for context\n"), []);
964
- // #1033's ACCEPTED LIMITATION, carried onto the engine gh applies — see
965
- // "a bold split landing MID-PHRASE misses the label" above for the ruling
966
- // and for what a widener owes. The [] rows are NOT a spec: they record
967
- // that a seam landing mid-phrase is missed, and being here says the gap
968
- // is a property of the alternation and not an engine artifact — measured,
969
- // both engines agree. The paired seam row is the control, so a scan broken
970
- // outright reds this block rather than confirming the limitation.
971
- assert.deepEqual(deps("**Blocked** by #12\n"), []);
972
- assert.deepEqual(deps("**Depends** on #5\n"), []);
973
- assert.deepEqual(deps("**Blocked by** #12\n"), [12]);
974
- // #383's inline-separator ruling: `[\t \p{Zs}*]*`, so a U+00A0 between
975
- // label and ref IS a separator and the ref IS a blocker. GFM renders
976
- // `Blocked by:<NBSP>#12` as `Blocked by: #12` and autolinks the ref, so a
977
- // reader sees a dependency — and an admission gate must not be the one
978
- // thing that misses it. This row asserted [] until the ruling: gojq's
979
- // answer, adopted because it was production's and never because anyone
980
- // ruled it correct. Both engines now say [12]. Keep the `\u00a0` escape:
981
- // a literal NBSP does not survive being copied.
982
- assert.deepEqual(deps("Blocked by:\u00a0#12\n"), [12]);
983
- // The same ruling through #439's widened separator, which now spells the
984
- // class instead of inheriting `\s`, so the bolded form agrees with the
985
- // bare one rather than inheriting a split.
986
- assert.deepEqual(deps("Blocked by:\u00a0**#12**\n"), [12]);
987
- // The separator BETWEEN refs takes that class too, so a chain does not
988
- // stop at the first ref when a NBSP follows the comma — gojq collected
989
- // [12] alone here.
990
- assert.deepEqual(deps("Blocked by #12,\u00a0#13\n"), [12, 13]);
991
- // `\d` → `[0-9]` (#383). Arabic-Indic digits are not refs anyone is trying
992
- // to support, and this is the one position where the divergence could take
993
- // the whole RUN down rather than change a row: under Oniguruma `\d` matched
994
- // these, the scan collected the ref, and `tonumber` then made the program
995
- // exit 5 — one ticket body took down the entire candidate listing. Exit 0
996
- // is half of what this pins, and `deps` asserts that before it returns.
997
- assert.deepEqual(deps("Blocked by: #\u0661\u0662\n"), []);
998
- // The after-marker ruling is the NARROWER `[ \t]`, not the separator class
999
- // above: GFM opens a heading on a space or a tab only, so this line is a
1000
- // PARAGRAPH, arms no section, and its bullet declares nothing. Two
1001
- // positions on one construct taking two different classes is exactly what
1002
- // ruling them per position buys.
1003
- assert.deepEqual(deps("##\u00a0Blocked by\n\n- #12\n"), []);
1004
- // The same class one level down, on the list item's own marker gap.
1005
- assert.deepEqual(deps("## Blocked by\n\n-\u00a0#12\n"), []);
1006
- // An ASCII NARROWING, and deliberate: `\f` was whitespace to `\s` on BOTH
1007
- // engines, so both used to collect [12] here. GFM renders `-<FF>#12` as a
1008
- // paragraph, so it is not a bullet and declares nothing.
1009
- assert.deepEqual(deps("## Blocked by\n\n-\f#12\n"), []);
1010
- // The same narrowing after the heading marker: `##<FF>Blocked by` is a
1011
- // paragraph, so it arms nothing and the bullet below is never read.
1012
- assert.deepEqual(deps("##\fBlocked by\n\n- #12\n"), []);
1013
- // `\b` → `(?:[^\p{L}\p{M}\p{N}_]|$)`, Unicode-aware where RE2's `\b` is
1014
- // ASCII. A letter continuing the word means a DIFFERENT word, so this
1015
- // heading does not arm — gojq armed it and collected [12].
1016
- assert.deepEqual(deps("## Blocked by\u00e9\n\n- #12\n"), []);
1017
- // End-of-heading-line ruling `[\t\f\r \p{Zs}]*`: the noun form must be the
1018
- // WHOLE heading, and U+00A0 padding does not stop it being that heading —
1019
- // GFM still renders an `<h2>`. gojq refused this and reported it as a near
1020
- // miss instead of collecting the blocker.
1021
- assert.deepEqual(deps("## Dependencies\u00a0\n\n1. #12\n"), [12]);
1022
- },
1023
- );
1024
-
1025
- // Findings 5-7 (review of #1746): the U+00A0/Arabic-Indic-digit/VERTICAL-TAB
1026
- // fixtures inside the three gojq-gated tests above and below are ALL
1027
- // gojq-only assertions, but RE2 was already ASCII-only at every one of these
1028
- // three positions before #383 too — so pre- and post-#383 agree under gojq
1029
- // on every one of them, and none of those gated assertions actually exercise
1030
- // the jq-vs-gojq divergence its own comment describes. That divergence is
1031
- // real, but only reachable on system jq (Oniguruma), which the gojq-gated
1032
- // tests structurally cannot run. This test reaches it directly, under plain
1033
- // system jq — no `JQ_BIN` override, no `SKIP_WITHOUT_GOJQ` guard — running
1034
- // OLD_JQ (pre-#383) through the STUB's `JQ_OVERRIDE` escape hatch to prove
1035
- // each old class actually MISBEHAVED here, paired with the current script
1036
- // (unmodified, no override) proving #383's narrower classes actually fixed
1037
- // it — on the one engine where the fix was ever needed.
1038
- test("system jq: the three class-position fixes actually changed behaviour on the engine that has the bug (#383)", () => {
1039
- // #1's after-marker toggle position, on the heading and on the list item's
1040
- // own marker gap: Oniguruma's old `\s` matched U+00A0, so `##<NBSP>Blocked
1041
- // by` and a bulleted `-<NBSP>#12` both wrongly armed/collected under jq.
1042
- const oldHeadingGap = run([ticket(1, "##\u00a0Blocked by\n\n- #12\n")], undefined, null, { JQ_OVERRIDE: OLD_JQ });
1043
- assert.deepEqual(oldHeadingGap.rows[0].d, [12], oldHeadingGap.stderr);
1044
- const oldListGap = run([ticket(2, "## Blocked by\n\n-\u00a0#12\n")], undefined, null, { JQ_OVERRIDE: OLD_JQ });
1045
- assert.deepEqual(oldListGap.rows[0].d, [12], oldListGap.stderr);
1046
- // #383's narrower `[ \t]` fixes both, on the same engine.
1047
- const newHeadingGap = run([ticket(1, "##\u00a0Blocked by\n\n- #12\n")]);
1048
- assert.deepEqual(newHeadingGap.rows[0].d, []);
1049
- const newListGap = run([ticket(2, "## Blocked by\n\n-\u00a0#12\n")]);
1050
- assert.deepEqual(newListGap.rows[0].d, []);
1051
-
1052
- // #2: Oniguruma's old `\d` matched Arabic-Indic digits, so `tonumber`
1053
- // crashed jq itself (exit 5), which candidates.mjs's own gh-call guard
1054
- // turns into a hard refusal rather than a silent per-ticket miss.
1055
- const oldDigits = run([ticket(3, "Blocked by: #\u0661\u0662\n")], undefined, null, { JQ_OVERRIDE: OLD_JQ });
1056
- assert.equal(oldDigits.status, 2, oldDigits.stderr);
1057
- assert.match(oldDigits.stderr, /gh issue list failed/);
1058
- // #383's `[0-9]` no longer matches them, so the same fixture no longer
1059
- // crashes the real CLI invocation — the actual bug #6 describes.
1060
- const newDigits = run([ticket(3, "Blocked by: #\u0661\u0662\n")]);
1061
- assert.equal(newDigits.status, 0, newDigits.stderr);
1062
- assert.deepEqual(newDigits.rows[0].d, []);
1063
-
1064
- // #3: Oniguruma's old `\s*$` trailing class matched VERTICAL TAB, so a
1065
- // `## User Stories<VT>` heading was wrongly to-spec and dropSpecs silently
1066
- // discarded the ticket. #999 keeps the queue non-empty either way, so #1's
1067
- // presence in rows is the verdict.
1068
- const oldVt = run(
1069
- [ticket(4, "## User Stories\u000b\n\nx\n"), ticket(999, "## What to build\n\nx\n")],
1070
- undefined,
1071
- null,
1072
- { JQ_OVERRIDE: OLD_JQ },
1073
- );
1074
- assert.equal(oldVt.status, 0, oldVt.stderr);
1075
- assert.ok(!oldVt.rows.some((r) => r.n === 4), "old class must have dropped #4 as a false spec");
1076
- // #383's class excludes VT on purpose (GFM renders it as U+FFFD, not
1077
- // blank), so the same heading is no longer to-spec and #4 survives.
1078
- const newVt = run([ticket(4, "## User Stories\u000b\n\nx\n"), ticket(999, "## What to build\n\nx\n")]);
1079
- assert.equal(newVt.status, 0, newVt.stderr);
1080
- assert.ok(newVt.rows.some((r) => r.n === 4), "the fixed class must not drop #4");
1081
- });
1082
-
1083
- // #1032's diagnostic, on the engine that actually runs it. The test above
1084
- // exercises `depmiss` under system jq only, and a `def` gojq parses differently
1085
- // would red the dependency-forms test next door (it asserts exit 0) — but a
1086
- // `test` RE2 merely reads differently would not: the diagnostic would go dead
1087
- // in production, green here, which is exactly the silent degrade the two
1088
- // gated tests either side of this one exist to refuse. The near-miss row and
1089
- // the armed row together, so a `depmiss` that answered [] and one that
1090
- // answered every heading are both caught.
1091
- test(
1092
- "the near-miss diagnostic fires under gojq too — the engine gh actually applies (#1032)",
1093
- SKIP_WITHOUT_GOJQ,
1094
- () => {
1095
- const r = run(
1096
- [
1097
- ticket(1, "## Dependencies (blocking)\n\n- #300\n"),
1098
- ticket(2, "## Dependencies\n\n- #12\n"),
1099
- ],
1100
- undefined,
1101
- null,
1102
- { JQ_BIN: GOJQ },
1103
- );
1104
- const { status, stderr } = r;
1105
- assert.equal(status, 0, stderr);
1106
- assertRanGojq(r);
1107
- const lines = missLines(stderr);
1108
- assert.deepEqual(lines.map((l) => l.match(/#(\d+)/)[1]), ["1"], stderr);
1109
- assert.ok(lines[0].includes("'## Dependencies (blocking)'"), lines[0]);
1110
- },
1111
- );
1112
-
1113
- // #65: the spec predicate's decisions, checked against gojq specifically —
1114
- // gojq is what gh applies, so gojq is where a leak would actually happen.
1115
- // The first three fixtures are plain ASCII (`#`, `[ \t]`, `\r`), so none of
1116
- // them is expected to diverge between engines; they exist to confirm that,
1117
- // not because a divergence was found. The predicate's trailing class still
1118
- // contains `\r` for #65's reason: b4739c8 widened it back from `[ \t]*` so a
1119
- // CRLF line's `\r` still reaches the anchor, which is the only reason the
1120
- // third fixture matches at all — drop `\r` from that class and the CRLF
1121
- // assertion is what reds. #383 then spelled the class out as
1122
- // `[\t\f\r \p{Zs}]*` instead of leaving it `\s*`, because `\s` was the one
1123
- // part of this predicate whose meaning changed with the engine (Oniguruma
1124
- // Unicode-aware, RE2 `[\t\n\f\r ]` — not even every ASCII whitespace, since
1125
- // a vertical tab diverged too). The rows at the end pin each edge of that
1126
- // ruling: `\p{Zs}` padding is still a spec, a VERTICAL TAB is not, and a
1127
- // U+00A0 after the `#` marker is not a heading at all. Which engine actually
1128
- // ran is no longer inferred from a disagreement — `assertRanGojq` reads the
1129
- // STUB's recording; see this file's header.
1130
- test(
1131
- "the spec predicate holds under gojq — depth widened, split heading not spanned, CRLF caught",
1132
- SKIP_WITHOUT_GOJQ,
1133
- () => {
1134
- const extraEnv = { JQ_BIN: GOJQ };
1135
- // #9 is the body under test; #999 keeps the queue non-empty either way, so
1136
- // a passing run always exits 0 and #9's presence/absence in rows is what
1137
- // reads the verdict — dropSpecs strips the `spec` field from surviving rows.
1138
- const isDroppedAsSpec = (body) => {
1139
- const r = run(
1140
- [ticket(9, body), ticket(999, "## What to build\n\nx\n")],
1141
- undefined,
1142
- null,
1143
- extraEnv,
1144
- );
1145
- assert.equal(r.status, 0, r.stderr);
1146
- assertRanGojq(r);
1147
- return !r.rows.some((row) => row.n === 9);
1148
- };
1149
- assert.equal(isDroppedAsSpec("### User Stories\n\nnested one level deeper\n"), true);
1150
- assert.equal(isDroppedAsSpec("##\nUser Stories\n\nsplit across lines\n"), false);
1151
- assert.equal(isDroppedAsSpec("## User Stories\r\n\r\n1. As a user…\r\n"), true);
1152
- // #383's end-of-heading-line ruling: `[\t\f\r \p{Zs}]*` before the anchor.
1153
- // `## User Stories` padded with U+00A0 still renders a real `<h2>` on
1154
- // GitHub, so it is still a to-spec spec and is still dropped. This row
1155
- // asserted `false` until the ruling — gojq's answer, i.e. production's,
1156
- // adopted because it was production's and never because anyone ruled it
1157
- // correct; under it the spec shipped as a claimable ticket, the under-fire
1158
- // direction an admission gate must never take. Both engines now say true.
1159
- // Keep the `\u00a0` escape: a literal NBSP does not survive being copied,
1160
- // and has already produced a false refutation of a correct finding in this
1161
- // repo (#200).
1162
- assert.equal(isDroppedAsSpec("## User Stories\u00a0\n\nx\n"), true);
1163
- // Any `\p{Zs}`, not NBSP alone: every Unicode space separator renders as a
1164
- // blank, so the rule is "looks like a space" rather than a list of code
1165
- // points anyone has met so far.
1166
- assert.equal(isDroppedAsSpec("## User Stories\u2003\n\nx\n"), true);
1167
- // VERTICAL TAB is excluded ON PURPOSE, and it is why the class is not
1168
- // simply "whitespace": GFM renders U+000B as U+FFFD, so a reader sees a
1169
- // replacement character sitting in the heading, not padding. Not a spec.
1170
- assert.equal(isDroppedAsSpec("## User Stories\u000b\n\nx\n"), false);
1171
- // The after-marker position takes the narrower `[ \t]`: GFM opens a
1172
- // heading on a space or a tab only, so this is a paragraph and not a
1173
- // heading at all. Same line, two positions, two classes.
1174
- assert.equal(isDroppedAsSpec("##\u00a0User Stories\n\nx\n"), false);
1175
- },
1176
- );
1177
-
1178
- // #1741's half of the program under the engine gh applies. `blockers` holds no
1179
- // regex, so no class ruling is at stake — what gojq could still refuse or
1180
- // read differently is the rest: `//` over a missing field, `error()` with an
1181
- // interpolated message, and `unique` over the union. The STUB cuts each
1182
- // fixture to the `--json` list with gojq too, so this runs gojq end to end.
1183
- test("native edges reduce the same under gojq — union, dedupe, closed kept, and the short-list refusal (#1741)", SKIP_WITHOUT_GOJQ, () => {
1184
- const extraEnv = { JQ_BIN: GOJQ };
1185
- const r = run(
1186
- [
1187
- ticket(592, BODY_593),
1188
- withEdges(ticket(593, BODY_593), edges([531, "OPEN"], [239, "CLOSED"])),
1189
- withEdges(ticket(9, "## What to build\n\nx\n"), edges([12, "OPEN"], [7, "CLOSED"])),
1190
- ],
1191
- undefined,
1192
- null,
1193
- extraEnv,
1194
- );
1195
- assert.equal(r.status, 0, r.stderr);
1196
- assertRanGojq(r);
1197
- assert.deepEqual(r.rows.map((row) => row.d), [[7, 12], [239], [239, 531]]);
1198
- const short = run([withEdges(ticket(12, "## What to build\n\nx\n"), SHORT_50_OF_51)], undefined, null, extraEnv);
1199
- assert.equal(short.status, 2, short.stderr);
1200
- assertRanGojq(short);
1201
- assert.match(short.stderr, /#12: blockedBy lists 50 of 51 blockers, so d would miss the rest/);
1202
- });
1203
-
1204
- test("the cap is checked before specs are dropped — filtering first hides truncation", () => {
1205
- // The ordering candidates.mjs calls load-bearing. Swap the two and this is
1206
- // the only thing that fails: dropSpecs shrinks the array below --limit, the
1207
- // cap check stops seeing a capped list, and the silent-truncation bug the
1208
- // "no silent caps" rule exists to close comes back with every test green.
1209
- const { status, stderr } = run(
1210
- [
1211
- ticket(10, "## User Stories\n\n1. As a user…\n"),
1212
- ticket(11, "## What to build\n\nx\n"),
1213
- ],
1214
- ["--require-label", "ready-for-agent", "--limit", "2"],
1215
- );
1216
- assert.equal(status, 2);
1217
- assert.match(stderr, /capped/);
1218
- });
1219
-
1220
- test("a labeled queue of only specs falls back to unfiltered — all filtered out is an empty queue", () => {
1221
- // The emptiness test gating --allow-fallback has to read the length AFTER
1222
- // the specs are dropped. Reading the raw one leaves a ready-for-agent queue
1223
- // holding nothing but to-spec specs reporting "no work" with the fallback
1224
- // untried, while #12 sits there claimable.
1225
- const { rows, status, stderr } = run(
1226
- [
1227
- ticket(10, "## Problem Statement\n\nx\n\n## User Stories\n\n1. As a user…\n"),
1228
- ticket(11, "## User Stories\n\n2. As a user…\n"),
1229
- ],
1230
- ["--require-label", "ready-for-agent", "--allow-fallback"],
1231
- [
1232
- ticket(12, "## What to build\n\nreal work\n", ["ready-for-human"]),
1233
- // A spec HERE is what makes the assertion below discriminate: delete the
1234
- // strip inside the fallback and #13 ships as a claimable ticket. A leaked
1235
- // spec passes phase 2's bail tests, so a member implements a whole spec.
1236
- ticket(13, "## User Stories\n\n4. As a user…\n", ["ready-for-human"]),
1237
- ],
1238
- );
1239
- assert.deepEqual(rows.map((r) => r.n), [12]);
1240
- // The BRANCH ran, not merely that the unfiltered fixture reached some query.
1241
- // Move the positive term to the front of `search` and the stub serves the
1242
- // LABELED query the unfiltered fixture: same payload, same status, fallback
1243
- // never entered. The payload alone pins the fixture, never the branch.
1244
- assert.match(stderr, /retrying unfiltered/);
1245
- // Each pass's drop line is attributable to the query it came from. Swap the
1246
- // two dropSpecs call-site labels and both lines still print, both still name
1247
- // a real spec, and nothing else in this file notices — these two asserts are
1248
- // what tell pass 1's drop from pass 2's. Deliberately not a count: drops are
1249
- // not deduplicated across passes (candidates.mjs's dropSpecs, #133), so a raw
1250
- // `grep -c 'dropped #'` reads 4 here whether or not the tags are present.
1251
- assert.match(stderr, /dropped #10 — to-spec spec, not a ticket \(User Stories heading\) \[label:ready-for-agent\]/);
1252
- assert.match(stderr, /dropped #13 — to-spec spec, not a ticket \(User Stories heading\) \[unfiltered \(fallback\)\]/);
1253
- // Two queries actually RAN. The announcement above prints before the query,
1254
- // so it pins the branch being entered, never that its answer shipped.
1255
- assert.equal(stderr.match(/^\$ gh /gm).length, 2);
1256
- // The strip inside the block ran. Delete that line and every assertion above
1257
- // still passes while `spec` — and any spec row — reaches the payload.
1258
- assert.deepEqual(Object.keys(rows[0]).sort(), ["d", "l", "n", "t"]);
1259
- // Exit 0, not 1: the payload and the "is there work" answer are one fact,
1260
- // and a caller that reads only the status must not still hear "empty".
1261
- assert.equal(status, 0);
1262
- });
1263
-
1264
- test("the cap is checked before specs are dropped in the fallback too, not only in the labeled query", () => {
1265
- // The cap-before-drop ordering of "the cap is checked before specs are
1266
- // dropped", one branch deeper — NOT the drop-before-emptiness test directly
1267
- // above, which pins the other half of the same wedge. The labeled query is
1268
- // under its cap on 1 row, empties out, and hands over to the fallback —
1269
- // whose 2 rows hit --limit 2 exactly. Drop first and one spec leaves, the
1270
- // cap check sees 1, and the truncated list ships as an answer.
1271
- const { status, stderr } = run(
1272
- [ticket(10, "## User Stories\n\n1. As a user…\n")],
1273
- ["--require-label", "ready-for-agent", "--allow-fallback", "--limit", "2"],
1274
- [
1275
- ticket(12, "## User Stories\n\n3. As a user…\n", ["ready-for-human"]),
1276
- ticket(13, "## What to build\n\nx\n", ["ready-for-human"]),
1277
- ],
1278
- );
1279
- assert.equal(status, 2);
1280
- // Pin which query refused. `/capped/` alone is satisfied by the labeled
1281
- // query dying, which is a different bug with the same exit code.
1282
- assert.match(stderr, /unfiltered \(fallback\)/);
1283
- });
1284
-
1285
- test("the spec predicate never reaches the payload — it is pure token cost downstream", () => {
1286
- const { rows } = run([ticket(12, "## What to build\n\nplain ticket\n")]);
1287
- assert.equal(rows.length, 1);
1288
- assert.deepEqual(Object.keys(rows[0]).sort(), ["d", "l", "n", "t"]);
1289
- });
1290
-
1291
- // `$ gh …` is echoed by query() before every invocation, so its absence is how
1292
- // these pin that the refusal came BEFORE any query ran. Status 2 alone does not:
1293
- // a query that ran and then died still exits 2, and for --require-label the
1294
- // query that runs is the widening one the whole file exists to prevent.
1295
- const queriesRun = (stderr) => (stderr.match(/^\$ gh /gm) ?? []).length;
1296
-
1297
- test("a trailing --limit refuses — an absent value is not a licence to use the default", () => {
1298
- const { status, stderr } = run(
1299
- [ticket(11, "## What to build\n\nx\n")],
1300
- ["--require-label", "ready-for-agent", "--limit"],
1301
- );
1302
- assert.equal(status, 2);
1303
- // Not 0-with-500-rows. `arg()` handed back `undefined`, `?? 500` read that as
1304
- // "flag absent", and the positive-integer guard never saw the malformed input
1305
- // it exists to catch.
1306
- assert.equal(queriesRun(stderr), 0);
1307
- });
1308
-
1309
- test("a trailing --require-label refuses — a falsy label runs the query --allow-fallback exists to gate", () => {
1310
- const { status, stderr } = run([ticket(11, "## What to build\n\nx\n")], ["--require-label"]);
1311
- // The worst of the three: query() reads a falsy label as "no label", so the
1312
- // UNFILTERED query shipped at exit 0 — the exact widening onto ready-for-human
1313
- // that run-team/SKILL.md's "do NOT pass `--allow-fallback`" prevents, reached
1314
- // without the flag. Asserting the status alone would pass on a run that
1315
- // widened and then happened to die.
1316
- assert.equal(queriesRun(stderr), 0);
1317
- assert.equal(status, 2);
1318
- });
1319
-
1320
- test("a --require-label whose value is the next flag refuses — it reaches the same widening", () => {
1321
- // Same harm as the trailing case, one keystroke away: `--allow-fallback` is
1322
- // consumed as the label, `label:--allow-fallback` matches nothing, and the
1323
- // empty result hands straight over to the unfiltered fallback at exit 0.
1324
- const { status, stderr } = run([ticket(11, "## What to build\n\nx\n")], ["--require-label", "--allow-fallback"]);
1325
- assert.equal(queriesRun(stderr), 0);
1326
- assert.equal(status, 2);
1327
- });
1328
-
1329
- test("a --require-label whose value is empty refuses — an unset shell variable is not 'no label'", () => {
1330
- // The likeliest spelling of all of them, and the only one the caller cannot
1331
- // see: `--require-label "$LABEL"` with the variable unset leaves an empty
1332
- // argv slot, not a missing one. `""` is falsy, so query() dropped the label
1333
- // term and ran the UNFILTERED search at exit 0 — and stderr said
1334
- // `N candidate(s)` with no label suffix, indistinguishable from a clean run.
1335
- const { status, stderr } = run([ticket(11, "## What to build\n\nx\n")], ["--require-label", ""]);
1336
- assert.equal(queriesRun(stderr), 0);
1337
- assert.equal(status, 2);
1338
- });
1339
-
1340
- test("a --require-label=value refuses — indexOf cannot see the = form, so the flag reads as absent", () => {
1341
- // Not the same route as the three above: `indexOf("--require-label")` misses
1342
- // `--require-label=x` entirely, so arg() returned null — "flag absent" — and
1343
- // the unfiltered query ran at exit 0 with the label the caller did pass.
1344
- const { status, stderr } = run(
1345
- [ticket(11, "## What to build\n\nx\n")],
1346
- ["--require-label=ready-for-agent"],
1347
- );
1348
- assert.equal(queriesRun(stderr), 0);
1349
- assert.equal(status, 2);
1350
- });
1351
-
1352
- test("an unknown flag refuses and names it — a flag that is merely ignored runs unfiltered", () => {
1353
- // `--label ready-for-agent` is not an invented typo: it is what run-team's
1354
- // phase 0 rule said, one line under a command spelling it `--require-label`.
1355
- // Ignored, the label term never reaches gh, so the ready-for-human queue
1356
- // ships at exit 0 — the widening `--require-label` exists to prevent, reached
1357
- // by following this repo's own instruction.
1358
- const { status, stderr } = run(
1359
- [ticket(11, "## What to build\n\nx\n")],
1360
- ["--label", "ready-for-agent"],
1361
- );
1362
- assert.equal(queriesRun(stderr), 0);
1363
- assert.equal(status, 2);
1364
- // The flag has to be NAMED. A bare "bad arguments" leaves the caller — markdown
1365
- // read by a model, with no way to see the script — re-guessing its own spelling.
1366
- // Line-anchored, per die()'s leading newline.
1367
- assert.match(stderr, /^candidates: Unknown option '--label'/m);
1368
- });
1369
-
1370
- // The test above is satisfied by a guard that only works when the bad flag is
1371
- // FIRST, which is what shipped: the catch tested
1372
- // `e.code === "ERR_PARSE_ARGS_UNKNOWN_OPTION"` and dropped every other
1373
- // parseArgs error. That is not a narrower guard but a disabled one. parseArgs
1374
- // stops at the FIRST offending argument, so an argument raising any other code
1375
- // was swallowed and the unknown flag behind it was never reached — the query
1376
- // then ran UNFILTERED at exit 0, which is the whole harm #173 exists to stop.
1377
- //
1378
- // Rows 1-2 are the two masking arguments on their own; rows 3-4 put a real
1379
- // `--label` BEHIND each of them, which is the shape the code test let through
1380
- // and the one no other test in this file covers. All four must refuse before
1381
- // any query — `status` alone is not enough, since exit 2 is also what a broken
1382
- // query produces after gh has already run.
1383
- //
1384
- // Rows 2 and 4 no longer reach parseArgs at all: #364 gave has() its own `=`
1385
- // guard, evaluated when `allowFallback` is assigned — above and long before
1386
- // the parseArgs call below — so the boolean case now dies with has()'s own
1387
- // boolean-specific wording (distinct from arg()'s "needs a space-separated
1388
- // value", since a boolean has no value to take), never Node's "does not take
1389
- // an argument". That does NOT make `ERR_PARSE_ARGS_INVALID_OPTION_VALUE` dead
1390
- // here, and an earlier draft of this comment claimed it did: arg()'s own `=`
1391
- // guard is indexOf-gated, firing only when the bare spelling is absent, so a
1392
- // REPEATED string flag whose later occurrence carries no value sails past both
1393
- // guards into parseArgs — measured, `--limit 5 --limit` raises exactly that
1394
- // code, with Node's wording (#462 review). The catch stays unconditional for
1395
- // that reason among others. The loop still refuses before any query and still
1396
- // names the offending flag, which is the property that matters; only the
1397
- // source moved.
1398
- for (const [what, args, refusal] of [
1399
- [
1400
- "a bare positional",
1401
- ["ready-for-agent"],
1402
- /^candidates: Unexpected argument 'ready-for-agent'\./m,
1403
- ],
1404
- [
1405
- "a boolean flag handed a =value",
1406
- ["--allow-fallback=true"],
1407
- /^candidates: --allow-fallback is a boolean flag, not --allow-fallback=/m,
1408
- ],
1409
- [
1410
- "an unknown flag behind a positional",
1411
- ["junk", "--label", "ready-for-agent"],
1412
- /^candidates: Unexpected argument 'junk'\./m,
1413
- ],
1414
- [
1415
- "an unknown flag behind a =value boolean",
1416
- ["--allow-fallback=true", "--label", "ready-for-agent"],
1417
- /^candidates: --allow-fallback is a boolean flag, not --allow-fallback=/m,
1418
- ],
1419
- ]) {
1420
- test(`${what} refuses before any query — every parseArgs error is a refusal, not just the unknown-name one`, () => {
1421
- const { status, stderr } = run([ticket(11, "## What to build\n\nx\n")], args);
1422
- assert.equal(queriesRun(stderr), 0);
1423
- assert.equal(status, 2);
1424
- // The refusal has to name the OFFENDING argument, not merely the accepted
1425
- // set: parseArgs reports the first offender, and a caller told only "bad
1426
- // arguments" cannot tell which of its four tokens was wrong.
1427
- assert.match(stderr, refusal);
1428
- });
1429
- }
1430
-
1431
- // Pre-#364, `--allow-fallback=<value>` was silently IGNORED rather than merely
1432
- // unchecked: `has()` compared exactly, so no `=` form ever matched and the
1433
- // fallback was disabled without a word. The fixtures below are the arrangement
1434
- // that showed it — the labeled query comes back empty and the unfiltered one
1435
- // has a row — so the flag was the only thing standing between exit 1 and exit
1436
- // 0. Measured on the pre-fix version, all three spellings here exited 1
1437
- // ("queue empty, query fine") for a run that never applied the flag it was
1438
- // handed, while the space-separated spelling exited 0 off the fallback.
1439
- // `notEqual(1)` is the load-bearing assertion, not decoration: it pins the
1440
- // exact wrong code the ignored form used to produce, now unreachable — has()
1441
- // refuses this argv outright, before `allowFallback` is even assigned. All
1442
- // three spellings are the ones the ticket's acceptance criteria name.
1443
- for (const v of ["=true", "=false", "="]) {
1444
- test(`--allow-fallback${v} refuses outright, rather than being silently dropped and falling through`, () => {
1445
- const { status, stderr } = run(
1446
- [],
1447
- ["--require-label", "nonexistent", `--allow-fallback${v}`],
1448
- [ticket(12, "## What to build\n\ny\n", ["ready-for-human"])],
1449
- );
1450
- assert.equal(queriesRun(stderr), 0);
1451
- assert.notEqual(status, 1);
1452
- assert.equal(status, 2);
1453
- assert.match(stderr, /^candidates: --allow-fallback is a boolean flag, not --allow-fallback=/m);
1454
- });
1455
- }
1456
-
1457
- // The control for the test above. Without it, "refuses the = form" is equally
1458
- // satisfied by a guard that refuses `--allow-fallback` in every spelling, which
1459
- // would break the one caller that legitimately passes it
1460
- // (next-ticket/SKILL.md's `## 1. Candidates` step). Same fixtures,
1461
- // space-separated: the fallback has
1462
- // to run and ship the unfiltered row at exit 0.
1463
- test("--allow-fallback still falls back in its space-separated spelling — the = refusal is not a blanket one", () => {
1464
- const { status, stderr, rows } = run(
1465
- [],
1466
- ["--require-label", "nonexistent", "--allow-fallback"],
1467
- [ticket(12, "## What to build\n\ny\n", ["ready-for-human"])],
1468
- );
1469
- assert.equal(queriesRun(stderr), 2);
1470
- assert.equal(status, 0);
1471
- assert.deepEqual(rows.map((r) => r.n), [12]);
1472
- });
1473
-
1474
- // The parseArgs call sits BELOW the value guards so their wording wins wherever
1475
- // both would refuse, and ORDER is the only thing that decides that. Node's
1476
- // message for this same argv is "Option '--require-label <value>' argument
1477
- // missing", which names the syntax; #169's names the flag the way every other
1478
- // refusal in this file does. Anchored at both ends so the suffix parseArgs'
1479
- // branch appends cannot satisfy it.
1480
- test("--require-label with no value keeps #169's wording — the name check runs after the value guards", () => {
1481
- const { status, stderr } = run([ticket(11, "## What to build\n\nx\n")], ["--require-label"]);
1482
- assert.equal(queriesRun(stderr), 0);
1483
- assert.equal(status, 2);
1484
- assert.match(stderr, /^candidates: --require-label needs a value$/m);
1485
- });
1486
-
1487
- test("gh output that is not an array refuses — a reduction that did not apply is not an empty queue", () => {
1488
- const { status, stderr } = run(
1489
- [ticket(11, "## What to build\n\nx\n")],
1490
- ["--require-label", "ready-for-agent"],
1491
- null,
1492
- { JQ_OVERRIDE: '{message:"Not Found"}' },
1493
- );
1494
- // 2, not 1. Unchecked, the object flows on and the first use of it throws —
1495
- // and an uncaught throw exits 1, the code reserved for "successful query, no
1496
- // survivors". A caller reading only the status hears "there is no work".
1497
- assert.equal(status, 2);
1498
- // Anchored on the die() prefix: `--jq` also appears in the echoed `$ gh` line,
1499
- // so an unanchored /--jq/ passes on a run that never refused at all.
1500
- assert.match(stderr, /^candidates: .*--jq/m);
1501
- });
1502
-
1503
- test("gh output that is empty refuses — the reduction emits an array for every input, including none", () => {
1504
- // The one shape the checks below cannot see, because it never reaches the
1505
- // parse. `[…]`-wrapped, the expression emits `[]` for an empty list, so no
1506
- // output at all means it did not run — the same fact as a wrong shape, and
1507
- // NOT an empty queue. Returning `[]` here spent that on exit 1, "no work".
1508
- const { status, stderr } = run(
1509
- [ticket(11, "## What to build\n\nx\n")],
1510
- ["--require-label", "ready-for-agent"],
1511
- null,
1512
- { JQ_OVERRIDE: "empty" },
1513
- );
1514
- assert.equal(status, 2);
1515
- assert.match(stderr, /^candidates: .*--jq/m);
1516
- });
1517
-
1518
- test("gh rows that were never reduced refuse — raw issues are not {n,t,l,d,dh,spec}", () => {
1519
- // What a gh that ignored `--jq` actually returns: the unreduced `--json`
1520
- // payload. An array, so an array check alone passes it through.
1521
- const { status, stderr } = run(
1522
- [ticket(11, "## What to build\n\nx\n")],
1523
- ["--require-label", "ready-for-agent"],
1524
- null,
1525
- { JQ_OVERRIDE: "." },
1526
- );
1527
- assert.equal(status, 2);
1528
- assert.match(stderr, /^candidates: .*--jq/m);
1529
- // The refusal names the row by index and never prints it. This is the one
1530
- // payload-shaped stderr path in the file, and an unreduced row carries the
1531
- // full issue body — the ~97% this script exists to not fetch. Without this,
1532
- // a die() that interpolated the row instead of its index kept the whole
1533
- // suite green while emitting every body it refused to pull.
1534
- assert.doesNotMatch(stderr, /What to build/);
1535
- });
1536
-
1537
- test("a failed gh query refuses without re-emitting gh's own stderr", () => {
1538
- // A program jq cannot compile: it exits non-zero, so execFileSync throws and
1539
- // the fail-closed die() runs. jq echoes the offending program in its own
1540
- // error, which is the marker below.
1541
- const { status, stderr } = run(
1542
- [ticket(11, "## What to build\n\nx\n")],
1543
- ["--require-label", "ready-for-agent"],
1544
- null,
1545
- { JQ_OVERRIDE: "MARKER_ZZZ(((" },
1546
- );
1547
- // 2, not 1: a broken query is not an empty queue — the archetypal fail-closed
1548
- // path, and the one die() in query() that had no test at all.
1549
- assert.equal(status, 2);
1550
- // Both halves of the contract, and neither is a count over the whole stream.
1551
- // Counting occurrences looks like the tighter pin but is coupled to jq's
1552
- // choice of error path: `(((` is unbalanced, so jq takes the SYNTAX-error
1553
- // branch, which echoes the program once. Balanced `MARKER_ZZZ` takes the
1554
- // undefined-function branch, which names the symbol AND echoes the program —
1555
- // two occurrences from the child alone, failing an exactly-1 assertion
1556
- // against correct code.
1557
- //
1558
- // So: the child's copy must arrive (execFileSync forwards it), and the
1559
- // refusal line must not carry it a second time (#176 — measured 7,700 B of
1560
- // gh stderr becoming 15,454 B, into a caller that is markdown read by a
1561
- // model). The same duplication this file's row-shape refusal already avoids.
1562
- assert.match(stderr, /MARKER_ZZZ/);
1563
- const refusal = stderr.match(/^candidates: .*$/m)[0];
1564
- assert.match(refusal, /^candidates: gh issue list failed/);
1565
- assert.doesNotMatch(refusal, /MARKER_ZZZ/);
1566
- });
1567
-
1568
- test("the exit code survives a gh stderr larger than the pipe buffer — the case that motivated it", () => {
1569
- // The failure mode the fix targets, at the size it actually happens. Once
1570
- // ~64 KiB of gh's forwarded stderr is already queued on a pipe, this fd is
1571
- // non-blocking and die()'s own writeSync can throw EAGAIN. Unfixed, that
1572
- // throw is uncaught: process.exit(2) never runs and the process falls
1573
- // through to exit 1 — "query fine, queue empty", the wrong fact (#299).
1574
- // Under the buffer every variant passes, so the program must stay large.
1575
- //
1576
- // Flaky by nature — it races the pipe reader — so it's pinned by repeated
1577
- // runs under Linux (this repo's CI target), not by a single local pass:
1578
- // unfixed, 7/15 runs inverted to exit 1; fixed, 0/20 did, 1000+ msgs run.
1579
- // On darwin EAGAIN never fires at all — unfixed, the status assertion below
1580
- // caught the revert 0/15 — so a local green confirms nothing and the source
1581
- // pin at the end is the only gate on this platform.
1582
- //
1583
- // Text, not just status: this only asserts the EXIT CODE, not the refusal
1584
- // wording. #889 gave die() a bounded EAGAIN retry loop (MAX_EAGAIN_RETRIES
1585
- // in arg.mjs) that resumes a short write and retries the exact EAGAIN this
1586
- // test forces, so the message is no longer lost on the first EAGAIN the way
1587
- // it was before #889 — only a reader that never drains past the retry cap
1588
- // can still lose it, and #299 is what keeps the exit code honest either way.
1589
- const { status, stderr } = run(
1590
- [ticket(11, "## What to build\n\nx\n")],
1591
- ["--require-label", "ready-for-agent"],
1592
- null,
1593
- { JQ_OVERRIDE: `${"z".repeat(100_000)}(((` },
1594
- );
1595
- assert.ok(stderr.length > 60_000, `gh stderr must exceed the buffer, got ${stderr.length}`);
1596
- assert.equal(status, 2);
1597
- // The guard's SHAPE, deterministically — the assertion above is the race.
1598
- // Stripped first, and this is not optional: against raw source a `die()`
1599
- // genuinely reverted to console.error with the good shape parked in a block
1600
- // comment passes 375/375 (measured), because a `^…/m` anchor matches inside
1601
- // the comment. Dropping this pin also drops the only pin on #176, which is
1602
- // what the refusal-text assertion this test used to carry was doing.
1603
- //
1604
- // #367 moved die() into arg.mjs's makeDie() — candidates.mjs's own source no
1605
- // longer contains the try/catch shape, so a lift pin against SCRIPT alone
1606
- // would now test a module nothing here calls (a source text-lift pin tests
1607
- // a COPY). The SHAPE is pinned here; the CALL SITE that wires it in is
1608
- // pinned for all seven consumers in arg.test.mjs, not here — pinning it here
1609
- // was vacuous: candidates.mjs can define a local console.error makeDie(),
1610
- // keep `const die = makeDie(NAME);` verbatim, and the suite stays green at
1611
- // 640/640 (measured). That line pins a NAME, not the module it resolves to.
1612
- //
1613
- // Each fragment anchored at a line start under /m — the old regex joined
1614
- // them with a bare `\s*` and no anchor at all. stripComments() blanks
1615
- // whole-line comments only (its own documented ceiling), so a trailing
1616
- // `code; // function die(msg) { try { writeSync(2,` survives stripping and
1617
- // satisfied the old regex with die() genuinely reverted (measured).
1618
- // Requiring each fragment at a line START refuses that — the decoy sits
1619
- // mid-line, and a whole-line decoy is blanked to "". `\s*^\s*` between the
1620
- // fragments rather than a literal `\n` on purpose: a future editor adding a
1621
- // comment line inside makeDie must not turn this pin red (measured — the
1622
- // adjacency form over-fired on exactly that). Do not terminate the lines
1623
- // with `$` either; that over-fires on a trailing comment (measured).
1624
- //
1625
- // #889/#1549: the loop that consumes writeSync's return value no longer
1626
- // lives in die() at all — it is arg.mjs's writeAll(), pinned ONCE at its
1627
- // own definition in arg.test.mjs. Three test files used to carry
1628
- // near-copies of that same loop regex, differing only in head line and fd,
1629
- // and keeping three hand-mirrored loops in step was what #1549 removed.
1630
- //
1631
- // What stays die()'s OWN, and is all this pins, is the wiring: the refusal
1632
- // goes through writeAll on fd 2, the template sits INSIDE the try, and
1633
- // process.exit(2) follows it unconditionally, outside the catch. That
1634
- // middle clause is load-bearing rather than decorative — hoisting the
1635
- // template out to writeAll's argument list puts msg's own string coercion
1636
- // back where a throwing msg escapes and skips the exit call, which is the
1637
- // #299/#328 inversion this file's die() promises never to repeat.
1638
- // Measured against that exact mutant: this pin reds, and the writeAll loop
1639
- // pin in arg.test.mjs stays green — the two discriminate different defects
1640
- // and neither substitutes for the other.
1641
- assert.match(
1642
- stripComments(readFileSync(ARG_MODULE, "utf8")),
1643
- /^\s*(?:return )?function die\(msg\) \{\s*^\s*try \{\s*^\s*writeAll\(2, `[^`]*`\);\s*^\s*\} catch \{\s*^\s*\}\s*^\s*process\.exit\(2\);/m,
1644
- );
1645
- });
1646
-
1647
- test("gh missing entirely is named as ENOENT — the shape that prints nothing at all", () => {
1648
- // The other side of `e.code ?? …`. Deleting the whole `e.code` half leaves
1649
- // the suite green without this: the jq test above only ever reaches the
1650
- // exit-status branch. ENOENT is the shape where die() is the SOLE diagnostic
1651
- // — nothing spawned, so no child stderr was forwarded to fall back on.
1652
- const { status, stderr } = run(
1653
- [ticket(11, "## What to build\n\nx\n")],
1654
- ["--require-label", "ready-for-agent"],
1655
- null,
1656
- { PATH: "/nonexistent" },
1657
- );
1658
- assert.equal(status, 2);
1659
- assert.match(stderr, /^candidates: gh issue list failed: ENOENT$/m);
1660
- });
1661
-
1662
- test("a signal-killed gh is named by its signal, not reported as `exit null`", () => {
1663
- // Third disjoint shape: `e.code` undefined AND `e.status` null, so a
1664
- // two-branch expression prints `exit null` and names nothing. Live for gh —
1665
- // the OOM killer on a large query, a SIGPIPE, a propagated Ctrl-C — and the
1666
- // child prints nothing on its way out, so this line is all the caller gets.
1667
- const { status, stderr } = runWithGh(`#!/bin/sh\nkill -TERM $$\n`);
1668
- assert.equal(status, 2);
1669
- assert.match(stderr, /^candidates: gh issue list failed: killed by SIGTERM$/m);
1670
- assert.doesNotMatch(stderr, /exit null/);
1671
- });
1672
-
1673
- test("a payload past the pipe buffer arrives whole — truncated JSON must never read as exit 0", () => {
1674
- // process.exit() discards queued writes, and stdout on a pipe (which is what
1675
- // spawnSync gives us, and what every real caller gives it) is async. 499 rows
1676
- // — one UNDER the default --limit, so refuseIfCapped never fires — measured
1677
- // 65,536 B and mid-JSON at exit 0: a corrupt payload typed as success, the
1678
- // one outcome this file's exit codes exist to make impossible.
1679
- //
1680
- // The pin is `rows`, which run() JSON.parses: truncation throws there. The
1681
- // byte assertion is what makes it meaningful — under the buffer this test
1682
- // passes against process.exit() too, so it must stay comfortably over.
1683
- const issues = Array.from({ length: 499 }, (_, i) => ({
1684
- number: i + 1,
1685
- title: "t".repeat(120),
1686
- labels: [{ name: "ready-for-agent" }, { name: "size:m" }],
1687
- body: "## What to build\n\ndepends on #7\n",
1688
- }));
1689
- const { status, stdout, rows } = run(issues);
1690
- assert.ok(stdout.length > 70_000, `payload must exceed the 64 KiB buffer, got ${stdout.length}`);
1691
- assert.equal(rows.length, 499);
1692
- assert.equal(status, 0);
1693
- });
1694
-
1695
- test("an empty queue is exit 1, not 2 — the query worked and there is no work", () => {
1696
- // The other half of the contract in this file's header. Every refusal above
1697
- // pins 2; nothing pinned 1, so a change spending 2 on an empty queue — the
1698
- // swap the header forbids — shipped green. Verified by mutation: flipping the
1699
- // `: 1` arm of `rows.length === 0 ? (allFilteredOut ? 3 : 1) : 0` to `: 2`
1700
- // fails this test, the #64 fallback-genuinely-empty test below, AND
1701
- // fleet-tick's "supply comes from candidates.mjs" — every reader of that arm,
1702
- // across both files. Before #64 the line read `rows.length === 0 ? 1 : 0`
1703
- // and this was indeed the only test that saw it; the exit-3 branch and its
1704
- // fallback test are what made that count stale.
1705
- const { status, stdout } = run([], ["--require-label", "ready-for-agent"]);
1706
- assert.equal(status, 1);
1707
- assert.equal(stdout.trim(), "[]");
1708
- });
1709
-
1710
- test("an all-filtered labeled queue with no fallback is exit 3, not 1 — the caller can tell filtered-empty from genuinely-empty (#64)", () => {
1711
- // No --allow-fallback, so this is a single pass: the labeled query returns
1712
- // rows and dropSpecs removes every one. Byte-identical to a genuinely empty
1713
- // queue on stdout; the exit code is the only place the two facts differ.
1714
- const { status, stdout } = run([
1715
- ticket(10, "## User Stories\n\n1. As a user…\n"),
1716
- ticket(11, "## Problem Statement\n\ny\n\n## User Stories\n\n2. As a user…\n"),
1717
- ]);
1718
- assert.equal(stdout.trim(), "[]");
1719
- assert.equal(status, 3);
1720
- });
1721
-
1722
- test("an all-filtered fallback pass is exit 3 even though the labeled pass was genuinely empty — the verdict is the fallback's own, not pass 1's (#64)", () => {
1723
- // Pass 1 (labeled) comes back with zero rows — genuinely empty, nothing to
1724
- // filter, so pass 1's own flag would be false. Pass 2 (fallback) comes back
1725
- // with rows and drops every one. If the two passes' flags were AND'd together
1726
- // instead of the fallback's replacing pass 1's outright, this would wrongly
1727
- // read pass 1's "false" through and report exit 1 — verified by mutation, and
1728
- // this is the only test in the suite that the AND breaks. OR is the
1729
- // complementary bug, which this test cannot see (false OR true is true
1730
- // either way); the test below is what pins that direction.
1731
- const { status, stdout, stderr } = run(
1732
- [],
1733
- ["--require-label", "ready-for-agent", "--allow-fallback"],
1734
- [
1735
- ticket(12, "## User Stories\n\n1. As a user…\n", ["ready-for-human"]),
1736
- ticket(13, "## Problem Statement\n\nz\n\n## User Stories\n\n2. As a user…\n", ["ready-for-human"]),
1737
- ],
1738
- );
1739
- assert.match(stderr, /retrying unfiltered/);
1740
- assert.equal(stdout.trim(), "[]");
1741
- assert.equal(status, 3);
1742
- });
1743
-
1744
- test("a labeled all-specs query that falls back to a genuinely empty unfiltered query is exit 1, not 3 — pass 1's drop must not leak into pass 2's verdict (#64)", () => {
1745
- // The edge case the ticket names explicitly. Pass 1 (labeled) has rows and
1746
- // drops every one as a spec — pass 1's own flag is true. Pass 2 (fallback)
1747
- // comes back with no rows at all — genuinely empty, nothing to filter. If
1748
- // pass 1's "true" carried forward (an OR instead of a plain reassignment),
1749
- // this would wrongly report exit 3 for a fallback that filtered nothing.
1750
- const { status, stdout, stderr } = run(
1751
- [ticket(10, "## User Stories\n\n1. As a user…\n")],
1752
- ["--require-label", "ready-for-agent", "--allow-fallback"],
1753
- [],
1754
- );
1755
- assert.match(stderr, /retrying unfiltered/);
1756
- assert.equal(stdout.trim(), "[]");
1757
- assert.equal(status, 1);
1758
- });
1759
-
1760
- test("candidates come back oldest first, whatever order gh returned them in", () => {
1761
- // gh defaults to created-desc, so newest-first is the realistic input.
1762
- const { rows } = run([
1763
- ticket(42, "## What to build\n\nc\n"),
1764
- ticket(19, "## What to build\n\nb\n"),
1765
- ticket(7, "## What to build\n\na\n"),
1766
- ]);
1767
- assert.deepEqual(rows.map((r) => r.n), [7, 19, 42]);
1768
- });
1769
-
1770
- // #175. The search term is built by interpolation, so the label's own
1771
- // characters are read by GitHub's query parser, not carried past it. The stub
1772
- // above models the one split that matters — an unquoted value ends at the
1773
- // first space. Neither test below is green on both trees, but they go red for
1774
- // different reasons. On the raw interpolation the first fails outright: the
1775
- // term resolves to `good`, the query matches nothing, and the empty answer
1776
- // surfaces as exit 1. The second's QUERY OUTCOME — status and rows — is
1777
- // identical either way, which is the whole point of it: quoting the label
1778
- // every current caller actually passes is a no-op for what comes back. Its
1779
- // wire-form assertion is the fix-dependent half, and reds on the raw
1780
- // interpolation exactly like the first.
1781
-
1782
- test("a multi-word label is quoted into the search term — unquoted, every word after the first is free text (#175)", () => {
1783
- // The ticket's headline symptom, and the one the exit-code contract makes
1784
- // expensive: the query silently did not mean what was asked, matched
1785
- // nothing, and its empty result read as exit 1 — "the query worked and
1786
- // there is no work" — against a queue that was not empty. Reachable in a
1787
- // repo that remapped its triage labels (docs/agents/triage-labels.md invites
1788
- // exactly that), where the AFK-ready role can be spelled `good first issue`.
1789
- const { status, rows } = run(
1790
- [ticket(11, "## What to build\n\nx\n", ["good first issue"])],
1791
- ["--require-label", "good first issue"],
1792
- null,
1793
- { LABEL_EXPECT: "good first issue" },
1794
- );
1795
- assert.equal(status, 0);
1796
- assert.deepEqual(rows.map((r) => r.n), [11]);
1797
- });
1798
-
1799
- test("a single-word label still resolves to itself — quoting must not disturb the only label every caller passes", () => {
1800
- // AC-3's other half, and the half a quoting fix can newly break: this term
1801
- // is what next-ticket/SKILL.md's `## 1. Candidates` step, shortlist.mjs (run-team/SKILL.md's
1802
- // **Build the Shortlist** step) and fleet-tick.mjs's `supply` all send, so a change that widened or narrowed it would
1803
- // empty the fleet's queue at exit 1 — #175's own defect, relocated. Measured
1804
- // against the live repo 2026-08-17: `label:ready-for-agent` and
1805
- // `label:"ready-for-agent"` return the same count, so the quoting is a no-op
1806
- // HERE, which is a fact about GitHub's parser and not one this stub can
1807
- // prove. What the stub can prove is the half that would break: the term
1808
- // still resolves to `ready-for-agent` and nothing else.
1809
- const { status, rows, stderr } = run(
1810
- [ticket(11, "## What to build\n\nx\n")],
1811
- ["--require-label", "ready-for-agent"],
1812
- null,
1813
- { LABEL_EXPECT: "ready-for-agent" },
1814
- );
1815
- assert.equal(status, 0);
1816
- assert.deepEqual(rows.map((r) => r.n), [11]);
1817
- // Anchored on the `--json` that follows, so this cannot be satisfied by a
1818
- // term with the label plus trailing free text — the exact shape #175 is.
1819
- assert.match(stderr, /--search -label:in-progress[^\n]* label:"ready-for-agent" --json/);
1820
- });
1821
-
1822
- test("a label with no space is quoted too — the wire form carries every label, not only the ones a space would break", () => {
1823
- // What this pins is the UNIFORMITY, not a second dangerous character. Space
1824
- // is how the ticket found the bug and — measured 2026-08-17 — it is the only
1825
- // character GitHub splits an unquoted value on: `label:auto:logs` and
1826
- // `label:"auto:logs"` agree on renovatebot/renovate, so `status:ready` would
1827
- // in fact survive unquoted. The fix quotes it anyway because quoting a value
1828
- // that needs none is a measured no-op, and one unconditional form beats a
1829
- // `/\s/` predicate that has to stay in step with GitHub's parser; quoting
1830
- // only when /\s/ matched would pass every other test in this suite unnoticed.
1831
- // Behaviour cannot pin this one — the stub's parser resolves the label under
1832
- // either form — so the wire form is the assertion.
1833
- const { stderr } = run(
1834
- [ticket(11, "## What to build\n\nx\n", ["status:ready"])],
1835
- ["--require-label", "status:ready"],
1836
- );
1837
- assert.match(stderr, /--search -label:in-progress[^\n]* label:"status:ready" --json/);
1838
- });
1839
-
1840
- test("a label containing a double quote refuses — an unrepresentable query must not run as an empty one", () => {
1841
- // What the fix itself newly makes possible: the term is `label:"<value>"`,
1842
- // so a `"` inside the value closes it early and the remainder becomes free
1843
- // text — #175's defect with the quoting applied. Refuse at exit 2 ("the
1844
- // query broke") rather than send it, because the alternative is the fail-open
1845
- // this file exists to prevent: a misparsed query whose empty result is
1846
- // indistinguishable from an empty queue. Escaping instead was not shipped
1847
- // because there is nothing to escape TO: GitHub does honour `\"` inside a
1848
- // qualifier, so a backslash cannot rescue a quote — it only moves the same
1849
- // misparse one character left, which is what the next test pins.
1850
- const { status, stderr } = run(
1851
- [ticket(11, "## What to build\n\nx\n")],
1852
- ["--require-label", 'say "hi"'],
1853
- );
1854
- assert.equal(status, 2);
1855
- assert.match(stderr, /^candidates: --require-label cannot contain a double quote/m);
1856
- // Refuses BEFORE gh, like every other value guard: a query that cannot be
1857
- // built correctly must not be sent at all.
1858
- assert.equal(queriesRun(stderr), 0);
1859
- });
1860
-
1861
- test("a label ending in a backslash refuses too — it eats the closing quote the fix adds", () => {
1862
- // The same term broken from the other side, and the side that leaves no
1863
- // trace. Measured against feigi/claude-config 2026-08-17: negating a label
1864
- // no issue carries is a no-op, so `-label:"zzz" label:"ready-for-agent"`
1865
- // returns the same count as `label:"ready-for-agent"` alone — but
1866
- // `-label:"zzz\" label:"ready-for-agent"` returns the count for
1867
- // `ready-for-agent` as FREE TEXT, the backslash having eaten the closing
1868
- // quote and swallowed the whole following qualifier. Doubling does not
1869
- // escape it. In the shape query() actually builds the label term comes
1870
- // LAST, so a trailing backslash leaves its value unterminated and the query
1871
- // answers zero rows at HTTP 200, no error — #175's silent empty, reached
1872
- // through the fix rather than around it. Without this case the widened
1873
- // guard is unpinned: narrowing it back to `"` alone leaves every other test
1874
- // in this file green.
1875
- const { status, stderr } = run(
1876
- [ticket(11, "## What to build\n\nx\n")],
1877
- ["--require-label", "trailing\\"],
1878
- );
1879
- assert.equal(status, 2);
1880
- assert.match(stderr, /^candidates: --require-label cannot contain a double quote or a backslash/m);
1881
- assert.equal(queriesRun(stderr), 0);
1882
- });
1883
-
1884
- // The other half of #173, and the half that produced the refused invocation
1885
- // above: the rule a controller reads lives in another file and was pinned by
1886
- // nothing, so it named a flag the script has never accepted.
1887
- const RUN_TEAM = readFileSync(join(import.meta.dirname, "..", "skills", "run-team", "SKILL.md"), "utf8");
1888
-
1889
- test("run-team's phase 0 rule names the flag candidates.mjs accepts", () => {
1890
- // The step-1 bullet alone. A slice any wider is vacuous for this claim: the
1891
- // `candidates.mjs` invocation in the same bullet already spells
1892
- // `--require-label` correctly, so a positive match anywhere in phase 0 stays
1893
- // green with the rule naming anything at all.
1894
- const step1 = between(RUN_TEAM, "1. **Build the Shortlist**", "\n2. ", "run-team/SKILL.md");
1895
- // Leading backtick, so this can only be satisfied by the RULE: in the command
1896
- // above it, `--require-label` is preceded by a line break, not a backtick.
1897
- // The gap is loose enough that rewording around `mandatory` stays green and
1898
- // tight enough that a different flag in the rule does not.
1899
- assert.match(
1900
- step1,
1901
- /`--require-label ready-for-agent`[\s\S]{0,20}mandatory/,
1902
- "run-team's mandatory-label rule no longer names --require-label",
1903
- );
1904
- // The positive pin cannot see a SECOND sentence naming the wrong flag, which
1905
- // is exactly what shipped. `--require-label` does not contain `--label`, so
1906
- // this needs no quoting to tell them apart.
1907
- assert.doesNotMatch(
1908
- step1,
1909
- /--label\b/,
1910
- "run-team's phase 0 step 1 names `--label`, which candidates.mjs refuses",
1911
- );
1912
- });
1913
-
1914
- // --- #240: the third carrier, and the one the pin above cannot see. The rule
1915
- // in run-team/SKILL.md is pinned; the design spec that rule descends from was
1916
- // not, and it still spelled the flag `--label`.
1917
- //
1918
- // The spelling was CORRECT the day that spec was written: at `02bfd59`, the
1919
- // commit that added it, `next-ticket` step 1 — the step the spec cites rather
1920
- // than restates — was a raw `gh issue list` reading "Add `--label
1921
- // ready-for-agent` first". `c3310fe` moved that query into this script the next
1922
- // day, and this script has read `arg("require-label")` in every commit of its
1923
- // life. That is why the spec carries a supersession note rather than a silent
1924
- // swap: these documents are dated records, and only a behavioural claim the
1925
- // code contradicts gets corrected in them.
1926
- const SPEC_RUN_TEAM_DESIGN = fileURLToPath(
1927
- new URL("../../docs/specs/2026-07-22-run-team-agent-fleet-design.md", import.meta.url),
1928
- );
1929
- const SPEC_COCKPIT_DESIGN = fileURLToPath(
1930
- new URL("../../docs/specs/2026-07-24-fleet-cockpit-design.md", import.meta.url),
1931
- );
1932
-
1933
- // Derived from the declaration, not restated from it: a rename of the accepted
1934
- // flag reddens here, which is the whole failure this ticket is an instance of —
1935
- // a hand-copied spelling that nothing made move with the script.
1936
- //
1937
- // Guarded at every step, and called from a test body rather than run at module
1938
- // scope, because an unguarded index here fails at IMPORT, and an import-time
1939
- // failure takes down every test in this file — the ones deriving the flag and
1940
- // the ones that have nothing to do with it alike — leaving one node-authored
1941
- // `TypeError` that names a file and a line but never names `OPTIONS` as the
1942
- // thing to look at. Measured on this tree by appending an unguarded
1943
- // module-scope `.match(...)[0]` to this file and one-lining `OPTIONS`:
1944
- // `node --test plugin/scripts/candidates.test.mjs` then reports one
1945
- // synthetic test, zero passes, and that TypeError.
1946
- //
1947
- // Reformatting `OPTIONS` — one-lining it, wrapping it in `Object.freeze`,
1948
- // anything that drops the literal `\n};` this regex needs — has zero
1949
- // behavioural effect, so what the guard buys is that such a reformat costs
1950
- // only the tests that derive the flag: measured by one-lining `OPTIONS` on
1951
- // this tree, those tests red with this guard's own `OPTIONS block no longer
1952
- // matches — update this test` message and every other test in this file still
1953
- // runs and passes.
1954
- //
1955
- // The `*label` match is global and required to be UNIQUE rather than indexed
1956
- // at the first hit, because indexing fails SILENTLY: an `"exclude-label"`
1957
- // declared above `"require-label"` would pin the specs below to a flag they do
1958
- // not mean, and a test passing for the wrong reason is this ticket's own defect
1959
- // class. Anchoring to the literal `require-label` would settle the ordering by
1960
- // restating the string this test exists to stop restating.
1961
- function declaredLabelFlag() {
1962
- const block = readFileSync(SCRIPT, "utf8").match(/const OPTIONS = \{[\s\S]*?\n\};/);
1963
- assert.ok(block, "candidates.mjs' OPTIONS block no longer matches — update this test");
1964
- const flags = [...block[0].matchAll(/"([a-z-]*label)":/g)];
1965
- assert.equal(
1966
- flags.length,
1967
- 1,
1968
- `candidates.mjs' OPTIONS declares ${flags.length} \`*label\` flags, not the 1 this test derives — update this test`,
1969
- );
1970
- return flags[0][1];
1971
- }
1972
-
1973
- // Which carriers of this flag's spelling carry a pin through this function,
1974
- // and what a rename owes the ones that deliberately do not (#1196).
1975
- //
1976
- // The line is LIVE versus FROZEN, never value-bearing versus bare. Every live
1977
- // carrier is pinned — a shipped skill or command, a live doc, and a contract
1978
- // table edited in place, which is live whatever date its file is named for —
1979
- // and a bare `candidates.mjs --<flag>` counts as much as one carrying a value:
1980
- // left behind by a rename, it still names a flag that no longer exists, in a
1981
- // file that is read as current. A rename edits each live carrier in place,
1982
- // with no note, and its pin reds until it does.
1983
- //
1984
- // Frozen records — an accepted ADR's body, a dated spec's narrative, a
1985
- // supersession note — stay unpinned, because a rename's correct handling of
1986
- // them never touches the line a pin would read, so a pin there could only red
1987
- // on a rename it cannot see handled. What a rename owes each kind:
1988
- // - a dated spec: a supersession note, never a silent swap. The old spelling
1989
- // now names an invocation this script refuses at exit 2, which is the
1990
- // behavioural contradiction #257 lets a dated record be corrected for —
1991
- // the note at the head of the run-team design spec (#240) is the shape.
1992
- // - an accepted ADR: an `Amended by #N: …` on its `**Status:**` line, its
1993
- // body left as ruled (AGENTS.md, *ADR amendments*).
1994
- // The one pin on a dated spec below stays: it protects a correction #240
1995
- // already made in place, rather than freezing a record.
1996
-
1997
- test("the run-team design spec's candidate-scan step names the flag this script declares", () => {
1998
- const DECLARED_LABEL_FLAG = declaredLabelFlag();
1999
- // The step alone, for the reason the run-team pin above gives: a match
2000
- // anywhere in the document is vacuous, and here it is worse than vacuous —
2001
- // the supersession note at the head cites `--label` deliberately, as the
2002
- // spelling this script refuses, so a file-wide negative would fail on the
2003
- // correction's own prose.
2004
- const step = between(
2005
- readFileSync(SPEC_RUN_TEAM_DESIGN, "utf8"),
2006
- "1. Candidate scan",
2007
- "\n2. ",
2008
- "the run-team design spec",
2009
- );
2010
-
2011
- assert.ok(
2012
- step.includes(`\`--${DECLARED_LABEL_FLAG} ready-for-agent\``),
2013
- `the spec's candidate-scan step no longer names --${DECLARED_LABEL_FLAG}, the label flag this script's OPTIONS declares`,
2014
- );
2015
- // `--require-label` does not contain `--label`, so this tells them apart with
2016
- // no quoting — same reasoning as the run-team pin above.
2017
- assert.doesNotMatch(
2018
- step,
2019
- /--label\b/,
2020
- "the spec's candidate-scan step names `--label`, which this script refuses at exit 2",
2021
- );
2022
- });
2023
-
2024
- // --- #851: the next-ticket carrier. `git grep -n -- --require-label docs
2025
- // plugin` lists every carrier of this flag's spelling, pinned and unpinned
2026
- // alike; which ones carry a pin, and why the rest deliberately do not, is the
2027
- // rule above `declaredLabelFlag()`. `next-ticket/SKILL.md` earns a pin here
2028
- // because it is what a solo caller reads to run the scan, and `--label` is
2029
- // what `gh issue list` accepts, so it is the plausible thing to write and the
2030
- // thing this script refuses at exit 2. Derived from `OPTIONS`, through the
2031
- // same guarded `declaredLabelFlag()` the run-team design-spec pin uses, so a
2032
- // rename of the flag reddens here instead of minting the next hand-copied
2033
- // spelling.
2034
- //
2035
- // What gets pinned is the TOKEN inside the candidate step, not the invocation
2036
- // that carries it. Measured, both legs: deleting that step's only fenced block
2037
- // and naming the flag in prose instead leaves the whole suite green, while
2038
- // deleting the same fence with NO mention of the flag reds with this pin's own
2039
- // message. So this holds "the flag named here is the one `OPTIONS` declares",
2040
- // never "there is a runnable command here" — a separate invariant #851 does
2041
- // not ask for and nothing else in this suite claims.
2042
- const NEXT_TICKET = readFileSync(join(import.meta.dirname, "..", "skills", "next-ticket", "SKILL.md"), "utf8");
2043
-
2044
- // Takes the text rather than reading the file, so the discrimination test below
2045
- // can put the same pin in front of the input it must red on AND the input it
2046
- // must not. A pin only shown to red is not shown to discriminate.
2047
- function assertScanRunsDeclaredFlag(text, where) {
2048
- const flag = declaredLabelFlag();
2049
- // The candidate step alone, for the reason both pins above give, and here
2050
- // with a live second reason: step 7 of this same skill says "Never fold
2051
- // `--label` into the create", where `--label` is `gh pr edit`'s own flag and
2052
- // correct (#375). A file-wide negative would fail on that sentence.
2053
- const step = between(text, "## 1. Candidates", "## 2. Dependencies", where);
2054
- assert.ok(
2055
- step.includes(`--${flag} ready-for-agent`),
2056
- `${where}'s candidate scan no longer runs the label flag candidates.mjs' OPTIONS declares`,
2057
- );
2058
- // `--require-label` does not contain `--label`, so this tells them apart with
2059
- // no quoting — same reasoning as the two pins above.
2060
- assert.doesNotMatch(
2061
- step,
2062
- /--label\b/,
2063
- `${where}'s candidate scan names \`--label\`, which candidates.mjs refuses at exit 2`,
2064
- );
2065
- }
2066
-
2067
- test("next-ticket's candidate scan runs the flag this script declares", () => {
2068
- assertScanRunsDeclaredFlag(NEXT_TICKET, "next-ticket/SKILL.md");
2069
- });
2070
-
2071
- test("the next-ticket pin reds on the refused spelling and stays green on text it must accept", () => {
2072
- // REDS on the swap it exists to catch — written through the declared flag
2073
- // rather than the literal, so a rename of the flag moves the mutant with it
2074
- // instead of leaving a mutant that tests nothing.
2075
- assert.throws(
2076
- () => assertScanRunsDeclaredFlag(NEXT_TICKET.replaceAll(`--${declaredLabelFlag()}`, "--label"), "the mutant"),
2077
- /the mutant's candidate scan/,
2078
- "the pin does not red when the candidate scan is given the spelling this script refuses",
2079
- );
2080
-
2081
- // GREEN, control 1: the live file, which carries a legitimate `--label`
2082
- // outside this slice. This is what separates a bounded pin from a whole-file
2083
- // one — the latter clears "it and only it reds" and still reds on prose it
2084
- // has no business reading.
2085
- assert.match(
2086
- NEXT_TICKET,
2087
- /Never fold `--label` into the create/,
2088
- "next-ticket/SKILL.md no longer carries a legitimate `--label` outside its candidate step — this control now proves nothing",
2089
- );
2090
-
2091
- // GREEN, control 2: a reflow. Rewrapping prose outside fenced blocks has zero
2092
- // behavioural effect on a skill, and a prose pin that reds on one costs its
2093
- // readers more than it holds.
2094
- //
2095
- // TWO reflows, in opposite directions, because one is not enough to stay
2096
- // honest: an unwrap applied to an already-unwrapped file changes nothing, and
2097
- // a control that asserts its own input changed would then red on the very
2098
- // maintainer edit it exists to bless — measured, by running this against an
2099
- // unwrapped copy of the skill. Requiring only that ONE of the two differs
2100
- // keeps the control non-vacuous whatever the file's current wrapping.
2101
- const outsideFences = (md, f) =>
2102
- md.split(/(```[\s\S]*?```)/).map((part, i) => (i % 2 ? part : f(part))).join("");
2103
- const unwrapped = outsideFences(NEXT_TICKET, (p) => p.replace(/(\S)\n(?=\S)/g, "$1 "));
2104
- const rewrapped = outsideFences(unwrapped, (p) => p.replace(/, (?=\S)/g, ",\n"));
2105
- assert.ok(
2106
- unwrapped !== NEXT_TICKET || rewrapped !== NEXT_TICKET,
2107
- "neither reflow changed next-ticket/SKILL.md — this control now proves nothing",
2108
- );
2109
- assertScanRunsDeclaredFlag(unwrapped, "an unwrapped next-ticket/SKILL.md");
2110
- assertScanRunsDeclaredFlag(rewrapped, "a rewrapped next-ticket/SKILL.md");
2111
- });
2112
-
2113
- // --- #1196: the remaining live carriers, one pin each, by the rule above
2114
- // `declaredLabelFlag()`. Each pin takes the text rather than reading the file,
2115
- // for the reason `assertScanRunsDeclaredFlag` gives, and each has a
2116
- // discrimination test that reds it on the refused spelling and keeps it green
2117
- // on the real text.
2118
- //
2119
- // The fleet-plugin design spec's script-surface table is a live contract,
2120
- // edited in place row by row, not a dated record — so its `candidates.mjs` row
2121
- // is pinned like any shipped file, and a rename edits it with no note.
2122
- const SPEC_PLUGIN_DESIGN = readFileSync(
2123
- fileURLToPath(new URL("../../docs/specs/2026-07-23-fleet-plugin-design.md", import.meta.url)),
2124
- "utf8",
2125
- );
2126
-
2127
- // The lookup candidates-exit3-prose.test.mjs's `candidatesRow()` makes, over
2128
- // a text passed in. A local copy rather than an import: importing a test
2129
- // module registers every test in it a second time in this process.
2130
- function candidatesRow(text, where) {
2131
- const line = text.split("\n").find((l) => l.startsWith("| `candidates.mjs` |"));
2132
- assert.ok(line, `${where}'s script-surface table no longer has a \`candidates.mjs\` row — update this test`);
2133
- return line;
2134
- }
2135
-
2136
- // The declared flag as a whole token: `--<flag>` with nothing word-like after
2137
- // it, so `--<flag>s` or `--<flag>-x` is not it.
2138
- const labelFlagToken = () => new RegExp(`--${declaredLabelFlag()}(?![\\w-])`);
2139
-
2140
- // Two tokens, both pinned: the In cell's synopsis and the Non-zero cell's
2141
- // exit-2 cause. Either one alone renamed is a refused spelling in the row a
2142
- // reader copies, and a positive on one cannot see the other. Each positive
2143
- // reads the TOKEN inside its own cell, never the cell's wording around it
2144
- // (#1196: "pin the token, not … the synopsis row's shape"), so a new
2145
- // placeholder or a reworded exit-2 cause stays green.
2146
- function assertRowNamesDeclaredFlag(text, where) {
2147
- const flag = declaredLabelFlag();
2148
- const token = labelFlagToken();
2149
- const row = candidatesRow(text, where);
2150
- const [, , inCell, , nonZeroCell] = row.split("|");
2151
- assert.match(
2152
- inCell,
2153
- token,
2154
- `${where}'s candidates row no longer names --${flag} in its In cell, the label flag candidates.mjs' OPTIONS declares`,
2155
- );
2156
- assert.match(
2157
- nonZeroCell ?? "",
2158
- token,
2159
- `${where}'s candidates row no longer names --${flag} in its Non-zero cell, where \`--allow-fallback\` needs it`,
2160
- );
2161
- // The row carries no `gh` invocation, so nothing in it spells `--label`
2162
- // correctly and a row-wide negative is safe.
2163
- assert.doesNotMatch(row, /--label\b/, `${where}'s candidates row names \`--label\`, which candidates.mjs refuses at exit 2`);
2164
- }
2165
-
2166
- test("the fleet-plugin design spec's candidates row names the flag this script declares", () => {
2167
- assertRowNamesDeclaredFlag(SPEC_PLUGIN_DESIGN, "the fleet-plugin design spec");
2168
- });
2169
-
2170
- // The spec's text with its `candidates.mjs` row's cells rewritten by `edit`.
2171
- // Mutants and controls are built from the cells rather than from quoted row
2172
- // wording, so a row edit that keeps the token cannot strand them.
2173
- function withRowCells(text, edit) {
2174
- const row = candidatesRow(text, "the fleet-plugin design spec");
2175
- const cells = row.split("|");
2176
- edit(cells);
2177
- return text.replace(row, cells.join("|"));
2178
- }
2179
-
2180
- test("the candidates-row pin reds on either token renamed alone and stays green on the real row", () => {
2181
- const flag = declaredLabelFlag();
2182
- for (const [cell, which] of [
2183
- [2, "the In cell"],
2184
- [4, "the Non-zero cell"],
2185
- ]) {
2186
- const mutant = withRowCells(SPEC_PLUGIN_DESIGN, (cells) => {
2187
- cells[cell] = cells[cell].replace(labelFlagToken(), "--label");
2188
- });
2189
- assert.notEqual(mutant, SPEC_PLUGIN_DESIGN, `${which} no longer carries --${flag} — this mutant now tests nothing`);
2190
- assert.throws(
2191
- () => assertRowNamesDeclaredFlag(mutant, "the mutant"),
2192
- /the mutant's candidates row/,
2193
- `the pin does not red when only ${which} is given the spelling this script refuses`,
2194
- );
2195
- }
2196
- // GREEN on a row that keeps the flag's spelling but not its wording — a new
2197
- // placeholder, a reworded exit-2 cause — the inputs a shape pin wrongly
2198
- // refuses.
2199
- for (const [cell, text, which] of [
2200
- [2, ` \`[--${flag} LABEL] [--limit N]\` `, "the In cell's placeholder"],
2201
- [4, ` exit 2 when \`--allow-fallback\` is given unless \`--${flag}\` is `, "the Non-zero cell's wording"],
2202
- ]) {
2203
- assertRowNamesDeclaredFlag(
2204
- withRowCells(SPEC_PLUGIN_DESIGN, (cells) => {
2205
- cells[cell] = text;
2206
- }),
2207
- `the spec with ${which} reworded`,
2208
- );
2209
- }
2210
- assertRowNamesDeclaredFlag(SPEC_PLUGIN_DESIGN, "the real spec");
2211
- });
2212
-
2213
- // The two component docs, each sliced to its candidate-scan list item. Matched
2214
- // through `phrase()` so a rewrap of the item stays green; the negative names
2215
- // `candidates.mjs` so it reads the invocation, not any `--label` near it.
2216
- const COMPONENT_SCANS = [
2217
- ["next-ticket.md", "1. **List candidates**", "\n2. **Filter.**"],
2218
- ["supply-and-shortlist.md", "1. **Scan.**", "\n2. **Filter**"],
2219
- ].map(([file, from, to]) => ({
2220
- where: `docs/components/${file}`,
2221
- text: readFileSync(fileURLToPath(new URL(`../../docs/components/${file}`, import.meta.url)), "utf8"),
2222
- from,
2223
- to,
2224
- }));
2225
-
2226
- function assertComponentScanNamesDeclaredFlag({ text, from, to }, where) {
2227
- const flag = declaredLabelFlag();
2228
- const item = between(text, from, to, where);
2229
- assert.match(
2230
- item,
2231
- phrase(`candidates.mjs --${flag} ready-for-agent`),
2232
- `${where}'s candidate-scan item no longer runs the label flag candidates.mjs' OPTIONS declares`,
2233
- );
2234
- assert.doesNotMatch(
2235
- item,
2236
- /candidates\.mjs\s+--label\b/,
2237
- `${where}'s candidate-scan item names \`candidates.mjs --label\`, which candidates.mjs refuses at exit 2`,
2238
- );
2239
- }
2240
-
2241
- for (const doc of COMPONENT_SCANS) {
2242
- test(`${doc.where}'s candidate scan names the flag this script declares`, () => {
2243
- assertComponentScanNamesDeclaredFlag(doc, doc.where);
2244
- });
2245
-
2246
- test(`the ${doc.where} pin reds on the refused spelling and stays green on the real text`, () => {
2247
- assert.throws(
2248
- () =>
2249
- assertComponentScanNamesDeclaredFlag(
2250
- { ...doc, text: doc.text.replaceAll(`--${declaredLabelFlag()}`, "--label") },
2251
- "the mutant",
2252
- ),
2253
- /the mutant's candidate-scan item/,
2254
- "the pin does not red when the candidate-scan item is given the spelling this script refuses",
2255
- );
2256
- assertComponentScanNamesDeclaredFlag(doc, doc.where);
2257
- });
2258
- }
2259
-
2260
- // review-and-fix step 5 names the flag BARE — `candidates.mjs --<flag>`, no
2261
- // value — and is pinned all the same: live, shipped, read as current. The
2262
- // negative names `candidates.mjs` rather than `/--label\b/`, because this same
2263
- // step correctly runs `gh issue create --label`, `gh`'s own flag.
2264
- const REVIEW_AND_FIX = readFileSync(join(import.meta.dirname, "..", "commands", "review-and-fix.md"), "utf8");
2265
-
2266
- function assertDeferralStepNamesDeclaredFlag(text, where) {
2267
- const flag = declaredLabelFlag();
2268
- const step5 = between(text, "5. File each deferred finding", "\n6. Diff-check green", `${where} step 5`);
2269
- assert.match(
2270
- step5,
2271
- new RegExp(`${phrase(`candidates.mjs --${flag}`).source}(?![\\w-])`),
2272
- `${where}'s step 5 no longer names the label flag candidates.mjs' OPTIONS declares`,
2273
- );
2274
- assert.doesNotMatch(
2275
- step5,
2276
- /candidates\.mjs\s+--label\b/,
2277
- `${where}'s step 5 names \`candidates.mjs --label\`, which candidates.mjs refuses at exit 2`,
2278
- );
2279
- }
2280
-
2281
- test("review-and-fix's deferral step names the flag this script declares", () => {
2282
- assertDeferralStepNamesDeclaredFlag(REVIEW_AND_FIX, "review-and-fix.md");
2283
- });
2284
-
2285
- test("the review-and-fix pin reds on the refused spelling and stays green beside gh's own --label", () => {
2286
- assert.throws(
2287
- () =>
2288
- assertDeferralStepNamesDeclaredFlag(REVIEW_AND_FIX.replaceAll(`--${declaredLabelFlag()}`, "--label"), "the mutant"),
2289
- /the mutant's step 5/,
2290
- "the pin does not red when step 5 is given the spelling this script refuses",
2291
- );
2292
- // GREEN on the real step, which carries a legitimate `--label` — the input
2293
- // a looser negative would wrongly refuse.
2294
- assert.match(
2295
- between(REVIEW_AND_FIX, "5. File each deferred finding", "\n6. Diff-check green", "review-and-fix.md step 5"),
2296
- /gh issue create --label/,
2297
- "review-and-fix step 5 no longer carries gh's own `--label` — this control now proves nothing",
2298
- );
2299
- assertDeferralStepNamesDeclaredFlag(REVIEW_AND_FIX, "review-and-fix.md");
2300
- });
2301
-
2302
- test("the cockpit spec's gh invocation keeps the flag gh accepts", () => {
2303
- // The other half of the same class, and the one a fix for it can break:
2304
- // `--label` is correct for `gh issue list`, so a sweep for the wrong spelling
2305
- // that did not distinguish the caller would rewrite a working invocation into
2306
- // a broken one. This is the input the check must ACCEPT.
2307
- assert.match(
2308
- readFileSync(SPEC_COCKPIT_DESIGN, "utf8"),
2309
- /`gh issue list --label ready-for-agent`/,
2310
- "the cockpit spec's `gh issue list --label ready-for-agent` changed — `--label` is gh's own flag and is correct there, so a sweep for this script's wrong spelling must leave it alone",
2311
- );
2312
- });