rungate 0.8.22 → 0.8.24
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.
- package/gates/.gate-salt +1 -1
- package/gates/gate-executor.ts +316 -68
- package/gates/orchestrator.test.ts +108 -2
- package/gates/orchestrator.ts +25 -5
- package/gates/prompts/prove-reproducer.md +14 -1
- package/gates/prompts.sha256 +1 -1
- package/gates/prove.test.ts +6 -1
- package/gates/schema.ts +21 -1
- package/gates/workflow.test.ts +12 -0
- package/lib/ai-review-suppression.ts +26 -2
- package/lib/artifact-schema.ts +37 -0
- package/lib/cli-commands.ts +10 -0
- package/lib/scaffold/ai-review-step.ts +14 -0
- package/lib/scaffold/intake-workflow.ts +120 -0
- package/lib/scaffold/owned-files.ts +10 -0
- package/lib/scaffold/steps.ts +14 -1
- package/lib/stage-ladder.ts +62 -0
- package/package.json +2 -1
- package/scripts/ai-review-findings.ts +6 -2
- package/scripts/intake.ts +101 -0
- package/config/prove-evidence.schema.json +0 -83
package/gates/prompts.sha256
CHANGED
|
@@ -1,3 +1,3 @@
|
|
|
1
1
|
ba56e1fcb00c7a9bc1f2b05beca701cb264785e017cd2bf54a2dd07cdde331df ac-adversary.md
|
|
2
2
|
7d4ff0ef4b787e68b59643c11ba51e62a9411f5e24509b97855a337598c67508 evidence-validator.md
|
|
3
|
-
|
|
3
|
+
6f15ff6891118f749894f55f18fa21234211f993cf6eadfb3c90561a64b91691 prove-reproducer.md
|
package/gates/prove.test.ts
CHANGED
|
@@ -1,7 +1,8 @@
|
|
|
1
1
|
import { test, expect, describe } from "bun:test";
|
|
2
2
|
import { readFileSync, existsSync, writeFileSync } from "fs";
|
|
3
3
|
import { execSync } from "child_process";
|
|
4
|
-
import { beforeEvidenceFor, comparisonSummaryFor, readBeforeState } from "./gate-executor";
|
|
4
|
+
import { beforeEvidenceFor, comparisonSummaryFor, readBeforeState, validateArtifactSchema } from "./gate-executor";
|
|
5
|
+
import { ProveEvidenceSchema } from "./schema";
|
|
5
6
|
|
|
6
7
|
const TEST_DIR = process.env.TEST_WORK_DIR || "";
|
|
7
8
|
const PROJECT_ROOT = process.env.PROJECT_ROOT || process.cwd();
|
|
@@ -145,6 +146,10 @@ describe("prove: run evidence commands from workflow-state", () => {
|
|
|
145
146
|
criteriaResults,
|
|
146
147
|
};
|
|
147
148
|
|
|
149
|
+
// #2254 / SC-2770: checked against the one contract before it is written,
|
|
150
|
+
// as B3's and prove.js's are. A required field this writer stops emitting
|
|
151
|
+
// fails here, with the field named, instead of in whatever reads it next.
|
|
152
|
+
expect(validateArtifactSchema(evidence, ProveEvidenceSchema)).toEqual([]);
|
|
148
153
|
writeFileSync(`${TEST_DIR}/prove-evidence.json`, JSON.stringify(evidence, null, 2));
|
|
149
154
|
expect(verdict).toBe("PROVEN");
|
|
150
155
|
});
|
package/gates/schema.ts
CHANGED
|
@@ -1193,7 +1193,10 @@ export const ShipEvidenceSchema = z.object({
|
|
|
1193
1193
|
status: z.enum(["PASS", "FAIL"]),
|
|
1194
1194
|
evidence: z.array(z.object({
|
|
1195
1195
|
criterion: z.string(),
|
|
1196
|
-
|
|
1196
|
+
// PENDING is a legitimate ship-time state (#2255): an OUTCOME criterion
|
|
1197
|
+
// waits for Quinn or /prove, and the verify gate passes it through
|
|
1198
|
+
// (checkVerifyPendingACs). The writer copies the criterion's verdict.
|
|
1199
|
+
result: z.enum(["PASS", "FAIL", "SKIP", "PENDING"]),
|
|
1197
1200
|
detail: z.any().optional(),
|
|
1198
1201
|
})),
|
|
1199
1202
|
}),
|
|
@@ -1252,9 +1255,26 @@ const BeforeEvidenceSchema = z.union([
|
|
|
1252
1255
|
}),
|
|
1253
1256
|
]);
|
|
1254
1257
|
|
|
1258
|
+
/**
|
|
1259
|
+
* One row of `prove-evidence.json`'s `criteriaResults` (#1559). The shape all
|
|
1260
|
+
* three writers already emit — B3's reproducer, `gates/prove.test.ts` and
|
|
1261
|
+
* `workflows/prove.js` — and the one the deleted
|
|
1262
|
+
* `config/prove-evidence.schema.json` declared. SKIP is the reproducer's
|
|
1263
|
+
* verdict for a UI criterion it has no browser to check.
|
|
1264
|
+
*/
|
|
1265
|
+
const CriterionResultSchema = z.object({
|
|
1266
|
+
scId: z.string(),
|
|
1267
|
+
verdict: z.enum(["PASS", "FAIL", "SKIP"]),
|
|
1268
|
+
evidence: z.string(),
|
|
1269
|
+
});
|
|
1270
|
+
|
|
1255
1271
|
export const ProveEvidenceSchema = z.object({
|
|
1256
1272
|
issueNumber: z.number(),
|
|
1257
1273
|
verdict: z.enum(["PROVEN", "UNPROVEN", "INCONCLUSIVE"]),
|
|
1274
|
+
// Required, but not `.min(1)`: an INCONCLUSIVE with no criteria is a record
|
|
1275
|
+
// `gates/prove.test.ts` legitimately writes. That a PROVEN needs a PASS row
|
|
1276
|
+
// is `buildProveEvidence`'s rule, because it is about two fields at once.
|
|
1277
|
+
criteriaResults: z.array(CriterionResultSchema),
|
|
1258
1278
|
commitSHA: z.string(),
|
|
1259
1279
|
capturedAt: z.string(),
|
|
1260
1280
|
beforeEvidence: BeforeEvidenceSchema,
|
package/gates/workflow.test.ts
CHANGED
|
@@ -453,6 +453,18 @@ export function verifyFanoutViolations(label: string, val: unknown, tier: unknow
|
|
|
453
453
|
`is an absence with extra steps (#126)`,
|
|
454
454
|
];
|
|
455
455
|
}
|
|
456
|
+
// #2259: the agent ran and its reply could not be read. Before, that reply
|
|
457
|
+
// left no record, which the absence branch above refused, so the ship
|
|
458
|
+
// outcome is unchanged. What changes is that the refusal names the cause
|
|
459
|
+
// instead of failing on a missing `gameable` further down.
|
|
460
|
+
if ((val as { status?: unknown }).status === "UNMEASURED") {
|
|
461
|
+
const reason = (val as { reason?: unknown }).reason;
|
|
462
|
+
return [
|
|
463
|
+
`gates.verify.${label} is UNMEASURED — the agent ran and its reply could not be read` +
|
|
464
|
+
(typeof reason === "string" && reason.trim() !== "" ? `: ${reason}` : "") +
|
|
465
|
+
`. An unread reply is not an approval (#126, #2259).`,
|
|
466
|
+
];
|
|
467
|
+
}
|
|
456
468
|
return [];
|
|
457
469
|
}
|
|
458
470
|
|
|
@@ -224,8 +224,19 @@ export function aiReviewSuppressionGate(inputs: SuppressionInputs): AiReviewSupp
|
|
|
224
224
|
const a = anchorOnlyKey(k);
|
|
225
225
|
if (a !== null) anchors.add(a);
|
|
226
226
|
}
|
|
227
|
+
// #2257: how many findings cite a line, and how many of those lines could
|
|
228
|
+
// be read. Read from the SAME calls the keys are built from, not a second
|
|
229
|
+
// pass, so the count cannot disagree with what the keys saw.
|
|
230
|
+
let located = 0;
|
|
231
|
+
let readable = 0;
|
|
227
232
|
const repeat = candidates.map(c => {
|
|
228
|
-
const
|
|
233
|
+
const cited = c as ReviewFinding;
|
|
234
|
+
const lineText = inputs.lineTextFor(cited);
|
|
235
|
+
if (typeof cited?.file === "string" && Number.isInteger(cited?.line) && (cited.line as number) >= 1) {
|
|
236
|
+
located++;
|
|
237
|
+
if (lineText !== undefined) readable++;
|
|
238
|
+
}
|
|
239
|
+
const key = carryKey(c, lineText);
|
|
229
240
|
// Unkeyable CARRIES. It cannot be shown to be a repeat, and the
|
|
230
241
|
// direction this is allowed to be wrong in is toward reporting.
|
|
231
242
|
if (key === null) return null;
|
|
@@ -282,7 +293,20 @@ export function aiReviewSuppressionGate(inputs: SuppressionInputs): AiReviewSupp
|
|
|
282
293
|
// this whole increment is a correction for.
|
|
283
294
|
return null;
|
|
284
295
|
});
|
|
285
|
-
|
|
296
|
+
// #2257. A gate that could read no cited line built every key from a line
|
|
297
|
+
// NUMBER, which never equals a key an earlier round recorded from line
|
|
298
|
+
// TEXT. "Nothing to suppress" from that gate is a reading of nothing, and
|
|
299
|
+
// it printed the same words as a gate that read every line (#1872: 0 of
|
|
300
|
+
// 420 in CI). Byte-key matches above still stand; this says the anchor
|
|
301
|
+
// half was blind.
|
|
302
|
+
const unmeasured =
|
|
303
|
+
located > 0 && readable === 0 && prior.keys.length > 0
|
|
304
|
+
? `no cited line could be read for any of this round's ${located} finding(s) that cite one, ` +
|
|
305
|
+
`so no key could match one an earlier round recorded from the line's text, and a ` +
|
|
306
|
+
`restated finding could not be recognised. The checkout the gate reads is probably not ` +
|
|
307
|
+
`the one under review (#2257).`
|
|
308
|
+
: null;
|
|
309
|
+
return { repeat, unmeasured };
|
|
286
310
|
};
|
|
287
311
|
}
|
|
288
312
|
|
|
@@ -0,0 +1,37 @@
|
|
|
1
|
+
/**
|
|
2
|
+
* Check a gate artifact against its declared schema, before it is written
|
|
3
|
+
* (#1559, #2255, SC-2770).
|
|
4
|
+
*
|
|
5
|
+
* Lives in `lib/` rather than beside its first caller in
|
|
6
|
+
* `gates/gate-executor.ts` because `gates/orchestrator.ts` needs it too, and
|
|
7
|
+
* gate-executor already imports orchestrator. `gates/gate-executor.ts`
|
|
8
|
+
* re-exports both names, so there is one definition and every existing import
|
|
9
|
+
* keeps working.
|
|
10
|
+
*/
|
|
11
|
+
|
|
12
|
+
/**
|
|
13
|
+
* Anything with zod's `safeParse` shape. Structural, so callers keep no
|
|
14
|
+
* direct zod import and a test can hand it a stricter copy of a schema.
|
|
15
|
+
*/
|
|
16
|
+
export interface ArtifactSchema {
|
|
17
|
+
safeParse(value: unknown): {
|
|
18
|
+
success: boolean;
|
|
19
|
+
error?: { issues: ReadonlyArray<{ path: ReadonlyArray<PropertyKey>; message: string }> };
|
|
20
|
+
};
|
|
21
|
+
}
|
|
22
|
+
|
|
23
|
+
/**
|
|
24
|
+
* One `path: message` line per violation, `[]` when it validates.
|
|
25
|
+
*
|
|
26
|
+
* `config/prove-evidence.schema.json` declared a contract nothing read, so
|
|
27
|
+
* writer and contract drifted apart: B3's artifact failed both the JSON
|
|
28
|
+
* schema and the zod one, and no check said so. A refusal that names no
|
|
29
|
+
* issue is still a refusal — an empty list there would read as a pass.
|
|
30
|
+
*/
|
|
31
|
+
export function validateArtifactSchema(artifact: unknown, schema: ArtifactSchema): string[] {
|
|
32
|
+
const result = schema.safeParse(artifact);
|
|
33
|
+
if (result.success) return [];
|
|
34
|
+
const issues = result.error?.issues ?? [];
|
|
35
|
+
if (issues.length === 0) return ["(root): the schema refused the artifact without naming a field"];
|
|
36
|
+
return issues.map((i) => `${i.path.length > 0 ? i.path.map(String).join(".") : "(root)"}: ${i.message}`);
|
|
37
|
+
}
|
package/lib/cli-commands.ts
CHANGED
|
@@ -299,6 +299,16 @@ export const CLI_COMMANDS: Record<string, CliCommand> = {
|
|
|
299
299
|
"rungate unaccounted-work [--project-root <dir>] [--repo <owner/name>] [--issues-fixture <file.json>] [--json] " +
|
|
300
300
|
"(exit 1 when work is unaccounted, 2 when a source could not be read)",
|
|
301
301
|
},
|
|
302
|
+
// #2248: the command #2233 (fix-commits on push to main) and #2234
|
|
303
|
+
// (gate-refusals at the end of a ship run) call, per #1126's decision that
|
|
304
|
+
// each intake source runs on its own trigger.
|
|
305
|
+
intake: {
|
|
306
|
+
script: "scripts/intake.ts",
|
|
307
|
+
summary: "Run each intake source and print its reading — MEASURED with a count, or UNMEASURED naming why",
|
|
308
|
+
usage:
|
|
309
|
+
"rungate intake [--project-root <dir>] [--source <id>] [--json] " +
|
|
310
|
+
"(advisory: an unreadable source prints UNMEASURED and exits 0; exit 2 on a bad argument)",
|
|
311
|
+
},
|
|
302
312
|
// Reads a `gh pr view`-shaped payload on stdin, which the dispatcher passes
|
|
303
313
|
// through (`stdio: "inherit"`), or `--fixture` offline.
|
|
304
314
|
"audit-closing-references": {
|
|
@@ -1822,6 +1822,20 @@ export const AI_REVIEW_LEDGER_SNAPSHOT_BASENAME = "ai-review-ledger-snapshot.txt
|
|
|
1822
1822
|
*/
|
|
1823
1823
|
export const AI_REVIEW_THREADS_PATH_VAR = "AI_REVIEW_THREADS_PATH";
|
|
1824
1824
|
|
|
1825
|
+
/**
|
|
1826
|
+
* The checkout the adjudicator reads cited lines from (#2257).
|
|
1827
|
+
*
|
|
1828
|
+
* The adjudicator runs from the payload's own directory, outside the
|
|
1829
|
+
* checkout, because `bunx` resolves bunfig.toml, .npmrc and package.json by
|
|
1830
|
+
* walking up from its working directory, so the PR under review must not be
|
|
1831
|
+
* the tree it starts in. That made `process.cwd()` the payload directory,
|
|
1832
|
+
* the suppression gate could read no cited line, and the anchor rule matched
|
|
1833
|
+
* 0 of 420 round-2+ findings in CI (#1872). The root has to arrive by name.
|
|
1834
|
+
* `scripts/ai-review-findings.ts` falls back to its working directory when
|
|
1835
|
+
* this is unset, which is what a step generated before it was set does.
|
|
1836
|
+
*/
|
|
1837
|
+
export const AI_REVIEW_CHECKOUT_ROOT_VAR = "AI_REVIEW_CHECKOUT_ROOT";
|
|
1838
|
+
|
|
1825
1839
|
/** The basename inside the findings directory. One spelling. */
|
|
1826
1840
|
export const AI_REVIEW_THREADS_BASENAME = "ai-review-threads.json";
|
|
1827
1841
|
|
|
@@ -0,0 +1,120 @@
|
|
|
1
|
+
/**
|
|
2
|
+
* `.github/workflows/intake.yml`, generated — #2233.
|
|
3
|
+
*
|
|
4
|
+
* Jason decided on #1126 (2026-11-15) that each intake source runs on its own
|
|
5
|
+
* trigger, not on scaffold: "scaffold" reads as a one-time setup step, so
|
|
6
|
+
* nobody re-runs it to re-mine, and intake should run when its data appears.
|
|
7
|
+
* For `fix-commits` the data is the history of the default branch, and the
|
|
8
|
+
* trigger is a push to it.
|
|
9
|
+
*
|
|
10
|
+
* ## Full history, or the reading is of the clone
|
|
11
|
+
*
|
|
12
|
+
* `mineFixCommits` reports `SHALLOW_HISTORY` on a shallow clone, and
|
|
13
|
+
* `actions/checkout` clones to depth 1 unless told otherwise. That depth is
|
|
14
|
+
* the only reason the source read UNMEASURED on a runner, so the checkout
|
|
15
|
+
* sets `fetch-depth: 0`.
|
|
16
|
+
*
|
|
17
|
+
* ## Advisory, and never silent
|
|
18
|
+
*
|
|
19
|
+
* `rungate intake` exits 0 on every reading (#2248), so a reading never turns
|
|
20
|
+
* the run red: an unreadable source says nothing about the push that fired
|
|
21
|
+
* it. UNMEASURED is printed as the command prints it and raised as a warning
|
|
22
|
+
* annotation, so it is visible on the run without being mistaken for zero.
|
|
23
|
+
* A non-zero exit is the command refusing its arguments or `bunx` failing,
|
|
24
|
+
* which is the tool broken, and the step goes red on it.
|
|
25
|
+
*
|
|
26
|
+
* ## It calls the published CLI
|
|
27
|
+
*
|
|
28
|
+
* A consumer has `node_modules/rungate`, not `scripts/`, so the step runs
|
|
29
|
+
* `bunx rungate@<version> intake`, at the generating version like every other
|
|
30
|
+
* pin. `intake` first shipped in 0.8.23 (#2250).
|
|
31
|
+
*/
|
|
32
|
+
|
|
33
|
+
/** The file this module generates, under `.github/workflows/`. */
|
|
34
|
+
export const INTAKE_WORKFLOW_FILE = "intake.yml";
|
|
35
|
+
|
|
36
|
+
/** The one job in it. */
|
|
37
|
+
export const INTAKE_JOB = "fix-commit-intake";
|
|
38
|
+
|
|
39
|
+
/**
|
|
40
|
+
* The CLI command the step runs. `lib/cli-commands.ts` must register it, or
|
|
41
|
+
* the workflow calls a command `bin/rungate.ts` refuses (#675's defect).
|
|
42
|
+
*/
|
|
43
|
+
export const INTAKE_COMMAND = "intake";
|
|
44
|
+
|
|
45
|
+
/**
|
|
46
|
+
* The step's `run:` body. Exported so `test/intake-workflow.test.ts` executes
|
|
47
|
+
* THIS text against a stub `bunx`, rather than a paraphrase.
|
|
48
|
+
*
|
|
49
|
+
* The output is captured so it can be both printed and inspected. Under
|
|
50
|
+
* `bash -e` a failing command substitution in an assignment stops the step,
|
|
51
|
+
* which is the red a broken tool should produce.
|
|
52
|
+
*/
|
|
53
|
+
export function buildIntakeScript(version: string): string {
|
|
54
|
+
return `out=$(bunx rungate@${version} ${INTAKE_COMMAND} --source fix-commits)
|
|
55
|
+
printf '%s\\n' "$out"
|
|
56
|
+
case "$out" in
|
|
57
|
+
*UNMEASURED*)
|
|
58
|
+
echo "::warning::UNMEASURED — the fix-commits intake reading could not be taken on this run; the reason is printed above. This is not zero candidates."
|
|
59
|
+
;;
|
|
60
|
+
esac
|
|
61
|
+
`;
|
|
62
|
+
}
|
|
63
|
+
|
|
64
|
+
export interface IntakeWorkflowOptions {
|
|
65
|
+
/** The `runs-on:` value, already JSON-quoted by the caller. */
|
|
66
|
+
runsOn: string;
|
|
67
|
+
/** The pinned `actions/checkout` reference. */
|
|
68
|
+
checkout: string;
|
|
69
|
+
/** The pinned `oven-sh/setup-bun` reference. */
|
|
70
|
+
setupBun: string;
|
|
71
|
+
/** The `with:` body of that step, at ten spaces — the project's single bun declaration (#622). */
|
|
72
|
+
setupBunWith: string;
|
|
73
|
+
/** The step's `env:` block, from `setupBunEnv` (#2039). */
|
|
74
|
+
setupBunEnv: string;
|
|
75
|
+
/** The published harness version the step runs through `bunx`. */
|
|
76
|
+
version: string;
|
|
77
|
+
}
|
|
78
|
+
|
|
79
|
+
/** The whole `intake.yml` body. */
|
|
80
|
+
export function buildIntakeWorkflow(opts: IntakeWorkflowOptions): string {
|
|
81
|
+
const indent = (text: string, pad: string) =>
|
|
82
|
+
text
|
|
83
|
+
.trimEnd()
|
|
84
|
+
.split("\n")
|
|
85
|
+
.map(line => (line === "" ? "" : `${pad}${line}`))
|
|
86
|
+
.join("\n");
|
|
87
|
+
|
|
88
|
+
return `# Managed by rungate — do not edit. Customize the ci section of .claude/rungate/config.json.
|
|
89
|
+
#
|
|
90
|
+
# Mines the default branch's fix commits into rule candidates on every push to
|
|
91
|
+
# it (#2233). Jason's #1126 decision: each intake source runs on its own
|
|
92
|
+
# trigger, never on scaffold. Advisory: a reading never fails the run, and an
|
|
93
|
+
# UNMEASURED one is raised as a warning rather than read as zero.
|
|
94
|
+
name: Intake — fix commits
|
|
95
|
+
|
|
96
|
+
on:
|
|
97
|
+
push:
|
|
98
|
+
branches: [main]
|
|
99
|
+
|
|
100
|
+
permissions:
|
|
101
|
+
contents: read
|
|
102
|
+
|
|
103
|
+
jobs:
|
|
104
|
+
${INTAKE_JOB}:
|
|
105
|
+
runs-on: ${opts.runsOn}
|
|
106
|
+
steps:
|
|
107
|
+
# Full history: a depth-1 clone is the only reason this source reads
|
|
108
|
+
# UNMEASURED (SHALLOW_HISTORY) on a runner.
|
|
109
|
+
- uses: ${opts.checkout}
|
|
110
|
+
with:
|
|
111
|
+
fetch-depth: 0
|
|
112
|
+
- uses: ${opts.setupBun}
|
|
113
|
+
with:
|
|
114
|
+
${opts.setupBunWith}
|
|
115
|
+
${opts.setupBunEnv}
|
|
116
|
+
- name: Mine fix commits into rule candidates
|
|
117
|
+
run: |
|
|
118
|
+
${indent(buildIntakeScript(opts.version), " ")}
|
|
119
|
+
`;
|
|
120
|
+
}
|
|
@@ -210,6 +210,16 @@ export const OWNED_FILES: OwnedFile[] = [
|
|
|
210
210
|
customize: "Set `board.id` in config and secret `RUNGATE_PROJECT_TOKEN`; add your own jobs alongside the generated `add-to-board` job",
|
|
211
211
|
avoid: "Don't edit the generated `add-to-board` job",
|
|
212
212
|
},
|
|
213
|
+
{
|
|
214
|
+
// #2233: fix-commit intake on push to main, with full history. Jason's
|
|
215
|
+
// #1126 decision put each intake source on its own trigger; a consumer has
|
|
216
|
+
// no scripts/, so the generator writes it and calls the published CLI.
|
|
217
|
+
path: ".github/workflows/intake.yml",
|
|
218
|
+
strategy: "job-preserving",
|
|
219
|
+
servedBy: "writeManagedFile (lib/scaffold/steps.ts) via mergeManagedWorkflow — the harness owns named jobs, not the file",
|
|
220
|
+
customize: "Advisory — a reading never fails the run; add your own jobs alongside the generated `fix-commit-intake` job",
|
|
221
|
+
avoid: "Don't edit the generated `fix-commit-intake` job, and don't drop `fetch-depth: 0` — a shallow clone reads UNMEASURED",
|
|
222
|
+
},
|
|
213
223
|
{
|
|
214
224
|
// `regenerate` until #655, and the declaration was never consulted:
|
|
215
225
|
// `generateAgentBriefsStep` called `writeFileSync` directly, so the row
|
package/lib/scaffold/steps.ts
CHANGED
|
@@ -50,6 +50,7 @@ import {
|
|
|
50
50
|
} from "./ai-review-unmeasured-workflow";
|
|
51
51
|
import { buildAdrAcceptedWorkflow } from "./adr-accepted-workflow";
|
|
52
52
|
import { buildClosingReferencesWorkflow } from "./closing-references-workflow";
|
|
53
|
+
import { buildIntakeWorkflow } from "./intake-workflow";
|
|
53
54
|
import { buildBoardReadyWorkflow } from "./board-ready-workflow";
|
|
54
55
|
import { buildBoardIntakeWorkflow } from "./board-intake-workflow";
|
|
55
56
|
import { CI_SUITE_FETCH_DEPTH, baseRefFetchDepth } from "../ci-history-depth";
|
|
@@ -4515,6 +4516,18 @@ ${buildAuthDetectStep(vertex)}${buildReviewRunStep(
|
|
|
4515
4516
|
version: harnessVersion(),
|
|
4516
4517
|
});
|
|
4517
4518
|
|
|
4519
|
+
// #2233: fix-commit intake on every push to main, with full history (Jason's
|
|
4520
|
+
// #1126 decision: each intake source on its own trigger, never on scaffold).
|
|
4521
|
+
// It calls the published CLI's `intake`, first shipped in 0.8.23.
|
|
4522
|
+
const intakeYml = buildIntakeWorkflow({
|
|
4523
|
+
runsOn,
|
|
4524
|
+
checkout: actionRef("actions/checkout"),
|
|
4525
|
+
setupBun: actionRef("oven-sh/setup-bun"),
|
|
4526
|
+
setupBunWith: bunStep,
|
|
4527
|
+
setupBunEnv: setupBunEnv(" "),
|
|
4528
|
+
version: harnessVersion(),
|
|
4529
|
+
});
|
|
4530
|
+
|
|
4518
4531
|
// #2166, ADR-014 row 4: new issues and pull requests onto the board. It
|
|
4519
4532
|
// names no board: the published CLI prints board.id from the default
|
|
4520
4533
|
// branch's config (#2173), and the add uses that node id.
|
|
@@ -4542,7 +4555,7 @@ ${buildAuthDetectStep(vertex)}${buildReviewRunStep(
|
|
|
4542
4555
|
// lib/conformity.ts reads back, and the two are held together by
|
|
4543
4556
|
// test/ai-review-workflow.test.ts AC-1, which looks for the constant's
|
|
4544
4557
|
// filename in what this loop actually wrote.
|
|
4545
|
-
for (const [name, content] of [["ci.yml", ciYml], ["gates.yml", gatesYml], ["ai-review.yml", aiReviewYml], ["ai-review-unmeasured.yml", aiReviewUnmeasuredYml], ["adr-accepted.yml", adrAcceptedYml], ["closing-references.yml", closingReferencesYml], ["board-ready.yml", boardReadyYml], ["board-intake.yml", boardIntakeYml]] as const) {
|
|
4558
|
+
for (const [name, content] of [["ci.yml", ciYml], ["gates.yml", gatesYml], ["ai-review.yml", aiReviewYml], ["ai-review-unmeasured.yml", aiReviewUnmeasuredYml], ["adr-accepted.yml", adrAcceptedYml], ["closing-references.yml", closingReferencesYml], ["board-ready.yml", boardReadyYml], ["board-intake.yml", boardIntakeYml], ["intake.yml", intakeYml]] as const) {
|
|
4546
4559
|
const label = `.github/workflows/${name}`;
|
|
4547
4560
|
try {
|
|
4548
4561
|
// Keyed by FILE, and the lookup is here because this loop is the only
|
package/lib/stage-ladder.ts
CHANGED
|
@@ -509,6 +509,68 @@ export function stageLadderViolations(raw: unknown): string[] {
|
|
|
509
509
|
});
|
|
510
510
|
});
|
|
511
511
|
|
|
512
|
+
out.push(...duplicateVerifyViolations(raw));
|
|
513
|
+
return out;
|
|
514
|
+
}
|
|
515
|
+
|
|
516
|
+
/**
|
|
517
|
+
* Two items proved by one command — #2236.
|
|
518
|
+
*
|
|
519
|
+
* They can never disagree, so the second is never checked as its own claim:
|
|
520
|
+
* whatever makes the first true makes the second true, which is the
|
|
521
|
+
* done-condition restated that AGENTS.md rule 10 forbids. On #2221, E4's first
|
|
522
|
+
* proof was B4's `criteria-field` reading copied byte for byte, and only
|
|
523
|
+
* review caught it.
|
|
524
|
+
*
|
|
525
|
+
* Across the WHOLE ladder, not within a stage, because that is where it
|
|
526
|
+
* happened. Compared after trimming, so a copied command with a stray newline
|
|
527
|
+
* is still the same command. A null verify is an item nobody has written a
|
|
528
|
+
* proof for yet, and two of those duplicate nothing.
|
|
529
|
+
*
|
|
530
|
+
* Retro improvements are in the same population (review of #2243): an
|
|
531
|
+
* improvement's verify proves a mechanism is in place, which is a claim of
|
|
532
|
+
* the same kind, and the live plan carried #1849's improvement in two retros
|
|
533
|
+
* with one command until #2245. An improvement is named by its stage and its
|
|
534
|
+
* 1-based position, `"C retro improvement 2"`, because it has no id.
|
|
535
|
+
*/
|
|
536
|
+
function duplicateVerifyViolations(raw: unknown[]): string[] {
|
|
537
|
+
const proofs: { id: string; verify: unknown }[] = [];
|
|
538
|
+
for (const s of raw) {
|
|
539
|
+
const stage = s as { id?: unknown; items?: unknown; retro?: unknown } | null;
|
|
540
|
+
if (stage === null || typeof stage !== "object") continue;
|
|
541
|
+
if (Array.isArray(stage.items)) {
|
|
542
|
+
for (const it of stage.items) {
|
|
543
|
+
const item = it as { id?: unknown; verify?: unknown } | null;
|
|
544
|
+
if (item !== null && typeof item === "object") proofs.push({ id: String(item.id), verify: item.verify });
|
|
545
|
+
}
|
|
546
|
+
}
|
|
547
|
+
const improvements = (stage.retro as { improvements?: unknown } | null)?.improvements;
|
|
548
|
+
if (Array.isArray(improvements)) {
|
|
549
|
+
improvements.forEach((imp, k) => {
|
|
550
|
+
const i = imp as { verify?: unknown } | null;
|
|
551
|
+
if (i !== null && typeof i === "object") {
|
|
552
|
+
proofs.push({ id: `${String(stage.id)} retro improvement ${k + 1}`, verify: i.verify });
|
|
553
|
+
}
|
|
554
|
+
});
|
|
555
|
+
}
|
|
556
|
+
}
|
|
557
|
+
const first = new Map<string, string>();
|
|
558
|
+
const out: string[] = [];
|
|
559
|
+
for (const { id, verify } of proofs) {
|
|
560
|
+
if (typeof verify !== "string") continue;
|
|
561
|
+
const cmd = verify.trim();
|
|
562
|
+
if (cmd === "") continue;
|
|
563
|
+
const earlier = first.get(cmd);
|
|
564
|
+
if (earlier === undefined) {
|
|
565
|
+
first.set(cmd, id);
|
|
566
|
+
continue;
|
|
567
|
+
}
|
|
568
|
+
out.push(
|
|
569
|
+
`"${earlier}" and "${id}" carry the same verify command, so they can never disagree and ` +
|
|
570
|
+
`"${id}" is never checked as its own claim — a done-condition restated (AGENTS.md rule 10). ` +
|
|
571
|
+
`Give "${id}" a command that reads what only it claims.`,
|
|
572
|
+
);
|
|
573
|
+
}
|
|
512
574
|
return out;
|
|
513
575
|
}
|
|
514
576
|
|
package/package.json
CHANGED
|
@@ -1,6 +1,6 @@
|
|
|
1
1
|
{
|
|
2
2
|
"name": "rungate",
|
|
3
|
-
"version": "0.8.
|
|
3
|
+
"version": "0.8.24",
|
|
4
4
|
"type": "module",
|
|
5
5
|
"description": "Ship harness — conformity tests, scaffold, and agent briefs for AI-first development",
|
|
6
6
|
"repository": {
|
|
@@ -37,6 +37,7 @@
|
|
|
37
37
|
"scripts/board-ready.ts",
|
|
38
38
|
"scripts/board-id.ts",
|
|
39
39
|
"scripts/unaccounted-work.ts",
|
|
40
|
+
"scripts/intake.ts",
|
|
40
41
|
"scripts/board-intake-move.ts",
|
|
41
42
|
"scripts/audit-closing-references.ts",
|
|
42
43
|
"scripts/create-spec.ts",
|
|
@@ -94,7 +94,7 @@
|
|
|
94
94
|
import { aiReviewJuryGate } from "../lib/ai-review-jury";
|
|
95
95
|
import { productionSuppression } from "../lib/ai-review-suppression";
|
|
96
96
|
import { main } from "../lib/scaffold/ai-review-findings";
|
|
97
|
-
import { AI_REVIEW_BALLOT_DIR_ENV, AI_REVIEW_THREADS_PATH_VAR } from "../lib/scaffold/ai-review-step";
|
|
97
|
+
import { AI_REVIEW_BALLOT_DIR_ENV, AI_REVIEW_CHECKOUT_ROOT_VAR, AI_REVIEW_THREADS_PATH_VAR } from "../lib/scaffold/ai-review-step";
|
|
98
98
|
|
|
99
99
|
const ballotDir = process.env[AI_REVIEW_BALLOT_DIR_ENV];
|
|
100
100
|
const jury =
|
|
@@ -105,9 +105,13 @@ const jury =
|
|
|
105
105
|
// Assembled in `lib/`, not here. This command and `ai-review-threads.ts` must
|
|
106
106
|
// form the SAME gate from the same file, and two hand-written literals is how
|
|
107
107
|
// `anchorOccurrences` reached one of them and not the other (#1861 round 4).
|
|
108
|
+
// #2257: the checkout by name. The generated step runs this from the payload's
|
|
109
|
+
// directory on purpose, so the working directory is not the checkout; it is
|
|
110
|
+
// the fallback only for a step generated before the variable was set.
|
|
111
|
+
const checkoutRoot = process.env[AI_REVIEW_CHECKOUT_ROOT_VAR];
|
|
108
112
|
const { gate: suppress } = productionSuppression(
|
|
109
113
|
process.env[AI_REVIEW_THREADS_PATH_VAR],
|
|
110
|
-
process.cwd(),
|
|
114
|
+
checkoutRoot !== undefined && checkoutRoot !== "" ? checkoutRoot : process.cwd(),
|
|
111
115
|
);
|
|
112
116
|
|
|
113
117
|
// One line, and `process.exit(main(` is load-bearing text: AC-4 in
|
|
@@ -0,0 +1,101 @@
|
|
|
1
|
+
#!/usr/bin/env bun
|
|
2
|
+
/**
|
|
3
|
+
* `rungate intake` — run the registered intake sources and print each one's
|
|
4
|
+
* reading (#2248).
|
|
5
|
+
*
|
|
6
|
+
* Before this, nothing outside `lib/intake-sources.ts` called
|
|
7
|
+
* `mineObservedFailures`, so no command ran intake at all. Jason decided on
|
|
8
|
+
* #1126 (2026-11-15) that each source runs on its own trigger: `fix-commits`
|
|
9
|
+
* in CI on push to main with full history (#2233), `gate-refusals` at the end
|
|
10
|
+
* of a ship run (#2234). Both call this.
|
|
11
|
+
*
|
|
12
|
+
* ADVISORY. A source that cannot be read here prints UNMEASURED with the
|
|
13
|
+
* producer's own reason and the command still exits 0: a shallow clone or an
|
|
14
|
+
* absent run store says nothing about the change that triggered the run, so
|
|
15
|
+
* it must never turn that run red. UNMEASURED is printed, never a silent
|
|
16
|
+
* zero, and the partial count of an unmeasured reading is never printed as
|
|
17
|
+
* the reading (`AggregateIntakeReading` withholds `candidates` there for the
|
|
18
|
+
* same reason). A bad argument is the only refusal.
|
|
19
|
+
*/
|
|
20
|
+
import { resolve } from "path";
|
|
21
|
+
import {
|
|
22
|
+
INTAKE_SOURCES,
|
|
23
|
+
defaultIntakeContext,
|
|
24
|
+
mineObservedFailures,
|
|
25
|
+
type IntakeContext,
|
|
26
|
+
} from "../lib/intake-sources";
|
|
27
|
+
|
|
28
|
+
/** A bad argument. The only non-zero exit: every reading exits 0. */
|
|
29
|
+
export const EXIT_REFUSED = 2;
|
|
30
|
+
|
|
31
|
+
const USAGE =
|
|
32
|
+
"rungate intake [--project-root <dir>] [--source <id>] [--json] " +
|
|
33
|
+
`(sources: ${INTAKE_SOURCES.map(s => s.id).join(", ")}; advisory — an unreadable source prints UNMEASURED and exits 0)`;
|
|
34
|
+
|
|
35
|
+
/**
|
|
36
|
+
* `deps.workStore` overrides the run store, so a test reads a fixture rather
|
|
37
|
+
* than this machine's `~/.rungate`. Absent, `defaultIntakeContext` decides,
|
|
38
|
+
* which reads `RUNGATE_WORK_DIR` through `lib/paths.ts`.
|
|
39
|
+
*/
|
|
40
|
+
export function main(argv: readonly string[], deps: { workStore?: string } = {}): number {
|
|
41
|
+
let rootArg: string | undefined;
|
|
42
|
+
let source: string | undefined;
|
|
43
|
+
let json = false;
|
|
44
|
+
for (let i = 0; i < argv.length; i++) {
|
|
45
|
+
const a = argv[i];
|
|
46
|
+
if (a === "--json") {
|
|
47
|
+
json = true;
|
|
48
|
+
} else if (a === "--project-root" || a === "--source") {
|
|
49
|
+
const next = argv[i + 1];
|
|
50
|
+
if (next === undefined || next.startsWith("--")) {
|
|
51
|
+
console.error(`intake: ${a} needs a value after it. Nothing was read.\n${USAGE}`);
|
|
52
|
+
return EXIT_REFUSED;
|
|
53
|
+
}
|
|
54
|
+
if (a === "--project-root") rootArg = next;
|
|
55
|
+
else source = next;
|
|
56
|
+
i++;
|
|
57
|
+
} else {
|
|
58
|
+
console.error(`intake: unknown argument ${JSON.stringify(a)}. Nothing was read.\n${USAGE}`);
|
|
59
|
+
return EXIT_REFUSED;
|
|
60
|
+
}
|
|
61
|
+
}
|
|
62
|
+
|
|
63
|
+
const sources = source === undefined ? INTAKE_SOURCES : INTAKE_SOURCES.filter(s => s.id === source);
|
|
64
|
+
if (sources.length === 0) {
|
|
65
|
+
console.error(
|
|
66
|
+
`intake: no source is named ${JSON.stringify(source)}. The registered sources are ` +
|
|
67
|
+
`${INTAKE_SOURCES.map(s => s.id).join(", ")}. Nothing was read.`,
|
|
68
|
+
);
|
|
69
|
+
return EXIT_REFUSED;
|
|
70
|
+
}
|
|
71
|
+
|
|
72
|
+
const root = resolve(rootArg ?? process.cwd());
|
|
73
|
+
const base = defaultIntakeContext(root);
|
|
74
|
+
const ctx: IntakeContext = deps.workStore === undefined ? base : { ...base, workStore: deps.workStore };
|
|
75
|
+
const reading = mineObservedFailures(ctx, sources);
|
|
76
|
+
|
|
77
|
+
if (json) {
|
|
78
|
+
console.log(JSON.stringify(reading, null, 2));
|
|
79
|
+
return 0;
|
|
80
|
+
}
|
|
81
|
+
|
|
82
|
+
for (const s of sources) {
|
|
83
|
+
const r = reading.bySource[s.id];
|
|
84
|
+
if (r.reading === "MEASURED") {
|
|
85
|
+
console.log(
|
|
86
|
+
`${s.id}: MEASURED — ${r.candidates.length} candidate(s) from ${r.commitsScanned} unit(s) of history read`,
|
|
87
|
+
);
|
|
88
|
+
} else {
|
|
89
|
+
console.log(`${s.id}: UNMEASURED — ${r.unmeasuredReasons.join("; ")}`);
|
|
90
|
+
}
|
|
91
|
+
}
|
|
92
|
+
console.log(
|
|
93
|
+
reading.reading === "MEASURED"
|
|
94
|
+
? `intake: MEASURED — ${reading.candidates.length} candidate(s) across ${sources.length} source(s)`
|
|
95
|
+
: `intake: UNMEASURED — ${reading.unmeasuredReasons.length} reason(s) above; ` +
|
|
96
|
+
`${reading.partialCandidates.length} candidate(s) came from the sources that could be read, and are not a reading`,
|
|
97
|
+
);
|
|
98
|
+
return 0;
|
|
99
|
+
}
|
|
100
|
+
|
|
101
|
+
if (import.meta.main) process.exitCode = main(process.argv.slice(2));
|