mandrel-platform 1.0.1 → 1.1.0
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/package.json +1 -1
- package/scripts/audit-check.mjs +112 -12
- package/scripts/audit-check.test.mjs +195 -0
- package/scripts/check-cancelled-provenance.test.mjs +493 -1
- package/scripts/check-docs-staleness.mjs +15 -2
- package/scripts/check-docs-staleness.test.mjs +114 -9
- package/scripts/job-cleanup-hook.test.mjs +234 -0
- package/templates/runbooks/runner-provisioning.md +12 -3
- package/templates/runner/.env.example +8 -2
- package/templates/runner/job-cleanup.sh +49 -20
|
@@ -3,16 +3,23 @@
|
|
|
3
3
|
* check-docs-staleness.test.mjs — node:test suite for the docs-staleness lint
|
|
4
4
|
* (Story #197).
|
|
5
5
|
*
|
|
6
|
-
* Focus: the
|
|
7
|
-
* years 2020–2024 (`/expires[:\s]+202[0-4]-\d{2}-\d{2}/i`), so an expiry that
|
|
8
|
-
* lapsed in 2025, 2026, or any later year sailed through the gate — a
|
|
9
|
-
* fail-open. The fix broadens the pattern to any 20xx year and defers the
|
|
10
|
-
* "is it actually in the past?" decision to `isExpiredDate`, so the rule stays
|
|
11
|
-
* correct as the calendar advances and never flags a still-valid future date.
|
|
6
|
+
* Focus: rule PRECISION — the two ways a staleness rule stops being useful.
|
|
12
7
|
*
|
|
13
|
-
*
|
|
14
|
-
* (`
|
|
15
|
-
*
|
|
8
|
+
* 1. Fail-open (`expired-placeholder` under-fires). The rule first hardcoded
|
|
9
|
+
* the years 2020–2024 (`/expires[:\s]+202[0-4]-\d{2}-\d{2}/i`), so an expiry
|
|
10
|
+
* that lapsed in 2025 or later sailed through. The fix broadens to any 20xx
|
|
11
|
+
* year and defers "is it actually in the past?" to `isExpiredDate`, so the
|
|
12
|
+
* rule stays correct as the calendar advances and never flags a still-valid
|
|
13
|
+
* future date. A second instance of the same class: the separator required
|
|
14
|
+
* `expires` to be followed directly by `:` or whitespace, which skipped the
|
|
15
|
+
* CVE allowlist's own JSON shape (`"expires": "…"`).
|
|
16
|
+
* 2. Fail-noisy (`quality-yml-ref` over-fires). A plain substring match meant
|
|
17
|
+
* the platform's own `pr-quality.yml` matched, so 58 of 59 findings were
|
|
18
|
+
* false positives — enough to bury a real error in the same run.
|
|
19
|
+
*
|
|
20
|
+
* These tests exercise the date logic directly (`isExpiredDate`), the compiled
|
|
21
|
+
* patterns, and end-to-end behaviour (`lintFile` against a real fixture file),
|
|
22
|
+
* pinning "today" via a fixed clock so they are deterministic.
|
|
16
23
|
*
|
|
17
24
|
* Run: node --test scripts/check-docs-staleness.test.mjs
|
|
18
25
|
*/
|
|
@@ -128,3 +135,101 @@ test('lintFile honours the staleness-ignore suppression comment for the year rul
|
|
|
128
135
|
},
|
|
129
136
|
);
|
|
130
137
|
});
|
|
138
|
+
|
|
139
|
+
// ---------------------------------------------------------------------------
|
|
140
|
+
// expired-placeholder — the JSON-quoted shape.
|
|
141
|
+
//
|
|
142
|
+
// Second fail-open of the same class as the 202[0-4] year window: the pattern
|
|
143
|
+
// required `expires` to be followed directly by `:` or whitespace, so the CVE
|
|
144
|
+
// allowlist's own JSON shape (`"expires": "2026-12-31"`, per audit-check.mjs)
|
|
145
|
+
// never matched — `expires"` is neither. Every documented allowlist entry was
|
|
146
|
+
// therefore invisible to the gate, including a lapsed one in
|
|
147
|
+
// docs/runbooks/dependency-update.md.
|
|
148
|
+
// ---------------------------------------------------------------------------
|
|
149
|
+
|
|
150
|
+
test('expired-placeholder matches the JSON-quoted allowlist shape', () => {
|
|
151
|
+
const rule = RULES.find((r) => r.id === 'expired-placeholder');
|
|
152
|
+
for (const line of [
|
|
153
|
+
' "expires": "2025-12-31",', // the CVE allowlist's real shape
|
|
154
|
+
" 'expires': '2025-12-31',", // single-quoted (YAML/JS)
|
|
155
|
+
' expires: "2025-12-31"', // quoted value, bare key
|
|
156
|
+
]) {
|
|
157
|
+
rule.pattern.lastIndex = 0;
|
|
158
|
+
assert.ok(rule.pattern.test(line), `pattern should match ${JSON.stringify(line)}`);
|
|
159
|
+
}
|
|
160
|
+
});
|
|
161
|
+
|
|
162
|
+
test('expired-placeholder still matches the unquoted shapes (no regression)', () => {
|
|
163
|
+
const rule = RULES.find((r) => r.id === 'expired-placeholder');
|
|
164
|
+
for (const line of [
|
|
165
|
+
'expires: 2025-01-01',
|
|
166
|
+
'# CVE-2022-3517 — expires 2025-06-01',
|
|
167
|
+
]) {
|
|
168
|
+
rule.pattern.lastIndex = 0;
|
|
169
|
+
assert.ok(rule.pattern.test(line), `pattern should match ${JSON.stringify(line)}`);
|
|
170
|
+
}
|
|
171
|
+
});
|
|
172
|
+
|
|
173
|
+
test('lintFile flags a lapsed JSON-quoted expiry end-to-end', () => {
|
|
174
|
+
withTempDoc('Allowlist entry.\n "expires": "2025-12-31",\nEnd.\n', (file) => {
|
|
175
|
+
const expired = lintFile(file).filter((f) => f.rule.id === 'expired-placeholder');
|
|
176
|
+
assert.equal(expired.length, 1);
|
|
177
|
+
assert.match(expired[0].match, /2025-12-31/);
|
|
178
|
+
});
|
|
179
|
+
});
|
|
180
|
+
|
|
181
|
+
test('lintFile does NOT flag a future JSON-quoted expiry', () => {
|
|
182
|
+
withTempDoc('Allowlist entry.\n "expires": "2099-12-31",\nEnd.\n', (file) => {
|
|
183
|
+
const expired = lintFile(file).filter((f) => f.rule.id === 'expired-placeholder');
|
|
184
|
+
assert.equal(expired.length, 0);
|
|
185
|
+
});
|
|
186
|
+
});
|
|
187
|
+
|
|
188
|
+
test('lintFile does NOT flag a placeholder expiry token', () => {
|
|
189
|
+
// The runbooks intentionally use `<YYYY-MM-DD>` rather than a concrete date,
|
|
190
|
+
// precisely so an example can never lapse into a finding.
|
|
191
|
+
withTempDoc('Allowlist entry.\n "expires": "<YYYY-MM-DD>",\nEnd.\n', (file) => {
|
|
192
|
+
const expired = lintFile(file).filter((f) => f.rule.id === 'expired-placeholder');
|
|
193
|
+
assert.equal(expired.length, 0);
|
|
194
|
+
});
|
|
195
|
+
});
|
|
196
|
+
|
|
197
|
+
// ---------------------------------------------------------------------------
|
|
198
|
+
// quality-yml-ref — bare filename only.
|
|
199
|
+
//
|
|
200
|
+
// The pattern was a plain substring match, so the platform's own
|
|
201
|
+
// `pr-quality.yml` matched: 58 of 59 findings against mandrel-platform's docs
|
|
202
|
+
// were that false positive, burying a real `expired-placeholder` error. The
|
|
203
|
+
// assertion below is on the invariant (a `<prefix>-quality.yml` is a different
|
|
204
|
+
// file) rather than on the single `pr-` spelling that motivated it.
|
|
205
|
+
// ---------------------------------------------------------------------------
|
|
206
|
+
|
|
207
|
+
test('quality-yml-ref does NOT match a prefixed <prefix>-quality.yml', () => {
|
|
208
|
+
const rule = RULES.find((r) => r.id === 'quality-yml-ref');
|
|
209
|
+
assert.ok(rule, 'quality-yml-ref rule must exist');
|
|
210
|
+
for (const line of [
|
|
211
|
+
'uses: dsj1984/mandrel-platform/.github/workflows/pr-quality.yml@abc123',
|
|
212
|
+
'the `pr-quality.yml` reusable workflow',
|
|
213
|
+
'a consumer that names its caller `ci-quality.yml`',
|
|
214
|
+
'see my_quality.yml for details',
|
|
215
|
+
]) {
|
|
216
|
+
rule.pattern.lastIndex = 0;
|
|
217
|
+
assert.equal(
|
|
218
|
+
rule.pattern.test(line),
|
|
219
|
+
false,
|
|
220
|
+
`pattern must not match ${JSON.stringify(line)}`,
|
|
221
|
+
);
|
|
222
|
+
}
|
|
223
|
+
});
|
|
224
|
+
|
|
225
|
+
test('quality-yml-ref still matches a bare quality.yml reference', () => {
|
|
226
|
+
const rule = RULES.find((r) => r.id === 'quality-yml-ref');
|
|
227
|
+
for (const line of [
|
|
228
|
+
"| athportal | `quality.yml` | `quality` |",
|
|
229
|
+
'the quality.yml workflow was renamed',
|
|
230
|
+
'quality.yml',
|
|
231
|
+
]) {
|
|
232
|
+
rule.pattern.lastIndex = 0;
|
|
233
|
+
assert.ok(rule.pattern.test(line), `pattern should match ${JSON.stringify(line)}`);
|
|
234
|
+
}
|
|
235
|
+
});
|
|
@@ -0,0 +1,234 @@
|
|
|
1
|
+
#!/usr/bin/env node
|
|
2
|
+
/**
|
|
3
|
+
* job-cleanup-hook.test.mjs — node:test suite for the ACTIONS_RUNNER_HOOK_JOB_STARTED
|
|
4
|
+
* hook shipped at `templates/runner/job-cleanup.sh` (Story #345).
|
|
5
|
+
*
|
|
6
|
+
* The hook runs INSIDE the job's clock on every job of every persistent
|
|
7
|
+
* self-hosted runner, so its cost is charged to `Set up runner` and a slow
|
|
8
|
+
* hook kills jobs against their own `timeout-minutes`. Issue #343 is exactly
|
|
9
|
+
* that failure: the hook enumerated the shared OS temp root, which had reached
|
|
10
|
+
* 841,690 entries on the swarm-os host, and `Set up runner` reached 5m29s.
|
|
11
|
+
*
|
|
12
|
+
* The load-bearing property this suite pins is therefore a NEGATIVE one — the
|
|
13
|
+
* hook must never read the shared temp root — and it is asserted two ways,
|
|
14
|
+
* because neither alone is sufficient:
|
|
15
|
+
*
|
|
16
|
+
* 1. Behaviourally: a decoy shared temp root is planted with entries that
|
|
17
|
+
* match the old sweep's patterns, and the hook must leave every one of
|
|
18
|
+
* them alone. This catches a sweep that still deletes there.
|
|
19
|
+
* 2. Structurally: the script text must contain no enumeration rooted at the
|
|
20
|
+
* shared temp root. This catches a sweep that reads the directory but
|
|
21
|
+
* happens to delete nothing — the exact shape of the #343 stall, which
|
|
22
|
+
* was pure cost with no observable effect.
|
|
23
|
+
*
|
|
24
|
+
* A wall-clock assertion was deliberately NOT used for (2): the cost is a
|
|
25
|
+
* function of host churn, so on a clean dev machine a full enumeration of a
|
|
26
|
+
* small decoy root is fast and would pass. Timing here would be a test that
|
|
27
|
+
* only fails on the machine that least needs it.
|
|
28
|
+
*
|
|
29
|
+
* The suite executes the real script with env fixtures, following the
|
|
30
|
+
* `scripts/resolve-diff-range.test.mjs` precedent for shell-under-test.
|
|
31
|
+
*
|
|
32
|
+
* Run: node --test scripts/job-cleanup-hook.test.mjs
|
|
33
|
+
*/
|
|
34
|
+
|
|
35
|
+
import assert from "node:assert/strict";
|
|
36
|
+
import { execFileSync } from "node:child_process";
|
|
37
|
+
import { mkdtempSync, mkdirSync, writeFileSync, readFileSync, readdirSync, rmSync, chmodSync, existsSync } from "node:fs";
|
|
38
|
+
import { tmpdir } from "node:os";
|
|
39
|
+
import { fileURLToPath } from "node:url";
|
|
40
|
+
import { dirname, join } from "node:path";
|
|
41
|
+
import { test, after } from "node:test";
|
|
42
|
+
|
|
43
|
+
const HERE = dirname(fileURLToPath(import.meta.url));
|
|
44
|
+
const SCRIPT = join(HERE, "..", "templates", "runner", "job-cleanup.sh");
|
|
45
|
+
|
|
46
|
+
/** Sandboxes created by the suite, torn down in `after`. */
|
|
47
|
+
const SANDBOXES = [];
|
|
48
|
+
|
|
49
|
+
after(() => {
|
|
50
|
+
for (const dir of SANDBOXES) {
|
|
51
|
+
// Restore any permissions the unwritable-temp case removed, or the
|
|
52
|
+
// recursive delete cannot descend.
|
|
53
|
+
try {
|
|
54
|
+
chmodSync(join(dir, "runner", "_work", "_temp"), 0o755);
|
|
55
|
+
} catch {
|
|
56
|
+
/* not every sandbox has one */
|
|
57
|
+
}
|
|
58
|
+
rmSync(dir, { recursive: true, force: true });
|
|
59
|
+
}
|
|
60
|
+
});
|
|
61
|
+
|
|
62
|
+
/**
|
|
63
|
+
* Build a sandbox holding a synthetic runner root and a decoy shared temp root.
|
|
64
|
+
*
|
|
65
|
+
* @param {{ sharedEntries?: number }} [opts]
|
|
66
|
+
* @returns {{ root: string, runnerDir: string, runnerTmp: string, sharedTmp: string }}
|
|
67
|
+
*/
|
|
68
|
+
function makeSandbox({ sharedEntries = 0 } = {}) {
|
|
69
|
+
const root = mkdtempSync(join(tmpdir(), "job-cleanup-test-"));
|
|
70
|
+
SANDBOXES.push(root);
|
|
71
|
+
|
|
72
|
+
const runnerDir = join(root, "runner");
|
|
73
|
+
const runnerTmp = join(runnerDir, "_work", "_temp");
|
|
74
|
+
const sharedTmp = join(root, "shared-tmp");
|
|
75
|
+
mkdirSync(runnerTmp, { recursive: true });
|
|
76
|
+
mkdirSync(join(runnerDir, "_work", "_tool"), { recursive: true });
|
|
77
|
+
mkdirSync(sharedTmp, { recursive: true });
|
|
78
|
+
|
|
79
|
+
// The decoy shared root carries entries matching BOTH the retired sweep's
|
|
80
|
+
// patterns and the current runner-scoped ones, so a sweep that kept the old
|
|
81
|
+
// root or reused the new globs against it is caught either way.
|
|
82
|
+
writeFileSync(join(sharedTmp, "gitleaks.tmp"), "decoy");
|
|
83
|
+
mkdirSync(join(sharedTmp, "gitleaks-8.30.1"), { recursive: true });
|
|
84
|
+
mkdirSync(join(sharedTmp, "gitleaks.AbCdEf"), { recursive: true });
|
|
85
|
+
mkdirSync(join(sharedTmp, "osv-scanner.AbCdEf"), { recursive: true });
|
|
86
|
+
for (let i = 0; i < sharedEntries; i += 1) {
|
|
87
|
+
writeFileSync(join(sharedTmp, `unrelated-${i}.tmp`), "x");
|
|
88
|
+
}
|
|
89
|
+
|
|
90
|
+
return { root, runnerDir, runnerTmp, sharedTmp };
|
|
91
|
+
}
|
|
92
|
+
|
|
93
|
+
/**
|
|
94
|
+
* Run the hook against a sandbox.
|
|
95
|
+
*
|
|
96
|
+
* @param {{ runnerDir: string, sharedTmp: string }} sandbox
|
|
97
|
+
* @returns {{ status: number, stdout: string }}
|
|
98
|
+
*/
|
|
99
|
+
function runHook({ runnerDir, sharedTmp }) {
|
|
100
|
+
try {
|
|
101
|
+
const stdout = execFileSync("bash", [SCRIPT], {
|
|
102
|
+
encoding: "utf8",
|
|
103
|
+
env: { ...process.env, RUNNER_DIR: runnerDir, TMPDIR: sharedTmp },
|
|
104
|
+
timeout: 60_000,
|
|
105
|
+
});
|
|
106
|
+
return { status: 0, stdout };
|
|
107
|
+
} catch (err) {
|
|
108
|
+
return { status: err.status ?? 1, stdout: err.stdout ?? "" };
|
|
109
|
+
}
|
|
110
|
+
}
|
|
111
|
+
|
|
112
|
+
test("removes this runner's own stale tool-download and pnpm-shim leftovers", () => {
|
|
113
|
+
const sandbox = makeSandbox();
|
|
114
|
+
const { runnerTmp } = sandbox;
|
|
115
|
+
|
|
116
|
+
// Every artifact the platform's actions now extract into runner.temp.
|
|
117
|
+
const owned = [
|
|
118
|
+
join(runnerTmp, "pnpm"),
|
|
119
|
+
join(runnerTmp, "setup-pnpm"),
|
|
120
|
+
join(runnerTmp, "gitleaks.AbCdEf"),
|
|
121
|
+
join(runnerTmp, "osv-scanner.AbCdEf"),
|
|
122
|
+
join(runnerTmp, "semgrep.AbCdEf"),
|
|
123
|
+
];
|
|
124
|
+
for (const dir of owned) {
|
|
125
|
+
mkdirSync(dir, { recursive: true });
|
|
126
|
+
writeFileSync(join(dir, "leftover"), "stale");
|
|
127
|
+
}
|
|
128
|
+
writeFileSync(join(runnerTmp, "gh-api-err.AbCdEf"), "stale");
|
|
129
|
+
|
|
130
|
+
const { status } = runHook(sandbox);
|
|
131
|
+
assert.equal(status, 0, "the hook must never fail a job");
|
|
132
|
+
|
|
133
|
+
for (const dir of owned) {
|
|
134
|
+
assert.equal(existsSync(dir), false, `runner-owned leftover not swept: ${dir}`);
|
|
135
|
+
}
|
|
136
|
+
assert.equal(existsSync(join(runnerTmp, "gh-api-err.AbCdEf")), false);
|
|
137
|
+
});
|
|
138
|
+
|
|
139
|
+
test("leaves the shared temp root untouched", () => {
|
|
140
|
+
// The decoy is deliberately SMALL. An earlier draft planted 5,000 entries to
|
|
141
|
+
// evoke the 841,690 seen on the swarm-os host, but that bought no signal: the
|
|
142
|
+
// assertion below is on effect (nothing deleted), which is count-independent,
|
|
143
|
+
// and the companion structural test — not a stopwatch — is what pins the
|
|
144
|
+
// cost property. Planting alone cost ~87s on a dev Mac whose endpoint
|
|
145
|
+
// security scans every write (~17ms/file), which would have made this the
|
|
146
|
+
// slowest test in the suite by two orders of magnitude, and a flaky one.
|
|
147
|
+
//
|
|
148
|
+
// What actually matters is pattern COVERAGE, and makeSandbox plants every
|
|
149
|
+
// shape — both the retired sweep's names and the current runner-scoped
|
|
150
|
+
// globs — so a sweep pointed back at the shared root is caught by the first
|
|
151
|
+
// entry it would match.
|
|
152
|
+
const sandbox = makeSandbox({ sharedEntries: 20 });
|
|
153
|
+
const { sharedTmp } = sandbox;
|
|
154
|
+
|
|
155
|
+
const before = readdirSync(sharedTmp).sort();
|
|
156
|
+
const { status } = runHook(sandbox);
|
|
157
|
+
|
|
158
|
+
assert.equal(status, 0);
|
|
159
|
+
assert.deepEqual(
|
|
160
|
+
readdirSync(sharedTmp).sort(),
|
|
161
|
+
before,
|
|
162
|
+
"the hook deleted from the shared temp root — it must only sweep runner-owned paths",
|
|
163
|
+
);
|
|
164
|
+
});
|
|
165
|
+
|
|
166
|
+
test("the script text contains no enumeration rooted at the shared temp root", () => {
|
|
167
|
+
const source = readFileSync(SCRIPT, "utf8");
|
|
168
|
+
|
|
169
|
+
// Assert the PROPERTY (the hook cannot reach the shared root at all), not
|
|
170
|
+
// one syntactic form of violating it. Pinning a specific `find "${TMP…"`
|
|
171
|
+
// shape would miss an unbraced `$TMPDIR`, an `ls | grep`, a `for f in
|
|
172
|
+
// "${TMP}"/…` loop, or a renamed intermediate variable — all of which
|
|
173
|
+
// reintroduce the unbounded read this Story removed.
|
|
174
|
+
//
|
|
175
|
+
// Every one of those must name the shared root to reach it, and the hook has
|
|
176
|
+
// no legitimate use for it: RUNNER_TMP derives from RUNNER_DIR. So the
|
|
177
|
+
// absence of the name is both necessary and sufficient, and it stays a true
|
|
178
|
+
// invariant rather than a regex chasing syntax.
|
|
179
|
+
//
|
|
180
|
+
// Comments are stripped first: the header documents the #343 incident by
|
|
181
|
+
// name, and that prose is the reason the next maintainer will not re-add the
|
|
182
|
+
// sweep. Asserting over raw text would force the fix to delete its own
|
|
183
|
+
// rationale.
|
|
184
|
+
const code = source
|
|
185
|
+
.split("\n")
|
|
186
|
+
.filter((line) => !/^\s*#/.test(line))
|
|
187
|
+
.join("\n");
|
|
188
|
+
|
|
189
|
+
assert.equal(
|
|
190
|
+
/TMPDIR/.test(code),
|
|
191
|
+
false,
|
|
192
|
+
"the hook references the shared temp root — its cost must scale with runner-owned state only",
|
|
193
|
+
);
|
|
194
|
+
|
|
195
|
+
// Second half of the property: no directory enumeration, anywhere. Naming
|
|
196
|
+
// the shared root is one way to reintroduce unbounded cost; a `find` rooted
|
|
197
|
+
// at an intermediate variable is another, and it would not have to mention
|
|
198
|
+
// TMPDIR on the same line. The hook has no legitimate need to enumerate —
|
|
199
|
+
// it addresses known paths under RUNNER_TMP directly — so the absence of
|
|
200
|
+
// `find` is a stronger and more durable invariant than any pattern match
|
|
201
|
+
// over its arguments.
|
|
202
|
+
assert.equal(
|
|
203
|
+
/\bfind\b/.test(code),
|
|
204
|
+
false,
|
|
205
|
+
"the hook enumerates a directory — address known runner-owned paths directly instead",
|
|
206
|
+
);
|
|
207
|
+
|
|
208
|
+
// The age gate existed solely to make deleting from the SHARED root safe.
|
|
209
|
+
// Runner-scoped paths are unreachable by a co-resident runner, so a
|
|
210
|
+
// surviving knob would be dead configuration the runbook still promises.
|
|
211
|
+
assert.equal(
|
|
212
|
+
source.includes("JOB_CLEANUP_STALE_MINUTES"),
|
|
213
|
+
false,
|
|
214
|
+
"the retired age-gate knob is still referenced",
|
|
215
|
+
);
|
|
216
|
+
});
|
|
217
|
+
|
|
218
|
+
test("exits 0 when the runner directory does not exist", () => {
|
|
219
|
+
const sandbox = makeSandbox();
|
|
220
|
+
const { status } = runHook({
|
|
221
|
+
runnerDir: join(sandbox.root, "no-such-runner"),
|
|
222
|
+
sharedTmp: sandbox.sharedTmp,
|
|
223
|
+
});
|
|
224
|
+
assert.equal(status, 0, "a missing runner root must degrade to a no-op, never fail the job");
|
|
225
|
+
});
|
|
226
|
+
|
|
227
|
+
test("exits 0 when the runner temp is unwritable", () => {
|
|
228
|
+
const sandbox = makeSandbox();
|
|
229
|
+
mkdirSync(join(sandbox.runnerTmp, "gitleaks.AbCdEf"), { recursive: true });
|
|
230
|
+
chmodSync(sandbox.runnerTmp, 0o500);
|
|
231
|
+
|
|
232
|
+
const { status } = runHook(sandbox);
|
|
233
|
+
assert.equal(status, 0, "an unwritable runner temp must not fail the job");
|
|
234
|
+
});
|
|
@@ -150,9 +150,18 @@ with the runner root's absolute path. The resulting file wires:
|
|
|
150
150
|
|
|
151
151
|
- `ACTIONS_RUNNER_HOOK_JOB_STARTED=<RUNNER_DIR>/job-cleanup.sh` — the
|
|
152
152
|
job-start hook. It reaps orphaned pnpm/node processes parented to **this**
|
|
153
|
-
runner's work tree, clears stale runner-scoped pnpm installs, and
|
|
154
|
-
|
|
155
|
-
(always exits 0) and never touches another runner's state.
|
|
153
|
+
runner's work tree, clears stale runner-scoped pnpm installs, and removes
|
|
154
|
+
leftover tool-download dirs from `<RUNNER_DIR>/_work/_temp`. It never fails
|
|
155
|
+
a job (always exits 0) and never touches another runner's state.
|
|
156
|
+
|
|
157
|
+
**Every path it reads is runner-scoped, and that is load-bearing** (issue
|
|
158
|
+
#343). The hook runs inside the *job's* clock, so its cost is charged to
|
|
159
|
+
`Set up runner` and counts against the job's own `timeout-minutes`. An
|
|
160
|
+
earlier version swept the host-shared OS temp root; on a host where that
|
|
161
|
+
directory had grown to ~840k entries, `Set up runner` reached 5m29s and
|
|
162
|
+
jobs were killed before their first real step — surfacing as `cancelled`
|
|
163
|
+
on unrelated diffs. If you add a sweep to this hook, root it at
|
|
164
|
+
`_work/_temp`, never at `$TMPDIR`.
|
|
156
165
|
- `RUNNER_TOOL_CACHE=<RUNNER_DIR>/_work/_tool` and
|
|
157
166
|
`AGENT_TOOLSDIRECTORY=<RUNNER_DIR>/_work/_tool` — runner-scoped tool cache
|
|
158
167
|
(two env names, one dir; some actions read the legacy name).
|
|
@@ -19,8 +19,14 @@ LANG=en_US.UTF-8
|
|
|
19
19
|
|
|
20
20
|
# Job-start hygiene hook. Runs templates/runner/job-cleanup.sh (installed
|
|
21
21
|
# into the runner root) before every job: reaps orphaned pnpm/node processes
|
|
22
|
-
# from THIS runner's work tree and clears stale install/temp artifacts
|
|
23
|
-
# The hook is runner-scoped and never fails the job (always
|
|
22
|
+
# from THIS runner's work tree and clears stale install/temp artifacts from
|
|
23
|
+
# `_work/_temp`. The hook is runner-scoped and never fails the job (always
|
|
24
|
+
# exits 0).
|
|
25
|
+
#
|
|
26
|
+
# Runner-scoped means CHEAP as well as safe: the hook runs inside the job's
|
|
27
|
+
# clock, so anything it reads is billed to `Set up runner` and counts against
|
|
28
|
+
# the job's `timeout-minutes`. It touches only this runner's own paths, so its
|
|
29
|
+
# cost never becomes a function of host-wide temp churn (issue #343).
|
|
24
30
|
ACTIONS_RUNNER_HOOK_JOB_STARTED=<RUNNER_DIR>/job-cleanup.sh
|
|
25
31
|
|
|
26
32
|
# Runner-scoped tool cache. Without this, actions/setup-node & friends
|
|
@@ -10,8 +10,8 @@
|
|
|
10
10
|
#
|
|
11
11
|
# - an orphaned `pnpm`/`node` process (e.g. a hung install or lint) still
|
|
12
12
|
# mutating the pnpm shim install, corrupting the pnpm CLI for the next job;
|
|
13
|
-
# - leftover
|
|
14
|
-
#
|
|
13
|
+
# - leftover tool-download temp dirs (gitleaks, OSV-scanner, the semgrep
|
|
14
|
+
# venv) accumulating in the runner's own job temp.
|
|
15
15
|
#
|
|
16
16
|
# Running this before every job gives each job a clean slate ("fresh per job"
|
|
17
17
|
# without the cost of re-registering an ephemeral runner).
|
|
@@ -35,9 +35,36 @@
|
|
|
35
35
|
# own work tree (`<RUNNER_DIR>/_work/...`). Every path below derives
|
|
36
36
|
# from RUNNER_DIR, which is unique per runner, so a co-resident
|
|
37
37
|
# runner's processes and files are never matched.
|
|
38
|
-
# 3.
|
|
39
|
-
#
|
|
40
|
-
#
|
|
38
|
+
# 3. NEVER READS the shared OS temp root, either. Reading is not free: the
|
|
39
|
+
# hook runs inside the JOB's clock, so any cost here is charged to
|
|
40
|
+
# `Set up runner` and counts against the job's own `timeout-minutes`.
|
|
41
|
+
# $TMPDIR is unbounded and shared with every other process on the host,
|
|
42
|
+
# so a sweep rooted there costs a function of how much UNRELATED junk
|
|
43
|
+
# the host has accumulated — see the incident note below.
|
|
44
|
+
#
|
|
45
|
+
# ── WHY THE SHARED-$TMPDIR SWEEP IS GONE (issue #343) ───────────────────────
|
|
46
|
+
#
|
|
47
|
+
# This hook used to age-gate two `find "$TMPDIR" -maxdepth 1 -name …` sweeps
|
|
48
|
+
# for `gitleaks.tmp` / `gitleaks-*`. `-maxdepth 1 -name <literal>` is a FULL
|
|
49
|
+
# directory enumeration for what is really an existence check, so its cost
|
|
50
|
+
# scaled with host churn. On the swarm-os runner host $TMPDIR reached 841,690
|
|
51
|
+
# entries; one scan measured 42s, the hook ran two of them, and up to 16
|
|
52
|
+
# co-resident runners ran it concurrently. `Set up runner` reached 5m29s, and
|
|
53
|
+
# every job whose `timeout-minutes` sat at or below that was killed before its
|
|
54
|
+
# first real step — surfacing as `cancelled` on an innocent diff.
|
|
55
|
+
#
|
|
56
|
+
# It was also a no-op: the platform's actions extract via `mktemp -d`, so
|
|
57
|
+
# nothing ever created `gitleaks.tmp` or `gitleaks-*`. The sweep paid an
|
|
58
|
+
# unbounded cost hunting names that never existed, while the dirs the actions
|
|
59
|
+
# DID leave went unswept.
|
|
60
|
+
#
|
|
61
|
+
# The fix is ownership, not tuning: every platform action now extracts into
|
|
62
|
+
# `${RUNNER_TEMP}` (== RUNNER_TMP below), which is unique per runner. A
|
|
63
|
+
# co-resident runner's in-flight download is therefore unreachable from here
|
|
64
|
+
# by construction — which is what retired the age gate outright (along with
|
|
65
|
+
# the stale-minutes env knob that tuned it), rather than merely shrinking its
|
|
66
|
+
# blast radius. Keep it that way: a sweep added here MUST be rooted at
|
|
67
|
+
# RUNNER_TMP.
|
|
41
68
|
#
|
|
42
69
|
# ── PARAMETERIZATION ────────────────────────────────────────────────────────
|
|
43
70
|
#
|
|
@@ -48,11 +75,8 @@
|
|
|
48
75
|
# runner root, next to config.sh / run.sh). Override via env
|
|
49
76
|
# only if you install the hook elsewhere.
|
|
50
77
|
# RUNNER_TMP — the runner's per-runner job temp (`runner.temp`), always
|
|
51
|
-
# `${RUNNER_DIR}/_work/_temp`.
|
|
52
|
-
#
|
|
53
|
-
# — age threshold (minutes) for the shared-$TMPDIR gitleaks
|
|
54
|
-
# sweep. Default 60. Artifacts younger than this are assumed
|
|
55
|
-
# in-flight and left alone.
|
|
78
|
+
# `${RUNNER_DIR}/_work/_temp`. Every path this hook touches
|
|
79
|
+
# lives under it.
|
|
56
80
|
#
|
|
57
81
|
# Configured via `ACTIONS_RUNNER_HOOK_JOB_STARTED=<RUNNER_DIR>/job-cleanup.sh`
|
|
58
82
|
# in the runner's `.env` (see .env.example in this directory).
|
|
@@ -64,8 +88,6 @@ set +e
|
|
|
64
88
|
RUNNER_DIR="${RUNNER_DIR:-$(cd "$(dirname "$0")" && pwd)}"
|
|
65
89
|
RUNNER_WORK="${RUNNER_DIR}/_work"
|
|
66
90
|
RUNNER_TMP="${RUNNER_WORK}/_temp"
|
|
67
|
-
TMP="${TMPDIR:-/tmp}"
|
|
68
|
-
STALE_MINUTES="${JOB_CLEANUP_STALE_MINUTES:-60}"
|
|
69
91
|
|
|
70
92
|
# 1) Reap orphaned pnpm/node processes from prior jobs — scoped to THIS
|
|
71
93
|
# runner's work tree only. The patterns target executable paths INSIDE the
|
|
@@ -85,13 +107,20 @@ pkill -9 -f "${RUNNER_WORK}/_tool/[^ ]*node_modules" 2>/dev/null
|
|
|
85
107
|
rm -rf "${RUNNER_TMP}/pnpm" 2>/dev/null
|
|
86
108
|
rm -rf "${RUNNER_TMP}/setup-pnpm" 2>/dev/null
|
|
87
109
|
|
|
88
|
-
# 3)
|
|
89
|
-
#
|
|
90
|
-
#
|
|
91
|
-
# runner's in-flight download is
|
|
92
|
-
|
|
93
|
-
|
|
94
|
-
|
|
95
|
-
|
|
110
|
+
# 3) Remove this runner's own leftover tool-download temp dirs. The platform's
|
|
111
|
+
# composite actions and workflows create these via
|
|
112
|
+
# `mktemp -d "${RUNNER_TEMP}/<tool>.XXXXXX"`, so every one of them is
|
|
113
|
+
# runner-scoped and a co-resident runner's in-flight download is
|
|
114
|
+
# unreachable here — no age gate is needed (see the issue #343 note above).
|
|
115
|
+
#
|
|
116
|
+
# Globbing is what keeps this bounded: the shell expands these against
|
|
117
|
+
# RUNNER_TMP alone, so the cost is a function of THIS runner's leftovers,
|
|
118
|
+
# never of host-wide churn. Do not replace it with a `find` over a parent.
|
|
119
|
+
# A glob that matches nothing stays literal, and `rm -rf` on a nonexistent
|
|
120
|
+
# path is silent — hence the nullglob-free form plus 2>/dev/null.
|
|
121
|
+
rm -rf "${RUNNER_TMP}"/gitleaks.* 2>/dev/null
|
|
122
|
+
rm -rf "${RUNNER_TMP}"/osv-scanner.* 2>/dev/null
|
|
123
|
+
rm -rf "${RUNNER_TMP}"/semgrep.* 2>/dev/null
|
|
124
|
+
rm -f "${RUNNER_TMP}"/gh-api-err.* 2>/dev/null
|
|
96
125
|
|
|
97
126
|
exit 0
|