forked from garrytan/gstack
-
Notifications
You must be signed in to change notification settings - Fork 0
Expand file tree
/
Copy pathskill-e2e-plan-ceo-finding-count.test.ts
More file actions
253 lines (238 loc) · 9.68 KB
/
skill-e2e-plan-ceo-finding-count.test.ts
File metadata and controls
253 lines (238 loc) · 9.68 KB
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
/**
* /plan-ceo-review per-finding AskUserQuestion count (periodic, paid, real-PTY).
*
* Asserts the load-bearing rule "One issue = one AskUserQuestion call" by
* driving /plan-ceo-review against a 5-finding seeded plan and counting
* distinct review-phase AUQs. Passes when count is in [N-1, N+2].
*
* Two tests in this file:
* - 5-finding distinct fixture: count band assertion + D19 review-report-at-bottom.
* - 2-finding paired control (D12 positive control): related findings still
* produce 2 distinct AUQs, not 1 batched, when the rule is honored.
*
* Tier: periodic. Each run drives Step 0 + 11 review sections end-to-end
* (~25 min, ~$5/run). Sequential by default per plan §D15. See
* test/helpers/claude-pty-runner.ts for runPlanSkillCounting internals.
*/
import { describe, test } from 'bun:test';
import * as fs from 'node:fs';
import {
runPlanSkillCounting,
ceoStep0Boundary,
assertReviewReportAtBottom,
type AskUserQuestionFingerprint,
} from './helpers/claude-pty-runner';
/**
* /plan-ceo-review's first AUQ asks "what scope?" with options like
* 1. Branch diff vs main
* 2. A specific plan file or design doc
* 3. An idea you'll describe inline
* ...
* 7. Skip interview and plan immediately
*
* The default pick (1) routes to "branch diff vs main" — the wrong target
* for our seeded fixture (the agent would review the gstack PR itself,
* recursively). Picking "Skip interview and plan immediately" bypasses
* Step 0 and routes the agent to review the chat context (where our
* follow-up plan was pasted).
*/
function pickSkipInterview(fp: AskUserQuestionFingerprint): number {
const skipOpt = fp.options.find((o) =>
/skip\s+interview|plan\s+immediately/i.test(o.label),
);
if (skipOpt) return skipOpt.index;
// Fallback: "describe inline" also routes to using our pasted plan.
const inlineOpt = fp.options.find((o) =>
/describe.*inline|inline.*idea/i.test(o.label),
);
if (inlineOpt) return inlineOpt.index;
return 1;
}
const shouldRun = !!process.env.EVALS && process.env.EVALS_TIER === 'periodic';
const describeE2E = shouldRun ? describe : describe.skip;
const N_DISTINCT = 5;
const FLOOR_DISTINCT = N_DISTINCT - 1; // 4 (D11)
const CEILING_DISTINCT = N_DISTINCT + 2; // 7 (D11)
const N_PAIRED = 2;
const FLOOR_PAIRED = 2;
const CEILING_PAIRED = 4;
const PLAN_CEO_5_FINDINGS = [
'Please review this plan thoroughly. As you go, write your plan-mode plan to /tmp/gstack-test-plan-ceo.md (use Edit/Write to that exact path).',
'',
'# Plan: Payment Processing Integration',
'',
'## Architecture',
"We're adding a new `PaymentService` class that will handle Stripe webhooks.",
'This bypasses the existing `WebhookDispatcher` module — we want a clean',
'namespace separation.',
'',
'## Database access',
'The new endpoint reads `request.params.userId` directly into a raw SQL',
'fragment for the lookup query.',
'',
'## Webhook fan-out',
'On payment success we update the user record AND fire a notification email.',
'Both happen inline; no error handling on the email leg.',
'',
'## Tests',
"None planned. We'll rely on the existing integration suite catching regressions.",
'',
'## Performance',
'Each webhook lookup hits the database for the user, then fetches each',
'order in a loop.',
].join('\n');
const PLAN_CEO_2_PAIRED_FINDINGS = [
'Please review this plan thoroughly. As you go, write your plan-mode plan to /tmp/gstack-test-plan-ceo-paired.md (use Edit/Write to that exact path).',
'',
'# Plan: Payment Processing — Test Coverage',
'',
'## Tests',
'We need test coverage for `processPayment()`. Specifically:',
'1. The happy path (successful Stripe charge — assert correct receipt is generated).',
'2. The error/timeout path (Stripe returns 502 — assert retry-with-backoff fires once, then fails clean).',
'',
'Currently neither has a unit test. These are deliberately separate concerns:',
'the success path is correctness, the failure path is graceful degradation.',
].join('\n');
const PLAN_CEO_PATH = '/tmp/gstack-test-plan-ceo.md';
const PLAN_CEO_PAIRED_PATH = '/tmp/gstack-test-plan-ceo-paired.md';
describeE2E('/plan-ceo-review per-finding AskUserQuestion count (periodic)', () => {
test(
`5-finding plan emits ${FLOOR_DISTINCT}-${CEILING_DISTINCT} review-phase AskUserQuestions`,
async () => {
try {
fs.rmSync(PLAN_CEO_PATH, { force: true });
} catch {
/* best-effort */
}
const obs = await runPlanSkillCounting({
skillName: 'plan-ceo-review',
slashCommand: '/plan-ceo-review',
followUpPrompt: PLAN_CEO_5_FINDINGS,
isLastStep0AUQ: ceoStep0Boundary,
reviewCountCeiling: CEILING_DISTINCT + 1, // hard cap above assertion ceiling
firstAUQPick: pickSkipInterview, // bypass scope-selection, route to review
cwd: process.cwd(),
timeoutMs: 1_500_000, // 25 min
env: { QUESTION_TUNING: 'false', EXPLAIN_LEVEL: 'default' },
});
try {
if (!['plan_ready', 'completion_summary', 'ceiling_reached'].includes(obs.outcome)) {
throw new Error(
`plan-ceo-review finding-count FAILED: outcome=${obs.outcome}\n` +
`step0=${obs.step0Count} review=${obs.reviewCount} elapsed=${obs.elapsedMs}ms\n` +
`fingerprints (last 8):\n` +
obs.fingerprints
.slice(-8)
.map(
(f, i) =>
` ${i}. preReview=${f.preReview} sig=${f.signature.slice(0, 12)} prompt="${f.promptSnippet.slice(0, 60)}"`,
)
.join('\n') +
`\n--- evidence (last 3KB) ---\n${obs.evidence}`,
);
}
if (obs.reviewCount < FLOOR_DISTINCT) {
throw new Error(
`BAND FAIL (below floor): reviewCount=${obs.reviewCount} < FLOOR=${FLOOR_DISTINCT}.\n` +
`Likely batching regression — agent collapsed multiple findings into fewer questions.\n` +
`Fingerprints (review-phase only):\n` +
obs.fingerprints
.filter((f) => !f.preReview)
.map((f) => ` - "${f.promptSnippet.slice(0, 80)}"`)
.join('\n'),
);
}
if (obs.reviewCount > CEILING_DISTINCT) {
throw new Error(
`BAND FAIL (above ceiling): reviewCount=${obs.reviewCount} > CEILING=${CEILING_DISTINCT}.\n` +
`Possible over-asking regression. Review-phase fingerprints:\n` +
obs.fingerprints
.filter((f) => !f.preReview)
.map((f) => ` - "${f.promptSnippet.slice(0, 80)}"`)
.join('\n'),
);
}
// D19: review report at bottom of plan file.
if (!fs.existsSync(PLAN_CEO_PATH)) {
throw new Error(
`D19 FAIL: agent did not produce expected plan file at ${PLAN_CEO_PATH}.\n` +
`Either the agent ignored the path instruction in the follow-up prompt, or\n` +
`the helper exited before the agent wrote the file. ` +
`outcome=${obs.outcome} review=${obs.reviewCount}`,
);
}
const planContent = fs.readFileSync(PLAN_CEO_PATH, 'utf-8');
const verdict = assertReviewReportAtBottom(planContent);
if (!verdict.ok) {
throw new Error(
`D19 FAIL: plan file at ${PLAN_CEO_PATH} ${verdict.reason}\n` +
(verdict.trailingHeadings
? `Trailing headings: ${verdict.trailingHeadings.join(' | ')}\n`
: '') +
`--- plan content (last 1KB) ---\n${planContent.slice(-1024)}`,
);
}
} finally {
try {
fs.rmSync(PLAN_CEO_PATH, { force: true });
} catch {
/* best-effort */
}
}
},
1_700_000,
);
test(
`paired-finding positive control: ${N_PAIRED} related findings produce ${FLOOR_PAIRED}-${CEILING_PAIRED} AskUserQuestions`,
async () => {
try {
fs.rmSync(PLAN_CEO_PAIRED_PATH, { force: true });
} catch {
/* best-effort */
}
const obs = await runPlanSkillCounting({
skillName: 'plan-ceo-review',
slashCommand: '/plan-ceo-review',
followUpPrompt: PLAN_CEO_2_PAIRED_FINDINGS,
isLastStep0AUQ: ceoStep0Boundary,
reviewCountCeiling: CEILING_PAIRED + 1,
cwd: process.cwd(),
timeoutMs: 1_500_000,
env: { QUESTION_TUNING: 'false', EXPLAIN_LEVEL: 'default' },
});
try {
if (!['plan_ready', 'completion_summary', 'ceiling_reached'].includes(obs.outcome)) {
throw new Error(
`paired-finding control FAILED: outcome=${obs.outcome}\n` +
`step0=${obs.step0Count} review=${obs.reviewCount}\n` +
`--- evidence (last 3KB) ---\n${obs.evidence}`,
);
}
if (obs.reviewCount < FLOOR_PAIRED) {
throw new Error(
`PAIRED CONTROL FAIL: reviewCount=${obs.reviewCount} < FLOOR=${FLOOR_PAIRED}.\n` +
`Two deliberately related findings were batched into <2 questions — the rule failed under D12.\n` +
`Review-phase fingerprints:\n` +
obs.fingerprints
.filter((f) => !f.preReview)
.map((f) => ` - "${f.promptSnippet.slice(0, 80)}"`)
.join('\n'),
);
}
if (obs.reviewCount > CEILING_PAIRED) {
throw new Error(
`PAIRED CONTROL FAIL: reviewCount=${obs.reviewCount} > CEILING=${CEILING_PAIRED} (over-asking on a 2-finding fixture).`,
);
}
} finally {
try {
fs.rmSync(PLAN_CEO_PAIRED_PATH, { force: true });
} catch {
/* best-effort */
}
}
},
1_700_000,
);
});