mirror of
https://github.com/garrytan/gstack.git
synced 2026-08-31 18:30:39 +02:00
Six paid PTY tests wrote their expected plan artifact to a FIXED shared
/tmp path ('/tmp/gstack-test-plan-<mode>.md') and rmSync'd it in
finally — under --retry 1, EVALS_JOBS>1, or two concurrent worktrees, a
sibling's cleanup deletes this run's artifact and the D19 'agent did
not produce expected plan file' assertion fires spuriously. Each test
now mkdtemps its own dir, interpolates the unique path into the agent
prompt (fixture-sourced prompts get a replaceAll + drift guard that
throws if the fixture's literal ever moves), and cleans up its own dir.
The 18 cwd:-into-the-live-repo sites were audited: all deliberate
(skill registry + hermetic pre-trusted dir, in-repo gen renders, git
history reads, slug resolution) — each now carries a
'// LIVE-REPO CWD: <reason>' comment so the next audit can tell
deliberate from accidental.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
107 lines
4.5 KiB
TypeScript
107 lines
4.5 KiB
TypeScript
/**
|
|
* /plan-eng-review multi-finding batching regression (periodic, paid, real-PTY).
|
|
*
|
|
* Catches the specific shape of the May 2026 transcript bug that the
|
|
* single-finding gate-tier floor test cannot detect: a model that fires
|
|
* one AskUserQuestion and then batches the remaining findings into a
|
|
* single "## Decisions to confirm" plan write + ExitPlanMode.
|
|
*
|
|
* Why a separate test from skill-e2e-plan-eng-finding-floor:
|
|
* - The gate-tier floor (runPlanSkillFloorCheck) exits on the first AUQ
|
|
* render and returns success. A model that fires once-then-batches
|
|
* would pass that test trivially.
|
|
* - This test uses runPlanSkillCounting at periodic tier (~25 min budget,
|
|
* N-AUQ tracking, ceiling-bounded retries) to actually count distinct
|
|
* review-phase AUQs and assert the model fires one per finding.
|
|
*
|
|
* Why a separate test from skill-e2e-plan-eng-finding-count (the existing
|
|
* 5-finding count test):
|
|
* - The fixture here mirrors the D1-D4 transcript shape (4 findings) and
|
|
* the floor matches that exact threshold (3, the [N-1] tolerance band).
|
|
* This is the tightest regression test for the original bug class —
|
|
* not a band-around-N test, but a "did the agent batch?" test.
|
|
*
|
|
* Tier: periodic (~25 min, ~$5/run). Sequential by default.
|
|
*/
|
|
|
|
import { test } from 'bun:test';
|
|
import { describeE2ETier } from './helpers/e2e-gate';
|
|
import * as fs from 'node:fs';
|
|
import * as os from 'node:os';
|
|
import * as path from 'node:path';
|
|
import {
|
|
runPlanSkillCounting,
|
|
engStep0Boundary,
|
|
} from './helpers/claude-pty-runner';
|
|
import { FORCING_BATCHING_ENG } from './fixtures/forcing-finding-seeds';
|
|
|
|
const describeE2E = describeE2ETier('periodic');
|
|
|
|
const N = 4;
|
|
const FLOOR = N - 1; // 3 — agent must fire at least one AUQ per non-batched finding
|
|
|
|
/** Plan-file target baked into the FORCING_BATCHING_ENG fixture prompt.
|
|
* Rewritten per-run to a mkdtemp path so concurrent runs (--retry,
|
|
* EVALS_JOBS>1, sibling worktrees) never share one /tmp artifact. */
|
|
const FIXTURE_PLAN_PATH = '/tmp/gstack-test-plan-eng-batching.md';
|
|
|
|
describeE2E('/plan-eng-review multi-finding batching regression (periodic)', () => {
|
|
test(
|
|
`4-finding plan emits >= ${FLOOR} review-phase AskUserQuestions (no batching)`,
|
|
async () => {
|
|
const tmpDir = fs.mkdtempSync(path.join(os.tmpdir(), 'gstack-e2e-plan-eng-batching-'));
|
|
const planPath = path.join(tmpDir, 'gstack-test-plan-eng-batching.md');
|
|
const followUpPrompt = FORCING_BATCHING_ENG.replaceAll(FIXTURE_PLAN_PATH, planPath);
|
|
if (!followUpPrompt.includes(planPath)) {
|
|
throw new Error(
|
|
`fixture drift: FORCING_BATCHING_ENG no longer contains ${FIXTURE_PLAN_PATH} — update FIXTURE_PLAN_PATH`,
|
|
);
|
|
}
|
|
|
|
try {
|
|
const obs = await runPlanSkillCounting({
|
|
skillName: 'plan-eng-review',
|
|
slashCommand: '/plan-eng-review',
|
|
followUpPrompt,
|
|
isLastStep0AUQ: engStep0Boundary,
|
|
reviewCountCeiling: N + 3, // hard cap above floor + tolerance
|
|
// LIVE-REPO CWD: PTY session needs the repo cwd — gstack skill
|
|
// registry + hermetic pre-trusted dir (hermetic-env trustedDirs).
|
|
cwd: process.cwd(),
|
|
timeoutMs: 1_500_000, // 25 min
|
|
env: { QUESTION_TUNING: 'false', EXPLAIN_LEVEL: 'default' },
|
|
});
|
|
|
|
if (!['plan_ready', 'completion_summary', 'ceiling_reached'].includes(obs.outcome)) {
|
|
throw new Error(
|
|
`multi-finding batching test FAILED: outcome=${obs.outcome}\n` +
|
|
`step0=${obs.step0Count} review=${obs.reviewCount} elapsed=${obs.elapsedMs}ms\n` +
|
|
`--- evidence (last 3KB) ---\n${obs.evidence}`,
|
|
);
|
|
}
|
|
if (obs.reviewCount < FLOOR) {
|
|
throw new Error(
|
|
`BATCHING REGRESSION: reviewCount=${obs.reviewCount} < FLOOR=${FLOOR}.\n` +
|
|
`Agent surfaced fewer review-phase AUQs than findings — this is the\n` +
|
|
`May 2026 transcript bug shape: model batched multiple findings into\n` +
|
|
`a single plan write + ExitPlanMode instead of asking one per finding.\n` +
|
|
`Review-phase fingerprints:\n` +
|
|
obs.fingerprints
|
|
.filter((f) => !f.preReview)
|
|
.map((f) => ` - "${f.promptSnippet.slice(0, 80)}"`)
|
|
.join('\n') +
|
|
`\n--- evidence (last 3KB) ---\n${obs.evidence}`,
|
|
);
|
|
}
|
|
} finally {
|
|
try {
|
|
fs.rmSync(tmpDir, { recursive: true, force: true });
|
|
} catch {
|
|
/* best-effort */
|
|
}
|
|
}
|
|
},
|
|
1_700_000,
|
|
);
|
|
});
|