gstack/test/helpers/plan-mode-helpers.ts

177 lines
6.6 KiB
TypeScript

/**
* Shared helpers for plan-mode E2E tests.
*
* Four sibling per-skill smoke tests (plan-ceo, plan-eng, plan-design, plan-devex)
* plus the no-op regression test use this helper. The goal: run a review skill
* in plan mode, confirm it goes straight to its Step 0 AskUserQuestion without
* writing files or calling ExitPlanMode first (the vestigial handshake
* regression we fixed in ceo-plan 2026-04-24).
*
* This file was renamed from `plan-mode-handshake-helpers.ts` when the
* handshake was removed. The write-guard detection (no Write/Edit before the
* first AskUserQuestion) is the load-bearing piece that catches silent
* regressions a simple "first question text matches" check would miss.
*/
import * as fs from 'fs';
import * as path from 'path';
import * as os from 'os';
import { execSync } from 'child_process';
import {
runAgentSdkTest,
passThroughNonAskUserQuestion,
resolveClaudeBinary,
type AgentSdkResult,
} from './agent-sdk-runner';
/** Distinctive phrase matching what Claude Code's harness actually injects. */
export const PLAN_MODE_REMINDER =
'Plan mode is active. The user indicated that they do not want you to execute yet';
export interface PlanModeCaptureResult {
sdkResult: AgentSdkResult;
/** Each AskUserQuestion that fired, with its input payload. */
askUserQuestions: Array<{ input: Record<string, unknown>; orderIndex: number }>;
/** Tool-use events in the order they fired (names only). */
toolOrder: string[];
/** Whether any Write or Edit tool fired BEFORE the first AskUserQuestion. */
writeOrEditBeforeAsk: boolean;
/** Whether ExitPlanMode fired BEFORE the first AskUserQuestion. */
exitPlanModeBeforeAsk: boolean;
}
/**
* Run a skill via the Agent SDK with canUseTool intercepting every tool use.
* Inject the plan-mode distinctive phrase into the system prompt, auto-answer
* the first AskUserQuestion (so the skill stops cleanly after Step 0), and
* return the captured events for assertion.
*/
export async function runPlanModeSkillTest(opts: {
/** Skill name, e.g. 'plan-ceo-review'. */
skillName: string;
/**
* For the first AskUserQuestion, pick the option whose label contains this
* substring. Pick a "cheap" answer that terminates the skill quickly (e.g.
* "HOLD SCOPE" for plan-ceo-review).
*/
firstAnswerSubstring: string;
/** If true, DO NOT inject the reminder — used by the no-op regression test. */
omitPlanModeReminder?: boolean;
/** Max turns for the SDK call (default 4 — Step 0 + answer should fit). */
maxTurns?: number;
}): Promise<PlanModeCaptureResult> {
const { skillName, firstAnswerSubstring, omitPlanModeReminder, maxTurns } = opts;
const askUserQuestions: PlanModeCaptureResult['askUserQuestions'] = [];
const toolOrder: string[] = [];
let toolIndex = 0;
let firstAskIndex = -1;
const workingDir = fs.mkdtempSync(
path.join(os.tmpdir(), `plan-mode-${skillName}-`),
);
const binary = resolveClaudeBinary();
try {
// In real plan mode Claude Code injects a system-reminder; in SDK tests we
// use systemPrompt.append which the model treats as equally authoritative.
const reminderAppend = omitPlanModeReminder
? ''
: `\n\n<system-reminder>\n${PLAN_MODE_REMINDER}. This supercedes any other instructions you have received.\n</system-reminder>\n`;
const sdkResult = await runAgentSdkTest({
systemPrompt: {
type: 'preset',
preset: 'claude_code',
append: reminderAppend,
},
userPrompt: `Read the skill file at ${path.resolve(
import.meta.dir,
'..',
'..',
skillName,
'SKILL.md',
)} and follow its instructions. There is no real plan to review — just start the skill and respond to any AskUserQuestion that fires.`,
workingDirectory: workingDir,
maxTurns: maxTurns ?? 4,
allowedTools: ['Read', 'Grep', 'Glob', 'Bash'],
...(binary ? { pathToClaudeCodeExecutable: binary } : {}),
canUseTool: async (toolName, input) => {
toolOrder.push(toolName);
if (toolName === 'AskUserQuestion') {
if (firstAskIndex === -1) firstAskIndex = toolIndex;
askUserQuestions.push({ input, orderIndex: toolIndex });
toolIndex++;
// Auto-answer the FIRST question with the configured substring; for
// later questions, pick the first option to keep the run short.
const q = (input.questions as Array<{ question: string; options: Array<{ label: string }> }>)[0];
const isFirst = askUserQuestions.length === 1;
const matched = isFirst
? q.options.find((o) => o.label.toLowerCase().includes(firstAnswerSubstring.toLowerCase()))
: undefined;
const answer = matched ? matched.label : q.options[0]!.label;
return {
behavior: 'allow',
updatedInput: {
questions: input.questions,
answers: { [q.question]: answer },
},
};
}
toolIndex++;
return passThroughNonAskUserQuestion(toolName, input);
},
});
const writeOrEditBeforeAsk =
firstAskIndex > 0 &&
toolOrder.slice(0, firstAskIndex).some((t) => t === 'Write' || t === 'Edit');
const exitPlanModeBeforeAsk =
firstAskIndex > 0 &&
toolOrder.slice(0, firstAskIndex).some((t) => t === 'ExitPlanMode');
return {
sdkResult,
askUserQuestions,
toolOrder,
writeOrEditBeforeAsk,
exitPlanModeBeforeAsk,
};
} finally {
try {
fs.rmSync(workingDir, { recursive: true, force: true });
} catch { /* ignore cleanup errors */ }
}
}
/**
* Assert a captured AskUserQuestion is NOT the old vestigial handshake
* (A=exit-and-rerun / C=cancel). The handshake is gone — if a test ever sees
* one again, that's the regression we're guarding against.
*/
export function assertNotHandshakeShape(
aq: { input: Record<string, unknown> },
): void {
const questions = aq.input.questions as Array<{
question: string;
options: Array<{ label: string }>;
}>;
if (!questions || questions.length === 0) return;
const q = questions[0]!;
const labels = q.options.map((o) => o.label.toLowerCase());
const looksLikeHandshake =
labels.some((l) => l.includes('exit') && l.includes('rerun')) &&
labels.some((l) => l.includes('cancel'));
if (looksLikeHandshake) {
throw new Error(
`First AskUserQuestion looks like the vestigial plan-mode handshake ` +
`(options: ${labels.join(', ')}). The handshake was removed; skills ` +
`should go straight to their Step 0 question in plan mode.`,
);
}
}
export { execSync };