feat(commander): add GPT prompt boundary lint

2026-05-23 13:24:15 +00:00
parent e2646763c0
commit 77a8c6b878
10 changed files with 434 additions and 5 deletions
@@ -0,0 +1,245 @@
+import { readFileSync } from "node:fs";
+
+export type CommanderPromptLintKind = "gpt55-pr";
+export type CommanderPromptLintRiskLevel = "low" | "medium" | "high";
+
+interface ClauseCheck {
+  id: string;
+  label: string;
+  required: boolean;
+  matched: boolean;
+  severity: "medium" | "high";
+  suggestion: string;
+}
+
+export interface CommanderPromptLintResult {
+  ok: boolean;
+  kind: CommanderPromptLintKind;
+  missingClauses: string[];
+  riskLevel: CommanderPromptLintRiskLevel;
+  suggestedPatchSnippet: string;
+  promptShape: {
+    chars: number;
+    lines: number;
+    textEchoed: false;
+  };
+  policy: {
+    advisoryOnly: true;
+    mutatesScheduler: false;
+    changesCodexSubmitDefault: false;
+    printsPromptText: false;
+    supportedInputs: Array<"--stdin" | "--prompt-file">;
+    reference: string;
+  };
+}
+
+interface ParsedPromptLintArgs {
+  kind: CommanderPromptLintKind;
+  prompt: string;
+}
+
+const gpt55PrSnippet = [
+  "GPT-5.5 runner boundary:",
+  "- You may create/update the branch and PR, resolve conflicts, rebase/update from the target branch, and self-merge/close an ordinary PR when checks pass and the task boundary is satisfied.",
+  "- You may use repo-owned CI/CD, publish, or equivalent controlled build paths to build/publish DEV images or artifacts, and must report commit, image tag, digest, artifact report, and validation evidence.",
+  "- DEV deploy apply, rollout, and live health verification are owned by the host commander unless this prompt explicitly contains ROLLOUT_OK.",
+  "- Without explicit ROLLOUT_OK, do not acquire the DEV CD lock, run deploy apply, perform rollout/restart, or compete with host commander live verification.",
+  "- Forbidden: PROD mutation, reading or printing secrets, manual database writes, destructive rollback, Code Queue backend/scheduler restart, or interrupt/cancel running tasks unless explicitly authorized.",
+].join("\n");
+
+function hasFlag(args: string[], flag: string): boolean {
+  return args.includes(flag);
+}
+
+function optionValue(args: string[], names: string[]): string | undefined {
+  for (const name of names) {
+    const index = args.indexOf(name);
+    if (index === -1) continue;
+    const value = args[index + 1];
+    if (value === undefined || value.startsWith("--")) throw new Error(`${name} requires a value`);
+    return value;
+  }
+  return undefined;
+}
+
+function assertKnownOptions(args: string[]): void {
+  const flags = new Set(["--stdin"]);
+  const valueOptions = new Set(["--kind", "--prompt-file", "--file"]);
+  for (let index = 0; index < args.length; index += 1) {
+    const arg = args[index] ?? "";
+    if (!arg.startsWith("--")) throw new Error(`commander prompt-lint does not accept positional prompt text; use --stdin or --prompt-file`);
+    if (flags.has(arg)) continue;
+    if (valueOptions.has(arg)) {
+      index += 1;
+      if (index >= args.length) throw new Error(`${arg} requires a value`);
+      continue;
+    }
+    throw new Error(`unknown commander prompt-lint option: ${arg}`);
+  }
+}
+
+function parseKind(raw: string | undefined): CommanderPromptLintKind {
+  const kind = raw ?? "gpt55-pr";
+  if (kind !== "gpt55-pr") throw new Error(`unsupported commander prompt-lint kind: ${kind}`);
+  return kind;
+}
+
+function readPromptFromArgs(args: string[]): string {
+  const promptFile = optionValue(args, ["--prompt-file", "--file"]);
+  const promptStdin = hasFlag(args, "--stdin");
+  const sources = [promptFile !== undefined, promptStdin].filter(Boolean).length;
+  if (sources !== 1) throw new Error("commander prompt-lint requires exactly one prompt source: --prompt-file or --stdin");
+  const prompt = promptFile !== undefined
+    ? (promptFile === "-" ? readFileSync(0, "utf8") : readFileSync(promptFile, "utf8"))
+    : readFileSync(0, "utf8");
+  if (prompt.trim().length === 0) throw new Error("commander prompt-lint prompt must not be empty");
+  return prompt;
+}
+
+export function parseCommanderPromptLintArgs(args: string[]): ParsedPromptLintArgs {
+  assertKnownOptions(args);
+  return {
+    kind: parseKind(optionValue(args, ["--kind"])),
+    prompt: readPromptFromArgs(args),
+  };
+}
+
+function any(patterns: RegExp[], prompt: string): boolean {
+  return patterns.some((pattern) => pattern.test(prompt));
+}
+
+function rolloutOkPresent(prompt: string): boolean {
+  return /\bROLLOUT_OK\b/u.test(prompt);
+}
+
+function gpt55PrClauseChecks(prompt: string): ClauseCheck[] {
+  const prAuthorization = any([
+    /\bPR\b[^\n。]{0,120}(?:create|update|push|branch|merge|close|rebase|conflict)/iu,
+    /(?:创建|更新|push|提交|合并|关闭|rebase|解决冲突)[^\n。]{0,80}\bPR\b/iu,
+  ], prompt) && any([
+    /self[- ]?(?:merge|close)|自行(?:合并|关闭|收口)|自合并|自收口/iu,
+    /merge\/close|合并\/关闭/iu,
+  ], prompt) && any([
+    /\brebase\b|\bupdate\b|解决冲突|更新(?:目标|base|master|分支)/iu,
+  ], prompt);
+
+  const artifactAuthorization = any([
+    /\b(?:build|publish)\b[^\n。]{0,80}\b(?:artifact|image|DEV image|镜像|制品)\b/iu,
+    /\b(?:artifact|image|镜像|制品)\b[^\n。]{0,80}\b(?:build|publish|tag|digest|构建|发布)\b/iu,
+    /(?:构建|发布)[^\n。]{0,80}(?:镜像|制品|artifact|image)/iu,
+  ], prompt) && any([
+    /\b(?:tag|digest|artifact report|image tag)\b/iu,
+    /(?:镜像\s*tag|digest|制品报告|artifact report)/iu,
+  ], prompt);
+
+  const hostOwnsDevRollout = any([
+    /DEV[^\n。]{0,120}(?:deploy apply|rollout|live health|live verification)[^\n。]{0,120}(?:host commander|commander|host|统一执行|统一处理|统一复验)/iu,
+    /(?:host commander|commander|host|指挥官)[^\n。]{0,120}DEV[^\n。]{0,120}(?:deploy apply|rollout|live health|live verification|发布|上线|复验)/iu,
+    /DEV[^\n。]{0,80}(?:发布|上线|rollout|复验)[^\n。]{0,80}(?:host|commander|指挥官|统一)/iu,
+  ], prompt);
+
+  const forbidsRunnerRolloutUnlessOk = any([
+    /unless[^\n。]{0,80}\bROLLOUT_OK\b/iu,
+    /未显式[^\n。]{0,60}\bROLLOUT_OK\b[^\n。]{0,80}(?:不要|不得|禁止|do not|don't|must not)/iu,
+    /\bROLLOUT_OK\b[^\n。]{0,80}(?:才|unless|only if|explicit)/iu,
+  ], prompt) && any([
+    /(?:do not|don't|must not|禁止|不要|不得)[^\n。]{0,100}(?:DEV CD lock|deploy apply|rollout|live health|rollout restart|竞争|抢)/iu,
+    /(?:DEV CD lock|deploy apply|rollout|live health|rollout restart|竞争|抢)[^\n。]{0,100}(?:do not|don't|must not|禁止|不要|不得)/iu,
+  ], prompt);
+
+  const prodForbidden = any([
+    /(?:禁止|不要|不得|must not|do not|don't|\bno\b|forbid|forbidden)[^\n。]{0,80}(?:PROD|prod|production|生产)/iu,
+    /(?:PROD|prod|production|生产)[^\n。]{0,80}(?:禁止|不要|不得|must not|do not|don't|\bno\b|forbid|forbidden|not allowed)/iu,
+  ], prompt);
+  const secretForbidden = any([
+    /(?:禁止|不要|不得|must not|do not|don't|\bno\b|forbid|forbidden)[^\n。]{0,80}(?:secret|token|credential|密钥|凭证)/iu,
+    /(?:secret|token|credential|密钥|凭证)[^\n。]{0,80}(?:禁止|不要|不得|must not|do not|don't|\bno\b|forbid|forbidden|not allowed|读取|打印)/iu,
+  ], prompt);
+  const dbForbidden = any([
+    /(?:禁止|不要|不得|must not|do not|don't|\bno\b|forbid|forbidden)[^\n。]{0,80}(?:database|DB|数据库)/iu,
+    /(?:database|DB|数据库)[^\n。]{0,80}(?:manual|手工|patch|write|写入|禁止|不要|不得|must not|do not|don't|\bno\b)/iu,
+  ], prompt);
+  const rollbackForbidden = any([
+    /(?:禁止|不要|不得|must not|do not|don't|\bno\b|forbid|forbidden)[^\n。]{0,80}(?:destructive rollback|破坏性回滚|回滚)/iu,
+    /(?:destructive rollback|破坏性回滚)[^\n。]{0,80}(?:禁止|不要|不得|must not|do not|don't|\bno\b|forbid|forbidden|not allowed)/iu,
+  ], prompt);
+
+  return [
+    {
+      id: "pr-self-merge-rebase-authorization",
+      label: "PR/自合并/rebase/update 授权文本",
+      required: true,
+      matched: prAuthorization,
+      severity: "high",
+      suggestion: "明确授权 runner 创建/更新 PR、rebase/update/解决冲突，并在普通 PR 满足任务边界和检查通过时自合并/关闭。",
+    },
+    {
+      id: "artifact-build-publish-authorization",
+      label: "build/publish artifact 授权文本",
+      required: true,
+      matched: artifactAuthorization,
+      severity: "high",
+      suggestion: "明确授权使用 repo-owned CI/CD 或受控构建路径发布 DEV image/artifact，并要求回报 tag、digest 和 artifact report。",
+    },
+    {
+      id: "host-owned-dev-rollout",
+      label: "DEV deploy apply/rollout/live verification 由 host 统一执行文本",
+      required: true,
+      matched: hostOwnsDevRollout,
+      severity: "high",
+      suggestion: "写明 DEV deploy apply、rollout 和 live health verification 默认由 host commander 统一执行。",
+    },
+    {
+      id: "runner-rollout-forbidden-without-rollout-ok",
+      label: "未显式 ROLLOUT_OK 时禁止 runner rollout",
+      required: true,
+      matched: forbidsRunnerRolloutUnlessOk,
+      severity: "high",
+      suggestion: "写明未显式包含 ROLLOUT_OK 时，runner 不得抢 DEV CD lock、deploy apply、rollout 或 live verification。",
+    },
+    {
+      id: "prod-secret-db-rollback-boundary",
+      label: "PROD/secret/DB/破坏性回滚边界",
+      required: true,
+      matched: prodForbidden && secretForbidden && dbForbidden && rollbackForbidden,
+      severity: "high",
+      suggestion: "同时禁止 PROD mutation、密钥读取/打印、数据库手工写入和破坏性回滚。",
+    },
+  ];
+}
+
+export function lintCommanderPrompt(prompt: string, kind: CommanderPromptLintKind = "gpt55-pr"): CommanderPromptLintResult {
+  const clauses = gpt55PrClauseChecks(prompt);
+  const missingClauses = clauses.filter((clause) => clause.required && !clause.matched).map((clause) => clause.id);
+  const highMissing = clauses.some((clause) => clause.severity === "high" && !clause.matched);
+  const rolloutOk = rolloutOkPresent(prompt);
+  const riskLevel: CommanderPromptLintRiskLevel = missingClauses.length === 0
+    ? rolloutOk ? "medium" : "low"
+    : highMissing ? "high" : "medium";
+
+  return {
+    ok: missingClauses.length === 0,
+    kind,
+    missingClauses,
+    riskLevel,
+    suggestedPatchSnippet: missingClauses.length === 0 ? "" : gpt55PrSnippet,
+    promptShape: {
+      chars: prompt.length,
+      lines: prompt.split(/\r\n|\r|\n/u).length,
+      textEchoed: false,
+    },
+    policy: {
+      advisoryOnly: true,
+      mutatesScheduler: false,
+      changesCodexSubmitDefault: false,
+      printsPromptText: false,
+      supportedInputs: ["--stdin", "--prompt-file"],
+      reference: "docs/reference/host-codex-commander.md",
+    },
+  };
+}
+
+export function runCommanderPromptLintCommand(args: string[]): CommanderPromptLintResult {
+  const options = parseCommanderPromptLintArgs(args);
+  return lintCommanderPrompt(options.prompt, options.kind);
+}