Jyotisha/frontend/tests/rectification-grounding-prompt-20260927.test.ts

/**
 * TASK-rectification-grounding-20260927 T6 (R4 + Skill slicing): what a real
 * rectification Agent sends the model on an evidence turn. Real
 * `runV9AgentTurn` + real `getRectificationV9Agent` (real bound Skill, real
 * tools) + a prompt-recording scripted model; case data is a real local
 * engine response for a public AA chart (9 candidates).
 *
 * Token estimate (for the size table): one token per CJK character plus one
 * per four other characters (`estimateTokens`). "Fixed overhead" is what every
 * call of the turn re-sends regardless of tool results: the system messages
 * plus the tool schemas.
 */
import assert from "node:assert/strict";
import { readFileSync } from "node:fs";
import test from "node:test";
import {
  CASE_ID,
  SESSION_ID,
  TURN_ID,
  USER_ID,
  dossierFixture,
  fakeAccounting,
  receiptHandlers,
} from "./rectification-v9-test-support.ts";
import {
  AA_EVIDENCE_ROWS,
  AA_RANGE,
  AA_TURNS,
  buildGoldenLatest,
  contentText,
  estimateTokens,
  scriptedModel,
  setFocusHandler,
  stubGoldenFetch,
  type RecordedCall,
} from "./rectification-grounding-support.ts";
import { getRectificationV9Agent } from "../src/mastra/agentic-rectification.ts";
import { runV9AgentTurn } from "../src/lib/rectification-agentic/v9/agent-run.ts";
import {
  RECTIFICATION_SKILL_SECTIONS_BY_ACTION,
  sliceRectificationSkill,
} from "../src/lib/rectification-agentic/v9/skill-slice.ts";
import {
  resolveActiveSkillPackage,
  resolveExactSkillPackage,
} from "../src/lib/skill-package-registry.ts";
import { resolve } from "node:path";
import { productDeliveryForDomain } from "../src/lib/rectification-agentic/v9/product-delivery.ts";

export type CallSize = Readonly<{
  call: number;
  systemChars: number;
  systemTokens: number;
  toolsBytes: number;
  toolsTokens: number;
  toolNames: string[];
  fixedTokens: number;
  totalTokens: number;
}>;

export function callSizes(calls: readonly RecordedCall[]): CallSize[] {
  return calls.map((call, index) => {
    const system = call.prompt.filter((message) => message.role === "system").map((message) => contentText(message.content)).join("\n");
    const tools = JSON.stringify(call.tools ?? []);
    const all = call.prompt.map((message) => contentText(message.content)).join("\n");
    return {
      call: index + 1,
      systemChars: system.length,
      systemTokens: estimateTokens(system),
      toolsBytes: Buffer.byteLength(tools, "utf8"),
      toolsTokens: estimateTokens(tools),
      toolNames: Array.isArray(call.tools) ? (call.tools as Array<{ name: string }>).map((tool) => tool.name) : [],
      fixedTokens: estimateTokens(system) + estimateTokens(tools),
      totalTokens: estimateTokens(all) + estimateTokens(tools),
    };
  });
}

/** One evidence turn: read-case, then one short body (no write claimed). */
export async function recordEvidenceTurn(productDomain?: string) {
  const { latest, compute } = await buildGoldenLatest();
  const { model, calls } = scriptedModel([
    { parts: [{ tool: { name: "rectification-read-case", input: { caseId: CASE_ID } } }], finish: "tool-calls" },
    { parts: [{ text: "2014 年 8 月结婚这件事拿去和星盘对照了。" }], finish: "stop" },
  ]);
  const active = resolveActiveSkillPackage("jyotish-birth-time-rectification");
  const accounting = fakeAccounting({
    ...receiptHandlers,
    // The case is bound to the active Skill (10.0.x layout), as a new case is.
    get_agentic_rectification_skill_identity: () => ({
      skill_name: active.name,
      skill_version: active.version,
      skill_sha256: active.sha256,
      skill_source_commit: active.sourceCommit,
    }),
    insert_agentic_rectification_skill_run_receipt: () => ({
      receipt_id: "88888888-8888-4888-8888-888888888888",
      skill_name: active.name,
      skill_version: active.version,
      skill_sha256: active.sha256,
      source_commit: active.sourceCommit,
    }),
    get_agentic_rectification_case_dossier: () => {
      const dossier = dossierFixture({
        evidence: AA_EVIDENCE_ROWS,
        latestResult: latest,
        turns: AA_TURNS as never,
        candidateRange: AA_RANGE as never,
      }) as { case: Record<string, unknown> };
      return { ...dossier, case: { ...dossier.case, skill_version: active.version,
        ...(productDomain ? { rectification_domain: productDomain } : {}),
      } };
    },
    get_agentic_rectification_case_compute: () => compute,
    set_agentic_rectification_conversation_focus: setFocusHandler,
    append_agentic_rectification_turn: () => ({ turn_id: TURN_ID, should_execute: true, status: "pending" }),
    finalize_agentic_rectification_turn: () => ({ turn_id: TURN_ID, status: "completed", idempotent: false }),
  }, { fallback: () => null });
  const restore = stubGoldenFetch();
  try {
    await runV9AgentTurn({
      userId: USER_ID,
      caseId: CASE_ID,
      sessionId: SESSION_ID,
      requestId: "req-grounding-prompt",
      action: "evidence",
      message: "2014 年 8 月结婚。",
      modelName: "fake",
      accounting: accounting.client as never,
      billing: { reserve: async () => ({ success: true, status: 200 }), complete: async () => true, release: async () => true },
      emit: () => {},
      expectedWrite: "none",
      buildAgent: async (turnId, skillPackage, attemptId) => getRectificationV9Agent(
        { id: "fake", label: "fake", description: "", creditCost: 1, isDefault: true, mode: "compatible", model: model as never } as never,
        { userId: USER_ID, caseId: CASE_ID, turnId, attemptId, userMessage: "2014 年 8 月结婚。", accounting: accounting.client as never,
          ...(productDomain ? { productDelivery: productDeliveryForDomain(productDomain) } : {}),
        },
        skillPackage,
      ),
    });
  } finally {
    restore();
  }
  return calls;
}

function activeSkillBody(): string {
  const skill = resolveActiveSkillPackage("jyotish-birth-time-rectification");
  const raw = readFileSync(resolve(skill.resolvedPath, "SKILL.md"), "utf8");
  return raw.replace(/^---\r?\n[\s\S]*?\r?\n---\r?\n/, "").trim();
}

test("R4: an evidence-turn call carries no <available_skills> block and no skill tools", async () => {
  const calls = await recordEvidenceTurn();
  assert.ok(calls.length >= 2);
  for (const call of calls) {
    const system = call.prompt.filter((message) => message.role === "system").map((message) => contentText(message.content)).join("\n");
    assert.doesNotMatch(system, /<available_skills>/);
    assert.doesNotMatch(system, /Skills are NOT tools|call the `skill` tool/);
    const names = Array.isArray(call.tools) ? (call.tools as Array<{ name: string }>).map((tool) => tool.name) : [];
    assert.equal(names.includes("skill"), false);
    assert.equal(names.includes("skill_search"), false);
  }
});

test("T6: the evidence-turn bootstrap carries Skill §5 and §7 only", async () => {
  const calls = await recordEvidenceTurn();
  const system = calls[0]!.prompt.filter((message) => message.role === "system").map((message) => contentText(message.content)).join("\n");
  assert.match(system, /## 5\. ConversationFocus 与意图承接/);
  assert.match(system, /## 7\. 批量证据与日期真实性/);
  for (const heading of ["## 1. ", "## 2. ", "## 3. ", "## 4. ", "## 6. ", "## 8. ", "## 9. ", "## 10. ", "## 11. "]) {
    assert.equal(system.includes(heading), false, `${heading} in ${system.slice(system.indexOf(heading) - 200, system.indexOf(heading) + 200)}`);
  }
  // full_diagnostics-only fields described in §6 are not in the evidence turn.
  assert.doesNotMatch(system, /guided_collect_windows|method_followup_plan/);
  assert.match(system, /本轮是证据轮/);
  assert.match(system, /# Jyotish 生时校正（V10）/);
});

test("T6: fixed overhead per evidence-turn call is under 8K estimated tokens", async () => {
  const sizes = callSizes(await recordEvidenceTurn());
  console.log(JSON.stringify({ scope: "T6 evidence turn sizes", sizes }));
  for (const size of sizes) assert.ok(size.fixedTokens < 8_000, JSON.stringify(size));
});

test("segment-product real model calls keep target identity in the evidence slice within the original 8K budget", async () => {
  const calls = await recordEvidenceTurn("career");
  assert.ok(calls.length >= 2);
  for (const call of calls) {
    const system = call.prompt.filter(message => message.role === "system").map(message => contentText(message.content)).join("\n");
    assert.match(system, /服务端持久产品身份/);
    assert.match(system, /segment-product-v1/);
    assert.match(system, /"targets":\["D1","D10"\]/);
    assert.match(system, /没有摘要仍是新产品/);
    assert.match(system, /代表候选签名只是诊断/);
    assert.match(system, /## 5\. ConversationFocus 与意图承接/);
    assert.match(system, /## 7\. 批量证据与日期真实性/);
    for (const heading of ["## 1. ", "## 2. ", "## 3. ", "## 4. ", "## 6. ", "## 8. ", "## 9. ", "## 10. ", "## 11. "]) assert.equal(system.includes(heading), false);
    assert.doesNotMatch(system, /<available_skills>/);
  }
  for (const size of callSizes(calls)) assert.ok(size.fixedTokens < 8_000, JSON.stringify(size));
});

test("T6: the slice rule is a code constant; other actions and unknown layouts get the whole Skill", () => {
  assert.deepEqual(RECTIFICATION_SKILL_SECTIONS_BY_ACTION.evidence, ["ConversationFocus 与意图承接", "批量证据与日期真实性"]);
  // 原值: opening / read_only 都是 "all"，开场轮发整本 Skill
  // 新值: 开场轮只附「必须先读、OpeningPolicy、ConversationFocus、可调用工具」，只读轮只附「必须先读、Case 状态、可调用工具、候选语言、输出与停止」，两者有「持久产品身份」时一并附上
  // 原因: 2026-10-07 校正口吻统一为知心姐姐、开场 / 只读轮只附用得上的 Skill 章节（BUG-1257）
  assert.deepEqual(RECTIFICATION_SKILL_SECTIONS_BY_ACTION.opening, ["必须先读与服务器权威", "OpeningPolicy", "ConversationFocus 与意图承接", "可调用工具与输入边界"]);
  assert.deepEqual(RECTIFICATION_SKILL_SECTIONS_BY_ACTION.read_only, ["必须先读与服务器权威", "Case 状态与只读边界", "可调用工具与输入边界", "candidate / accepted / confirmed 语言边界", "输出与停止条件"]);
  const body = activeSkillBody();
  const opening = sliceRectificationSkill(body, "opening");
  assert.ok(opening.text.length < body.length / 2, `${opening.text.length} of ${body.length}`);
  assert.match(opening.text, /本轮是开场轮/);
  // Dropped as sections (their names may still be mentioned inside kept text).
  for (const dropped of ["CaseConversationSummary 与长会话记忆", "上游同步边界", "触发条件与方法学归属"]) {
    const heading = new RegExp(`^## (?:\\d+\\. )?${dropped}`, "m");
    assert.doesNotMatch(opening.text, heading, dropped);
    assert.doesNotMatch(sliceRectificationSkill(body, "read_only").text, heading, dropped);
  }
  if (body.includes("持久产品身份与盘型交付")) {
    assert.match(opening.text, /持久产品身份与盘型交付/);
    assert.match(sliceRectificationSkill(body, "read_only").text, /持久产品身份与盘型交付/);
  }
  const evidence = sliceRectificationSkill(body, "evidence");
  assert.deepEqual(evidence.sections, ["## 5. ConversationFocus 与意图承接", "## 7. 批量证据与日期真实性"]);
  assert.ok(evidence.text.length < body.length / 2);
  const oldLayout = "# 旧版\n\n## 概述\n\n没有编号的章节。";
  assert.equal(sliceRectificationSkill(oldLayout, "evidence").text, oldLayout);
  // BUG-621: the oldest deprecated snapshot still resolves and still slices safely.
  const oldest = resolveExactSkillPackage(
    "jyotish-birth-time-rectification",
    "9.0.0",
    "5acb3103e80993ea611b93d2c1746e70b74aff8f8636bcffc80a837b954b470d",
  );
  const oldestBody = readFileSync(resolve(oldest.resolvedPath, "SKILL.md"), "utf8");
  // 9.0.0 numbers its sections differently (its §5 is the candidate language):
  // no title match, so the whole bound body goes out unchanged.
  assert.equal(sliceRectificationSkill(oldestBody, "evidence").text, oldestBody);
  assert.equal(sliceRectificationSkill(oldestBody, "opening").text, oldestBody);
  assert.equal(sliceRectificationSkill(oldestBody, "read_only").text, oldestBody);
});