agentboss 0.1.0 → 0.1.2

package/server/analysis/scoring-v2.js

@@ -1,72 +1,126 @@
-'use strict';
-
-const { classifySession } = require('./difficulty');
-const problemDef    = require('./dimensions/problem-definition');
-const judgement     = require('./dimensions/judgement');
-const aiTools       = require('./dimensions/ai-tools');
-const aiKnowledge   = require('./dimensions/ai-knowledge');
-const outputQuality = require('./dimensions/output-quality');
-const { analyzeSessionLLM } = require('../llm/session-analyzer');
-const { dimensionSource } = require('./dimensions/llm-merge');
-const { rollupDimension, scoreToLevel } = require('./thresholds-v2');
-
-/**
- * Score a single session through the v2.1 dimensions.
- * ONE consolidated LLM call per session returns BOTH the dimension scores
- * AND the collaboration advice.  All difficulties (including trivial) are
- * assessed; the prompt is difficulty-conditioned so short sessions are
- * graded on an easier curve.
- *
- * @param {object} db
- * @param {object} session
- * @param {object} [opts]   { force?: boolean }  bypass the analyzer cache
- */
-async function analyzeSessionV2(db, session, opts = {}) {
-  const difficulty = classifySession(session);
-
-  // E2 (tool coverage) is structural — always computed.
-  const e2 = safe(() => aiTools.analyze(db, session, difficulty.bucket));
-
-  // One consolidated LLM call → { scores, advice } (null if disabled/failed).
-  let llm = null;
-  try { llm = await analyzeSessionLLM(db, session, { force: opts.force === true }); }
-  catch (_) { llm = null; }
-  const cells = (llm && llm.scores) || {};
-
-  const h1 = safe(() => problemDef.analyze(db, session, difficulty.bucket, cells.H1));
-  const h2 = safe(() => judgement.analyze(db, session, difficulty.bucket, cells.H2));
-  const e1 = safe(() => aiKnowledge.analyze(db, session, difficulty.bucket, cells.E1));
-  const o1 = safe(() => outputQuality.analyze(db, session, difficulty.bucket, cells.O1));
-
-  const judgeSrc = dimensionSource([h1, h2, e1, o1].map((r) => r && r.judgeSource));
-  return assemble(difficulty, { h1, h2, e1, e2, o1 }, llm, judgeSrc);
-}
-
-/** Build the standard result object from the per-dimension scorer outputs. */
-function assemble(difficulty, dims, llm, judgeSource) {
-  const { h1, h2, e1, e2, o1 } = dims;
-  const pick = (r, f) => (r ? r[f] : null);
-
-  const subScores = { H1: pick(h1, 'subScores'), H2: pick(h2, 'subScores'), H3: null, E1: pick(e1, 'subScores'), E2: pick(e2, 'subScores'), O1: pick(o1, 'subScores') };
-  const subLevels = { H1: pick(h1, 'subLevels'), H2: pick(h2, 'subLevels'), H3: null, E1: pick(e1, 'subLevels'), E2: pick(e2, 'subLevels'), O1: pick(o1, 'subLevels') };
-  const subEvidence = { H1: pick(h1, 'subEvidence'), H2: pick(h2, 'subEvidence'), H3: null, E1: pick(e1, 'subEvidence'), E2: pick(e2, 'subEvidence'), O1: pick(o1, 'subEvidence') };
-
-  const scores = { H1: pick(h1, 'score'), H2: pick(h2, 'score'), H3: null, E1: pick(e1, 'score'), E2: pick(e2, 'score'), O1: pick(o1, 'score') };
-  const levels = { H1: scoreToLevel(scores.H1), H2: scoreToLevel(scores.H2), H3: null, E1: scoreToLevel(scores.E1), E2: scoreToLevel(scores.E2), O1: scoreToLevel(scores.O1) };
-
-  return {
-    difficulty, subScores, subLevels, subEvidence, scores, levels, judgeSource,
-    // Full analyzer payload (scores+advice) cached verbatim in llm_judge_v2.
-    llmJudge: llm || null,
-    // Raw advice + the meta the advice normaliser needs (job persists to llm_advice).
-    llmAdvice: llm && llm.advice ? llm.advice : null,
-    llmAdviceMeta: llm ? { msgCount: llm.msgCount, cli: llm.cli } : null,
-  };
-}
-
-function safe(fn) {
-  try { return fn(); }
-  catch (err) { console.warn('[scoring-v2] dimension error:', err.message); return null; }
-}
-
-module.exports = { analyzeSessionV2, rollupDimension };
+'use strict';
+
+const { classifySession } = require('./difficulty');
+const { analyzeSessionLLM } = require('../llm/session-analyzer');
+const { dimensionSource, mergeIndicator } = require('./dimensions/llm-merge');
+const { rollupDimension, scoreToLevel } = require('./thresholds-v2');
+
+/**
+ * Score a single session through the v2.1 dimensions.
+ * ONE consolidated LLM call per session returns BOTH the dimension scores
+ * AND the collaboration advice.  All difficulties (including trivial) are
+ * assessed; the prompt is difficulty-conditioned so short sessions are
+ * graded on an easier curve.
+ *
+ * @param {object} db
+ * @param {object} session
+ * @param {object} [opts]   { force?: boolean }  bypass the analyzer cache
+ */
+async function analyzeSessionV2(db, session, opts = {}) {
+  const difficulty = classifySession(session);
+
+  // One consolidated LLM call → { scores, advice } (null if disabled/failed).
+  let llm = null;
+  try { llm = await analyzeSessionLLM(db, session, { force: opts.force === true }); }
+  catch (_) { llm = null; }
+  const cells = (llm && llm.scores) || {};
+
+  // All five dimensions are fully LLM-judged — no rule fallback.  A
+  // sub-indicator the LLM can't score stays null ("未评估") rather than
+  // degrading to a keyword heuristic.
+  const h1  = safe(() => buildLlmDimension(cells.H1, H1_LABELS, 'H1'));
+  const h2  = safe(() => buildLlmDimension(cells.H2, H2_LABELS, 'H2'));
+  const h3  = safe(() => buildLlmDimension(normaliseH3(cells.H3), H3_LABELS, 'H3'));
+  // ENV — merged AI-capability environment (knowledge / tooling / currency).
+  // Persisted into the legacy E1 slot (E2 retired).
+  const env = safe(() => buildLlmDimension(cells.ENV, ENV_LABELS, 'ENV'));
+  const o1  = safe(() => buildLlmDimension(cells.O1, O1_LABELS, 'O1'));
+
+  const judgeSrc = dimensionSource([h1, h2, h3, env, o1].map((r) => r && r.judgeSource));
+  return assemble(difficulty, { h1, h2, h3, e1: env, e2: null, o1 }, llm, judgeSrc);
+}
+
+const H1_LABELS  = { clarity: '初始指令清晰度', converge: '任务收敛轮次', drift: '方向变更次数' };
+const H2_LABELS  = { challenge: '合理质疑率', override: '推翻率', accept_rate: '顺从/采纳判断' };
+const O1_LABELS  = { first_take: '一次采纳率', code_style: '代码规范性', completeness: '方案完备性' };
+const H3_LABELS  = { abstraction: '抽象层级', reuse: '复用意识', standard: '规范约束' };
+const ENV_LABELS = { knowledge: '知识覆盖', tooling: '工具运用', currency: '时效性' };
+
+/** Tolerate the older single-cell H3 shape ({system:{…}}). */
+function normaliseH3(group) {
+  const g = group || {};
+  if (g.system && !g.abstraction) g.abstraction = g.system;
+  return g;
+}
+
+/**
+ * Build a fully LLM-judged dimension from a group of sub-indicator cells
+ * (no rule fallback — these dimensions have no single-session structural
+ * proxy).  Returns null when the LLM scored none of them.
+ *
+ * @param {object|null} group   { <subKey>: {score,confidence,evidence}, … }
+ * @param {object} labels       { <subKey>: 中文label }
+ * @param {string} dimKey       rollup weights key (H3 / ENV)
+ */
+function buildLlmDimension(group, labels, dimKey) {
+  const g = group || {};
+  const subScores = {};
+  const subLevels = {};
+  const subEvidence = {};
+  const sources = [];
+  let any = false;
+
+  for (const k of Object.keys(labels)) {
+    const m = mergeIndicator(g[k], null, null);
+    subScores[k] = m.score;
+    subLevels[k] = m.level;
+    subEvidence[k] = {
+      key: k, label: labels[k],
+      what: m.evidence || '未发现相关证据',
+      score: m.score, level: m.level,
+      source: m.source,
+    };
+    sources.push(m.source);
+    if (m.score != null) any = true;
+  }
+  if (!any) return null;
+
+  const score = rollupDimension(dimKey, subScores);
+  return {
+    subScores,
+    subLevels,
+    subEvidence,
+    score,
+    level: scoreToLevel(score),
+    judgeSource: dimensionSource(sources),
+  };
+}
+
+/** Build the standard result object from the per-dimension scorer outputs. */
+function assemble(difficulty, dims, llm, judgeSource) {
+  const { h1, h2, h3, e1, e2, o1 } = dims;
+  const pick = (r, f) => (r ? r[f] : null);
+
+  const subScores = { H1: pick(h1, 'subScores'), H2: pick(h2, 'subScores'), H3: pick(h3, 'subScores'), E1: pick(e1, 'subScores'), E2: pick(e2, 'subScores'), O1: pick(o1, 'subScores') };
+  const subLevels = { H1: pick(h1, 'subLevels'), H2: pick(h2, 'subLevels'), H3: pick(h3, 'subLevels'), E1: pick(e1, 'subLevels'), E2: pick(e2, 'subLevels'), O1: pick(o1, 'subLevels') };
+  const subEvidence = { H1: pick(h1, 'subEvidence'), H2: pick(h2, 'subEvidence'), H3: pick(h3, 'subEvidence'), E1: pick(e1, 'subEvidence'), E2: pick(e2, 'subEvidence'), O1: pick(o1, 'subEvidence') };
+
+  const scores = { H1: pick(h1, 'score'), H2: pick(h2, 'score'), H3: pick(h3, 'score'), E1: pick(e1, 'score'), E2: pick(e2, 'score'), O1: pick(o1, 'score') };
+  const levels = { H1: scoreToLevel(scores.H1), H2: scoreToLevel(scores.H2), H3: scoreToLevel(scores.H3), E1: scoreToLevel(scores.E1), E2: scoreToLevel(scores.E2), O1: scoreToLevel(scores.O1) };
+
+  return {
+    difficulty, subScores, subLevels, subEvidence, scores, levels, judgeSource,
+    // Full analyzer payload (scores+advice) cached verbatim in llm_judge_v2.
+    llmJudge: llm || null,
+    // Raw advice + the meta the advice normaliser needs (job persists to llm_advice).
+    llmAdvice: llm && llm.advice ? llm.advice : null,
+    llmAdviceMeta: llm ? { msgCount: llm.msgCount, cli: llm.cli } : null,
+  };
+}
+
+function safe(fn) {
+  try { return fn(); }
+  catch (err) { console.warn('[scoring-v2] dimension error:', err.message); return null; }
+}
+
+module.exports = { analyzeSessionV2, rollupDimension };