portal/tools/observer-transcript-parser.mjs

#!/usr/bin/env node
/**
 * Transcript parser for the brain governance observer.
 * Deterministically extracts episode fields from a Claude Code session
 * transcript (JSONL). No LLM — pure parsing.
 *
 * Scope: the last turn (from the last real user prompt to end of file) —
 * one episode == one prompt→response cycle.
 *
 * Reasoning fields (triggers_matched / candidates_considered /
 * boundaries_applied) are NOT recoverable from a transcript and stay [];
 * their capture is a separate design question (ADR-011 follow-up).
 *
 * Security Guidance #40: pure parsing — no exec/execSync.
 * Per ADR-011 §6 + spec v1.1 §5.2.1.
 */

const SUPERPOWERS_PREFIX = 'superpowers:';

function parseLines(text) {
  const entries = [];
  let broken = 0;
  let total = 0;
  for (const line of String(text || '').split('\n')) {
    const trimmed = line.trim();
    if (!trimmed) continue;
    total += 1;
    try {
      entries.push(JSON.parse(trimmed));
    } catch {
      broken += 1; // broken line — counted for parse_gap, never thrown
    }
  }
  return { entries, broken, total };
}

// A genuine user prompt (turn boundary) — not a tool_result carrier message.
function isRealUserPrompt(entry) {
  const msg = entry && entry.message;
  if (!msg || msg.role !== 'user') return false;
  const c = msg.content;
  if (typeof c === 'string') return c.trim().length > 0;
  if (Array.isArray(c)) {
    const hasToolResult = c.some((b) => b && b.type === 'tool_result');
    const hasText = c.some((b) => b && b.type === 'text');
    return hasText && !hasToolResult;
  }
  return false;
}

function findTurnStart(entries) {
  for (let i = entries.length - 1; i >= 0; i--) {
    if (isRealUserPrompt(entries[i])) return i;
  }
  return 0;
}

function promptText(entry) {
  const c = entry && entry.message && entry.message.content;
  if (typeof c === 'string') return c;
  if (Array.isArray(c)) {
    return c
      .filter((b) => b && b.type === 'text')
      .map((b) => b.text || '')
      .join(' ');
  }
  return '';
}

export function classifyTask(text) {
  const t = String(text || '').toLowerCase();
  if (/рефактор|refactor/.test(t)) return 'refactor';
  if (/баг|bug|почини|исправ|fix\b|сломан|broken/.test(t)) return 'bugfix';
  if (/фич|feature|добав|implement|реализ|создай|create|новый|new /.test(t)) return 'feature';
  if (/докум|readme|\bdocs?\b/.test(t)) return 'docs';
  if (/\?|как |что |почему|зачем|why|how |what /.test(t)) return 'question';
  return 'other';
}

function collectToolUse(entries) {
  const skills = [];
  const counts = {};
  let errorCount = 0;
  for (const e of entries) {
    const content = e && e.message && Array.isArray(e.message.content) ? e.message.content : [];
    for (const block of content) {
      if (!block || typeof block !== 'object') continue;
      if (block.type === 'tool_use') {
        const name = block.name || 'unknown';
        counts[name] = (counts[name] || 0) + 1;
        if (name === 'Skill') {
          skills.push((block.input && block.input.skill) || 'unknown');
        }
      } else if (block.type === 'tool_result' && block.is_error === true) {
        errorCount += 1;
      }
    }
  }
  return { skills, counts, errorCount };
}

const FILE_TOOLS = new Set(['Read', 'Edit', 'Write', 'MultiEdit', 'NotebookEdit']);

/**
 * Deterministic environment factors for the turn that starts at turnStartIdx.
 * economy_level / parallel_session are scanned from the stringified turn;
 * model / post_compaction / session_turn from structural fields.
 */
export function extractEnvironment(allEntries, turnStartIdx) {
  const turn = allEntries.slice(turnStartIdx);
  const rawTurn = JSON.stringify(turn);

  const econ = rawTurn.match(/=== ECONOMY MODE:\s*(\d+)\s*%/);
  const economy_level = econ ? Number(econ[1]) : null;

  let model = null;
  for (const e of turn) {
    if (e && e.message && e.message.model) {
      model = e.message.model;
      break;
    }
  }

  let post_compaction = false;
  for (let i = 0; i < turnStartIdx && i < allEntries.length; i++) {
    if (allEntries[i] && allEntries[i].isCompactSummary === true) {
      post_compaction = true;
      break;
    }
  }

  let session_turn = 0;
  for (let i = 0; i <= turnStartIdx && i < allEntries.length; i++) {
    if (isRealUserPrompt(allEntries[i])) session_turn += 1;
  }

  const parallel_session = /параллельн|parallel session|чужой staged|foreign git index/i.test(rawTurn);

  return { economy_level, model, post_compaction, session_turn, parallel_session };
}

/** Task size: total tool calls + unique file paths touched (per spec §3, gap-resolution 2). */
export function extractTaskSize(turn) {
  let tool_calls = 0;
  const files = new Set();
  for (const e of turn) {
    const content = e && e.message && Array.isArray(e.message.content) ? e.message.content : [];
    for (const b of content) {
      if (b && b.type === 'tool_use') {
        tool_calls += 1;
        if (FILE_TOOLS.has(b.name) && b.input) {
          const p = b.input.file_path || b.input.notebook_path;
          if (p) files.add(String(p));
        }
      }
    }
  }
  return { tool_calls, files_touched: files.size, files: [...files] };
}

/** Classify the opening user-prompt sentiment (per spec §6 / gap-resolution 1). */
export function classifyPromptSignal(text) {
  const t = String(text || '').toLowerCase().trim();
  if (/не то\b|не так\b|переделай|отбой|\bстоп\b|почему ты|неверно|не верно|это не /.test(t)) {
    return 'correction';
  }
  if (/^(ок|окей|ok|спасибо|супер|отлично|готово|дальше|идеально)([,\s]|$)/.test(t)) {
    return 'approval';
  }
  if (classifyTask(t) !== 'other' && t.length > 15) return 'new_task';
  return 'neutral';
}

/**
 * Parse a transcript JSONL string into observer episode fields.
 * @param {string} transcriptText - Raw JSONL transcript contents.
 * @param {string|null} fallbackSessionId - Used when the transcript has no sessionId.
 * @returns {object} Episode with 5 mandatory fields + events.
 */
export function parseTranscript(transcriptText, fallbackSessionId = null) {
  const { entries } = parseLines(transcriptText);

  const withSession = entries.find((e) => e && e.sessionId);
  const sessionId =
    (withSession && withSession.sessionId) || fallbackSessionId || `unknown-${Date.now()}`;

  const start = findTurnStart(entries);
  const turn = entries.slice(start);

  const stamps = turn.map((e) => e && e.timestamp).filter(Boolean);
  const started_at = stamps[0] || new Date().toISOString();
  const ended_at = stamps[stamps.length - 1] || started_at;

  const { skills, counts, errorCount } = collectToolUse(turn);

  const events = [];
  for (const skill of skills) events.push({ kind: 'skill_invoked', skill });
  if (Object.keys(counts).length > 0) events.push({ kind: 'tool_summary', counts });
  for (let i = 0; i < errorCount; i++) {
    events.push({ kind: 'error', message: 'tool_result reported is_error' });
  }

  const usedSuperpowers = skills.some((s) => String(s).startsWith(SUPERPOWERS_PREFIX));

  return {
    task_id: sessionId,
    timestamps: { started_at, ended_at },
    path_type: usedSuperpowers ? 'regulated' : 'improvised',
    outcome: 'success',
    primary_rationale: {
      step: 1,
      node_chosen: skills.length > 0 ? skills[0] : 'direct',
      triggers_matched: [],
      candidates_considered: [],
      boundaries_applied: [],
      hard_floor: usedSuperpowers
        ? { invoked: true, rules: ['Pravila §12'] }
        : { invoked: false, rules: [] },
      task_classification: classifyTask(promptText(entries[start])),
    },
    events,
  };
}