buildanything 2.0.0 → 2.1.2

package/src/graph/parser/product-spec.ts

@@ -0,0 +1,930 @@
+// Deterministic, schema-aware extractor for product-spec.md.
+// Source of truth: docs/graph/04-slice1-schema.md + protocols/product-spec-schema.md.
+// No LLM, no I/O. Caller passes the markdown content and path.
+
+import { ids, kebab, sha256Hex } from "../ids.js";
+import type {
+  AcceptanceCriterionNode,
+  BusinessRuleNode,
+  ExtractError,
+  ExtractResult,
+  FailureModeNode,
+  FeatureNode,
+  GraphEdge,
+  GraphFragment,
+  GraphNode,
+  PersonaConstraintNode,
+  PersonaNode,
+  Relation,
+  ScreenNode,
+  StateNode,
+  TransitionNode,
+} from "../types.js";
+
+const PRODUCED_BY = "product-spec-writer";
+const PRODUCED_AT_STEP = "1.6";
+const META_STATE_NAMES = new Set([
+  "loading",
+  "empty",
+  "error",
+  "stale",
+  "offline",
+  "disabled",
+  "permission-denied",
+]);
+
+interface Line {
+  n: number; // 1-based line number
+  text: string; // raw text, no trailing newline
+}
+
+interface Section {
+  heading: string; // e.g. "App Overview", "Feature: Checkout", "States"
+  level: number; // 2 = ##, 3 = ###
+  startLine: number; // line number of the heading itself
+  bodyLines: Line[]; // all lines until the next heading at <= same level
+}
+
+interface Ctx {
+  mdPath: string;
+  errors: ExtractError[];
+  nodes: GraphNode[];
+  edges: GraphEdge[];
+  personasByKey: Map<string, PersonaNode>; // canonical persona label → node
+}
+
+function loc(line: number): string {
+  return `L${line}`;
+}
+
+function pushError(ctx: Ctx, line: number, message: string): void {
+  ctx.errors.push({ line, message });
+}
+
+function splitLines(content: string): Line[] {
+  const raw = content.split(/\r?\n/);
+  return raw.map((text, i) => ({ n: i + 1, text }));
+}
+
+// Walk lines and produce sections at the requested heading levels (2 or 3).
+// A section ends at the next heading of equal or shallower level.
+function partitionSections(lines: Line[], level: number, start: number, end: number): Section[] {
+  const sections: Section[] = [];
+  const headingPrefix = "#".repeat(level) + " ";
+  let i = start;
+  while (i < end) {
+    const line = lines[i];
+    if (isHeadingAtLevel(line.text, level)) {
+      const heading = line.text.slice(headingPrefix.length).trim();
+      const bodyStart = i + 1;
+      let j = bodyStart;
+      while (j < end && !isHeadingAtOrAbove(lines[j].text, level)) j++;
+      sections.push({
+        heading,
+        level,
+        startLine: line.n,
+        bodyLines: lines.slice(bodyStart, j),
+      });
+      i = j;
+    } else {
+      i++;
+    }
+  }
+  return sections;
+}
+
+function isHeadingAtLevel(text: string, level: number): boolean {
+  const prefix = "#".repeat(level) + " ";
+  if (!text.startsWith(prefix)) return false;
+  // make sure it's exactly this level, not deeper
+  return text[level] !== "#";
+}
+
+function isHeadingAtOrAbove(text: string, level: number): boolean {
+  for (let l = 1; l <= level; l++) {
+    if (isHeadingAtLevel(text, l)) return true;
+  }
+  return false;
+}
+
+// Parse a markdown pipe table from a slice of lines starting at the first
+// non-blank, non-comment line. Returns headers (lowercased, trimmed) +
+// rows (each row keyed by header). Each cell tracks its source line number
+// (the row line, not per-cell — multi-line cells are not supported here).
+interface TableRow {
+  cells: Record<string, string>;
+  line: number;
+}
+
+function parseTable(body: Line[]): { headers: string[]; rows: TableRow[] } | null {
+  // Find a header+separator pair anywhere in the body. A section may have a
+  // grounding paragraph before the table.
+  const sepRe = /^\s*\|?\s*[-:| ]+\s*\|?\s*$/;
+  const significant = body.filter((l) => l.text.trim().length > 0);
+  let headerIdx = -1;
+  for (let i = 0; i < significant.length - 1; i++) {
+    if (
+      significant[i].text.includes("|") &&
+      significant[i + 1].text.includes("|") &&
+      sepRe.test(significant[i + 1].text)
+    ) {
+      headerIdx = i;
+      break;
+    }
+  }
+  if (headerIdx < 0) return null;
+
+  const headerLine = significant[headerIdx];
+  const headers = splitRow(headerLine.text).map((h) => h.toLowerCase());
+  const rows: TableRow[] = [];
+  for (let i = headerIdx + 2; i < significant.length; i++) {
+    const ln = significant[i];
+    if (!ln.text.includes("|")) break;
+    const cells = splitRow(ln.text);
+    if (cells.length === 0) continue;
+    const row: Record<string, string> = {};
+    for (let c = 0; c < headers.length; c++) {
+      row[headers[c]] = (cells[c] ?? "").trim();
+    }
+    rows.push({ cells: row, line: ln.n });
+  }
+  return { headers, rows };
+}
+
+function splitRow(text: string): string[] {
+  // Strip leading/trailing pipes, then split.
+  let s = text.trim();
+  if (s.startsWith("|")) s = s.slice(1);
+  if (s.endsWith("|")) s = s.slice(0, -1);
+  return s.split("|").map((c) => c.trim());
+}
+
+// =============================================================================
+// Top-level: App Overview persona table
+// =============================================================================
+
+function parsePersonas(ctx: Ctx, section: Section): void {
+  const table = parseTable(section.bodyLines);
+  if (!table) {
+    pushError(ctx, section.startLine, "App Overview: persona table not found or malformed");
+    return;
+  }
+  const required = ["persona", "role", "primary jtbd", "relationship to other personas"];
+  for (const r of required) {
+    if (!table.headers.includes(r)) {
+      pushError(
+        ctx,
+        section.startLine,
+        `App Overview: persona table missing required column "${r}"`,
+      );
+      return;
+    }
+  }
+  if (table.rows.length < 1) {
+    pushError(ctx, section.startLine, "App Overview: persona table has zero rows");
+    return;
+  }
+
+  let primaryCount = 0;
+  for (const row of table.rows) {
+    const rawName = row.cells["persona"] ?? "";
+    if (!rawName) {
+      pushError(ctx, row.line, "Persona row missing name");
+      continue;
+    }
+    const isPrimary = /\(primary\)/i.test(rawName);
+    if (isPrimary) primaryCount++;
+    const cleanLabel = rawName.replace(/\(primary\)/gi, "").trim();
+    const role = row.cells["role"] ?? "";
+    const jtbd = row.cells["primary jtbd"] ?? "";
+    const relationship = row.cells["relationship to other personas"] ?? "";
+    if (!role || !jtbd) {
+      pushError(ctx, row.line, `Persona "${cleanLabel}" missing role or JTBD`);
+      continue;
+    }
+    const node: PersonaNode = {
+      id: ids.persona(cleanLabel),
+      label: cleanLabel,
+      entity_type: "persona",
+      source_file: ctx.mdPath,
+      source_location: loc(row.line),
+      confidence: "EXTRACTED",
+      description: relationship,
+      role,
+      is_primary: isPrimary,
+      primary_jtbd: jtbd,
+    };
+    ctx.nodes.push(node);
+    ctx.personasByKey.set(cleanLabel.toLowerCase(), node);
+  }
+
+  if (primaryCount !== 1) {
+    pushError(
+      ctx,
+      section.startLine,
+      `App Overview: expected exactly one (primary) persona, found ${primaryCount}`,
+    );
+  }
+}
+
+// =============================================================================
+// Top-level: Screen Inventory
+// =============================================================================
+
+interface ScreenInfo {
+  rawName: string;
+  description: string;
+  featureNames: string[];
+}
+
+function parseScreenInventory(ctx: Ctx, section: Section): ScreenInfo[] {
+  const table = parseTable(section.bodyLines);
+  if (!table) {
+    pushError(ctx, section.startLine, "Screen Inventory: table not found or malformed");
+    return [];
+  }
+  const required = ["screen", "description", "features"];
+  for (const r of required) {
+    if (!table.headers.includes(r)) {
+      pushError(
+        ctx,
+        section.startLine,
+        `Screen Inventory: missing required column "${r}"`,
+      );
+      return [];
+    }
+  }
+
+  const out: ScreenInfo[] = [];
+  for (const row of table.rows) {
+    const rawName = row.cells["screen"] ?? "";
+    const description = row.cells["description"] ?? "";
+    const features = row.cells["features"] ?? "";
+    if (!rawName) {
+      pushError(ctx, row.line, "Screen Inventory: row missing screen name");
+      continue;
+    }
+    const featureNames = features
+      .split(",")
+      .map((s) => s.trim())
+      .filter((s) => s.length > 0);
+
+    const countMatch = rawName.match(/\((\d+)\s+screens?\)/i);
+    const cleanName = rawName.replace(/\(\d+\s+screens?\)/i, "").trim();
+    const count = countMatch ? Number.parseInt(countMatch[1], 10) : undefined;
+
+    const node: ScreenNode = {
+      id: ids.screen(cleanName),
+      label: cleanName,
+      entity_type: "screen",
+      source_file: ctx.mdPath,
+      source_location: loc(row.line),
+      confidence: "EXTRACTED",
+      description,
+      feature_ids: featureNames.map((f) => ids.feature(f)),
+      ...(count ? { count } : {}),
+    };
+    ctx.nodes.push(node);
+    out.push({ rawName: cleanName, description, featureNames });
+
+    // has_screen edges per feature attribution
+    for (const fname of featureNames) {
+      ctx.edges.push({
+        source: ids.feature(fname),
+        target: node.id,
+        relation: "has_screen",
+        confidence: "EXTRACTED",
+        source_file: ctx.mdPath,
+        source_location: loc(row.line),
+        produced_by_agent: PRODUCED_BY,
+        produced_at_step: PRODUCED_AT_STEP,
+      });
+    }
+  }
+  return out;
+}
+
+// =============================================================================
+// Top-level: Cross-Feature Interactions
+// =============================================================================
+
+function parseCrossFeature(ctx: Ctx, section: Section): void {
+  // Bullets like "- Auth → Checkout: user must be authenticated"
+  // Persona-crossing variants: "- Order Placement (Buyer) → Order Notification (Seller): ..."
+  const arrow = /^\s*-\s+(.+?)\s*(?:→|->)\s*(.+?)\s*:\s*(.*)$/u;
+  for (const line of section.bodyLines) {
+    const m = line.text.match(arrow);
+    if (!m) continue;
+    const lhs = m[1].trim();
+    const rhs = m[2].trim();
+    const ruleText = m[3].trim() || undefined;
+    // Strip trailing "(Persona)" annotations from each side; the feature
+    // name is what precedes any parenthetical.
+    const lhsFeature = lhs.replace(/\s*\([^)]*\)\s*$/u, "").trim();
+    const rhsFeatureRaw = rhs.replace(/\s*\([^)]*\)\s*$/u, "").trim();
+    // RHS may also list multiple comma-separated features (e.g.
+    // "Auth, Checkout, Dashboard"). Fan out one edge per target.
+    const rhsFeatures = rhsFeatureRaw
+      .split(",")
+      .map((s) => s.trim())
+      .filter((s) => s.length > 0);
+    for (const target of rhsFeatures) {
+      ctx.edges.push({
+        source: ids.feature(lhsFeature),
+        target: ids.feature(target),
+        relation: "depends_on",
+        confidence: "EXTRACTED",
+        source_file: ctx.mdPath,
+        source_location: loc(line.n),
+        produced_by_agent: PRODUCED_BY,
+        produced_at_step: PRODUCED_AT_STEP,
+        ...(ruleText ? { label: ruleText } : {}),
+      });
+    }
+  }
+}
+
+// =============================================================================
+// Per-feature
+// =============================================================================
+
+function parseFeature(ctx: Ctx, section: Section): void {
+  const headingLabel = section.heading.replace(/^Feature:\s*/i, "").trim();
+  if (!headingLabel) {
+    pushError(ctx, section.startLine, "Feature heading missing name");
+    return;
+  }
+  const featureNode: FeatureNode = {
+    id: ids.feature(headingLabel),
+    label: headingLabel,
+    entity_type: "feature",
+    source_file: ctx.mdPath,
+    source_location: loc(section.startLine),
+    confidence: "EXTRACTED",
+    name: headingLabel,
+    kebab_anchor: kebab(headingLabel),
+  };
+  ctx.nodes.push(featureNode);
+
+  // Re-tokenize this feature's body to extract ### subsections.
+  const subSections = partitionSubsections(section.bodyLines);
+  const stateNamesInOrder: { name: string; line: number; isInitial: boolean }[] = [];
+
+  for (const sub of subSections) {
+    const lower = sub.heading.toLowerCase();
+    if (lower === "states") {
+      parseStates(ctx, featureNode, sub, stateNamesInOrder);
+    } else if (lower === "transitions") {
+      parseTransitions(ctx, featureNode, sub);
+    } else if (lower === "business rules") {
+      parseBusinessRules(ctx, featureNode, sub);
+    } else if (lower === "failure modes") {
+      parseFailureModes(ctx, featureNode, sub);
+    } else if (lower === "acceptance criteria") {
+      parseAcceptanceCriteria(ctx, featureNode, sub);
+    } else if (lower === "persona constraints") {
+      parsePersonaConstraints(ctx, featureNode, sub);
+    }
+    // Other subsections (Data Requirements, Happy Path, Empty States,
+    // Notification Triggers, etc.) are intentionally ignored in Slice 1.
+  }
+}
+
+function partitionSubsections(body: Line[]): Section[] {
+  // Treat the body as a virtual document; reuse partitionSections at level 3.
+  // Need to feed line objects with their absolute line numbers.
+  const sections: Section[] = [];
+  let i = 0;
+  while (i < body.length) {
+    const line = body[i];
+    if (isHeadingAtLevel(line.text, 3)) {
+      const heading = line.text.slice(4).trim();
+      const start = i + 1;
+      let j = start;
+      while (j < body.length && !isHeadingAtOrAbove(body[j].text, 3)) j++;
+      sections.push({
+        heading,
+        level: 3,
+        startLine: line.n,
+        bodyLines: body.slice(start, j),
+      });
+      i = j;
+    } else {
+      i++;
+    }
+  }
+  return sections;
+}
+
+// ----- States -----
+
+function parseStates(
+  ctx: Ctx,
+  feature: FeatureNode,
+  section: Section,
+  collect: { name: string; line: number; isInitial: boolean }[],
+): void {
+  // Two accepted forms:
+  //   "States: idle (initial), loading, loaded, empty, error"
+  //   bullets:  "- idle (initial)"   "- loading"
+  const inline = section.bodyLines.find((l) => /^\s*states\s*:/i.test(l.text));
+  let entries: { name: string; isInitial: boolean; line: number }[] = [];
+
+  if (inline) {
+    const after = inline.text.replace(/^\s*states\s*:\s*/i, "");
+    entries = after
+      .split(",")
+      .map((s) => s.trim())
+      .filter((s) => s.length > 0)
+      .map((s) => ({
+        name: s.replace(/\(initial\)/i, "").trim(),
+        isInitial: /\(initial\)/i.test(s),
+        line: inline.n,
+      }));
+  } else {
+    for (const line of section.bodyLines) {
+      const m = line.text.match(/^\s*-\s+(.+?)\s*$/);
+      if (!m) continue;
+      const raw = m[1];
+      entries.push({
+        name: raw.replace(/\(initial\)/i, "").trim(),
+        isInitial: /\(initial\)/i.test(raw),
+        line: line.n,
+      });
+    }
+  }
+
+  if (entries.length === 0) {
+    pushError(ctx, section.startLine, `Feature "${feature.label}": States section is empty`);
+    return;
+  }
+
+  // If nothing is explicitly marked (initial), the first entry is initial.
+  if (!entries.some((e) => e.isInitial)) {
+    entries[0].isInitial = true;
+  }
+
+  for (const e of entries) {
+    const stateNode: StateNode = {
+      id: ids.state(feature.name, e.name),
+      label: e.name,
+      entity_type: "state",
+      source_file: ctx.mdPath,
+      source_location: loc(e.line),
+      confidence: "EXTRACTED",
+      feature_id: feature.id,
+      is_initial: e.isInitial,
+      meta_state: META_STATE_NAMES.has(kebab(e.name)),
+    };
+    ctx.nodes.push(stateNode);
+    ctx.edges.push({
+      source: feature.id,
+      target: stateNode.id,
+      relation: "has_state",
+      confidence: "EXTRACTED",
+      source_file: ctx.mdPath,
+      source_location: loc(e.line),
+      produced_by_agent: PRODUCED_BY,
+      produced_at_step: PRODUCED_AT_STEP,
+    });
+    if (e.isInitial) {
+      ctx.edges.push({
+        source: feature.id,
+        target: stateNode.id,
+        relation: "has_initial_state",
+        confidence: "EXTRACTED",
+        source_file: ctx.mdPath,
+        source_location: loc(e.line),
+        produced_by_agent: PRODUCED_BY,
+        produced_at_step: PRODUCED_AT_STEP,
+      });
+    }
+    collect.push({ name: e.name, line: e.line, isInitial: e.isInitial });
+  }
+}
+
+// ----- Transitions -----
+
+function parseTransitions(ctx: Ctx, feature: FeatureNode, section: Section): void {
+  const table = parseTable(section.bodyLines);
+  if (!table) {
+    pushError(
+      ctx,
+      section.startLine,
+      `Feature "${feature.label}": Transitions table not found or malformed`,
+    );
+    return;
+  }
+  // header may be "from → to" with arrow, or "from->to". Find the from→to col.
+  const fromToKey = table.headers.find((h) => /from\s*(?:→|->)\s*to/u.test(h));
+  if (!fromToKey) {
+    pushError(
+      ctx,
+      section.startLine,
+      `Feature "${feature.label}": Transitions table missing "From → To" column`,
+    );
+    return;
+  }
+  for (const row of table.rows) {
+    const fromTo = row.cells[fromToKey] ?? "";
+    const m = fromTo.match(/^(.+?)\s*(?:→|->)\s*(.+)$/u);
+    if (!m) {
+      pushError(ctx, row.line, `Transition row malformed: "${fromTo}"`);
+      continue;
+    }
+    const from = m[1].trim();
+    const to = m[2].trim();
+    const trigger = row.cells["trigger"] ?? "";
+    const preconditions = row.cells["preconditions"] ?? "";
+    const sideEffects = row.cells["side effects"] ?? "";
+
+    const fromId = ids.state(feature.name, from);
+    const toId = ids.state(feature.name, to);
+    const transitionNode: TransitionNode = {
+      id: ids.transition(feature.name, from, to),
+      label: `${from} → ${to}`,
+      entity_type: "transition",
+      source_file: ctx.mdPath,
+      source_location: loc(row.line),
+      confidence: "EXTRACTED",
+      from_state_id: fromId,
+      to_state_id: toId,
+      trigger,
+      preconditions,
+      side_effects: sideEffects,
+    };
+    ctx.nodes.push(transitionNode);
+    ctx.edges.push({
+      source: fromId,
+      target: toId,
+      relation: "transitions_to",
+      confidence: "EXTRACTED",
+      source_file: ctx.mdPath,
+      source_location: loc(row.line),
+      produced_by_agent: PRODUCED_BY,
+      produced_at_step: PRODUCED_AT_STEP,
+    });
+    ctx.edges.push({
+      source: fromId,
+      target: transitionNode.id,
+      relation: "triggered_by_transition",
+      confidence: "EXTRACTED",
+      source_file: ctx.mdPath,
+      source_location: loc(row.line),
+      produced_by_agent: PRODUCED_BY,
+      produced_at_step: PRODUCED_AT_STEP,
+    });
+  }
+}
+
+// ----- Business Rules -----
+
+function parseBusinessRules(ctx: Ctx, feature: FeatureNode, section: Section): void {
+  const bullets = collectBullets(section.bodyLines);
+  for (const b of bullets) {
+    const text = b.text.trim();
+    if (!text) continue;
+    const decisionNeeded = /\[DECISION NEEDED/i.test(text);
+    const value = extractRuleValue(text);
+    const node: BusinessRuleNode = {
+      id: ids.businessRule(feature.name, text),
+      label: text.length > 80 ? text.slice(0, 77) + "..." : text,
+      entity_type: "business_rule",
+      source_file: ctx.mdPath,
+      source_location: loc(b.line),
+      confidence: "EXTRACTED",
+      feature_id: feature.id,
+      text,
+      value,
+      decision_needed: decisionNeeded,
+    };
+    ctx.nodes.push(node);
+    ctx.edges.push({
+      source: feature.id,
+      target: node.id,
+      relation: "has_rule",
+      confidence: "EXTRACTED",
+      source_file: ctx.mdPath,
+      source_location: loc(b.line),
+      produced_by_agent: PRODUCED_BY,
+      produced_at_step: PRODUCED_AT_STEP,
+    });
+  }
+}
+
+function extractRuleValue(text: string): string | null {
+  // Prefer "= <value>" explicit assignment, else first numeric+unit run.
+  const eq = text.match(/=\s*([^.\[]+?)(?:\s*\[|$)/);
+  if (eq) return eq[1].trim();
+  const num = text.match(
+    /(\d+(?:\.\d+)?)\s*(seconds?|minutes?|hours?|days?|items?|orders?|requests?|%|percent|MB|GB|KB|ms)/i,
+  );
+  if (num) return `${num[1]} ${num[2]}`;
+  return null;
+}
+
+// ----- Failure Modes -----
+
+function parseFailureModes(ctx: Ctx, feature: FeatureNode, section: Section): void {
+  // Multi-line bullet structure:
+  //   - <trigger> →
+  //     User sees: "..."
+  //     User can: ...
+  //     System: ...
+  // Bullets are separated by either a blank line or the start of a new "- " bullet.
+  const blocks = splitFailureBlocks(section.bodyLines);
+  for (const block of blocks) {
+    if (block.length === 0) continue;
+    const headLine = block[0];
+    const headText = headLine.text.replace(/^\s*-\s+/, "").trim();
+    // Trigger is the part before a trailing arrow if present.
+    const trigger = headText.replace(/\s*(?:→|->)\s*$/u, "").trim();
+    if (!trigger) continue;
+
+    const userSees = readLabel(block, /^\s*user\s+sees\s*:/i);
+    const userCan = readLabel(block, /^\s*user\s+can\s*:/i);
+    const systemDoes = readLabel(block, /^\s*system\s*:/i);
+
+    const node: FailureModeNode = {
+      id: ids.failureMode(feature.name, trigger),
+      label: trigger,
+      entity_type: "failure_mode",
+      source_file: ctx.mdPath,
+      source_location: loc(headLine.n),
+      confidence: "EXTRACTED",
+      feature_id: feature.id,
+      trigger,
+      user_sees: userSees,
+      user_can: userCan,
+      system_does: systemDoes,
+    };
+    ctx.nodes.push(node);
+    ctx.edges.push({
+      source: feature.id,
+      target: node.id,
+      relation: "has_failure_mode",
+      confidence: "EXTRACTED",
+      source_file: ctx.mdPath,
+      source_location: loc(headLine.n),
+      produced_by_agent: PRODUCED_BY,
+      produced_at_step: PRODUCED_AT_STEP,
+    });
+  }
+}
+
+function splitFailureBlocks(body: Line[]): Line[][] {
+  const blocks: Line[][] = [];
+  let current: Line[] = [];
+  for (const line of body) {
+    const isBulletStart = /^\s*-\s+/.test(line.text);
+    if (isBulletStart) {
+      if (current.length > 0) blocks.push(current);
+      current = [line];
+    } else if (line.text.trim() === "") {
+      if (current.length > 0) {
+        blocks.push(current);
+        current = [];
+      }
+    } else if (current.length > 0) {
+      current.push(line);
+    }
+  }
+  if (current.length > 0) blocks.push(current);
+  return blocks;
+}
+
+function readLabel(block: Line[], pattern: RegExp): string {
+  const idx = block.findIndex((l) => pattern.test(l.text));
+  if (idx < 0) return "";
+  // Take the text after the label, plus any non-labelled continuation lines
+  // until we hit another label or end of block.
+  const first = block[idx].text.replace(pattern, "").trim();
+  const labelStartRe = /^\s*(user\s+sees|user\s+can|system)\s*:/i;
+  const parts: string[] = [];
+  if (first) parts.push(first.replace(/^"+|"+$/g, ""));
+  for (let i = idx + 1; i < block.length; i++) {
+    const t = block[i].text.trim();
+    if (!t) break;
+    if (labelStartRe.test(t)) break;
+    parts.push(t.replace(/^"+|"+$/g, ""));
+  }
+  return parts.join(" ").trim();
+}
+
+// ----- Acceptance Criteria -----
+
+function parseAcceptanceCriteria(ctx: Ctx, feature: FeatureNode, section: Section): void {
+  const re = /^\s*-\s+\[\s\]\s+(.+)$/;
+  for (const line of section.bodyLines) {
+    const m = line.text.match(re);
+    if (!m) continue;
+    const text = m[1].trim();
+    if (!text) continue;
+    const node: AcceptanceCriterionNode = {
+      id: ids.acceptanceCriterion(feature.name, text),
+      label: text.length > 80 ? text.slice(0, 77) + "..." : text,
+      entity_type: "acceptance_criterion",
+      source_file: ctx.mdPath,
+      source_location: loc(line.n),
+      confidence: "EXTRACTED",
+      feature_id: feature.id,
+      text,
+      verified: false,
+    };
+    ctx.nodes.push(node);
+    ctx.edges.push({
+      source: feature.id,
+      target: node.id,
+      relation: "has_acceptance",
+      confidence: "EXTRACTED",
+      source_file: ctx.mdPath,
+      source_location: loc(line.n),
+      produced_by_agent: PRODUCED_BY,
+      produced_at_step: PRODUCED_AT_STEP,
+    });
+  }
+}
+
+// ----- Persona Constraints -----
+
+function parsePersonaConstraints(ctx: Ctx, feature: FeatureNode, section: Section): void {
+  // Block structure (per protocols/product-spec-schema.md):
+  //   - Persona: <name> [(primary)] — <one-liner> [src]
+  //     Constraint: <text> [src]
+  //     Constraint: <text> [src]
+  //   - Persona: <other> ...
+  let currentPersona: PersonaNode | null = null;
+
+  for (const line of section.bodyLines) {
+    const text = line.text;
+    const personaMatch = text.match(/^\s*-\s+Persona\s*:\s*(.+?)\s+(?:—|--|-)\s+(.*)$/u);
+    if (personaMatch) {
+      const rawName = personaMatch[1].trim();
+      const cleanName = rawName.replace(/\(primary\)/gi, "").trim();
+      const found = ctx.personasByKey.get(cleanName.toLowerCase());
+      if (!found) {
+        pushError(
+          ctx,
+          line.n,
+          `Feature "${feature.label}": persona "${cleanName}" not found in App Overview persona table`,
+        );
+        currentPersona = null;
+      } else {
+        currentPersona = found;
+      }
+      continue;
+    }
+    const constraintMatch = text.match(/^\s*Constraint\s*:\s*(.+)$/i);
+    if (constraintMatch && currentPersona) {
+      const full = constraintMatch[1].trim();
+      // Last [src] suffix is the citation; everything before is the constraint.
+      const citeMatch = full.match(/\[([^\]]+)\]\s*$/);
+      const constraintText = citeMatch ? full.slice(0, citeMatch.index).trim() : full;
+      const citedSource = citeMatch ? citeMatch[1].trim() : "";
+      const node: PersonaConstraintNode = {
+        id: ids.personaConstraint(feature.name, constraintText),
+        label:
+          constraintText.length > 80 ? constraintText.slice(0, 77) + "..." : constraintText,
+        entity_type: "persona_constraint",
+        source_file: ctx.mdPath,
+        source_location: loc(line.n),
+        confidence: "EXTRACTED",
+        feature_id: feature.id,
+        persona_id: currentPersona.id,
+        constraint_text: constraintText,
+        cited_source: citedSource,
+      };
+      ctx.nodes.push(node);
+      ctx.edges.push({
+        source: node.id,
+        target: feature.id,
+        relation: "constrains",
+        confidence: "EXTRACTED",
+        source_file: ctx.mdPath,
+        source_location: loc(line.n),
+        produced_by_agent: PRODUCED_BY,
+        produced_at_step: PRODUCED_AT_STEP,
+      });
+      ctx.edges.push({
+        source: node.id,
+        target: currentPersona.id,
+        relation: "applies_to_persona",
+        confidence: "EXTRACTED",
+        source_file: ctx.mdPath,
+        source_location: loc(line.n),
+        produced_by_agent: PRODUCED_BY,
+        produced_at_step: PRODUCED_AT_STEP,
+      });
+    }
+  }
+}
+
+// =============================================================================
+// Bullet collector (treats indented continuations as part of the same bullet)
+// =============================================================================
+
+function collectBullets(body: Line[]): { text: string; line: number }[] {
+  const out: { text: string; line: number }[] = [];
+  let current: { text: string; line: number } | null = null;
+  for (const line of body) {
+    const m = line.text.match(/^(\s*)-\s+(.+)$/);
+    if (m) {
+      if (current) out.push(current);
+      current = { text: m[2].trim(), line: line.n };
+    } else if (current && /^\s+\S/.test(line.text)) {
+      current.text += " " + line.text.trim();
+    } else if (line.text.trim() === "") {
+      if (current) {
+        out.push(current);
+        current = null;
+      }
+    }
+  }
+  if (current) out.push(current);
+  return out;
+}
+
+// =============================================================================
+// Determinism: stable sort
+// =============================================================================
+
+function sortNodes(nodes: GraphNode[]): GraphNode[] {
+  return [...nodes].sort((a, b) => (a.id < b.id ? -1 : a.id > b.id ? 1 : 0));
+}
+
+function sortEdges(edges: GraphEdge[]): GraphEdge[] {
+  return [...edges].sort((a, b) => {
+    const k = (e: GraphEdge): string =>
+      `${e.relation}${e.source}${e.target}${e.source_location ?? ""}`;
+    const ka = k(a);
+    const kb = k(b);
+    return ka < kb ? -1 : ka > kb ? 1 : 0;
+  });
+}
+
+// =============================================================================
+// Public entrypoint
+// =============================================================================
+
+export function extractProductSpec(input: { mdPath: string; mdContent: string }): ExtractResult {
+  const { mdPath, mdContent } = input;
+  const lines = splitLines(mdContent);
+  const ctx: Ctx = {
+    mdPath,
+    errors: [],
+    nodes: [],
+    edges: [],
+    personasByKey: new Map(),
+  };
+
+  const topSections = partitionSections(lines, 2, 0, lines.length);
+  const byHeading = (name: string): Section | undefined =>
+    topSections.find((s) => s.heading.trim().toLowerCase() === name.toLowerCase());
+
+  const required = ["App Overview", "Screen Inventory", "Cross-Feature Interactions"];
+  for (const r of required) {
+    if (!byHeading(r)) {
+      pushError(ctx, 1, `Missing required top-level section: "## ${r}"`);
+    }
+  }
+  if (ctx.errors.length > 0) {
+    return { ok: false, errors: ctx.errors };
+  }
+
+  // 1. Personas first — features need persona lookup.
+  const overviewSection = byHeading("App Overview")!;
+  parsePersonas(ctx, overviewSection);
+
+  // 2. Features (we want the feature nodes to exist before screen edges so
+  //    that has_screen edges target real feature ids; though screens encode
+  //    the target id directly via kebab, this also catches stray feature
+  //    names that don't match any feature heading — out of scope for fail
+  //    loud, but the feature_ids array reflects the inventory verbatim).
+  const featureSections = topSections.filter((s) => /^Feature\s*:/i.test(s.heading));
+  if (featureSections.length === 0) {
+    pushError(ctx, 1, "No `## Feature: ...` sections found");
+  }
+  for (const fs of featureSections) parseFeature(ctx, fs);
+
+  // 3. Screen inventory (after personas; doesn't depend on features but
+  //    emits feature-targeted edges).
+  parseScreenInventory(ctx, byHeading("Screen Inventory")!);
+
+  // 4. Cross-feature interactions (depends on nothing, but emits feature edges).
+  parseCrossFeature(ctx, byHeading("Cross-Feature Interactions")!);
+
+  if (ctx.errors.length > 0) {
+    return { ok: false, errors: ctx.errors };
+  }
+
+  const fragment: GraphFragment = {
+    version: 1,
+    schema: "buildanything-slice-1",
+    source_file: mdPath,
+    source_sha: sha256Hex(mdContent),
+    produced_at: new Date().toISOString(),
+    nodes: sortNodes(ctx.nodes),
+    edges: sortEdges(ctx.edges),
+  };
+  return { ok: true, fragment, errors: [] };
+}