memorix 1.4.3 → 1.4.5

package/src/orchestrate/task-graph.ts

@@ -1,389 +1,389 @@
-/**
- * Task Graph — Phase 6b: Structured task plan schema + DAG utilities.
- *
- * Defines the Zod schema for planner output, validates structure and
- * semantics, and provides DAG operations (topological sort, cycle
- * detection, parallel group discovery).
- *
- * The planner returns a TaskGraph; the coordinator materializes it
- * into TeamStore tasks with system-injected pipelineId.
- */
-
-import { z } from 'zod';
-
-// ── Zod Schema ─────────────────────────────────────────────────────
-
-export const TaskNodeSchema = z.object({
-  tempId: z.string().min(1).describe('Temporary ID like "t1", "t2"'),
-  role: z.enum(['pm', 'engineer', 'qa', 'reviewer']),
-  description: z.string().min(30).describe('Self-contained task description with all context'),
-  deps: z.array(z.string()).describe('tempIds of prerequisite tasks'),
-  files: z.array(z.string()).optional().describe('Files this task will create or modify'),
-  /** Phase 7: Natural-language acceptance criteria (Goal-First Testing) */
-  goals: z.array(z.string()).optional().describe('Acceptance criteria for task verification'),
-});
-
-export type TaskNode = z.infer<typeof TaskNodeSchema>;
-
-export const TaskGraphSchema = z.object({
-  summary: z.string().min(10).describe('One-paragraph plan summary'),
-  tasks: z.array(TaskNodeSchema).min(2).max(30),
-}).superRefine((data, ctx) => {
-  const ids = new Set(data.tasks.map(t => t.tempId));
-
-  // Validate: all dep references exist
-  for (const task of data.tasks) {
-    for (const dep of task.deps) {
-      if (!ids.has(dep)) {
-        ctx.addIssue({
-          code: z.ZodIssueCode.custom,
-          message: `Task "${task.tempId}" references unknown dependency "${dep}"`,
-          path: ['tasks'],
-        });
-      }
-    }
-    // Self-reference check
-    if (task.deps.includes(task.tempId)) {
-      ctx.addIssue({
-        code: z.ZodIssueCode.custom,
-        message: `Task "${task.tempId}" depends on itself`,
-        path: ['tasks'],
-      });
-    }
-  }
-
-  // Validate: last task must be a reviewer
-  if (data.tasks.length > 0 && data.tasks[data.tasks.length - 1].role !== 'reviewer') {
-    ctx.addIssue({
-      code: z.ZodIssueCode.custom,
-      message: 'Last task must have role "reviewer"',
-      path: ['tasks'],
-    });
-  }
-
-  // Validate: no cycles
-  if (hasCycle(data.tasks)) {
-    ctx.addIssue({
-      code: z.ZodIssueCode.custom,
-      message: 'Task graph contains a cycle',
-      path: ['tasks'],
-    });
-  }
-
-  // Validate: unique tempIds
-  if (ids.size !== data.tasks.length) {
-    ctx.addIssue({
-      code: z.ZodIssueCode.custom,
-      message: 'Duplicate tempId values detected',
-      path: ['tasks'],
-    });
-  }
-});
-
-export type TaskGraph = z.infer<typeof TaskGraphSchema>;
-
-// ── Parsing ────────────────────────────────────────────────────────
-
-/**
- * Parse and validate a JSON string into a TaskGraph.
- * Returns { success, data, error }.
- */
-export function parseTaskGraph(raw: string): {
-  success: true; data: TaskGraph; warnings: string[];
-} | {
-  success: false; error: string;
-} {
-  // Try to extract JSON from fenced code blocks or raw text
-  const json = extractJson(raw);
-  if (!json) {
-    return { success: false, error: 'No valid JSON found in planner output' };
-  }
-
-  let parsed: unknown;
-  try {
-    parsed = JSON.parse(json);
-  } catch (e) {
-    return { success: false, error: `JSON parse error: ${(e as Error).message}` };
-  }
-
-  const result = TaskGraphSchema.safeParse(parsed);
-  if (!result.success) {
-    const issues = result.error.issues.map(i => `${i.path.join('.')}: ${i.message}`);
-    return { success: false, error: `Validation failed:\n${issues.join('\n')}` };
-  }
-
-  const warnings = validateSemantics(result.data);
-  return { success: true, data: result.data, warnings };
-}
-
-// ── DAG Utilities ──────────────────────────────────────────────────
-
-/**
- * Detect cycles in the task graph using DFS.
- */
-export function hasCycle(tasks: Pick<TaskNode, 'tempId' | 'deps'>[]): boolean {
-  const WHITE = 0, GRAY = 1, BLACK = 2;
-  const color = new Map<string, number>();
-  for (const t of tasks) color.set(t.tempId, WHITE);
-
-  const adj = new Map<string, string[]>();
-  for (const t of tasks) adj.set(t.tempId, t.deps);
-
-  function dfs(node: string): boolean {
-    color.set(node, GRAY);
-    for (const dep of adj.get(node) ?? []) {
-      const c = color.get(dep);
-      if (c === GRAY) return true; // back edge → cycle
-      if (c === WHITE && dfs(dep)) return true;
-    }
-    color.set(node, BLACK);
-    return false;
-  }
-
-  for (const t of tasks) {
-    if (color.get(t.tempId) === WHITE && dfs(t.tempId)) return true;
-  }
-  return false;
-}
-
-/**
- * Topological sort (Kahn's algorithm). Returns tasks in execution order.
- * Throws if graph has a cycle.
- */
-export function topologicalSort(tasks: TaskNode[]): TaskNode[] {
-  const inDegree = new Map<string, number>();
-  const taskMap = new Map<string, TaskNode>();
-  const adj = new Map<string, string[]>(); // dep → dependents
-
-  for (const t of tasks) {
-    taskMap.set(t.tempId, t);
-    inDegree.set(t.tempId, t.deps.length);
-    if (!adj.has(t.tempId)) adj.set(t.tempId, []);
-    for (const dep of t.deps) {
-      if (!adj.has(dep)) adj.set(dep, []);
-      adj.get(dep)!.push(t.tempId);
-    }
-  }
-
-  const queue: string[] = [];
-  for (const [id, deg] of inDegree) {
-    if (deg === 0) queue.push(id);
-  }
-
-  const sorted: TaskNode[] = [];
-  while (queue.length > 0) {
-    const id = queue.shift()!;
-    sorted.push(taskMap.get(id)!);
-    for (const dependent of adj.get(id) ?? []) {
-      const newDeg = (inDegree.get(dependent) ?? 1) - 1;
-      inDegree.set(dependent, newDeg);
-      if (newDeg === 0) queue.push(dependent);
-    }
-  }
-
-  if (sorted.length !== tasks.length) {
-    throw new Error('Task graph contains a cycle — topological sort failed');
-  }
-  return sorted;
-}
-
-/**
- * Find groups of tasks that can execute in parallel.
- * Tasks in the same group have no mutual dependencies and all their
- * deps are satisfied at the same "level".
- */
-export function findParallelGroups(tasks: TaskNode[]): string[][] {
-  const inDegree = new Map<string, number>();
-  const adj = new Map<string, string[]>();
-
-  for (const t of tasks) {
-    inDegree.set(t.tempId, t.deps.length);
-    if (!adj.has(t.tempId)) adj.set(t.tempId, []);
-    for (const dep of t.deps) {
-      if (!adj.has(dep)) adj.set(dep, []);
-      adj.get(dep)!.push(t.tempId);
-    }
-  }
-
-  const groups: string[][] = [];
-  const remaining = new Set(tasks.map(t => t.tempId));
-
-  while (remaining.size > 0) {
-    const ready: string[] = [];
-    for (const id of remaining) {
-      if ((inDegree.get(id) ?? 0) === 0) ready.push(id);
-    }
-    if (ready.length === 0) break; // cycle guard
-    groups.push(ready);
-    for (const id of ready) {
-      remaining.delete(id);
-      for (const dep of adj.get(id) ?? []) {
-        inDegree.set(dep, (inDegree.get(dep) ?? 1) - 1);
-      }
-    }
-  }
-
-  return groups;
-}
-
-/**
- * Length of the longest dependency chain.
- */
-export function longestChain(tasks: Pick<TaskNode, 'tempId' | 'deps'>[]): number {
-  const memo = new Map<string, number>();
-  const taskMap = new Map(tasks.map(t => [t.tempId, t]));
-
-  function depth(id: string): number {
-    if (memo.has(id)) return memo.get(id)!;
-    const t = taskMap.get(id);
-    if (!t || t.deps.length === 0) { memo.set(id, 1); return 1; }
-    const maxDep = Math.max(...t.deps.map(d => depth(d)));
-    const d = maxDep + 1;
-    memo.set(id, d);
-    return d;
-  }
-
-  let max = 0;
-  for (const t of tasks) max = Math.max(max, depth(t.tempId));
-  return max;
-}
-
-// ── Semantic Validation (warnings, not errors) ─────────────────────
-
-function validateSemantics(graph: TaskGraph): string[] {
-  const warnings: string[] = [];
-
-  // Check for duplicate descriptions
-  const descs = graph.tasks.map(t => t.description.slice(0, 80));
-  if (new Set(descs).size < descs.length) {
-    warnings.push('Some tasks have very similar descriptions — consider making them more distinct');
-  }
-
-  // Check for overly linear graph (missed parallelism)
-  const chain = longestChain(graph.tasks);
-  if (chain > graph.tasks.length * 0.8 && graph.tasks.length > 3) {
-    warnings.push(`Task graph is nearly linear (chain=${chain}/${graph.tasks.length}) — independent tasks could run in parallel`);
-  }
-
-  // Check for very short descriptions
-  for (const t of graph.tasks) {
-    if (t.description.length < 80) {
-      warnings.push(`Task "${t.tempId}" has a short description (${t.description.length} chars) — agents may lack context`);
-    }
-  }
-
-  return warnings;
-}
-
-// ── Helpers ─────────────────────────────────────────────────────────
-
-/**
- * Extract JSON from raw LLM output — handles fenced code blocks,
- * bare JSON objects, and markdown-wrapped output.
- */
-function extractJson(raw: string): string | null {
-  // Handle Claude CLI output formats:
-  // 1. --output-format json: single {"type":"result","result":"...text..."}
-  // 2. --output-format stream-json: multiple lines, plan is in assistant text blocks
-  // 3. Plain text with fenced code blocks or bare JSON
-  let text = raw;
-
-  // Try single JSON envelope first
-  try {
-    const envelope = JSON.parse(raw.trim());
-    if (envelope && typeof envelope.result === 'string' && envelope.type === 'result') {
-      text = envelope.result;
-    }
-  } catch { /* not a single JSON envelope */ }
-
-  // Handle stream-json: extract text content from assistant messages
-  // Each line is a separate JSON object — collect all text blocks
-  if (text === raw && raw.includes('"type"')) {
-    const textBlocks: string[] = [];
-    for (const line of raw.split('\n')) {
-      const trimmed = line.trim();
-      if (!trimmed.startsWith('{')) continue;
-      try {
-        const obj = JSON.parse(trimmed);
-        // Claude stream-json assistant message with text content
-        if (obj.type === 'assistant' && Array.isArray(obj.message?.content)) {
-          for (const block of obj.message.content) {
-            if (block.type === 'text' && typeof block.text === 'string') {
-              textBlocks.push(block.text);
-            }
-          }
-        }
-        // Also check content_block_delta (streaming text chunks)
-        if (obj.type === 'content_block_delta' && obj.delta?.type === 'text_delta') {
-          textBlocks.push(obj.delta.text);
-        }
-        // Result message with subtype text
-        if (obj.type === 'result' && typeof obj.result === 'string') {
-          textBlocks.push(obj.result);
-        }
-      } catch { /* skip unparseable lines */ }
-    }
-    if (textBlocks.length > 0) {
-      text = textBlocks.join('');
-    }
-  }
-
-  // Try fenced code block first: ```json ... ``` or ``` ... ```
-  const fenced = text.match(/```(?:json)?\s*\n?([\s\S]*?)```/);
-  if (fenced) {
-    const candidate = fenced[1].trim();
-    if (candidate.startsWith('{')) return candidate;
-  }
-
-  // Try bare JSON object — use brace-counting to find matching closing brace
-  // This handles trailing text after the JSON (common with Gemini and other LLMs)
-  const braceStart = text.indexOf('{');
-  if (braceStart !== -1) {
-    const extracted = extractBalancedJson(text, braceStart);
-    if (extracted) return extracted;
-  }
-
-  return null;
-}
-
-/**
- * Extract a balanced JSON object starting at `startIdx` using brace-counting.
- * Respects string literals (skips braces inside quotes) and escape sequences.
- * Returns the substring if a balanced object is found, else null.
- */
-function extractBalancedJson(text: string, startIdx: number): string | null {
-  let depth = 0;
-  let inString = false;
-  let escape = false;
-
-  for (let i = startIdx; i < text.length; i++) {
-    const ch = text[i];
-
-    if (escape) {
-      escape = false;
-      continue;
-    }
-
-    if (ch === '\\' && inString) {
-      escape = true;
-      continue;
-    }
-
-    if (ch === '"') {
-      inString = !inString;
-      continue;
-    }
-
-    if (inString) continue;
-
-    if (ch === '{') depth++;
-    else if (ch === '}') {
-      depth--;
-      if (depth === 0) {
-        return text.slice(startIdx, i + 1);
-      }
-    }
-  }
-
-  return null; // unbalanced
-}
+/**
+ * Task Graph — Phase 6b: Structured task plan schema + DAG utilities.
+ *
+ * Defines the Zod schema for planner output, validates structure and
+ * semantics, and provides DAG operations (topological sort, cycle
+ * detection, parallel group discovery).
+ *
+ * The planner returns a TaskGraph; the coordinator materializes it
+ * into TeamStore tasks with system-injected pipelineId.
+ */
+
+import { z } from 'zod';
+
+// ── Zod Schema ─────────────────────────────────────────────────────
+
+export const TaskNodeSchema = z.object({
+  tempId: z.string().min(1).describe('Temporary ID like "t1", "t2"'),
+  role: z.enum(['pm', 'engineer', 'qa', 'reviewer']),
+  description: z.string().min(30).describe('Self-contained task description with all context'),
+  deps: z.array(z.string()).describe('tempIds of prerequisite tasks'),
+  files: z.array(z.string()).optional().describe('Files this task will create or modify'),
+  /** Phase 7: Natural-language acceptance criteria (Goal-First Testing) */
+  goals: z.array(z.string()).optional().describe('Acceptance criteria for task verification'),
+});
+
+export type TaskNode = z.infer<typeof TaskNodeSchema>;
+
+export const TaskGraphSchema = z.object({
+  summary: z.string().min(10).describe('One-paragraph plan summary'),
+  tasks: z.array(TaskNodeSchema).min(2).max(30),
+}).superRefine((data, ctx) => {
+  const ids = new Set(data.tasks.map(t => t.tempId));
+
+  // Validate: all dep references exist
+  for (const task of data.tasks) {
+    for (const dep of task.deps) {
+      if (!ids.has(dep)) {
+        ctx.addIssue({
+          code: z.ZodIssueCode.custom,
+          message: `Task "${task.tempId}" references unknown dependency "${dep}"`,
+          path: ['tasks'],
+        });
+      }
+    }
+    // Self-reference check
+    if (task.deps.includes(task.tempId)) {
+      ctx.addIssue({
+        code: z.ZodIssueCode.custom,
+        message: `Task "${task.tempId}" depends on itself`,
+        path: ['tasks'],
+      });
+    }
+  }
+
+  // Validate: last task must be a reviewer
+  if (data.tasks.length > 0 && data.tasks[data.tasks.length - 1].role !== 'reviewer') {
+    ctx.addIssue({
+      code: z.ZodIssueCode.custom,
+      message: 'Last task must have role "reviewer"',
+      path: ['tasks'],
+    });
+  }
+
+  // Validate: no cycles
+  if (hasCycle(data.tasks)) {
+    ctx.addIssue({
+      code: z.ZodIssueCode.custom,
+      message: 'Task graph contains a cycle',
+      path: ['tasks'],
+    });
+  }
+
+  // Validate: unique tempIds
+  if (ids.size !== data.tasks.length) {
+    ctx.addIssue({
+      code: z.ZodIssueCode.custom,
+      message: 'Duplicate tempId values detected',
+      path: ['tasks'],
+    });
+  }
+});
+
+export type TaskGraph = z.infer<typeof TaskGraphSchema>;
+
+// ── Parsing ────────────────────────────────────────────────────────
+
+/**
+ * Parse and validate a JSON string into a TaskGraph.
+ * Returns { success, data, error }.
+ */
+export function parseTaskGraph(raw: string): {
+  success: true; data: TaskGraph; warnings: string[];
+} | {
+  success: false; error: string;
+} {
+  // Try to extract JSON from fenced code blocks or raw text
+  const json = extractJson(raw);
+  if (!json) {
+    return { success: false, error: 'No valid JSON found in planner output' };
+  }
+
+  let parsed: unknown;
+  try {
+    parsed = JSON.parse(json);
+  } catch (e) {
+    return { success: false, error: `JSON parse error: ${(e as Error).message}` };
+  }
+
+  const result = TaskGraphSchema.safeParse(parsed);
+  if (!result.success) {
+    const issues = result.error.issues.map(i => `${i.path.join('.')}: ${i.message}`);
+    return { success: false, error: `Validation failed:\n${issues.join('\n')}` };
+  }
+
+  const warnings = validateSemantics(result.data);
+  return { success: true, data: result.data, warnings };
+}
+
+// ── DAG Utilities ──────────────────────────────────────────────────
+
+/**
+ * Detect cycles in the task graph using DFS.
+ */
+export function hasCycle(tasks: Pick<TaskNode, 'tempId' | 'deps'>[]): boolean {
+  const WHITE = 0, GRAY = 1, BLACK = 2;
+  const color = new Map<string, number>();
+  for (const t of tasks) color.set(t.tempId, WHITE);
+
+  const adj = new Map<string, string[]>();
+  for (const t of tasks) adj.set(t.tempId, t.deps);
+
+  function dfs(node: string): boolean {
+    color.set(node, GRAY);
+    for (const dep of adj.get(node) ?? []) {
+      const c = color.get(dep);
+      if (c === GRAY) return true; // back edge → cycle
+      if (c === WHITE && dfs(dep)) return true;
+    }
+    color.set(node, BLACK);
+    return false;
+  }
+
+  for (const t of tasks) {
+    if (color.get(t.tempId) === WHITE && dfs(t.tempId)) return true;
+  }
+  return false;
+}
+
+/**
+ * Topological sort (Kahn's algorithm). Returns tasks in execution order.
+ * Throws if graph has a cycle.
+ */
+export function topologicalSort(tasks: TaskNode[]): TaskNode[] {
+  const inDegree = new Map<string, number>();
+  const taskMap = new Map<string, TaskNode>();
+  const adj = new Map<string, string[]>(); // dep → dependents
+
+  for (const t of tasks) {
+    taskMap.set(t.tempId, t);
+    inDegree.set(t.tempId, t.deps.length);
+    if (!adj.has(t.tempId)) adj.set(t.tempId, []);
+    for (const dep of t.deps) {
+      if (!adj.has(dep)) adj.set(dep, []);
+      adj.get(dep)!.push(t.tempId);
+    }
+  }
+
+  const queue: string[] = [];
+  for (const [id, deg] of inDegree) {
+    if (deg === 0) queue.push(id);
+  }
+
+  const sorted: TaskNode[] = [];
+  while (queue.length > 0) {
+    const id = queue.shift()!;
+    sorted.push(taskMap.get(id)!);
+    for (const dependent of adj.get(id) ?? []) {
+      const newDeg = (inDegree.get(dependent) ?? 1) - 1;
+      inDegree.set(dependent, newDeg);
+      if (newDeg === 0) queue.push(dependent);
+    }
+  }
+
+  if (sorted.length !== tasks.length) {
+    throw new Error('Task graph contains a cycle — topological sort failed');
+  }
+  return sorted;
+}
+
+/**
+ * Find groups of tasks that can execute in parallel.
+ * Tasks in the same group have no mutual dependencies and all their
+ * deps are satisfied at the same "level".
+ */
+export function findParallelGroups(tasks: TaskNode[]): string[][] {
+  const inDegree = new Map<string, number>();
+  const adj = new Map<string, string[]>();
+
+  for (const t of tasks) {
+    inDegree.set(t.tempId, t.deps.length);
+    if (!adj.has(t.tempId)) adj.set(t.tempId, []);
+    for (const dep of t.deps) {
+      if (!adj.has(dep)) adj.set(dep, []);
+      adj.get(dep)!.push(t.tempId);
+    }
+  }
+
+  const groups: string[][] = [];
+  const remaining = new Set(tasks.map(t => t.tempId));
+
+  while (remaining.size > 0) {
+    const ready: string[] = [];
+    for (const id of remaining) {
+      if ((inDegree.get(id) ?? 0) === 0) ready.push(id);
+    }
+    if (ready.length === 0) break; // cycle guard
+    groups.push(ready);
+    for (const id of ready) {
+      remaining.delete(id);
+      for (const dep of adj.get(id) ?? []) {
+        inDegree.set(dep, (inDegree.get(dep) ?? 1) - 1);
+      }
+    }
+  }
+
+  return groups;
+}
+
+/**
+ * Length of the longest dependency chain.
+ */
+export function longestChain(tasks: Pick<TaskNode, 'tempId' | 'deps'>[]): number {
+  const memo = new Map<string, number>();
+  const taskMap = new Map(tasks.map(t => [t.tempId, t]));
+
+  function depth(id: string): number {
+    if (memo.has(id)) return memo.get(id)!;
+    const t = taskMap.get(id);
+    if (!t || t.deps.length === 0) { memo.set(id, 1); return 1; }
+    const maxDep = Math.max(...t.deps.map(d => depth(d)));
+    const d = maxDep + 1;
+    memo.set(id, d);
+    return d;
+  }
+
+  let max = 0;
+  for (const t of tasks) max = Math.max(max, depth(t.tempId));
+  return max;
+}
+
+// ── Semantic Validation (warnings, not errors) ─────────────────────
+
+function validateSemantics(graph: TaskGraph): string[] {
+  const warnings: string[] = [];
+
+  // Check for duplicate descriptions
+  const descs = graph.tasks.map(t => t.description.slice(0, 80));
+  if (new Set(descs).size < descs.length) {
+    warnings.push('Some tasks have very similar descriptions — consider making them more distinct');
+  }
+
+  // Check for overly linear graph (missed parallelism)
+  const chain = longestChain(graph.tasks);
+  if (chain > graph.tasks.length * 0.8 && graph.tasks.length > 3) {
+    warnings.push(`Task graph is nearly linear (chain=${chain}/${graph.tasks.length}) — independent tasks could run in parallel`);
+  }
+
+  // Check for very short descriptions
+  for (const t of graph.tasks) {
+    if (t.description.length < 80) {
+      warnings.push(`Task "${t.tempId}" has a short description (${t.description.length} chars) — agents may lack context`);
+    }
+  }
+
+  return warnings;
+}
+
+// ── Helpers ─────────────────────────────────────────────────────────
+
+/**
+ * Extract JSON from raw LLM output — handles fenced code blocks,
+ * bare JSON objects, and markdown-wrapped output.
+ */
+function extractJson(raw: string): string | null {
+  // Handle Claude CLI output formats:
+  // 1. --output-format json: single {"type":"result","result":"...text..."}
+  // 2. --output-format stream-json: multiple lines, plan is in assistant text blocks
+  // 3. Plain text with fenced code blocks or bare JSON
+  let text = raw;
+
+  // Try single JSON envelope first
+  try {
+    const envelope = JSON.parse(raw.trim());
+    if (envelope && typeof envelope.result === 'string' && envelope.type === 'result') {
+      text = envelope.result;
+    }
+  } catch { /* not a single JSON envelope */ }
+
+  // Handle stream-json: extract text content from assistant messages
+  // Each line is a separate JSON object — collect all text blocks
+  if (text === raw && raw.includes('"type"')) {
+    const textBlocks: string[] = [];
+    for (const line of raw.split('\n')) {
+      const trimmed = line.trim();
+      if (!trimmed.startsWith('{')) continue;
+      try {
+        const obj = JSON.parse(trimmed);
+        // Claude stream-json assistant message with text content
+        if (obj.type === 'assistant' && Array.isArray(obj.message?.content)) {
+          for (const block of obj.message.content) {
+            if (block.type === 'text' && typeof block.text === 'string') {
+              textBlocks.push(block.text);
+            }
+          }
+        }
+        // Also check content_block_delta (streaming text chunks)
+        if (obj.type === 'content_block_delta' && obj.delta?.type === 'text_delta') {
+          textBlocks.push(obj.delta.text);
+        }
+        // Result message with subtype text
+        if (obj.type === 'result' && typeof obj.result === 'string') {
+          textBlocks.push(obj.result);
+        }
+      } catch { /* skip unparseable lines */ }
+    }
+    if (textBlocks.length > 0) {
+      text = textBlocks.join('');
+    }
+  }
+
+  // Try fenced code block first: ```json ... ``` or ``` ... ```
+  const fenced = text.match(/```(?:json)?\s*\n?([\s\S]*?)```/);
+  if (fenced) {
+    const candidate = fenced[1].trim();
+    if (candidate.startsWith('{')) return candidate;
+  }
+
+  // Try bare JSON object — use brace-counting to find matching closing brace
+  // This handles trailing text after the JSON (common with Gemini and other LLMs)
+  const braceStart = text.indexOf('{');
+  if (braceStart !== -1) {
+    const extracted = extractBalancedJson(text, braceStart);
+    if (extracted) return extracted;
+  }
+
+  return null;
+}
+
+/**
+ * Extract a balanced JSON object starting at `startIdx` using brace-counting.
+ * Respects string literals (skips braces inside quotes) and escape sequences.
+ * Returns the substring if a balanced object is found, else null.
+ */
+function extractBalancedJson(text: string, startIdx: number): string | null {
+  let depth = 0;
+  let inString = false;
+  let escape = false;
+
+  for (let i = startIdx; i < text.length; i++) {
+    const ch = text[i];
+
+    if (escape) {
+      escape = false;
+      continue;
+    }
+
+    if (ch === '\\' && inString) {
+      escape = true;
+      continue;
+    }
+
+    if (ch === '"') {
+      inString = !inString;
+      continue;
+    }
+
+    if (inString) continue;
+
+    if (ch === '{') depth++;
+    else if (ch === '}') {
+      depth--;
+      if (depth === 0) {
+        return text.slice(startIdx, i + 1);
+      }
+    }
+  }
+
+  return null; // unbalanced
+}