opencode-swarm-plugin 0.37.0 → 0.39.1

package/dist/plugin.js

@@ -22178,6 +22178,383 @@ Codebase context considered: ${args.codebase_context.slice(0, 200)}...`;
   };
 });
 
+// src/eval-capture.ts
+var exports_eval_capture = {};
+__export(exports_eval_capture, {
+  updateEvalRecord: () => updateEvalRecord,
+  saveSession: () => saveSession,
+  readSessionEvents: () => readSessionEvents,
+  readPartialRecords: () => readPartialRecords,
+  readEvalRecords: () => readEvalRecords,
+  getSessionPath: () => getSessionPath,
+  getSessionDir: () => getSessionDir,
+  getEvalDataStats: () => getEvalDataStats,
+  getEvalDataPath: () => getEvalDataPath,
+  finalizeEvalRecord: () => finalizeEvalRecord,
+  exportForEvalite: () => exportForEvalite,
+  ensureSessionDir: () => ensureSessionDir,
+  ensureEvalDataDir: () => ensureEvalDataDir,
+  captureSubtaskOutcome: () => captureSubtaskOutcome,
+  captureHumanFeedback: () => captureHumanFeedback,
+  captureDecomposition: () => captureDecomposition,
+  captureCoordinatorEvent: () => captureCoordinatorEvent,
+  captureCompactionEvent: () => captureCompactionEvent,
+  appendEvalRecord: () => appendEvalRecord,
+  SubtaskOutcomeSchema: () => SubtaskOutcomeSchema,
+  EvalRecordSchema: () => EvalRecordSchema,
+  DEFAULT_EVAL_DATA_PATH: () => DEFAULT_EVAL_DATA_PATH,
+  CoordinatorSessionSchema: () => CoordinatorSessionSchema,
+  CoordinatorEventSchema: () => CoordinatorEventSchema
+});
+import * as fs from "node:fs";
+import * as os from "node:os";
+import * as path from "node:path";
+function getEvalDataPath(projectPath) {
+  return path.join(projectPath, DEFAULT_EVAL_DATA_PATH);
+}
+function ensureEvalDataDir(projectPath) {
+  const evalPath = getEvalDataPath(projectPath);
+  const dir = path.dirname(evalPath);
+  if (!fs.existsSync(dir)) {
+    fs.mkdirSync(dir, { recursive: true });
+  }
+}
+function appendEvalRecord(projectPath, record2) {
+  ensureEvalDataDir(projectPath);
+  const evalPath = getEvalDataPath(projectPath);
+  const line = `${JSON.stringify(record2)}
+`;
+  fs.appendFileSync(evalPath, line, "utf-8");
+}
+function readEvalRecords(projectPath) {
+  const evalPath = getEvalDataPath(projectPath);
+  if (!fs.existsSync(evalPath)) {
+    return [];
+  }
+  const content = fs.readFileSync(evalPath, "utf-8");
+  const lines = content.trim().split(`
+`).filter(Boolean);
+  return lines.map((line) => {
+    const parsed = JSON.parse(line);
+    return EvalRecordSchema.parse(parsed);
+  });
+}
+function readPartialRecords(projectPath) {
+  const evalPath = getEvalDataPath(projectPath);
+  if (!fs.existsSync(evalPath)) {
+    return [];
+  }
+  const content = fs.readFileSync(evalPath, "utf-8");
+  const lines = content.trim().split(`
+`).filter(Boolean);
+  return lines.map((line) => JSON.parse(line));
+}
+function updateEvalRecord(projectPath, id, updates) {
+  const records = readPartialRecords(projectPath);
+  const index = records.findIndex((r) => r.id === id);
+  if (index === -1) {
+    return false;
+  }
+  records[index] = { ...records[index], ...updates };
+  const evalPath = getEvalDataPath(projectPath);
+  const content = `${records.map((r) => JSON.stringify(r)).join(`
+`)}
+`;
+  fs.writeFileSync(evalPath, content, "utf-8");
+  return true;
+}
+function captureDecomposition(params) {
+  const record2 = {
+    id: params.epicId,
+    timestamp: new Date().toISOString(),
+    project_path: params.projectPath,
+    task: params.task,
+    context: params.context,
+    strategy: params.strategy,
+    subtask_count: params.subtasks.length,
+    epic_title: params.epicTitle,
+    epic_description: params.epicDescription,
+    subtasks: params.subtasks,
+    outcomes: []
+  };
+  inProgressRecords.set(params.epicId, record2);
+  appendEvalRecord(params.projectPath, record2);
+  return record2;
+}
+function captureSubtaskOutcome(params) {
+  const outcome = {
+    bead_id: params.beadId,
+    title: params.title,
+    planned_files: params.plannedFiles,
+    actual_files: params.actualFiles,
+    duration_ms: params.durationMs,
+    error_count: params.errorCount,
+    retry_count: params.retryCount,
+    success: params.success,
+    failure_mode: params.failureMode
+  };
+  const record2 = inProgressRecords.get(params.epicId);
+  if (record2) {
+    record2.outcomes = record2.outcomes || [];
+    record2.outcomes.push(outcome);
+  }
+  updateEvalRecord(params.projectPath, params.epicId, {
+    outcomes: record2?.outcomes
+  });
+}
+function finalizeEvalRecord(params) {
+  const record2 = inProgressRecords.get(params.epicId);
+  if (!record2 || !record2.outcomes || record2.outcomes.length === 0) {
+    return null;
+  }
+  const outcomes = record2.outcomes;
+  const overallSuccess = outcomes.every((o) => o.success);
+  const totalDurationMs = outcomes.reduce((sum, o) => sum + o.duration_ms, 0);
+  const totalErrors = outcomes.reduce((sum, o) => sum + o.error_count, 0);
+  const allPlannedFiles = record2.subtasks?.flatMap((s) => s.files) || [];
+  const fileOccurrences = new Map;
+  for (const file2 of allPlannedFiles) {
+    fileOccurrences.set(file2, (fileOccurrences.get(file2) || 0) + 1);
+  }
+  const fileOverlapCount = Array.from(fileOccurrences.values()).filter((count) => count > 1).length;
+  const plannedFileSet = new Set(allPlannedFiles);
+  const actualFileSet = new Set(outcomes.flatMap((o) => o.actual_files));
+  const scopeAccuracy = plannedFileSet.size > 0 ? actualFileSet.size / plannedFileSet.size : 1;
+  const durations = outcomes.map((o) => o.duration_ms).filter((d) => d > 0);
+  const timeBalanceRatio = durations.length > 1 ? Math.max(...durations) / Math.min(...durations) : 1;
+  const finalRecord = {
+    ...record2,
+    overall_success: overallSuccess,
+    total_duration_ms: totalDurationMs,
+    total_errors: totalErrors,
+    file_overlap_count: fileOverlapCount,
+    scope_accuracy: scopeAccuracy,
+    time_balance_ratio: timeBalanceRatio
+  };
+  updateEvalRecord(params.projectPath, params.epicId, finalRecord);
+  inProgressRecords.delete(params.epicId);
+  return finalRecord;
+}
+function captureHumanFeedback(params) {
+  updateEvalRecord(params.projectPath, params.epicId, {
+    human_accepted: params.accepted,
+    human_modified: params.modified,
+    human_notes: params.notes
+  });
+}
+function exportForEvalite(projectPath) {
+  const records = readEvalRecords(projectPath);
+  return records.filter((r) => r.outcomes && r.outcomes.length > 0).map((record2) => ({
+    input: {
+      task: record2.task,
+      context: record2.context
+    },
+    expected: {
+      minSubtasks: 2,
+      subtaskCount: record2.subtask_count,
+      requiredFiles: record2.subtasks.flatMap((s) => s.files),
+      overallSuccess: record2.overall_success
+    },
+    actual: record2
+  }));
+}
+function getEvalDataStats(projectPath) {
+  const records = readEvalRecords(projectPath);
+  const complete = records.filter((r) => r.outcomes && r.outcomes.length > 0);
+  if (complete.length === 0) {
+    return {
+      totalRecords: records.length,
+      completeRecords: 0,
+      successRate: 0,
+      avgSubtasks: 0,
+      avgDurationMs: 0,
+      avgScopeAccuracy: 0,
+      avgTimeBalance: 0
+    };
+  }
+  const successCount = complete.filter((r) => r.overall_success).length;
+  const avgSubtasks = complete.reduce((sum, r) => sum + (r.outcomes?.length || 0), 0) / complete.length;
+  const avgDurationMs = complete.reduce((sum, r) => sum + (r.total_duration_ms || 0), 0) / complete.length;
+  const avgScopeAccuracy = complete.reduce((sum, r) => sum + (r.scope_accuracy || 1), 0) / complete.length;
+  const avgTimeBalance = complete.reduce((sum, r) => sum + (r.time_balance_ratio || 1), 0) / complete.length;
+  return {
+    totalRecords: records.length,
+    completeRecords: complete.length,
+    successRate: successCount / complete.length,
+    avgSubtasks,
+    avgDurationMs,
+    avgScopeAccuracy,
+    avgTimeBalance
+  };
+}
+function getSessionDir() {
+  return path.join(os.homedir(), ".config", "swarm-tools", "sessions");
+}
+function getSessionPath(sessionId) {
+  return path.join(getSessionDir(), `${sessionId}.jsonl`);
+}
+function ensureSessionDir() {
+  const sessionDir = getSessionDir();
+  if (!fs.existsSync(sessionDir)) {
+    fs.mkdirSync(sessionDir, { recursive: true });
+  }
+}
+function captureCoordinatorEvent(event) {
+  CoordinatorEventSchema.parse(event);
+  ensureSessionDir();
+  const sessionPath = getSessionPath(event.session_id);
+  const line = `${JSON.stringify(event)}
+`;
+  fs.appendFileSync(sessionPath, line, "utf-8");
+}
+function captureCompactionEvent(params) {
+  const event = {
+    session_id: params.session_id,
+    epic_id: params.epic_id,
+    timestamp: new Date().toISOString(),
+    event_type: "COMPACTION",
+    compaction_type: params.compaction_type,
+    payload: params.payload
+  };
+  captureCoordinatorEvent(event);
+}
+function readSessionEvents(sessionId) {
+  const sessionPath = getSessionPath(sessionId);
+  if (!fs.existsSync(sessionPath)) {
+    return [];
+  }
+  const content = fs.readFileSync(sessionPath, "utf-8");
+  const lines = content.trim().split(`
+`).filter(Boolean);
+  return lines.map((line) => {
+    const parsed = JSON.parse(line);
+    return CoordinatorEventSchema.parse(parsed);
+  });
+}
+function saveSession(params) {
+  const events = readSessionEvents(params.session_id);
+  if (events.length === 0) {
+    return null;
+  }
+  const timestamps = events.map((e) => new Date(e.timestamp).getTime());
+  const startTime = new Date(Math.min(...timestamps)).toISOString();
+  const endTime = new Date(Math.max(...timestamps)).toISOString();
+  const session = {
+    session_id: params.session_id,
+    epic_id: params.epic_id,
+    start_time: startTime,
+    end_time: endTime,
+    events
+  };
+  return session;
+}
+var SubtaskOutcomeSchema, EvalRecordSchema, CoordinatorEventSchema, CoordinatorSessionSchema, DEFAULT_EVAL_DATA_PATH = ".opencode/eval-data.jsonl", inProgressRecords;
+var init_eval_capture = __esm(() => {
+  init_zod();
+  SubtaskOutcomeSchema = exports_external.object({
+    bead_id: exports_external.string(),
+    title: exports_external.string(),
+    planned_files: exports_external.array(exports_external.string()),
+    actual_files: exports_external.array(exports_external.string()),
+    duration_ms: exports_external.number().int().min(0),
+    error_count: exports_external.number().int().min(0),
+    retry_count: exports_external.number().int().min(0),
+    success: exports_external.boolean(),
+    failure_mode: exports_external.string().optional()
+  });
+  EvalRecordSchema = exports_external.object({
+    id: exports_external.string(),
+    timestamp: exports_external.string(),
+    project_path: exports_external.string(),
+    task: exports_external.string(),
+    context: exports_external.string().optional(),
+    strategy: exports_external.enum(["file-based", "feature-based", "risk-based", "auto"]),
+    subtask_count: exports_external.number().int().min(1),
+    epic_title: exports_external.string(),
+    epic_description: exports_external.string().optional(),
+    subtasks: exports_external.array(exports_external.object({
+      title: exports_external.string(),
+      description: exports_external.string().optional(),
+      files: exports_external.array(exports_external.string()),
+      dependencies: exports_external.array(exports_external.number()).optional(),
+      estimated_complexity: exports_external.number().int().min(1).max(5).optional()
+    })),
+    outcomes: exports_external.array(SubtaskOutcomeSchema).optional(),
+    overall_success: exports_external.boolean().optional(),
+    total_duration_ms: exports_external.number().int().min(0).optional(),
+    total_errors: exports_external.number().int().min(0).optional(),
+    human_accepted: exports_external.boolean().optional(),
+    human_modified: exports_external.boolean().optional(),
+    human_notes: exports_external.string().optional(),
+    file_overlap_count: exports_external.number().int().min(0).optional(),
+    scope_accuracy: exports_external.number().min(0).max(2).optional(),
+    time_balance_ratio: exports_external.number().min(1).optional()
+  });
+  CoordinatorEventSchema = exports_external.discriminatedUnion("event_type", [
+    exports_external.object({
+      session_id: exports_external.string(),
+      epic_id: exports_external.string(),
+      timestamp: exports_external.string(),
+      event_type: exports_external.literal("DECISION"),
+      decision_type: exports_external.enum([
+        "strategy_selected",
+        "worker_spawned",
+        "review_completed",
+        "decomposition_complete"
+      ]),
+      payload: exports_external.any()
+    }),
+    exports_external.object({
+      session_id: exports_external.string(),
+      epic_id: exports_external.string(),
+      timestamp: exports_external.string(),
+      event_type: exports_external.literal("VIOLATION"),
+      violation_type: exports_external.enum([
+        "coordinator_edited_file",
+        "coordinator_ran_tests",
+        "coordinator_reserved_files",
+        "no_worker_spawned"
+      ]),
+      payload: exports_external.any()
+    }),
+    exports_external.object({
+      session_id: exports_external.string(),
+      epic_id: exports_external.string(),
+      timestamp: exports_external.string(),
+      event_type: exports_external.literal("OUTCOME"),
+      outcome_type: exports_external.enum([
+        "subtask_success",
+        "subtask_retry",
+        "subtask_failed",
+        "epic_complete"
+      ]),
+      payload: exports_external.any()
+    }),
+    exports_external.object({
+      session_id: exports_external.string(),
+      epic_id: exports_external.string(),
+      timestamp: exports_external.string(),
+      event_type: exports_external.literal("COMPACTION"),
+      compaction_type: exports_external.enum([
+        "detection_complete",
+        "prompt_generated",
+        "context_injected",
+        "resumption_started",
+        "tool_call_tracked"
+      ]),
+      payload: exports_external.any()
+    })
+  ]);
+  CoordinatorSessionSchema = exports_external.object({
+    session_id: exports_external.string(),
+    epic_id: exports_external.string(),
+    start_time: exports_external.string(),
+    end_time: exports_external.string().optional(),
+    events: exports_external.array(CoordinatorEventSchema)
+  });
+  inProgressRecords = new Map;
+});
+
 // src/learning.ts
 var exports_learning = {};
 __export(exports_learning, {
@@ -39165,6 +39542,71 @@ var hive_ready = tool({
     }
   }
 });
+var hive_cells = tool({
+  description: `Query cells from the hive database with flexible filtering.
+
+USE THIS TOOL TO:
+- List all open cells: hive_cells()
+- Find cells by status: hive_cells({ status: "in_progress" })
+- Find cells by type: hive_cells({ type: "bug" })
+- Get a specific cell by partial ID: hive_cells({ id: "mjkmd" })
+- Get the next ready (unblocked) cell: hive_cells({ ready: true })
+- Combine filters: hive_cells({ status: "open", type: "task" })
+
+RETURNS: Array of cells with id, title, status, priority, type, parent_id, created_at, updated_at
+
+PREFER THIS OVER hive_query when you need to:
+- See what work is available
+- Check status of multiple cells
+- Find cells matching criteria
+- Look up a cell by partial ID`,
+  args: {
+    id: tool.schema.string().optional().describe("Partial or full cell ID to look up"),
+    status: tool.schema.enum(["open", "in_progress", "blocked", "closed"]).optional().describe("Filter by status"),
+    type: tool.schema.enum(["task", "bug", "feature", "epic", "chore"]).optional().describe("Filter by type"),
+    ready: tool.schema.boolean().optional().describe("If true, return only the next unblocked cell"),
+    limit: tool.schema.number().optional().describe("Max cells to return (default 20)")
+  },
+  async execute(args, ctx) {
+    const projectKey = getHiveWorkingDirectory();
+    const adapter = await getHiveAdapter(projectKey);
+    try {
+      if (args.id) {
+        const fullId = await resolvePartialId(adapter, projectKey, args.id) || args.id;
+        const cell = await adapter.getCell(projectKey, fullId);
+        if (!cell) {
+          throw new HiveError(`No cell found matching ID '${args.id}'`, "hive_cells");
+        }
+        const formatted2 = formatCellForOutput(cell);
+        return JSON.stringify([formatted2], null, 2);
+      }
+      if (args.ready) {
+        const ready = await adapter.getNextReadyCell(projectKey);
+        if (!ready) {
+          return JSON.stringify([], null, 2);
+        }
+        const formatted2 = formatCellForOutput(ready);
+        return JSON.stringify([formatted2], null, 2);
+      }
+      const cells = await adapter.queryCells(projectKey, {
+        status: args.status,
+        type: args.type,
+        limit: args.limit || 20
+      });
+      const formatted = cells.map((c) => formatCellForOutput(c));
+      return JSON.stringify(formatted, null, 2);
+    } catch (error45) {
+      const message = error45 instanceof Error ? error45.message : String(error45);
+      if (message.includes("Ambiguous hash")) {
+        throw new HiveError(`Ambiguous ID '${args.id}': multiple cells match. Please provide more characters.`, "hive_cells");
+      }
+      if (message.includes("Bead not found") || message.includes("Cell not found")) {
+        throw new HiveError(`No cell found matching ID '${args.id || "unknown"}'`, "hive_cells");
+      }
+      throw new HiveError(`Failed to query cells: ${message}`, "hive_cells");
+    }
+  }
+});
 var hive_sync = tool({
   description: "Sync hive to git and push (MANDATORY at session end)",
   args: {
@@ -39306,6 +39748,7 @@ var hiveTools = {
   hive_close,
   hive_start,
   hive_ready,
+  hive_cells,
   hive_sync,
   hive_link_thread
 };
@@ -41464,122 +41907,7 @@ init_swarm_strategies();
 init_dist();
 init_zod();
 init_swarm_strategies();
-
-// src/eval-capture.ts
-init_zod();
-import * as fs from "node:fs";
-import * as os from "node:os";
-import * as path from "node:path";
-var SubtaskOutcomeSchema = exports_external.object({
-  bead_id: exports_external.string(),
-  title: exports_external.string(),
-  planned_files: exports_external.array(exports_external.string()),
-  actual_files: exports_external.array(exports_external.string()),
-  duration_ms: exports_external.number().int().min(0),
-  error_count: exports_external.number().int().min(0),
-  retry_count: exports_external.number().int().min(0),
-  success: exports_external.boolean(),
-  failure_mode: exports_external.string().optional()
-});
-var EvalRecordSchema = exports_external.object({
-  id: exports_external.string(),
-  timestamp: exports_external.string(),
-  project_path: exports_external.string(),
-  task: exports_external.string(),
-  context: exports_external.string().optional(),
-  strategy: exports_external.enum(["file-based", "feature-based", "risk-based", "auto"]),
-  subtask_count: exports_external.number().int().min(1),
-  epic_title: exports_external.string(),
-  epic_description: exports_external.string().optional(),
-  subtasks: exports_external.array(exports_external.object({
-    title: exports_external.string(),
-    description: exports_external.string().optional(),
-    files: exports_external.array(exports_external.string()),
-    dependencies: exports_external.array(exports_external.number()).optional(),
-    estimated_complexity: exports_external.number().int().min(1).max(5).optional()
-  })),
-  outcomes: exports_external.array(SubtaskOutcomeSchema).optional(),
-  overall_success: exports_external.boolean().optional(),
-  total_duration_ms: exports_external.number().int().min(0).optional(),
-  total_errors: exports_external.number().int().min(0).optional(),
-  human_accepted: exports_external.boolean().optional(),
-  human_modified: exports_external.boolean().optional(),
-  human_notes: exports_external.string().optional(),
-  file_overlap_count: exports_external.number().int().min(0).optional(),
-  scope_accuracy: exports_external.number().min(0).max(2).optional(),
-  time_balance_ratio: exports_external.number().min(1).optional()
-});
-var CoordinatorEventSchema = exports_external.discriminatedUnion("event_type", [
-  exports_external.object({
-    session_id: exports_external.string(),
-    epic_id: exports_external.string(),
-    timestamp: exports_external.string(),
-    event_type: exports_external.literal("DECISION"),
-    decision_type: exports_external.enum([
-      "strategy_selected",
-      "worker_spawned",
-      "review_completed",
-      "decomposition_complete"
-    ]),
-    payload: exports_external.any()
-  }),
-  exports_external.object({
-    session_id: exports_external.string(),
-    epic_id: exports_external.string(),
-    timestamp: exports_external.string(),
-    event_type: exports_external.literal("VIOLATION"),
-    violation_type: exports_external.enum([
-      "coordinator_edited_file",
-      "coordinator_ran_tests",
-      "coordinator_reserved_files",
-      "no_worker_spawned"
-    ]),
-    payload: exports_external.any()
-  }),
-  exports_external.object({
-    session_id: exports_external.string(),
-    epic_id: exports_external.string(),
-    timestamp: exports_external.string(),
-    event_type: exports_external.literal("OUTCOME"),
-    outcome_type: exports_external.enum([
-      "subtask_success",
-      "subtask_retry",
-      "subtask_failed",
-      "epic_complete"
-    ]),
-    payload: exports_external.any()
-  })
-]);
-var CoordinatorSessionSchema = exports_external.object({
-  session_id: exports_external.string(),
-  epic_id: exports_external.string(),
-  start_time: exports_external.string(),
-  end_time: exports_external.string().optional(),
-  events: exports_external.array(CoordinatorEventSchema)
-});
-var inProgressRecords = new Map;
-function getSessionDir() {
-  return path.join(os.homedir(), ".config", "swarm-tools", "sessions");
-}
-function getSessionPath(sessionId) {
-  return path.join(getSessionDir(), `${sessionId}.jsonl`);
-}
-function ensureSessionDir() {
-  const sessionDir = getSessionDir();
-  if (!fs.existsSync(sessionDir)) {
-    fs.mkdirSync(sessionDir, { recursive: true });
-  }
-}
-function captureCoordinatorEvent(event) {
-  CoordinatorEventSchema.parse(event);
-  ensureSessionDir();
-  const sessionPath = getSessionPath(event.session_id);
-  const line = `${JSON.stringify(event)}
-`;
-  fs.appendFileSync(sessionPath, line, "utf-8");
-}
-
-// src/swarm-decompose.ts
+init_eval_capture();
 var DECOMPOSITION_PROMPT = `You are decomposing a task into parallelizable subtasks for a swarm of agents.
 
 ## Task
@@ -41897,9 +42225,14 @@ ${fullContext}` : `## Additional Context
   }
 });
 var swarm_validate_decomposition = tool({
-  description: "Validate a decomposition response against CellTreeSchema",
+  description: "Validate a decomposition response against CellTreeSchema and capture for eval",
   args: {
-    response: tool.schema.string().describe("JSON response from agent (CellTree format)")
+    response: tool.schema.string().describe("JSON response from agent (CellTree format)"),
+    project_path: tool.schema.string().optional().describe("Project path for eval capture"),
+    task: tool.schema.string().optional().describe("Original task description for eval capture"),
+    context: tool.schema.string().optional().describe("Context provided for decomposition"),
+    strategy: tool.schema.enum(["file-based", "feature-based", "risk-based", "auto"]).optional().describe("Decomposition strategy used"),
+    epic_id: tool.schema.string().optional().describe("Epic ID for eval capture")
   },
   async execute(args) {
     try {
@@ -41933,6 +42266,29 @@ var swarm_validate_decomposition = tool({
         }
       }
       const instructionConflicts = detectInstructionConflicts(validated.subtasks);
+      if (args.project_path && args.task && args.strategy && args.epic_id) {
+        try {
+          const { captureDecomposition: captureDecomposition2 } = await Promise.resolve().then(() => (init_eval_capture(), exports_eval_capture));
+          captureDecomposition2({
+            epicId: args.epic_id,
+            projectPath: args.project_path,
+            task: args.task,
+            context: args.context,
+            strategy: args.strategy,
+            epicTitle: validated.epic.title,
+            epicDescription: validated.epic.description,
+            subtasks: validated.subtasks.map((s) => ({
+              title: s.title,
+              description: s.description,
+              files: s.files,
+              dependencies: s.dependencies,
+              estimated_complexity: s.estimated_complexity
+            }))
+          });
+        } catch (error45) {
+          console.warn("[swarm_validate_decomposition] Failed to capture decomposition:", error45);
+        }
+      }
       return JSON.stringify({
         valid: true,
         cell_tree: validated,
@@ -41973,7 +42329,7 @@ var swarm_delegate_planning = tool({
     strategy: tool.schema.enum(["auto", "file-based", "feature-based", "risk-based"]).optional().default("auto").describe("Decomposition strategy (default: auto-detect)"),
     query_cass: tool.schema.boolean().optional().default(true).describe("Query CASS for similar past tasks (default: true)")
   },
-  async execute(args) {
+  async execute(args, _ctx) {
     const { selectStrategy: selectStrategy2, formatStrategyGuidelines: formatStrategyGuidelines2 } = await Promise.resolve().then(() => (init_swarm_strategies(), exports_swarm_strategies));
     const { formatMemoryQueryForDecomposition: formatMemoryQueryForDecomposition2 } = await Promise.resolve().then(() => (init_learning(), exports_learning));
     const { listSkills: listSkills2, getSkillsContextForSwarm: getSkillsContextForSwarm2, findRelevantSkills: findRelevantSkills2 } = await Promise.resolve().then(() => (init_skills(), exports_skills));
@@ -41989,7 +42345,7 @@ var swarm_delegate_planning = tool({
     }
     try {
       captureCoordinatorEvent({
-        session_id: process.env.OPENCODE_SESSION_ID || "unknown",
+        session_id: _ctx.sessionID || "unknown",
         epic_id: "planning",
         timestamp: new Date().toISOString(),
         event_type: "DECISION",
@@ -44080,6 +44436,7 @@ var worktreeTools = {
 init_dist();
 init_zod();
 import { sendSwarmMessage as sendSwarmMessage2 } from "swarm-mail";
+init_eval_capture();
 var ReviewIssueSchema = exports_external.object({
   file: exports_external.string(),
   line: exports_external.number().optional(),
@@ -44304,7 +44661,7 @@ var swarm_review_feedback = tool({
     summary: exports_external.string().optional().describe("Review summary"),
     issues: exports_external.string().optional().describe("JSON array of ReviewIssue objects (for needs_changes)")
   },
-  async execute(args) {
+  async execute(args, _ctx) {
     let parsedIssues = [];
     if (args.issues) {
       try {
@@ -44327,7 +44684,7 @@ var swarm_review_feedback = tool({
       markReviewApproved(args.task_id);
       try {
         captureCoordinatorEvent({
-          session_id: process.env.OPENCODE_SESSION_ID || "unknown",
+          session_id: _ctx.sessionID || "unknown",
           epic_id: epicId,
           timestamp: new Date().toISOString(),
           event_type: "DECISION",
@@ -44365,7 +44722,7 @@ You may now complete the task with \`swarm_complete\`.`,
     const remaining = MAX_REVIEW_ATTEMPTS - attemptNumber;
     try {
       captureCoordinatorEvent({
-        session_id: process.env.OPENCODE_SESSION_ID || "unknown",
+        session_id: _ctx.sessionID || "unknown",
         epic_id: epicId,
         timestamp: new Date().toISOString(),
         event_type: "DECISION",
@@ -44436,6 +44793,7 @@ var reviewTools = {
 };
 
 // src/swarm-orchestrate.ts
+init_eval_capture();
 function generateWorkerHandoff(params) {
   const handoff = {
     contract: {
@@ -45336,10 +45694,29 @@ Files touched: ${args.files_touched?.join(", ") || "none recorded"}`,
           reason: "No files_owned contract found (non-epic subtask or decomposition event missing)"
         }
       };
+      try {
+        const { captureSubtaskOutcome: captureSubtaskOutcome2 } = await Promise.resolve().then(() => (init_eval_capture(), exports_eval_capture));
+        const durationMs2 = args.start_time ? Date.now() - args.start_time : 0;
+        const evalEpicId = cell.parent_id || epicId2;
+        captureSubtaskOutcome2({
+          epicId: evalEpicId,
+          projectPath: args.project_key,
+          beadId: args.bead_id,
+          title: cell.title,
+          plannedFiles: args.planned_files || [],
+          actualFiles: args.files_touched || [],
+          durationMs: durationMs2,
+          errorCount: args.error_count || 0,
+          retryCount: args.retry_count || 0,
+          success: true
+        });
+      } catch (error45) {
+        console.warn("[swarm_complete] Failed to capture subtask outcome:", error45);
+      }
       try {
         const durationMs2 = args.start_time ? Date.now() - args.start_time : 0;
         captureCoordinatorEvent({
-          session_id: process.env.OPENCODE_SESSION_ID || "unknown",
+          session_id: _ctx.sessionID || "unknown",
           epic_id: epicId2,
           timestamp: new Date().toISOString(),
           event_type: "OUTCOME",
@@ -45421,7 +45798,7 @@ ${errorStack.slice(0, 1000)}
       try {
         const durationMs = args.start_time ? Date.now() - args.start_time : 0;
         captureCoordinatorEvent({
-          session_id: process.env.OPENCODE_SESSION_ID || "unknown",
+          session_id: _ctx.sessionID || "unknown",
           epic_id: epicId,
           timestamp: new Date().toISOString(),
           event_type: "OUTCOME",
@@ -45488,7 +45865,9 @@ var swarm_record_outcome = tool({
       "user_cancelled",
       "unknown"
     ]).optional().describe("Failure classification (only when success=false). Auto-classified if not provided."),
-    failure_details: tool.schema.string().optional().describe("Detailed failure context (error message, stack trace, etc.)")
+    failure_details: tool.schema.string().optional().describe("Detailed failure context (error message, stack trace, etc.)"),
+    project_path: tool.schema.string().optional().describe("Project path (for finalizing eval records when all subtasks complete)"),
+    epic_id: tool.schema.string().optional().describe("Epic ID (for finalizing eval records when all subtasks complete)")
   },
   async execute(args) {
     const signals = {
@@ -45510,6 +45889,18 @@ var swarm_record_outcome = tool({
     const validated = OutcomeSignalsSchema.parse(signals);
     const scored = scoreImplicitFeedback(validated, DEFAULT_LEARNING_CONFIG);
     const errorStats = await globalErrorAccumulator.getErrorStats(args.bead_id);
+    let finalizedRecord = null;
+    if (args.project_path && args.epic_id) {
+      try {
+        const { finalizeEvalRecord: finalizeEvalRecord2 } = await Promise.resolve().then(() => (init_eval_capture(), exports_eval_capture));
+        finalizedRecord = finalizeEvalRecord2({
+          epicId: args.epic_id,
+          projectPath: args.project_path
+        });
+      } catch (error45) {
+        console.warn("[swarm_record_outcome] Failed to finalize eval record:", error45);
+      }
+    }
     const criteriaToScore = args.criteria ?? [
       "type_safe",
       "no_bugs",
@@ -45551,6 +45942,7 @@ var swarm_record_outcome = tool({
         accumulated_errors: errorStats.total,
         unresolved_errors: errorStats.unresolved
       },
+      finalized_eval_record: finalizedRecord || undefined,
       note: "Feedback events should be stored for criterion weight calculation. Use learning.ts functions to apply weights."
     }, null, 2);
   }
@@ -45582,12 +45974,31 @@ async function runResearchPhase(task, projectPath, options2) {
   if (techStack.length === 0) {
     return {
       tech_stack: [],
+      spawn_instructions: [],
       summaries: {},
       memory_ids: []
     };
   }
+  const spawnInstructions = [];
+  for (const tech of techStack) {
+    const researchId = `research-${tech}-${Date.now()}-${Math.random().toString(36).slice(2, 9)}`;
+    const prompt = formatResearcherPrompt({
+      research_id: researchId,
+      epic_id: "standalone-research",
+      tech_stack: [tech],
+      project_path: projectPath,
+      check_upgrades: options2?.checkUpgrades ?? false
+    });
+    spawnInstructions.push({
+      research_id: researchId,
+      tech,
+      prompt,
+      subagent_type: "swarm/researcher"
+    });
+  }
   return {
     tech_stack: techStack,
+    spawn_instructions: spawnInstructions,
     summaries: {},
     memory_ids: []
   };
@@ -46024,6 +46435,7 @@ var orchestrateTools = {
 };
 
 // src/swarm-prompts.ts
+init_eval_capture();
 var STRATEGY_DECOMPOSITION_PROMPT2 = `You are decomposing a task into parallelizable subtasks for a swarm of agents.
 
 ## Task
@@ -46820,7 +47232,7 @@ var swarm_spawn_subtask = tool({
     }).optional().describe("Recovery context from checkpoint compaction"),
     model: tool.schema.string().optional().describe("Optional explicit model override (auto-selected if not provided)")
   },
-  async execute(args) {
+  async execute(args, _ctx) {
     const prompt = formatSubtaskPromptV2({
       bead_id: args.bead_id,
       epic_id: args.epic_id,
@@ -46849,7 +47261,7 @@ var swarm_spawn_subtask = tool({
     const postCompletionInstructions = COORDINATOR_POST_WORKER_CHECKLIST.replace(/{project_key}/g, args.project_path || "$PWD").replace(/{epic_id}/g, args.epic_id).replace(/{task_id}/g, args.bead_id).replace(/{files_touched}/g, filesJoined).replace(/{worker_id}/g, "worker");
     try {
       captureCoordinatorEvent({
-        session_id: process.env.OPENCODE_SESSION_ID || "unknown",
+        session_id: _ctx.sessionID || "unknown",
         epic_id: args.epic_id,
         timestamp: new Date().toISOString(),
         event_type: "DECISION",
@@ -62712,6 +63124,7 @@ function guardrailOutput(toolName, output, config2 = DEFAULT_GUARDRAIL_CONFIG) {
 }
 
 // src/planning-guardrails.ts
+init_eval_capture();
 var FILE_MODIFICATION_PATTERNS = [
   /\bimplement\b/i,
   /\bcreate\b.*\.(ts|js|tsx|jsx|py|rs|go|java|rb|swift|kt)/i,
@@ -62997,9 +63410,21 @@ function getLog() {
   }
   return _logger;
 }
-var SWARM_COMPACTION_CONTEXT = `## \uD83D\uDC1D SWARM ACTIVE - You Are The COORDINATOR
+var SWARM_COMPACTION_CONTEXT = `
+┌─────────────────────────────────────────────────────────────┐
+│                                                             │
+│             \uD83D\uDC1D  YOU ARE THE COORDINATOR  \uD83D\uDC1D                 │
+│                                                             │
+│             NOT A WORKER. NOT AN IMPLEMENTER.               │
+│                  YOU ORCHESTRATE.                           │
+│                                                             │
+└─────────────────────────────────────────────────────────────┘
+
+## \uD83C\uDFAF NON-NEGOTIABLE: YOU ARE THE COORDINATOR
 
-Context was compacted but the swarm is still running. You are the **COORDINATOR**.
+Context was compacted but the swarm is still running. **YOU ARE THE COORDINATOR.**
+
+Your role is ORCHESTRATION, not implementation. When you catch yourself about to do work directly, STOP.
 
 ### ⛔ NEVER DO THESE (Coordinator Anti-Patterns)
 
@@ -63010,9 +63435,27 @@ Context was compacted but the swarm is still running. You are the **COORDINATOR*
 - ❌ **NEVER** implement features yourself - SPAWN A WORKER
 - ❌ **NEVER** "just do it myself to save time" - NO. SPAWN A WORKER.
 - ❌ **NEVER** reserve files with \`swarmmail_reserve\` - Workers reserve files
+- ❌ **NEVER** fetch files/docs directly - SPAWN A RESEARCHER
 
 **If you catch yourself about to edit a file, STOP. Use \`swarm_spawn_subtask\` instead.**
 
+### \uD83D\uDEAB FORBIDDEN TOOLS (Coordinators MUST delegate these)
+
+**NEVER use these tools directly. ALWAYS spawn a researcher worker via \`swarm_spawn_researcher\`:**
+
+**Repository fetching:**
+- \`repo-crawl_file\`, \`repo-crawl_readme\`, \`repo-crawl_search\`, \`repo-crawl_structure\`, \`repo-crawl_tree\`
+- \`repo-autopsy_*\` (all repo-autopsy tools)
+
+**Web/documentation fetching:**
+- \`webfetch\`, \`fetch_fetch\`
+- \`context7_resolve-library-id\`, \`context7_get-library-docs\`
+
+**Knowledge base:**
+- \`pdf-brain_search\`, \`pdf-brain_read\`
+
+**If you need external data:** Use \`swarm_spawn_researcher\` with a clear research task. The researcher will fetch, summarize, and return findings.
+
 ### ✅ ALWAYS DO THESE (Coordinator Checklist)
 
 On resume, execute this checklist IN ORDER:
@@ -63062,6 +63505,87 @@ Extract from session context:
 - **Review work** - Use \`swarm_review\` and \`swarm_review_feedback\` for completed work
 - **Close the loop** - When all subtasks done, verify and close the epic
 
+**You are the COORDINATOR. You orchestrate. You do NOT implement. Spawn workers.**
+
+---
+
+## \uD83D\uDCCB FULL COORDINATOR WORKFLOW (Reference)
+
+You are ALWAYS swarming. Here is the complete workflow for any new work:
+
+### Phase 1.5: Research Phase (FOR COMPLEX TASKS)
+
+**If the task requires understanding unfamiliar technologies, spawn a researcher FIRST:**
+
+\`\`\`
+swarm_spawn_researcher(
+  research_id="research-<topic>",
+  epic_id="<epic-id>",
+  tech_stack=["<technology>"],
+  project_path="<path>"
+)
+// Then spawn with Task(subagent_type="swarm/researcher", prompt="<from above>")
+\`\`\`
+
+### Phase 2: Knowledge Gathering
+
+\`\`\`
+semantic-memory_find(query="<task keywords>", limit=5)   # Past learnings
+cass_search(query="<task description>", limit=5)         # Similar past tasks  
+skills_list()                                            # Available skills
+\`\`\`
+
+### Phase 3: Decompose
+
+\`\`\`
+swarm_select_strategy(task="<task>")
+swarm_plan_prompt(task="<task>", context="<synthesized knowledge>")
+swarm_validate_decomposition(response="<CellTree JSON>")
+\`\`\`
+
+### Phase 4: Create Cells
+
+\`hive_create_epic(epic_title="<task>", subtasks=[...])\`
+
+### Phase 5: DO NOT Reserve Files
+
+> **⚠️ Coordinator NEVER reserves files.** Workers reserve their own files.
+
+### Phase 6: Spawn Workers
+
+\`\`\`
+swarm_spawn_subtask(bead_id, epic_id, title, files, shared_context, project_path)
+Task(subagent_type="swarm/worker", prompt="<from above>")
+\`\`\`
+
+### Phase 7: MANDATORY Review Loop
+
+**AFTER EVERY Task() RETURNS:**
+
+1. \`swarmmail_inbox()\` - Check for messages
+2. \`swarm_review(project_key, epic_id, task_id, files_touched)\` - Generate review
+3. Evaluate against epic goals
+4. \`swarm_review_feedback(project_key, task_id, worker_id, status, issues)\`
+
+**If needs_changes:**
+\`\`\`
+swarm_spawn_retry(bead_id, epic_id, original_prompt, attempt, issues, diff, files, project_path)
+// Spawn NEW worker with Task() using retry prompt
+// Max 3 attempts before marking task blocked
+\`\`\`
+
+### Phase 8: Complete
+
+\`hive_sync()\` - Sync all cells to git
+
+## Strategy Reference
+
+| Strategy       | Best For                 | Keywords                               |
+| -------------- | ------------------------ | -------------------------------------- |
+| file-based     | Refactoring, migrations  | refactor, migrate, rename, update all  |
+| feature-based  | New features             | add, implement, build, create, feature |
+| risk-based     | Bug fixes, security      | fix, bug, security, critical, urgent   |
+
 **You are the COORDINATOR. You orchestrate. You do NOT implement. Spawn workers.**
 `;
 var SWARM_DETECTION_FALLBACK = `## \uD83D\uDC1D Swarm Detection - Check Your Context
@@ -63736,7 +64260,7 @@ var SwarmPlugin = async (input) => {
       if (isInCoordinatorContext()) {
         const ctx = getCoordinatorContext();
         const violation = detectCoordinatorViolation({
-          sessionId: ctx.sessionId || "unknown",
+          sessionId: input2.sessionID || "unknown",
           epicId: ctx.epicId || "unknown",
           toolName,
           toolArgs: output.args,