opencode-swarm-plugin 0.37.0 → 0.39.1

package/evals/scorers/coordinator-discipline.evalite-test.ts

@@ -12,7 +12,7 @@ import {
 } from "./coordinator-discipline.js";
 
 describe("violationCount", () => {
-	it("scores 1.0 for zero violations", () => {
+	it("scores 1.0 for zero violations", async () => {
 		const session: CoordinatorSession = {
 			session_id: "test-session",
 			epic_id: "test-epic",
@@ -30,9 +30,10 @@ describe("violationCount", () => {
 			],
 		};
 
-		const result = violationCount.scorer({
+		const result = await violationCount({
 			output: JSON.stringify(session),
 			expected: {},
+			input: undefined,
 		});
 
 		expect(result.score).toBe(1.0);

package/evals/scorers/coordinator-discipline.ts

@@ -270,14 +270,14 @@ export const timeToFirstSpawn = createScorer({
 export const overallDiscipline = createScorer({
   name: "Overall Coordinator Discipline",
   description: "Composite score for coordinator protocol adherence",
-  scorer: ({ output, expected }) => {
+  scorer: async ({ output, expected, input }) => {
     try {
       // Run all scorers
       const scores = {
-        violations: violationCount.scorer({ output, expected }),
-        spawn: spawnEfficiency.scorer({ output, expected }),
-        review: reviewThoroughness.scorer({ output, expected }),
-        speed: timeToFirstSpawn.scorer({ output, expected }),
+        violations: await violationCount({ output, expected, input }),
+        spawn: await spawnEfficiency({ output, expected, input }),
+        review: await reviewThoroughness({ output, expected, input }),
+        speed: await timeToFirstSpawn({ output, expected, input }),
       };
 
       // Weighted average
@@ -289,16 +289,16 @@ export const overallDiscipline = createScorer({
       };
 
       const totalScore =
-        scores.violations.score * weights.violations +
-        scores.spawn.score * weights.spawn +
-        scores.review.score * weights.review +
-        scores.speed.score * weights.speed;
+        (scores.violations.score ?? 0) * weights.violations +
+        (scores.spawn.score ?? 0) * weights.spawn +
+        (scores.review.score ?? 0) * weights.review +
+        (scores.speed.score ?? 0) * weights.speed;
 
       const details = [
-        `Violations: ${(scores.violations.score * 100).toFixed(0)}%`,
-        `Spawn: ${(scores.spawn.score * 100).toFixed(0)}%`,
-        `Review: ${(scores.review.score * 100).toFixed(0)}%`,
-        `Speed: ${(scores.speed.score * 100).toFixed(0)}%`,
+        `Violations: ${((scores.violations.score ?? 0) * 100).toFixed(0)}%`,
+        `Spawn: ${((scores.spawn.score ?? 0) * 100).toFixed(0)}%`,
+        `Review: ${((scores.review.score ?? 0) * 100).toFixed(0)}%`,
+        `Speed: ${((scores.speed.score ?? 0) * 100).toFixed(0)}%`,
       ].join(", ");
 
       return {

package/examples/plugin-wrapper-template.ts

@@ -65,6 +65,42 @@ function logCompaction(
   }
 }
 
+/**
+ * Capture compaction event for evals (non-fatal dynamic import)
+ * 
+ * Uses dynamic import to avoid circular dependencies and keep the plugin wrapper
+ * self-contained. Captures COMPACTION events to session JSONL for eval analysis.
+ * 
+ * @param sessionID - Session ID
+ * @param epicID - Epic ID (or "unknown" if not detected)
+ * @param compactionType - Event type (detection_complete, prompt_generated, context_injected)
+ * @param payload - Event-specific data (full prompts, detection results, etc.)
+ */
+async function captureCompaction(
+  sessionID: string,
+  epicID: string,
+  compactionType: "detection_complete" | "prompt_generated" | "context_injected",
+  payload: any,
+): Promise<void> {
+  try {
+    // Dynamic import to avoid circular deps (plugin wrapper → src → plugin wrapper)
+    const { captureCompactionEvent } = await import("../src/eval-capture");
+    captureCompactionEvent({
+      session_id: sessionID,
+      epic_id: epicID,
+      compaction_type: compactionType,
+      payload,
+    });
+  } catch (err) {
+    // Non-fatal - capture failures shouldn't break compaction
+    logCompaction("warn", "compaction_capture_failed", {
+      session_id: sessionID,
+      compaction_type: compactionType,
+      error: err instanceof Error ? err.message : String(err),
+    });
+  }
+}
+
 // Module-level project directory - set during plugin initialization
 // This is CRITICAL: without it, the CLI uses process.cwd() which may be wrong
 let projectDirectory: string = process.cwd();
@@ -305,6 +341,34 @@ const hive_sync = tool({
   execute: (args, ctx) => execTool("hive_sync", args, ctx),
 });
 
+const 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)"),
+  },
+  execute: (args, ctx) => execTool("hive_cells", args, ctx),
+});
+
 const beads_link_thread = tool({
   description: "Add metadata linking bead to Agent Mail thread",
   args: {
@@ -1202,9 +1266,18 @@ ${JSON.stringify(snapshot, null, 2)}
 
 Generate a prompt following this structure:
 
+┌─────────────────────────────────────────────────────────────┐
+│                                                             │
+│             🐝  YOU ARE THE COORDINATOR  🐝                 │
+│                                                             │
+│             NOT A WORKER. NOT AN IMPLEMENTER.               │
+│                  YOU ORCHESTRATE.                           │
+│                                                             │
+└─────────────────────────────────────────────────────────────┘
+
 # 🐝 Swarm Continuation - [Epic Title or "Unknown"]
 
-You are resuming coordination of an active swarm that was interrupted by context compaction.
+**NON-NEGOTIABLE: YOU ARE THE COORDINATOR.** You resumed after context compaction.
 
 ## Epic State
 
@@ -1231,15 +1304,29 @@ You are resuming coordination of an active swarm that was interrupted by context
 
 [List 3-5 concrete actions with actual commands, using real IDs from the state]
 
-## Coordinator Reminders
+## 🎯 COORDINATOR MANDATES (NON-NEGOTIABLE)
+
+**YOU ARE THE COORDINATOR. NOT A WORKER.**
+
+### ⛔ FORBIDDEN - NEVER do these:
+- ❌ NEVER use \`edit\`, \`write\`, or \`bash\` for implementation - SPAWN A WORKER
+- ❌ NEVER fetch directly with \`repo-crawl_*\`, \`repo-autopsy_*\`, \`webfetch\`, \`fetch_fetch\` - SPAWN A RESEARCHER
+- ❌ NEVER use \`context7_*\` or \`pdf-brain_*\` directly - SPAWN A RESEARCHER
+- ❌ NEVER reserve files - Workers reserve files
+
+### ✅ ALWAYS do these:
+- ✅ ALWAYS check \`swarm_status\` and \`swarmmail_inbox\` first
+- ✅ ALWAYS use \`swarm_spawn_subtask\` for implementation work
+- ✅ ALWAYS use \`swarm_spawn_researcher\` for external data fetching
+- ✅ ALWAYS review worker output with \`swarm_review\` → \`swarm_review_feedback\`
+- ✅ ALWAYS monitor actively - Check messages every ~10 minutes
+- ✅ ALWAYS unblock aggressively - Resolve dependencies immediately
 
-- **You are the coordinator** - Don't wait for instructions, orchestrate
-- **Monitor actively** - Check messages every ~10 minutes
-- **Unblock aggressively** - Resolve dependencies immediately
-- **Review thoroughly** - 3-strike rule enforced
-- **Ship it** - When all subtasks done, close the epic
+**If you need external data:** Use \`swarm_spawn_researcher\` with a clear research task. The researcher will fetch, summarize, and return findings.
 
-Keep the prompt concise but actionable. Use actual data from the snapshot, not placeholders.`;
+**3-strike rule enforced:** Workers get 3 review attempts. After 3 rejections, escalate to human.
+
+Keep the prompt concise but actionable. Use actual data from the snapshot, not placeholders. Include the ASCII header and ALL coordinator mandates.`;
 
     logCompaction("debug", "generate_compaction_prompt_calling_llm", {
       session_id: snapshot.sessionID,
@@ -1896,6 +1983,7 @@ const SwarmPlugin: Plugin = async (
       hive_close,
       hive_start,
       hive_ready,
+      hive_cells,
       hive_sync,
       beads_link_thread,
       // Swarm Mail (Embedded)
@@ -2100,6 +2188,25 @@ const SwarmPlugin: Plugin = async (
             full_snapshot: snapshot, // Log the entire snapshot
           });
 
+          // =======================================================================
+          // CAPTURE POINT 1: Detection complete - record confidence and reasons
+          // =======================================================================
+          await captureCompaction(
+            input.sessionID,
+            snapshot.epic?.id || "unknown",
+            "detection_complete",
+            {
+              confidence: snapshot.detection.confidence,
+              detected: detection.detected,
+              reasons: snapshot.detection.reasons,
+              session_scan_contributed: sessionScan.swarmDetected,
+              session_scan_reasons: sessionScan.reasons,
+              epic_id: snapshot.epic?.id,
+              epic_title: snapshot.epic?.title,
+              subtask_count: snapshot.epic?.subtasks?.length ?? 0,
+            },
+          );
+
           // Level 2: Generate prompt with LLM
           const llmStart = Date.now();
           const llmPrompt = await generateCompactionPrompt(snapshot);
@@ -2113,6 +2220,23 @@ const SwarmPlugin: Plugin = async (
             prompt_preview: llmPrompt?.substring(0, 500),
           });
 
+          // =======================================================================
+          // CAPTURE POINT 2: Prompt generated - record FULL prompt content
+          // =======================================================================
+          if (llmPrompt) {
+            await captureCompaction(
+              input.sessionID,
+              snapshot.epic?.id || "unknown",
+              "prompt_generated",
+              {
+                prompt_length: llmPrompt.length,
+                full_prompt: llmPrompt, // FULL content, not truncated
+                context_type: "llm_generated",
+                duration_ms: llmDuration,
+              },
+            );
+          }
+
           if (llmPrompt) {
             // SUCCESS: Use LLM-generated prompt
             const header = `[Swarm compaction: LLM-generated, ${detection.reasons.join(", ")}]\n\n`;
@@ -2136,6 +2260,21 @@ const SwarmPlugin: Plugin = async (
               });
             }
 
+            // =======================================================================
+            // CAPTURE POINT 3a: Context injected (LLM path) - record FULL content
+            // =======================================================================
+            await captureCompaction(
+              input.sessionID,
+              snapshot.epic?.id || "unknown",
+              "context_injected",
+              {
+                full_content: fullContent, // FULL content, not truncated
+                content_length: fullContent.length,
+                injection_method: "prompt" in output ? "output.prompt" : "output.context.push",
+                context_type: "llm_generated",
+              },
+            );
+
             const totalDuration = Date.now() - startTime;
             logCompaction("info", "compaction_complete_llm_success", {
               session_id: input.sessionID,
@@ -2171,6 +2310,21 @@ const SwarmPlugin: Plugin = async (
         const staticContent = header + SWARM_COMPACTION_CONTEXT;
         output.context.push(staticContent);
 
+        // =======================================================================
+        // CAPTURE POINT 3b: Context injected (static fallback) - record FULL content
+        // =======================================================================
+        await captureCompaction(
+          input.sessionID,
+          "unknown", // No snapshot available in this path
+          "context_injected",
+          {
+            full_content: staticContent,
+            content_length: staticContent.length,
+            injection_method: "output.context.push",
+            context_type: "static_swarm_context",
+          },
+        );
+
         const totalDuration = Date.now() - startTime;
         logCompaction("info", "compaction_complete_static_fallback", {
           session_id: input.sessionID,
@@ -2186,6 +2340,21 @@ const SwarmPlugin: Plugin = async (
         const fallbackContent = header + SWARM_DETECTION_FALLBACK;
         output.context.push(fallbackContent);
 
+        // =======================================================================
+        // CAPTURE POINT 3c: Context injected (detection fallback) - record FULL content
+        // =======================================================================
+        await captureCompaction(
+          input.sessionID,
+          "unknown", // No snapshot for low confidence
+          "context_injected",
+          {
+            full_content: fallbackContent,
+            content_length: fallbackContent.length,
+            injection_method: "output.context.push",
+            context_type: "detection_fallback",
+          },
+        );
+
         const totalDuration = Date.now() - startTime;
         logCompaction("info", "compaction_complete_detection_fallback", {
           session_id: input.sessionID,

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "opencode-swarm-plugin",
-  "version": "0.37.0",
+  "version": "0.39.1",
   "description": "Multi-agent swarm coordination for OpenCode with learning capabilities, beads integration, and Agent Mail",
   "type": "module",
   "main": "./dist/index.js",
@@ -30,6 +30,11 @@
     "test:all": "bun test --timeout 60000 src/",
     "test:watch": "bun test --watch src/",
     "typecheck": "tsc --noEmit",
+    "eval:run": "bun --env-file=.env run bunx evalite run evals/",
+    "eval:decomposition": "bun --env-file=.env run bunx evalite run evals/swarm-decomposition.eval.ts",
+    "eval:coordinator": "bun --env-file=.env run bunx evalite run evals/coordinator-session.eval.ts",
+    "eval:compaction": "bun --env-file=.env run bunx evalite run evals/compaction-prompt.eval.ts",
+    "migrate:sessions": "bun run scripts/migrate-unknown-sessions.ts",
     "postinstall": "node -e \"console.log('\\n\\x1b[33m  Run \\x1b[36mswarm setup\\x1b[33m to configure OpenCode integration\\x1b[0m\\n')\""
   },
   "dependencies": {
@@ -41,7 +46,7 @@
     "minimatch": "^10.1.1",
     "pino": "^9.6.0",
     "pino-roll": "^1.3.0",
-    "swarm-mail": "1.5.1",
+    "swarm-mail": "1.5.2",
     "yaml": "^2.8.2",
     "zod": "4.1.8"
   },