opencode-swarm-plugin 0.18.0 → 0.20.0

This diff represents the content of publicly available package versions that have been released to one of the supported registries. The information contained in this diff is provided for informational purposes only and reflects changes between package versions as they appear in their respective public registries.
package/src/swarm.ts CHANGED
@@ -1,3888 +1,35 @@
1
1
  /**
2
2
  * Swarm Module - High-level swarm coordination
3
3
  *
4
- * Orchestrates beads, Agent Mail, and structured validation for parallel task execution.
5
- * The actual agent spawning happens via OpenCode's Task tool - this module provides
6
- * the primitives and prompts that /swarm command uses.
4
+ * This module re-exports from focused submodules for backward compatibility.
5
+ * For new code, prefer importing from specific modules:
6
+ * - swarm-strategies.ts - Strategy selection
7
+ * - swarm-decompose.ts - Task decomposition
8
+ * - swarm-prompts.ts - Prompt templates
9
+ * - swarm-orchestrate.ts - Status and completion
7
10
  *
8
- * Key responsibilities:
9
- * - Task decomposition into bead trees with file assignments
10
- * - Swarm status tracking via beads + Agent Mail
11
- * - Progress reporting and completion handling
12
- * - Prompt templates for decomposition, subtasks, and evaluation
11
+ * @module swarm
13
12
  */
14
- import { tool } from "@opencode-ai/plugin";
15
- import { z } from "zod";
16
- import {
17
- BeadTreeSchema,
18
- SwarmStatusSchema,
19
- AgentProgressSchema,
20
- EvaluationSchema,
21
- BeadSchema,
22
- type SwarmStatus,
23
- type AgentProgress,
24
- type Evaluation,
25
- type SpawnedAgent,
26
- type Bead,
27
- } from "./schemas";
28
- import {
29
- sendSwarmMessage,
30
- getSwarmInbox,
31
- readSwarmMessage,
32
- releaseSwarmFiles,
33
- } from "./streams/swarm-mail";
34
- import {
35
- OutcomeSignalsSchema,
36
- DecompositionStrategySchema,
37
- scoreImplicitFeedback,
38
- outcomeToFeedback,
39
- ErrorAccumulator,
40
- ErrorEntrySchema,
41
- formatMemoryStoreOnSuccess,
42
- formatMemoryStoreOn3Strike,
43
- formatMemoryQueryForDecomposition,
44
- formatMemoryValidationHint,
45
- type OutcomeSignals,
46
- type ScoredOutcome,
47
- type FeedbackEvent,
48
- type ErrorEntry,
49
- type ErrorType,
50
- type DecompositionStrategy as LearningDecompositionStrategy,
51
- DEFAULT_LEARNING_CONFIG,
52
- } from "./learning";
53
- import {
54
- isToolAvailable,
55
- warnMissingTool,
56
- checkAllTools,
57
- formatToolAvailability,
58
- type ToolName,
59
- } from "./tool-availability";
60
- import {
61
- getSkillsContextForSwarm,
62
- findRelevantSkills,
63
- listSkills,
64
- } from "./skills";
65
13
 
66
- // ============================================================================
67
- // Conflict Detection
68
- // ============================================================================
14
+ // Re-export everything for backward compatibility
15
+ export * from "./swarm-strategies";
16
+ export * from "./swarm-decompose";
17
+ export * from "./swarm-prompts";
18
+ export * from "./swarm-orchestrate";
69
19
 
70
- /**
71
- * Marker words that indicate positive directives
72
- */
73
- const POSITIVE_MARKERS = [
74
- "always",
75
- "must",
76
- "required",
77
- "ensure",
78
- "use",
79
- "prefer",
80
- ];
81
-
82
- /**
83
- * Marker words that indicate negative directives
84
- */
85
- const NEGATIVE_MARKERS = [
86
- "never",
87
- "dont",
88
- "don't",
89
- "avoid",
90
- "forbid",
91
- "no ",
92
- "not ",
93
- ];
94
-
95
- /**
96
- * A detected conflict between subtask instructions
97
- */
98
- export interface InstructionConflict {
99
- subtask_a: number;
100
- subtask_b: number;
101
- directive_a: string;
102
- directive_b: string;
103
- conflict_type: "positive_negative" | "contradictory";
104
- description: string;
105
- }
106
-
107
- /**
108
- * Extract directives from text based on marker words
109
- */
110
- function extractDirectives(text: string): {
111
- positive: string[];
112
- negative: string[];
113
- } {
114
- const sentences = text.split(/[.!?\n]+/).map((s) => s.trim().toLowerCase());
115
- const positive: string[] = [];
116
- const negative: string[] = [];
117
-
118
- for (const sentence of sentences) {
119
- if (!sentence) continue;
120
-
121
- const hasPositive = POSITIVE_MARKERS.some((m) => sentence.includes(m));
122
- const hasNegative = NEGATIVE_MARKERS.some((m) => sentence.includes(m));
123
-
124
- if (hasPositive && !hasNegative) {
125
- positive.push(sentence);
126
- } else if (hasNegative) {
127
- negative.push(sentence);
128
- }
129
- }
130
-
131
- return { positive, negative };
132
- }
133
-
134
- /**
135
- * Check if two directives conflict
136
- *
137
- * Simple heuristic: look for common subjects with opposite polarity
138
- */
139
- function directivesConflict(positive: string, negative: string): boolean {
140
- // Extract key nouns/concepts (simple word overlap check)
141
- const positiveWords = new Set(
142
- positive.split(/\s+/).filter((w) => w.length > 3),
143
- );
144
- const negativeWords = negative.split(/\s+/).filter((w) => w.length > 3);
145
-
146
- // If they share significant words, they might conflict
147
- const overlap = negativeWords.filter((w) => positiveWords.has(w));
148
- return overlap.length >= 2;
149
- }
150
-
151
- /**
152
- * Detect conflicts between subtask instructions
153
- *
154
- * Looks for cases where one subtask says "always use X" and another says "avoid X".
155
- *
156
- * @param subtasks - Array of subtask descriptions
157
- * @returns Array of detected conflicts
158
- *
159
- * @see https://github.com/Dicklesworthstone/cass_memory_system/blob/main/src/curate.ts#L36-L89
160
- */
161
- export function detectInstructionConflicts(
162
- subtasks: Array<{ title: string; description?: string }>,
163
- ): InstructionConflict[] {
164
- const conflicts: InstructionConflict[] = [];
165
-
166
- // Extract directives from each subtask
167
- const subtaskDirectives = subtasks.map((s, i) => ({
168
- index: i,
169
- title: s.title,
170
- ...extractDirectives(`${s.title} ${s.description || ""}`),
171
- }));
172
-
173
- // Compare each pair of subtasks
174
- for (let i = 0; i < subtaskDirectives.length; i++) {
175
- for (let j = i + 1; j < subtaskDirectives.length; j++) {
176
- const a = subtaskDirectives[i];
177
- const b = subtaskDirectives[j];
178
-
179
- // Check if A's positive conflicts with B's negative
180
- for (const posA of a.positive) {
181
- for (const negB of b.negative) {
182
- if (directivesConflict(posA, negB)) {
183
- conflicts.push({
184
- subtask_a: i,
185
- subtask_b: j,
186
- directive_a: posA,
187
- directive_b: negB,
188
- conflict_type: "positive_negative",
189
- description: `Subtask ${i} says "${posA}" but subtask ${j} says "${negB}"`,
190
- });
191
- }
192
- }
193
- }
194
-
195
- // Check if B's positive conflicts with A's negative
196
- for (const posB of b.positive) {
197
- for (const negA of a.negative) {
198
- if (directivesConflict(posB, negA)) {
199
- conflicts.push({
200
- subtask_a: j,
201
- subtask_b: i,
202
- directive_a: posB,
203
- directive_b: negA,
204
- conflict_type: "positive_negative",
205
- description: `Subtask ${j} says "${posB}" but subtask ${i} says "${negA}"`,
206
- });
207
- }
208
- }
209
- }
210
- }
211
- }
212
-
213
- return conflicts;
214
- }
215
-
216
- // ============================================================================
217
- // Strategy Definitions
218
- // ============================================================================
219
-
220
- /**
221
- * Decomposition strategy types
222
- */
223
- export type DecompositionStrategy =
224
- | "file-based"
225
- | "feature-based"
226
- | "risk-based"
227
- | "research-based"
228
- | "auto";
229
-
230
- /**
231
- * Strategy definition with keywords, guidelines, and anti-patterns
232
- */
233
- export interface StrategyDefinition {
234
- name: DecompositionStrategy;
235
- description: string;
236
- keywords: string[];
237
- guidelines: string[];
238
- antiPatterns: string[];
239
- examples: string[];
240
- }
241
-
242
- /**
243
- * Strategy definitions for task decomposition
244
- */
245
- export const STRATEGIES: Record<
246
- Exclude<DecompositionStrategy, "auto">,
247
- StrategyDefinition
248
- > = {
249
- "file-based": {
250
- name: "file-based",
251
- description:
252
- "Group by file type or directory. Best for refactoring, migrations, and pattern changes across codebase.",
253
- keywords: [
254
- "refactor",
255
- "migrate",
256
- "update all",
257
- "rename",
258
- "replace",
259
- "convert",
260
- "upgrade",
261
- "deprecate",
262
- "remove",
263
- "cleanup",
264
- "lint",
265
- "format",
266
- ],
267
- guidelines: [
268
- "Group files by directory or type (e.g., all components, all tests)",
269
- "Minimize cross-directory dependencies within a subtask",
270
- "Handle shared types/utilities first if they change",
271
- "Each subtask should be a complete transformation of its file set",
272
- "Consider import/export relationships when grouping",
273
- ],
274
- antiPatterns: [
275
- "Don't split tightly coupled files across subtasks",
276
- "Don't group files that have no relationship",
277
- "Don't forget to update imports when moving/renaming",
278
- ],
279
- examples: [
280
- "Migrate all components to new API → split by component directory",
281
- "Rename userId to accountId → split by module (types first, then consumers)",
282
- "Update all tests to use new matcher → split by test directory",
283
- ],
284
- },
285
- "feature-based": {
286
- name: "feature-based",
287
- description:
288
- "Vertical slices with UI + API + data. Best for new features and adding functionality.",
289
- keywords: [
290
- "add",
291
- "implement",
292
- "build",
293
- "create",
294
- "feature",
295
- "new",
296
- "integrate",
297
- "connect",
298
- "enable",
299
- "support",
300
- ],
301
- guidelines: [
302
- "Each subtask is a complete vertical slice (UI + logic + data)",
303
- "Start with data layer/types, then logic, then UI",
304
- "Keep related components together (form + validation + submission)",
305
- "Separate concerns that can be developed independently",
306
- "Consider user-facing features as natural boundaries",
307
- ],
308
- antiPatterns: [
309
- "Don't split a single feature across multiple subtasks",
310
- "Don't create subtasks that can't be tested independently",
311
- "Don't forget integration points between features",
312
- ],
313
- examples: [
314
- "Add user auth → [OAuth setup, Session management, Protected routes]",
315
- "Build dashboard → [Data fetching, Chart components, Layout/navigation]",
316
- "Add search → [Search API, Search UI, Results display]",
317
- ],
318
- },
319
- "risk-based": {
320
- name: "risk-based",
321
- description:
322
- "Isolate high-risk changes, add tests first. Best for bug fixes, security issues, and critical changes.",
323
- keywords: [
324
- "fix",
325
- "bug",
326
- "security",
327
- "vulnerability",
328
- "critical",
329
- "urgent",
330
- "hotfix",
331
- "patch",
332
- "audit",
333
- "review",
334
- ],
335
- guidelines: [
336
- "Write tests FIRST to capture expected behavior",
337
- "Isolate the risky change to minimize blast radius",
338
- "Add monitoring/logging around the change",
339
- "Create rollback plan as part of the task",
340
- "Audit similar code for the same issue",
341
- ],
342
- antiPatterns: [
343
- "Don't make multiple risky changes in one subtask",
344
- "Don't skip tests for 'simple' fixes",
345
- "Don't forget to check for similar issues elsewhere",
346
- ],
347
- examples: [
348
- "Fix auth bypass → [Add regression test, Fix vulnerability, Audit similar endpoints]",
349
- "Fix race condition → [Add test reproducing issue, Implement fix, Add concurrency tests]",
350
- "Security audit → [Scan for vulnerabilities, Fix critical issues, Document remaining risks]",
351
- ],
352
- },
353
- "research-based": {
354
- name: "research-based",
355
- description:
356
- "Parallel search across multiple sources, then synthesize. Best for investigation, learning, and discovery tasks.",
357
- keywords: [
358
- "research",
359
- "investigate",
360
- "explore",
361
- "find out",
362
- "discover",
363
- "understand",
364
- "learn about",
365
- "analyze",
366
- "what is",
367
- "what are",
368
- "how does",
369
- "how do",
370
- "why does",
371
- "why do",
372
- "compare",
373
- "evaluate",
374
- "study",
375
- "look up",
376
- "look into",
377
- "search for",
378
- "dig into",
379
- "figure out",
380
- "debug options",
381
- "debug levers",
382
- "configuration options",
383
- "environment variables",
384
- "available options",
385
- "documentation",
386
- ],
387
- guidelines: [
388
- "Split by information source (PDFs, repos, history, web)",
389
- "Each agent searches with different query angles",
390
- "Include a synthesis subtask that depends on all search subtasks",
391
- "Use pdf-brain for documentation/books if available",
392
- "Use repo-crawl for GitHub repos if URL provided",
393
- "Use cass for past agent session history",
394
- "Assign NO files to research subtasks (read-only)",
395
- ],
396
- antiPatterns: [
397
- "Don't have one agent search everything sequentially",
398
- "Don't skip synthesis - raw search results need consolidation",
399
- "Don't forget to check tool availability before assigning sources",
400
- ],
401
- examples: [
402
- "Research auth patterns → [Search PDFs, Search repos, Search history, Synthesize]",
403
- "Investigate error → [Search cass for similar errors, Search repo for error handling, Synthesize]",
404
- "Learn about library → [Search docs, Search examples, Search issues, Synthesize findings]",
405
- ],
406
- },
407
- };
408
-
409
- /**
410
- * Analyze task description and select best decomposition strategy
411
- *
412
- * @param task - Task description
413
- * @returns Selected strategy with reasoning
414
- */
415
- export function selectStrategy(task: string): {
416
- strategy: Exclude<DecompositionStrategy, "auto">;
417
- confidence: number;
418
- reasoning: string;
419
- alternatives: Array<{
420
- strategy: Exclude<DecompositionStrategy, "auto">;
421
- score: number;
422
- }>;
423
- } {
424
- const taskLower = task.toLowerCase();
425
-
426
- // Score each strategy based on keyword matches
427
- const scores: Record<Exclude<DecompositionStrategy, "auto">, number> = {
428
- "file-based": 0,
429
- "feature-based": 0,
430
- "risk-based": 0,
431
- "research-based": 0,
432
- };
433
-
434
- for (const [strategyName, definition] of Object.entries(STRATEGIES)) {
435
- const name = strategyName as Exclude<DecompositionStrategy, "auto">;
436
- for (const keyword of definition.keywords) {
437
- // Use word boundary matching to avoid "debug" matching "bug"
438
- // For multi-word keywords, just check includes (they're specific enough)
439
- if (keyword.includes(" ")) {
440
- if (taskLower.includes(keyword)) {
441
- scores[name] += 1;
442
- }
443
- } else {
444
- // Single word: use word boundary regex
445
- const regex = new RegExp(`\\b${keyword}\\b`, "i");
446
- if (regex.test(taskLower)) {
447
- scores[name] += 1;
448
- }
449
- }
450
- }
451
- }
452
-
453
- // Find the winner
454
- const entries = Object.entries(scores) as Array<
455
- [Exclude<DecompositionStrategy, "auto">, number]
456
- >;
457
- entries.sort((a, b) => b[1] - a[1]);
458
-
459
- const [winner, winnerScore] = entries[0];
460
- const [runnerUp, runnerUpScore] = entries[1] || [null, 0];
461
-
462
- // Calculate confidence based on margin
463
- const totalScore = entries.reduce((sum, [, score]) => sum + score, 0);
464
- const confidence =
465
- totalScore > 0
466
- ? Math.min(0.95, 0.5 + (winnerScore - runnerUpScore) / totalScore)
467
- : 0.5; // Default to 50% if no keywords matched
468
-
469
- // Build reasoning
470
- let reasoning: string;
471
- if (winnerScore === 0) {
472
- reasoning = `No strong keyword signals. Defaulting to feature-based as it's most versatile.`;
473
- } else {
474
- const matchedKeywords = STRATEGIES[winner].keywords.filter((k) =>
475
- taskLower.includes(k),
476
- );
477
- reasoning = `Matched keywords: ${matchedKeywords.join(", ")}. ${STRATEGIES[winner].description}`;
478
- }
479
-
480
- // If no keywords matched, default to feature-based
481
- const finalStrategy = winnerScore === 0 ? "feature-based" : winner;
482
-
483
- return {
484
- strategy: finalStrategy,
485
- confidence,
486
- reasoning,
487
- alternatives: entries
488
- .filter(([s]) => s !== finalStrategy)
489
- .map(([strategy, score]) => ({ strategy, score })),
490
- };
491
- }
492
-
493
- /**
494
- * Format strategy-specific guidelines for the decomposition prompt
495
- */
496
- export function formatStrategyGuidelines(
497
- strategy: Exclude<DecompositionStrategy, "auto">,
498
- ): string {
499
- const def = STRATEGIES[strategy];
500
-
501
- const guidelines = def.guidelines.map((g) => `- ${g}`).join("\n");
502
- const antiPatterns = def.antiPatterns.map((a) => `- ${a}`).join("\n");
503
- const examples = def.examples.map((e) => `- ${e}`).join("\n");
504
-
505
- return `## Strategy: ${strategy}
506
-
507
- ${def.description}
508
-
509
- ### Guidelines
510
- ${guidelines}
511
-
512
- ### Anti-Patterns (Avoid These)
513
- ${antiPatterns}
514
-
515
- ### Examples
516
- ${examples}`;
517
- }
518
-
519
- // ============================================================================
520
- // Prompt Templates
521
- // ============================================================================
522
-
523
- /**
524
- * Prompt for decomposing a task into parallelizable subtasks.
525
- *
526
- * Used by swarm:decompose to instruct the agent on how to break down work.
527
- * The agent responds with a BeadTree that gets validated.
528
- */
529
- export const DECOMPOSITION_PROMPT = `You are decomposing a task into parallelizable subtasks for a swarm of agents.
530
-
531
- ## Task
532
- {task}
533
-
534
- {context_section}
535
-
536
- ## MANDATORY: Beads Issue Tracking
537
-
538
- **Every subtask MUST become a bead.** This is non-negotiable.
539
-
540
- After decomposition, the coordinator will:
541
- 1. Create an epic bead for the overall task
542
- 2. Create child beads for each subtask
543
- 3. Track progress through bead status updates
544
- 4. Close beads with summaries when complete
545
-
546
- Agents MUST update their bead status as they work. No silent progress.
547
-
548
- ## Requirements
549
-
550
- 1. **Break into 2-{max_subtasks} independent subtasks** that can run in parallel
551
- 2. **Assign files** - each subtask must specify which files it will modify
552
- 3. **No file overlap** - files cannot appear in multiple subtasks (they get exclusive locks)
553
- 4. **Order by dependency** - if subtask B needs subtask A's output, A must come first in the array
554
- 5. **Estimate complexity** - 1 (trivial) to 5 (complex)
555
- 6. **Plan aggressively** - break down more than you think necessary, smaller is better
556
-
557
- ## Response Format
558
-
559
- Respond with a JSON object matching this schema:
560
-
561
- \`\`\`typescript
562
- {
563
- epic: {
564
- title: string, // Epic title for the beads tracker
565
- description?: string // Brief description of the overall goal
566
- },
567
- subtasks: [
568
- {
569
- title: string, // What this subtask accomplishes
570
- description?: string, // Detailed instructions for the agent
571
- files: string[], // Files this subtask will modify (globs allowed)
572
- dependencies: number[], // Indices of subtasks this depends on (0-indexed)
573
- estimated_complexity: 1-5 // Effort estimate
574
- },
575
- // ... more subtasks
576
- ]
577
- }
578
- \`\`\`
579
-
580
- ## Guidelines
581
-
582
- - **Plan aggressively** - when in doubt, split further. 3 small tasks > 1 medium task
583
- - **Prefer smaller, focused subtasks** over large complex ones
584
- - **Include test files** in the same subtask as the code they test
585
- - **Consider shared types** - if multiple files share types, handle that first
586
- - **Think about imports** - changes to exported APIs affect downstream files
587
- - **Explicit > implicit** - spell out what each subtask should do, don't assume
588
-
589
- ## File Assignment Examples
590
-
591
- - Schema change: \`["src/schemas/user.ts", "src/schemas/index.ts"]\`
592
- - Component + test: \`["src/components/Button.tsx", "src/components/Button.test.tsx"]\`
593
- - API route: \`["src/app/api/users/route.ts"]\`
594
-
595
- Now decompose the task:`;
596
-
597
- /**
598
- * Prompt template for spawned subtask agents.
599
- *
600
- * Each agent receives this prompt with their specific subtask details filled in.
601
- * The prompt establishes context, constraints, and expectations.
602
- */
603
- export const SUBTASK_PROMPT = `You are a swarm agent working on a subtask of a larger epic.
604
-
605
- ## Your Identity
606
- - **Agent Name**: {agent_name}
607
- - **Bead ID**: {bead_id}
608
- - **Epic ID**: {epic_id}
609
-
610
- ## Your Subtask
611
- **Title**: {subtask_title}
612
-
613
- {subtask_description}
614
-
615
- ## File Scope
616
- You have exclusive reservations for these files:
617
- {file_list}
618
-
619
- **CRITICAL**: Only modify files in your reservation. If you need to modify other files,
620
- send a message to the coordinator requesting the change.
621
-
622
- ## Shared Context
623
- {shared_context}
624
-
625
- ## MANDATORY: Beads Tracking
626
-
627
- You MUST keep your bead updated as you work:
628
-
629
- 1. **Your bead is already in_progress** - don't change this unless blocked
630
- 2. **If blocked**: \`bd update {bead_id} --status blocked\` and message coordinator
631
- 3. **When done**: Use \`swarm_complete\` - it closes your bead automatically
632
- 4. **Discovered issues**: Create new beads with \`bd create "issue" -t bug\`
633
-
634
- **Never work silently.** Your bead status is how the swarm tracks progress.
635
-
636
- ## MANDATORY: Swarm Mail Communication
637
-
638
- You MUST communicate with other agents:
639
-
640
- 1. **Report progress** every significant milestone (not just at the end)
641
- 2. **Ask questions** if requirements are unclear - don't guess
642
- 3. **Announce blockers** immediately - don't spin trying to fix alone
643
- 4. **Coordinate on shared concerns** - if you see something affecting other agents, say so
644
-
645
- Use Swarm Mail for all communication:
646
- \`\`\`
647
- swarmmail_send(
648
- to: ["coordinator" or specific agent],
649
- subject: "Brief subject",
650
- body: "Message content",
651
- thread_id: "{epic_id}"
652
- )
653
- \`\`\`
654
-
655
- ## Coordination Protocol
656
-
657
- 1. **Start**: Your bead is already marked in_progress
658
- 2. **Progress**: Use swarm_progress to report status updates
659
- 3. **Blocked**: Report immediately via Swarm Mail - don't spin
660
- 4. **Complete**: Use swarm_complete when done - it handles:
661
- - Closing your bead with a summary
662
- - Releasing file reservations
663
- - Notifying the coordinator
664
-
665
- ## Self-Evaluation
666
-
667
- Before calling swarm_complete, evaluate your work:
668
- - Type safety: Does it compile without errors?
669
- - No obvious bugs: Did you handle edge cases?
670
- - Follows patterns: Does it match existing code style?
671
- - Readable: Would another developer understand it?
672
-
673
- If evaluation fails, fix the issues before completing.
674
-
675
- ## Planning Your Work
676
-
677
- Before writing code:
678
- 1. **Read the files** you're assigned to understand current state
679
- 2. **Plan your approach** - what changes, in what order?
680
- 3. **Identify risks** - what could go wrong? What dependencies?
681
- 4. **Communicate your plan** via Swarm Mail if non-trivial
682
-
683
- Begin work on your subtask now.`;
684
-
685
- /**
686
- * Streamlined subtask prompt (V2) - uses Swarm Mail and beads
687
- *
688
- * This is a cleaner version of SUBTASK_PROMPT that's easier to parse.
689
- * Agents MUST use Swarm Mail for communication and beads for tracking.
690
- *
691
- * Supports {error_context} placeholder for retry prompts.
692
- */
693
- export const SUBTASK_PROMPT_V2 = `You are a swarm agent working on: **{subtask_title}**
694
-
695
- ## [IDENTITY]
696
- Agent: (assigned at spawn)
697
- Bead: {bead_id}
698
- Epic: {epic_id}
699
-
700
- ## [TASK]
701
- {subtask_description}
702
-
703
- ## [FILES]
704
- Reserved (exclusive):
705
- {file_list}
706
-
707
- Only modify these files. Need others? Message the coordinator.
708
-
709
- ## [CONTEXT]
710
- {shared_context}
711
-
712
- {compressed_context}
713
-
714
- {error_context}
715
-
716
- ## [MANDATORY: SWARM MAIL]
717
-
718
- **YOU MUST USE SWARM MAIL FOR ALL COORDINATION.** This is non-negotiable.
719
-
720
- ### Initialize FIRST (before any work)
721
- \`\`\`
722
- swarmmail_init(project_path="$PWD", task_description="{subtask_title}")
723
- \`\`\`
724
-
725
- ### Reserve Files (if not already reserved by coordinator)
726
- \`\`\`
727
- swarmmail_reserve(paths=[...files...], reason="{bead_id}: {subtask_title}")
728
- \`\`\`
729
-
730
- ### Check Inbox Regularly
731
- \`\`\`
732
- swarmmail_inbox() # Check for coordinator messages
733
- swarmmail_read_message(message_id=N) # Read specific message
734
- \`\`\`
735
-
736
- ### Report Progress (REQUIRED - don't work silently)
737
- \`\`\`
738
- swarmmail_send(
739
- to=["coordinator"],
740
- subject="Progress: {bead_id}",
741
- body="<what you did, blockers, questions>",
742
- thread_id="{epic_id}"
743
- )
744
- \`\`\`
745
-
746
- ### When Blocked
747
- \`\`\`
748
- swarmmail_send(
749
- to=["coordinator"],
750
- subject="BLOCKED: {bead_id}",
751
- body="<blocker description, what you need>",
752
- importance="high",
753
- thread_id="{epic_id}"
754
- )
755
- beads_update(id="{bead_id}", status="blocked")
756
- \`\`\`
757
-
758
- ### Release Files When Done
759
- \`\`\`
760
- swarmmail_release() # Or let swarm_complete handle it
761
- \`\`\`
762
-
763
- ## [OTHER TOOLS]
764
- ### Beads
765
- - beads_update(id, status) - Mark blocked if stuck
766
- - beads_create(title, type) - Log new bugs found
767
-
768
- ### Skills (if available)
769
- - skills_list() - Discover available skills
770
- - skills_use(name) - Activate skill for specialized guidance
771
-
772
- ### Completion (REQUIRED)
773
- - swarm_complete(project_key, agent_name, bead_id, summary, files_touched)
774
-
775
- ## [LEARNING]
776
- As you work, note reusable patterns, best practices, or domain insights:
777
- - If you discover something that would help future agents, consider creating a skill
778
- - Use skills_create to codify patterns for the project
779
- - Good skills have clear "when to use" descriptions with actionable instructions
780
- - Skills make swarms smarter over time
781
-
782
- ## [WORKFLOW]
783
- 1. **swarmmail_init** - Initialize session FIRST
784
- 2. Read assigned files
785
- 3. Implement changes
786
- 4. **swarmmail_send** - Report progress to coordinator
787
- 5. Verify (typecheck)
788
- 6. **swarm_complete** - Mark done, release reservations
789
-
790
- **CRITICAL: Never work silently. Send progress updates via swarmmail_send every significant milestone.**
791
-
792
- Begin now.`;
793
-
794
- /**
795
- * Format the V2 subtask prompt for a specific agent
796
- */
797
- export function formatSubtaskPromptV2(params: {
798
- bead_id: string;
799
- epic_id: string;
800
- subtask_title: string;
801
- subtask_description: string;
802
- files: string[];
803
- shared_context?: string;
804
- compressed_context?: string;
805
- error_context?: string;
806
- }): string {
807
- const fileList =
808
- params.files.length > 0
809
- ? params.files.map((f) => `- \`${f}\``).join("\n")
810
- : "(no specific files - use judgment)";
811
-
812
- const compressedSection = params.compressed_context
813
- ? params.compressed_context
814
- : "";
815
-
816
- const errorSection = params.error_context ? params.error_context : "";
817
-
818
- return SUBTASK_PROMPT_V2.replace(/{bead_id}/g, params.bead_id)
819
- .replace(/{epic_id}/g, params.epic_id)
820
- .replace("{subtask_title}", params.subtask_title)
821
- .replace(
822
- "{subtask_description}",
823
- params.subtask_description || "(see title)",
824
- )
825
- .replace("{file_list}", fileList)
826
- .replace("{shared_context}", params.shared_context || "(none)")
827
- .replace("{compressed_context}", compressedSection)
828
- .replace("{error_context}", errorSection);
829
- }
830
-
831
- /**
832
- * Prompt for self-evaluation before completing a subtask.
833
- *
834
- * Agents use this to assess their work quality before marking complete.
835
- */
836
- export const EVALUATION_PROMPT = `Evaluate the work completed for this subtask.
837
-
838
- ## Subtask
839
- **Bead ID**: {bead_id}
840
- **Title**: {subtask_title}
841
-
842
- ## Files Modified
843
- {files_touched}
844
-
845
- ## Evaluation Criteria
846
-
847
- For each criterion, assess passed/failed and provide brief feedback:
848
-
849
- 1. **type_safe**: Code compiles without TypeScript errors
850
- 2. **no_bugs**: No obvious bugs, edge cases handled
851
- 3. **patterns**: Follows existing codebase patterns and conventions
852
- 4. **readable**: Code is clear and maintainable
853
-
854
- ## Response Format
855
-
856
- \`\`\`json
857
- {
858
- "passed": boolean, // Overall pass/fail
859
- "criteria": {
860
- "type_safe": { "passed": boolean, "feedback": string },
861
- "no_bugs": { "passed": boolean, "feedback": string },
862
- "patterns": { "passed": boolean, "feedback": string },
863
- "readable": { "passed": boolean, "feedback": string }
864
- },
865
- "overall_feedback": string,
866
- "retry_suggestion": string | null // If failed, what to fix
867
- }
868
- \`\`\`
869
-
870
- If any criterion fails, the overall evaluation fails and retry_suggestion
871
- should describe what needs to be fixed.`;
872
-
873
- // ============================================================================
874
- // Errors
875
- // ============================================================================
876
-
877
- export class SwarmError extends Error {
878
- constructor(
879
- message: string,
880
- public readonly operation: string,
881
- public readonly details?: unknown,
882
- ) {
883
- super(message);
884
- this.name = "SwarmError";
885
- }
886
- }
887
-
888
- export class DecompositionError extends SwarmError {
889
- constructor(
890
- message: string,
891
- public readonly zodError?: z.ZodError,
892
- ) {
893
- super(message, "decompose", zodError?.issues);
894
- }
895
- }
896
-
897
- // ============================================================================
898
- // Helper Functions
899
- // ============================================================================
900
-
901
- /**
902
- * Format the decomposition prompt with actual values
903
- */
904
- function formatDecompositionPrompt(
905
- task: string,
906
- maxSubtasks: number,
907
- context?: string,
908
- ): string {
909
- const contextSection = context
910
- ? `## Additional Context\n${context}`
911
- : "## Additional Context\n(none provided)";
912
-
913
- return DECOMPOSITION_PROMPT.replace("{task}", task)
914
- .replace("{max_subtasks}", maxSubtasks.toString())
915
- .replace("{context_section}", contextSection);
916
- }
917
-
918
- /**
919
- * Format the subtask prompt for a specific agent
920
- */
921
- export function formatSubtaskPrompt(params: {
922
- agent_name: string;
923
- bead_id: string;
924
- epic_id: string;
925
- subtask_title: string;
926
- subtask_description: string;
927
- files: string[];
928
- shared_context?: string;
929
- }): string {
930
- const fileList = params.files.map((f) => `- \`${f}\``).join("\n");
931
-
932
- return SUBTASK_PROMPT.replace("{agent_name}", params.agent_name)
933
- .replace("{bead_id}", params.bead_id)
934
- .replace(/{epic_id}/g, params.epic_id)
935
- .replace("{subtask_title}", params.subtask_title)
936
- .replace("{subtask_description}", params.subtask_description || "(none)")
937
- .replace("{file_list}", fileList || "(no files assigned)")
938
- .replace("{shared_context}", params.shared_context || "(none)");
939
- }
940
-
941
- /**
942
- * Format the evaluation prompt
943
- */
944
- export function formatEvaluationPrompt(params: {
945
- bead_id: string;
946
- subtask_title: string;
947
- files_touched: string[];
948
- }): string {
949
- const filesList = params.files_touched.map((f) => `- \`${f}\``).join("\n");
950
-
951
- return EVALUATION_PROMPT.replace("{bead_id}", params.bead_id)
952
- .replace("{subtask_title}", params.subtask_title)
953
- .replace("{files_touched}", filesList || "(no files recorded)");
954
- }
955
-
956
- /**
957
- * Query beads for subtasks of an epic
958
- */
959
- async function queryEpicSubtasks(epicId: string): Promise<Bead[]> {
960
- // Check if beads is available
961
- const beadsAvailable = await isToolAvailable("beads");
962
- if (!beadsAvailable) {
963
- warnMissingTool("beads");
964
- return []; // Return empty - swarm can still function without status tracking
965
- }
966
-
967
- const result = await Bun.$`bd list --parent ${epicId} --json`
968
- .quiet()
969
- .nothrow();
970
-
971
- if (result.exitCode !== 0) {
972
- // Don't throw - just return empty and log error prominently
973
- console.error(
974
- `[swarm] ERROR: Failed to query subtasks for epic ${epicId}:`,
975
- result.stderr.toString(),
976
- );
977
- return [];
978
- }
979
-
980
- try {
981
- const parsed = JSON.parse(result.stdout.toString());
982
- return z.array(BeadSchema).parse(parsed);
983
- } catch (error) {
984
- if (error instanceof z.ZodError) {
985
- console.error(
986
- `[swarm] ERROR: Invalid bead data for epic ${epicId}:`,
987
- error.message,
988
- );
989
- return [];
990
- }
991
- console.error(
992
- `[swarm] ERROR: Failed to parse beads for epic ${epicId}:`,
993
- error,
994
- );
995
- throw error;
996
- }
997
- }
998
-
999
- /**
1000
- * Query Agent Mail for swarm thread messages
1001
- */
1002
- async function querySwarmMessages(
1003
- projectKey: string,
1004
- threadId: string,
1005
- ): Promise<number> {
1006
- // Check if agent-mail is available
1007
- const agentMailAvailable = await isToolAvailable("agent-mail");
1008
- if (!agentMailAvailable) {
1009
- // Don't warn here - it's checked elsewhere
1010
- return 0;
1011
- }
1012
-
1013
- try {
1014
- // Use embedded swarm-mail inbox to count messages in thread
1015
- const inbox = await getSwarmInbox({
1016
- projectPath: projectKey,
1017
- agentName: "coordinator", // Dummy agent name for thread query
1018
- limit: 5,
1019
- includeBodies: false,
1020
- });
1021
-
1022
- // Count messages that match the thread ID
1023
- const threadMessages = inbox.messages.filter(
1024
- (m) => m.thread_id === threadId,
1025
- );
1026
- return threadMessages.length;
1027
- } catch (error) {
1028
- // Thread might not exist yet, or query failed
1029
- console.warn(
1030
- `[swarm] Failed to query swarm messages for thread ${threadId}:`,
1031
- error,
1032
- );
1033
- return 0;
1034
- }
1035
- }
1036
-
1037
- /**
1038
- * Format a progress message for Agent Mail
1039
- */
1040
- function formatProgressMessage(progress: AgentProgress): string {
1041
- const lines = [
1042
- `**Status**: ${progress.status}`,
1043
- progress.progress_percent !== undefined
1044
- ? `**Progress**: ${progress.progress_percent}%`
1045
- : null,
1046
- progress.message ? `**Message**: ${progress.message}` : null,
1047
- progress.files_touched && progress.files_touched.length > 0
1048
- ? `**Files touched**:\n${progress.files_touched.map((f) => `- \`${f}\``).join("\n")}`
1049
- : null,
1050
- progress.blockers && progress.blockers.length > 0
1051
- ? `**Blockers**:\n${progress.blockers.map((b) => `- ${b}`).join("\n")}`
1052
- : null,
1053
- ];
1054
-
1055
- return lines.filter(Boolean).join("\n\n");
1056
- }
1057
-
1058
- // ============================================================================
1059
- // CASS History Integration
1060
- // ============================================================================
1061
-
1062
- /**
1063
- * CASS search result from similar past tasks
1064
- */
1065
- interface CassSearchResult {
1066
- query: string;
1067
- results: Array<{
1068
- source_path: string;
1069
- line: number;
1070
- agent: string;
1071
- preview: string;
1072
- score: number;
1073
- }>;
1074
- }
1075
-
1076
- /**
1077
- * CASS query result with status
1078
- */
1079
- type CassQueryResult =
1080
- | { status: "unavailable" }
1081
- | { status: "failed"; error?: string }
1082
- | { status: "empty"; query: string }
1083
- | { status: "success"; data: CassSearchResult };
1084
-
1085
- /**
1086
- * Query CASS for similar past tasks
1087
- *
1088
- * @param task - Task description to search for
1089
- * @param limit - Maximum results to return
1090
- * @returns Structured result with status indicator
1091
- */
1092
- async function queryCassHistory(
1093
- task: string,
1094
- limit: number = 3,
1095
- ): Promise<CassQueryResult> {
1096
- // Check if CASS is available first
1097
- const cassAvailable = await isToolAvailable("cass");
1098
- if (!cassAvailable) {
1099
- warnMissingTool("cass");
1100
- return { status: "unavailable" };
1101
- }
1102
-
1103
- try {
1104
- const result = await Bun.$`cass search ${task} --limit ${limit} --json`
1105
- .quiet()
1106
- .nothrow();
1107
-
1108
- if (result.exitCode !== 0) {
1109
- const error = result.stderr.toString();
1110
- console.warn(
1111
- `[swarm] CASS search failed (exit ${result.exitCode}):`,
1112
- error,
1113
- );
1114
- return { status: "failed", error };
1115
- }
1116
-
1117
- const output = result.stdout.toString();
1118
- if (!output.trim()) {
1119
- return { status: "empty", query: task };
1120
- }
1121
-
1122
- try {
1123
- const parsed = JSON.parse(output);
1124
- const searchResult: CassSearchResult = {
1125
- query: task,
1126
- results: Array.isArray(parsed) ? parsed : parsed.results || [],
1127
- };
1128
-
1129
- if (searchResult.results.length === 0) {
1130
- return { status: "empty", query: task };
1131
- }
1132
-
1133
- return { status: "success", data: searchResult };
1134
- } catch (error) {
1135
- console.warn(`[swarm] Failed to parse CASS output:`, error);
1136
- return { status: "failed", error: String(error) };
1137
- }
1138
- } catch (error) {
1139
- console.error(`[swarm] CASS query error:`, error);
1140
- return { status: "failed", error: String(error) };
1141
- }
1142
- }
1143
-
1144
- /**
1145
- * Format CASS history for inclusion in decomposition prompt
1146
- */
1147
- function formatCassHistoryForPrompt(history: CassSearchResult): string {
1148
- if (history.results.length === 0) {
1149
- return "";
1150
- }
1151
-
1152
- const lines = [
1153
- "## Similar Past Tasks",
1154
- "",
1155
- "These similar tasks were found in agent history:",
1156
- "",
1157
- ...history.results.slice(0, 3).map((r, i) => {
1158
- const preview = r.preview.slice(0, 200).replace(/\n/g, " ");
1159
- return `${i + 1}. [${r.agent}] ${preview}...`;
1160
- }),
1161
- "",
1162
- "Consider patterns that worked in these past tasks.",
1163
- "",
1164
- ];
1165
-
1166
- return lines.join("\n");
1167
- }
1168
-
1169
- // ============================================================================
1170
- // Tool Definitions
1171
- // ============================================================================
1172
-
1173
- /**
1174
- * Select the best decomposition strategy for a task
1175
- *
1176
- * Analyzes task description and recommends a strategy with reasoning.
1177
- * Use this before swarm_plan_prompt to understand the recommended approach.
1178
- */
1179
- export const swarm_select_strategy = tool({
1180
- description:
1181
- "Analyze task and recommend decomposition strategy (file-based, feature-based, or risk-based)",
1182
- args: {
1183
- task: tool.schema.string().min(1).describe("Task description to analyze"),
1184
- codebase_context: tool.schema
1185
- .string()
1186
- .optional()
1187
- .describe("Optional codebase context (file structure, tech stack, etc.)"),
1188
- },
1189
- async execute(args) {
1190
- const result = selectStrategy(args.task);
1191
-
1192
- // Enhance reasoning with codebase context if provided
1193
- let enhancedReasoning = result.reasoning;
1194
- if (args.codebase_context) {
1195
- enhancedReasoning += `\n\nCodebase context considered: ${args.codebase_context.slice(0, 200)}...`;
1196
- }
1197
-
1198
- return JSON.stringify(
1199
- {
1200
- strategy: result.strategy,
1201
- confidence: Math.round(result.confidence * 100) / 100,
1202
- reasoning: enhancedReasoning,
1203
- description: STRATEGIES[result.strategy].description,
1204
- guidelines: STRATEGIES[result.strategy].guidelines,
1205
- anti_patterns: STRATEGIES[result.strategy].antiPatterns,
1206
- alternatives: result.alternatives.map((alt) => ({
1207
- strategy: alt.strategy,
1208
- description: STRATEGIES[alt.strategy].description,
1209
- score: alt.score,
1210
- })),
1211
- },
1212
- null,
1213
- 2,
1214
- );
1215
- },
1216
- });
1217
-
1218
- /**
1219
- * Strategy-specific decomposition prompt template
1220
- */
1221
- const STRATEGY_DECOMPOSITION_PROMPT = `You are decomposing a task into parallelizable subtasks for a swarm of agents.
1222
-
1223
- ## Task
1224
- {task}
1225
-
1226
- {strategy_guidelines}
1227
-
1228
- {context_section}
1229
-
1230
- {cass_history}
1231
-
1232
- {skills_context}
1233
-
1234
- ## MANDATORY: Beads Issue Tracking
1235
-
1236
- **Every subtask MUST become a bead.** This is non-negotiable.
1237
-
1238
- After decomposition, the coordinator will:
1239
- 1. Create an epic bead for the overall task
1240
- 2. Create child beads for each subtask
1241
- 3. Track progress through bead status updates
1242
- 4. Close beads with summaries when complete
1243
-
1244
- Agents MUST update their bead status as they work. No silent progress.
1245
-
1246
- ## Requirements
1247
-
1248
- 1. **Break into 2-{max_subtasks} independent subtasks** that can run in parallel
1249
- 2. **Assign files** - each subtask must specify which files it will modify
1250
- 3. **No file overlap** - files cannot appear in multiple subtasks (they get exclusive locks)
1251
- 4. **Order by dependency** - if subtask B needs subtask A's output, A must come first in the array
1252
- 5. **Estimate complexity** - 1 (trivial) to 5 (complex)
1253
- 6. **Plan aggressively** - break down more than you think necessary, smaller is better
1254
-
1255
- ## Response Format
1256
-
1257
- Respond with a JSON object matching this schema:
1258
-
1259
- \`\`\`typescript
1260
- {
1261
- epic: {
1262
- title: string, // Epic title for the beads tracker
1263
- description?: string // Brief description of the overall goal
1264
- },
1265
- subtasks: [
1266
- {
1267
- title: string, // What this subtask accomplishes
1268
- description?: string, // Detailed instructions for the agent
1269
- files: string[], // Files this subtask will modify (globs allowed)
1270
- dependencies: number[], // Indices of subtasks this depends on (0-indexed)
1271
- estimated_complexity: 1-5 // Effort estimate
1272
- },
1273
- // ... more subtasks
1274
- ]
1275
- }
1276
- \`\`\`
1277
-
1278
- Now decompose the task:`;
1279
-
1280
- /**
1281
- * Generate a strategy-specific planning prompt
1282
- *
1283
- * Higher-level than swarm_decompose - includes strategy selection and guidelines.
1284
- * Use this when you want the full planning experience with strategy-specific advice.
1285
- */
1286
- export const swarm_plan_prompt = tool({
1287
- description:
1288
- "Generate strategy-specific decomposition prompt. Auto-selects strategy or uses provided one. Queries CASS for similar tasks.",
1289
- args: {
1290
- task: tool.schema.string().min(1).describe("Task description to decompose"),
1291
- strategy: tool.schema
1292
- .enum(["file-based", "feature-based", "risk-based", "auto"])
1293
- .optional()
1294
- .describe("Decomposition strategy (default: auto-detect)"),
1295
- max_subtasks: tool.schema
1296
- .number()
1297
- .int()
1298
- .min(2)
1299
-
1300
- .default(5)
1301
- .describe("Maximum number of subtasks (default: 5)"),
1302
- context: tool.schema
1303
- .string()
1304
- .optional()
1305
- .describe("Additional context (codebase info, constraints, etc.)"),
1306
- query_cass: tool.schema
1307
- .boolean()
1308
- .optional()
1309
- .describe("Query CASS for similar past tasks (default: true)"),
1310
- cass_limit: tool.schema
1311
- .number()
1312
- .int()
1313
- .min(1)
1314
-
1315
- .optional()
1316
- .describe("Max CASS results to include (default: 3)"),
1317
- include_skills: tool.schema
1318
- .boolean()
1319
- .optional()
1320
- .describe("Include available skills in context (default: true)"),
1321
- },
1322
- async execute(args) {
1323
- // Select strategy
1324
- let selectedStrategy: Exclude<DecompositionStrategy, "auto">;
1325
- let strategyReasoning: string;
1326
-
1327
- if (args.strategy && args.strategy !== "auto") {
1328
- selectedStrategy = args.strategy;
1329
- strategyReasoning = `User-specified strategy: ${selectedStrategy}`;
1330
- } else {
1331
- const selection = selectStrategy(args.task);
1332
- selectedStrategy = selection.strategy;
1333
- strategyReasoning = selection.reasoning;
1334
- }
1335
-
1336
- // Query CASS for similar past tasks
1337
- let cassContext = "";
1338
- let cassResultInfo: {
1339
- queried: boolean;
1340
- results_found?: number;
1341
- included_in_context?: boolean;
1342
- reason?: string;
1343
- };
1344
-
1345
- if (args.query_cass !== false) {
1346
- const cassResult = await queryCassHistory(
1347
- args.task,
1348
- args.cass_limit ?? 3,
1349
- );
1350
- if (cassResult.status === "success") {
1351
- cassContext = formatCassHistoryForPrompt(cassResult.data);
1352
- cassResultInfo = {
1353
- queried: true,
1354
- results_found: cassResult.data.results.length,
1355
- included_in_context: true,
1356
- };
1357
- } else {
1358
- cassResultInfo = {
1359
- queried: true,
1360
- results_found: 0,
1361
- included_in_context: false,
1362
- reason: cassResult.status,
1363
- };
1364
- }
1365
- } else {
1366
- cassResultInfo = { queried: false, reason: "disabled" };
1367
- }
1368
-
1369
- // Fetch skills context
1370
- let skillsContext = "";
1371
- let skillsInfo: { included: boolean; count?: number; relevant?: string[] } =
1372
- {
1373
- included: false,
1374
- };
1375
-
1376
- if (args.include_skills !== false) {
1377
- const allSkills = await listSkills();
1378
- if (allSkills.length > 0) {
1379
- skillsContext = await getSkillsContextForSwarm();
1380
- const relevantSkills = await findRelevantSkills(args.task);
1381
- skillsInfo = {
1382
- included: true,
1383
- count: allSkills.length,
1384
- relevant: relevantSkills,
1385
- };
1386
-
1387
- // Add suggestion for relevant skills
1388
- if (relevantSkills.length > 0) {
1389
- skillsContext += `\n\n**Suggested skills for this task**: ${relevantSkills.join(", ")}`;
1390
- }
1391
- }
1392
- }
1393
-
1394
- // Format strategy guidelines
1395
- const strategyGuidelines = formatStrategyGuidelines(selectedStrategy);
1396
-
1397
- // Combine user context
1398
- const contextSection = args.context
1399
- ? `## Additional Context\n${args.context}`
1400
- : "## Additional Context\n(none provided)";
1401
-
1402
- // Build the prompt
1403
- const prompt = STRATEGY_DECOMPOSITION_PROMPT.replace("{task}", args.task)
1404
- .replace("{strategy_guidelines}", strategyGuidelines)
1405
- .replace("{context_section}", contextSection)
1406
- .replace("{cass_history}", cassContext || "")
1407
- .replace("{skills_context}", skillsContext || "")
1408
- .replace("{max_subtasks}", (args.max_subtasks ?? 5).toString());
1409
-
1410
- return JSON.stringify(
1411
- {
1412
- prompt,
1413
- strategy: {
1414
- selected: selectedStrategy,
1415
- reasoning: strategyReasoning,
1416
- guidelines: STRATEGIES[selectedStrategy].guidelines,
1417
- anti_patterns: STRATEGIES[selectedStrategy].antiPatterns,
1418
- },
1419
- expected_schema: "BeadTree",
1420
- schema_hint: {
1421
- epic: { title: "string", description: "string?" },
1422
- subtasks: [
1423
- {
1424
- title: "string",
1425
- description: "string?",
1426
- files: "string[]",
1427
- dependencies: "number[]",
1428
- estimated_complexity: "1-5",
1429
- },
1430
- ],
1431
- },
1432
- validation_note:
1433
- "Parse agent response as JSON and validate with swarm_validate_decomposition",
1434
- cass_history: cassResultInfo,
1435
- skills: skillsInfo,
1436
- // Add semantic-memory query instruction
1437
- memory_query: formatMemoryQueryForDecomposition(args.task, 3),
1438
- },
1439
- null,
1440
- 2,
1441
- );
1442
- },
1443
- });
1444
-
1445
- /**
1446
- * Delegate task decomposition to a swarm/planner subagent
1447
- *
1448
- * Returns a prompt for spawning a planner agent that will handle all decomposition
1449
- * reasoning. This keeps the coordinator context lean by offloading:
1450
- * - Strategy selection
1451
- * - CASS queries
1452
- * - Skills discovery
1453
- * - File analysis
1454
- * - BeadTree generation
1455
- *
1456
- * The planner returns ONLY structured BeadTree JSON, which the coordinator
1457
- * validates and uses to create beads.
1458
- *
1459
- * @example
1460
- * ```typescript
1461
- * // Coordinator workflow:
1462
- * const delegateResult = await swarm_delegate_planning({
1463
- * task: "Add user authentication",
1464
- * context: "Next.js 14 app",
1465
- * });
1466
- *
1467
- * // Parse the result
1468
- * const { prompt, subagent_type } = JSON.parse(delegateResult);
1469
- *
1470
- * // Spawn subagent using Task tool
1471
- * const plannerResponse = await Task(prompt, subagent_type);
1472
- *
1473
- * // Validate the response
1474
- * await swarm_validate_decomposition({ response: plannerResponse });
1475
- * ```
1476
- */
1477
- export const swarm_delegate_planning = tool({
1478
- description:
1479
- "Delegate task decomposition to a swarm/planner subagent. Returns a prompt to spawn the planner. Use this to keep coordinator context lean - all planning reasoning happens in the subagent.",
1480
- args: {
1481
- task: tool.schema.string().min(1).describe("The task to decompose"),
1482
- context: tool.schema
1483
- .string()
1484
- .optional()
1485
- .describe("Additional context to include"),
1486
- max_subtasks: tool.schema
1487
- .number()
1488
- .int()
1489
- .min(2)
1490
- .max(10)
1491
- .optional()
1492
- .default(5)
1493
- .describe("Maximum number of subtasks (default: 5)"),
1494
- strategy: tool.schema
1495
- .enum(["auto", "file-based", "feature-based", "risk-based"])
1496
- .optional()
1497
- .default("auto")
1498
- .describe("Decomposition strategy (default: auto-detect)"),
1499
- query_cass: tool.schema
1500
- .boolean()
1501
- .optional()
1502
- .default(true)
1503
- .describe("Query CASS for similar past tasks (default: true)"),
1504
- },
1505
- async execute(args) {
1506
- // Select strategy
1507
- let selectedStrategy: Exclude<DecompositionStrategy, "auto">;
1508
- let strategyReasoning: string;
1509
-
1510
- if (args.strategy && args.strategy !== "auto") {
1511
- selectedStrategy = args.strategy;
1512
- strategyReasoning = `User-specified strategy: ${selectedStrategy}`;
1513
- } else {
1514
- const selection = selectStrategy(args.task);
1515
- selectedStrategy = selection.strategy;
1516
- strategyReasoning = selection.reasoning;
1517
- }
1518
-
1519
- // Query CASS for similar past tasks
1520
- let cassContext = "";
1521
- let cassResultInfo: {
1522
- queried: boolean;
1523
- results_found?: number;
1524
- included_in_context?: boolean;
1525
- reason?: string;
1526
- };
1527
-
1528
- if (args.query_cass !== false) {
1529
- const cassResult = await queryCassHistory(args.task, 3);
1530
- if (cassResult.status === "success") {
1531
- cassContext = formatCassHistoryForPrompt(cassResult.data);
1532
- cassResultInfo = {
1533
- queried: true,
1534
- results_found: cassResult.data.results.length,
1535
- included_in_context: true,
1536
- };
1537
- } else {
1538
- cassResultInfo = {
1539
- queried: true,
1540
- results_found: 0,
1541
- included_in_context: false,
1542
- reason: cassResult.status,
1543
- };
1544
- }
1545
- } else {
1546
- cassResultInfo = { queried: false, reason: "disabled" };
1547
- }
1548
-
1549
- // Fetch skills context
1550
- let skillsContext = "";
1551
- let skillsInfo: { included: boolean; count?: number; relevant?: string[] } =
1552
- {
1553
- included: false,
1554
- };
1555
-
1556
- const allSkills = await listSkills();
1557
- if (allSkills.length > 0) {
1558
- skillsContext = await getSkillsContextForSwarm();
1559
- const relevantSkills = await findRelevantSkills(args.task);
1560
- skillsInfo = {
1561
- included: true,
1562
- count: allSkills.length,
1563
- relevant: relevantSkills,
1564
- };
1565
-
1566
- // Add suggestion for relevant skills
1567
- if (relevantSkills.length > 0) {
1568
- skillsContext += `\n\n**Suggested skills for this task**: ${relevantSkills.join(", ")}`;
1569
- }
1570
- }
1571
-
1572
- // Format strategy guidelines
1573
- const strategyGuidelines = formatStrategyGuidelines(selectedStrategy);
1574
-
1575
- // Combine user context
1576
- const contextSection = args.context
1577
- ? `## Additional Context\n${args.context}`
1578
- : "## Additional Context\n(none provided)";
1579
-
1580
- // Build the planning prompt with clear instructions for JSON-only output
1581
- const planningPrompt = STRATEGY_DECOMPOSITION_PROMPT.replace(
1582
- "{task}",
1583
- args.task,
1584
- )
1585
- .replace("{strategy_guidelines}", strategyGuidelines)
1586
- .replace("{context_section}", contextSection)
1587
- .replace("{cass_history}", cassContext || "")
1588
- .replace("{skills_context}", skillsContext || "")
1589
- .replace("{max_subtasks}", (args.max_subtasks ?? 5).toString());
1590
-
1591
- // Add strict JSON-only instructions for the subagent
1592
- const subagentInstructions = `
1593
- ## CRITICAL: Output Format
1594
-
1595
- You are a planner subagent. Your ONLY output must be valid JSON matching the BeadTree schema.
1596
-
1597
- DO NOT include:
1598
- - Explanatory text before or after the JSON
1599
- - Markdown code fences (\`\`\`json)
1600
- - Commentary or reasoning
1601
-
1602
- OUTPUT ONLY the raw JSON object.
1603
-
1604
- ## Example Output
1605
-
1606
- {
1607
- "epic": {
1608
- "title": "Add user authentication",
1609
- "description": "Implement OAuth-based authentication system"
1610
- },
1611
- "subtasks": [
1612
- {
1613
- "title": "Set up OAuth provider",
1614
- "description": "Configure OAuth client credentials and redirect URLs",
1615
- "files": ["src/auth/oauth.ts", "src/config/auth.ts"],
1616
- "dependencies": [],
1617
- "estimated_complexity": 2
1618
- },
1619
- {
1620
- "title": "Create auth routes",
1621
- "description": "Implement login, logout, and callback routes",
1622
- "files": ["src/app/api/auth/[...nextauth]/route.ts"],
1623
- "dependencies": [0],
1624
- "estimated_complexity": 3
1625
- }
1626
- ]
1627
- }
1628
-
1629
- Now generate the BeadTree for the given task.`;
1630
-
1631
- const fullPrompt = `${planningPrompt}\n\n${subagentInstructions}`;
1632
-
1633
- // Return structured output for coordinator
1634
- return JSON.stringify(
1635
- {
1636
- prompt: fullPrompt,
1637
- subagent_type: "swarm/planner",
1638
- description: "Task decomposition planning",
1639
- strategy: {
1640
- selected: selectedStrategy,
1641
- reasoning: strategyReasoning,
1642
- },
1643
- expected_output: "BeadTree JSON (raw JSON, no markdown)",
1644
- next_steps: [
1645
- "1. Spawn subagent with Task tool using returned prompt",
1646
- "2. Parse subagent response as JSON",
1647
- "3. Validate with swarm_validate_decomposition",
1648
- "4. Create beads with beads_create_epic",
1649
- ],
1650
- cass_history: cassResultInfo,
1651
- skills: skillsInfo,
1652
- // Add semantic-memory query instruction
1653
- memory_query: formatMemoryQueryForDecomposition(args.task, 3),
1654
- },
1655
- null,
1656
- 2,
1657
- );
1658
- },
1659
- });
1660
-
1661
- /**
1662
- * Decompose a task into a bead tree
1663
- *
1664
- * This is a PROMPT tool - it returns a prompt for the agent to respond to.
1665
- * The agent's response (JSON) should be validated with BeadTreeSchema.
1666
- *
1667
- * Optionally queries CASS for similar past tasks to inform decomposition.
1668
- */
1669
- export const swarm_decompose = tool({
1670
- description:
1671
- "Generate decomposition prompt for breaking task into parallelizable subtasks. Optionally queries CASS for similar past tasks.",
1672
- args: {
1673
- task: tool.schema.string().min(1).describe("Task description to decompose"),
1674
- max_subtasks: tool.schema
1675
- .number()
1676
- .int()
1677
- .min(2)
1678
-
1679
- .default(5)
1680
- .describe("Maximum number of subtasks (default: 5)"),
1681
- context: tool.schema
1682
- .string()
1683
- .optional()
1684
- .describe("Additional context (codebase info, constraints, etc.)"),
1685
- query_cass: tool.schema
1686
- .boolean()
1687
- .optional()
1688
- .describe("Query CASS for similar past tasks (default: true)"),
1689
- cass_limit: tool.schema
1690
- .number()
1691
- .int()
1692
- .min(1)
1693
-
1694
- .optional()
1695
- .describe("Max CASS results to include (default: 3)"),
1696
- },
1697
- async execute(args) {
1698
- // Query CASS for similar past tasks
1699
- let cassContext = "";
1700
- let cassResultInfo: {
1701
- queried: boolean;
1702
- results_found?: number;
1703
- included_in_context?: boolean;
1704
- reason?: string;
1705
- };
1706
-
1707
- if (args.query_cass !== false) {
1708
- const cassResult = await queryCassHistory(
1709
- args.task,
1710
- args.cass_limit ?? 3,
1711
- );
1712
- if (cassResult.status === "success") {
1713
- cassContext = formatCassHistoryForPrompt(cassResult.data);
1714
- cassResultInfo = {
1715
- queried: true,
1716
- results_found: cassResult.data.results.length,
1717
- included_in_context: true,
1718
- };
1719
- } else {
1720
- cassResultInfo = {
1721
- queried: true,
1722
- results_found: 0,
1723
- included_in_context: false,
1724
- reason: cassResult.status,
1725
- };
1726
- }
1727
- } else {
1728
- cassResultInfo = { queried: false, reason: "disabled" };
1729
- }
1730
-
1731
- // Combine user context with CASS history
1732
- const fullContext = [args.context, cassContext]
1733
- .filter(Boolean)
1734
- .join("\n\n");
1735
-
1736
- const prompt = formatDecompositionPrompt(
1737
- args.task,
1738
- args.max_subtasks ?? 5,
1739
- fullContext || undefined,
1740
- );
1741
-
1742
- // Return the prompt and schema info for the caller
1743
- return JSON.stringify(
1744
- {
1745
- prompt,
1746
- expected_schema: "BeadTree",
1747
- schema_hint: {
1748
- epic: { title: "string", description: "string?" },
1749
- subtasks: [
1750
- {
1751
- title: "string",
1752
- description: "string?",
1753
- files: "string[]",
1754
- dependencies: "number[]",
1755
- estimated_complexity: "1-5",
1756
- },
1757
- ],
1758
- },
1759
- validation_note:
1760
- "Parse agent response as JSON and validate with BeadTreeSchema from schemas/bead.ts",
1761
- cass_history: cassResultInfo,
1762
- // Add semantic-memory query instruction
1763
- memory_query: formatMemoryQueryForDecomposition(args.task, 3),
1764
- },
1765
- null,
1766
- 2,
1767
- );
1768
- },
1769
- });
1770
-
1771
- /**
1772
- * Validate a decomposition response from an agent
1773
- *
1774
- * Use this after the agent responds to swarm:decompose to validate the structure.
1775
- */
1776
- export const swarm_validate_decomposition = tool({
1777
- description: "Validate a decomposition response against BeadTreeSchema",
1778
- args: {
1779
- response: tool.schema
1780
- .string()
1781
- .describe("JSON response from agent (BeadTree format)"),
1782
- },
1783
- async execute(args) {
1784
- try {
1785
- const parsed = JSON.parse(args.response);
1786
- const validated = BeadTreeSchema.parse(parsed);
1787
-
1788
- // Additional validation: check for file conflicts
1789
- const allFiles = new Set<string>();
1790
- const conflicts: string[] = [];
1791
-
1792
- for (const subtask of validated.subtasks) {
1793
- for (const file of subtask.files) {
1794
- if (allFiles.has(file)) {
1795
- conflicts.push(file);
1796
- }
1797
- allFiles.add(file);
1798
- }
1799
- }
1800
-
1801
- if (conflicts.length > 0) {
1802
- return JSON.stringify(
1803
- {
1804
- valid: false,
1805
- error: `File conflicts detected: ${conflicts.join(", ")}`,
1806
- hint: "Each file can only be assigned to one subtask",
1807
- },
1808
- null,
1809
- 2,
1810
- );
1811
- }
1812
-
1813
- // Check dependency indices are valid
1814
- for (let i = 0; i < validated.subtasks.length; i++) {
1815
- const deps = validated.subtasks[i].dependencies;
1816
- for (const dep of deps) {
1817
- // Check bounds first
1818
- if (dep < 0 || dep >= validated.subtasks.length) {
1819
- return JSON.stringify(
1820
- {
1821
- valid: false,
1822
- error: `Invalid dependency: subtask ${i} depends on ${dep}, but only ${validated.subtasks.length} subtasks exist (indices 0-${validated.subtasks.length - 1})`,
1823
- hint: "Dependency index is out of bounds",
1824
- },
1825
- null,
1826
- 2,
1827
- );
1828
- }
1829
- // Check forward references
1830
- if (dep >= i) {
1831
- return JSON.stringify(
1832
- {
1833
- valid: false,
1834
- error: `Invalid dependency: subtask ${i} depends on ${dep}, but dependencies must be earlier in the array`,
1835
- hint: "Reorder subtasks so dependencies come before dependents",
1836
- },
1837
- null,
1838
- 2,
1839
- );
1840
- }
1841
- }
1842
- }
1843
-
1844
- // Check for instruction conflicts between subtasks
1845
- const instructionConflicts = detectInstructionConflicts(
1846
- validated.subtasks,
1847
- );
1848
-
1849
- return JSON.stringify(
1850
- {
1851
- valid: true,
1852
- bead_tree: validated,
1853
- stats: {
1854
- subtask_count: validated.subtasks.length,
1855
- total_files: allFiles.size,
1856
- total_complexity: validated.subtasks.reduce(
1857
- (sum, s) => sum + s.estimated_complexity,
1858
- 0,
1859
- ),
1860
- },
1861
- // Include conflicts as warnings (not blocking)
1862
- warnings:
1863
- instructionConflicts.length > 0
1864
- ? {
1865
- instruction_conflicts: instructionConflicts,
1866
- hint: "Review these potential conflicts between subtask instructions",
1867
- }
1868
- : undefined,
1869
- },
1870
- null,
1871
- 2,
1872
- );
1873
- } catch (error) {
1874
- if (error instanceof z.ZodError) {
1875
- return JSON.stringify(
1876
- {
1877
- valid: false,
1878
- error: "Schema validation failed",
1879
- details: error.issues,
1880
- },
1881
- null,
1882
- 2,
1883
- );
1884
- }
1885
- if (error instanceof SyntaxError) {
1886
- return JSON.stringify(
1887
- {
1888
- valid: false,
1889
- error: "Invalid JSON",
1890
- details: error.message,
1891
- },
1892
- null,
1893
- 2,
1894
- );
1895
- }
1896
- throw error;
1897
- }
1898
- },
1899
- });
1900
-
1901
- /**
1902
- * Get status of a swarm by epic ID
1903
- *
1904
- * Requires project_key to query Agent Mail for message counts.
1905
- */
1906
- export const swarm_status = tool({
1907
- description: "Get status of a swarm by epic ID",
1908
- args: {
1909
- epic_id: tool.schema.string().describe("Epic bead ID (e.g., bd-abc123)"),
1910
- project_key: tool.schema
1911
- .string()
1912
- .describe("Project path (for Agent Mail queries)"),
1913
- },
1914
- async execute(args) {
1915
- // Query subtasks from beads
1916
- const subtasks = await queryEpicSubtasks(args.epic_id);
1917
-
1918
- // Count statuses
1919
- const statusCounts = {
1920
- running: 0,
1921
- completed: 0,
1922
- failed: 0,
1923
- blocked: 0,
1924
- };
1925
-
1926
- const agents: SpawnedAgent[] = [];
1927
-
1928
- for (const bead of subtasks) {
1929
- // Map bead status to agent status
1930
- let agentStatus: SpawnedAgent["status"] = "pending";
1931
- switch (bead.status) {
1932
- case "in_progress":
1933
- agentStatus = "running";
1934
- statusCounts.running++;
1935
- break;
1936
- case "closed":
1937
- agentStatus = "completed";
1938
- statusCounts.completed++;
1939
- break;
1940
- case "blocked":
1941
- agentStatus = "pending"; // Blocked treated as pending for swarm
1942
- statusCounts.blocked++;
1943
- break;
1944
- default:
1945
- // open = pending
1946
- break;
1947
- }
1948
-
1949
- agents.push({
1950
- bead_id: bead.id,
1951
- agent_name: "", // We don't track this in beads
1952
- status: agentStatus,
1953
- files: [], // Would need to parse from description
1954
- });
1955
- }
1956
-
1957
- // Query Agent Mail for message activity
1958
- const messageCount = await querySwarmMessages(
1959
- args.project_key,
1960
- args.epic_id,
1961
- );
1962
-
1963
- const status: SwarmStatus = {
1964
- epic_id: args.epic_id,
1965
- total_agents: subtasks.length,
1966
- running: statusCounts.running,
1967
- completed: statusCounts.completed,
1968
- failed: statusCounts.failed,
1969
- blocked: statusCounts.blocked,
1970
- agents,
1971
- last_update: new Date().toISOString(),
1972
- };
1973
-
1974
- // Validate and return
1975
- const validated = SwarmStatusSchema.parse(status);
1976
-
1977
- return JSON.stringify(
1978
- {
1979
- ...validated,
1980
- message_count: messageCount,
1981
- progress_percent:
1982
- subtasks.length > 0
1983
- ? Math.round((statusCounts.completed / subtasks.length) * 100)
1984
- : 0,
1985
- },
1986
- null,
1987
- 2,
1988
- );
1989
- },
1990
- });
1991
-
1992
- /**
1993
- * Report progress on a subtask
1994
- *
1995
- * Takes explicit agent identity since tools don't have persistent state.
1996
- */
1997
- export const swarm_progress = tool({
1998
- description: "Report progress on a subtask to coordinator",
1999
- args: {
2000
- project_key: tool.schema.string().describe("Project path"),
2001
- agent_name: tool.schema.string().describe("Your Agent Mail name"),
2002
- bead_id: tool.schema.string().describe("Subtask bead ID"),
2003
- status: tool.schema
2004
- .enum(["in_progress", "blocked", "completed", "failed"])
2005
- .describe("Current status"),
2006
- message: tool.schema
2007
- .string()
2008
- .optional()
2009
- .describe("Progress message or blockers"),
2010
- progress_percent: tool.schema
2011
- .number()
2012
- .min(0)
2013
- .max(100)
2014
- .optional()
2015
- .describe("Completion percentage"),
2016
- files_touched: tool.schema
2017
- .array(tool.schema.string())
2018
- .optional()
2019
- .describe("Files modified so far"),
2020
- },
2021
- async execute(args) {
2022
- // Build progress report
2023
- const progress: AgentProgress = {
2024
- bead_id: args.bead_id,
2025
- agent_name: args.agent_name,
2026
- status: args.status,
2027
- progress_percent: args.progress_percent,
2028
- message: args.message,
2029
- files_touched: args.files_touched,
2030
- timestamp: new Date().toISOString(),
2031
- };
2032
-
2033
- // Validate
2034
- const validated = AgentProgressSchema.parse(progress);
2035
-
2036
- // Update bead status if needed
2037
- if (args.status === "blocked" || args.status === "in_progress") {
2038
- const beadStatus = args.status === "blocked" ? "blocked" : "in_progress";
2039
- await Bun.$`bd update ${args.bead_id} --status ${beadStatus} --json`
2040
- .quiet()
2041
- .nothrow();
2042
- }
2043
-
2044
- // Extract epic ID from bead ID (e.g., bd-abc123.1 -> bd-abc123)
2045
- const epicId = args.bead_id.includes(".")
2046
- ? args.bead_id.split(".")[0]
2047
- : args.bead_id;
2048
-
2049
- // Send progress message to thread using embedded swarm-mail
2050
- await sendSwarmMessage({
2051
- projectPath: args.project_key,
2052
- fromAgent: args.agent_name,
2053
- toAgents: [], // Coordinator will pick it up from thread
2054
- subject: `Progress: ${args.bead_id} - ${args.status}`,
2055
- body: formatProgressMessage(validated),
2056
- threadId: epicId,
2057
- importance: args.status === "blocked" ? "high" : "normal",
2058
- });
2059
-
2060
- return `Progress reported: ${args.status}${args.progress_percent !== undefined ? ` (${args.progress_percent}%)` : ""}`;
2061
- },
2062
- });
2063
-
2064
- /**
2065
- * UBS scan result schema
2066
- */
2067
- interface UbsScanResult {
2068
- exitCode: number;
2069
- bugs: Array<{
2070
- file: string;
2071
- line: number;
2072
- severity: string;
2073
- message: string;
2074
- category: string;
2075
- }>;
2076
- summary: {
2077
- total: number;
2078
- critical: number;
2079
- high: number;
2080
- medium: number;
2081
- low: number;
2082
- };
2083
- }
2084
-
2085
- // ============================================================================
2086
- // Verification Gate
2087
- // ============================================================================
2088
-
2089
- /**
2090
- * Verification Gate result - tracks each verification step
2091
- *
2092
- * Based on the Gate Function from superpowers:
2093
- * 1. IDENTIFY: What command proves this claim?
2094
- * 2. RUN: Execute the FULL command (fresh, complete)
2095
- * 3. READ: Full output, check exit code, count failures
2096
- * 4. VERIFY: Does output confirm the claim?
2097
- * 5. ONLY THEN: Make the claim
2098
- */
2099
- interface VerificationStep {
2100
- name: string;
2101
- command: string;
2102
- passed: boolean;
2103
- exitCode: number;
2104
- output?: string;
2105
- error?: string;
2106
- skipped?: boolean;
2107
- skipReason?: string;
2108
- }
2109
-
2110
- interface VerificationGateResult {
2111
- passed: boolean;
2112
- steps: VerificationStep[];
2113
- summary: string;
2114
- blockers: string[];
2115
- }
2116
-
2117
- /**
2118
- * Run typecheck verification
2119
- *
2120
- * Attempts to run TypeScript type checking on the project.
2121
- * Falls back gracefully if tsc is not available.
2122
- */
2123
- async function runTypecheckVerification(): Promise<VerificationStep> {
2124
- const step: VerificationStep = {
2125
- name: "typecheck",
2126
- command: "tsc --noEmit",
2127
- passed: false,
2128
- exitCode: -1,
2129
- };
2130
-
2131
- try {
2132
- // Check if tsconfig.json exists in current directory
2133
- const tsconfigExists = await Bun.file("tsconfig.json").exists();
2134
- if (!tsconfigExists) {
2135
- step.skipped = true;
2136
- step.skipReason = "No tsconfig.json found";
2137
- step.passed = true; // Don't block if no TypeScript
2138
- return step;
2139
- }
2140
-
2141
- const result = await Bun.$`tsc --noEmit`.quiet().nothrow();
2142
- step.exitCode = result.exitCode;
2143
- step.passed = result.exitCode === 0;
2144
-
2145
- if (!step.passed) {
2146
- step.error = result.stderr.toString().slice(0, 1000); // Truncate for context
2147
- step.output = result.stdout.toString().slice(0, 1000);
2148
- }
2149
- } catch (error) {
2150
- step.skipped = true;
2151
- step.skipReason = `tsc not available: ${error instanceof Error ? error.message : String(error)}`;
2152
- step.passed = true; // Don't block if tsc unavailable
2153
- }
2154
-
2155
- return step;
2156
- }
2157
-
2158
- /**
2159
- * Run test verification for specific files
2160
- *
2161
- * Attempts to find and run tests related to the touched files.
2162
- * Uses common test patterns (*.test.ts, *.spec.ts, __tests__/).
2163
- */
2164
- async function runTestVerification(
2165
- filesTouched: string[],
2166
- ): Promise<VerificationStep> {
2167
- const step: VerificationStep = {
2168
- name: "tests",
2169
- command: "bun test <related-files>",
2170
- passed: false,
2171
- exitCode: -1,
2172
- };
2173
-
2174
- if (filesTouched.length === 0) {
2175
- step.skipped = true;
2176
- step.skipReason = "No files touched";
2177
- step.passed = true;
2178
- return step;
2179
- }
2180
-
2181
- // Find test files related to touched files
2182
- const testPatterns: string[] = [];
2183
- for (const file of filesTouched) {
2184
- // Skip if already a test file
2185
- if (file.includes(".test.") || file.includes(".spec.")) {
2186
- testPatterns.push(file);
2187
- continue;
2188
- }
2189
-
2190
- // Look for corresponding test file
2191
- const baseName = file.replace(/\.(ts|tsx|js|jsx)$/, "");
2192
- testPatterns.push(`${baseName}.test.ts`);
2193
- testPatterns.push(`${baseName}.test.tsx`);
2194
- testPatterns.push(`${baseName}.spec.ts`);
2195
- }
2196
-
2197
- // Check if any test files exist
2198
- const existingTests: string[] = [];
2199
- for (const pattern of testPatterns) {
2200
- try {
2201
- const exists = await Bun.file(pattern).exists();
2202
- if (exists) {
2203
- existingTests.push(pattern);
2204
- }
2205
- } catch {
2206
- // File doesn't exist, skip
2207
- }
2208
- }
2209
-
2210
- if (existingTests.length === 0) {
2211
- step.skipped = true;
2212
- step.skipReason = "No related test files found";
2213
- step.passed = true;
2214
- return step;
2215
- }
2216
-
2217
- try {
2218
- step.command = `bun test ${existingTests.join(" ")}`;
2219
- const result = await Bun.$`bun test ${existingTests}`.quiet().nothrow();
2220
- step.exitCode = result.exitCode;
2221
- step.passed = result.exitCode === 0;
2222
-
2223
- if (!step.passed) {
2224
- step.error = result.stderr.toString().slice(0, 1000);
2225
- step.output = result.stdout.toString().slice(0, 1000);
2226
- }
2227
- } catch (error) {
2228
- step.skipped = true;
2229
- step.skipReason = `Test runner failed: ${error instanceof Error ? error.message : String(error)}`;
2230
- step.passed = true; // Don't block if test runner unavailable
2231
- }
2232
-
2233
- return step;
2234
- }
2235
-
2236
- /**
2237
- * Run the full Verification Gate
2238
- *
2239
- * Implements the Gate Function (IDENTIFY → RUN → READ → VERIFY → CLAIM):
2240
- * 1. UBS scan (already exists)
2241
- * 2. Typecheck
2242
- * 3. Tests for touched files
2243
- *
2244
- * All steps must pass (or be skipped with valid reason) to proceed.
2245
- */
2246
- async function runVerificationGate(
2247
- filesTouched: string[],
2248
- skipUbs: boolean = false,
2249
- ): Promise<VerificationGateResult> {
2250
- const steps: VerificationStep[] = [];
2251
- const blockers: string[] = [];
2252
-
2253
- // Step 1: UBS scan
2254
- if (!skipUbs && filesTouched.length > 0) {
2255
- const ubsResult = await runUbsScan(filesTouched);
2256
- if (ubsResult) {
2257
- const ubsStep: VerificationStep = {
2258
- name: "ubs_scan",
2259
- command: `ubs scan ${filesTouched.join(" ")}`,
2260
- passed: ubsResult.summary.critical === 0,
2261
- exitCode: ubsResult.exitCode,
2262
- };
2263
-
2264
- if (!ubsStep.passed) {
2265
- ubsStep.error = `Found ${ubsResult.summary.critical} critical bugs`;
2266
- blockers.push(
2267
- `UBS found ${ubsResult.summary.critical} critical bug(s). Try: Run 'ubs scan ${filesTouched.join(" ")}' to see details, fix critical bugs in reported files, or use skip_ubs_scan=true to bypass (not recommended).`,
2268
- );
2269
- }
2270
-
2271
- steps.push(ubsStep);
2272
- } else {
2273
- steps.push({
2274
- name: "ubs_scan",
2275
- command: "ubs scan",
2276
- passed: true,
2277
- exitCode: 0,
2278
- skipped: true,
2279
- skipReason: "UBS not available",
2280
- });
2281
- }
2282
- }
2283
-
2284
- // Step 2: Typecheck
2285
- const typecheckStep = await runTypecheckVerification();
2286
- steps.push(typecheckStep);
2287
- if (!typecheckStep.passed && !typecheckStep.skipped) {
2288
- blockers.push(
2289
- `Typecheck failed: ${typecheckStep.error?.slice(0, 100) || "type errors found"}. Try: Run 'tsc --noEmit' to see full errors, check tsconfig.json configuration, or fix reported type errors in modified files.`,
2290
- );
2291
- }
2292
-
2293
- // Step 3: Tests
2294
- const testStep = await runTestVerification(filesTouched);
2295
- steps.push(testStep);
2296
- if (!testStep.passed && !testStep.skipped) {
2297
- blockers.push(
2298
- `Tests failed: ${testStep.error?.slice(0, 100) || "test failures"}. Try: Run 'bun test ${testStep.command.split(" ").slice(2).join(" ")}' to see full output, check test assertions, or fix failing tests in modified files.`,
2299
- );
2300
- }
2301
-
2302
- // Build summary
2303
- const passedCount = steps.filter((s) => s.passed).length;
2304
- const skippedCount = steps.filter((s) => s.skipped).length;
2305
- const failedCount = steps.filter((s) => !s.passed && !s.skipped).length;
2306
-
2307
- const summary =
2308
- failedCount === 0
2309
- ? `Verification passed: ${passedCount} checks passed, ${skippedCount} skipped`
2310
- : `Verification FAILED: ${failedCount} checks failed, ${passedCount} passed, ${skippedCount} skipped`;
2311
-
2312
- return {
2313
- passed: failedCount === 0,
2314
- steps,
2315
- summary,
2316
- blockers,
2317
- };
2318
- }
2319
-
2320
- /**
2321
- * Run UBS scan on files before completion
2322
- *
2323
- * @param files - Files to scan
2324
- * @returns Scan result or null if UBS not available
2325
- */
2326
- async function runUbsScan(files: string[]): Promise<UbsScanResult | null> {
2327
- if (files.length === 0) {
2328
- return null;
2329
- }
2330
-
2331
- // Check if UBS is available first
2332
- const ubsAvailable = await isToolAvailable("ubs");
2333
- if (!ubsAvailable) {
2334
- warnMissingTool("ubs");
2335
- return null;
2336
- }
2337
-
2338
- try {
2339
- // Run UBS scan with JSON output
2340
- const result = await Bun.$`ubs scan ${files.join(" ")} --json`
2341
- .quiet()
2342
- .nothrow();
2343
-
2344
- const output = result.stdout.toString();
2345
- if (!output.trim()) {
2346
- return {
2347
- exitCode: result.exitCode,
2348
- bugs: [],
2349
- summary: { total: 0, critical: 0, high: 0, medium: 0, low: 0 },
2350
- };
2351
- }
2352
-
2353
- try {
2354
- const parsed = JSON.parse(output);
2355
-
2356
- // Basic validation of structure
2357
- if (typeof parsed !== "object" || parsed === null) {
2358
- throw new Error("UBS output is not an object");
2359
- }
2360
- if (!Array.isArray(parsed.bugs)) {
2361
- console.warn("[swarm] UBS output missing bugs array, using empty");
2362
- }
2363
- if (typeof parsed.summary !== "object" || parsed.summary === null) {
2364
- console.warn("[swarm] UBS output missing summary object, using empty");
2365
- }
2366
-
2367
- return {
2368
- exitCode: result.exitCode,
2369
- bugs: Array.isArray(parsed.bugs) ? parsed.bugs : [],
2370
- summary: parsed.summary || {
2371
- total: 0,
2372
- critical: 0,
2373
- high: 0,
2374
- medium: 0,
2375
- low: 0,
2376
- },
2377
- };
2378
- } catch (error) {
2379
- // UBS output wasn't JSON - this is an error condition
2380
- console.error(
2381
- `[swarm] CRITICAL: UBS scan failed to parse JSON output because output is malformed:`,
2382
- error,
2383
- );
2384
- console.error(
2385
- `[swarm] Raw output: ${output}. Try: Run 'ubs doctor' to check installation, verify UBS version with 'ubs --version' (need v1.0.0+), or check if UBS supports --json flag.`,
2386
- );
2387
- return {
2388
- exitCode: result.exitCode,
2389
- bugs: [],
2390
- summary: { total: 0, critical: 0, high: 0, medium: 0, low: 0 },
2391
- };
2392
- }
2393
- } catch {
2394
- return null;
2395
- }
2396
- }
2397
-
2398
- /**
2399
- * Broadcast context updates to all agents in the epic
2400
- *
2401
- * Enables mid-task coordination by sharing discoveries, warnings, or blockers
2402
- * with all agents working on the same epic. Agents can broadcast without
2403
- * waiting for task completion.
2404
- *
2405
- * Based on "Patterns for Building AI Agents" p.31: "Ensure subagents can share context along the way"
2406
- */
2407
- export const swarm_broadcast = tool({
2408
- description:
2409
- "Broadcast context update to all agents working on the same epic",
2410
- args: {
2411
- project_path: tool.schema
2412
- .string()
2413
- .describe("Absolute path to project root"),
2414
- agent_name: tool.schema
2415
- .string()
2416
- .describe("Name of the agent broadcasting the message"),
2417
- epic_id: tool.schema.string().describe("Epic ID (e.g., bd-abc123)"),
2418
- message: tool.schema
2419
- .string()
2420
- .describe("Context update to share (what changed, what was learned)"),
2421
- importance: tool.schema
2422
- .enum(["info", "warning", "blocker"])
2423
- .default("info")
2424
- .describe("Priority level (default: info)"),
2425
- files_affected: tool.schema
2426
- .array(tool.schema.string())
2427
- .optional()
2428
- .describe("Files this context relates to"),
2429
- },
2430
- async execute(args, ctx) {
2431
- // Extract bead_id from context if available (for traceability)
2432
- const beadId = (ctx as { beadId?: string }).beadId || "unknown";
2433
-
2434
- // Format the broadcast message
2435
- const body = [
2436
- `## Context Update`,
2437
- "",
2438
- `**From**: ${args.agent_name} (${beadId})`,
2439
- `**Priority**: ${args.importance.toUpperCase()}`,
2440
- "",
2441
- args.message,
2442
- "",
2443
- args.files_affected && args.files_affected.length > 0
2444
- ? `**Files affected**:\n${args.files_affected.map((f) => `- \`${f}\``).join("\n")}`
2445
- : "",
2446
- ]
2447
- .filter(Boolean)
2448
- .join("\n");
2449
-
2450
- // Map importance to Agent Mail importance
2451
- const mailImportance =
2452
- args.importance === "blocker"
2453
- ? "urgent"
2454
- : args.importance === "warning"
2455
- ? "high"
2456
- : "normal";
2457
-
2458
- // Send as broadcast to thread using embedded swarm-mail
2459
- await sendSwarmMessage({
2460
- projectPath: args.project_path,
2461
- fromAgent: args.agent_name,
2462
- toAgents: [], // Broadcast to thread
2463
- subject: `[${args.importance.toUpperCase()}] Context update from ${args.agent_name}`,
2464
- body,
2465
- threadId: args.epic_id,
2466
- importance: mailImportance,
2467
- ackRequired: args.importance === "blocker",
2468
- });
2469
-
2470
- return JSON.stringify(
2471
- {
2472
- broadcast: true,
2473
- epic_id: args.epic_id,
2474
- from: args.agent_name,
2475
- bead_id: beadId,
2476
- importance: args.importance,
2477
- recipients: "all agents in epic",
2478
- ack_required: args.importance === "blocker",
2479
- },
2480
- null,
2481
- 2,
2482
- );
2483
- },
2484
- });
2485
-
2486
- /**
2487
- * Mark a subtask as complete
2488
- *
2489
- * Implements the Verification Gate (from superpowers):
2490
- * 1. IDENTIFY: What commands prove this claim?
2491
- * 2. RUN: Execute verification (UBS, typecheck, tests)
2492
- * 3. READ: Check exit codes and output
2493
- * 4. VERIFY: All checks must pass
2494
- * 5. ONLY THEN: Close the bead
2495
- *
2496
- * Closes bead, releases reservations, notifies coordinator.
2497
- */
2498
- export const swarm_complete = tool({
2499
- description:
2500
- "Mark subtask complete with Verification Gate. Runs UBS scan, typecheck, and tests before allowing completion.",
2501
- args: {
2502
- project_key: tool.schema.string().describe("Project path"),
2503
- agent_name: tool.schema.string().describe("Your Agent Mail name"),
2504
- bead_id: tool.schema.string().describe("Subtask bead ID"),
2505
- summary: tool.schema.string().describe("Brief summary of work done"),
2506
- evaluation: tool.schema
2507
- .string()
2508
- .optional()
2509
- .describe("Self-evaluation JSON (Evaluation schema)"),
2510
- files_touched: tool.schema
2511
- .array(tool.schema.string())
2512
- .optional()
2513
- .describe("Files modified - will be verified (UBS, typecheck, tests)"),
2514
- skip_ubs_scan: tool.schema
2515
- .boolean()
2516
- .optional()
2517
- .describe("Skip UBS bug scan (default: false)"),
2518
- skip_verification: tool.schema
2519
- .boolean()
2520
- .optional()
2521
- .describe(
2522
- "Skip ALL verification (UBS, typecheck, tests). Use sparingly! (default: false)",
2523
- ),
2524
- },
2525
- async execute(args) {
2526
- // Run Verification Gate unless explicitly skipped
2527
- let verificationResult: VerificationGateResult | null = null;
2528
-
2529
- if (!args.skip_verification && args.files_touched?.length) {
2530
- verificationResult = await runVerificationGate(
2531
- args.files_touched,
2532
- args.skip_ubs_scan ?? false,
2533
- );
2534
-
2535
- // Block completion if verification failed
2536
- if (!verificationResult.passed) {
2537
- return JSON.stringify(
2538
- {
2539
- success: false,
2540
- error: "Verification Gate FAILED - fix issues before completing",
2541
- verification: {
2542
- passed: false,
2543
- summary: verificationResult.summary,
2544
- blockers: verificationResult.blockers,
2545
- steps: verificationResult.steps.map((s) => ({
2546
- name: s.name,
2547
- passed: s.passed,
2548
- skipped: s.skipped,
2549
- skipReason: s.skipReason,
2550
- error: s.error?.slice(0, 200),
2551
- })),
2552
- },
2553
- hint:
2554
- verificationResult.blockers.length > 0
2555
- ? `Fix these issues: ${verificationResult.blockers.map((b, i) => `${i + 1}. ${b}`).join(", ")}. Use skip_verification=true only as last resort.`
2556
- : "Fix the failing checks and try again. Use skip_verification=true only as last resort.",
2557
- gate_function:
2558
- "IDENTIFY → RUN → READ → VERIFY → CLAIM (you are at VERIFY, claim blocked)",
2559
- },
2560
- null,
2561
- 2,
2562
- );
2563
- }
2564
- }
2565
-
2566
- // Legacy UBS-only path for backward compatibility (when no files_touched)
2567
- let ubsResult: UbsScanResult | null = null;
2568
- if (
2569
- !args.skip_verification &&
2570
- !verificationResult &&
2571
- args.files_touched?.length &&
2572
- !args.skip_ubs_scan
2573
- ) {
2574
- ubsResult = await runUbsScan(args.files_touched);
2575
-
2576
- // Block completion if critical bugs found
2577
- if (ubsResult && ubsResult.summary.critical > 0) {
2578
- return JSON.stringify(
2579
- {
2580
- success: false,
2581
- error: `UBS found ${ubsResult.summary.critical} critical bug(s) that must be fixed before completing`,
2582
- ubs_scan: {
2583
- critical_count: ubsResult.summary.critical,
2584
- bugs: ubsResult.bugs.filter((b) => b.severity === "critical"),
2585
- },
2586
- hint: `Fix these critical bugs: ${ubsResult.bugs
2587
- .filter((b) => b.severity === "critical")
2588
- .map((b) => `${b.file}:${b.line} - ${b.message}`)
2589
- .slice(0, 3)
2590
- .join(
2591
- "; ",
2592
- )}. Try: Run 'ubs scan ${args.files_touched?.join(" ") || "."} --json' for full report, fix reported issues, or use skip_ubs_scan=true to bypass (not recommended).`,
2593
- },
2594
- null,
2595
- 2,
2596
- );
2597
- }
2598
- }
2599
-
2600
- // Parse and validate evaluation if provided
2601
- let parsedEvaluation: Evaluation | undefined;
2602
- if (args.evaluation) {
2603
- try {
2604
- parsedEvaluation = EvaluationSchema.parse(JSON.parse(args.evaluation));
2605
- } catch (error) {
2606
- return JSON.stringify(
2607
- {
2608
- success: false,
2609
- error: "Invalid evaluation format",
2610
- details: error instanceof z.ZodError ? error.issues : String(error),
2611
- },
2612
- null,
2613
- 2,
2614
- );
2615
- }
2616
-
2617
- // If evaluation failed, don't complete
2618
- if (!parsedEvaluation.passed) {
2619
- return JSON.stringify(
2620
- {
2621
- success: false,
2622
- error: "Self-evaluation failed",
2623
- retry_suggestion: parsedEvaluation.retry_suggestion,
2624
- feedback: parsedEvaluation.overall_feedback,
2625
- },
2626
- null,
2627
- 2,
2628
- );
2629
- }
2630
- }
2631
-
2632
- // Close the bead
2633
- const closeResult =
2634
- await Bun.$`bd close ${args.bead_id} --reason ${args.summary} --json`
2635
- .quiet()
2636
- .nothrow();
2637
-
2638
- if (closeResult.exitCode !== 0) {
2639
- throw new SwarmError(
2640
- `Failed to close bead because bd close command failed: ${closeResult.stderr.toString()}. Try: Verify bead exists and is not already closed with 'bd show ${args.bead_id}', check if bead ID is correct with 'beads_query()', or use beads_close tool directly.`,
2641
- "complete",
2642
- );
2643
- }
2644
-
2645
- // Release file reservations for this agent using embedded swarm-mail
2646
- try {
2647
- await releaseSwarmFiles({
2648
- projectPath: args.project_key,
2649
- agentName: args.agent_name,
2650
- // Release all reservations for this agent
2651
- });
2652
- } catch (error) {
2653
- // Release might fail (e.g., no reservations existed)
2654
- // This is non-fatal - log and continue
2655
- console.warn(
2656
- `[swarm] Failed to release file reservations for ${args.agent_name}:`,
2657
- error,
2658
- );
2659
- }
2660
-
2661
- // Extract epic ID
2662
- const epicId = args.bead_id.includes(".")
2663
- ? args.bead_id.split(".")[0]
2664
- : args.bead_id;
2665
-
2666
- // Send completion message using embedded swarm-mail
2667
- const completionBody = [
2668
- `## Subtask Complete: ${args.bead_id}`,
2669
- "",
2670
- `**Summary**: ${args.summary}`,
2671
- "",
2672
- parsedEvaluation
2673
- ? `**Self-Evaluation**: ${parsedEvaluation.passed ? "PASSED" : "FAILED"}`
2674
- : "",
2675
- parsedEvaluation?.overall_feedback
2676
- ? `**Feedback**: ${parsedEvaluation.overall_feedback}`
2677
- : "",
2678
- ]
2679
- .filter(Boolean)
2680
- .join("\n");
2681
-
2682
- await sendSwarmMessage({
2683
- projectPath: args.project_key,
2684
- fromAgent: args.agent_name,
2685
- toAgents: [], // Thread broadcast
2686
- subject: `Complete: ${args.bead_id}`,
2687
- body: completionBody,
2688
- threadId: epicId,
2689
- importance: "normal",
2690
- });
2691
-
2692
- // Build success response with semantic-memory integration
2693
- const response = {
2694
- success: true,
2695
- bead_id: args.bead_id,
2696
- closed: true,
2697
- reservations_released: true,
2698
- message_sent: true,
2699
- verification_gate: verificationResult
2700
- ? {
2701
- passed: true,
2702
- summary: verificationResult.summary,
2703
- steps: verificationResult.steps.map((s) => ({
2704
- name: s.name,
2705
- passed: s.passed,
2706
- skipped: s.skipped,
2707
- skipReason: s.skipReason,
2708
- })),
2709
- }
2710
- : args.skip_verification
2711
- ? { skipped: true, reason: "skip_verification=true" }
2712
- : { skipped: true, reason: "no files_touched provided" },
2713
- ubs_scan: ubsResult
2714
- ? {
2715
- ran: true,
2716
- bugs_found: ubsResult.summary.total,
2717
- summary: ubsResult.summary,
2718
- warnings: ubsResult.bugs.filter((b) => b.severity !== "critical"),
2719
- }
2720
- : verificationResult
2721
- ? { ran: true, included_in_verification_gate: true }
2722
- : {
2723
- ran: false,
2724
- reason: args.skip_ubs_scan
2725
- ? "skipped"
2726
- : "no files or ubs unavailable",
2727
- },
2728
- learning_prompt: `## Reflection
2729
-
2730
- Did you learn anything reusable during this subtask? Consider:
2731
-
2732
- 1. **Patterns**: Any code patterns or approaches that worked well?
2733
- 2. **Gotchas**: Edge cases or pitfalls to warn future agents about?
2734
- 3. **Best Practices**: Domain-specific guidelines worth documenting?
2735
- 4. **Tool Usage**: Effective ways to use tools for this type of task?
2736
-
2737
- If you discovered something valuable, use \`swarm_learn\` or \`skills_create\` to preserve it as a skill for future swarms.
2738
-
2739
- Files touched: ${args.files_touched?.join(", ") || "none recorded"}`,
2740
- // Add semantic-memory integration on success
2741
- memory_store: formatMemoryStoreOnSuccess(
2742
- args.bead_id,
2743
- args.summary,
2744
- args.files_touched || [],
2745
- ),
2746
- };
2747
-
2748
- return JSON.stringify(response, null, 2);
2749
- },
2750
- });
2751
-
2752
- /**
2753
- * Classify failure based on error message heuristics
2754
- *
2755
- * Simple pattern matching to categorize why a task failed.
2756
- * Used when failure_mode is not explicitly provided.
2757
- *
2758
- * @param error - Error object or message
2759
- * @returns FailureMode classification
2760
- */
2761
- function classifyFailure(error: Error | string): string {
2762
- const msg = (typeof error === "string" ? error : error.message).toLowerCase();
2763
-
2764
- if (msg.includes("timeout")) return "timeout";
2765
- if (msg.includes("conflict") || msg.includes("reservation"))
2766
- return "conflict";
2767
- if (msg.includes("validation") || msg.includes("schema")) return "validation";
2768
- if (msg.includes("context") || msg.includes("token"))
2769
- return "context_overflow";
2770
- if (msg.includes("blocked") || msg.includes("dependency"))
2771
- return "dependency_blocked";
2772
- if (msg.includes("cancel")) return "user_cancelled";
2773
-
2774
- // Check for tool failure patterns
2775
- if (
2776
- msg.includes("tool") ||
2777
- msg.includes("command") ||
2778
- msg.includes("failed to execute")
2779
- ) {
2780
- return "tool_failure";
2781
- }
2782
-
2783
- return "unknown";
2784
- }
2785
-
2786
- /**
2787
- * Record outcome signals from a completed subtask
2788
- *
2789
- * Tracks implicit feedback (duration, errors, retries) to score
2790
- * decomposition quality over time. This data feeds into criterion
2791
- * weight calculations.
2792
- *
2793
- * Strategy tracking enables learning about which decomposition strategies
2794
- * work best for different task types.
2795
- *
2796
- * @see src/learning.ts for scoring logic
2797
- */
2798
- export const swarm_record_outcome = tool({
2799
- description:
2800
- "Record subtask outcome for implicit feedback scoring. Tracks duration, errors, retries to learn decomposition quality.",
2801
- args: {
2802
- bead_id: tool.schema.string().describe("Subtask bead ID"),
2803
- duration_ms: tool.schema
2804
- .number()
2805
- .int()
2806
- .min(0)
2807
- .describe("Duration in milliseconds"),
2808
- error_count: tool.schema
2809
- .number()
2810
- .int()
2811
- .min(0)
2812
- .default(0)
2813
- .describe("Number of errors encountered"),
2814
- retry_count: tool.schema
2815
- .number()
2816
- .int()
2817
- .min(0)
2818
- .default(0)
2819
- .describe("Number of retry attempts"),
2820
- success: tool.schema.boolean().describe("Whether the subtask succeeded"),
2821
- files_touched: tool.schema
2822
- .array(tool.schema.string())
2823
- .optional()
2824
- .describe("Files that were modified"),
2825
- criteria: tool.schema
2826
- .array(tool.schema.string())
2827
- .optional()
2828
- .describe(
2829
- "Criteria to generate feedback for (default: all default criteria)",
2830
- ),
2831
- strategy: tool.schema
2832
- .enum(["file-based", "feature-based", "risk-based", "research-based"])
2833
- .optional()
2834
- .describe("Decomposition strategy used for this task"),
2835
- failure_mode: tool.schema
2836
- .enum([
2837
- "timeout",
2838
- "conflict",
2839
- "validation",
2840
- "tool_failure",
2841
- "context_overflow",
2842
- "dependency_blocked",
2843
- "user_cancelled",
2844
- "unknown",
2845
- ])
2846
- .optional()
2847
- .describe(
2848
- "Failure classification (only when success=false). Auto-classified if not provided.",
2849
- ),
2850
- failure_details: tool.schema
2851
- .string()
2852
- .optional()
2853
- .describe("Detailed failure context (error message, stack trace, etc.)"),
2854
- },
2855
- async execute(args) {
2856
- // Build outcome signals
2857
- const signals: OutcomeSignals = {
2858
- bead_id: args.bead_id,
2859
- duration_ms: args.duration_ms,
2860
- error_count: args.error_count ?? 0,
2861
- retry_count: args.retry_count ?? 0,
2862
- success: args.success,
2863
- files_touched: args.files_touched ?? [],
2864
- timestamp: new Date().toISOString(),
2865
- strategy: args.strategy as LearningDecompositionStrategy | undefined,
2866
- failure_mode: args.failure_mode,
2867
- failure_details: args.failure_details,
2868
- };
2869
-
2870
- // If task failed but no failure_mode provided, try to classify from failure_details
2871
- if (!args.success && !args.failure_mode && args.failure_details) {
2872
- signals.failure_mode = classifyFailure(args.failure_details) as any;
2873
- }
2874
-
2875
- // Validate signals
2876
- const validated = OutcomeSignalsSchema.parse(signals);
2877
-
2878
- // Score the outcome
2879
- const scored: ScoredOutcome = scoreImplicitFeedback(
2880
- validated,
2881
- DEFAULT_LEARNING_CONFIG,
2882
- );
2883
-
2884
- // Get error patterns from accumulator
2885
- const errorStats = await globalErrorAccumulator.getErrorStats(args.bead_id);
2886
-
2887
- // Generate feedback events for each criterion
2888
- const criteriaToScore = args.criteria ?? [
2889
- "type_safe",
2890
- "no_bugs",
2891
- "patterns",
2892
- "readable",
2893
- ];
2894
- const feedbackEvents: FeedbackEvent[] = criteriaToScore.map((criterion) => {
2895
- const event = outcomeToFeedback(scored, criterion);
2896
- // Include strategy in feedback context for future analysis
2897
- if (args.strategy) {
2898
- event.context =
2899
- `${event.context || ""} [strategy: ${args.strategy}]`.trim();
2900
- }
2901
- // Include error patterns in feedback context
2902
- if (errorStats.total > 0) {
2903
- const errorSummary = Object.entries(errorStats.by_type)
2904
- .map(([type, count]) => `${type}:${count}`)
2905
- .join(", ");
2906
- event.context =
2907
- `${event.context || ""} [errors: ${errorSummary}]`.trim();
2908
- }
2909
- return event;
2910
- });
2911
-
2912
- return JSON.stringify(
2913
- {
2914
- success: true,
2915
- outcome: {
2916
- signals: validated,
2917
- scored: {
2918
- type: scored.type,
2919
- decayed_value: scored.decayed_value,
2920
- reasoning: scored.reasoning,
2921
- },
2922
- },
2923
- feedback_events: feedbackEvents,
2924
- error_patterns: errorStats,
2925
- summary: {
2926
- feedback_type: scored.type,
2927
- duration_seconds: Math.round(args.duration_ms / 1000),
2928
- error_count: args.error_count ?? 0,
2929
- retry_count: args.retry_count ?? 0,
2930
- success: args.success,
2931
- strategy: args.strategy,
2932
- failure_mode: validated.failure_mode,
2933
- failure_details: validated.failure_details,
2934
- accumulated_errors: errorStats.total,
2935
- unresolved_errors: errorStats.unresolved,
2936
- },
2937
- note: "Feedback events should be stored for criterion weight calculation. Use learning.ts functions to apply weights.",
2938
- },
2939
- null,
2940
- 2,
2941
- );
2942
- },
2943
- });
2944
-
2945
- /**
2946
- * Generate subtask prompt for a spawned agent
2947
- */
2948
- export const swarm_subtask_prompt = tool({
2949
- description: "Generate the prompt for a spawned subtask agent",
2950
- args: {
2951
- agent_name: tool.schema.string().describe("Agent Mail name for the agent"),
2952
- bead_id: tool.schema.string().describe("Subtask bead ID"),
2953
- epic_id: tool.schema.string().describe("Epic bead ID"),
2954
- subtask_title: tool.schema.string().describe("Subtask title"),
2955
- subtask_description: tool.schema
2956
- .string()
2957
- .optional()
2958
- .describe("Detailed subtask instructions"),
2959
- files: tool.schema
2960
- .array(tool.schema.string())
2961
- .describe("Files assigned to this subtask"),
2962
- shared_context: tool.schema
2963
- .string()
2964
- .optional()
2965
- .describe("Context shared across all agents"),
2966
- },
2967
- async execute(args) {
2968
- const prompt = formatSubtaskPrompt({
2969
- agent_name: args.agent_name,
2970
- bead_id: args.bead_id,
2971
- epic_id: args.epic_id,
2972
- subtask_title: args.subtask_title,
2973
- subtask_description: args.subtask_description || "",
2974
- files: args.files,
2975
- shared_context: args.shared_context,
2976
- });
2977
-
2978
- return prompt;
2979
- },
2980
- });
2981
-
2982
- /**
2983
- * Prepare a subtask for spawning with Task tool (V2 prompt)
2984
- *
2985
- * Generates a streamlined prompt that tells agents to USE Agent Mail and beads.
2986
- * Returns JSON that can be directly used with Task tool.
2987
- */
2988
- export const swarm_spawn_subtask = tool({
2989
- description:
2990
- "Prepare a subtask for spawning. Returns prompt with Agent Mail/beads instructions.",
2991
- args: {
2992
- bead_id: tool.schema.string().describe("Subtask bead ID"),
2993
- epic_id: tool.schema.string().describe("Parent epic bead ID"),
2994
- subtask_title: tool.schema.string().describe("Subtask title"),
2995
- subtask_description: tool.schema
2996
- .string()
2997
- .optional()
2998
- .describe("Detailed subtask instructions"),
2999
- files: tool.schema
3000
- .array(tool.schema.string())
3001
- .describe("Files assigned to this subtask"),
3002
- shared_context: tool.schema
3003
- .string()
3004
- .optional()
3005
- .describe("Context shared across all agents"),
3006
- },
3007
- async execute(args) {
3008
- const prompt = formatSubtaskPromptV2({
3009
- bead_id: args.bead_id,
3010
- epic_id: args.epic_id,
3011
- subtask_title: args.subtask_title,
3012
- subtask_description: args.subtask_description || "",
3013
- files: args.files,
3014
- shared_context: args.shared_context,
3015
- });
3016
-
3017
- return JSON.stringify(
3018
- {
3019
- prompt,
3020
- bead_id: args.bead_id,
3021
- epic_id: args.epic_id,
3022
- files: args.files,
3023
- },
3024
- null,
3025
- 2,
3026
- );
3027
- },
3028
- });
3029
-
3030
- /**
3031
- * Schema for task agent result
3032
- */
3033
- const TaskResultSchema = z.object({
3034
- success: z.boolean(),
3035
- summary: z.string(),
3036
- files_modified: z.array(z.string()).optional().default([]),
3037
- files_created: z.array(z.string()).optional().default([]),
3038
- issues_found: z.array(z.string()).optional().default([]),
3039
- tests_passed: z.boolean().optional(),
3040
- notes: z.string().optional(),
3041
- blocker: z.string().optional(),
3042
- suggestions: z.array(z.string()).optional(),
3043
- });
3044
-
3045
- type TaskResult = z.infer<typeof TaskResultSchema>;
3046
-
3047
- /**
3048
- * Handle subtask completion from a Task agent
3049
- *
3050
- * This tool is for coordinators to process the result after a Task subagent
3051
- * returns. It parses the JSON result, closes the bead on success, and
3052
- * creates new beads for any issues discovered.
3053
- *
3054
- * @example
3055
- * // Task agent returns JSON:
3056
- * // { "success": true, "summary": "Added auth", "files_modified": ["src/auth.ts"], "issues_found": ["Missing tests"] }
3057
- * //
3058
- * // Coordinator calls:
3059
- * swarm_complete_subtask(bead_id="bd-123.1", task_result=<agent_response>)
3060
- */
3061
- export const swarm_complete_subtask = tool({
3062
- description:
3063
- "Handle subtask completion after Task agent returns. Parses result JSON, closes bead on success, creates new beads for issues found.",
3064
- args: {
3065
- bead_id: z.string().describe("Subtask bead ID to close"),
3066
- task_result: z
3067
- .string()
3068
- .describe("JSON result from the Task agent (TaskResult schema)"),
3069
- files_touched: z
3070
- .array(z.string())
3071
- .optional()
3072
- .describe(
3073
- "Override files touched (uses task_result.files_modified if not provided)",
3074
- ),
3075
- },
3076
- async execute(args) {
3077
- // Parse the task result JSON
3078
- let result: TaskResult;
3079
- try {
3080
- const parsed = JSON.parse(args.task_result);
3081
- result = TaskResultSchema.parse(parsed);
3082
- } catch (error) {
3083
- // Handle parse errors gracefully
3084
- const errorMessage =
3085
- error instanceof SyntaxError
3086
- ? `Invalid JSON: ${error.message}`
3087
- : error instanceof z.ZodError
3088
- ? `Schema validation failed: ${error.issues.map((i) => i.message).join(", ")}`
3089
- : String(error);
3090
-
3091
- return JSON.stringify(
3092
- {
3093
- success: false,
3094
- error: "Failed to parse task result",
3095
- details: errorMessage,
3096
- hint: "Task agent should return JSON matching TaskResult schema: { success, summary, files_modified?, issues_found?, ... }",
3097
- },
3098
- null,
3099
- 2,
3100
- );
3101
- }
3102
-
3103
- const filesTouched = args.files_touched ?? [
3104
- ...result.files_modified,
3105
- ...result.files_created,
3106
- ];
3107
- const issuesCreated: Array<{ title: string; id?: string }> = [];
3108
-
3109
- // If task failed, don't close the bead - return info for coordinator to handle
3110
- if (!result.success) {
3111
- return JSON.stringify(
3112
- {
3113
- success: false,
3114
- bead_id: args.bead_id,
3115
- task_failed: true,
3116
- summary: result.summary,
3117
- blocker: result.blocker,
3118
- suggestions: result.suggestions,
3119
- files_touched: filesTouched,
3120
- action_needed:
3121
- "Task failed - review blocker and decide whether to retry or close as failed",
3122
- },
3123
- null,
3124
- 2,
3125
- );
3126
- }
3127
-
3128
- // Task succeeded - close the bead
3129
- const closeReason = result.summary.slice(0, 200); // Truncate for safety
3130
- await Bun.$`bd close ${args.bead_id} -r "${closeReason}"`.quiet().nothrow();
3131
-
3132
- // Create new beads for each issue found
3133
- if (result.issues_found.length > 0) {
3134
- for (const issue of result.issues_found) {
3135
- const issueTitle = issue.slice(0, 100); // Truncate long titles
3136
- const createResult = await Bun.$`bd create "${issueTitle}" -t bug`
3137
- .quiet()
3138
- .nothrow();
3139
-
3140
- if (createResult.exitCode === 0) {
3141
- // Try to parse the bead ID from output
3142
- const output = createResult.stdout.toString();
3143
- const idMatch = output.match(/bd-[a-z0-9]+/);
3144
- issuesCreated.push({
3145
- title: issueTitle,
3146
- id: idMatch?.[0],
3147
- });
3148
- } else {
3149
- issuesCreated.push({
3150
- title: issueTitle,
3151
- id: undefined, // Failed to create
3152
- });
3153
- }
3154
- }
3155
- }
3156
-
3157
- return JSON.stringify(
3158
- {
3159
- success: true,
3160
- bead_id: args.bead_id,
3161
- bead_closed: true,
3162
- summary: result.summary,
3163
- files_touched: filesTouched,
3164
- tests_passed: result.tests_passed,
3165
- notes: result.notes,
3166
- issues_created: issuesCreated.length > 0 ? issuesCreated : undefined,
3167
- issues_count: issuesCreated.length,
3168
- },
3169
- null,
3170
- 2,
3171
- );
3172
- },
3173
- });
3174
-
3175
- /**
3176
- * Generate self-evaluation prompt
3177
- */
3178
- export const swarm_evaluation_prompt = tool({
3179
- description: "Generate self-evaluation prompt for a completed subtask",
3180
- args: {
3181
- bead_id: tool.schema.string().describe("Subtask bead ID"),
3182
- subtask_title: tool.schema.string().describe("Subtask title"),
3183
- files_touched: tool.schema
3184
- .array(tool.schema.string())
3185
- .describe("Files that were modified"),
3186
- },
3187
- async execute(args) {
3188
- const prompt = formatEvaluationPrompt({
3189
- bead_id: args.bead_id,
3190
- subtask_title: args.subtask_title,
3191
- files_touched: args.files_touched,
3192
- });
3193
-
3194
- return JSON.stringify(
3195
- {
3196
- prompt,
3197
- expected_schema: "Evaluation",
3198
- schema_hint: {
3199
- passed: "boolean",
3200
- criteria: {
3201
- type_safe: { passed: "boolean", feedback: "string" },
3202
- no_bugs: { passed: "boolean", feedback: "string" },
3203
- patterns: { passed: "boolean", feedback: "string" },
3204
- readable: { passed: "boolean", feedback: "string" },
3205
- },
3206
- overall_feedback: "string",
3207
- retry_suggestion: "string | null",
3208
- },
3209
- },
3210
- null,
3211
- 2,
3212
- );
3213
- },
3214
- });
3215
-
3216
- // ============================================================================
3217
- // Swarm Learning
3218
- // ============================================================================
3219
-
3220
- /**
3221
- * Learn from completed work and optionally create a skill
3222
- *
3223
- * This tool helps agents reflect on patterns, best practices, or domain
3224
- * knowledge discovered during task execution and codify them into reusable
3225
- * skills for future swarms.
3226
- *
3227
- * Implements the "learning swarm" pattern where swarms get smarter over time.
3228
- */
3229
- export const swarm_learn = tool({
3230
- description: `Analyze completed work and optionally create a skill from learned patterns.
3231
-
3232
- Use after completing a subtask when you've discovered:
3233
- - Reusable code patterns or approaches
3234
- - Domain-specific best practices
3235
- - Gotchas or edge cases to warn about
3236
- - Effective tool usage patterns
3237
-
3238
- This tool helps you formalize learnings into a skill that future agents can discover and use.`,
3239
- args: {
3240
- summary: tool.schema
3241
- .string()
3242
- .describe("Brief summary of what was learned (1-2 sentences)"),
3243
- pattern_type: tool.schema
3244
- .enum([
3245
- "code-pattern",
3246
- "best-practice",
3247
- "gotcha",
3248
- "tool-usage",
3249
- "domain-knowledge",
3250
- "workflow",
3251
- ])
3252
- .describe("Category of the learning"),
3253
- details: tool.schema
3254
- .string()
3255
- .describe("Detailed explanation of the pattern or practice"),
3256
- example: tool.schema
3257
- .string()
3258
- .optional()
3259
- .describe("Code example or concrete illustration"),
3260
- when_to_use: tool.schema
3261
- .string()
3262
- .describe("When should an agent apply this knowledge?"),
3263
- files_context: tool.schema
3264
- .array(tool.schema.string())
3265
- .optional()
3266
- .describe("Files that exemplify this pattern"),
3267
- create_skill: tool.schema
3268
- .boolean()
3269
- .optional()
3270
- .describe(
3271
- "Create a skill from this learning (default: false, just document)",
3272
- ),
3273
- skill_name: tool.schema
3274
- .string()
3275
- .regex(/^[a-z0-9-]+$/)
3276
- .max(64)
3277
- .optional()
3278
- .describe("Skill name if creating (required if create_skill=true)"),
3279
- skill_tags: tool.schema
3280
- .array(tool.schema.string())
3281
- .optional()
3282
- .describe("Tags for the skill if creating"),
3283
- },
3284
- async execute(args) {
3285
- // Format the learning as structured documentation
3286
- const learning = {
3287
- summary: args.summary,
3288
- type: args.pattern_type,
3289
- details: args.details,
3290
- example: args.example,
3291
- when_to_use: args.when_to_use,
3292
- files_context: args.files_context,
3293
- recorded_at: new Date().toISOString(),
3294
- };
3295
-
3296
- // If creating a skill, generate and create it
3297
- if (args.create_skill) {
3298
- if (!args.skill_name) {
3299
- return JSON.stringify(
3300
- {
3301
- success: false,
3302
- error: "skill_name is required when create_skill=true",
3303
- learning: learning,
3304
- },
3305
- null,
3306
- 2,
3307
- );
3308
- }
3309
-
3310
- // Build skill body from learning
3311
- const skillBody = `# ${args.summary}
3312
-
3313
- ## When to Use
3314
- ${args.when_to_use}
3315
-
3316
- ## ${args.pattern_type.replace(/-/g, " ").replace(/\b\w/g, (c) => c.toUpperCase())}
3317
-
3318
- ${args.details}
3319
-
3320
- ${args.example ? `## Example\n\n\`\`\`\n${args.example}\n\`\`\`\n` : ""}
3321
- ${args.files_context && args.files_context.length > 0 ? `## Reference Files\n\n${args.files_context.map((f) => `- \`${f}\``).join("\n")}\n` : ""}
3322
-
3323
- ---
3324
- *Learned from swarm execution on ${new Date().toISOString().split("T")[0]}*`;
3325
-
3326
- // Import skills_create functionality
3327
- const { getSkill, invalidateSkillsCache } = await import("./skills");
3328
- const { mkdir, writeFile } = await import("fs/promises");
3329
- const { join } = await import("path");
3330
-
3331
- // Check if skill exists
3332
- const existing = await getSkill(args.skill_name);
3333
- if (existing) {
3334
- return JSON.stringify(
3335
- {
3336
- success: false,
3337
- error: `Skill '${args.skill_name}' already exists`,
3338
- existing_path: existing.path,
3339
- learning: learning,
3340
- suggestion:
3341
- "Use skills_update to add to existing skill, or choose a different name",
3342
- },
3343
- null,
3344
- 2,
3345
- );
3346
- }
3347
-
3348
- // Create skill directory and file
3349
- const skillDir = join(
3350
- process.cwd(),
3351
- ".opencode",
3352
- "skills",
3353
- args.skill_name,
3354
- );
3355
- const skillPath = join(skillDir, "SKILL.md");
3356
-
3357
- const frontmatter = [
3358
- "---",
3359
- `name: ${args.skill_name}`,
3360
- `description: ${args.when_to_use.slice(0, 200)}${args.when_to_use.length > 200 ? "..." : ""}`,
3361
- "tags:",
3362
- ` - ${args.pattern_type}`,
3363
- ` - learned`,
3364
- ...(args.skill_tags || []).map((t) => ` - ${t}`),
3365
- "---",
3366
- ].join("\n");
3367
-
3368
- try {
3369
- await mkdir(skillDir, { recursive: true });
3370
- await writeFile(skillPath, `${frontmatter}\n\n${skillBody}`, "utf-8");
3371
- invalidateSkillsCache();
3372
-
3373
- return JSON.stringify(
3374
- {
3375
- success: true,
3376
- skill_created: true,
3377
- skill: {
3378
- name: args.skill_name,
3379
- path: skillPath,
3380
- type: args.pattern_type,
3381
- },
3382
- learning: learning,
3383
- message: `Created skill '${args.skill_name}' from learned pattern. Future agents can discover it with skills_list.`,
3384
- },
3385
- null,
3386
- 2,
3387
- );
3388
- } catch (error) {
3389
- return JSON.stringify(
3390
- {
3391
- success: false,
3392
- error: `Failed to create skill: ${error instanceof Error ? error.message : String(error)}`,
3393
- learning: learning,
3394
- },
3395
- null,
3396
- 2,
3397
- );
3398
- }
3399
- }
3400
-
3401
- // Just document the learning without creating a skill
3402
- return JSON.stringify(
3403
- {
3404
- success: true,
3405
- skill_created: false,
3406
- learning: learning,
3407
- message:
3408
- "Learning documented. Use create_skill=true to persist as a skill for future agents.",
3409
- suggested_skill_name:
3410
- args.skill_name ||
3411
- args.summary
3412
- .toLowerCase()
3413
- .replace(/[^a-z0-9\s-]/g, "")
3414
- .replace(/\s+/g, "-")
3415
- .slice(0, 64),
3416
- },
3417
- null,
3418
- 2,
3419
- );
3420
- },
3421
- });
3422
-
3423
- // ============================================================================
3424
- // Error Accumulator
3425
- // ============================================================================
3426
-
3427
- /**
3428
- * Global error accumulator for tracking errors across subtasks
3429
- *
3430
- * This is a session-level singleton that accumulates errors during
3431
- * swarm execution for feeding into retry prompts.
3432
- */
3433
- const globalErrorAccumulator = new ErrorAccumulator();
3434
-
3435
- /**
3436
- * Record an error during subtask execution
3437
- *
3438
- * Implements pattern from "Patterns for Building AI Agents" p.40:
3439
- * "Good agents examine and correct errors when something goes wrong"
3440
- *
3441
- * Errors are accumulated and can be fed into retry prompts to help
3442
- * agents learn from past failures.
3443
- */
3444
- export const swarm_accumulate_error = tool({
3445
- description:
3446
- "Record an error during subtask execution. Errors feed into retry prompts.",
3447
- args: {
3448
- bead_id: tool.schema.string().describe("Bead ID where error occurred"),
3449
- error_type: tool.schema
3450
- .enum(["validation", "timeout", "conflict", "tool_failure", "unknown"])
3451
- .describe("Category of error"),
3452
- message: tool.schema.string().describe("Human-readable error message"),
3453
- stack_trace: tool.schema
3454
- .string()
3455
- .optional()
3456
- .describe("Stack trace for debugging"),
3457
- tool_name: tool.schema.string().optional().describe("Tool that failed"),
3458
- context: tool.schema
3459
- .string()
3460
- .optional()
3461
- .describe("What was happening when error occurred"),
3462
- },
3463
- async execute(args) {
3464
- const entry = await globalErrorAccumulator.recordError(
3465
- args.bead_id,
3466
- args.error_type as ErrorType,
3467
- args.message,
3468
- {
3469
- stack_trace: args.stack_trace,
3470
- tool_name: args.tool_name,
3471
- context: args.context,
3472
- },
3473
- );
3474
-
3475
- return JSON.stringify(
3476
- {
3477
- success: true,
3478
- error_id: entry.id,
3479
- bead_id: entry.bead_id,
3480
- error_type: entry.error_type,
3481
- message: entry.message,
3482
- timestamp: entry.timestamp,
3483
- note: "Error recorded for retry context. Use swarm_get_error_context to retrieve accumulated errors.",
3484
- },
3485
- null,
3486
- 2,
3487
- );
3488
- },
3489
- });
3490
-
3491
- /**
3492
- * Get accumulated errors for a bead to feed into retry prompts
3493
- *
3494
- * Returns formatted error context that can be injected into retry prompts
3495
- * to help agents learn from past failures.
3496
- */
3497
- export const swarm_get_error_context = tool({
3498
- description:
3499
- "Get accumulated errors for a bead. Returns formatted context for retry prompts.",
3500
- args: {
3501
- bead_id: tool.schema.string().describe("Bead ID to get errors for"),
3502
- include_resolved: tool.schema
3503
- .boolean()
3504
- .optional()
3505
- .describe("Include resolved errors (default: false)"),
3506
- },
3507
- async execute(args) {
3508
- const errorContext = await globalErrorAccumulator.getErrorContext(
3509
- args.bead_id,
3510
- args.include_resolved ?? false,
3511
- );
3512
-
3513
- const stats = await globalErrorAccumulator.getErrorStats(args.bead_id);
3514
-
3515
- return JSON.stringify(
3516
- {
3517
- bead_id: args.bead_id,
3518
- error_context: errorContext,
3519
- stats: {
3520
- total_errors: stats.total,
3521
- unresolved: stats.unresolved,
3522
- by_type: stats.by_type,
3523
- },
3524
- has_errors: errorContext.length > 0,
3525
- usage:
3526
- "Inject error_context into retry prompt using {error_context} placeholder",
3527
- },
3528
- null,
3529
- 2,
3530
- );
3531
- },
3532
- });
3533
-
3534
- /**
3535
- * Mark an error as resolved
3536
- *
3537
- * Call this after an agent successfully addresses an error to update
3538
- * the accumulator state.
3539
- */
3540
- export const swarm_resolve_error = tool({
3541
- description:
3542
- "Mark an error as resolved after fixing it. Updates error accumulator state.",
3543
- args: {
3544
- error_id: tool.schema.string().describe("Error ID to mark as resolved"),
3545
- },
3546
- async execute(args) {
3547
- await globalErrorAccumulator.resolveError(args.error_id);
3548
-
3549
- return JSON.stringify(
3550
- {
3551
- success: true,
3552
- error_id: args.error_id,
3553
- resolved: true,
3554
- },
3555
- null,
3556
- 2,
3557
- );
3558
- },
3559
- });
20
+ // Import tools from each module
21
+ import { strategyTools } from "./swarm-strategies";
22
+ import { decomposeTools } from "./swarm-decompose";
23
+ import { promptTools } from "./swarm-prompts";
24
+ import { orchestrateTools } from "./swarm-orchestrate";
3560
25
 
3561
26
  /**
3562
- * Initialize swarm and check tool availability
3563
- *
3564
- * Call this at the start of a swarm session to see what tools are available,
3565
- * what skills exist in the project, and what features will be degraded.
3566
- *
3567
- * Skills are automatically discovered from:
3568
- * - .opencode/skills/
3569
- * - .claude/skills/
3570
- * - skills/
27
+ * Combined swarm tools for plugin registration.
28
+ * Includes all tools from strategy, decompose, prompt, and orchestrate modules.
3571
29
  */
3572
- export const swarm_init = tool({
3573
- description:
3574
- "Initialize swarm session: discovers available skills, checks tool availability. ALWAYS call at swarm start.",
3575
- args: {
3576
- project_path: tool.schema
3577
- .string()
3578
- .optional()
3579
- .describe("Project path (for Agent Mail init)"),
3580
- },
3581
- async execute(args) {
3582
- // Check all tools
3583
- const availability = await checkAllTools();
3584
-
3585
- // Build status report
3586
- const report = formatToolAvailability(availability);
3587
-
3588
- // Check critical tools
3589
- const beadsAvailable = availability.get("beads")?.status.available ?? false;
3590
- const agentMailAvailable =
3591
- availability.get("agent-mail")?.status.available ?? false;
3592
-
3593
- // Build warnings
3594
- const warnings: string[] = [];
3595
- const degradedFeatures: string[] = [];
3596
-
3597
- if (!beadsAvailable) {
3598
- warnings.push(
3599
- "⚠️ beads (bd) not available - issue tracking disabled, swarm coordination will be limited",
3600
- );
3601
- degradedFeatures.push("issue tracking", "progress persistence");
3602
- }
3603
-
3604
- if (!agentMailAvailable) {
3605
- warnings.push(
3606
- "⚠️ agent-mail not available - multi-agent communication disabled",
3607
- );
3608
- degradedFeatures.push("agent communication", "file reservations");
3609
- }
3610
-
3611
- if (!availability.get("cass")?.status.available) {
3612
- degradedFeatures.push("historical context from past sessions");
3613
- }
3614
-
3615
- if (!availability.get("ubs")?.status.available) {
3616
- degradedFeatures.push("pre-completion bug scanning");
3617
- }
3618
-
3619
- if (!availability.get("semantic-memory")?.status.available) {
3620
- degradedFeatures.push("persistent learning (using in-memory fallback)");
3621
- }
3622
-
3623
- // Discover available skills
3624
- const availableSkills = await listSkills();
3625
- const skillsInfo = {
3626
- count: availableSkills.length,
3627
- available: availableSkills.length > 0,
3628
- skills: availableSkills.map((s) => ({
3629
- name: s.name,
3630
- description: s.description,
3631
- hasScripts: s.hasScripts,
3632
- })),
3633
- };
3634
-
3635
- // Add skills guidance if available
3636
- let skillsGuidance: string | undefined;
3637
- if (availableSkills.length > 0) {
3638
- skillsGuidance = `Found ${availableSkills.length} skill(s). Use skills_list to see details, skills_use to activate.`;
3639
- } else {
3640
- skillsGuidance =
3641
- "No skills found. Add skills to .opencode/skills/ or .claude/skills/ for specialized guidance.";
3642
- }
3643
-
3644
- return JSON.stringify(
3645
- {
3646
- ready: true,
3647
- tool_availability: Object.fromEntries(
3648
- Array.from(availability.entries()).map(([k, v]) => [
3649
- k,
3650
- {
3651
- available: v.status.available,
3652
- fallback: v.status.available ? null : v.fallbackBehavior,
3653
- },
3654
- ]),
3655
- ),
3656
- skills: skillsInfo,
3657
- warnings: warnings.length > 0 ? warnings : undefined,
3658
- degraded_features:
3659
- degradedFeatures.length > 0 ? degradedFeatures : undefined,
3660
- recommendations: {
3661
- skills: skillsGuidance,
3662
- beads: beadsAvailable
3663
- ? "✓ Use beads for all task tracking"
3664
- : "Install beads: npm i -g @joelhooks/beads",
3665
- agent_mail: agentMailAvailable
3666
- ? "✓ Use Agent Mail for coordination"
3667
- : "Start Agent Mail: agent-mail serve",
3668
- },
3669
- report,
3670
- },
3671
- null,
3672
- 2,
3673
- );
3674
- },
3675
- });
3676
-
3677
- // ============================================================================
3678
- // Export all tools
3679
- // ============================================================================
3680
-
3681
30
  export const swarmTools = {
3682
- swarm_init: swarm_init,
3683
- swarm_select_strategy: swarm_select_strategy,
3684
- swarm_plan_prompt: swarm_plan_prompt,
3685
- swarm_delegate_planning: swarm_delegate_planning,
3686
- swarm_decompose: swarm_decompose,
3687
- swarm_validate_decomposition: swarm_validate_decomposition,
3688
- swarm_status: swarm_status,
3689
- swarm_progress: swarm_progress,
3690
- swarm_broadcast: swarm_broadcast,
3691
- swarm_complete: swarm_complete,
3692
- swarm_learn: swarm_learn,
3693
- swarm_record_outcome: swarm_record_outcome,
3694
- swarm_subtask_prompt: swarm_subtask_prompt,
3695
- swarm_spawn_subtask: swarm_spawn_subtask,
3696
- swarm_complete_subtask: swarm_complete_subtask,
3697
- swarm_evaluation_prompt: swarm_evaluation_prompt,
3698
- swarm_accumulate_error: swarm_accumulate_error,
3699
- swarm_get_error_context: swarm_get_error_context,
3700
- swarm_resolve_error: swarm_resolve_error,
31
+ ...strategyTools,
32
+ ...decomposeTools,
33
+ ...promptTools,
34
+ ...orchestrateTools,
3701
35
  };
3702
-
3703
- // ============================================================================
3704
- // 3-Strike Detection
3705
- // ============================================================================
3706
-
3707
- /**
3708
- * Global strike storage for tracking consecutive fix failures
3709
- */
3710
- import {
3711
- InMemoryStrikeStorage,
3712
- addStrike,
3713
- getStrikes,
3714
- isStrikedOut,
3715
- getArchitecturePrompt,
3716
- clearStrikes,
3717
- type StrikeStorage,
3718
- } from "./learning";
3719
-
3720
- const globalStrikeStorage: StrikeStorage = new InMemoryStrikeStorage();
3721
-
3722
- /**
3723
- * Check if a bead has struck out (3 consecutive failures)
3724
- *
3725
- * The 3-Strike Rule:
3726
- * IF 3+ fixes have failed:
3727
- * STOP → Question the architecture
3728
- * DON'T attempt Fix #4
3729
- * Discuss with human partner
3730
- *
3731
- * This is NOT a failed hypothesis.
3732
- * This is a WRONG ARCHITECTURE.
3733
- *
3734
- * Use this tool to:
3735
- * - Check strike count before attempting a fix
3736
- * - Get architecture review prompt if struck out
3737
- * - Record a strike when a fix fails
3738
- * - Clear strikes when a fix succeeds
3739
- */
3740
- export const swarm_check_strikes = tool({
3741
- description:
3742
- "Check 3-strike status for a bead. Records failures, detects architectural problems, generates architecture review prompts.",
3743
- args: {
3744
- bead_id: tool.schema.string().describe("Bead ID to check"),
3745
- action: tool.schema
3746
- .enum(["check", "add_strike", "clear", "get_prompt"])
3747
- .describe(
3748
- "Action: check count, add strike, clear strikes, or get prompt",
3749
- ),
3750
- attempt: tool.schema
3751
- .string()
3752
- .optional()
3753
- .describe("Description of fix attempt (required for add_strike)"),
3754
- reason: tool.schema
3755
- .string()
3756
- .optional()
3757
- .describe("Why the fix failed (required for add_strike)"),
3758
- },
3759
- async execute(args) {
3760
- switch (args.action) {
3761
- case "check": {
3762
- const count = await getStrikes(args.bead_id, globalStrikeStorage);
3763
- const strikedOut = await isStrikedOut(
3764
- args.bead_id,
3765
- globalStrikeStorage,
3766
- );
3767
-
3768
- return JSON.stringify(
3769
- {
3770
- bead_id: args.bead_id,
3771
- strike_count: count,
3772
- is_striked_out: strikedOut,
3773
- message: strikedOut
3774
- ? "⚠️ STRUCK OUT: 3 strikes reached. Use get_prompt action for architecture review."
3775
- : count === 0
3776
- ? "No strikes. Clear to proceed."
3777
- : `${count} strike${count > 1 ? "s" : ""}. ${3 - count} remaining before architecture review required.`,
3778
- next_action: strikedOut
3779
- ? "Call with action=get_prompt to get architecture review questions"
3780
- : "Continue with fix attempt",
3781
- },
3782
- null,
3783
- 2,
3784
- );
3785
- }
3786
-
3787
- case "add_strike": {
3788
- if (!args.attempt || !args.reason) {
3789
- return JSON.stringify(
3790
- {
3791
- error: "add_strike requires 'attempt' and 'reason' parameters",
3792
- },
3793
- null,
3794
- 2,
3795
- );
3796
- }
3797
-
3798
- const record = await addStrike(
3799
- args.bead_id,
3800
- args.attempt,
3801
- args.reason,
3802
- globalStrikeStorage,
3803
- );
3804
-
3805
- const strikedOut = record.strike_count >= 3;
3806
-
3807
- // Build response with memory storage hint on 3-strike
3808
- const response: Record<string, unknown> = {
3809
- bead_id: args.bead_id,
3810
- strike_count: record.strike_count,
3811
- is_striked_out: strikedOut,
3812
- failures: record.failures,
3813
- message: strikedOut
3814
- ? "⚠️ STRUCK OUT: 3 strikes reached. STOP and question the architecture."
3815
- : `Strike ${record.strike_count} recorded. ${3 - record.strike_count} remaining.`,
3816
- warning: strikedOut
3817
- ? "DO NOT attempt Fix #4. Call with action=get_prompt for architecture review."
3818
- : undefined,
3819
- };
3820
-
3821
- // Add semantic-memory storage hint on 3-strike
3822
- if (strikedOut) {
3823
- response.memory_store = formatMemoryStoreOn3Strike(
3824
- args.bead_id,
3825
- record.failures,
3826
- );
3827
- }
3828
-
3829
- return JSON.stringify(response, null, 2);
3830
- }
3831
-
3832
- case "clear": {
3833
- await clearStrikes(args.bead_id, globalStrikeStorage);
3834
-
3835
- return JSON.stringify(
3836
- {
3837
- bead_id: args.bead_id,
3838
- strike_count: 0,
3839
- is_striked_out: false,
3840
- message: "Strikes cleared. Fresh start.",
3841
- },
3842
- null,
3843
- 2,
3844
- );
3845
- }
3846
-
3847
- case "get_prompt": {
3848
- const prompt = await getArchitecturePrompt(
3849
- args.bead_id,
3850
- globalStrikeStorage,
3851
- );
3852
-
3853
- if (!prompt) {
3854
- return JSON.stringify(
3855
- {
3856
- bead_id: args.bead_id,
3857
- has_prompt: false,
3858
- message: "No architecture prompt (not struck out yet)",
3859
- },
3860
- null,
3861
- 2,
3862
- );
3863
- }
3864
-
3865
- return JSON.stringify(
3866
- {
3867
- bead_id: args.bead_id,
3868
- has_prompt: true,
3869
- architecture_review_prompt: prompt,
3870
- message:
3871
- "Architecture review required. Present this prompt to the human partner.",
3872
- },
3873
- null,
3874
- 2,
3875
- );
3876
- }
3877
-
3878
- default:
3879
- return JSON.stringify(
3880
- {
3881
- error: `Unknown action: ${args.action}`,
3882
- },
3883
- null,
3884
- 2,
3885
- );
3886
- }
3887
- },
3888
- });