opencode-swarm-plugin 0.43.0 → 0.44.1

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.
Files changed (208) hide show
  1. package/bin/cass.characterization.test.ts +422 -0
  2. package/bin/swarm.serve.test.ts +6 -4
  3. package/bin/swarm.test.ts +68 -0
  4. package/bin/swarm.ts +81 -8
  5. package/dist/compaction-prompt-scoring.js +139 -0
  6. package/dist/contributor-tools.d.ts +42 -0
  7. package/dist/contributor-tools.d.ts.map +1 -0
  8. package/dist/eval-capture.js +12811 -0
  9. package/dist/hive.d.ts.map +1 -1
  10. package/dist/index.d.ts +12 -0
  11. package/dist/index.d.ts.map +1 -1
  12. package/dist/index.js +7728 -62590
  13. package/dist/plugin.js +23833 -78695
  14. package/dist/sessions/agent-discovery.d.ts +59 -0
  15. package/dist/sessions/agent-discovery.d.ts.map +1 -0
  16. package/dist/sessions/index.d.ts +10 -0
  17. package/dist/sessions/index.d.ts.map +1 -0
  18. package/dist/swarm-orchestrate.d.ts.map +1 -1
  19. package/dist/swarm-prompts.d.ts.map +1 -1
  20. package/dist/swarm-review.d.ts.map +1 -1
  21. package/package.json +17 -5
  22. package/.changeset/swarm-insights-data-layer.md +0 -63
  23. package/.hive/analysis/eval-failure-analysis-2025-12-25.md +0 -331
  24. package/.hive/analysis/session-data-quality-audit.md +0 -320
  25. package/.hive/eval-results.json +0 -483
  26. package/.hive/issues.jsonl +0 -138
  27. package/.hive/memories.jsonl +0 -729
  28. package/.opencode/eval-history.jsonl +0 -327
  29. package/.turbo/turbo-build.log +0 -9
  30. package/CHANGELOG.md +0 -2255
  31. package/SCORER-ANALYSIS.md +0 -598
  32. package/docs/analysis/subagent-coordination-patterns.md +0 -902
  33. package/docs/analysis-socratic-planner-pattern.md +0 -504
  34. package/docs/planning/ADR-001-monorepo-structure.md +0 -171
  35. package/docs/planning/ADR-002-package-extraction.md +0 -393
  36. package/docs/planning/ADR-003-performance-improvements.md +0 -451
  37. package/docs/planning/ADR-004-message-queue-features.md +0 -187
  38. package/docs/planning/ADR-005-devtools-observability.md +0 -202
  39. package/docs/planning/ADR-007-swarm-enhancements-worktree-review.md +0 -168
  40. package/docs/planning/ADR-008-worker-handoff-protocol.md +0 -293
  41. package/docs/planning/ADR-009-oh-my-opencode-patterns.md +0 -353
  42. package/docs/planning/ROADMAP.md +0 -368
  43. package/docs/semantic-memory-cli-syntax.md +0 -123
  44. package/docs/swarm-mail-architecture.md +0 -1147
  45. package/docs/testing/context-recovery-test.md +0 -470
  46. package/evals/ARCHITECTURE.md +0 -1189
  47. package/evals/README.md +0 -768
  48. package/evals/compaction-prompt.eval.ts +0 -149
  49. package/evals/compaction-resumption.eval.ts +0 -289
  50. package/evals/coordinator-behavior.eval.ts +0 -307
  51. package/evals/coordinator-session.eval.ts +0 -154
  52. package/evals/evalite.config.ts.bak +0 -15
  53. package/evals/example.eval.ts +0 -31
  54. package/evals/fixtures/compaction-cases.ts +0 -350
  55. package/evals/fixtures/compaction-prompt-cases.ts +0 -311
  56. package/evals/fixtures/coordinator-sessions.ts +0 -328
  57. package/evals/fixtures/decomposition-cases.ts +0 -105
  58. package/evals/lib/compaction-loader.test.ts +0 -248
  59. package/evals/lib/compaction-loader.ts +0 -320
  60. package/evals/lib/data-loader.evalite-test.ts +0 -289
  61. package/evals/lib/data-loader.test.ts +0 -345
  62. package/evals/lib/data-loader.ts +0 -281
  63. package/evals/lib/llm.ts +0 -115
  64. package/evals/scorers/compaction-prompt-scorers.ts +0 -145
  65. package/evals/scorers/compaction-scorers.ts +0 -305
  66. package/evals/scorers/coordinator-discipline.evalite-test.ts +0 -539
  67. package/evals/scorers/coordinator-discipline.ts +0 -325
  68. package/evals/scorers/index.test.ts +0 -146
  69. package/evals/scorers/index.ts +0 -328
  70. package/evals/scorers/outcome-scorers.evalite-test.ts +0 -27
  71. package/evals/scorers/outcome-scorers.ts +0 -349
  72. package/evals/swarm-decomposition.eval.ts +0 -121
  73. package/examples/commands/swarm.md +0 -745
  74. package/examples/plugin-wrapper-template.ts +0 -2426
  75. package/examples/skills/hive-workflow/SKILL.md +0 -212
  76. package/examples/skills/skill-creator/SKILL.md +0 -223
  77. package/examples/skills/swarm-coordination/SKILL.md +0 -292
  78. package/global-skills/cli-builder/SKILL.md +0 -344
  79. package/global-skills/cli-builder/references/advanced-patterns.md +0 -244
  80. package/global-skills/learning-systems/SKILL.md +0 -644
  81. package/global-skills/skill-creator/LICENSE.txt +0 -202
  82. package/global-skills/skill-creator/SKILL.md +0 -352
  83. package/global-skills/skill-creator/references/output-patterns.md +0 -82
  84. package/global-skills/skill-creator/references/workflows.md +0 -28
  85. package/global-skills/swarm-coordination/SKILL.md +0 -995
  86. package/global-skills/swarm-coordination/references/coordinator-patterns.md +0 -235
  87. package/global-skills/swarm-coordination/references/strategies.md +0 -138
  88. package/global-skills/system-design/SKILL.md +0 -213
  89. package/global-skills/testing-patterns/SKILL.md +0 -430
  90. package/global-skills/testing-patterns/references/dependency-breaking-catalog.md +0 -586
  91. package/opencode-swarm-plugin-0.30.7.tgz +0 -0
  92. package/opencode-swarm-plugin-0.31.0.tgz +0 -0
  93. package/scripts/cleanup-test-memories.ts +0 -346
  94. package/scripts/init-skill.ts +0 -222
  95. package/scripts/migrate-unknown-sessions.ts +0 -349
  96. package/scripts/validate-skill.ts +0 -204
  97. package/src/agent-mail.ts +0 -1724
  98. package/src/anti-patterns.test.ts +0 -1167
  99. package/src/anti-patterns.ts +0 -448
  100. package/src/compaction-capture.integration.test.ts +0 -257
  101. package/src/compaction-hook.test.ts +0 -838
  102. package/src/compaction-hook.ts +0 -1204
  103. package/src/compaction-observability.integration.test.ts +0 -139
  104. package/src/compaction-observability.test.ts +0 -187
  105. package/src/compaction-observability.ts +0 -324
  106. package/src/compaction-prompt-scorers.test.ts +0 -475
  107. package/src/compaction-prompt-scoring.ts +0 -300
  108. package/src/dashboard.test.ts +0 -611
  109. package/src/dashboard.ts +0 -462
  110. package/src/error-enrichment.test.ts +0 -403
  111. package/src/error-enrichment.ts +0 -219
  112. package/src/eval-capture.test.ts +0 -1015
  113. package/src/eval-capture.ts +0 -929
  114. package/src/eval-gates.test.ts +0 -306
  115. package/src/eval-gates.ts +0 -218
  116. package/src/eval-history.test.ts +0 -508
  117. package/src/eval-history.ts +0 -214
  118. package/src/eval-learning.test.ts +0 -378
  119. package/src/eval-learning.ts +0 -360
  120. package/src/eval-runner.test.ts +0 -223
  121. package/src/eval-runner.ts +0 -402
  122. package/src/export-tools.test.ts +0 -476
  123. package/src/export-tools.ts +0 -257
  124. package/src/hive.integration.test.ts +0 -2241
  125. package/src/hive.ts +0 -1628
  126. package/src/index.ts +0 -935
  127. package/src/learning.integration.test.ts +0 -1815
  128. package/src/learning.ts +0 -1079
  129. package/src/logger.test.ts +0 -189
  130. package/src/logger.ts +0 -135
  131. package/src/mandate-promotion.test.ts +0 -473
  132. package/src/mandate-promotion.ts +0 -239
  133. package/src/mandate-storage.integration.test.ts +0 -601
  134. package/src/mandate-storage.test.ts +0 -578
  135. package/src/mandate-storage.ts +0 -794
  136. package/src/mandates.ts +0 -540
  137. package/src/memory-tools.test.ts +0 -195
  138. package/src/memory-tools.ts +0 -344
  139. package/src/memory.integration.test.ts +0 -334
  140. package/src/memory.test.ts +0 -158
  141. package/src/memory.ts +0 -527
  142. package/src/model-selection.test.ts +0 -188
  143. package/src/model-selection.ts +0 -68
  144. package/src/observability-tools.test.ts +0 -359
  145. package/src/observability-tools.ts +0 -871
  146. package/src/output-guardrails.test.ts +0 -438
  147. package/src/output-guardrails.ts +0 -381
  148. package/src/pattern-maturity.test.ts +0 -1160
  149. package/src/pattern-maturity.ts +0 -525
  150. package/src/planning-guardrails.test.ts +0 -491
  151. package/src/planning-guardrails.ts +0 -438
  152. package/src/plugin.ts +0 -23
  153. package/src/post-compaction-tracker.test.ts +0 -251
  154. package/src/post-compaction-tracker.ts +0 -237
  155. package/src/query-tools.test.ts +0 -636
  156. package/src/query-tools.ts +0 -324
  157. package/src/rate-limiter.integration.test.ts +0 -466
  158. package/src/rate-limiter.ts +0 -774
  159. package/src/replay-tools.test.ts +0 -496
  160. package/src/replay-tools.ts +0 -240
  161. package/src/repo-crawl.integration.test.ts +0 -441
  162. package/src/repo-crawl.ts +0 -610
  163. package/src/schemas/cell-events.test.ts +0 -347
  164. package/src/schemas/cell-events.ts +0 -807
  165. package/src/schemas/cell.ts +0 -257
  166. package/src/schemas/evaluation.ts +0 -166
  167. package/src/schemas/index.test.ts +0 -199
  168. package/src/schemas/index.ts +0 -286
  169. package/src/schemas/mandate.ts +0 -232
  170. package/src/schemas/swarm-context.ts +0 -115
  171. package/src/schemas/task.ts +0 -161
  172. package/src/schemas/worker-handoff.test.ts +0 -302
  173. package/src/schemas/worker-handoff.ts +0 -131
  174. package/src/skills.integration.test.ts +0 -1192
  175. package/src/skills.test.ts +0 -643
  176. package/src/skills.ts +0 -1549
  177. package/src/storage.integration.test.ts +0 -341
  178. package/src/storage.ts +0 -884
  179. package/src/structured.integration.test.ts +0 -817
  180. package/src/structured.test.ts +0 -1046
  181. package/src/structured.ts +0 -762
  182. package/src/swarm-decompose.test.ts +0 -188
  183. package/src/swarm-decompose.ts +0 -1302
  184. package/src/swarm-deferred.integration.test.ts +0 -157
  185. package/src/swarm-deferred.test.ts +0 -38
  186. package/src/swarm-insights.test.ts +0 -214
  187. package/src/swarm-insights.ts +0 -459
  188. package/src/swarm-mail.integration.test.ts +0 -970
  189. package/src/swarm-mail.ts +0 -739
  190. package/src/swarm-orchestrate.integration.test.ts +0 -282
  191. package/src/swarm-orchestrate.test.ts +0 -548
  192. package/src/swarm-orchestrate.ts +0 -3084
  193. package/src/swarm-prompts.test.ts +0 -1270
  194. package/src/swarm-prompts.ts +0 -2077
  195. package/src/swarm-research.integration.test.ts +0 -701
  196. package/src/swarm-research.test.ts +0 -698
  197. package/src/swarm-research.ts +0 -472
  198. package/src/swarm-review.integration.test.ts +0 -285
  199. package/src/swarm-review.test.ts +0 -879
  200. package/src/swarm-review.ts +0 -709
  201. package/src/swarm-strategies.ts +0 -407
  202. package/src/swarm-worktree.test.ts +0 -501
  203. package/src/swarm-worktree.ts +0 -575
  204. package/src/swarm.integration.test.ts +0 -2377
  205. package/src/swarm.ts +0 -38
  206. package/src/tool-adapter.integration.test.ts +0 -1221
  207. package/src/tool-availability.ts +0 -461
  208. package/tsconfig.json +0 -28
@@ -1,929 +0,0 @@
1
- /**
2
- * Eval Data Capture - Captures real swarm execution data for evals
3
- *
4
- * Records decomposition inputs, outputs, and outcomes to JSONL files
5
- * that can be used as ground truth for Evalite evals.
6
- *
7
- * Data flow:
8
- * 1. swarm_decompose captures: task, context, generated decomposition
9
- * 2. swarm_complete captures: outcome signals per subtask
10
- * 3. swarm_record_outcome captures: learning signals
11
- * 4. Human feedback (optional): accept/reject/modify
12
- * 5. Coordinator events: decisions, violations, outcomes, compaction
13
- * 6. Session capture: full coordinator session to ~/.config/swarm-tools/sessions/
14
- *
15
- * Event types:
16
- * - DECISION: strategy_selected, worker_spawned, review_completed, decomposition_complete, researcher_spawned, skill_loaded, inbox_checked, blocker_resolved, scope_change_approved, scope_change_rejected
17
- * - VIOLATION: coordinator_edited_file, coordinator_ran_tests, coordinator_reserved_files, no_worker_spawned
18
- * - OUTCOME: subtask_success, subtask_retry, subtask_failed, epic_complete, blocker_detected
19
- * - COMPACTION: detection_complete, prompt_generated, context_injected, resumption_started, tool_call_tracked
20
- *
21
- * @module eval-capture
22
- */
23
- import * as fs from "node:fs";
24
- import * as os from "node:os";
25
- import * as path from "node:path";
26
- import { z } from "zod";
27
-
28
- // ============================================================================
29
- // Schemas
30
- // ============================================================================
31
-
32
- /**
33
- * Subtask outcome - what actually happened
34
- */
35
- export const SubtaskOutcomeSchema = z.object({
36
- /** Subtask bead ID */
37
- bead_id: z.string(),
38
- /** Subtask title */
39
- title: z.string(),
40
- /** Planned files */
41
- planned_files: z.array(z.string()),
42
- /** Actual files touched */
43
- actual_files: z.array(z.string()),
44
- /** Duration in ms */
45
- duration_ms: z.number().int().min(0),
46
- /** Error count */
47
- error_count: z.number().int().min(0),
48
- /** Retry count */
49
- retry_count: z.number().int().min(0),
50
- /** Success */
51
- success: z.boolean(),
52
- /** Failure mode if failed */
53
- failure_mode: z.string().optional(),
54
- });
55
- export type SubtaskOutcome = z.infer<typeof SubtaskOutcomeSchema>;
56
-
57
- /**
58
- * Complete eval record - input, output, and outcome
59
- */
60
- export const EvalRecordSchema = z.object({
61
- /** Unique ID for this eval record */
62
- id: z.string(),
63
- /** Timestamp when decomposition was generated */
64
- timestamp: z.string(), // ISO-8601
65
- /** Project path */
66
- project_path: z.string(),
67
-
68
- // INPUT
69
- /** Original task description */
70
- task: z.string(),
71
- /** Context provided (codebase info, CASS results, etc.) */
72
- context: z.string().optional(),
73
- /** Strategy used for decomposition */
74
- strategy: z.enum(["file-based", "feature-based", "risk-based", "auto"]),
75
- /** Number of subtasks generated */
76
- subtask_count: z.number().int().min(1),
77
-
78
- // OUTPUT (the decomposition)
79
- /** Epic title */
80
- epic_title: z.string(),
81
- /** Epic description */
82
- epic_description: z.string().optional(),
83
- /** Generated subtasks */
84
- subtasks: z.array(
85
- z.object({
86
- title: z.string(),
87
- description: z.string().optional(),
88
- files: z.array(z.string()),
89
- dependencies: z.array(z.number()).optional(),
90
- estimated_complexity: z.number().int().min(1).max(5).optional(),
91
- }),
92
- ),
93
-
94
- // OUTCOME (what actually happened)
95
- /** Subtask outcomes */
96
- outcomes: z.array(SubtaskOutcomeSchema).optional(),
97
- /** Overall success (all subtasks succeeded) */
98
- overall_success: z.boolean().optional(),
99
- /** Total duration (sum of all subtasks) */
100
- total_duration_ms: z.number().int().min(0).optional(),
101
- /** Total errors across all subtasks */
102
- total_errors: z.number().int().min(0).optional(),
103
-
104
- // HUMAN FEEDBACK (optional)
105
- /** Human accepted the decomposition as-is */
106
- human_accepted: z.boolean().optional(),
107
- /** Human modified the decomposition */
108
- human_modified: z.boolean().optional(),
109
- /** Human feedback notes */
110
- human_notes: z.string().optional(),
111
-
112
- // COMPUTED METRICS
113
- /** File overlap between subtasks (should be 0) */
114
- file_overlap_count: z.number().int().min(0).optional(),
115
- /** Scope accuracy: actual files / planned files */
116
- scope_accuracy: z.number().min(0).max(2).optional(),
117
- /** Time balance: max duration / min duration (lower is better) */
118
- time_balance_ratio: z.number().min(1).optional(),
119
- });
120
- export type EvalRecord = z.infer<typeof EvalRecordSchema>;
121
-
122
- /**
123
- * Partial record for in-progress capture
124
- */
125
- export type PartialEvalRecord = Partial<EvalRecord> & {
126
- id: string;
127
- timestamp: string;
128
- task: string;
129
- };
130
-
131
- /**
132
- * Coordinator Event - captures coordinator decisions, violations, outcomes, and compaction
133
- */
134
- export const CoordinatorEventSchema = z.discriminatedUnion("event_type", [
135
- // DECISION events
136
- z.object({
137
- session_id: z.string(),
138
- epic_id: z.string(),
139
- timestamp: z.string(),
140
- event_type: z.literal("DECISION"),
141
- decision_type: z.enum([
142
- "strategy_selected",
143
- "worker_spawned",
144
- "review_completed",
145
- "decomposition_complete",
146
- "researcher_spawned",
147
- "skill_loaded",
148
- "inbox_checked",
149
- "blocker_resolved",
150
- "scope_change_approved",
151
- "scope_change_rejected",
152
- ]),
153
- payload: z.any(),
154
- }),
155
- // VIOLATION events
156
- z.object({
157
- session_id: z.string(),
158
- epic_id: z.string(),
159
- timestamp: z.string(),
160
- event_type: z.literal("VIOLATION"),
161
- violation_type: z.enum([
162
- "coordinator_edited_file",
163
- "coordinator_ran_tests",
164
- "coordinator_reserved_files",
165
- "no_worker_spawned",
166
- ]),
167
- payload: z.any(),
168
- }),
169
- // OUTCOME events
170
- z.object({
171
- session_id: z.string(),
172
- epic_id: z.string(),
173
- timestamp: z.string(),
174
- event_type: z.literal("OUTCOME"),
175
- outcome_type: z.enum([
176
- "subtask_success",
177
- "subtask_retry",
178
- "subtask_failed",
179
- "epic_complete",
180
- "blocker_detected",
181
- ]),
182
- payload: z.any(),
183
- }),
184
- // COMPACTION events
185
- z.object({
186
- session_id: z.string(),
187
- epic_id: z.string(),
188
- timestamp: z.string(),
189
- event_type: z.literal("COMPACTION"),
190
- compaction_type: z.enum([
191
- "detection_complete",
192
- "prompt_generated",
193
- "context_injected",
194
- "resumption_started",
195
- "tool_call_tracked",
196
- ]),
197
- payload: z.any(),
198
- }),
199
- ]);
200
- export type CoordinatorEvent = z.infer<typeof CoordinatorEventSchema>;
201
-
202
- /**
203
- * Coordinator Session - wraps a full coordinator session
204
- */
205
- export const CoordinatorSessionSchema = z.object({
206
- session_id: z.string(),
207
- epic_id: z.string(),
208
- start_time: z.string(),
209
- end_time: z.string().optional(),
210
- events: z.array(CoordinatorEventSchema),
211
- });
212
- export type CoordinatorSession = z.infer<typeof CoordinatorSessionSchema>;
213
-
214
- // ============================================================================
215
- // Storage
216
- // ============================================================================
217
-
218
- /**
219
- * Default path for eval data
220
- */
221
- export const DEFAULT_EVAL_DATA_PATH = ".opencode/eval-data.jsonl";
222
-
223
- /**
224
- * Get the eval data file path for a project
225
- */
226
- export function getEvalDataPath(projectPath: string): string {
227
- return path.join(projectPath, DEFAULT_EVAL_DATA_PATH);
228
- }
229
-
230
- /**
231
- * Ensure the eval data directory exists
232
- */
233
- export function ensureEvalDataDir(projectPath: string): void {
234
- const evalPath = getEvalDataPath(projectPath);
235
- const dir = path.dirname(evalPath);
236
- if (!fs.existsSync(dir)) {
237
- fs.mkdirSync(dir, { recursive: true });
238
- }
239
- }
240
-
241
- /**
242
- * Append an eval record to the JSONL file
243
- */
244
- export function appendEvalRecord(
245
- projectPath: string,
246
- record: EvalRecord | PartialEvalRecord,
247
- ): void {
248
- ensureEvalDataDir(projectPath);
249
- const evalPath = getEvalDataPath(projectPath);
250
- const line = `${JSON.stringify(record)}\n`;
251
- fs.appendFileSync(evalPath, line, "utf-8");
252
- }
253
-
254
- /**
255
- * Read all eval records from a project
256
- */
257
- export function readEvalRecords(projectPath: string): EvalRecord[] {
258
- const evalPath = getEvalDataPath(projectPath);
259
- if (!fs.existsSync(evalPath)) {
260
- return [];
261
- }
262
-
263
- const content = fs.readFileSync(evalPath, "utf-8");
264
- const lines = content.trim().split("\n").filter(Boolean);
265
-
266
- return lines.map((line) => {
267
- const parsed = JSON.parse(line);
268
- return EvalRecordSchema.parse(parsed);
269
- });
270
- }
271
-
272
- /**
273
- * Read partial records (for updating in-progress records)
274
- */
275
- export function readPartialRecords(projectPath: string): PartialEvalRecord[] {
276
- const evalPath = getEvalDataPath(projectPath);
277
- if (!fs.existsSync(evalPath)) {
278
- return [];
279
- }
280
-
281
- const content = fs.readFileSync(evalPath, "utf-8");
282
- const lines = content.trim().split("\n").filter(Boolean);
283
-
284
- return lines.map((line) => JSON.parse(line) as PartialEvalRecord);
285
- }
286
-
287
- /**
288
- * Update an existing record by ID
289
- */
290
- export function updateEvalRecord(
291
- projectPath: string,
292
- id: string,
293
- updates: Partial<EvalRecord>,
294
- ): boolean {
295
- const records = readPartialRecords(projectPath);
296
- const index = records.findIndex((r) => r.id === id);
297
-
298
- if (index === -1) {
299
- return false;
300
- }
301
-
302
- records[index] = { ...records[index], ...updates };
303
-
304
- // Rewrite the file
305
- const evalPath = getEvalDataPath(projectPath);
306
- const content = `${records.map((r) => JSON.stringify(r)).join("\n")}\n`;
307
- fs.writeFileSync(evalPath, content, "utf-8");
308
-
309
- return true;
310
- }
311
-
312
- // ============================================================================
313
- // Capture Functions
314
- // ============================================================================
315
-
316
- /**
317
- * In-memory store for in-progress records (keyed by epic ID)
318
- */
319
- const inProgressRecords = new Map<string, PartialEvalRecord>();
320
-
321
- /**
322
- * Start capturing a decomposition
323
- *
324
- * Called when swarm_decompose generates a decomposition.
325
- * Creates a partial record that will be completed when outcomes arrive.
326
- */
327
- export function captureDecomposition(params: {
328
- epicId: string;
329
- projectPath: string;
330
- task: string;
331
- context?: string;
332
- strategy: "file-based" | "feature-based" | "risk-based" | "auto";
333
- epicTitle: string;
334
- epicDescription?: string;
335
- subtasks: Array<{
336
- title: string;
337
- description?: string;
338
- files: string[];
339
- dependencies?: number[];
340
- estimated_complexity?: number;
341
- }>;
342
- }): PartialEvalRecord {
343
- const record: PartialEvalRecord = {
344
- id: params.epicId,
345
- timestamp: new Date().toISOString(),
346
- project_path: params.projectPath,
347
- task: params.task,
348
- context: params.context,
349
- strategy: params.strategy,
350
- subtask_count: params.subtasks.length,
351
- epic_title: params.epicTitle,
352
- epic_description: params.epicDescription,
353
- subtasks: params.subtasks,
354
- outcomes: [],
355
- };
356
-
357
- // Store in memory for later updates
358
- inProgressRecords.set(params.epicId, record);
359
-
360
- // Also persist to disk (partial)
361
- appendEvalRecord(params.projectPath, record);
362
-
363
- return record;
364
- }
365
-
366
- /**
367
- * Capture a subtask outcome
368
- *
369
- * Called when swarm_complete finishes a subtask.
370
- * Updates the in-progress record with outcome data.
371
- */
372
- export function captureSubtaskOutcome(params: {
373
- epicId: string;
374
- projectPath: string;
375
- beadId: string;
376
- title: string;
377
- plannedFiles: string[];
378
- actualFiles: string[];
379
- durationMs: number;
380
- errorCount: number;
381
- retryCount: number;
382
- success: boolean;
383
- failureMode?: string;
384
- }): void {
385
- const outcome: SubtaskOutcome = {
386
- bead_id: params.beadId,
387
- title: params.title,
388
- planned_files: params.plannedFiles,
389
- actual_files: params.actualFiles,
390
- duration_ms: params.durationMs,
391
- error_count: params.errorCount,
392
- retry_count: params.retryCount,
393
- success: params.success,
394
- failure_mode: params.failureMode,
395
- };
396
-
397
- // Update in-memory record
398
- const record = inProgressRecords.get(params.epicId);
399
- if (record) {
400
- record.outcomes = record.outcomes || [];
401
- record.outcomes.push(outcome);
402
- }
403
-
404
- // Update on disk
405
- updateEvalRecord(params.projectPath, params.epicId, {
406
- outcomes: record?.outcomes,
407
- });
408
- }
409
-
410
- /**
411
- * Finalize an eval record
412
- *
413
- * Called when all subtasks are complete.
414
- * Computes aggregate metrics and marks record as complete.
415
- */
416
- export function finalizeEvalRecord(params: {
417
- epicId: string;
418
- projectPath: string;
419
- }): EvalRecord | null {
420
- const record = inProgressRecords.get(params.epicId);
421
- if (!record || !record.outcomes || record.outcomes.length === 0) {
422
- return null;
423
- }
424
-
425
- // Compute aggregate metrics
426
- const outcomes = record.outcomes;
427
-
428
- const overallSuccess = outcomes.every((o) => o.success);
429
- const totalDurationMs = outcomes.reduce((sum, o) => sum + o.duration_ms, 0);
430
- const totalErrors = outcomes.reduce((sum, o) => sum + o.error_count, 0);
431
-
432
- // File overlap: count files that appear in multiple subtasks
433
- const allPlannedFiles = record.subtasks?.flatMap((s) => s.files) || [];
434
- const fileOccurrences = new Map<string, number>();
435
- for (const file of allPlannedFiles) {
436
- fileOccurrences.set(file, (fileOccurrences.get(file) || 0) + 1);
437
- }
438
- const fileOverlapCount = Array.from(fileOccurrences.values()).filter(
439
- (count) => count > 1,
440
- ).length;
441
-
442
- // Scope accuracy: actual files / planned files
443
- const plannedFileSet = new Set(allPlannedFiles);
444
- const actualFileSet = new Set(outcomes.flatMap((o) => o.actual_files));
445
- const scopeAccuracy =
446
- plannedFileSet.size > 0 ? actualFileSet.size / plannedFileSet.size : 1;
447
-
448
- // Time balance: max duration / min duration
449
- const durations = outcomes.map((o) => o.duration_ms).filter((d) => d > 0);
450
- const timeBalanceRatio =
451
- durations.length > 1 ? Math.max(...durations) / Math.min(...durations) : 1;
452
-
453
- // Update record with computed metrics
454
- const finalRecord: EvalRecord = {
455
- ...(record as EvalRecord),
456
- overall_success: overallSuccess,
457
- total_duration_ms: totalDurationMs,
458
- total_errors: totalErrors,
459
- file_overlap_count: fileOverlapCount,
460
- scope_accuracy: scopeAccuracy,
461
- time_balance_ratio: timeBalanceRatio,
462
- };
463
-
464
- // Update on disk
465
- updateEvalRecord(params.projectPath, params.epicId, finalRecord);
466
-
467
- // Remove from in-progress
468
- inProgressRecords.delete(params.epicId);
469
-
470
- return finalRecord;
471
- }
472
-
473
- /**
474
- * Capture human feedback on a decomposition
475
- */
476
- export function captureHumanFeedback(params: {
477
- epicId: string;
478
- projectPath: string;
479
- accepted: boolean;
480
- modified: boolean;
481
- notes?: string;
482
- }): void {
483
- updateEvalRecord(params.projectPath, params.epicId, {
484
- human_accepted: params.accepted,
485
- human_modified: params.modified,
486
- human_notes: params.notes,
487
- });
488
- }
489
-
490
- // ============================================================================
491
- // Eval Data Export
492
- // ============================================================================
493
-
494
- /**
495
- * Export eval records as Evalite-compatible test cases
496
- *
497
- * Filters to only complete records with outcomes.
498
- */
499
- export function exportForEvalite(projectPath: string): Array<{
500
- input: { task: string; context?: string };
501
- expected: {
502
- minSubtasks: number;
503
- subtaskCount: number;
504
- requiredFiles?: string[];
505
- overallSuccess?: boolean;
506
- };
507
- actual: EvalRecord;
508
- }> {
509
- const records = readEvalRecords(projectPath);
510
-
511
- return records
512
- .filter((r) => r.outcomes && r.outcomes.length > 0)
513
- .map((record) => ({
514
- input: {
515
- task: record.task,
516
- context: record.context,
517
- },
518
- expected: {
519
- minSubtasks: 2,
520
- subtaskCount: record.subtask_count,
521
- requiredFiles: record.subtasks.flatMap((s) => s.files),
522
- overallSuccess: record.overall_success,
523
- },
524
- actual: record,
525
- }));
526
- }
527
-
528
- /**
529
- * Get statistics about captured eval data
530
- */
531
- export function getEvalDataStats(projectPath: string): {
532
- totalRecords: number;
533
- completeRecords: number;
534
- successRate: number;
535
- avgSubtasks: number;
536
- avgDurationMs: number;
537
- avgScopeAccuracy: number;
538
- avgTimeBalance: number;
539
- } {
540
- const records = readEvalRecords(projectPath);
541
- const complete = records.filter((r) => r.outcomes && r.outcomes.length > 0);
542
-
543
- if (complete.length === 0) {
544
- return {
545
- totalRecords: records.length,
546
- completeRecords: 0,
547
- successRate: 0,
548
- avgSubtasks: 0,
549
- avgDurationMs: 0,
550
- avgScopeAccuracy: 0,
551
- avgTimeBalance: 0,
552
- };
553
- }
554
-
555
- const successCount = complete.filter((r) => r.overall_success).length;
556
- const avgSubtasks =
557
- complete.reduce((sum, r) => sum + (r.outcomes?.length || 0), 0) /
558
- complete.length;
559
- const avgDurationMs =
560
- complete.reduce((sum, r) => sum + (r.total_duration_ms || 0), 0) /
561
- complete.length;
562
- const avgScopeAccuracy =
563
- complete.reduce((sum, r) => sum + (r.scope_accuracy || 1), 0) /
564
- complete.length;
565
- const avgTimeBalance =
566
- complete.reduce((sum, r) => sum + (r.time_balance_ratio || 1), 0) /
567
- complete.length;
568
-
569
- return {
570
- totalRecords: records.length,
571
- completeRecords: complete.length,
572
- successRate: successCount / complete.length,
573
- avgSubtasks,
574
- avgDurationMs,
575
- avgScopeAccuracy,
576
- avgTimeBalance,
577
- };
578
- }
579
-
580
- // ============================================================================
581
- // Coordinator Session Capture
582
- // ============================================================================
583
-
584
- /**
585
- * Get the session directory path
586
- */
587
- export function getSessionDir(): string {
588
- return path.join(os.homedir(), ".config", "swarm-tools", "sessions");
589
- }
590
-
591
- /**
592
- * Get the session file path for a session ID
593
- */
594
- export function getSessionPath(sessionId: string): string {
595
- return path.join(getSessionDir(), `${sessionId}.jsonl`);
596
- }
597
-
598
- /**
599
- * Ensure the session directory exists
600
- */
601
- export function ensureSessionDir(): void {
602
- const sessionDir = getSessionDir();
603
- if (!fs.existsSync(sessionDir)) {
604
- fs.mkdirSync(sessionDir, { recursive: true });
605
- }
606
- }
607
-
608
- /**
609
- * Capture a coordinator event to the session file
610
- *
611
- * Appends the event as a JSONL line to ~/.config/swarm-tools/sessions/{session_id}.jsonl
612
- */
613
- export function captureCoordinatorEvent(event: CoordinatorEvent): void {
614
- // Validate event
615
- CoordinatorEventSchema.parse(event);
616
-
617
- // Ensure directory exists
618
- ensureSessionDir();
619
-
620
- // Append to session file
621
- const sessionPath = getSessionPath(event.session_id);
622
- const line = `${JSON.stringify(event)}\n`;
623
- fs.appendFileSync(sessionPath, line, "utf-8");
624
- }
625
-
626
- /**
627
- * Capture a compaction event to the session file
628
- *
629
- * Helper for capturing COMPACTION events with automatic timestamp generation.
630
- * Tracks compaction hook lifecycle: detection → prompt generation → context injection → resumption.
631
- *
632
- * **Part of eval-driven development pipeline:** Compaction events are used by `compaction-prompt.eval.ts`
633
- * to score prompt quality (ID specificity, actionability, coordinator identity).
634
- *
635
- * **Lifecycle stages:**
636
- * - `detection_complete` - Compaction detected (confidence level, context type)
637
- * - `prompt_generated` - Continuation prompt created (FULL content stored for eval)
638
- * - `context_injected` - Prompt injected into OpenCode context
639
- * - `resumption_started` - Coordinator resumed from checkpoint
640
- * - `tool_call_tracked` - First tool called post-compaction (measures discipline)
641
- *
642
- * @param params - Compaction event parameters
643
- * @param params.session_id - Coordinator session ID
644
- * @param params.epic_id - Epic ID being coordinated
645
- * @param params.compaction_type - Stage of compaction lifecycle
646
- * @param params.payload - Event-specific data (full prompt content, detection results, etc.)
647
- *
648
- * @example
649
- * // Capture detection complete
650
- * captureCompactionEvent({
651
- * session_id: "session-123",
652
- * epic_id: "bd-456",
653
- * compaction_type: "detection_complete",
654
- * payload: {
655
- * confidence: "high",
656
- * context_type: "full",
657
- * epic_id: "bd-456",
658
- * },
659
- * });
660
- *
661
- * @example
662
- * // Capture prompt generated (with full content for eval)
663
- * captureCompactionEvent({
664
- * session_id: "session-123",
665
- * epic_id: "bd-456",
666
- * compaction_type: "prompt_generated",
667
- * payload: {
668
- * prompt_length: 5000,
669
- * full_prompt: "You are a coordinator...", // Full prompt, not truncated - used for quality scoring
670
- * context_type: "full",
671
- * },
672
- * });
673
- */
674
- export function captureCompactionEvent(params: {
675
- session_id: string;
676
- epic_id: string;
677
- compaction_type:
678
- | "detection_complete"
679
- | "prompt_generated"
680
- | "context_injected"
681
- | "resumption_started"
682
- | "tool_call_tracked";
683
- payload: any;
684
- }): void {
685
- const event: CoordinatorEvent = {
686
- session_id: params.session_id,
687
- epic_id: params.epic_id,
688
- timestamp: new Date().toISOString(),
689
- event_type: "COMPACTION",
690
- compaction_type: params.compaction_type,
691
- payload: params.payload,
692
- };
693
-
694
- captureCoordinatorEvent(event);
695
- }
696
-
697
- /**
698
- * Capture a researcher spawned event
699
- *
700
- * Called when coordinator spawns a swarm-researcher to handle unfamiliar technology
701
- * or gather documentation before decomposition.
702
- */
703
- export function captureResearcherSpawned(params: {
704
- session_id: string;
705
- epic_id: string;
706
- researcher_id: string;
707
- research_topic: string;
708
- tools_used?: string[];
709
- }): void {
710
- const event: CoordinatorEvent = {
711
- session_id: params.session_id,
712
- epic_id: params.epic_id,
713
- timestamp: new Date().toISOString(),
714
- event_type: "DECISION",
715
- decision_type: "researcher_spawned",
716
- payload: {
717
- researcher_id: params.researcher_id,
718
- research_topic: params.research_topic,
719
- tools_used: params.tools_used || [],
720
- },
721
- };
722
-
723
- captureCoordinatorEvent(event);
724
- }
725
-
726
- /**
727
- * Capture a skill loaded event
728
- *
729
- * Called when coordinator loads domain knowledge via skills_use().
730
- */
731
- export function captureSkillLoaded(params: {
732
- session_id: string;
733
- epic_id: string;
734
- skill_name: string;
735
- context?: string;
736
- }): void {
737
- const event: CoordinatorEvent = {
738
- session_id: params.session_id,
739
- epic_id: params.epic_id,
740
- timestamp: new Date().toISOString(),
741
- event_type: "DECISION",
742
- decision_type: "skill_loaded",
743
- payload: {
744
- skill_name: params.skill_name,
745
- context: params.context,
746
- },
747
- };
748
-
749
- captureCoordinatorEvent(event);
750
- }
751
-
752
- /**
753
- * Capture an inbox checked event
754
- *
755
- * Called when coordinator checks swarmmail inbox for worker messages.
756
- * Tracks monitoring frequency and responsiveness.
757
- */
758
- export function captureInboxChecked(params: {
759
- session_id: string;
760
- epic_id: string;
761
- message_count: number;
762
- urgent_count: number;
763
- }): void {
764
- const event: CoordinatorEvent = {
765
- session_id: params.session_id,
766
- epic_id: params.epic_id,
767
- timestamp: new Date().toISOString(),
768
- event_type: "DECISION",
769
- decision_type: "inbox_checked",
770
- payload: {
771
- message_count: params.message_count,
772
- urgent_count: params.urgent_count,
773
- },
774
- };
775
-
776
- captureCoordinatorEvent(event);
777
- }
778
-
779
- /**
780
- * Capture a blocker resolved event
781
- *
782
- * Called when coordinator successfully unblocks a worker.
783
- */
784
- export function captureBlockerResolved(params: {
785
- session_id: string;
786
- epic_id: string;
787
- worker_id: string;
788
- subtask_id: string;
789
- blocker_type: string;
790
- resolution: string;
791
- }): void {
792
- const event: CoordinatorEvent = {
793
- session_id: params.session_id,
794
- epic_id: params.epic_id,
795
- timestamp: new Date().toISOString(),
796
- event_type: "DECISION",
797
- decision_type: "blocker_resolved",
798
- payload: {
799
- worker_id: params.worker_id,
800
- subtask_id: params.subtask_id,
801
- blocker_type: params.blocker_type,
802
- resolution: params.resolution,
803
- },
804
- };
805
-
806
- captureCoordinatorEvent(event);
807
- }
808
-
809
- /**
810
- * Capture a scope change decision event
811
- *
812
- * Called when coordinator approves or rejects a worker's scope expansion request.
813
- */
814
- export function captureScopeChangeDecision(params: {
815
- session_id: string;
816
- epic_id: string;
817
- worker_id: string;
818
- subtask_id: string;
819
- approved: boolean;
820
- original_scope?: string;
821
- new_scope?: string;
822
- requested_scope?: string;
823
- rejection_reason?: string;
824
- estimated_time_add?: number;
825
- }): void {
826
- const event: CoordinatorEvent = {
827
- session_id: params.session_id,
828
- epic_id: params.epic_id,
829
- timestamp: new Date().toISOString(),
830
- event_type: "DECISION",
831
- decision_type: params.approved ? "scope_change_approved" : "scope_change_rejected",
832
- payload: params.approved
833
- ? {
834
- worker_id: params.worker_id,
835
- subtask_id: params.subtask_id,
836
- original_scope: params.original_scope,
837
- new_scope: params.new_scope,
838
- estimated_time_add: params.estimated_time_add,
839
- }
840
- : {
841
- worker_id: params.worker_id,
842
- subtask_id: params.subtask_id,
843
- requested_scope: params.requested_scope,
844
- rejection_reason: params.rejection_reason,
845
- },
846
- };
847
-
848
- captureCoordinatorEvent(event);
849
- }
850
-
851
- /**
852
- * Capture a blocker detected event
853
- *
854
- * Called when a worker reports being blocked (OUTCOME event, not DECISION).
855
- */
856
- export function captureBlockerDetected(params: {
857
- session_id: string;
858
- epic_id: string;
859
- worker_id: string;
860
- subtask_id: string;
861
- blocker_type: string;
862
- blocker_description: string;
863
- }): void {
864
- const event: CoordinatorEvent = {
865
- session_id: params.session_id,
866
- epic_id: params.epic_id,
867
- timestamp: new Date().toISOString(),
868
- event_type: "OUTCOME",
869
- outcome_type: "blocker_detected",
870
- payload: {
871
- worker_id: params.worker_id,
872
- subtask_id: params.subtask_id,
873
- blocker_type: params.blocker_type,
874
- blocker_description: params.blocker_description,
875
- reported_at: new Date().toISOString(),
876
- },
877
- };
878
-
879
- captureCoordinatorEvent(event);
880
- }
881
-
882
- /**
883
- * Read all events from a session file
884
- */
885
- export function readSessionEvents(sessionId: string): CoordinatorEvent[] {
886
- const sessionPath = getSessionPath(sessionId);
887
- if (!fs.existsSync(sessionPath)) {
888
- return [];
889
- }
890
-
891
- const content = fs.readFileSync(sessionPath, "utf-8");
892
- const lines = content.trim().split("\n").filter(Boolean);
893
-
894
- return lines.map((line) => {
895
- const parsed = JSON.parse(line);
896
- return CoordinatorEventSchema.parse(parsed);
897
- });
898
- }
899
-
900
- /**
901
- * Save a session - wraps all events in a CoordinatorSession structure
902
- *
903
- * Reads all events from the session file and wraps them in a session object.
904
- * Returns null if the session file doesn't exist.
905
- */
906
- export function saveSession(params: {
907
- session_id: string;
908
- epic_id: string;
909
- }): CoordinatorSession | null {
910
- const events = readSessionEvents(params.session_id);
911
- if (events.length === 0) {
912
- return null;
913
- }
914
-
915
- // Get timestamps from events
916
- const timestamps = events.map((e) => new Date(e.timestamp).getTime());
917
- const startTime = new Date(Math.min(...timestamps)).toISOString();
918
- const endTime = new Date(Math.max(...timestamps)).toISOString();
919
-
920
- const session: CoordinatorSession = {
921
- session_id: params.session_id,
922
- epic_id: params.epic_id,
923
- start_time: startTime,
924
- end_time: endTime,
925
- events,
926
- };
927
-
928
- return session;
929
- }