crewswarm 0.9.1 → 0.9.3

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 (210) hide show
  1. package/README.md +22 -9
  2. package/apps/dashboard/dist/assets/{chat-core-Cx4sTxDd.js → chat-core-3KirthZA.js} +1 -1
  3. package/apps/dashboard/dist/assets/index-GSWxxEPO.js +2 -0
  4. package/apps/dashboard/dist/assets/{tab-pm-loop-tab-Bfd449B4.js → tab-pm-loop-tab-DiAPTJXu.js} +1 -1
  5. package/apps/dashboard/dist/assets/{tab-projects-tab-DhNWnlzt.js → tab-projects-tab-SFH4E--a.js} +1 -1
  6. package/apps/dashboard/dist/assets/tab-settings-tab-BselH1c0.js +1 -0
  7. package/apps/dashboard/dist/index.html +82 -11
  8. package/apps/vibe/README.md +2 -2
  9. package/apps/vibe/package.json +1 -1
  10. package/apps/vibe/server.mjs +3 -3
  11. package/crew-lead.mjs +48 -5
  12. package/lib/bridges/gateway-ws.mjs +4 -0
  13. package/lib/bridges/tmux-bridge.mjs +200 -0
  14. package/lib/cli-process-tracker.mjs +2 -1
  15. package/lib/crew-lead/chat-handler.mjs +34 -0
  16. package/lib/crew-lead/http-server.mjs +340 -14
  17. package/lib/crew-lead/llm-caller.mjs +24 -8
  18. package/lib/crew-lead/prompts.mjs +7 -0
  19. package/lib/crew-lead/wave-dispatcher.mjs +53 -3
  20. package/lib/crew-lead/ws-router.mjs +219 -27
  21. package/lib/engines/engine-registry.mjs +9 -0
  22. package/lib/engines/rt-envelope.mjs +1 -0
  23. package/lib/engines/runners.mjs +26 -2
  24. package/lib/runtime/config.mjs +7 -0
  25. package/lib/runtime/paths.mjs +12 -8
  26. package/lib/sessions/session-manager.mjs +287 -0
  27. package/package.json +35 -15
  28. package/scripts/capture-build-flow.mjs +118 -0
  29. package/scripts/coverage-report.mjs +209 -0
  30. package/scripts/coverage-summary.mjs +47 -0
  31. package/scripts/dashboard-validation.mjs +74 -0
  32. package/scripts/dashboard.mjs +560 -70
  33. package/scripts/live-bridge-matrix.mjs +79 -0
  34. package/scripts/live-cli-matrix.mjs +166 -0
  35. package/scripts/live-crewchat-check.mjs +42 -0
  36. package/scripts/live-engine-matrix.mjs +50 -0
  37. package/scripts/live-provider-failover-matrix.mjs +107 -0
  38. package/scripts/live-provider-matrix.mjs +228 -0
  39. package/scripts/restart-all-from-repo.sh +4 -4
  40. package/scripts/smoke-dispatch.mjs +4 -1
  41. package/scripts/test-blast-radius.mjs +204 -0
  42. package/scripts/test-report-summary.mjs +88 -0
  43. package/scripts/test-reporter.mjs +651 -0
  44. package/scripts/test-rerun.mjs +136 -0
  45. package/scripts/tmux-bridge +130 -0
  46. package/apps/dashboard/dist/assets/chat-core-Cx4sTxDd.js.br +0 -0
  47. package/apps/dashboard/dist/assets/cli-process-COMRNPqr.js.br +0 -0
  48. package/apps/dashboard/dist/assets/components-BS9fQjE_.js.br +0 -0
  49. package/apps/dashboard/dist/assets/core-utils-CmOkXgzi.js.br +0 -0
  50. package/apps/dashboard/dist/assets/index-CF0aJRtC.css.br +0 -0
  51. package/apps/dashboard/dist/assets/index-DnClJ1ee.js +0 -2
  52. package/apps/dashboard/dist/assets/index-DnClJ1ee.js.br +0 -0
  53. package/apps/dashboard/dist/assets/orchestration-Ca2DLWN-.js.br +0 -0
  54. package/apps/dashboard/dist/assets/setup-wizard-CA0Or47w.js.br +0 -0
  55. package/apps/dashboard/dist/assets/tab-agents-tab-BgpIsjkw.js.br +0 -0
  56. package/apps/dashboard/dist/assets/tab-comms-tab-kguqTIzD.js.br +0 -0
  57. package/apps/dashboard/dist/assets/tab-contacts-tab-DiOyMYth.js.br +0 -0
  58. package/apps/dashboard/dist/assets/tab-engines-tab-BsdZVvU0.js.br +0 -0
  59. package/apps/dashboard/dist/assets/tab-memory-tab-Cu6u13EQ.js.br +0 -0
  60. package/apps/dashboard/dist/assets/tab-models-tab-BLEjmd19.js.br +0 -0
  61. package/apps/dashboard/dist/assets/tab-pm-loop-tab-Bfd449B4.js.br +0 -0
  62. package/apps/dashboard/dist/assets/tab-projects-tab-DhNWnlzt.js.br +0 -0
  63. package/apps/dashboard/dist/assets/tab-prompts-tab-DVkUNaJd.js.br +0 -0
  64. package/apps/dashboard/dist/assets/tab-services-tab-DU_LH3uG.js.br +0 -0
  65. package/apps/dashboard/dist/assets/tab-settings-tab-Bn4nXtDe.js +0 -1
  66. package/apps/dashboard/dist/assets/tab-settings-tab-Bn4nXtDe.js.br +0 -0
  67. package/apps/dashboard/dist/assets/tab-skills-tab-BpY0uZHW.js.br +0 -0
  68. package/apps/dashboard/dist/assets/tab-spending-tab-DEccQHnt.js.br +0 -0
  69. package/apps/dashboard/dist/assets/tab-swarm-chat-tab-BNrd88-r.js.br +0 -0
  70. package/apps/dashboard/dist/assets/tab-swarm-tab-B1AcjL1W.js.br +0 -0
  71. package/apps/dashboard/dist/assets/tab-usage-tab-BIOOnB-Y.js.br +0 -0
  72. package/apps/dashboard/dist/assets/tab-waves-tab-SaJDkb4x.js.br +0 -0
  73. package/apps/dashboard/dist/assets/tab-workflows-tab-B-soSy1k.js.br +0 -0
  74. package/apps/dashboard/dist/index.html.br +0 -0
  75. package/apps/dashboard/index.html +0 -6459
  76. package/apps/dashboard/package.json +0 -15
  77. package/apps/dashboard/src/app.js +0 -2823
  78. package/apps/dashboard/src/app.js.br +0 -0
  79. package/apps/dashboard/src/app.js.gz +0 -0
  80. package/apps/dashboard/src/chat/chat-actions.js +0 -1847
  81. package/apps/dashboard/src/chat/chat-actions.js.br +0 -0
  82. package/apps/dashboard/src/chat/unified-messages.js +0 -327
  83. package/apps/dashboard/src/chat/unified-messages.js.br +0 -0
  84. package/apps/dashboard/src/cli-process.js +0 -208
  85. package/apps/dashboard/src/cli-process.js.br +0 -0
  86. package/apps/dashboard/src/cli-process.js.gz +0 -0
  87. package/apps/dashboard/src/components/active-tasks-panel.js +0 -175
  88. package/apps/dashboard/src/components/active-tasks-panel.js.br +0 -0
  89. package/apps/dashboard/src/core/api.js +0 -18
  90. package/apps/dashboard/src/core/api.js.br +0 -0
  91. package/apps/dashboard/src/core/dom.js +0 -228
  92. package/apps/dashboard/src/core/dom.js.br +0 -0
  93. package/apps/dashboard/src/core/state.js +0 -91
  94. package/apps/dashboard/src/core/state.js.br +0 -0
  95. package/apps/dashboard/src/core/task-manager.js +0 -134
  96. package/apps/dashboard/src/core/task-manager.js.br +0 -0
  97. package/apps/dashboard/src/orchestration-status.js +0 -127
  98. package/apps/dashboard/src/orchestration-status.js.br +0 -0
  99. package/apps/dashboard/src/setup-wizard.js +0 -562
  100. package/apps/dashboard/src/setup-wizard.js.br +0 -0
  101. package/apps/dashboard/src/styles.css +0 -2085
  102. package/apps/dashboard/src/styles.css.br +0 -0
  103. package/apps/dashboard/src/styles.css.gz +0 -0
  104. package/apps/dashboard/src/tabs/agents-tab.js +0 -2237
  105. package/apps/dashboard/src/tabs/agents-tab.js.br +0 -0
  106. package/apps/dashboard/src/tabs/benchmarks-tab.js +0 -229
  107. package/apps/dashboard/src/tabs/benchmarks-tab.js.br +0 -0
  108. package/apps/dashboard/src/tabs/comms-tab.js +0 -955
  109. package/apps/dashboard/src/tabs/comms-tab.js.br +0 -0
  110. package/apps/dashboard/src/tabs/contacts-tab.js +0 -654
  111. package/apps/dashboard/src/tabs/contacts-tab.js.br +0 -0
  112. package/apps/dashboard/src/tabs/engines-tab.js +0 -175
  113. package/apps/dashboard/src/tabs/engines-tab.js.br +0 -0
  114. package/apps/dashboard/src/tabs/memory-tab.js +0 -182
  115. package/apps/dashboard/src/tabs/memory-tab.js.br +0 -0
  116. package/apps/dashboard/src/tabs/models-tab.js +0 -450
  117. package/apps/dashboard/src/tabs/models-tab.js.br +0 -0
  118. package/apps/dashboard/src/tabs/pm-loop-tab.js +0 -185
  119. package/apps/dashboard/src/tabs/pm-loop-tab.js.br +0 -0
  120. package/apps/dashboard/src/tabs/projects-tab.js +0 -663
  121. package/apps/dashboard/src/tabs/projects-tab.js.br +0 -0
  122. package/apps/dashboard/src/tabs/projects-tab.js.gz +0 -0
  123. package/apps/dashboard/src/tabs/prompts-tab.js +0 -160
  124. package/apps/dashboard/src/tabs/prompts-tab.js.br +0 -0
  125. package/apps/dashboard/src/tabs/services-tab.js +0 -202
  126. package/apps/dashboard/src/tabs/services-tab.js.br +0 -0
  127. package/apps/dashboard/src/tabs/settings-tab.js +0 -803
  128. package/apps/dashboard/src/tabs/settings-tab.js.br +0 -0
  129. package/apps/dashboard/src/tabs/skills-tab.js +0 -284
  130. package/apps/dashboard/src/tabs/skills-tab.js.br +0 -0
  131. package/apps/dashboard/src/tabs/spending-tab.js +0 -173
  132. package/apps/dashboard/src/tabs/spending-tab.js.br +0 -0
  133. package/apps/dashboard/src/tabs/swarm-chat-tab.js +0 -660
  134. package/apps/dashboard/src/tabs/swarm-chat-tab.js.br +0 -0
  135. package/apps/dashboard/src/tabs/swarm-tab.js +0 -538
  136. package/apps/dashboard/src/tabs/swarm-tab.js.br +0 -0
  137. package/apps/dashboard/src/tabs/usage-tab.js +0 -390
  138. package/apps/dashboard/src/tabs/usage-tab.js.br +0 -0
  139. package/apps/dashboard/src/tabs/waves-tab.js +0 -238
  140. package/apps/dashboard/src/tabs/waves-tab.js.br +0 -0
  141. package/apps/dashboard/src/tabs/workflows-tab.js +0 -747
  142. package/apps/dashboard/src/tabs/workflows-tab.js.br +0 -0
  143. package/apps/vibe/.crew/agent-memory/pipeline.json +0 -304
  144. package/apps/vibe/.crew/cost.json +0 -17
  145. package/apps/vibe/.crew/json-parse-metrics.jsonl +0 -27
  146. package/apps/vibe/.crew/pipeline-metrics.jsonl +0 -27
  147. package/apps/vibe/.crew/pipeline-runs/pipeline-0f90c392-2425-4ae5-850c-bd9d17b1d690.jsonl +0 -5
  148. package/apps/vibe/.crew/pipeline-runs/pipeline-1c269dd9-a63f-4fba-af81-5cf08048ef06.jsonl +0 -5
  149. package/apps/vibe/.crew/pipeline-runs/pipeline-288a7765-da24-4a22-89bc-1f3cc9b0562c.jsonl +0 -5
  150. package/apps/vibe/.crew/pipeline-runs/pipeline-2c78fd22-a657-4bd1-bc49-0679fb384409.jsonl +0 -5
  151. package/apps/vibe/.crew/pipeline-runs/pipeline-3da23550-22ed-4904-9a0a-8e79c1f3024c.jsonl +0 -5
  152. package/apps/vibe/.crew/pipeline-runs/pipeline-3e6fe08d-3264-404a-8df3-aab7efef10e7.jsonl +0 -5
  153. package/apps/vibe/.crew/pipeline-runs/pipeline-42eec610-57fe-4e09-9e7e-b315038495c2.jsonl +0 -5
  154. package/apps/vibe/.crew/pipeline-runs/pipeline-4438eb4c-ae13-42b1-90e2-b043d8983be8.jsonl +0 -5
  155. package/apps/vibe/.crew/pipeline-runs/pipeline-4740a9f5-86e7-44b6-a394-de433e291727.jsonl +0 -5
  156. package/apps/vibe/.crew/pipeline-runs/pipeline-49e1da6a-957e-48fd-9220-415019e4f8e2.jsonl +0 -5
  157. package/apps/vibe/.crew/pipeline-runs/pipeline-4c9251db-be68-427b-a3fc-a264f2b5778d.jsonl +0 -5
  158. package/apps/vibe/.crew/pipeline-runs/pipeline-6413fa33-a802-4b57-a8c0-a9056ad67842.jsonl +0 -5
  159. package/apps/vibe/.crew/pipeline-runs/pipeline-65e29a57-664d-4196-8109-017e364f182e.jsonl +0 -5
  160. package/apps/vibe/.crew/pipeline-runs/pipeline-6aa04bc5-9593-4b1f-b58d-3bf2978cb602.jsonl +0 -5
  161. package/apps/vibe/.crew/pipeline-runs/pipeline-6e1cba53-9b70-457e-99e0-59199149dd21.jsonl +0 -5
  162. package/apps/vibe/.crew/pipeline-runs/pipeline-749f41cc-4dac-4204-be64-873a6080a0d2.jsonl +0 -5
  163. package/apps/vibe/.crew/pipeline-runs/pipeline-74d68121-e181-4864-bd9a-c3211341dfaf.jsonl +0 -5
  164. package/apps/vibe/.crew/pipeline-runs/pipeline-8509bc24-142d-4e07-b44a-a50bf99d1103.jsonl +0 -5
  165. package/apps/vibe/.crew/pipeline-runs/pipeline-960339c6-07ca-43ce-9900-f6e1702b39b9.jsonl +0 -5
  166. package/apps/vibe/.crew/pipeline-runs/pipeline-9bef2dd2-6122-42e5-b3d9-19f4d80f9e40.jsonl +0 -5
  167. package/apps/vibe/.crew/pipeline-runs/pipeline-9c6480a9-7031-4146-b241-825b9a2d1de1.jsonl +0 -5
  168. package/apps/vibe/.crew/pipeline-runs/pipeline-9fd42426-8492-4157-9d5f-e1537c060489.jsonl +0 -2
  169. package/apps/vibe/.crew/pipeline-runs/pipeline-ad6d40a3-2f5e-46a9-a345-47caaccc51aa.jsonl +0 -5
  170. package/apps/vibe/.crew/pipeline-runs/pipeline-bc606133-8d5b-4535-8d85-f1a29cdaa981.jsonl +0 -5
  171. package/apps/vibe/.crew/pipeline-runs/pipeline-c1418f4e-b773-4ca1-84a3-216acf36e2f2.jsonl +0 -5
  172. package/apps/vibe/.crew/pipeline-runs/pipeline-c1a13ccd-634a-4d01-a4a7-1177b8a752ff.jsonl +0 -5
  173. package/apps/vibe/.crew/pipeline-runs/pipeline-c7d27b42-249e-4bd4-8f26-6aa998110b8a.jsonl +0 -5
  174. package/apps/vibe/.crew/pipeline-runs/pipeline-cca2e9b9-4a34-4d25-a311-5c793fa7e91e.jsonl +0 -5
  175. package/apps/vibe/.crew/sandbox.json +0 -7
  176. package/apps/vibe/.crew/session.json +0 -330
  177. package/apps/vibe/.crew/training-data.jsonl +0 -0
  178. package/apps/vibe/.github/workflows/studio-quality.yml +0 -37
  179. package/apps/vibe/.studio-data/project-messages/chuck-norris.jsonl +0 -18
  180. package/apps/vibe/.studio-data/project-messages/general.jsonl +0 -81
  181. package/apps/vibe/.studio-data/project-messages/studio-local.jsonl +0 -18
  182. package/apps/vibe/ARCHITECTURE.md +0 -3393
  183. package/apps/vibe/QUICK-REFERENCE.md +0 -211
  184. package/apps/vibe/ROADMAP.md +0 -41
  185. package/apps/vibe/STUDIO-SETUP-COMPLETE.md +0 -35
  186. package/apps/vibe/VISUAL-GUIDE.md +0 -378
  187. package/apps/vibe/capture-demo.mjs +0 -160
  188. package/apps/vibe/capture-full-demo.mjs +0 -255
  189. package/apps/vibe/capture-quickstart.mjs +0 -256
  190. package/apps/vibe/capture-vibe-assets.mjs +0 -71
  191. package/apps/vibe/capture-vibe-video.mjs +0 -260
  192. package/apps/vibe/check-buttons.js +0 -41
  193. package/apps/vibe/diagnose.html +0 -106
  194. package/apps/vibe/fix-buttons.js +0 -103
  195. package/apps/vibe/index.html +0 -3404
  196. package/apps/vibe/package-lock.json +0 -920
  197. package/apps/vibe/scripts/studio-pty-host.py +0 -117
  198. package/apps/vibe/src/main.js +0 -2940
  199. package/apps/vibe/src/register-all-languages.js +0 -98
  200. package/apps/vibe/start-studio.sh +0 -11
  201. package/apps/vibe/test/accessibility-tests.js +0 -77
  202. package/apps/vibe/test/browser-performance-audit.mjs +0 -205
  203. package/apps/vibe/test/performance-tests.js +0 -120
  204. package/apps/vibe/test/security-tests.js +0 -213
  205. package/apps/vibe/tests/e2e.local.mjs +0 -54
  206. package/apps/vibe/tests/server.smoke.mjs +0 -106
  207. package/apps/vibe/update_website.mjs +0 -74
  208. package/apps/vibe/vite.config.js +0 -19
  209. package/lib/crew-lead/chat-handler.mjs.bak +0 -1274
  210. package/lib/engines/rt-envelope.mjs.backup-current +0 -870
@@ -0,0 +1,651 @@
1
+ #!/usr/bin/env node
2
+
3
+ /**
4
+ * Custom Node.js test reporter that logs every test result (pass/fail/skip)
5
+ * with timestamps to a JSONL file at test-results/test-log.jsonl.
6
+ *
7
+ * Also updates test-results/.last-run.json with summary + timestamp.
8
+ *
9
+ * Usage: node --test --test-reporter=./scripts/test-reporter.mjs test/unit/*.test.mjs
10
+ */
11
+
12
+ import fs from "node:fs";
13
+ import os from "node:os";
14
+ import path from "node:path";
15
+ import crypto from "node:crypto";
16
+ import { execSync } from "node:child_process";
17
+ import { buildDependencySnapshot, getWorkspaceState } from "./test-blast-radius.mjs";
18
+
19
+ const ROOT = process.env.TEST_RESULTS_DIR || path.join(process.cwd(), "test-results");
20
+ const LOG_PATH = path.join(ROOT, "test-log.jsonl");
21
+ const SUMMARY_PATH = path.join(ROOT, ".last-run.json");
22
+ const CURRENT_RUN_PATH = path.join(ROOT, ".current-run.json");
23
+ const RUNS_DIR = path.join(ROOT, "runs");
24
+
25
+ fs.mkdirSync(ROOT, { recursive: true });
26
+ fs.mkdirSync(RUNS_DIR, { recursive: true });
27
+
28
+ const runId = new Date().toISOString().replace(/[:.]/g, "-");
29
+ const results = [];
30
+ let passed = 0;
31
+ let failed = 0;
32
+ let skipped = 0;
33
+ let totalDuration = 0;
34
+ const fileFingerprintCache = new Map();
35
+ const testArtifacts = new Map();
36
+ const workspaceStateAtRunStart = getWorkspaceState();
37
+
38
+ function safeExec(command) {
39
+ try {
40
+ return execSync(command, {
41
+ cwd: process.cwd(),
42
+ encoding: "utf8",
43
+ stdio: ["ignore", "pipe", "ignore"],
44
+ timeout: 5000,
45
+ }).trim();
46
+ } catch {
47
+ return null;
48
+ }
49
+ }
50
+
51
+ function truncate(value, limit = 800) {
52
+ const text = String(value ?? "");
53
+ return text.length <= limit ? text : `${text.slice(0, limit)}...`;
54
+ }
55
+
56
+ function compactText(value) {
57
+ return truncate(String(value ?? "").replace(/\s+/g, " ").trim());
58
+ }
59
+
60
+ function shellQuote(value) {
61
+ return `'${String(value).replace(/'/g, `'\\''`)}'`;
62
+ }
63
+
64
+ function slugify(value) {
65
+ return String(value || "unnamed")
66
+ .toLowerCase()
67
+ .replace(/[^a-z0-9]+/g, "-")
68
+ .replace(/^-+|-+$/g, "")
69
+ .slice(0, 80) || "unnamed";
70
+ }
71
+
72
+ function buildTestId(filePath, name) {
73
+ return `${slugify(path.relative(process.cwd(), filePath || "no-file"))}__${slugify(name || "unnamed-test")}`;
74
+ }
75
+
76
+ function buildRerunCommand(filePath, testName) {
77
+ return `node --test --test-reporter=./scripts/test-reporter.mjs --test-name-pattern=${shellQuote(testName)} ${shellQuote(filePath)}`;
78
+ }
79
+
80
+ function buildSelector(filePath, testName) {
81
+ return {
82
+ file: filePath || null,
83
+ test_name: testName || null,
84
+ test_name_pattern: testName || null,
85
+ command: buildRerunCommand(filePath, testName),
86
+ };
87
+ }
88
+
89
+ function getArtifactDir(testId) {
90
+ return path.join(RUNS_DIR, runId, testId);
91
+ }
92
+
93
+ function ensureArtifactDir(testId) {
94
+ const artifactDir = getArtifactDir(testId);
95
+ fs.mkdirSync(artifactDir, { recursive: true });
96
+ return artifactDir;
97
+ }
98
+
99
+ function writeArtifactJson(testId, filename, value) {
100
+ const artifactDir = ensureArtifactDir(testId);
101
+ fs.writeFileSync(path.join(artifactDir, filename), JSON.stringify(value, null, 2) + "\n");
102
+ }
103
+
104
+ function artifactPath(testId, filename) {
105
+ return path.join(getArtifactDir(testId), filename);
106
+ }
107
+
108
+ function upsertTestArtifact(entry) {
109
+ const current = testArtifacts.get(entry.testId) || {};
110
+ testArtifacts.set(entry.testId, { ...current, ...entry });
111
+ }
112
+
113
+ function fingerprintFile(filePath) {
114
+ if (!filePath) return null;
115
+ if (fileFingerprintCache.has(filePath)) return fileFingerprintCache.get(filePath);
116
+ try {
117
+ const stat = fs.statSync(filePath);
118
+ const content = fs.readFileSync(filePath);
119
+ const fingerprint = {
120
+ file: filePath,
121
+ relative_file: path.relative(process.cwd(), filePath),
122
+ size_bytes: stat.size,
123
+ mtime: stat.mtime.toISOString(),
124
+ sha256: crypto.createHash("sha256").update(content).digest("hex"),
125
+ git_blob: safeExec(`git hash-object ${JSON.stringify(filePath)}`),
126
+ };
127
+ fileFingerprintCache.set(filePath, fingerprint);
128
+ return fingerprint;
129
+ } catch {
130
+ return null;
131
+ }
132
+ }
133
+
134
+ function readEvidence(testId) {
135
+ const evidencePath = artifactPath(testId, "evidence.jsonl");
136
+ if (!fs.existsSync(evidencePath)) return [];
137
+ try {
138
+ return fs.readFileSync(evidencePath, "utf8")
139
+ .split("\n")
140
+ .filter(Boolean)
141
+ .map((line) => JSON.parse(line));
142
+ } catch {
143
+ return [];
144
+ }
145
+ }
146
+
147
+ function collectTestContext(testId) {
148
+ const evidence = readEvidence(testId);
149
+ const engineContext = [...evidence].reverse().find((entry) => entry.category === "engine_context") || null;
150
+ const latestHttp = [...evidence].reverse().find((entry) => entry.category === "http") || null;
151
+ const latestFileVerification = [...evidence].reverse().find((entry) => entry.category === "file_verification") || null;
152
+ const latestTimeout = [...evidence].reverse().find((entry) =>
153
+ entry.category === "task_timeout" || entry.category === "passthrough_timeout"
154
+ ) || null;
155
+ const latestPassthroughError = [...evidence].reverse().find((entry) => entry.category === "passthrough_error") || null;
156
+ const latestWorkflowEvent = [...evidence].reverse().find((entry) =>
157
+ /workflow|dispatch|task/i.test(String(entry.category || ""))
158
+ ) || null;
159
+ return {
160
+ evidence_count: evidence.length,
161
+ engine_context: engineContext,
162
+ latest_http: latestHttp,
163
+ latest_file_verification: latestFileVerification,
164
+ latest_timeout: latestTimeout,
165
+ latest_passthrough_error: latestPassthroughError,
166
+ latest_workflow_event: latestWorkflowEvent,
167
+ };
168
+ }
169
+
170
+ function extractSkipReason(data = {}) {
171
+ return data.skipReason ||
172
+ data.skip ||
173
+ data.details?.skipReason ||
174
+ data.details?.skip ||
175
+ (data.details?.skipped ? data.details?.message || "skipped" : null) ||
176
+ (data.details?.status === "skipped" ? data.details?.message || "skipped" : null) ||
177
+ null;
178
+ }
179
+
180
+ function classifyFailure(entry, context) {
181
+ const errorText = compactText([
182
+ entry.error,
183
+ entry.error_name,
184
+ entry.error_code,
185
+ context.latest_timeout?.error,
186
+ context.latest_passthrough_error?.error,
187
+ context.latest_http?.error,
188
+ context.latest_workflow_event?.last_result_preview,
189
+ ].filter(Boolean).join(" | "));
190
+
191
+ const checks = [
192
+ {
193
+ code: "timeout",
194
+ matches: /timeout|timed out|did not finish|deadline exceeded/i,
195
+ summary: "test or engine timed out",
196
+ },
197
+ {
198
+ code: "cancelled",
199
+ matches: /cancelled|canceled|abort|aborted/i,
200
+ summary: "test was cancelled",
201
+ },
202
+ {
203
+ code: "agent_unreachable",
204
+ matches: /rt bus not connected|agent unreachable|econnrefused|connect econrefused|socket hang up/i,
205
+ summary: "agent or service was unreachable",
206
+ },
207
+ {
208
+ code: "missing_task_id",
209
+ matches: /no taskid|missing taskid/i,
210
+ summary: "dispatch returned no task id",
211
+ },
212
+ {
213
+ code: "file_missing",
214
+ matches: /should exist|file not found|enoent/i,
215
+ summary: "expected output file was missing",
216
+ },
217
+ {
218
+ code: "empty_response",
219
+ matches: /non-empty response|returned nothing|response length/i,
220
+ summary: "engine returned an empty or unusable response",
221
+ },
222
+ {
223
+ code: "http_error",
224
+ matches: /dispatch failed: \d+|http|status code|unexpected status/i,
225
+ summary: "http request or api contract failed",
226
+ },
227
+ {
228
+ code: "assertion",
229
+ matches: /assert|expected .* got/i,
230
+ summary: "assertion failed",
231
+ },
232
+ ];
233
+
234
+ const match = checks.find((item) => item.matches.test(errorText));
235
+ return {
236
+ reason_code: match?.code || "unknown_failure",
237
+ reason_summary: match?.summary || "unclassified failure",
238
+ reason_detail: errorText || "no detailed failure message captured",
239
+ };
240
+ }
241
+
242
+ function classifySkip(entry) {
243
+ const reasonText = compactText(entry.skip_reason || "skipped");
244
+ const checks = [
245
+ {
246
+ code: "service_down",
247
+ matches: /not running|service unavailable|health check/i,
248
+ summary: "required local service was not running",
249
+ },
250
+ {
251
+ code: "engine_unavailable",
252
+ matches: /not available|not installed|missing/i,
253
+ summary: "required engine or dependency was unavailable",
254
+ },
255
+ {
256
+ code: "explicit_skip",
257
+ matches: /skip|skipped/i,
258
+ summary: "test was skipped intentionally",
259
+ },
260
+ ];
261
+ const match = checks.find((item) => item.matches.test(reasonText));
262
+ return {
263
+ reason_code: match?.code || "unknown_skip",
264
+ reason_summary: match?.summary || "unclassified skip",
265
+ reason_detail: reasonText,
266
+ };
267
+ }
268
+
269
+ export { buildSelector, classifyFailure, classifySkip };
270
+
271
+ function buildFailureBundle(entry) {
272
+ const context = collectTestContext(entry.testId);
273
+ const classification = classifyFailure(entry, context);
274
+ const engineRuntime = context.engine_context?.engine_runtime || {};
275
+ const agentRuntime = context.engine_context?.agent_runtime || {};
276
+ return {
277
+ ...entry,
278
+ ...classification,
279
+ selector: buildSelector(entry.file, entry.name),
280
+ workspace_state_at_run_start: workspaceStateAtRunStart,
281
+ isolation: {
282
+ isolated_rerun_command: entry.rerun_command,
283
+ latest_http_status: context.latest_http?.status ?? null,
284
+ latest_http_operation: context.latest_http?.operation ?? null,
285
+ latest_timeout_ms: context.latest_timeout?.timeout_ms ?? context.engine_context?.timeout_ms ?? null,
286
+ evidence_count: context.evidence_count,
287
+ },
288
+ engine: {
289
+ engine: context.engine_context?.engine || entry.engine || null,
290
+ provider: engineRuntime.provider || agentRuntime.provider || null,
291
+ model: agentRuntime.model || agentRuntime.cursorCliModel || agentRuntime.opencodeModel || null,
292
+ agent: context.engine_context?.agent || null,
293
+ binary_path: engineRuntime.binary?.path || null,
294
+ binary_version: engineRuntime.binary?.version || null,
295
+ route_flags: agentRuntime.routeFlags || null,
296
+ enabled_route: agentRuntime.enabledRoute || null,
297
+ },
298
+ artifacts: {
299
+ artifact_dir: entry.artifactDir,
300
+ manifest_json: artifactPath(entry.testId, "manifest.json"),
301
+ failure_json: artifactPath(entry.testId, "failure.json"),
302
+ evidence_jsonl: artifactPath(entry.testId, "evidence.jsonl"),
303
+ triage_json: artifactPath(entry.testId, "triage.json"),
304
+ },
305
+ latest_evidence: {
306
+ http: context.latest_http || null,
307
+ file_verification: context.latest_file_verification || null,
308
+ timeout: context.latest_timeout || null,
309
+ passthrough_error: context.latest_passthrough_error || null,
310
+ workflow: context.latest_workflow_event || null,
311
+ },
312
+ };
313
+ }
314
+
315
+ function buildSkipBundle(entry) {
316
+ const classification = classifySkip(entry);
317
+ return {
318
+ ...entry,
319
+ ...classification,
320
+ selector: buildSelector(entry.file, entry.name),
321
+ workspace_state_at_run_start: workspaceStateAtRunStart,
322
+ isolation: {
323
+ isolated_rerun_command: entry.rerun_command,
324
+ },
325
+ artifacts: {
326
+ artifact_dir: entry.artifactDir,
327
+ manifest_json: artifactPath(entry.testId, "manifest.json"),
328
+ evidence_jsonl: artifactPath(entry.testId, "evidence.jsonl"),
329
+ triage_json: artifactPath(entry.testId, "triage.json"),
330
+ },
331
+ };
332
+ }
333
+
334
+ const runMeta = {
335
+ runId,
336
+ entry_type: "run",
337
+ phase: "start",
338
+ timestamp: new Date().toISOString(),
339
+ cwd: process.cwd(),
340
+ hostname: os.hostname(),
341
+ platform: process.platform,
342
+ arch: process.arch,
343
+ node_version: process.version,
344
+ test_results_dir: ROOT,
345
+ test_command: process.argv.join(" "),
346
+ git_branch: safeExec("git branch --show-current"),
347
+ git_commit: safeExec("git rev-parse HEAD"),
348
+ git_dirty: !!safeExec("git status --short"),
349
+ workspace_state: workspaceStateAtRunStart,
350
+ package_version: (() => {
351
+ try {
352
+ return JSON.parse(fs.readFileSync(path.join(process.cwd(), "package.json"), "utf8")).version || null;
353
+ } catch {
354
+ return null;
355
+ }
356
+ })(),
357
+ };
358
+
359
+ fs.writeFileSync(CURRENT_RUN_PATH, JSON.stringify(runMeta, null, 2) + "\n");
360
+ fs.mkdirSync(path.join(RUNS_DIR, runId), { recursive: true });
361
+ fs.writeFileSync(path.join(RUNS_DIR, runId, "run.json"), JSON.stringify(runMeta, null, 2) + "\n");
362
+
363
+ export default async function* reporter(source) {
364
+ const logStream = fs.createWriteStream(LOG_PATH, { flags: "a" });
365
+ logStream.write(JSON.stringify(runMeta) + "\n");
366
+
367
+ for await (const event of source) {
368
+ if (event.type === "test:start" && event.data?.details?.type !== "suite") {
369
+ const testId = buildTestId(event.data.file, event.data.name);
370
+ const artifactDir = ensureArtifactDir(testId);
371
+ const startEntry = {
372
+ runId,
373
+ timestamp: new Date().toISOString(),
374
+ entry_type: "result",
375
+ phase: "start",
376
+ status: "start",
377
+ testId,
378
+ artifactDir,
379
+ name: event.data.name,
380
+ file: event.data.file,
381
+ file_fingerprint: fingerprintFile(event.data.file),
382
+ details_type: event.data.details?.type || null,
383
+ nesting: event.data.nesting ?? null,
384
+ line: event.data.line ?? null,
385
+ column: event.data.column ?? null,
386
+ rerun_command: buildRerunCommand(event.data.file, event.data.name),
387
+ selector: buildSelector(event.data.file, event.data.name),
388
+ dependency_snapshot: buildDependencySnapshot(event.data.file),
389
+ };
390
+ upsertTestArtifact({
391
+ testId,
392
+ artifactDir,
393
+ name: event.data.name,
394
+ file: event.data.file,
395
+ status: "start",
396
+ line: event.data.line ?? null,
397
+ column: event.data.column ?? null,
398
+ dependency_snapshot: startEntry.dependency_snapshot,
399
+ });
400
+ writeArtifactJson(testId, "manifest.json", startEntry);
401
+ logStream.write(JSON.stringify(startEntry) + "\n");
402
+ }
403
+
404
+ if (event.type === "test:pass" || event.type === "test:fail") {
405
+ // Only log leaf tests (not suite/describe wrappers)
406
+ if (event.data.details?.type === "suite") continue;
407
+
408
+ const inferredSkipReason = event.type === "test:pass" ? extractSkipReason(event.data) : null;
409
+ const status = inferredSkipReason ? "skip" : event.type === "test:pass" ? "pass" : "fail";
410
+ const duration = event.data.details?.duration_ms ?? 0;
411
+ totalDuration += duration;
412
+
413
+ if (status === "pass") passed++;
414
+ else if (status === "skip") skipped++;
415
+ else failed++;
416
+
417
+ const entry = {
418
+ runId,
419
+ timestamp: new Date().toISOString(),
420
+ entry_type: "result",
421
+ phase: "finish",
422
+ status,
423
+ testId: buildTestId(event.data.file, event.data.name),
424
+ artifactDir: getArtifactDir(buildTestId(event.data.file, event.data.name)),
425
+ name: event.data.name,
426
+ file: event.data.file,
427
+ file_fingerprint: fingerprintFile(event.data.file),
428
+ duration_ms: Math.round(duration * 100) / 100,
429
+ details_type: event.data.details?.type || null,
430
+ nesting: event.data.nesting ?? null,
431
+ line: event.data.line ?? null,
432
+ column: event.data.column ?? null,
433
+ rerun_command: buildRerunCommand(event.data.file, event.data.name),
434
+ selector: buildSelector(event.data.file, event.data.name),
435
+ dependency_snapshot: (testArtifacts.get(buildTestId(event.data.file, event.data.name)) || {}).dependency_snapshot || buildDependencySnapshot(event.data.file),
436
+ ...(status === "skip" ? { skip_reason: truncate(inferredSkipReason || "skipped") } : {}),
437
+ ...(status === "fail" && event.data.details?.error
438
+ ? {
439
+ error: truncate(event.data.details.error.message || event.data.details.error),
440
+ error_name: event.data.details.error.name || null,
441
+ error_code: event.data.details.error.code || null,
442
+ error_stack: truncate(event.data.details.error.stack || ""),
443
+ timeout_detected: /timeout|cancelled|did not finish/i.test(
444
+ String(event.data.details.error.message || event.data.details.error)
445
+ ),
446
+ }
447
+ : {}),
448
+ };
449
+
450
+ results.push(entry);
451
+ upsertTestArtifact({
452
+ testId: entry.testId,
453
+ artifactDir: entry.artifactDir,
454
+ name: entry.name,
455
+ file: entry.file,
456
+ status: entry.status,
457
+ duration_ms: entry.duration_ms,
458
+ error: entry.error,
459
+ error_name: entry.error_name,
460
+ timeout_detected: entry.timeout_detected,
461
+ });
462
+ writeArtifactJson(entry.testId, "manifest.json", {
463
+ ...testArtifacts.get(entry.testId),
464
+ result: entry,
465
+ });
466
+ if (status === "fail") {
467
+ writeArtifactJson(entry.testId, "failure.json", entry);
468
+ writeArtifactJson(entry.testId, "triage.json", buildFailureBundle(entry));
469
+ } else if (status === "skip") {
470
+ writeArtifactJson(entry.testId, "triage.json", buildSkipBundle(entry));
471
+ }
472
+ logStream.write(JSON.stringify(entry) + "\n");
473
+ }
474
+
475
+ if (event.type === "test:skip") {
476
+ if (event.data.details?.type === "suite") continue;
477
+ skipped++;
478
+ const entry = {
479
+ runId,
480
+ timestamp: new Date().toISOString(),
481
+ entry_type: "result",
482
+ phase: "finish",
483
+ status: "skip",
484
+ testId: buildTestId(event.data.file, event.data.name),
485
+ artifactDir: getArtifactDir(buildTestId(event.data.file, event.data.name)),
486
+ name: event.data.name,
487
+ file: event.data.file,
488
+ skip_reason: truncate(extractSkipReason(event.data) || "skipped"),
489
+ rerun_command: buildRerunCommand(event.data.file, event.data.name),
490
+ selector: buildSelector(event.data.file, event.data.name),
491
+ dependency_snapshot: (testArtifacts.get(buildTestId(event.data.file, event.data.name)) || {}).dependency_snapshot || buildDependencySnapshot(event.data.file),
492
+ };
493
+ results.push(entry);
494
+ upsertTestArtifact({
495
+ testId: entry.testId,
496
+ artifactDir: entry.artifactDir,
497
+ name: entry.name,
498
+ file: entry.file,
499
+ status: "skip",
500
+ });
501
+ writeArtifactJson(entry.testId, "manifest.json", entry);
502
+ writeArtifactJson(entry.testId, "triage.json", buildSkipBundle(entry));
503
+ logStream.write(JSON.stringify(entry) + "\n");
504
+ }
505
+
506
+ // Pass through to default spec output
507
+ if (event.type === "test:pass" && extractSkipReason(event.data)) {
508
+ yield ` - ${event.data.name} (skipped)\n`;
509
+ } else if (event.type === "test:pass") {
510
+ yield ` ✓ ${event.data.name}\n`;
511
+ } else if (event.type === "test:fail") {
512
+ yield ` ✗ ${event.data.name}\n`;
513
+ if (event.data.details?.error) {
514
+ const msg = String(event.data.details.error.message || event.data.details.error);
515
+ yield ` ${msg.split("\n")[0]}\n`;
516
+ }
517
+ } else if (event.type === "test:skip") {
518
+ yield ` - ${event.data.name} (skipped)\n`;
519
+ } else if (event.type === "test:diagnostic") {
520
+ yield `${event.data.message}\n`;
521
+ } else if (event.type === "test:start") {
522
+ if (event.data.file) yield `\n${path.basename(event.data.file)}\n`;
523
+ }
524
+ }
525
+
526
+ // Write summary
527
+ const summary = {
528
+ runId,
529
+ timestamp: new Date().toISOString(),
530
+ entry_type: "run",
531
+ phase: "finish",
532
+ status: failed > 0 ? "failed" : "passed",
533
+ passed,
534
+ failed,
535
+ skipped,
536
+ total: passed + failed + skipped,
537
+ duration_ms: Math.round(totalDuration * 100) / 100,
538
+ workspace_state_at_run_start: workspaceStateAtRunStart,
539
+ tests: results.map((r) => ({
540
+ testId: r.testId,
541
+ name: r.name,
542
+ file: r.file,
543
+ status: r.status,
544
+ artifactDir: r.artifactDir,
545
+ rerun_command: r.rerun_command,
546
+ reason_code: r.reason_code || null,
547
+ reason_summary: r.reason_summary || null,
548
+ skip_reason: r.skip_reason || null,
549
+ dependency_snapshot: r.dependency_snapshot || null,
550
+ engine: r.engine || null,
551
+ workspace_state_at_run_start: workspaceStateAtRunStart,
552
+ })),
553
+ failedTests: results
554
+ .filter((r) => r.status === "fail")
555
+ .map((r) => buildFailureBundle(r)),
556
+ skippedTests: results
557
+ .filter((r) => r.status === "skip")
558
+ .map((r) => buildSkipBundle(r)),
559
+ };
560
+ fs.writeFileSync(SUMMARY_PATH, JSON.stringify(summary, null, 2) + "\n");
561
+ fs.writeFileSync(path.join(RUNS_DIR, runId, "summary.json"), JSON.stringify(summary, null, 2) + "\n");
562
+ fs.writeFileSync(path.join(RUNS_DIR, runId, "failures.json"), JSON.stringify(summary.failedTests, null, 2) + "\n");
563
+ fs.writeFileSync(path.join(RUNS_DIR, runId, "skips.json"), JSON.stringify(summary.skippedTests, null, 2) + "\n");
564
+ fs.writeFileSync(path.join(RUNS_DIR, runId, "rerun-plan.json"), JSON.stringify({
565
+ runId,
566
+ generated_at: new Date().toISOString(),
567
+ failed: summary.failedTests.map((test) => ({
568
+ testId: test.testId,
569
+ name: test.name,
570
+ reason_code: test.reason_code,
571
+ reason_summary: test.reason_summary,
572
+ rerun_command: test.rerun_command,
573
+ })),
574
+ skipped: summary.skippedTests.map((test) => ({
575
+ testId: test.testId,
576
+ name: test.name,
577
+ reason_code: test.reason_code,
578
+ reason_summary: test.reason_summary,
579
+ rerun_command: test.rerun_command,
580
+ })),
581
+ }, null, 2) + "\n");
582
+
583
+ const markdown = [
584
+ `# Test Run ${runId}`,
585
+ "",
586
+ `- Status: ${summary.status}`,
587
+ `- Passed: ${summary.passed}`,
588
+ `- Failed: ${summary.failed}`,
589
+ `- Skipped: ${summary.skipped}`,
590
+ `- Duration: ${(summary.duration_ms / 1000).toFixed(2)}s`,
591
+ `- Commit: ${runMeta.git_commit || "unknown"}`,
592
+ `- Branch: ${runMeta.git_branch || "unknown"}`,
593
+ `- Node: ${runMeta.node_version}`,
594
+ "",
595
+ "## Failed Tests",
596
+ ];
597
+ if (summary.failedTests.length === 0) {
598
+ markdown.push("", "None.");
599
+ } else {
600
+ for (const failure of summary.failedTests) {
601
+ markdown.push(
602
+ "",
603
+ `### ${failure.name}`,
604
+ `- Test ID: \`${failure.testId}\``,
605
+ `- File: \`${failure.file}\``,
606
+ `- Error: ${failure.error || "unknown"}`,
607
+ `- Error Name: ${failure.error_name || "n/a"}`,
608
+ `- Reason Code: ${failure.reason_code}`,
609
+ `- Reason Summary: ${failure.reason_summary}`,
610
+ `- Timeout: ${failure.timeout_detected ? "yes" : "no"}`,
611
+ `- Engine: ${failure.engine.engine || "n/a"}`,
612
+ `- Provider: ${failure.engine.provider || "n/a"}`,
613
+ `- Model: ${failure.engine.model || "n/a"}`,
614
+ `- Agent: ${failure.engine.agent || "n/a"}`,
615
+ `- Artifacts: \`${failure.artifactDir}\``,
616
+ `- Re-run: \`${failure.rerun_command}\``
617
+ );
618
+ }
619
+ }
620
+ markdown.push("", "## Skipped Tests");
621
+ if (summary.skippedTests.length === 0) {
622
+ markdown.push("", "None.");
623
+ } else {
624
+ for (const skippedTest of summary.skippedTests) {
625
+ markdown.push(
626
+ "",
627
+ `### ${skippedTest.name}`,
628
+ `- Test ID: \`${skippedTest.testId}\``,
629
+ `- File: \`${skippedTest.file}\``,
630
+ `- Reason Code: ${skippedTest.reason_code}`,
631
+ `- Reason: ${skippedTest.skip_reason}`,
632
+ `- Artifacts: \`${skippedTest.artifactDir}\``,
633
+ `- Re-run: \`${skippedTest.rerun_command}\``
634
+ );
635
+ }
636
+ }
637
+ markdown.push("", "## Artifact Directories", "");
638
+ for (const artifact of [...testArtifacts.values()].sort((a, b) => a.name.localeCompare(b.name))) {
639
+ markdown.push(`- \`${artifact.name}\` → \`${artifact.artifactDir}\``);
640
+ }
641
+ fs.writeFileSync(path.join(RUNS_DIR, runId, "summary.md"), markdown.join("\n") + "\n");
642
+
643
+ logStream.write(JSON.stringify(summary) + "\n");
644
+ logStream.end();
645
+
646
+ yield `\n─────────────────────────────────────\n`;
647
+ yield ` ${passed} passed, ${failed} failed, ${skipped} skipped (${results.length} total)\n`;
648
+ yield ` Duration: ${(totalDuration / 1000).toFixed(1)}s\n`;
649
+ yield ` Results: ${LOG_PATH}\n`;
650
+ yield `─────────────────────────────────────\n`;
651
+ }