@sentry/warden 0.30.0 → 0.31.0

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.
Files changed (185) hide show
  1. package/CHANGELOG.md +246 -0
  2. package/README.md +2 -1
  3. package/coverage/lcov-report/base.css +224 -0
  4. package/coverage/lcov-report/block-navigation.js +87 -0
  5. package/coverage/lcov-report/favicon.png +0 -0
  6. package/coverage/lcov-report/index.html +446 -0
  7. package/coverage/lcov-report/prettify.css +1 -0
  8. package/coverage/lcov-report/prettify.js +2 -0
  9. package/coverage/lcov-report/sort-arrow-sprite.png +0 -0
  10. package/coverage/lcov-report/sorter.js +210 -0
  11. package/coverage/lcov-report/src/action/checks/index.html +116 -0
  12. package/coverage/lcov-report/src/action/checks/manager.ts.html +499 -0
  13. package/coverage/lcov-report/src/action/error-reporting.ts.html +181 -0
  14. package/coverage/lcov-report/src/action/fix-evaluation/github.ts.html +508 -0
  15. package/coverage/lcov-report/src/action/fix-evaluation/index.html +161 -0
  16. package/coverage/lcov-report/src/action/fix-evaluation/index.ts.html +1243 -0
  17. package/coverage/lcov-report/src/action/fix-evaluation/judge.ts.html +817 -0
  18. package/coverage/lcov-report/src/action/fix-evaluation/types.ts.html +274 -0
  19. package/coverage/lcov-report/src/action/index.html +161 -0
  20. package/coverage/lcov-report/src/action/inputs.ts.html +529 -0
  21. package/coverage/lcov-report/src/action/pi-ncc-compat.ts.html +292 -0
  22. package/coverage/lcov-report/src/action/review/coordination.ts.html +208 -0
  23. package/coverage/lcov-report/src/action/review/index.html +131 -0
  24. package/coverage/lcov-report/src/action/review/poster.ts.html +1051 -0
  25. package/coverage/lcov-report/src/action/review-state.ts.html +280 -0
  26. package/coverage/lcov-report/src/action/triggers/executor.ts.html +856 -0
  27. package/coverage/lcov-report/src/action/triggers/index.html +116 -0
  28. package/coverage/lcov-report/src/action/workflow/base.ts.html +1135 -0
  29. package/coverage/lcov-report/src/action/workflow/index.html +146 -0
  30. package/coverage/lcov-report/src/action/workflow/pr-workflow.ts.html +2713 -0
  31. package/coverage/lcov-report/src/action/workflow/schedule.ts.html +1027 -0
  32. package/coverage/lcov-report/src/cli/args.ts.html +1657 -0
  33. package/coverage/lcov-report/src/cli/commands/add.ts.html +1567 -0
  34. package/coverage/lcov-report/src/cli/commands/build.ts.html +1618 -0
  35. package/coverage/lcov-report/src/cli/commands/index.html +191 -0
  36. package/coverage/lcov-report/src/cli/commands/init.ts.html +1357 -0
  37. package/coverage/lcov-report/src/cli/commands/runs.ts.html +2560 -0
  38. package/coverage/lcov-report/src/cli/commands/setup-app/browser.ts.html +214 -0
  39. package/coverage/lcov-report/src/cli/commands/setup-app/credentials.ts.html +220 -0
  40. package/coverage/lcov-report/src/cli/commands/setup-app/index.html +161 -0
  41. package/coverage/lcov-report/src/cli/commands/setup-app/manifest.ts.html +235 -0
  42. package/coverage/lcov-report/src/cli/commands/setup-app/server.ts.html +601 -0
  43. package/coverage/lcov-report/src/cli/commands/setup-app.ts.html +571 -0
  44. package/coverage/lcov-report/src/cli/commands/sync.ts.html +475 -0
  45. package/coverage/lcov-report/src/cli/context.ts.html +532 -0
  46. package/coverage/lcov-report/src/cli/diff-apply.ts.html +334 -0
  47. package/coverage/lcov-report/src/cli/files.ts.html +904 -0
  48. package/coverage/lcov-report/src/cli/fix.ts.html +1060 -0
  49. package/coverage/lcov-report/src/cli/git.ts.html +1201 -0
  50. package/coverage/lcov-report/src/cli/help.ts.html +1696 -0
  51. package/coverage/lcov-report/src/cli/index.html +281 -0
  52. package/coverage/lcov-report/src/cli/input.ts.html +397 -0
  53. package/coverage/lcov-report/src/cli/log-cleanup.ts.html +337 -0
  54. package/coverage/lcov-report/src/cli/main.ts.html +5107 -0
  55. package/coverage/lcov-report/src/cli/output/box.ts.html +790 -0
  56. package/coverage/lcov-report/src/cli/output/formatters.ts.html +1099 -0
  57. package/coverage/lcov-report/src/cli/output/icons.ts.html +142 -0
  58. package/coverage/lcov-report/src/cli/output/index.html +281 -0
  59. package/coverage/lcov-report/src/cli/output/index.ts.html +295 -0
  60. package/coverage/lcov-report/src/cli/output/ink-runner.tsx.html +1597 -0
  61. package/coverage/lcov-report/src/cli/output/jsonl-schema-gen.ts.html +355 -0
  62. package/coverage/lcov-report/src/cli/output/jsonl.ts.html +2137 -0
  63. package/coverage/lcov-report/src/cli/output/live-status.tsx.html +916 -0
  64. package/coverage/lcov-report/src/cli/output/reporter.ts.html +1630 -0
  65. package/coverage/lcov-report/src/cli/output/tasks.ts.html +3193 -0
  66. package/coverage/lcov-report/src/cli/output/tty.ts.html +310 -0
  67. package/coverage/lcov-report/src/cli/output/verbosity.ts.html +181 -0
  68. package/coverage/lcov-report/src/cli/signals.ts.html +199 -0
  69. package/coverage/lcov-report/src/cli/terminal.ts.html +1060 -0
  70. package/coverage/lcov-report/src/config/index.html +161 -0
  71. package/coverage/lcov-report/src/config/index.ts.html +94 -0
  72. package/coverage/lcov-report/src/config/loader.ts.html +1753 -0
  73. package/coverage/lcov-report/src/config/schema.ts.html +910 -0
  74. package/coverage/lcov-report/src/config/writer.ts.html +538 -0
  75. package/coverage/lcov-report/src/diff/apply.ts.html +241 -0
  76. package/coverage/lcov-report/src/diff/classify.ts.html +361 -0
  77. package/coverage/lcov-report/src/diff/coalesce.ts.html +1123 -0
  78. package/coverage/lcov-report/src/diff/context.ts.html +952 -0
  79. package/coverage/lcov-report/src/diff/index.html +191 -0
  80. package/coverage/lcov-report/src/diff/index.ts.html +100 -0
  81. package/coverage/lcov-report/src/diff/parser.ts.html +532 -0
  82. package/coverage/lcov-report/src/evals/harness.ts.html +700 -0
  83. package/coverage/lcov-report/src/evals/index.html +221 -0
  84. package/coverage/lcov-report/src/evals/index.ts.html +955 -0
  85. package/coverage/lcov-report/src/evals/names.ts.html +121 -0
  86. package/coverage/lcov-report/src/evals/runner.ts.html +664 -0
  87. package/coverage/lcov-report/src/evals/scaffold.ts.html +970 -0
  88. package/coverage/lcov-report/src/evals/types.ts.html +661 -0
  89. package/coverage/lcov-report/src/evals/usage.ts.html +178 -0
  90. package/coverage/lcov-report/src/evals/verify.ts.html +1054 -0
  91. package/coverage/lcov-report/src/event/context.ts.html +490 -0
  92. package/coverage/lcov-report/src/event/index.html +116 -0
  93. package/coverage/lcov-report/src/index.html +116 -0
  94. package/coverage/lcov-report/src/output/dedup.ts.html +2950 -0
  95. package/coverage/lcov-report/src/output/github-checks.ts.html +1579 -0
  96. package/coverage/lcov-report/src/output/index.html +161 -0
  97. package/coverage/lcov-report/src/output/renderer.ts.html +988 -0
  98. package/coverage/lcov-report/src/output/stale.ts.html +640 -0
  99. package/coverage/lcov-report/src/sdk/analyze.ts.html +3010 -0
  100. package/coverage/lcov-report/src/sdk/auth.ts.html +217 -0
  101. package/coverage/lcov-report/src/sdk/circuit-breaker.ts.html +304 -0
  102. package/coverage/lcov-report/src/sdk/errors.ts.html +658 -0
  103. package/coverage/lcov-report/src/sdk/extract.ts.html +1810 -0
  104. package/coverage/lcov-report/src/sdk/fix-quality.ts.html +841 -0
  105. package/coverage/lcov-report/src/sdk/haiku.ts.html +1456 -0
  106. package/coverage/lcov-report/src/sdk/index.html +416 -0
  107. package/coverage/lcov-report/src/sdk/json-output.ts.html +508 -0
  108. package/coverage/lcov-report/src/sdk/model-pricing.json.html +499 -0
  109. package/coverage/lcov-report/src/sdk/otel.ts.html +433 -0
  110. package/coverage/lcov-report/src/sdk/post-process.ts.html +412 -0
  111. package/coverage/lcov-report/src/sdk/prepare.ts.html +400 -0
  112. package/coverage/lcov-report/src/sdk/pricing.ts.html +778 -0
  113. package/coverage/lcov-report/src/sdk/prompt-sections.ts.html +595 -0
  114. package/coverage/lcov-report/src/sdk/prompt.ts.html +451 -0
  115. package/coverage/lcov-report/src/sdk/report-files.ts.html +166 -0
  116. package/coverage/lcov-report/src/sdk/retry.ts.html +202 -0
  117. package/coverage/lcov-report/src/sdk/runner.ts.html +400 -0
  118. package/coverage/lcov-report/src/sdk/runtimes/claude.ts.html +1693 -0
  119. package/coverage/lcov-report/src/sdk/runtimes/index.html +176 -0
  120. package/coverage/lcov-report/src/sdk/runtimes/index.ts.html +244 -0
  121. package/coverage/lcov-report/src/sdk/runtimes/model-selectors.ts.html +328 -0
  122. package/coverage/lcov-report/src/sdk/runtimes/pi.ts.html +2104 -0
  123. package/coverage/lcov-report/src/sdk/runtimes/types.ts.html +508 -0
  124. package/coverage/lcov-report/src/sdk/types.ts.html +730 -0
  125. package/coverage/lcov-report/src/sdk/usage.ts.html +502 -0
  126. package/coverage/lcov-report/src/sdk/verify.ts.html +1027 -0
  127. package/coverage/lcov-report/src/sentry.ts.html +994 -0
  128. package/coverage/lcov-report/src/skill-builder/agentic.ts.html +1063 -0
  129. package/coverage/lcov-report/src/skill-builder/authoring-provider.ts.html +334 -0
  130. package/coverage/lcov-report/src/skill-builder/definition.ts.html +748 -0
  131. package/coverage/lcov-report/src/skill-builder/index.html +236 -0
  132. package/coverage/lcov-report/src/skill-builder/outline-contract.ts.html +400 -0
  133. package/coverage/lcov-report/src/skill-builder/outline-state.ts.html +382 -0
  134. package/coverage/lcov-report/src/skill-builder/outline.ts.html +1402 -0
  135. package/coverage/lcov-report/src/skill-builder/skill-contract.ts.html +382 -0
  136. package/coverage/lcov-report/src/skill-builder/skill-prompts.ts.html +1180 -0
  137. package/coverage/lcov-report/src/skill-builder/skill.ts.html +2398 -0
  138. package/coverage/lcov-report/src/skills/index.html +131 -0
  139. package/coverage/lcov-report/src/skills/loader.ts.html +1693 -0
  140. package/coverage/lcov-report/src/skills/remote.ts.html +1909 -0
  141. package/coverage/lcov-report/src/triggers/index.html +116 -0
  142. package/coverage/lcov-report/src/triggers/matcher.ts.html +841 -0
  143. package/coverage/lcov-report/src/types/index.html +116 -0
  144. package/coverage/lcov-report/src/types/index.ts.html +1240 -0
  145. package/coverage/lcov-report/src/utils/async.ts.html +331 -0
  146. package/coverage/lcov-report/src/utils/exec.ts.html +496 -0
  147. package/coverage/lcov-report/src/utils/index.html +176 -0
  148. package/coverage/lcov-report/src/utils/index.ts.html +313 -0
  149. package/coverage/lcov-report/src/utils/path.ts.html +202 -0
  150. package/coverage/lcov-report/src/utils/version.ts.html +136 -0
  151. package/coverage/lcov.info +19134 -0
  152. package/dist/cli/commands/runs.d.ts +1 -1
  153. package/dist/cli/commands/runs.d.ts.map +1 -1
  154. package/dist/cli/commands/runs.js +1 -1
  155. package/dist/cli/commands/runs.js.map +1 -1
  156. package/dist/cli/files.d.ts.map +1 -1
  157. package/dist/cli/files.js +10 -11
  158. package/dist/cli/files.js.map +1 -1
  159. package/dist/sdk/verify.d.ts.map +1 -1
  160. package/dist/sdk/verify.js +3 -2
  161. package/dist/sdk/verify.js.map +1 -1
  162. package/dist/skill-builder/outline.d.ts.map +1 -1
  163. package/dist/skill-builder/outline.js.map +1 -1
  164. package/package.json +8 -3
  165. package/test-results.junit.xml +4329 -0
  166. package/dist/evals/index.d.ts +0 -22
  167. package/dist/evals/index.d.ts.map +0 -1
  168. package/dist/evals/index.js +0 -92
  169. package/dist/evals/index.js.map +0 -1
  170. package/dist/evals/judge.d.ts +0 -12
  171. package/dist/evals/judge.d.ts.map +0 -1
  172. package/dist/evals/judge.js +0 -171
  173. package/dist/evals/judge.js.map +0 -1
  174. package/dist/evals/runner.d.ts +0 -18
  175. package/dist/evals/runner.d.ts.map +0 -1
  176. package/dist/evals/runner.js +0 -149
  177. package/dist/evals/runner.js.map +0 -1
  178. package/dist/evals/setup.d.ts +0 -2
  179. package/dist/evals/setup.d.ts.map +0 -1
  180. package/dist/evals/setup.js +0 -22
  181. package/dist/evals/setup.js.map +0 -1
  182. package/dist/evals/types.d.ts +0 -160
  183. package/dist/evals/types.d.ts.map +0 -1
  184. package/dist/evals/types.js +0 -148
  185. package/dist/evals/types.js.map +0 -1
@@ -0,0 +1,700 @@
1
+
2
+ <!doctype html>
3
+ <html lang="en">
4
+
5
+ <head>
6
+ <title>Code coverage report for src/evals/harness.ts</title>
7
+ <meta charset="utf-8" />
8
+ <link rel="stylesheet" href="../../prettify.css" />
9
+ <link rel="stylesheet" href="../../base.css" />
10
+ <link rel="shortcut icon" type="image/x-icon" href="../../favicon.png" />
11
+ <meta name="viewport" content="width=device-width, initial-scale=1" />
12
+ <style type='text/css'>
13
+ .coverage-summary .sorter {
14
+ background-image: url(../../sort-arrow-sprite.png);
15
+ }
16
+ </style>
17
+ </head>
18
+
19
+ <body>
20
+ <div class='wrapper'>
21
+ <div class='pad1'>
22
+ <h1><a href="../../index.html">All files</a> / <a href="index.html">src/evals</a> harness.ts</h1>
23
+ <div class='clearfix'>
24
+
25
+ <div class='fl pad1y space-right2'>
26
+ <span class="strong">62.5% </span>
27
+ <span class="quiet">Statements</span>
28
+ <span class='fraction'>30/48</span>
29
+ </div>
30
+
31
+
32
+ <div class='fl pad1y space-right2'>
33
+ <span class="strong">38.63% </span>
34
+ <span class="quiet">Branches</span>
35
+ <span class='fraction'>17/44</span>
36
+ </div>
37
+
38
+
39
+ <div class='fl pad1y space-right2'>
40
+ <span class="strong">50% </span>
41
+ <span class="quiet">Functions</span>
42
+ <span class='fraction'>4/8</span>
43
+ </div>
44
+
45
+
46
+ <div class='fl pad1y space-right2'>
47
+ <span class="strong">60.86% </span>
48
+ <span class="quiet">Lines</span>
49
+ <span class='fraction'>28/46</span>
50
+ </div>
51
+
52
+
53
+ </div>
54
+ <p class="quiet">
55
+ Press <em>n</em> or <em>j</em> to go to the next uncovered block, <em>b</em>, <em>p</em> or <em>k</em> for the previous block.
56
+ </p>
57
+ <template id="filterTemplate">
58
+ <div class="quiet">
59
+ Filter:
60
+ <input type="search" id="fileSearch">
61
+ </div>
62
+ </template>
63
+ </div>
64
+ <div class='status-line medium'></div>
65
+ <pre><table class="coverage">
66
+ <tr><td class="line-count quiet"><a name='L1'></a><a href='#L1'>1</a>
67
+ <a name='L2'></a><a href='#L2'>2</a>
68
+ <a name='L3'></a><a href='#L3'>3</a>
69
+ <a name='L4'></a><a href='#L4'>4</a>
70
+ <a name='L5'></a><a href='#L5'>5</a>
71
+ <a name='L6'></a><a href='#L6'>6</a>
72
+ <a name='L7'></a><a href='#L7'>7</a>
73
+ <a name='L8'></a><a href='#L8'>8</a>
74
+ <a name='L9'></a><a href='#L9'>9</a>
75
+ <a name='L10'></a><a href='#L10'>10</a>
76
+ <a name='L11'></a><a href='#L11'>11</a>
77
+ <a name='L12'></a><a href='#L12'>12</a>
78
+ <a name='L13'></a><a href='#L13'>13</a>
79
+ <a name='L14'></a><a href='#L14'>14</a>
80
+ <a name='L15'></a><a href='#L15'>15</a>
81
+ <a name='L16'></a><a href='#L16'>16</a>
82
+ <a name='L17'></a><a href='#L17'>17</a>
83
+ <a name='L18'></a><a href='#L18'>18</a>
84
+ <a name='L19'></a><a href='#L19'>19</a>
85
+ <a name='L20'></a><a href='#L20'>20</a>
86
+ <a name='L21'></a><a href='#L21'>21</a>
87
+ <a name='L22'></a><a href='#L22'>22</a>
88
+ <a name='L23'></a><a href='#L23'>23</a>
89
+ <a name='L24'></a><a href='#L24'>24</a>
90
+ <a name='L25'></a><a href='#L25'>25</a>
91
+ <a name='L26'></a><a href='#L26'>26</a>
92
+ <a name='L27'></a><a href='#L27'>27</a>
93
+ <a name='L28'></a><a href='#L28'>28</a>
94
+ <a name='L29'></a><a href='#L29'>29</a>
95
+ <a name='L30'></a><a href='#L30'>30</a>
96
+ <a name='L31'></a><a href='#L31'>31</a>
97
+ <a name='L32'></a><a href='#L32'>32</a>
98
+ <a name='L33'></a><a href='#L33'>33</a>
99
+ <a name='L34'></a><a href='#L34'>34</a>
100
+ <a name='L35'></a><a href='#L35'>35</a>
101
+ <a name='L36'></a><a href='#L36'>36</a>
102
+ <a name='L37'></a><a href='#L37'>37</a>
103
+ <a name='L38'></a><a href='#L38'>38</a>
104
+ <a name='L39'></a><a href='#L39'>39</a>
105
+ <a name='L40'></a><a href='#L40'>40</a>
106
+ <a name='L41'></a><a href='#L41'>41</a>
107
+ <a name='L42'></a><a href='#L42'>42</a>
108
+ <a name='L43'></a><a href='#L43'>43</a>
109
+ <a name='L44'></a><a href='#L44'>44</a>
110
+ <a name='L45'></a><a href='#L45'>45</a>
111
+ <a name='L46'></a><a href='#L46'>46</a>
112
+ <a name='L47'></a><a href='#L47'>47</a>
113
+ <a name='L48'></a><a href='#L48'>48</a>
114
+ <a name='L49'></a><a href='#L49'>49</a>
115
+ <a name='L50'></a><a href='#L50'>50</a>
116
+ <a name='L51'></a><a href='#L51'>51</a>
117
+ <a name='L52'></a><a href='#L52'>52</a>
118
+ <a name='L53'></a><a href='#L53'>53</a>
119
+ <a name='L54'></a><a href='#L54'>54</a>
120
+ <a name='L55'></a><a href='#L55'>55</a>
121
+ <a name='L56'></a><a href='#L56'>56</a>
122
+ <a name='L57'></a><a href='#L57'>57</a>
123
+ <a name='L58'></a><a href='#L58'>58</a>
124
+ <a name='L59'></a><a href='#L59'>59</a>
125
+ <a name='L60'></a><a href='#L60'>60</a>
126
+ <a name='L61'></a><a href='#L61'>61</a>
127
+ <a name='L62'></a><a href='#L62'>62</a>
128
+ <a name='L63'></a><a href='#L63'>63</a>
129
+ <a name='L64'></a><a href='#L64'>64</a>
130
+ <a name='L65'></a><a href='#L65'>65</a>
131
+ <a name='L66'></a><a href='#L66'>66</a>
132
+ <a name='L67'></a><a href='#L67'>67</a>
133
+ <a name='L68'></a><a href='#L68'>68</a>
134
+ <a name='L69'></a><a href='#L69'>69</a>
135
+ <a name='L70'></a><a href='#L70'>70</a>
136
+ <a name='L71'></a><a href='#L71'>71</a>
137
+ <a name='L72'></a><a href='#L72'>72</a>
138
+ <a name='L73'></a><a href='#L73'>73</a>
139
+ <a name='L74'></a><a href='#L74'>74</a>
140
+ <a name='L75'></a><a href='#L75'>75</a>
141
+ <a name='L76'></a><a href='#L76'>76</a>
142
+ <a name='L77'></a><a href='#L77'>77</a>
143
+ <a name='L78'></a><a href='#L78'>78</a>
144
+ <a name='L79'></a><a href='#L79'>79</a>
145
+ <a name='L80'></a><a href='#L80'>80</a>
146
+ <a name='L81'></a><a href='#L81'>81</a>
147
+ <a name='L82'></a><a href='#L82'>82</a>
148
+ <a name='L83'></a><a href='#L83'>83</a>
149
+ <a name='L84'></a><a href='#L84'>84</a>
150
+ <a name='L85'></a><a href='#L85'>85</a>
151
+ <a name='L86'></a><a href='#L86'>86</a>
152
+ <a name='L87'></a><a href='#L87'>87</a>
153
+ <a name='L88'></a><a href='#L88'>88</a>
154
+ <a name='L89'></a><a href='#L89'>89</a>
155
+ <a name='L90'></a><a href='#L90'>90</a>
156
+ <a name='L91'></a><a href='#L91'>91</a>
157
+ <a name='L92'></a><a href='#L92'>92</a>
158
+ <a name='L93'></a><a href='#L93'>93</a>
159
+ <a name='L94'></a><a href='#L94'>94</a>
160
+ <a name='L95'></a><a href='#L95'>95</a>
161
+ <a name='L96'></a><a href='#L96'>96</a>
162
+ <a name='L97'></a><a href='#L97'>97</a>
163
+ <a name='L98'></a><a href='#L98'>98</a>
164
+ <a name='L99'></a><a href='#L99'>99</a>
165
+ <a name='L100'></a><a href='#L100'>100</a>
166
+ <a name='L101'></a><a href='#L101'>101</a>
167
+ <a name='L102'></a><a href='#L102'>102</a>
168
+ <a name='L103'></a><a href='#L103'>103</a>
169
+ <a name='L104'></a><a href='#L104'>104</a>
170
+ <a name='L105'></a><a href='#L105'>105</a>
171
+ <a name='L106'></a><a href='#L106'>106</a>
172
+ <a name='L107'></a><a href='#L107'>107</a>
173
+ <a name='L108'></a><a href='#L108'>108</a>
174
+ <a name='L109'></a><a href='#L109'>109</a>
175
+ <a name='L110'></a><a href='#L110'>110</a>
176
+ <a name='L111'></a><a href='#L111'>111</a>
177
+ <a name='L112'></a><a href='#L112'>112</a>
178
+ <a name='L113'></a><a href='#L113'>113</a>
179
+ <a name='L114'></a><a href='#L114'>114</a>
180
+ <a name='L115'></a><a href='#L115'>115</a>
181
+ <a name='L116'></a><a href='#L116'>116</a>
182
+ <a name='L117'></a><a href='#L117'>117</a>
183
+ <a name='L118'></a><a href='#L118'>118</a>
184
+ <a name='L119'></a><a href='#L119'>119</a>
185
+ <a name='L120'></a><a href='#L120'>120</a>
186
+ <a name='L121'></a><a href='#L121'>121</a>
187
+ <a name='L122'></a><a href='#L122'>122</a>
188
+ <a name='L123'></a><a href='#L123'>123</a>
189
+ <a name='L124'></a><a href='#L124'>124</a>
190
+ <a name='L125'></a><a href='#L125'>125</a>
191
+ <a name='L126'></a><a href='#L126'>126</a>
192
+ <a name='L127'></a><a href='#L127'>127</a>
193
+ <a name='L128'></a><a href='#L128'>128</a>
194
+ <a name='L129'></a><a href='#L129'>129</a>
195
+ <a name='L130'></a><a href='#L130'>130</a>
196
+ <a name='L131'></a><a href='#L131'>131</a>
197
+ <a name='L132'></a><a href='#L132'>132</a>
198
+ <a name='L133'></a><a href='#L133'>133</a>
199
+ <a name='L134'></a><a href='#L134'>134</a>
200
+ <a name='L135'></a><a href='#L135'>135</a>
201
+ <a name='L136'></a><a href='#L136'>136</a>
202
+ <a name='L137'></a><a href='#L137'>137</a>
203
+ <a name='L138'></a><a href='#L138'>138</a>
204
+ <a name='L139'></a><a href='#L139'>139</a>
205
+ <a name='L140'></a><a href='#L140'>140</a>
206
+ <a name='L141'></a><a href='#L141'>141</a>
207
+ <a name='L142'></a><a href='#L142'>142</a>
208
+ <a name='L143'></a><a href='#L143'>143</a>
209
+ <a name='L144'></a><a href='#L144'>144</a>
210
+ <a name='L145'></a><a href='#L145'>145</a>
211
+ <a name='L146'></a><a href='#L146'>146</a>
212
+ <a name='L147'></a><a href='#L147'>147</a>
213
+ <a name='L148'></a><a href='#L148'>148</a>
214
+ <a name='L149'></a><a href='#L149'>149</a>
215
+ <a name='L150'></a><a href='#L150'>150</a>
216
+ <a name='L151'></a><a href='#L151'>151</a>
217
+ <a name='L152'></a><a href='#L152'>152</a>
218
+ <a name='L153'></a><a href='#L153'>153</a>
219
+ <a name='L154'></a><a href='#L154'>154</a>
220
+ <a name='L155'></a><a href='#L155'>155</a>
221
+ <a name='L156'></a><a href='#L156'>156</a>
222
+ <a name='L157'></a><a href='#L157'>157</a>
223
+ <a name='L158'></a><a href='#L158'>158</a>
224
+ <a name='L159'></a><a href='#L159'>159</a>
225
+ <a name='L160'></a><a href='#L160'>160</a>
226
+ <a name='L161'></a><a href='#L161'>161</a>
227
+ <a name='L162'></a><a href='#L162'>162</a>
228
+ <a name='L163'></a><a href='#L163'>163</a>
229
+ <a name='L164'></a><a href='#L164'>164</a>
230
+ <a name='L165'></a><a href='#L165'>165</a>
231
+ <a name='L166'></a><a href='#L166'>166</a>
232
+ <a name='L167'></a><a href='#L167'>167</a>
233
+ <a name='L168'></a><a href='#L168'>168</a>
234
+ <a name='L169'></a><a href='#L169'>169</a>
235
+ <a name='L170'></a><a href='#L170'>170</a>
236
+ <a name='L171'></a><a href='#L171'>171</a>
237
+ <a name='L172'></a><a href='#L172'>172</a>
238
+ <a name='L173'></a><a href='#L173'>173</a>
239
+ <a name='L174'></a><a href='#L174'>174</a>
240
+ <a name='L175'></a><a href='#L175'>175</a>
241
+ <a name='L176'></a><a href='#L176'>176</a>
242
+ <a name='L177'></a><a href='#L177'>177</a>
243
+ <a name='L178'></a><a href='#L178'>178</a>
244
+ <a name='L179'></a><a href='#L179'>179</a>
245
+ <a name='L180'></a><a href='#L180'>180</a>
246
+ <a name='L181'></a><a href='#L181'>181</a>
247
+ <a name='L182'></a><a href='#L182'>182</a>
248
+ <a name='L183'></a><a href='#L183'>183</a>
249
+ <a name='L184'></a><a href='#L184'>184</a>
250
+ <a name='L185'></a><a href='#L185'>185</a>
251
+ <a name='L186'></a><a href='#L186'>186</a>
252
+ <a name='L187'></a><a href='#L187'>187</a>
253
+ <a name='L188'></a><a href='#L188'>188</a>
254
+ <a name='L189'></a><a href='#L189'>189</a>
255
+ <a name='L190'></a><a href='#L190'>190</a>
256
+ <a name='L191'></a><a href='#L191'>191</a>
257
+ <a name='L192'></a><a href='#L192'>192</a>
258
+ <a name='L193'></a><a href='#L193'>193</a>
259
+ <a name='L194'></a><a href='#L194'>194</a>
260
+ <a name='L195'></a><a href='#L195'>195</a>
261
+ <a name='L196'></a><a href='#L196'>196</a>
262
+ <a name='L197'></a><a href='#L197'>197</a>
263
+ <a name='L198'></a><a href='#L198'>198</a>
264
+ <a name='L199'></a><a href='#L199'>199</a>
265
+ <a name='L200'></a><a href='#L200'>200</a>
266
+ <a name='L201'></a><a href='#L201'>201</a>
267
+ <a name='L202'></a><a href='#L202'>202</a>
268
+ <a name='L203'></a><a href='#L203'>203</a>
269
+ <a name='L204'></a><a href='#L204'>204</a>
270
+ <a name='L205'></a><a href='#L205'>205</a>
271
+ <a name='L206'></a><a href='#L206'>206</a></td><td class="line-coverage quiet"><span class="cline-any cline-neutral">&nbsp;</span>
272
+ <span class="cline-any cline-neutral">&nbsp;</span>
273
+ <span class="cline-any cline-neutral">&nbsp;</span>
274
+ <span class="cline-any cline-neutral">&nbsp;</span>
275
+ <span class="cline-any cline-neutral">&nbsp;</span>
276
+ <span class="cline-any cline-neutral">&nbsp;</span>
277
+ <span class="cline-any cline-neutral">&nbsp;</span>
278
+ <span class="cline-any cline-neutral">&nbsp;</span>
279
+ <span class="cline-any cline-neutral">&nbsp;</span>
280
+ <span class="cline-any cline-neutral">&nbsp;</span>
281
+ <span class="cline-any cline-neutral">&nbsp;</span>
282
+ <span class="cline-any cline-neutral">&nbsp;</span>
283
+ <span class="cline-any cline-neutral">&nbsp;</span>
284
+ <span class="cline-any cline-neutral">&nbsp;</span>
285
+ <span class="cline-any cline-neutral">&nbsp;</span>
286
+ <span class="cline-any cline-neutral">&nbsp;</span>
287
+ <span class="cline-any cline-yes">1x</span>
288
+ <span class="cline-any cline-neutral">&nbsp;</span>
289
+ <span class="cline-any cline-neutral">&nbsp;</span>
290
+ <span class="cline-any cline-neutral">&nbsp;</span>
291
+ <span class="cline-any cline-neutral">&nbsp;</span>
292
+ <span class="cline-any cline-neutral">&nbsp;</span>
293
+ <span class="cline-any cline-neutral">&nbsp;</span>
294
+ <span class="cline-any cline-neutral">&nbsp;</span>
295
+ <span class="cline-any cline-neutral">&nbsp;</span>
296
+ <span class="cline-any cline-neutral">&nbsp;</span>
297
+ <span class="cline-any cline-neutral">&nbsp;</span>
298
+ <span class="cline-any cline-neutral">&nbsp;</span>
299
+ <span class="cline-any cline-neutral">&nbsp;</span>
300
+ <span class="cline-any cline-yes">3x</span>
301
+ <span class="cline-any cline-no">&nbsp;</span>
302
+ <span class="cline-any cline-neutral">&nbsp;</span>
303
+ <span class="cline-any cline-neutral">&nbsp;</span>
304
+ <span class="cline-any cline-yes">3x</span>
305
+ <span class="cline-any cline-neutral">&nbsp;</span>
306
+ <span class="cline-any cline-neutral">&nbsp;</span>
307
+ <span class="cline-any cline-neutral">&nbsp;</span>
308
+ <span class="cline-any cline-neutral">&nbsp;</span>
309
+ <span class="cline-any cline-neutral">&nbsp;</span>
310
+ <span class="cline-any cline-neutral">&nbsp;</span>
311
+ <span class="cline-any cline-neutral">&nbsp;</span>
312
+ <span class="cline-any cline-no">&nbsp;</span>
313
+ <span class="cline-any cline-neutral">&nbsp;</span>
314
+ <span class="cline-any cline-neutral">&nbsp;</span>
315
+ <span class="cline-any cline-neutral">&nbsp;</span>
316
+ <span class="cline-any cline-neutral">&nbsp;</span>
317
+ <span class="cline-any cline-neutral">&nbsp;</span>
318
+ <span class="cline-any cline-neutral">&nbsp;</span>
319
+ <span class="cline-any cline-neutral">&nbsp;</span>
320
+ <span class="cline-any cline-neutral">&nbsp;</span>
321
+ <span class="cline-any cline-neutral">&nbsp;</span>
322
+ <span class="cline-any cline-neutral">&nbsp;</span>
323
+ <span class="cline-any cline-neutral">&nbsp;</span>
324
+ <span class="cline-any cline-neutral">&nbsp;</span>
325
+ <span class="cline-any cline-neutral">&nbsp;</span>
326
+ <span class="cline-any cline-neutral">&nbsp;</span>
327
+ <span class="cline-any cline-neutral">&nbsp;</span>
328
+ <span class="cline-any cline-neutral">&nbsp;</span>
329
+ <span class="cline-any cline-yes">2x</span>
330
+ <span class="cline-any cline-neutral">&nbsp;</span>
331
+ <span class="cline-any cline-yes">2x</span>
332
+ <span class="cline-any cline-yes">3x</span>
333
+ <span class="cline-any cline-yes">3x</span>
334
+ <span class="cline-any cline-yes">1x</span>
335
+ <span class="cline-any cline-neutral">&nbsp;</span>
336
+ <span class="cline-any cline-neutral">&nbsp;</span>
337
+ <span class="cline-any cline-yes">2x</span>
338
+ <span class="cline-any cline-yes">2x</span>
339
+ <span class="cline-any cline-no">&nbsp;</span>
340
+ <span class="cline-any cline-no">&nbsp;</span>
341
+ <span class="cline-any cline-neutral">&nbsp;</span>
342
+ <span class="cline-any cline-neutral">&nbsp;</span>
343
+ <span class="cline-any cline-yes">2x</span>
344
+ <span class="cline-any cline-neutral">&nbsp;</span>
345
+ <span class="cline-any cline-neutral">&nbsp;</span>
346
+ <span class="cline-any cline-yes">3x</span>
347
+ <span class="cline-any cline-no">&nbsp;</span>
348
+ <span class="cline-any cline-yes">2x</span>
349
+ <span class="cline-any cline-yes">1x</span>
350
+ <span class="cline-any cline-yes">1x</span>
351
+ <span class="cline-any cline-no">&nbsp;</span>
352
+ <span class="cline-any cline-no">&nbsp;</span>
353
+ <span class="cline-any cline-neutral">&nbsp;</span>
354
+ <span class="cline-any cline-neutral">&nbsp;</span>
355
+ <span class="cline-any cline-neutral">&nbsp;</span>
356
+ <span class="cline-any cline-neutral">&nbsp;</span>
357
+ <span class="cline-any cline-neutral">&nbsp;</span>
358
+ <span class="cline-any cline-neutral">&nbsp;</span>
359
+ <span class="cline-any cline-yes">2x</span>
360
+ <span class="cline-any cline-no">&nbsp;</span>
361
+ <span class="cline-any cline-no">&nbsp;</span>
362
+ <span class="cline-any cline-no">&nbsp;</span>
363
+ <span class="cline-any cline-neutral">&nbsp;</span>
364
+ <span class="cline-any cline-neutral">&nbsp;</span>
365
+ <span class="cline-any cline-neutral">&nbsp;</span>
366
+ <span class="cline-any cline-yes">2x</span>
367
+ <span class="cline-any cline-neutral">&nbsp;</span>
368
+ <span class="cline-any cline-neutral">&nbsp;</span>
369
+ <span class="cline-any cline-neutral">&nbsp;</span>
370
+ <span class="cline-any cline-no">&nbsp;</span>
371
+ <span class="cline-any cline-neutral">&nbsp;</span>
372
+ <span class="cline-any cline-neutral">&nbsp;</span>
373
+ <span class="cline-any cline-no">&nbsp;</span>
374
+ <span class="cline-any cline-neutral">&nbsp;</span>
375
+ <span class="cline-any cline-neutral">&nbsp;</span>
376
+ <span class="cline-any cline-no">&nbsp;</span>
377
+ <span class="cline-any cline-neutral">&nbsp;</span>
378
+ <span class="cline-any cline-neutral">&nbsp;</span>
379
+ <span class="cline-any cline-no">&nbsp;</span>
380
+ <span class="cline-any cline-neutral">&nbsp;</span>
381
+ <span class="cline-any cline-neutral">&nbsp;</span>
382
+ <span class="cline-any cline-neutral">&nbsp;</span>
383
+ <span class="cline-any cline-no">&nbsp;</span>
384
+ <span class="cline-any cline-neutral">&nbsp;</span>
385
+ <span class="cline-any cline-neutral">&nbsp;</span>
386
+ <span class="cline-any cline-neutral">&nbsp;</span>
387
+ <span class="cline-any cline-neutral">&nbsp;</span>
388
+ <span class="cline-any cline-no">&nbsp;</span>
389
+ <span class="cline-any cline-neutral">&nbsp;</span>
390
+ <span class="cline-any cline-no">&nbsp;</span>
391
+ <span class="cline-any cline-neutral">&nbsp;</span>
392
+ <span class="cline-any cline-neutral">&nbsp;</span>
393
+ <span class="cline-any cline-neutral">&nbsp;</span>
394
+ <span class="cline-any cline-neutral">&nbsp;</span>
395
+ <span class="cline-any cline-neutral">&nbsp;</span>
396
+ <span class="cline-any cline-neutral">&nbsp;</span>
397
+ <span class="cline-any cline-neutral">&nbsp;</span>
398
+ <span class="cline-any cline-neutral">&nbsp;</span>
399
+ <span class="cline-any cline-neutral">&nbsp;</span>
400
+ <span class="cline-any cline-neutral">&nbsp;</span>
401
+ <span class="cline-any cline-neutral">&nbsp;</span>
402
+ <span class="cline-any cline-neutral">&nbsp;</span>
403
+ <span class="cline-any cline-neutral">&nbsp;</span>
404
+ <span class="cline-any cline-neutral">&nbsp;</span>
405
+ <span class="cline-any cline-neutral">&nbsp;</span>
406
+ <span class="cline-any cline-neutral">&nbsp;</span>
407
+ <span class="cline-any cline-neutral">&nbsp;</span>
408
+ <span class="cline-any cline-neutral">&nbsp;</span>
409
+ <span class="cline-any cline-neutral">&nbsp;</span>
410
+ <span class="cline-any cline-neutral">&nbsp;</span>
411
+ <span class="cline-any cline-neutral">&nbsp;</span>
412
+ <span class="cline-any cline-neutral">&nbsp;</span>
413
+ <span class="cline-any cline-neutral">&nbsp;</span>
414
+ <span class="cline-any cline-neutral">&nbsp;</span>
415
+ <span class="cline-any cline-neutral">&nbsp;</span>
416
+ <span class="cline-any cline-neutral">&nbsp;</span>
417
+ <span class="cline-any cline-neutral">&nbsp;</span>
418
+ <span class="cline-any cline-neutral">&nbsp;</span>
419
+ <span class="cline-any cline-neutral">&nbsp;</span>
420
+ <span class="cline-any cline-neutral">&nbsp;</span>
421
+ <span class="cline-any cline-neutral">&nbsp;</span>
422
+ <span class="cline-any cline-neutral">&nbsp;</span>
423
+ <span class="cline-any cline-neutral">&nbsp;</span>
424
+ <span class="cline-any cline-neutral">&nbsp;</span>
425
+ <span class="cline-any cline-neutral">&nbsp;</span>
426
+ <span class="cline-any cline-neutral">&nbsp;</span>
427
+ <span class="cline-any cline-neutral">&nbsp;</span>
428
+ <span class="cline-any cline-neutral">&nbsp;</span>
429
+ <span class="cline-any cline-neutral">&nbsp;</span>
430
+ <span class="cline-any cline-neutral">&nbsp;</span>
431
+ <span class="cline-any cline-neutral">&nbsp;</span>
432
+ <span class="cline-any cline-neutral">&nbsp;</span>
433
+ <span class="cline-any cline-neutral">&nbsp;</span>
434
+ <span class="cline-any cline-neutral">&nbsp;</span>
435
+ <span class="cline-any cline-neutral">&nbsp;</span>
436
+ <span class="cline-any cline-neutral">&nbsp;</span>
437
+ <span class="cline-any cline-neutral">&nbsp;</span>
438
+ <span class="cline-any cline-yes">3x</span>
439
+ <span class="cline-any cline-yes">3x</span>
440
+ <span class="cline-any cline-yes">3x</span>
441
+ <span class="cline-any cline-no">&nbsp;</span>
442
+ <span class="cline-any cline-neutral">&nbsp;</span>
443
+ <span class="cline-any cline-neutral">&nbsp;</span>
444
+ <span class="cline-any cline-neutral">&nbsp;</span>
445
+ <span class="cline-any cline-neutral">&nbsp;</span>
446
+ <span class="cline-any cline-neutral">&nbsp;</span>
447
+ <span class="cline-any cline-neutral">&nbsp;</span>
448
+ <span class="cline-any cline-neutral">&nbsp;</span>
449
+ <span class="cline-any cline-yes">3x</span>
450
+ <span class="cline-any cline-yes">3x</span>
451
+ <span class="cline-any cline-yes">3x</span>
452
+ <span class="cline-any cline-yes">3x</span>
453
+ <span class="cline-any cline-yes">1x</span>
454
+ <span class="cline-any cline-neutral">&nbsp;</span>
455
+ <span class="cline-any cline-neutral">&nbsp;</span>
456
+ <span class="cline-any cline-neutral">&nbsp;</span>
457
+ <span class="cline-any cline-neutral">&nbsp;</span>
458
+ <span class="cline-any cline-neutral">&nbsp;</span>
459
+ <span class="cline-any cline-neutral">&nbsp;</span>
460
+ <span class="cline-any cline-neutral">&nbsp;</span>
461
+ <span class="cline-any cline-neutral">&nbsp;</span>
462
+ <span class="cline-any cline-neutral">&nbsp;</span>
463
+ <span class="cline-any cline-yes">2x</span>
464
+ <span class="cline-any cline-yes">2x</span>
465
+ <span class="cline-any cline-neutral">&nbsp;</span>
466
+ <span class="cline-any cline-yes">2x</span>
467
+ <span class="cline-any cline-neutral">&nbsp;</span>
468
+ <span class="cline-any cline-neutral">&nbsp;</span>
469
+ <span class="cline-any cline-neutral">&nbsp;</span>
470
+ <span class="cline-any cline-neutral">&nbsp;</span>
471
+ <span class="cline-any cline-neutral">&nbsp;</span>
472
+ <span class="cline-any cline-neutral">&nbsp;</span>
473
+ <span class="cline-any cline-neutral">&nbsp;</span>
474
+ <span class="cline-any cline-neutral">&nbsp;</span>
475
+ <span class="cline-any cline-neutral">&nbsp;</span>
476
+ <span class="cline-any cline-neutral">&nbsp;</span></td><td class="text"><pre class="prettyprint lang-js">import { z } from 'zod';
477
+ import { namedJudge } from 'vitest-evals';
478
+ import type { JudgeContext } from 'vitest-evals';
479
+ import {
480
+ normalizeContent,
481
+ normalizeMetadata,
482
+ toJsonValue,
483
+ type Harness,
484
+ } from 'vitest-evals/harness';
485
+ import { runJudge } from './judge.js';
486
+ import { runEvalSkill, type RunEvalOptions } from './runner.js';
487
+ import { evalPassed, type EvalMeta, type JudgeResponse } from './types.js';
488
+ import { usageToSummary } from './usage.js';
489
+ import { FindingSchema } from '../types/index.js';
490
+ import type { Finding, SkillReport, UsageStats } from '../types/index.js';
491
+ &nbsp;
492
+ export const WardenEvalOutputSchema = z.object({
493
+ name: z.string(),
494
+ summary: z.string(),
495
+ skill: z.string(),
496
+ runtime: z.string().optional(),
497
+ model: z.string().optional(),
498
+ findings: z.array(FindingSchema),
499
+ failedHunks: z.number().int().nonnegative().optional(),
500
+ failedExtractions: z.number().int().nonnegative().optional(),
501
+ });
502
+ export type WardenEvalOutput = z.infer&lt;typeof WardenEvalOutputSchema&gt;;
503
+ &nbsp;
504
+ function usageMetadata(usage: UsageStats | undefined): Record&lt;string, unknown&gt; | undefined {
505
+ <span class="missing-if-branch" title="if path not taken" >I</span>if (!usage) {
506
+ <span class="cstat-no" title="statement not covered" > return undefined;</span>
507
+ }
508
+ &nbsp;
509
+ return {
510
+ inputTokens: usage.inputTokens,
511
+ outputTokens: usage.outputTokens,
512
+ costUSD: usage.costUSD,
513
+ };
514
+ }
515
+ &nbsp;
516
+ function <span class="fstat-no" title="function not covered" >reportToOutput(n</span>ame: string, report: SkillReport): WardenEvalOutput {
517
+ <span class="cstat-no" title="statement not covered" > return {</span>
518
+ name,
519
+ summary: report.summary,
520
+ skill: report.skill,
521
+ runtime: report.runtime,
522
+ model: report.model,
523
+ findings: report.findings,
524
+ failedHunks: report.failedHunks,
525
+ failedExtractions: report.failedExtractions,
526
+ };
527
+ }
528
+ &nbsp;
529
+ function failedJudgeReasons(
530
+ meta: EvalMeta,
531
+ response: JudgeResponse,
532
+ findings: Finding[]
533
+ ): string[] {
534
+ const reasons: string[] = [];
535
+ &nbsp;
536
+ for (let i = 0; i &lt; meta.should_find.length; i++) {
537
+ const assertion = meta.should_find[i];
538
+ if (!assertion?.required) {
539
+ continue;
540
+ }
541
+ &nbsp;
542
+ const verdict = response.expectations[i];
543
+ <span class="missing-if-branch" title="if path not taken" >I</span>if (!verdict) {
544
+ <span class="cstat-no" title="statement not covered" > reasons.push(`missing verdict for should_find[${i}]`);</span>
545
+ <span class="cstat-no" title="statement not covered" > continue;</span>
546
+ }
547
+ &nbsp;
548
+ const matchedFinding = verdict.matchedFindingIndex === null
549
+ ? undefined
550
+ : findings[verdict.matchedFindingIndex];
551
+ <span class="missing-if-branch" title="if path not taken" >I</span>if (!verdict.met) {
552
+ <span class="cstat-no" title="statement not covered" > reasons.push(`should_find[${i}] not met: ${verdict.reasoning}`);</span>
553
+ } else if (assertion.severity) {
554
+ if (!matchedFinding) {
555
+ reasons.push(`should_find[${i}] severity could not be checked: no matched finding`);
556
+ <span class="cstat-no" title="statement not covered" > <span class="missing-if-branch" title="else path not taken" >E</span>} else if (matchedFinding.severity !== assertion.severity) {</span>
557
+ <span class="cstat-no" title="statement not covered" > reasons.push(</span>
558
+ `should_find[${i}] severity mismatch: expected ${assertion.severity}, got ${matchedFinding.severity}`
559
+ );
560
+ }
561
+ }
562
+ }
563
+ &nbsp;
564
+ for (let i = 0; i &lt; meta.should_not_find.length; i++) {
565
+ const verdict = <span class="cstat-no" title="statement not covered" >response.antiExpectations[i];</span>
566
+ <span class="cstat-no" title="statement not covered" > if (verdict?.violated) {</span>
567
+ <span class="cstat-no" title="statement not covered" > reasons.push(`should_not_find[${i}] violated: ${verdict.reasoning}`);</span>
568
+ }
569
+ }
570
+ &nbsp;
571
+ return reasons;
572
+ }
573
+ &nbsp;
574
+ export function <span class="fstat-no" title="function not covered" >createWardenEvalHarness(o</span>ptions: RunEvalOptions): Harness&lt;EvalMeta&gt; {
575
+ <span class="cstat-no" title="statement not covered" > return {</span>
576
+ name: 'warden',
577
+ prompt: <span class="fstat-no" title="function not covered" >async () =&gt; {</span>
578
+ <span class="cstat-no" title="statement not covered" > throw new Error('Warden eval judges use runJudge directly.');</span>
579
+ },
580
+ run: <span class="fstat-no" title="function not covered" >async (m</span>eta, context) =&gt; {
581
+ const modelOverride = <span class="cstat-no" title="statement not covered" >typeof context.metadata['model'] === 'string'</span>
582
+ ? context.metadata['model']
583
+ : undefined;
584
+ const runtimeOverride = <span class="cstat-no" title="statement not covered" >context.metadata['runtime'] === 'claude'</span>
585
+ || context.metadata['runtime'] === 'pi'
586
+ ? context.metadata['runtime']
587
+ : undefined;
588
+ const result = <span class="cstat-no" title="statement not covered" >await runEvalSkill(meta, {</span>
589
+ ...options,
590
+ model: modelOverride ?? options.model,
591
+ runtime: runtimeOverride ?? options.runtime,
592
+ });
593
+ const output = <span class="cstat-no" title="statement not covered" >reportToOutput(result.name, result.report);</span>
594
+ &nbsp;
595
+ <span class="cstat-no" title="statement not covered" > return {</span>
596
+ output: toJsonValue(output),
597
+ session: {
598
+ messages: [
599
+ {
600
+ role: 'user',
601
+ content: normalizeContent({
602
+ name: result.name,
603
+ given: meta.given,
604
+ shouldFind: meta.should_find,
605
+ shouldNotFind: meta.should_not_find,
606
+ }),
607
+ },
608
+ {
609
+ role: 'assistant',
610
+ content: normalizeContent(output),
611
+ },
612
+ ],
613
+ outputText: result.report.summary,
614
+ provider: result.report.runtime,
615
+ model: result.report.model,
616
+ metadata: normalizeMetadata({
617
+ category: meta.category,
618
+ scenario: meta.name,
619
+ skill: result.report.skill,
620
+ }),
621
+ },
622
+ usage: usageToSummary({
623
+ provider: result.report.runtime ?? 'unknown',
624
+ model: result.report.model ?? 'unknown',
625
+ usage: result.report.usage,
626
+ }),
627
+ timings: { totalMs: result.durationMs },
628
+ artifacts: {
629
+ logs: toJsonValue(result.logs) ?? [],
630
+ },
631
+ errors: result.report.error
632
+ ? [{
633
+ type: result.report.error.code,
634
+ message: result.report.error.message,
635
+ }]
636
+ : [],
637
+ };
638
+ },
639
+ };
640
+ }
641
+ &nbsp;
642
+ export function createWardenEvalJudge(apiKey: string) {
643
+ return namedJudge&lt;JudgeContext&lt;EvalMeta&gt;&gt;('WardenEvalJudge', async ({ inputValue, run }) =&gt; {
644
+ const output = WardenEvalOutputSchema.safeParse(run.output);
645
+ <span class="missing-if-branch" title="if path not taken" >I</span>if (!output.success) {
646
+ <span class="cstat-no" title="statement not covered" > return {</span>
647
+ score: 0,
648
+ metadata: {
649
+ rationale: `Invalid Warden harness output: ${output.error.message}`,
650
+ },
651
+ };
652
+ }
653
+ &nbsp;
654
+ const meta = inputValue;
655
+ const findings = output.data.findings;
656
+ const judgeResult = await runJudge(meta, findings, apiKey);
657
+ if (judgeResult.error) {
658
+ return {
659
+ score: 0,
660
+ metadata: {
661
+ rationale: `Judge failed: ${judgeResult.error}`,
662
+ output: judgeResult.response,
663
+ usage: usageMetadata(judgeResult.usage),
664
+ },
665
+ };
666
+ }
667
+ &nbsp;
668
+ const passed = evalPassed(meta, judgeResult.response, findings);
669
+ const reasons = failedJudgeReasons(meta, judgeResult.response, findings);
670
+ &nbsp;
671
+ return {
672
+ score: passed ? 1 : 0,
673
+ metadata: {
674
+ rationale: reasons.length &gt; 0 ? reasons.join('; ') : 'All eval assertions passed.',
675
+ output: judgeResult.response,
676
+ usage: usageMetadata(judgeResult.usage),
677
+ },
678
+ };
679
+ });
680
+ }
681
+ &nbsp;</pre></td></tr></table></pre>
682
+
683
+ <div class='push'></div><!-- for sticky footer -->
684
+ </div><!-- /wrapper -->
685
+ <div class='footer quiet pad2 space-top1 center small'>
686
+ Code coverage generated by
687
+ <a href="https://istanbul.js.org/" target="_blank" rel="noopener noreferrer">istanbul</a>
688
+ at 2026-05-17T04:35:52.381Z
689
+ </div>
690
+ <script src="../../prettify.js"></script>
691
+ <script>
692
+ window.onload = function () {
693
+ prettyPrint();
694
+ };
695
+ </script>
696
+ <script src="../../sorter.js"></script>
697
+ <script src="../../block-navigation.js"></script>
698
+ </body>
699
+ </html>
700
+