@langwatch/mcp-server 1.0.0 → 2.0.0

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 (125) hide show
  1. package/.python-version +1 -0
  2. package/CHANGELOG.md +21 -0
  3. package/CONTRIBUTING.md +1 -1
  4. package/README.md +55 -1
  5. package/dist/{archive-project-MK3673A7.js → archive-project-4EQGLBJJ.js} +5 -5
  6. package/dist/{archive-scenario-KATSXRIN.js → archive-scenario-W6BXYDZ3.js} +5 -5
  7. package/dist/{archive-suite-C3KX6IIZ.js → archive-suite-4T6BSRS7.js} +5 -5
  8. package/dist/{assign-prompt-tag-E3C5W6PI.js → assign-prompt-tag-VLXUWEKU.js} +4 -4
  9. package/dist/{chunk-VYPSKRPN.js → chunk-4QC7SA2L.js} +91 -91
  10. package/dist/{chunk-T7WJVS2N.js → chunk-5CCZFUXS.js} +2 -2
  11. package/dist/{chunk-J45SVWPR.js → chunk-77BNWQ3U.js} +2 -2
  12. package/dist/{chunk-XY4STOBE.js → chunk-7RHG56OH.js} +2 -2
  13. package/dist/{chunk-5UWY24RP.js → chunk-A2LEVYQ5.js} +2 -2
  14. package/dist/{chunk-UXXA4D6K.js → chunk-AQJO2EI5.js} +2 -2
  15. package/dist/{chunk-EU22L52K.js → chunk-CDYQOY5U.js} +2 -2
  16. package/dist/{chunk-O3LWVN67.js → chunk-J42FWRJ2.js} +2 -2
  17. package/dist/{chunk-BJ7B4R26.js → chunk-SAS5OB7F.js} +18 -2
  18. package/dist/{chunk-DAIQKS27.js → chunk-SFB4MU6H.js} +2 -2
  19. package/dist/{chunk-WWMDHZZV.js → chunk-TJCXFNDT.js} +2 -2
  20. package/dist/{chunk-CE6QMZOP.js → chunk-TOXJQDIZ.js} +4 -4
  21. package/dist/chunk-UJPXFA26.js +267 -0
  22. package/dist/{chunk-KOO5733S.js → chunk-WT4PO4A3.js} +8 -2
  23. package/dist/{chunk-RYRB2YVJ.js → chunk-YX3A3CFM.js} +2 -2
  24. package/dist/{chunk-WDH5P5RF.js → chunk-YZLYHLLI.js} +2 -2
  25. package/dist/{chunk-Y3HDZ6OW.js → chunk-ZWIL4EFL.js} +2 -2
  26. package/dist/config.js +2 -2
  27. package/dist/{create-agent-7W57J7MT.js → create-agent-7XHO6UIS.js} +5 -5
  28. package/dist/{create-annotation-W2XELBCQ.js → create-annotation-YQXRQC6A.js} +5 -5
  29. package/dist/{create-api-key-MKRXKFJM.js → create-api-key-YUS36KBU.js} +5 -5
  30. package/dist/{create-dashboard-QGL7IODC.js → create-dashboard-4Z5JO5LB.js} +5 -5
  31. package/dist/{create-dataset-3PH7A4HG.js → create-dataset-E6NMSVX7.js} +5 -5
  32. package/dist/{create-dataset-records-4DZS3VRL.js → create-dataset-records-SKHPRD2K.js} +5 -5
  33. package/dist/{create-evaluator-YWH6NC4M.js → create-evaluator-JXF5PF33.js} +5 -5
  34. package/dist/create-mcp-server.js +4 -4
  35. package/dist/{create-project-BOBQ32C5.js → create-project-CS2CB57Y.js} +5 -5
  36. package/dist/{create-prompt-FLCFGITE.js → create-prompt-F5RMX3L3.js} +4 -4
  37. package/dist/{create-prompt-tag-EMPNIKEA.js → create-prompt-tag-GMRWTEFR.js} +4 -4
  38. package/dist/{create-scenario-BSBIHUL6.js → create-scenario-JDR2VB2H.js} +5 -5
  39. package/dist/{create-suite-TWF6BBR4.js → create-suite-LZS6QIAF.js} +5 -5
  40. package/dist/{delete-agent-7MSYNY5E.js → delete-agent-O3EEH4S5.js} +5 -5
  41. package/dist/{delete-annotation-ZT5IGN24.js → delete-annotation-XP6FQM7A.js} +5 -5
  42. package/dist/{delete-dashboard-BTIY5S3L.js → delete-dashboard-RFUNYCNP.js} +5 -5
  43. package/dist/{delete-dataset-U45HQHPR.js → delete-dataset-3MJFRDLP.js} +5 -5
  44. package/dist/{delete-dataset-records-R43IMAWB.js → delete-dataset-records-WPMU5GL5.js} +5 -5
  45. package/dist/{delete-evaluator-U6JMQ6C7.js → delete-evaluator-33SP2JFX.js} +5 -5
  46. package/dist/{delete-prompt-tag-4EV4TDHB.js → delete-prompt-tag-KNMZVSPB.js} +4 -4
  47. package/dist/{delete-workflow-5QYALXEL.js → delete-workflow-3NUXO46U.js} +5 -5
  48. package/dist/{discover-evaluator-schema-MRGHVNTT.js → discover-evaluator-schema-XTAD4X7D.js} +52 -36
  49. package/dist/{discover-scenario-schema-UPBTCWPG.js → discover-scenario-schema-FE64QK2D.js} +2 -2
  50. package/dist/{discover-schema-ICGXPHUW.js → discover-schema-L6QDPON3.js} +2 -2
  51. package/dist/{get-agent-7QJVPH72.js → get-agent-Y7MWZ65F.js} +5 -5
  52. package/dist/{get-analytics-PVERWOA5.js → get-analytics-TETCSOK6.js} +4 -4
  53. package/dist/{get-dashboard-ZZPC3WMD.js → get-dashboard-KHBEGZGL.js} +5 -5
  54. package/dist/{get-dataset-25TB2LQH.js → get-dataset-RCMU5VAH.js} +5 -5
  55. package/dist/{get-evaluator-3FTRQKEC.js → get-evaluator-ADMKVWYZ.js} +5 -5
  56. package/dist/{get-project-5OAXYC57.js → get-project-RAK7URQJ.js} +5 -5
  57. package/dist/{get-prompt-I753LZKZ.js → get-prompt-OY5B7TRI.js} +4 -4
  58. package/dist/{get-scenario-DSSTR6FI.js → get-scenario-ZTAJJL5T.js} +5 -5
  59. package/dist/{get-simulation-run-RQPMS2MC.js → get-simulation-run-T2ZQWYH2.js} +5 -5
  60. package/dist/{get-suite-GTOCY5PL.js → get-suite-J7JVWO6O.js} +5 -5
  61. package/dist/{get-trace-O5VT3NAC.js → get-trace-K7U234Q4.js} +4 -4
  62. package/dist/{get-workflow-3EZDOEIF.js → get-workflow-L5L2PSK3.js} +5 -5
  63. package/dist/http-security-EU3CHTY7.js +26 -0
  64. package/dist/{http-server-2ACJ3CXW.js → http-server-YXHFPPGA.js} +1033 -693
  65. package/dist/index.js +3066 -2105
  66. package/dist/{langwatch-api-annotations-E65MI3WZ.js → langwatch-api-annotations-FDAS6NQ7.js} +5 -5
  67. package/dist/{langwatch-api-dashboards-QCDECRCY.js → langwatch-api-dashboards-JTVCS7ZJ.js} +5 -5
  68. package/dist/{langwatch-api-monitors-UIO2KTDV.js → langwatch-api-monitors-3J7UIQME.js} +4 -4
  69. package/dist/{langwatch-api-secrets-4T2EKSJ7.js → langwatch-api-secrets-BUBSNCNO.js} +4 -4
  70. package/dist/{langwatch-api-triggers-MPOEKPUY.js → langwatch-api-triggers-EDFUCJN6.js} +4 -4
  71. package/dist/langwatch-api-workflows-NBJX5KOG.js +16 -0
  72. package/dist/{list-agents-TIJTP566.js → list-agents-KFHKGPXV.js} +5 -5
  73. package/dist/{list-annotations-CGX5BT7E.js → list-annotations-ZTVP4EEL.js} +5 -5
  74. package/dist/{list-api-keys-JLWJPFRX.js → list-api-keys-CWD4UQT5.js} +5 -5
  75. package/dist/{list-dashboards-RNIWX3TX.js → list-dashboards-PWG6P6KR.js} +5 -5
  76. package/dist/{list-dataset-records-B5MHZR66.js → list-dataset-records-IOE7LQXS.js} +5 -5
  77. package/dist/{list-datasets-BVCAMQCP.js → list-datasets-ZT4JYUDT.js} +5 -5
  78. package/dist/{list-evaluators-65FBTL2U.js → list-evaluators-SJHC7VBH.js} +5 -5
  79. package/dist/{list-model-providers-GRK3I3VV.js → list-model-providers-PPD2WPD2.js} +5 -5
  80. package/dist/{list-projects-JLMBFMW3.js → list-projects-K7QAIX2I.js} +5 -5
  81. package/dist/{list-prompt-tags-UHNWDGQI.js → list-prompt-tags-FOO6TNHB.js} +4 -4
  82. package/dist/{list-prompts-7254B3U7.js → list-prompts-SC2HAUFG.js} +4 -4
  83. package/dist/{list-scenarios-46YNICKM.js → list-scenarios-F2WYNRIE.js} +5 -5
  84. package/dist/{list-simulation-runs-LGNEFNPN.js → list-simulation-runs-KF5THG5H.js} +5 -5
  85. package/dist/{list-suites-5VK4HXEK.js → list-suites-CRY4VWW6.js} +5 -5
  86. package/dist/{list-workflows-MLQDM6CP.js → list-workflows-QSJWURCK.js} +5 -5
  87. package/dist/{rename-prompt-tag-RHQMNEQL.js → rename-prompt-tag-EKWU5QF3.js} +4 -4
  88. package/dist/{report-issue-DXJ6TPTC.js → report-issue-XYVAPAYJ.js} +7 -7
  89. package/dist/{revoke-api-key-3Y35EMBT.js → revoke-api-key-FTRTZMOF.js} +5 -5
  90. package/dist/{run-agent-RBSIFKRT.js → run-agent-6SM4XNA7.js} +5 -5
  91. package/dist/{run-suite-K3DNQ3SE.js → run-suite-NKZCJ2IN.js} +5 -5
  92. package/dist/{run-workflow-5LVBZYLS.js → run-workflow-QPXC627O.js} +5 -5
  93. package/dist/{search-traces-U4HLOCZ2.js → search-traces-A6HFKPLJ.js} +4 -4
  94. package/dist/{set-model-provider-V3M5CDIY.js → set-model-provider-3WKT63ZN.js} +5 -5
  95. package/dist/{update-agent-47RZIHS7.js → update-agent-SMAEHLXH.js} +5 -5
  96. package/dist/{update-dataset-G6QJJGS5.js → update-dataset-BDUK7TUO.js} +5 -5
  97. package/dist/{update-dataset-record-RMWLSWCS.js → update-dataset-record-4ZSDRBSW.js} +5 -5
  98. package/dist/{update-evaluator-SFRRI5Q3.js → update-evaluator-LS6TIBY7.js} +5 -5
  99. package/dist/{update-project-PI5DWBVK.js → update-project-LKVYXENH.js} +5 -5
  100. package/dist/{update-prompt-RQFTEO7Y.js → update-prompt-ODCRQRN6.js} +4 -4
  101. package/dist/{update-scenario-G3IU2UIB.js → update-scenario-NEXCAQI7.js} +5 -5
  102. package/dist/{update-suite-TRGGLNR7.js → update-suite-SNO3GS7T.js} +5 -5
  103. package/package.json +3 -3
  104. package/pyproject.toml +1 -1
  105. package/src/__tests__/config.unit.test.ts +33 -0
  106. package/src/__tests__/docker.integration.test.ts +91 -7
  107. package/src/__tests__/documentation-fetch-security.integration.test.ts +8 -2
  108. package/src/__tests__/feature-map-tool-drift.unit.test.ts +93 -0
  109. package/src/__tests__/http-security-origin.unit.test.ts +97 -0
  110. package/src/__tests__/http-security-sessions.unit.test.ts +209 -0
  111. package/src/__tests__/http-security-verifier.unit.test.ts +248 -0
  112. package/src/__tests__/http-server-allocation.integration.test.ts +279 -0
  113. package/src/__tests__/http-server-origin.integration.test.ts +169 -0
  114. package/src/__tests__/http-server-session-auth.integration.test.ts +341 -0
  115. package/src/__tests__/http-transport.integration.test.ts +85 -87
  116. package/src/__tests__/support/http-server-harness.ts +138 -0
  117. package/src/config.ts +17 -2
  118. package/src/http-security.ts +492 -0
  119. package/src/http-server.ts +536 -152
  120. package/src/index.ts +49 -2
  121. package/src/tools/discover-evaluator-schema.ts +2 -2
  122. package/tsconfig.json +2 -2
  123. package/uv.lock +2 -2
  124. package/dist/langwatch-api-workflows-5HRZ7HR4.js +0 -16
  125. package/pnpm-workspace.yaml +0 -34
@@ -1,10 +1,10 @@
1
1
  import {
2
2
  createEvaluator,
3
3
  getEvaluatorType
4
- } from "./chunk-EU22L52K.js";
5
- import "./chunk-5UWY24RP.js";
6
- import "./chunk-KOO5733S.js";
7
- import "./chunk-BJ7B4R26.js";
4
+ } from "./chunk-CDYQOY5U.js";
5
+ import "./chunk-A2LEVYQ5.js";
6
+ import "./chunk-WT4PO4A3.js";
7
+ import "./chunk-SAS5OB7F.js";
8
8
 
9
9
  // src/tools/create-evaluator.ts
10
10
  async function handleCreateEvaluator(params) {
@@ -25,4 +25,4 @@ async function handleCreateEvaluator(params) {
25
25
  export {
26
26
  handleCreateEvaluator
27
27
  };
28
- //# sourceMappingURL=create-evaluator-YWH6NC4M.js.map
28
+ //# sourceMappingURL=create-evaluator-JXF5PF33.js.map
@@ -1,9 +1,9 @@
1
1
  import {
2
2
  createMcpServer
3
- } from "./chunk-VYPSKRPN.js";
4
- import "./chunk-5UWY24RP.js";
5
- import "./chunk-KOO5733S.js";
6
- import "./chunk-BJ7B4R26.js";
3
+ } from "./chunk-4QC7SA2L.js";
4
+ import "./chunk-A2LEVYQ5.js";
5
+ import "./chunk-WT4PO4A3.js";
6
+ import "./chunk-SAS5OB7F.js";
7
7
  export {
8
8
  createMcpServer
9
9
  };
@@ -1,9 +1,9 @@
1
1
  import {
2
2
  createProject
3
- } from "./chunk-XY4STOBE.js";
4
- import "./chunk-5UWY24RP.js";
5
- import "./chunk-KOO5733S.js";
6
- import "./chunk-BJ7B4R26.js";
3
+ } from "./chunk-7RHG56OH.js";
4
+ import "./chunk-A2LEVYQ5.js";
5
+ import "./chunk-WT4PO4A3.js";
6
+ import "./chunk-SAS5OB7F.js";
7
7
 
8
8
  // src/tools/create-project.ts
9
9
  async function handleCreateProject(params) {
@@ -42,4 +42,4 @@ async function handleCreateProject(params) {
42
42
  export {
43
43
  handleCreateProject
44
44
  };
45
- //# sourceMappingURL=create-project-BOBQ32C5.js.map
45
+ //# sourceMappingURL=create-project-CS2CB57Y.js.map
@@ -1,8 +1,8 @@
1
1
  import {
2
2
  createPrompt
3
- } from "./chunk-5UWY24RP.js";
4
- import "./chunk-KOO5733S.js";
5
- import "./chunk-BJ7B4R26.js";
3
+ } from "./chunk-A2LEVYQ5.js";
4
+ import "./chunk-WT4PO4A3.js";
5
+ import "./chunk-SAS5OB7F.js";
6
6
 
7
7
  // src/tools/create-prompt.ts
8
8
  var HANDLE_PATTERN = /^[a-z0-9_-]+(?:\/[a-z0-9_-]+)?$/;
@@ -37,4 +37,4 @@ async function handleCreatePrompt(params) {
37
37
  export {
38
38
  handleCreatePrompt
39
39
  };
40
- //# sourceMappingURL=create-prompt-FLCFGITE.js.map
40
+ //# sourceMappingURL=create-prompt-F5RMX3L3.js.map
@@ -1,8 +1,8 @@
1
1
  import {
2
2
  createPromptTag
3
- } from "./chunk-5UWY24RP.js";
4
- import "./chunk-KOO5733S.js";
5
- import "./chunk-BJ7B4R26.js";
3
+ } from "./chunk-A2LEVYQ5.js";
4
+ import "./chunk-WT4PO4A3.js";
5
+ import "./chunk-SAS5OB7F.js";
6
6
 
7
7
  // src/tools/create-prompt-tag.ts
8
8
  async function handleCreatePromptTag(params) {
@@ -15,4 +15,4 @@ async function handleCreatePromptTag(params) {
15
15
  export {
16
16
  handleCreatePromptTag
17
17
  };
18
- //# sourceMappingURL=create-prompt-tag-EMPNIKEA.js.map
18
+ //# sourceMappingURL=create-prompt-tag-GMRWTEFR.js.map
@@ -1,9 +1,9 @@
1
1
  import {
2
2
  createScenario
3
- } from "./chunk-O3LWVN67.js";
4
- import "./chunk-5UWY24RP.js";
5
- import "./chunk-KOO5733S.js";
6
- import "./chunk-BJ7B4R26.js";
3
+ } from "./chunk-J42FWRJ2.js";
4
+ import "./chunk-A2LEVYQ5.js";
5
+ import "./chunk-WT4PO4A3.js";
6
+ import "./chunk-SAS5OB7F.js";
7
7
 
8
8
  // src/tools/create-scenario.ts
9
9
  async function handleCreateScenario(params) {
@@ -24,4 +24,4 @@ async function handleCreateScenario(params) {
24
24
  export {
25
25
  handleCreateScenario
26
26
  };
27
- //# sourceMappingURL=create-scenario-BSBIHUL6.js.map
27
+ //# sourceMappingURL=create-scenario-JDR2VB2H.js.map
@@ -1,9 +1,9 @@
1
1
  import {
2
2
  createSuite
3
- } from "./chunk-WDH5P5RF.js";
4
- import "./chunk-5UWY24RP.js";
5
- import "./chunk-KOO5733S.js";
6
- import "./chunk-BJ7B4R26.js";
3
+ } from "./chunk-YZLYHLLI.js";
4
+ import "./chunk-A2LEVYQ5.js";
5
+ import "./chunk-WT4PO4A3.js";
6
+ import "./chunk-SAS5OB7F.js";
7
7
 
8
8
  // src/tools/create-suite.ts
9
9
  async function handleCreateSuite(params) {
@@ -36,4 +36,4 @@ async function handleCreateSuite(params) {
36
36
  export {
37
37
  handleCreateSuite
38
38
  };
39
- //# sourceMappingURL=create-suite-TWF6BBR4.js.map
39
+ //# sourceMappingURL=create-suite-LZS6QIAF.js.map
@@ -1,9 +1,9 @@
1
1
  import {
2
2
  deleteAgent
3
- } from "./chunk-CE6QMZOP.js";
4
- import "./chunk-5UWY24RP.js";
5
- import "./chunk-KOO5733S.js";
6
- import "./chunk-BJ7B4R26.js";
3
+ } from "./chunk-TOXJQDIZ.js";
4
+ import "./chunk-A2LEVYQ5.js";
5
+ import "./chunk-WT4PO4A3.js";
6
+ import "./chunk-SAS5OB7F.js";
7
7
 
8
8
  // src/tools/delete-agent.ts
9
9
  async function handleDeleteAgent(params) {
@@ -16,4 +16,4 @@ async function handleDeleteAgent(params) {
16
16
  export {
17
17
  handleDeleteAgent
18
18
  };
19
- //# sourceMappingURL=delete-agent-7MSYNY5E.js.map
19
+ //# sourceMappingURL=delete-agent-O3EEH4S5.js.map
@@ -1,9 +1,9 @@
1
1
  import {
2
2
  deleteAnnotation
3
- } from "./chunk-RYRB2YVJ.js";
4
- import "./chunk-5UWY24RP.js";
5
- import "./chunk-KOO5733S.js";
6
- import "./chunk-BJ7B4R26.js";
3
+ } from "./chunk-YX3A3CFM.js";
4
+ import "./chunk-A2LEVYQ5.js";
5
+ import "./chunk-WT4PO4A3.js";
6
+ import "./chunk-SAS5OB7F.js";
7
7
 
8
8
  // src/tools/delete-annotation.ts
9
9
  async function handleDeleteAnnotation(params) {
@@ -15,4 +15,4 @@ async function handleDeleteAnnotation(params) {
15
15
  export {
16
16
  handleDeleteAnnotation
17
17
  };
18
- //# sourceMappingURL=delete-annotation-ZT5IGN24.js.map
18
+ //# sourceMappingURL=delete-annotation-XP6FQM7A.js.map
@@ -1,9 +1,9 @@
1
1
  import {
2
2
  deleteDashboard
3
- } from "./chunk-UXXA4D6K.js";
4
- import "./chunk-5UWY24RP.js";
5
- import "./chunk-KOO5733S.js";
6
- import "./chunk-BJ7B4R26.js";
3
+ } from "./chunk-AQJO2EI5.js";
4
+ import "./chunk-A2LEVYQ5.js";
5
+ import "./chunk-WT4PO4A3.js";
6
+ import "./chunk-SAS5OB7F.js";
7
7
 
8
8
  // src/tools/delete-dashboard.ts
9
9
  async function handleDeleteDashboard(params) {
@@ -16,4 +16,4 @@ async function handleDeleteDashboard(params) {
16
16
  export {
17
17
  handleDeleteDashboard
18
18
  };
19
- //# sourceMappingURL=delete-dashboard-BTIY5S3L.js.map
19
+ //# sourceMappingURL=delete-dashboard-RFUNYCNP.js.map
@@ -1,9 +1,9 @@
1
1
  import {
2
2
  deleteDataset
3
- } from "./chunk-J45SVWPR.js";
4
- import "./chunk-5UWY24RP.js";
5
- import "./chunk-KOO5733S.js";
6
- import "./chunk-BJ7B4R26.js";
3
+ } from "./chunk-77BNWQ3U.js";
4
+ import "./chunk-A2LEVYQ5.js";
5
+ import "./chunk-WT4PO4A3.js";
6
+ import "./chunk-SAS5OB7F.js";
7
7
 
8
8
  // src/tools/delete-dataset.ts
9
9
  async function handleDeleteDataset(params) {
@@ -18,4 +18,4 @@ async function handleDeleteDataset(params) {
18
18
  export {
19
19
  handleDeleteDataset
20
20
  };
21
- //# sourceMappingURL=delete-dataset-U45HQHPR.js.map
21
+ //# sourceMappingURL=delete-dataset-3MJFRDLP.js.map
@@ -1,9 +1,9 @@
1
1
  import {
2
2
  deleteDatasetRecords
3
- } from "./chunk-J45SVWPR.js";
4
- import "./chunk-5UWY24RP.js";
5
- import "./chunk-KOO5733S.js";
6
- import "./chunk-BJ7B4R26.js";
3
+ } from "./chunk-77BNWQ3U.js";
4
+ import "./chunk-A2LEVYQ5.js";
5
+ import "./chunk-WT4PO4A3.js";
6
+ import "./chunk-SAS5OB7F.js";
7
7
 
8
8
  // src/tools/delete-dataset-records.ts
9
9
  async function handleDeleteDatasetRecords(params) {
@@ -19,4 +19,4 @@ async function handleDeleteDatasetRecords(params) {
19
19
  export {
20
20
  handleDeleteDatasetRecords
21
21
  };
22
- //# sourceMappingURL=delete-dataset-records-R43IMAWB.js.map
22
+ //# sourceMappingURL=delete-dataset-records-WPMU5GL5.js.map
@@ -1,9 +1,9 @@
1
1
  import {
2
2
  deleteEvaluator
3
- } from "./chunk-EU22L52K.js";
4
- import "./chunk-5UWY24RP.js";
5
- import "./chunk-KOO5733S.js";
6
- import "./chunk-BJ7B4R26.js";
3
+ } from "./chunk-CDYQOY5U.js";
4
+ import "./chunk-A2LEVYQ5.js";
5
+ import "./chunk-WT4PO4A3.js";
6
+ import "./chunk-SAS5OB7F.js";
7
7
 
8
8
  // src/tools/delete-evaluator.ts
9
9
  async function handleDeleteEvaluator(params) {
@@ -13,4 +13,4 @@ async function handleDeleteEvaluator(params) {
13
13
  export {
14
14
  handleDeleteEvaluator
15
15
  };
16
- //# sourceMappingURL=delete-evaluator-U6JMQ6C7.js.map
16
+ //# sourceMappingURL=delete-evaluator-33SP2JFX.js.map
@@ -1,8 +1,8 @@
1
1
  import {
2
2
  deletePromptTag
3
- } from "./chunk-5UWY24RP.js";
4
- import "./chunk-KOO5733S.js";
5
- import "./chunk-BJ7B4R26.js";
3
+ } from "./chunk-A2LEVYQ5.js";
4
+ import "./chunk-WT4PO4A3.js";
5
+ import "./chunk-SAS5OB7F.js";
6
6
 
7
7
  // src/tools/delete-prompt-tag.ts
8
8
  async function handleDeletePromptTag(params) {
@@ -15,4 +15,4 @@ async function handleDeletePromptTag(params) {
15
15
  export {
16
16
  handleDeletePromptTag
17
17
  };
18
- //# sourceMappingURL=delete-prompt-tag-4EV4TDHB.js.map
18
+ //# sourceMappingURL=delete-prompt-tag-KNMZVSPB.js.map
@@ -1,9 +1,9 @@
1
1
  import {
2
2
  deleteWorkflow
3
- } from "./chunk-T7WJVS2N.js";
4
- import "./chunk-5UWY24RP.js";
5
- import "./chunk-KOO5733S.js";
6
- import "./chunk-BJ7B4R26.js";
3
+ } from "./chunk-5CCZFUXS.js";
4
+ import "./chunk-A2LEVYQ5.js";
5
+ import "./chunk-WT4PO4A3.js";
6
+ import "./chunk-SAS5OB7F.js";
7
7
 
8
8
  // src/tools/delete-workflow.ts
9
9
  async function handleDeleteWorkflow(params) {
@@ -15,4 +15,4 @@ async function handleDeleteWorkflow(params) {
15
15
  export {
16
16
  handleDeleteWorkflow
17
17
  };
18
- //# sourceMappingURL=delete-workflow-5QYALXEL.js.map
18
+ //# sourceMappingURL=delete-workflow-3NUXO46U.js.map
@@ -1,6 +1,6 @@
1
- import "./chunk-BJ7B4R26.js";
1
+ import "./chunk-SAS5OB7F.js";
2
2
 
3
- // ../langevals/ts-integration/evaluators.generated.ts
3
+ // ../../services/langevals/ts-integration/evaluators.generated.ts
4
4
  import { z } from "zod";
5
5
  var evaluatorTypesSchema = z.string();
6
6
  var moneySchema = z.object({
@@ -580,6 +580,12 @@ var evaluatorsSchema = z.object({
580
580
  include_metrics: z.array(z.union([z.literal("cost"), z.literal("duration")])).describe("Per-candidate metrics to inject into the judge prompt").default([])
581
581
  })
582
582
  }),
583
+ "langevals/query_resolution": z.object({
584
+ settings: z.object({
585
+ model: z.string().describe("The model to use for evaluation").default("openai/gpt-5-mini"),
586
+ max_tokens: z.number().describe("Max tokens allowed for evaluation").default(128e3)
587
+ })
588
+ }),
583
589
  "langevals/select_best_compare": z.object({
584
590
  settings: z.object({
585
591
  model: z.string().describe("The model to use for evaluation").default("openai/gpt-5-mini"),
@@ -609,13 +615,13 @@ slot label. Use "tie" only when no candidate is clearly better.
609
615
  allow_tie: z.boolean().describe(
610
616
  "Allow the judge to return 'tie' when candidates are equivalent"
611
617
  ).default(true),
612
- include_metrics: z.array(z.union([z.literal("cost"), z.literal("duration")])).describe("Per-candidate metrics to inject into the judge prompt").default([])
613
- })
614
- }),
615
- "langevals/query_resolution": z.object({
616
- settings: z.object({
617
- model: z.string().describe("The model to use for evaluation").default("openai/gpt-5-mini"),
618
- max_tokens: z.number().describe("Max tokens allowed for evaluation").default(128e3)
618
+ include_metrics: z.array(z.union([z.literal("cost"), z.literal("duration")])).describe("Per-candidate metrics to inject into the judge prompt").default([]),
619
+ temperature: z.number().describe(
620
+ "Sampling temperature for the judge call. Lower is more deterministic."
621
+ ).default(0),
622
+ swap_and_reconcile: z.boolean().describe(
623
+ "Check each row twice, the second time with the candidate order reversed, and record no result for that row when the two disagree \u2014 rather than a tie, which would claim the candidates are equally good. Doubles judge-call cost per row (literature: swapping candidate order alone flips 10-30% of verdicts on contested rows)."
624
+ ).default(true)
619
625
  })
620
626
  }),
621
627
  "langevals/sentiment": z.object({
@@ -1866,19 +1872,44 @@ better one. Superseded by Comparison, which judges two or more candidates.
1866
1872
  }
1867
1873
  }
1868
1874
  },
1875
+ "langevals/query_resolution": {
1876
+ name: `Query Resolution`,
1877
+ description: `
1878
+ This evaluator checks if all the user queries in the conversation were resolved. Useful to detect when the bot doesn't know how to answer or can't help the user.
1879
+ `,
1880
+ category: "quality",
1881
+ docsUrl: "",
1882
+ isGuardrail: false,
1883
+ requiredFields: ["conversation"],
1884
+ optionalFields: [],
1885
+ settings: {
1886
+ model: {
1887
+ description: "The model to use for evaluation",
1888
+ default: "openai/gpt-5-mini"
1889
+ },
1890
+ max_tokens: {
1891
+ description: "Max tokens allowed for evaluation",
1892
+ default: 128e3
1893
+ }
1894
+ },
1895
+ envVars: [],
1896
+ result: {}
1897
+ },
1869
1898
  "langevals/select_best_compare": {
1870
1899
  name: `Comparison`,
1871
1900
  description: `
1872
1901
  Compare two or more candidate outputs and pick the best one, optionally
1873
1902
  against a reference answer. The judge sees every candidate at once and
1874
1903
  explains why the winner is better. Candidate order is shuffled so that a
1875
- candidate's position never sways the verdict.
1904
+ candidate's position never sways the verdict, and by default each row is
1905
+ checked twice \u2014 once more with the order reversed \u2014 recording no result
1906
+ for that row when the two checks disagree, rather than guessing.
1876
1907
  `,
1877
1908
  category: "quality",
1878
1909
  docsUrl: "",
1879
1910
  isGuardrail: false,
1880
- requiredFields: ["candidates"],
1881
- optionalFields: ["input", "golden", "row_index"],
1911
+ requiredFields: [],
1912
+ optionalFields: ["input", "golden", "candidates", "row_index"],
1882
1913
  settings: {
1883
1914
  model: {
1884
1915
  description: "The model to use for evaluation",
@@ -1918,6 +1949,14 @@ slot label. Use "tie" only when no candidate is clearly better.
1918
1949
  include_metrics: {
1919
1950
  description: "Per-candidate metrics to inject into the judge prompt",
1920
1951
  default: []
1952
+ },
1953
+ temperature: {
1954
+ description: "Sampling temperature for the judge call. Lower is more deterministic.",
1955
+ default: 0
1956
+ },
1957
+ swap_and_reconcile: {
1958
+ description: "Check each row twice, the second time with the candidate order reversed, and record no result for that row when the two disagree \u2014 rather than a tie, which would claim the candidates are equally good. Doubles judge-call cost per row (literature: swapping candidate order alone flips 10-30% of verdicts on contested rows).",
1959
+ default: true
1921
1960
  }
1922
1961
  },
1923
1962
  envVars: [],
@@ -1930,29 +1969,6 @@ slot label. Use "tie" only when no candidate is clearly better.
1930
1969
  }
1931
1970
  }
1932
1971
  },
1933
- "langevals/query_resolution": {
1934
- name: `Query Resolution`,
1935
- description: `
1936
- This evaluator checks if all the user queries in the conversation were resolved. Useful to detect when the bot doesn't know how to answer or can't help the user.
1937
- `,
1938
- category: "quality",
1939
- docsUrl: "",
1940
- isGuardrail: false,
1941
- requiredFields: ["conversation"],
1942
- optionalFields: [],
1943
- settings: {
1944
- model: {
1945
- description: "The model to use for evaluation",
1946
- default: "openai/gpt-5-mini"
1947
- },
1948
- max_tokens: {
1949
- description: "Max tokens allowed for evaluation",
1950
- default: 128e3
1951
- }
1952
- },
1953
- envVars: [],
1954
- result: {}
1955
- },
1956
1972
  "langevals/sentiment": {
1957
1973
  name: `Sentiment Evaluator`,
1958
1974
  description: `
@@ -2172,4 +2188,4 @@ function extractFirstLine(description) {
2172
2188
  export {
2173
2189
  formatEvaluatorSchema
2174
2190
  };
2175
- //# sourceMappingURL=discover-evaluator-schema-MRGHVNTT.js.map
2191
+ //# sourceMappingURL=discover-evaluator-schema-XTAD4X7D.js.map
@@ -1,4 +1,4 @@
1
- import "./chunk-BJ7B4R26.js";
1
+ import "./chunk-SAS5OB7F.js";
2
2
 
3
3
  // src/tools/discover-scenario-schema.ts
4
4
  function formatScenarioSchema() {
@@ -64,4 +64,4 @@ function formatScenarioSchema() {
64
64
  export {
65
65
  formatScenarioSchema
66
66
  };
67
- //# sourceMappingURL=discover-scenario-schema-UPBTCWPG.js.map
67
+ //# sourceMappingURL=discover-scenario-schema-FE64QK2D.js.map
@@ -1,4 +1,4 @@
1
- import "./chunk-BJ7B4R26.js";
1
+ import "./chunk-SAS5OB7F.js";
2
2
 
3
3
  // src/schemas/filter-fields.ts
4
4
  var filterFields = [
@@ -445,4 +445,4 @@ function formatGroups() {
445
445
  export {
446
446
  formatSchema
447
447
  };
448
- //# sourceMappingURL=discover-schema-ICGXPHUW.js.map
448
+ //# sourceMappingURL=discover-schema-L6QDPON3.js.map
@@ -1,9 +1,9 @@
1
1
  import {
2
2
  getAgent
3
- } from "./chunk-CE6QMZOP.js";
4
- import "./chunk-5UWY24RP.js";
5
- import "./chunk-KOO5733S.js";
6
- import "./chunk-BJ7B4R26.js";
3
+ } from "./chunk-TOXJQDIZ.js";
4
+ import "./chunk-A2LEVYQ5.js";
5
+ import "./chunk-WT4PO4A3.js";
6
+ import "./chunk-SAS5OB7F.js";
7
7
 
8
8
  // src/tools/get-agent.ts
9
9
  async function handleGetAgent(params) {
@@ -26,4 +26,4 @@ async function handleGetAgent(params) {
26
26
  export {
27
27
  handleGetAgent
28
28
  };
29
- //# sourceMappingURL=get-agent-7QJVPH72.js.map
29
+ //# sourceMappingURL=get-agent-Y7MWZ65F.js.map
@@ -3,9 +3,9 @@ import {
3
3
  } from "./chunk-ZXKLPC2E.js";
4
4
  import {
5
5
  getAnalyticsTimeseries
6
- } from "./chunk-5UWY24RP.js";
7
- import "./chunk-KOO5733S.js";
8
- import "./chunk-BJ7B4R26.js";
6
+ } from "./chunk-A2LEVYQ5.js";
7
+ import "./chunk-WT4PO4A3.js";
8
+ import "./chunk-SAS5OB7F.js";
9
9
 
10
10
  // src/tools/get-analytics.ts
11
11
  function getGroupedData(bucket, groupBy) {
@@ -70,4 +70,4 @@ async function handleGetAnalytics(params) {
70
70
  export {
71
71
  handleGetAnalytics
72
72
  };
73
- //# sourceMappingURL=get-analytics-PVERWOA5.js.map
73
+ //# sourceMappingURL=get-analytics-TETCSOK6.js.map
@@ -1,9 +1,9 @@
1
1
  import {
2
2
  getDashboard
3
- } from "./chunk-UXXA4D6K.js";
4
- import "./chunk-5UWY24RP.js";
5
- import "./chunk-KOO5733S.js";
6
- import "./chunk-BJ7B4R26.js";
3
+ } from "./chunk-AQJO2EI5.js";
4
+ import "./chunk-A2LEVYQ5.js";
5
+ import "./chunk-WT4PO4A3.js";
6
+ import "./chunk-SAS5OB7F.js";
7
7
 
8
8
  // src/tools/get-dashboard.ts
9
9
  async function handleGetDashboard(params) {
@@ -27,4 +27,4 @@ async function handleGetDashboard(params) {
27
27
  export {
28
28
  handleGetDashboard
29
29
  };
30
- //# sourceMappingURL=get-dashboard-ZZPC3WMD.js.map
30
+ //# sourceMappingURL=get-dashboard-KHBEGZGL.js.map
@@ -3,10 +3,10 @@ import {
3
3
  } from "./chunk-NUFPTHKU.js";
4
4
  import {
5
5
  getDataset
6
- } from "./chunk-J45SVWPR.js";
7
- import "./chunk-5UWY24RP.js";
8
- import "./chunk-KOO5733S.js";
9
- import "./chunk-BJ7B4R26.js";
6
+ } from "./chunk-77BNWQ3U.js";
7
+ import "./chunk-A2LEVYQ5.js";
8
+ import "./chunk-WT4PO4A3.js";
9
+ import "./chunk-SAS5OB7F.js";
10
10
 
11
11
  // src/tools/get-dataset.ts
12
12
  function formatDatasetResponse(dataset) {
@@ -46,4 +46,4 @@ export {
46
46
  formatDatasetResponse,
47
47
  handleGetDataset
48
48
  };
49
- //# sourceMappingURL=get-dataset-25TB2LQH.js.map
49
+ //# sourceMappingURL=get-dataset-RCMU5VAH.js.map
@@ -1,10 +1,10 @@
1
1
  import {
2
2
  getEvaluator,
3
3
  getEvaluatorType
4
- } from "./chunk-EU22L52K.js";
5
- import "./chunk-5UWY24RP.js";
6
- import "./chunk-KOO5733S.js";
7
- import "./chunk-BJ7B4R26.js";
4
+ } from "./chunk-CDYQOY5U.js";
5
+ import "./chunk-A2LEVYQ5.js";
6
+ import "./chunk-WT4PO4A3.js";
7
+ import "./chunk-SAS5OB7F.js";
8
8
 
9
9
  // src/tools/get-evaluator.ts
10
10
  async function handleGetEvaluator(params) {
@@ -45,4 +45,4 @@ async function handleGetEvaluator(params) {
45
45
  export {
46
46
  handleGetEvaluator
47
47
  };
48
- //# sourceMappingURL=get-evaluator-3FTRQKEC.js.map
48
+ //# sourceMappingURL=get-evaluator-ADMKVWYZ.js.map
@@ -1,9 +1,9 @@
1
1
  import {
2
2
  getProject
3
- } from "./chunk-XY4STOBE.js";
4
- import "./chunk-5UWY24RP.js";
5
- import "./chunk-KOO5733S.js";
6
- import "./chunk-BJ7B4R26.js";
3
+ } from "./chunk-7RHG56OH.js";
4
+ import "./chunk-A2LEVYQ5.js";
5
+ import "./chunk-WT4PO4A3.js";
6
+ import "./chunk-SAS5OB7F.js";
7
7
 
8
8
  // src/tools/get-project.ts
9
9
  async function handleGetProject(params) {
@@ -24,4 +24,4 @@ async function handleGetProject(params) {
24
24
  export {
25
25
  handleGetProject
26
26
  };
27
- //# sourceMappingURL=get-project-5OAXYC57.js.map
27
+ //# sourceMappingURL=get-project-RAK7URQJ.js.map
@@ -1,8 +1,8 @@
1
1
  import {
2
2
  getPrompt
3
- } from "./chunk-5UWY24RP.js";
4
- import "./chunk-KOO5733S.js";
5
- import "./chunk-BJ7B4R26.js";
3
+ } from "./chunk-A2LEVYQ5.js";
4
+ import "./chunk-WT4PO4A3.js";
5
+ import "./chunk-SAS5OB7F.js";
6
6
 
7
7
  // src/tools/get-prompt.ts
8
8
  async function handleGetPrompt(params) {
@@ -43,4 +43,4 @@ async function handleGetPrompt(params) {
43
43
  export {
44
44
  handleGetPrompt
45
45
  };
46
- //# sourceMappingURL=get-prompt-I753LZKZ.js.map
46
+ //# sourceMappingURL=get-prompt-OY5B7TRI.js.map
@@ -1,9 +1,9 @@
1
1
  import {
2
2
  getScenario
3
- } from "./chunk-O3LWVN67.js";
4
- import "./chunk-5UWY24RP.js";
5
- import "./chunk-KOO5733S.js";
6
- import "./chunk-BJ7B4R26.js";
3
+ } from "./chunk-J42FWRJ2.js";
4
+ import "./chunk-A2LEVYQ5.js";
5
+ import "./chunk-WT4PO4A3.js";
6
+ import "./chunk-SAS5OB7F.js";
7
7
 
8
8
  // src/tools/get-scenario.ts
9
9
  async function handleGetScenario(params) {
@@ -31,4 +31,4 @@ async function handleGetScenario(params) {
31
31
  export {
32
32
  handleGetScenario
33
33
  };
34
- //# sourceMappingURL=get-scenario-DSSTR6FI.js.map
34
+ //# sourceMappingURL=get-scenario-ZTAJJL5T.js.map
@@ -1,9 +1,9 @@
1
1
  import {
2
2
  getSimulationRun
3
- } from "./chunk-Y3HDZ6OW.js";
4
- import "./chunk-5UWY24RP.js";
5
- import "./chunk-KOO5733S.js";
6
- import "./chunk-BJ7B4R26.js";
3
+ } from "./chunk-ZWIL4EFL.js";
4
+ import "./chunk-A2LEVYQ5.js";
5
+ import "./chunk-WT4PO4A3.js";
6
+ import "./chunk-SAS5OB7F.js";
7
7
 
8
8
  // src/tools/get-simulation-run.ts
9
9
  async function handleGetSimulationRun(params) {
@@ -54,4 +54,4 @@ async function handleGetSimulationRun(params) {
54
54
  export {
55
55
  handleGetSimulationRun
56
56
  };
57
- //# sourceMappingURL=get-simulation-run-RQPMS2MC.js.map
57
+ //# sourceMappingURL=get-simulation-run-T2ZQWYH2.js.map