self-bench 0.3.6 → 0.3.8
This diff represents the content of publicly available package versions that have been released to one of the supported registries. The information contained in this diff is provided for informational purposes only and reflects changes between package versions as they appear in their respective public registries.
- package/.dockerignore +11 -1
- package/Dockerfile +9 -9
- package/Dockerfile.sandbox +5 -4
- package/README.md +27 -12
- package/bun.lock +462 -95
- package/compose.yaml +64 -20
- package/dist/agent-feed.d.ts +8 -0
- package/dist/agent-feed.d.ts.map +1 -0
- package/dist/agent-feed.js +56 -0
- package/dist/agent-feed.js.map +1 -0
- package/dist/api/http.d.ts +18 -0
- package/dist/api/http.d.ts.map +1 -0
- package/dist/api/http.js +111 -0
- package/dist/api/http.js.map +1 -0
- package/dist/api/run-request.d.ts +37 -0
- package/dist/api/run-request.d.ts.map +1 -0
- package/dist/api/run-request.js +59 -0
- package/dist/api/run-request.js.map +1 -0
- package/dist/api/status.d.ts +4 -0
- package/dist/api/status.d.ts.map +1 -0
- package/dist/api/status.js +36 -0
- package/dist/api/status.js.map +1 -0
- package/dist/api/viewer-routes.d.ts +9 -0
- package/dist/api/viewer-routes.d.ts.map +1 -0
- package/dist/api/viewer-routes.js +98 -0
- package/dist/api/viewer-routes.js.map +1 -0
- package/dist/api-main.js +5 -1
- package/dist/api-main.js.map +1 -1
- package/dist/api.d.ts +7 -25
- package/dist/api.d.ts.map +1 -1
- package/dist/api.js +80 -148
- package/dist/api.js.map +1 -1
- package/dist/archive.d.ts +14 -0
- package/dist/archive.d.ts.map +1 -0
- package/dist/archive.js +114 -0
- package/dist/archive.js.map +1 -0
- package/dist/artifacts/common.d.ts +15 -0
- package/dist/artifacts/common.d.ts.map +1 -0
- package/dist/artifacts/common.js +58 -0
- package/dist/artifacts/common.js.map +1 -0
- package/dist/artifacts/gcs.d.ts +18 -0
- package/dist/artifacts/gcs.d.ts.map +1 -0
- package/dist/artifacts/gcs.js +178 -0
- package/dist/artifacts/gcs.js.map +1 -0
- package/dist/artifacts/local.d.ts +17 -0
- package/dist/artifacts/local.d.ts.map +1 -0
- package/dist/artifacts/local.js +205 -0
- package/dist/artifacts/local.js.map +1 -0
- package/dist/artifacts/types.d.ts +18 -0
- package/dist/artifacts/types.d.ts.map +1 -0
- package/dist/artifacts/types.js +2 -0
- package/dist/artifacts/types.js.map +1 -0
- package/dist/artifacts.d.ts +5 -28
- package/dist/artifacts.d.ts.map +1 -1
- package/dist/artifacts.js +5 -302
- package/dist/artifacts.js.map +1 -1
- package/dist/audit.d.ts.map +1 -1
- package/dist/audit.js +8 -2
- package/dist/audit.js.map +1 -1
- package/dist/auth/api-key-routes.d.ts +17 -0
- package/dist/auth/api-key-routes.d.ts.map +1 -0
- package/dist/auth/api-key-routes.js +57 -0
- package/dist/auth/api-key-routes.js.map +1 -0
- package/dist/auth/api-keys.d.ts +47 -0
- package/dist/auth/api-keys.d.ts.map +1 -0
- package/dist/auth/api-keys.js +102 -0
- package/dist/auth/api-keys.js.map +1 -0
- package/dist/auth/config.d.ts +17 -0
- package/dist/auth/config.d.ts.map +1 -0
- package/dist/auth/config.js +38 -0
- package/dist/auth/config.js.map +1 -0
- package/dist/auth/cookies.d.ts +12 -0
- package/dist/auth/cookies.d.ts.map +1 -0
- package/dist/auth/cookies.js +45 -0
- package/dist/auth/cookies.js.map +1 -0
- package/dist/auth/crypto.d.ts +19 -0
- package/dist/auth/crypto.d.ts.map +1 -0
- package/dist/auth/crypto.js +48 -0
- package/dist/auth/crypto.js.map +1 -0
- package/dist/auth/github.d.ts +37 -0
- package/dist/auth/github.d.ts.map +1 -0
- package/dist/auth/github.js +127 -0
- package/dist/auth/github.js.map +1 -0
- package/dist/auth/routes.d.ts +41 -0
- package/dist/auth/routes.d.ts.map +1 -0
- package/dist/auth/routes.js +153 -0
- package/dist/auth/routes.js.map +1 -0
- package/dist/auth/session-expired.d.ts +4 -0
- package/dist/auth/session-expired.d.ts.map +1 -0
- package/dist/auth/session-expired.js +17 -0
- package/dist/auth/session-expired.js.map +1 -0
- package/dist/auth/session.d.ts +23 -0
- package/dist/auth/session.d.ts.map +1 -0
- package/dist/auth/session.js +56 -0
- package/dist/auth/session.js.map +1 -0
- package/dist/auth/users.d.ts +49 -0
- package/dist/auth/users.d.ts.map +1 -0
- package/dist/auth/users.js +125 -0
- package/dist/auth/users.js.map +1 -0
- package/dist/build-metadata.js +1 -1
- package/dist/cli/api-client.d.ts +8 -0
- package/dist/cli/api-client.d.ts.map +1 -0
- package/dist/cli/api-client.js +71 -0
- package/dist/cli/api-client.js.map +1 -0
- package/dist/cli/associate.d.ts +2 -0
- package/dist/cli/associate.d.ts.map +1 -0
- package/dist/cli/associate.js +51 -0
- package/dist/cli/associate.js.map +1 -0
- package/dist/cli/help.d.ts +2 -0
- package/dist/cli/help.d.ts.map +1 -0
- package/dist/cli/help.js +52 -0
- package/dist/cli/help.js.map +1 -0
- package/dist/cli/main.d.ts +2 -0
- package/dist/cli/main.d.ts.map +1 -0
- package/dist/cli/main.js +55 -0
- package/dist/cli/main.js.map +1 -0
- package/dist/cli/replay.d.ts +3 -0
- package/dist/cli/replay.d.ts.map +1 -0
- package/dist/cli/replay.js +59 -0
- package/dist/cli/replay.js.map +1 -0
- package/dist/cli/repository.d.ts +6 -0
- package/dist/cli/repository.d.ts.map +1 -0
- package/dist/cli/repository.js +36 -0
- package/dist/cli/repository.js.map +1 -0
- package/dist/cli/run.d.ts +2 -0
- package/dist/cli/run.d.ts.map +1 -0
- package/dist/cli/run.js +93 -0
- package/dist/cli/run.js.map +1 -0
- package/dist/cli/stack-environment.d.ts +26 -0
- package/dist/cli/stack-environment.d.ts.map +1 -0
- package/dist/cli/stack-environment.js +86 -0
- package/dist/cli/stack-environment.js.map +1 -0
- package/dist/cli/stack.d.ts +6 -0
- package/dist/cli/stack.d.ts.map +1 -0
- package/dist/cli/stack.js +131 -0
- package/dist/cli/stack.js.map +1 -0
- package/dist/cli/values.d.ts +8 -0
- package/dist/cli/values.d.ts.map +1 -0
- package/dist/cli/values.js +30 -0
- package/dist/cli/values.js.map +1 -0
- package/dist/cli/view.d.ts +2 -0
- package/dist/cli/view.d.ts.map +1 -0
- package/dist/cli/view.js +40 -0
- package/dist/cli/view.js.map +1 -0
- package/dist/cli.js +2 -429
- package/dist/cli.js.map +1 -1
- package/dist/config.d.ts +16 -2
- package/dist/config.d.ts.map +1 -1
- package/dist/config.js +60 -22
- package/dist/config.js.map +1 -1
- package/dist/contracts/common.d.ts +22 -0
- package/dist/contracts/common.d.ts.map +1 -0
- package/dist/contracts/common.js +29 -0
- package/dist/contracts/common.js.map +1 -0
- package/dist/contracts/results.d.ts +66 -0
- package/dist/contracts/results.d.ts.map +1 -0
- package/dist/contracts/results.js +2 -0
- package/dist/contracts/results.js.map +1 -0
- package/dist/contracts/run.d.ts +145 -0
- package/dist/contracts/run.d.ts.map +1 -0
- package/dist/contracts/run.js +74 -0
- package/dist/contracts/run.js.map +1 -0
- package/dist/contracts/task.d.ts +232 -0
- package/dist/contracts/task.d.ts.map +1 -0
- package/dist/contracts/task.js +198 -0
- package/dist/contracts/task.js.map +1 -0
- package/dist/contracts/verify.d.ts +331 -0
- package/dist/contracts/verify.d.ts.map +1 -0
- package/dist/contracts/verify.js +84 -0
- package/dist/contracts/verify.js.map +1 -0
- package/dist/contracts.d.ts +5 -243
- package/dist/contracts.d.ts.map +1 -1
- package/dist/contracts.js +5 -128
- package/dist/contracts.js.map +1 -1
- package/dist/db/client.d.ts +12 -0
- package/dist/db/client.d.ts.map +1 -0
- package/dist/db/client.js +16 -0
- package/dist/db/client.js.map +1 -0
- package/dist/db/schema.d.ts +1256 -0
- package/dist/db/schema.d.ts.map +1 -0
- package/dist/db/schema.js +128 -0
- package/dist/db/schema.js.map +1 -0
- package/dist/environment.d.ts +9 -0
- package/dist/environment.d.ts.map +1 -0
- package/dist/environment.js +101 -0
- package/dist/environment.js.map +1 -0
- package/dist/evaluation/account.d.ts +30 -0
- package/dist/evaluation/account.d.ts.map +1 -0
- package/dist/evaluation/account.js +34 -0
- package/dist/evaluation/account.js.map +1 -0
- package/dist/evaluation/activities.d.ts +9 -0
- package/dist/evaluation/activities.d.ts.map +1 -0
- package/dist/evaluation/activities.js +40 -0
- package/dist/evaluation/activities.js.map +1 -0
- package/dist/evaluation/catalog-pricing.d.ts +3 -0
- package/dist/evaluation/catalog-pricing.d.ts.map +1 -0
- package/dist/evaluation/catalog-pricing.js +37 -0
- package/dist/evaluation/catalog-pricing.js.map +1 -0
- package/dist/evaluation/catalog.d.ts +17 -0
- package/dist/evaluation/catalog.d.ts.map +1 -0
- package/dist/evaluation/catalog.js +126 -0
- package/dist/evaluation/catalog.js.map +1 -0
- package/dist/evaluation/codex-login-process.d.ts +12 -0
- package/dist/evaluation/codex-login-process.d.ts.map +1 -0
- package/dist/evaluation/codex-login-process.js +121 -0
- package/dist/evaluation/codex-login-process.js.map +1 -0
- package/dist/evaluation/codex-login-routes.d.ts +5 -0
- package/dist/evaluation/codex-login-routes.d.ts.map +1 -0
- package/dist/evaluation/codex-login-routes.js +55 -0
- package/dist/evaluation/codex-login-routes.js.map +1 -0
- package/dist/evaluation/codex-login.d.ts +42 -0
- package/dist/evaluation/codex-login.d.ts.map +1 -0
- package/dist/evaluation/codex-login.js +143 -0
- package/dist/evaluation/codex-login.js.map +1 -0
- package/dist/evaluation/comparisons.d.ts +65 -0
- package/dist/evaluation/comparisons.d.ts.map +1 -0
- package/dist/evaluation/comparisons.js +161 -0
- package/dist/evaluation/comparisons.js.map +1 -0
- package/dist/evaluation/config.d.ts +43 -0
- package/dist/evaluation/config.d.ts.map +1 -0
- package/dist/evaluation/config.js +116 -0
- package/dist/evaluation/config.js.map +1 -0
- package/dist/evaluation/cost.d.ts +3 -0
- package/dist/evaluation/cost.d.ts.map +1 -0
- package/dist/evaluation/cost.js +95 -0
- package/dist/evaluation/cost.js.map +1 -0
- package/dist/evaluation/credential-execution.d.ts +10 -0
- package/dist/evaluation/credential-execution.d.ts.map +1 -0
- package/dist/evaluation/credential-execution.js +65 -0
- package/dist/evaluation/credential-execution.js.map +1 -0
- package/dist/evaluation/credentials.d.ts +40 -0
- package/dist/evaluation/credentials.d.ts.map +1 -0
- package/dist/evaluation/credentials.js +146 -0
- package/dist/evaluation/credentials.js.map +1 -0
- package/dist/evaluation/encrypted-records.d.ts +15 -0
- package/dist/evaluation/encrypted-records.d.ts.map +1 -0
- package/dist/evaluation/encrypted-records.js +57 -0
- package/dist/evaluation/encrypted-records.js.map +1 -0
- package/dist/evaluation/gateway-execution.d.ts +18 -0
- package/dist/evaluation/gateway-execution.d.ts.map +1 -0
- package/dist/evaluation/gateway-execution.js +38 -0
- package/dist/evaluation/gateway-execution.js.map +1 -0
- package/dist/evaluation/harbor-cost.d.ts +3 -0
- package/dist/evaluation/harbor-cost.d.ts.map +1 -0
- package/dist/evaluation/harbor-cost.js +39 -0
- package/dist/evaluation/harbor-cost.js.map +1 -0
- package/dist/evaluation/harbor_e2b.py +20 -0
- package/dist/evaluation/harbor_gateway.py +22 -0
- package/dist/evaluation/harnesses.d.ts +8 -0
- package/dist/evaluation/harnesses.d.ts.map +1 -0
- package/dist/evaluation/harnesses.js +10 -0
- package/dist/evaluation/harnesses.js.map +1 -0
- package/dist/evaluation/model-options.d.ts +9 -0
- package/dist/evaluation/model-options.d.ts.map +1 -0
- package/dist/evaluation/model-options.js +81 -0
- package/dist/evaluation/model-options.js.map +1 -0
- package/dist/evaluation/org-credential-routes.d.ts +5 -0
- package/dist/evaluation/org-credential-routes.d.ts.map +1 -0
- package/dist/evaluation/org-credential-routes.js +56 -0
- package/dist/evaluation/org-credential-routes.js.map +1 -0
- package/dist/evaluation/org-records.d.ts +4 -0
- package/dist/evaluation/org-records.d.ts.map +1 -0
- package/dist/evaluation/org-records.js +14 -0
- package/dist/evaluation/org-records.js.map +1 -0
- package/dist/evaluation/output.d.ts +9 -0
- package/dist/evaluation/output.d.ts.map +1 -0
- package/dist/evaluation/output.js +171 -0
- package/dist/evaluation/output.js.map +1 -0
- package/dist/evaluation/platform-routes.d.ts +5 -0
- package/dist/evaluation/platform-routes.d.ts.map +1 -0
- package/dist/evaluation/platform-routes.js +123 -0
- package/dist/evaluation/platform-routes.js.map +1 -0
- package/dist/evaluation/profile-routes.d.ts +16 -0
- package/dist/evaluation/profile-routes.d.ts.map +1 -0
- package/dist/evaluation/profile-routes.js +47 -0
- package/dist/evaluation/profile-routes.js.map +1 -0
- package/dist/evaluation/profiles.d.ts +48 -0
- package/dist/evaluation/profiles.d.ts.map +1 -0
- package/dist/evaluation/profiles.js +134 -0
- package/dist/evaluation/profiles.js.map +1 -0
- package/dist/evaluation/providers.d.ts +11 -0
- package/dist/evaluation/providers.d.ts.map +1 -0
- package/dist/evaluation/providers.js +30 -0
- package/dist/evaluation/providers.js.map +1 -0
- package/dist/evaluation/routes.d.ts +23 -0
- package/dist/evaluation/routes.d.ts.map +1 -0
- package/dist/evaluation/routes.js +212 -0
- package/dist/evaluation/routes.js.map +1 -0
- package/dist/evaluation/runner.d.ts +16 -0
- package/dist/evaluation/runner.d.ts.map +1 -0
- package/dist/evaluation/runner.js +241 -0
- package/dist/evaluation/runner.js.map +1 -0
- package/dist/evaluation/store.d.ts +8 -0
- package/dist/evaluation/store.d.ts.map +1 -0
- package/dist/evaluation/store.js +60 -0
- package/dist/evaluation/store.js.map +1 -0
- package/dist/evaluation/types.d.ts +98 -0
- package/dist/evaluation/types.d.ts.map +1 -0
- package/dist/evaluation/types.js +2 -0
- package/dist/evaluation/types.js.map +1 -0
- package/dist/evaluation/worker-main.d.ts +2 -0
- package/dist/evaluation/worker-main.d.ts.map +1 -0
- package/dist/evaluation/worker-main.js +26 -0
- package/dist/evaluation/worker-main.js.map +1 -0
- package/dist/evaluation/worker-records.d.ts +5 -0
- package/dist/evaluation/worker-records.d.ts.map +1 -0
- package/dist/evaluation/worker-records.js +11 -0
- package/dist/evaluation/worker-records.js.map +1 -0
- package/dist/evaluation/workflow.d.ts +3 -0
- package/dist/evaluation/workflow.d.ts.map +1 -0
- package/dist/evaluation/workflow.js +20 -0
- package/dist/evaluation/workflow.js.map +1 -0
- package/dist/execution-environment.d.ts +3 -0
- package/dist/execution-environment.d.ts.map +1 -0
- package/dist/execution-environment.js +9 -0
- package/dist/execution-environment.js.map +1 -0
- package/dist/extension-authoring.bundle.js +347 -0
- package/dist/extension-verifier.bundle.js +73 -0
- package/dist/extensions/authoring.d.ts.map +1 -1
- package/dist/extensions/authoring.js +70 -59
- package/dist/extensions/authoring.js.map +1 -1
- package/dist/extensions/shared/deliverable.d.ts +33 -0
- package/dist/extensions/shared/deliverable.d.ts.map +1 -0
- package/dist/extensions/shared/deliverable.js +123 -0
- package/dist/extensions/shared/deliverable.js.map +1 -0
- package/dist/extensions/shared/mailbox.d.ts +33 -0
- package/dist/extensions/shared/mailbox.d.ts.map +1 -0
- package/dist/extensions/shared/mailbox.js +102 -0
- package/dist/extensions/shared/mailbox.js.map +1 -0
- package/dist/extensions/shared/schemas.d.ts +90 -0
- package/dist/extensions/shared/schemas.d.ts.map +1 -0
- package/dist/extensions/shared/schemas.js +71 -0
- package/dist/extensions/shared/schemas.js.map +1 -0
- package/dist/extensions/shared/static-check.d.ts +38 -0
- package/dist/extensions/shared/static-check.d.ts.map +1 -0
- package/dist/extensions/shared/static-check.js +52 -0
- package/dist/extensions/shared/static-check.js.map +1 -0
- package/dist/extensions/verifier.d.ts +4 -0
- package/dist/extensions/verifier.d.ts.map +1 -0
- package/dist/extensions/verifier.js +70 -0
- package/dist/extensions/verifier.js.map +1 -0
- package/dist/github.d.ts +1 -1
- package/dist/github.d.ts.map +1 -1
- package/dist/github.js +3 -3
- package/dist/github.js.map +1 -1
- package/dist/harbor-environment.d.ts +13 -1
- package/dist/harbor-environment.d.ts.map +1 -1
- package/dist/harbor-environment.js +37 -4
- package/dist/harbor-environment.js.map +1 -1
- package/dist/harbor-results.d.ts +2 -0
- package/dist/harbor-results.d.ts.map +1 -1
- package/dist/harbor-results.js +2 -0
- package/dist/harbor-results.js.map +1 -1
- package/dist/harbor-task/compiler.d.ts +10 -0
- package/dist/harbor-task/compiler.d.ts.map +1 -0
- package/dist/harbor-task/compiler.js +174 -0
- package/dist/harbor-task/compiler.js.map +1 -0
- package/dist/harbor-task/constants.d.ts +3 -0
- package/dist/harbor-task/constants.d.ts.map +1 -0
- package/dist/harbor-task/constants.js +3 -0
- package/dist/harbor-task/constants.js.map +1 -0
- package/dist/harbor-task/dependencies.d.ts +3 -0
- package/dist/harbor-task/dependencies.d.ts.map +1 -0
- package/dist/harbor-task/dependencies.js +20 -0
- package/dist/harbor-task/dependencies.js.map +1 -0
- package/dist/harbor-task/paths.d.ts +6 -0
- package/dist/harbor-task/paths.d.ts.map +1 -0
- package/dist/harbor-task/paths.js +37 -0
- package/dist/harbor-task/paths.js.map +1 -0
- package/dist/harbor-task/render.d.ts +11 -0
- package/dist/harbor-task/render.d.ts.map +1 -0
- package/dist/harbor-task/render.js +194 -0
- package/dist/harbor-task/render.js.map +1 -0
- package/dist/harbor-task/runtime/command.sh +26 -0
- package/dist/harbor-task/runtime/junit.py +69 -0
- package/dist/harbor-task/runtime-assets.d.ts +2 -0
- package/dist/harbor-task/runtime-assets.d.ts.map +1 -0
- package/dist/harbor-task/runtime-assets.js +8 -0
- package/dist/harbor-task/runtime-assets.js.map +1 -0
- package/dist/harbor-task/test-patch.d.ts +4 -0
- package/dist/harbor-task/test-patch.d.ts.map +1 -0
- package/dist/harbor-task/test-patch.js +15 -0
- package/dist/harbor-task/test-patch.js.map +1 -0
- package/dist/harbor-task/verifier.d.ts +4 -0
- package/dist/harbor-task/verifier.d.ts.map +1 -0
- package/dist/harbor-task/verifier.js +105 -0
- package/dist/harbor-task/verifier.js.map +1 -0
- package/dist/harbor-task.d.ts +3 -11
- package/dist/harbor-task.d.ts.map +1 -1
- package/dist/harbor-task.js +2 -463
- package/dist/harbor-task.js.map +1 -1
- package/dist/log-excerpt.d.ts +15 -0
- package/dist/log-excerpt.d.ts.map +1 -0
- package/dist/log-excerpt.js +104 -0
- package/dist/log-excerpt.js.map +1 -0
- package/dist/patch-check.d.ts +26 -0
- package/dist/patch-check.d.ts.map +1 -0
- package/dist/patch-check.js +137 -0
- package/dist/patch-check.js.map +1 -0
- package/dist/pi-event-feed.d.ts +16 -0
- package/dist/pi-event-feed.d.ts.map +1 -0
- package/dist/pi-event-feed.js +86 -0
- package/dist/pi-event-feed.js.map +1 -0
- package/dist/pi-session.d.ts +32 -0
- package/dist/pi-session.d.ts.map +1 -0
- package/dist/pi-session.js +175 -0
- package/dist/pi-session.js.map +1 -0
- package/dist/process.d.ts +5 -0
- package/dist/process.d.ts.map +1 -1
- package/dist/process.js +31 -9
- package/dist/process.js.map +1 -1
- package/dist/provenance/github.d.ts +4 -0
- package/dist/provenance/github.d.ts.map +1 -0
- package/dist/provenance/github.js +65 -0
- package/dist/provenance/github.js.map +1 -0
- package/dist/provenance/local.d.ts +4 -0
- package/dist/provenance/local.d.ts.map +1 -0
- package/dist/provenance/local.js +82 -0
- package/dist/provenance/local.js.map +1 -0
- package/dist/provenance/redact.d.ts +2 -0
- package/dist/provenance/redact.d.ts.map +1 -0
- package/dist/provenance/redact.js +28 -0
- package/dist/provenance/redact.js.map +1 -0
- package/dist/provenance/selection.d.ts +4 -0
- package/dist/provenance/selection.d.ts.map +1 -0
- package/dist/provenance/selection.js +29 -0
- package/dist/provenance/selection.js.map +1 -0
- package/dist/provenance/session.d.ts +3 -0
- package/dist/provenance/session.d.ts.map +1 -0
- package/dist/provenance/session.js +214 -0
- package/dist/provenance/session.js.map +1 -0
- package/dist/provenance/shared.d.ts +7 -0
- package/dist/provenance/shared.d.ts.map +1 -0
- package/dist/provenance/shared.js +10 -0
- package/dist/provenance/shared.js.map +1 -0
- package/dist/provenance/types.d.ts +34 -0
- package/dist/provenance/types.d.ts.map +1 -0
- package/dist/provenance/types.js +22 -0
- package/dist/provenance/types.js.map +1 -0
- package/dist/provenance-associations/apply.d.ts +3 -0
- package/dist/provenance-associations/apply.d.ts.map +1 -0
- package/dist/provenance-associations/apply.js +66 -0
- package/dist/provenance-associations/apply.js.map +1 -0
- package/dist/provenance-associations/shared.d.ts +45 -0
- package/dist/provenance-associations/shared.d.ts.map +1 -0
- package/dist/provenance-associations/shared.js +42 -0
- package/dist/provenance-associations/shared.js.map +1 -0
- package/dist/provenance-associations.d.ts +3 -30
- package/dist/provenance-associations.d.ts.map +1 -1
- package/dist/provenance-associations.js +4 -103
- package/dist/provenance-associations.js.map +1 -1
- package/dist/provenance.d.ts +7 -41
- package/dist/provenance.d.ts.map +1 -1
- package/dist/provenance.js +6 -432
- package/dist/provenance.js.map +1 -1
- package/dist/providers.d.ts +25 -2
- package/dist/providers.d.ts.map +1 -1
- package/dist/providers.js +23 -2
- package/dist/providers.js.map +1 -1
- package/dist/repair.d.ts +1 -6
- package/dist/repair.d.ts.map +1 -1
- package/dist/repair.js +17 -27
- package/dist/repair.js.map +1 -1
- package/dist/review/assets/index-BHHztQ9x.js +1679 -0
- package/dist/review/assets/index-KQ1eFQ0t.css +1 -0
- package/dist/review/index.html +11 -8
- package/dist/runtime/command.sh +26 -0
- package/dist/runtime/junit.py +69 -0
- package/dist/sandbox/contracts.d.ts +36 -1
- package/dist/sandbox/contracts.d.ts.map +1 -1
- package/dist/sandbox/contracts.js +3 -0
- package/dist/sandbox/contracts.js.map +1 -1
- package/dist/sandbox/index.d.ts +1 -1
- package/dist/sandbox/index.d.ts.map +1 -1
- package/dist/sandbox/index.js +13 -2
- package/dist/sandbox/index.js.map +1 -1
- package/dist/sandbox/live.d.ts +18 -0
- package/dist/sandbox/live.d.ts.map +1 -0
- package/dist/sandbox/live.js +55 -0
- package/dist/sandbox/live.js.map +1 -0
- package/dist/sandbox/output-retry.d.ts +16 -0
- package/dist/sandbox/output-retry.d.ts.map +1 -0
- package/dist/sandbox/output-retry.js +36 -0
- package/dist/sandbox/output-retry.js.map +1 -0
- package/dist/sandbox/programs/author.js +16 -6
- package/dist/sandbox/programs/author.js.map +1 -1
- package/dist/sandbox/programs/check.d.ts +3 -0
- package/dist/sandbox/programs/check.d.ts.map +1 -0
- package/dist/sandbox/programs/check.js +76 -0
- package/dist/sandbox/programs/check.js.map +1 -0
- package/dist/sandbox/programs/prepare-task.d.ts +12 -0
- package/dist/sandbox/programs/prepare-task.d.ts.map +1 -0
- package/dist/sandbox/programs/{prepare-repair.js → prepare-task.js} +11 -9
- package/dist/sandbox/programs/prepare-task.js.map +1 -0
- package/dist/sandbox/programs/verifier.d.ts +3 -0
- package/dist/sandbox/programs/verifier.d.ts.map +1 -0
- package/dist/sandbox/programs/verifier.js +9 -0
- package/dist/sandbox/programs/verifier.js.map +1 -0
- package/dist/sandbox/providers/docker/executor.d.ts +4 -1
- package/dist/sandbox/providers/docker/executor.d.ts.map +1 -1
- package/dist/sandbox/providers/docker/executor.js +75 -25
- package/dist/sandbox/providers/docker/executor.js.map +1 -1
- package/dist/sandbox/providers/e2b/bytes.d.ts +3 -0
- package/dist/sandbox/providers/e2b/bytes.d.ts.map +1 -0
- package/dist/sandbox/providers/e2b/bytes.js +8 -0
- package/dist/sandbox/providers/e2b/bytes.js.map +1 -0
- package/dist/sandbox/providers/e2b/cleanup.d.ts +8 -0
- package/dist/sandbox/providers/e2b/cleanup.d.ts.map +1 -0
- package/dist/sandbox/providers/e2b/cleanup.js +124 -0
- package/dist/sandbox/providers/e2b/cleanup.js.map +1 -0
- package/dist/sandbox/providers/e2b/command.d.ts +23 -0
- package/dist/sandbox/providers/e2b/command.d.ts.map +1 -0
- package/dist/sandbox/providers/e2b/command.js +160 -0
- package/dist/sandbox/providers/e2b/command.js.map +1 -0
- package/dist/sandbox/providers/e2b/config.d.ts +13 -0
- package/dist/sandbox/providers/e2b/config.d.ts.map +1 -0
- package/dist/sandbox/providers/e2b/config.js +2 -0
- package/dist/sandbox/providers/e2b/config.js.map +1 -0
- package/dist/sandbox/providers/e2b/executor.d.ts +15 -0
- package/dist/sandbox/providers/e2b/executor.d.ts.map +1 -0
- package/dist/sandbox/providers/e2b/executor.js +207 -0
- package/dist/sandbox/providers/e2b/executor.js.map +1 -0
- package/dist/sandbox/providers/e2b/lifecycle.d.ts +8 -0
- package/dist/sandbox/providers/e2b/lifecycle.d.ts.map +1 -0
- package/dist/sandbox/providers/e2b/lifecycle.js +37 -0
- package/dist/sandbox/providers/e2b/lifecycle.js.map +1 -0
- package/dist/sandbox/providers/e2b/outcome.d.ts +11 -0
- package/dist/sandbox/providers/e2b/outcome.d.ts.map +1 -0
- package/dist/sandbox/providers/e2b/outcome.js +72 -0
- package/dist/sandbox/providers/e2b/outcome.js.map +1 -0
- package/dist/sandbox/providers/e2b/stage-files.d.ts +10 -0
- package/dist/sandbox/providers/e2b/stage-files.d.ts.map +1 -0
- package/dist/sandbox/providers/e2b/stage-files.js +45 -0
- package/dist/sandbox/providers/e2b/stage-files.js.map +1 -0
- package/dist/sandbox/providers/e2b/startup.d.ts +10 -0
- package/dist/sandbox/providers/e2b/startup.d.ts.map +1 -0
- package/dist/sandbox/providers/e2b/startup.js +26 -0
- package/dist/sandbox/providers/e2b/startup.js.map +1 -0
- package/dist/sandbox/providers/e2b/types.d.ts +9 -0
- package/dist/sandbox/providers/e2b/types.d.ts.map +1 -0
- package/dist/sandbox/providers/e2b/types.js +2 -0
- package/dist/sandbox/providers/e2b/types.js.map +1 -0
- package/dist/sandbox/providers/e2b/validation.d.ts +14 -0
- package/dist/sandbox/providers/e2b/validation.d.ts.map +1 -0
- package/dist/sandbox/providers/e2b/validation.js +60 -0
- package/dist/sandbox/providers/e2b/validation.js.map +1 -0
- package/dist/{modal-auth.d.ts → sandbox/providers/modal/auth.d.ts} +1 -1
- package/dist/sandbox/providers/modal/auth.d.ts.map +1 -0
- package/dist/{modal-auth.js → sandbox/providers/modal/auth.js} +1 -1
- package/dist/sandbox/providers/modal/auth.js.map +1 -0
- package/dist/sandbox/providers/modal/executor.d.ts +4 -1
- package/dist/sandbox/providers/modal/executor.d.ts.map +1 -1
- package/dist/sandbox/providers/modal/executor.js +33 -9
- package/dist/sandbox/providers/modal/executor.js.map +1 -1
- package/dist/sandbox/providers/modal/live.d.ts +7 -0
- package/dist/sandbox/providers/modal/live.d.ts.map +1 -0
- package/dist/sandbox/providers/modal/live.js +50 -0
- package/dist/sandbox/providers/modal/live.js.map +1 -0
- package/dist/sandbox/providers/vercel/command.d.ts +6 -0
- package/dist/sandbox/providers/vercel/command.d.ts.map +1 -1
- package/dist/sandbox/providers/vercel/command.js +44 -11
- package/dist/sandbox/providers/vercel/command.js.map +1 -1
- package/dist/sandbox/providers/vercel/executor.d.ts +6 -7
- package/dist/sandbox/providers/vercel/executor.d.ts.map +1 -1
- package/dist/sandbox/providers/vercel/executor.js +21 -187
- package/dist/sandbox/providers/vercel/executor.js.map +1 -1
- package/dist/sandbox/providers/vercel/lifecycle.d.ts +13 -0
- package/dist/sandbox/providers/vercel/lifecycle.d.ts.map +1 -0
- package/dist/sandbox/providers/vercel/lifecycle.js +140 -0
- package/dist/sandbox/providers/vercel/lifecycle.js.map +1 -0
- package/dist/sandbox/providers/vercel/validation.d.ts +11 -0
- package/dist/sandbox/providers/vercel/validation.d.ts.map +1 -0
- package/dist/sandbox/providers/vercel/validation.js +30 -0
- package/dist/sandbox/providers/vercel/validation.js.map +1 -0
- package/dist/sandbox/remote-files.d.ts +11 -0
- package/dist/sandbox/remote-files.d.ts.map +1 -0
- package/dist/sandbox/remote-files.js +41 -0
- package/dist/sandbox/remote-files.js.map +1 -0
- package/dist/sandbox/request-validation.d.ts +5 -0
- package/dist/sandbox/request-validation.d.ts.map +1 -0
- package/dist/sandbox/request-validation.js +43 -0
- package/dist/sandbox/request-validation.js.map +1 -0
- package/dist/sandbox/supervisor.d.ts +42 -0
- package/dist/sandbox/supervisor.d.ts.map +1 -0
- package/dist/sandbox/supervisor.js +126 -0
- package/dist/sandbox/supervisor.js.map +1 -0
- package/dist/sandbox/timeout.d.ts +6 -1
- package/dist/sandbox/timeout.d.ts.map +1 -1
- package/dist/sandbox/timeout.js +12 -0
- package/dist/sandbox/timeout.js.map +1 -1
- package/dist/sandbox-author.bundle.js +47 -14879
- package/dist/{sandbox-validation-repair.bundle.js → sandbox-check.bundle.js} +7261 -1893
- package/dist/sandbox-verifier.bundle.js +307 -0
- package/dist/setup/e2b/build.d.ts +24 -0
- package/dist/setup/e2b/build.d.ts.map +1 -0
- package/dist/setup/e2b/build.js +46 -0
- package/dist/setup/e2b/build.js.map +1 -0
- package/dist/setup/e2b/cli.d.ts +2 -0
- package/dist/setup/e2b/cli.d.ts.map +1 -0
- package/dist/setup/e2b/cli.js +51 -0
- package/dist/setup/e2b/cli.js.map +1 -0
- package/dist/setup/e2b/index.d.ts +5 -0
- package/dist/setup/e2b/index.d.ts.map +1 -0
- package/dist/setup/e2b/index.js +4 -0
- package/dist/setup/e2b/index.js.map +1 -0
- package/dist/setup/e2b/template.d.ts +4 -0
- package/dist/setup/e2b/template.d.ts.map +1 -0
- package/dist/setup/e2b/template.js +53 -0
- package/dist/setup/e2b/template.js.map +1 -0
- package/dist/setup/vercel/cli-helpers.d.ts +22 -0
- package/dist/setup/vercel/cli-helpers.d.ts.map +1 -0
- package/dist/setup/vercel/cli-helpers.js +87 -0
- package/dist/setup/vercel/cli-helpers.js.map +1 -0
- package/dist/setup/vercel/cli-runner.d.ts +14 -0
- package/dist/setup/vercel/cli-runner.d.ts.map +1 -0
- package/dist/setup/vercel/cli-runner.js +24 -0
- package/dist/setup/vercel/cli-runner.js.map +1 -0
- package/dist/setup/vercel/cli-types.d.ts +24 -0
- package/dist/setup/vercel/cli-types.d.ts.map +1 -0
- package/dist/setup/vercel/cli-types.js +18 -0
- package/dist/setup/vercel/cli-types.js.map +1 -0
- package/dist/setup/vercel/cli.d.ts +5 -31
- package/dist/setup/vercel/cli.d.ts.map +1 -1
- package/dist/setup/vercel/cli.js +15 -125
- package/dist/setup/vercel/cli.js.map +1 -1
- package/dist/setup/vercel/profile-files.d.ts +6 -0
- package/dist/setup/vercel/profile-files.d.ts.map +1 -0
- package/dist/setup/vercel/profile-files.js +100 -0
- package/dist/setup/vercel/profile-files.js.map +1 -0
- package/dist/setup/vercel/profile-schema.d.ts +48 -0
- package/dist/setup/vercel/profile-schema.d.ts.map +1 -0
- package/dist/setup/vercel/profile-schema.js +39 -0
- package/dist/setup/vercel/profile-schema.js.map +1 -0
- package/dist/setup/vercel/profile.d.ts +3 -19
- package/dist/setup/vercel/profile.d.ts.map +1 -1
- package/dist/setup/vercel/profile.js +4 -130
- package/dist/setup/vercel/profile.js.map +1 -1
- package/dist/setup/vercel/selection.d.ts +18 -0
- package/dist/setup/vercel/selection.d.ts.map +1 -0
- package/dist/setup/vercel/selection.js +113 -0
- package/dist/setup/vercel/selection.js.map +1 -0
- package/dist/setup/vercel/setup-types.d.ts +34 -0
- package/dist/setup/vercel/setup-types.d.ts.map +1 -0
- package/dist/setup/vercel/setup-types.js +2 -0
- package/dist/setup/vercel/setup-types.js.map +1 -0
- package/dist/setup/vercel/setup.d.ts +3 -28
- package/dist/setup/vercel/setup.d.ts.map +1 -1
- package/dist/setup/vercel/setup.js +3 -111
- package/dist/setup/vercel/setup.js.map +1 -1
- package/dist/site/batch-activity.d.ts +5 -0
- package/dist/site/batch-activity.d.ts.map +1 -0
- package/dist/site/batch-activity.js +31 -0
- package/dist/site/batch-activity.js.map +1 -0
- package/dist/site/batch-progress.d.ts +16 -0
- package/dist/site/batch-progress.d.ts.map +1 -0
- package/dist/site/batch-progress.js +52 -0
- package/dist/site/batch-progress.js.map +1 -0
- package/dist/site/batch-routes.d.ts +30 -0
- package/dist/site/batch-routes.d.ts.map +1 -0
- package/dist/site/batch-routes.js +117 -0
- package/dist/site/batch-routes.js.map +1 -0
- package/dist/site/batch-start.d.ts +60 -0
- package/dist/site/batch-start.d.ts.map +1 -0
- package/dist/site/batch-start.js +39 -0
- package/dist/site/batch-start.js.map +1 -0
- package/dist/site/connected-repos.d.ts +25 -0
- package/dist/site/connected-repos.d.ts.map +1 -0
- package/dist/site/connected-repos.js +109 -0
- package/dist/site/connected-repos.js.map +1 -0
- package/dist/site/evaluation-start.d.ts +4 -0
- package/dist/site/evaluation-start.d.ts.map +1 -0
- package/dist/site/evaluation-start.js +19 -0
- package/dist/site/evaluation-start.js.map +1 -0
- package/dist/site/generation-config.d.ts +4 -0
- package/dist/site/generation-config.d.ts.map +1 -0
- package/dist/site/generation-config.js +16 -0
- package/dist/site/generation-config.js.map +1 -0
- package/dist/site/generation-credentials.d.ts +6 -0
- package/dist/site/generation-credentials.d.ts.map +1 -0
- package/dist/site/generation-credentials.js +101 -0
- package/dist/site/generation-credentials.js.map +1 -0
- package/dist/site/generation-run.d.ts +6 -0
- package/dist/site/generation-run.d.ts.map +1 -0
- package/dist/site/generation-run.js +22 -0
- package/dist/site/generation-run.js.map +1 -0
- package/dist/site/generation-settings.d.ts +80 -0
- package/dist/site/generation-settings.d.ts.map +1 -0
- package/dist/site/generation-settings.js +84 -0
- package/dist/site/generation-settings.js.map +1 -0
- package/dist/site/generation-subscription.d.ts +3 -0
- package/dist/site/generation-subscription.d.ts.map +1 -0
- package/dist/site/generation-subscription.js +38 -0
- package/dist/site/generation-subscription.js.map +1 -0
- package/dist/site/github-repos.d.ts +35 -0
- package/dist/site/github-repos.d.ts.map +1 -0
- package/dist/site/github-repos.js +121 -0
- package/dist/site/github-repos.js.map +1 -0
- package/dist/site/pr-candidate.d.ts +26 -0
- package/dist/site/pr-candidate.d.ts.map +1 -0
- package/dist/site/pr-candidate.js +101 -0
- package/dist/site/pr-candidate.js.map +1 -0
- package/dist/site/pr-list.d.ts +13 -0
- package/dist/site/pr-list.d.ts.map +1 -0
- package/dist/site/pr-list.js +24 -0
- package/dist/site/pr-list.js.map +1 -0
- package/dist/site/pr-routes.d.ts +26 -0
- package/dist/site/pr-routes.d.ts.map +1 -0
- package/dist/site/pr-routes.js +139 -0
- package/dist/site/pr-routes.js.map +1 -0
- package/dist/site/repo-store.d.ts +40 -0
- package/dist/site/repo-store.d.ts.map +1 -0
- package/dist/site/repo-store.js +74 -0
- package/dist/site/repo-store.js.map +1 -0
- package/dist/site/run-store.d.ts +17 -0
- package/dist/site/run-store.d.ts.map +1 -0
- package/dist/site/run-store.js +42 -0
- package/dist/site/run-store.js.map +1 -0
- package/dist/site/runtime.d.ts +27 -0
- package/dist/site/runtime.d.ts.map +1 -0
- package/dist/site/runtime.js +97 -0
- package/dist/site/runtime.js.map +1 -0
- package/dist/site/task-deletion.d.ts +7 -0
- package/dist/site/task-deletion.d.ts.map +1 -0
- package/dist/site/task-deletion.js +29 -0
- package/dist/site/task-deletion.js.map +1 -0
- package/dist/site/task-start-reservation.d.ts +4 -0
- package/dist/site/task-start-reservation.d.ts.map +1 -0
- package/dist/site/task-start-reservation.js +18 -0
- package/dist/site/task-start-reservation.js.map +1 -0
- package/dist/site/task-start.d.ts +33 -0
- package/dist/site/task-start.d.ts.map +1 -0
- package/dist/site/task-start.js +50 -0
- package/dist/site/task-start.js.map +1 -0
- package/dist/site/task-status.d.ts +36 -0
- package/dist/site/task-status.d.ts.map +1 -0
- package/dist/site/task-status.js +113 -0
- package/dist/site/task-status.js.map +1 -0
- package/dist/site/task-store.d.ts +78 -0
- package/dist/site/task-store.d.ts.map +1 -0
- package/dist/site/task-store.js +186 -0
- package/dist/site/task-store.js.map +1 -0
- package/dist/site/task-sync.d.ts +30 -0
- package/dist/site/task-sync.d.ts.map +1 -0
- package/dist/site/task-sync.js +121 -0
- package/dist/site/task-sync.js.map +1 -0
- package/dist/site/tasks.d.ts +47 -0
- package/dist/site/tasks.d.ts.map +1 -0
- package/dist/site/tasks.js +164 -0
- package/dist/site/tasks.js.map +1 -0
- package/dist/site/temporal-status.d.ts +8 -0
- package/dist/site/temporal-status.d.ts.map +1 -0
- package/dist/site/temporal-status.js +41 -0
- package/dist/site/temporal-status.js.map +1 -0
- package/dist/site/tenant.d.ts +4 -0
- package/dist/site/tenant.d.ts.map +1 -0
- package/dist/site/tenant.js +6 -0
- package/dist/site/tenant.js.map +1 -0
- package/dist/static-check.d.ts +38 -0
- package/dist/static-check.d.ts.map +1 -0
- package/dist/static-check.js +131 -0
- package/dist/static-check.js.map +1 -0
- package/dist/submission-hash.d.ts +18 -0
- package/dist/submission-hash.d.ts.map +1 -0
- package/dist/submission-hash.js +16 -0
- package/dist/submission-hash.js.map +1 -0
- package/dist/subscription-auth.d.ts +0 -7
- package/dist/subscription-auth.d.ts.map +1 -1
- package/dist/subscription-auth.js +3 -23
- package/dist/subscription-auth.js.map +1 -1
- package/dist/temporal/activities/agent-feed.d.ts +5 -0
- package/dist/temporal/activities/agent-feed.d.ts.map +1 -0
- package/dist/temporal/activities/agent-feed.js +41 -0
- package/dist/temporal/activities/agent-feed.js.map +1 -0
- package/dist/temporal/activities/agent-scripts.d.ts +16 -0
- package/dist/temporal/activities/agent-scripts.d.ts.map +1 -0
- package/dist/temporal/activities/agent-scripts.js +162 -0
- package/dist/temporal/activities/agent-scripts.js.map +1 -0
- package/dist/temporal/activities/authoring-round.d.ts +13 -0
- package/dist/temporal/activities/authoring-round.d.ts.map +1 -0
- package/dist/temporal/activities/authoring-round.js +160 -0
- package/dist/temporal/activities/authoring-round.js.map +1 -0
- package/dist/temporal/activities/constants.d.ts +5 -0
- package/dist/temporal/activities/constants.d.ts.map +1 -0
- package/dist/temporal/activities/constants.js +5 -0
- package/dist/temporal/activities/constants.js.map +1 -0
- package/dist/temporal/activities/discovery.d.ts +17 -0
- package/dist/temporal/activities/discovery.d.ts.map +1 -0
- package/dist/temporal/activities/discovery.js +153 -0
- package/dist/temporal/activities/discovery.js.map +1 -0
- package/dist/temporal/activities/drafts.d.ts +15 -0
- package/dist/temporal/activities/drafts.d.ts.map +1 -0
- package/dist/temporal/activities/drafts.js +51 -0
- package/dist/temporal/activities/drafts.js.map +1 -0
- package/dist/temporal/activities/excluded-source-prs.d.ts +9 -0
- package/dist/temporal/activities/excluded-source-prs.d.ts.map +1 -0
- package/dist/temporal/activities/excluded-source-prs.js +132 -0
- package/dist/temporal/activities/excluded-source-prs.js.map +1 -0
- package/dist/temporal/activities/export.d.ts +38 -0
- package/dist/temporal/activities/export.d.ts.map +1 -0
- package/dist/temporal/activities/export.js +79 -0
- package/dist/temporal/activities/export.js.map +1 -0
- package/dist/temporal/activities/factory.d.ts +5 -0
- package/dist/temporal/activities/factory.d.ts.map +1 -0
- package/dist/temporal/activities/factory.js +26 -0
- package/dist/temporal/activities/factory.js.map +1 -0
- package/dist/temporal/activities/gate-logs.d.ts +20 -0
- package/dist/temporal/activities/gate-logs.d.ts.map +1 -0
- package/dist/temporal/activities/gate-logs.js +62 -0
- package/dist/temporal/activities/gate-logs.js.map +1 -0
- package/dist/temporal/activities/generation-runtime.d.ts +6 -0
- package/dist/temporal/activities/generation-runtime.d.ts.map +1 -0
- package/dist/temporal/activities/generation-runtime.js +56 -0
- package/dist/temporal/activities/generation-runtime.js.map +1 -0
- package/dist/temporal/activities/harbor.d.ts +11 -0
- package/dist/temporal/activities/harbor.d.ts.map +1 -0
- package/dist/temporal/activities/harbor.js +112 -0
- package/dist/temporal/activities/harbor.js.map +1 -0
- package/dist/temporal/activities/modal-build-log.d.ts +12 -0
- package/dist/temporal/activities/modal-build-log.d.ts.map +1 -0
- package/dist/temporal/activities/modal-build-log.js +57 -0
- package/dist/temporal/activities/modal-build-log.js.map +1 -0
- package/dist/temporal/activities/prompts-authoring.d.ts +6 -0
- package/dist/temporal/activities/prompts-authoring.d.ts.map +1 -0
- package/dist/temporal/activities/prompts-authoring.js +87 -0
- package/dist/temporal/activities/prompts-authoring.js.map +1 -0
- package/dist/temporal/activities/prompts-verifier.d.ts +17 -0
- package/dist/temporal/activities/prompts-verifier.d.ts.map +1 -0
- package/dist/temporal/activities/prompts-verifier.js +66 -0
- package/dist/temporal/activities/prompts-verifier.js.map +1 -0
- package/dist/temporal/activities/provenance.d.ts +4 -0
- package/dist/temporal/activities/provenance.d.ts.map +1 -0
- package/dist/temporal/activities/provenance.js +18 -0
- package/dist/temporal/activities/provenance.js.map +1 -0
- package/dist/temporal/activities/replay.d.ts +20 -0
- package/dist/temporal/activities/replay.d.ts.map +1 -0
- package/dist/temporal/activities/replay.js +135 -0
- package/dist/temporal/activities/replay.js.map +1 -0
- package/dist/temporal/activities/round-outcome.d.ts +63 -0
- package/dist/temporal/activities/round-outcome.d.ts.map +1 -0
- package/dist/temporal/activities/round-outcome.js +127 -0
- package/dist/temporal/activities/round-outcome.js.map +1 -0
- package/dist/temporal/activities/runtime.d.ts +38 -0
- package/dist/temporal/activities/runtime.d.ts.map +1 -0
- package/dist/temporal/activities/runtime.js +146 -0
- package/dist/temporal/activities/runtime.js.map +1 -0
- package/dist/temporal/activities/session-verify.d.ts +45 -0
- package/dist/temporal/activities/session-verify.d.ts.map +1 -0
- package/dist/temporal/activities/session-verify.js +91 -0
- package/dist/temporal/activities/session-verify.js.map +1 -0
- package/dist/temporal/activities/submissions.d.ts +9 -0
- package/dist/temporal/activities/submissions.d.ts.map +1 -0
- package/dist/temporal/activities/submissions.js +28 -0
- package/dist/temporal/activities/submissions.js.map +1 -0
- package/dist/temporal/activities/task-compiler.d.ts +21 -0
- package/dist/temporal/activities/task-compiler.d.ts.map +1 -0
- package/dist/temporal/activities/task-compiler.js +63 -0
- package/dist/temporal/activities/task-compiler.js.map +1 -0
- package/dist/temporal/activities/types.d.ts +57 -0
- package/dist/temporal/activities/types.d.ts.map +1 -0
- package/dist/temporal/activities/types.js +2 -0
- package/dist/temporal/activities/types.js.map +1 -0
- package/dist/temporal/activities/verifier-material.d.ts +18 -0
- package/dist/temporal/activities/verifier-material.d.ts.map +1 -0
- package/dist/temporal/activities/verifier-material.js +49 -0
- package/dist/temporal/activities/verifier-material.js.map +1 -0
- package/dist/temporal/activities/verifier-outcome.d.ts +19 -0
- package/dist/temporal/activities/verifier-outcome.d.ts.map +1 -0
- package/dist/temporal/activities/verifier-outcome.js +70 -0
- package/dist/temporal/activities/verifier-outcome.js.map +1 -0
- package/dist/temporal/activities/verifier-round.d.ts +8 -0
- package/dist/temporal/activities/verifier-round.d.ts.map +1 -0
- package/dist/temporal/activities/verifier-round.js +110 -0
- package/dist/temporal/activities/verifier-round.js.map +1 -0
- package/dist/temporal/activities/verify-harbor.d.ts +15 -0
- package/dist/temporal/activities/verify-harbor.d.ts.map +1 -0
- package/dist/temporal/activities/verify-harbor.js +193 -0
- package/dist/temporal/activities/verify-harbor.js.map +1 -0
- package/dist/temporal/activities/verify.d.ts +11 -0
- package/dist/temporal/activities/verify.d.ts.map +1 -0
- package/dist/temporal/activities/verify.js +151 -0
- package/dist/temporal/activities/verify.js.map +1 -0
- package/dist/temporal/activities.d.ts +3 -51
- package/dist/temporal/activities.d.ts.map +1 -1
- package/dist/temporal/activities.js +2 -886
- package/dist/temporal/activities.js.map +1 -1
- package/dist/temporal/connection.d.ts.map +1 -1
- package/dist/temporal/connection.js +3 -2
- package/dist/temporal/connection.js.map +1 -1
- package/dist/temporal/worker-main.js +19 -3
- package/dist/temporal/worker-main.js.map +1 -1
- package/dist/temporal/workflow/activity-proxies.d.ts +3 -0
- package/dist/temporal/workflow/activity-proxies.d.ts.map +1 -0
- package/dist/temporal/workflow/activity-proxies.js +45 -0
- package/dist/temporal/workflow/activity-proxies.js.map +1 -0
- package/dist/temporal/workflow/authoring-stage.d.ts +5 -0
- package/dist/temporal/workflow/authoring-stage.d.ts.map +1 -0
- package/dist/temporal/workflow/authoring-stage.js +64 -0
- package/dist/temporal/workflow/authoring-stage.js.map +1 -0
- package/dist/temporal/workflow/candidate-tracker.d.ts +28 -0
- package/dist/temporal/workflow/candidate-tracker.d.ts.map +1 -0
- package/dist/temporal/workflow/candidate-tracker.js +85 -0
- package/dist/temporal/workflow/candidate-tracker.js.map +1 -0
- package/dist/temporal/workflow/candidate.d.ts +6 -0
- package/dist/temporal/workflow/candidate.d.ts.map +1 -0
- package/dist/temporal/workflow/candidate.js +54 -0
- package/dist/temporal/workflow/candidate.js.map +1 -0
- package/dist/temporal/workflow/discovery.d.ts +19 -0
- package/dist/temporal/workflow/discovery.d.ts.map +1 -0
- package/dist/temporal/workflow/discovery.js +137 -0
- package/dist/temporal/workflow/discovery.js.map +1 -0
- package/dist/temporal/workflow/failures.d.ts +5 -0
- package/dist/temporal/workflow/failures.d.ts.map +1 -0
- package/dist/temporal/workflow/failures.js +28 -0
- package/dist/temporal/workflow/failures.js.map +1 -0
- package/dist/temporal/workflow/run.d.ts +5 -0
- package/dist/temporal/workflow/run.d.ts.map +1 -0
- package/dist/temporal/workflow/run.js +154 -0
- package/dist/temporal/workflow/run.js.map +1 -0
- package/dist/temporal/workflow/stage.d.ts +26 -0
- package/dist/temporal/workflow/stage.d.ts.map +1 -0
- package/dist/temporal/workflow/stage.js +16 -0
- package/dist/temporal/workflow/stage.js.map +1 -0
- package/dist/temporal/workflow/verification-stage.d.ts +11 -0
- package/dist/temporal/workflow/verification-stage.d.ts.map +1 -0
- package/dist/temporal/workflow/verification-stage.js +20 -0
- package/dist/temporal/workflow/verification-stage.js.map +1 -0
- package/dist/temporal/workflow.d.ts +12 -4
- package/dist/temporal/workflow.d.ts.map +1 -1
- package/dist/temporal/workflow.js +38 -445
- package/dist/temporal/workflow.js.map +1 -1
- package/dist/verifier-fix.d.ts +20 -0
- package/dist/verifier-fix.d.ts.map +1 -0
- package/dist/verifier-fix.js +62 -0
- package/dist/verifier-fix.js.map +1 -0
- package/dist/verify-report.d.ts +9 -0
- package/dist/verify-report.d.ts.map +1 -0
- package/dist/verify-report.js +135 -0
- package/dist/verify-report.js.map +1 -0
- package/dist/viewer/archived-keys.d.ts +9 -0
- package/dist/viewer/archived-keys.d.ts.map +1 -0
- package/dist/viewer/archived-keys.js +36 -0
- package/dist/viewer/archived-keys.js.map +1 -0
- package/dist/viewer/archived.d.ts +20 -0
- package/dist/viewer/archived.d.ts.map +1 -0
- package/dist/viewer/archived.js +238 -0
- package/dist/viewer/archived.js.map +1 -0
- package/dist/viewer/artifacts.d.ts +8 -0
- package/dist/viewer/artifacts.d.ts.map +1 -0
- package/dist/viewer/artifacts.js +96 -0
- package/dist/viewer/artifacts.js.map +1 -0
- package/dist/viewer/bundle.d.ts +9 -0
- package/dist/viewer/bundle.d.ts.map +1 -0
- package/dist/viewer/bundle.js +79 -0
- package/dist/viewer/bundle.js.map +1 -0
- package/dist/viewer/candidates.d.ts +9 -0
- package/dist/viewer/candidates.d.ts.map +1 -0
- package/dist/viewer/candidates.js +137 -0
- package/dist/viewer/candidates.js.map +1 -0
- package/dist/viewer/local-server.d.ts +11 -0
- package/dist/viewer/local-server.d.ts.map +1 -0
- package/dist/viewer/local-server.js +64 -0
- package/dist/viewer/local-server.js.map +1 -0
- package/dist/viewer/task-files.d.ts +10 -0
- package/dist/viewer/task-files.d.ts.map +1 -0
- package/dist/viewer/task-files.js +120 -0
- package/dist/viewer/task-files.js.map +1 -0
- package/dist/viewer/types.d.ts +78 -0
- package/dist/viewer/types.d.ts.map +1 -0
- package/dist/viewer/types.js +36 -0
- package/dist/viewer/types.js.map +1 -0
- package/docs/api.md +168 -0
- package/docs/evaluations.md +42 -15
- package/docs/operations.md +254 -26
- package/docs/solver-evaluations.md +63 -0
- package/docs/task-construction.md +21 -11
- package/drizzle/0000_glossy_johnny_blaze.sql +97 -0
- package/drizzle/0001_evaluation_records.sql +5 -0
- package/drizzle/0002_task_deletion.sql +1 -0
- package/drizzle/0003_drop_repo_runs.sql +3 -0
- package/drizzle/0004_preserve_batch_runs.sql +8 -0
- package/drizzle/0005_api_keys.sql +15 -0
- package/drizzle/meta/0000_snapshot.json +776 -0
- package/drizzle/meta/0001_snapshot.json +807 -0
- package/drizzle/meta/0002_snapshot.json +813 -0
- package/drizzle/meta/0003_snapshot.json +726 -0
- package/drizzle/meta/0004_snapshot.json +813 -0
- package/drizzle/meta/0005_snapshot.json +919 -0
- package/drizzle/meta/_journal.json +48 -0
- package/package.json +39 -25
- package/src/extensions/authoring.ts +83 -72
- package/src/extensions/shared/deliverable.ts +153 -0
- package/src/extensions/shared/mailbox.ts +128 -0
- package/src/extensions/shared/schemas.ts +101 -0
- package/src/extensions/shared/static-check.ts +103 -0
- package/src/extensions/verifier.ts +88 -0
- package/src/skills/selfbench/SKILL.md +71 -13
- package/biome.json +0 -24
- package/dist/agent-smoke-main.d.ts +0 -3
- package/dist/agent-smoke-main.d.ts.map +0 -1
- package/dist/agent-smoke-main.js +0 -55
- package/dist/agent-smoke-main.js.map +0 -1
- package/dist/agent-smoke.d.ts +0 -18
- package/dist/agent-smoke.d.ts.map +0 -1
- package/dist/agent-smoke.js +0 -103
- package/dist/agent-smoke.js.map +0 -1
- package/dist/codex-review.d.ts +0 -41
- package/dist/codex-review.d.ts.map +0 -1
- package/dist/codex-review.js +0 -193
- package/dist/codex-review.js.map +0 -1
- package/dist/eval-main.d.ts +0 -3
- package/dist/eval-main.d.ts.map +0 -1
- package/dist/eval-main.js +0 -81
- package/dist/eval-main.js.map +0 -1
- package/dist/evaluate.d.ts +0 -25
- package/dist/evaluate.d.ts.map +0 -1
- package/dist/evaluate.js +0 -214
- package/dist/evaluate.js.map +0 -1
- package/dist/modal-auth.d.ts.map +0 -1
- package/dist/modal-auth.js.map +0 -1
- package/dist/reaudit-main.d.ts +0 -3
- package/dist/reaudit-main.d.ts.map +0 -1
- package/dist/reaudit-main.js +0 -158
- package/dist/reaudit-main.js.map +0 -1
- package/dist/repair-main.d.ts +0 -3
- package/dist/repair-main.d.ts.map +0 -1
- package/dist/repair-main.js +0 -171
- package/dist/repair-main.js.map +0 -1
- package/dist/review/assets/index-CZ3G759O.js +0 -1568
- package/dist/review/assets/index-DgSLzX8l.css +0 -1
- package/dist/sandbox/programs/prepare-repair.d.ts +0 -7
- package/dist/sandbox/programs/prepare-repair.d.ts.map +0 -1
- package/dist/sandbox/programs/prepare-repair.js.map +0 -1
- package/dist/sandbox/programs/repair.d.ts +0 -3
- package/dist/sandbox/programs/repair.d.ts.map +0 -1
- package/dist/sandbox/programs/repair.js +0 -100
- package/dist/sandbox/programs/repair.js.map +0 -1
- package/dist/sandbox/programs/review.d.ts +0 -2
- package/dist/sandbox/programs/review.d.ts.map +0 -1
- package/dist/sandbox/programs/review.js +0 -18
- package/dist/sandbox/programs/review.js.map +0 -1
- package/dist/sandbox/programs/validation-repair.d.ts +0 -3
- package/dist/sandbox/programs/validation-repair.d.ts.map +0 -1
- package/dist/sandbox/programs/validation-repair.js +0 -115
- package/dist/sandbox/programs/validation-repair.js.map +0 -1
- package/dist/sandbox-repair.bundle.js +0 -323
- package/dist/sandbox-review.bundle.js +0 -14462
- package/dist/validate-main.d.ts +0 -3
- package/dist/validate-main.d.ts.map +0 -1
- package/dist/validate-main.js +0 -143
- package/dist/validate-main.js.map +0 -1
- package/dist/validation-repair.d.ts +0 -10
- package/dist/validation-repair.d.ts.map +0 -1
- package/dist/validation-repair.js +0 -72
- package/dist/validation-repair.js.map +0 -1
- package/review/index.html +0 -14
- package/review/src/App.tsx +0 -338
- package/review/src/export-loader.test.ts +0 -57
- package/review/src/export-loader.ts +0 -151
- package/review/src/main.tsx +0 -15
- package/review/src/styles.css +0 -162
- package/review/src/types.ts +0 -47
- package/review/tsconfig.json +0 -16
- package/review/vite.config.ts +0 -20
- package/scripts/build-sandbox-programs.ts +0 -22
- package/scripts/verify-package.ts +0 -102
- package/scripts/write-build-metadata.ts +0 -27
- package/src/agent-smoke-main.ts +0 -64
- package/src/agent-smoke.ts +0 -138
- package/src/api-main.ts +0 -12
- package/src/api.ts +0 -291
- package/src/artifacts.ts +0 -357
- package/src/audit.ts +0 -106
- package/src/build-metadata.ts +0 -3
- package/src/cli.ts +0 -504
- package/src/codex-review.ts +0 -229
- package/src/config.ts +0 -239
- package/src/contracts.ts +0 -233
- package/src/coupling.ts +0 -259
- package/src/eval-main.ts +0 -93
- package/src/evaluate.ts +0 -300
- package/src/github.ts +0 -26
- package/src/harbor-environment.ts +0 -17
- package/src/harbor-results.ts +0 -144
- package/src/harbor-task.ts +0 -533
- package/src/hash.ts +0 -5
- package/src/modal-auth.ts +0 -11
- package/src/parallel.ts +0 -24
- package/src/process.ts +0 -167
- package/src/project-paths.ts +0 -15
- package/src/provenance-associations.ts +0 -311
- package/src/provenance.ts +0 -561
- package/src/providers.ts +0 -22
- package/src/reaudit-main.ts +0 -198
- package/src/repair-main.ts +0 -209
- package/src/repair.ts +0 -55
- package/src/run-wait.ts +0 -40
- package/src/sandbox/contracts.ts +0 -52
- package/src/sandbox/index.ts +0 -23
- package/src/sandbox/programs/author.ts +0 -19
- package/src/sandbox/programs/prepare-repair.ts +0 -44
- package/src/sandbox/programs/repair.ts +0 -132
- package/src/sandbox/programs/review.ts +0 -19
- package/src/sandbox/programs/validation-repair.ts +0 -153
- package/src/sandbox/providers/docker/executor.ts +0 -115
- package/src/sandbox/providers/modal/executor.ts +0 -217
- package/src/sandbox/providers/vercel/command.ts +0 -127
- package/src/sandbox/providers/vercel/executor.ts +0 -448
- package/src/sandbox/providers/vercel/fetch.ts +0 -45
- package/src/sandbox/timeout.ts +0 -56
- package/src/setup/vercel/cli.ts +0 -403
- package/src/setup/vercel/index.ts +0 -3
- package/src/setup/vercel/probe.ts +0 -232
- package/src/setup/vercel/profile.ts +0 -322
- package/src/setup/vercel/runtime-image.ts +0 -204
- package/src/setup/vercel/setup.ts +0 -419
- package/src/subscription-auth.ts +0 -80
- package/src/temporal/activities.ts +0 -1348
- package/src/temporal/connection.ts +0 -23
- package/src/temporal/worker-main.ts +0 -28
- package/src/temporal/workflow.ts +0 -568
- package/src/terminal-prompts.ts +0 -197
- package/src/terminal-reporter.ts +0 -126
- package/src/validate-main.ts +0 -177
- package/src/validation-repair.ts +0 -94
- package/tsconfig.build.json +0 -13
- package/tsconfig.json +0 -21
|
@@ -0,0 +1,11 @@
|
|
|
1
|
+
import type { Harness } from "./types.js";
|
|
2
|
+
export declare const providerIds: readonly ["openai", "anthropic", "google", "groq", "huggingface", "mistral", "openrouter", "xai"];
|
|
3
|
+
export type EvaluationProvider = (typeof providerIds)[number];
|
|
4
|
+
export declare const providers: Record<EvaluationProvider, {
|
|
5
|
+
label: string;
|
|
6
|
+
credential: string;
|
|
7
|
+
harnesses: Harness[];
|
|
8
|
+
}>;
|
|
9
|
+
export declare const modelIdPattern: RegExp;
|
|
10
|
+
export declare function modelProvider(modelName: string): EvaluationProvider | undefined;
|
|
11
|
+
//# sourceMappingURL=providers.d.ts.map
|
|
@@ -0,0 +1 @@
|
|
|
1
|
+
{"version":3,"file":"providers.d.ts","sourceRoot":"","sources":["../../src/evaluation/providers.ts"],"names":[],"mappings":"AAAA,OAAO,KAAK,EAAE,OAAO,EAAE,MAAM,YAAY,CAAC;AAE1C,eAAO,MAAM,WAAW,mGASd,CAAC;AACX,MAAM,MAAM,kBAAkB,GAAG,CAAC,OAAO,WAAW,CAAC,CAAC,MAAM,CAAC,CAAC;AAC9D,eAAO,MAAM,SAAS,EAAE,MAAM,CAC5B,kBAAkB,EAClB;IAAE,KAAK,EAAE,MAAM,CAAC;IAAC,UAAU,EAAE,MAAM,CAAC;IAAC,SAAS,EAAE,OAAO,EAAE,CAAA;CAAE,CAc5D,CAAC;AACF,eAAO,MAAM,cAAc,QAAyC,CAAC;AACrE,wBAAgB,aAAa,CAAC,SAAS,EAAE,MAAM,GAAG,kBAAkB,GAAG,SAAS,CAG/E"}
|
|
@@ -0,0 +1,30 @@
|
|
|
1
|
+
export const providerIds = [
|
|
2
|
+
"openai",
|
|
3
|
+
"anthropic",
|
|
4
|
+
"google",
|
|
5
|
+
"groq",
|
|
6
|
+
"huggingface",
|
|
7
|
+
"mistral",
|
|
8
|
+
"openrouter",
|
|
9
|
+
"xai",
|
|
10
|
+
];
|
|
11
|
+
export const providers = {
|
|
12
|
+
openai: { label: "OpenAI", credential: "OPENAI_API_KEY", harnesses: ["codex", "pi"] },
|
|
13
|
+
anthropic: {
|
|
14
|
+
label: "Anthropic",
|
|
15
|
+
credential: "ANTHROPIC_API_KEY",
|
|
16
|
+
harnesses: ["claude-code", "pi"],
|
|
17
|
+
},
|
|
18
|
+
google: { label: "Google Gemini", credential: "GEMINI_API_KEY", harnesses: ["pi"] },
|
|
19
|
+
groq: { label: "Groq", credential: "GROQ_API_KEY", harnesses: ["pi"] },
|
|
20
|
+
huggingface: { label: "Hugging Face", credential: "HF_TOKEN", harnesses: ["pi"] },
|
|
21
|
+
mistral: { label: "Mistral", credential: "MISTRAL_API_KEY", harnesses: ["pi"] },
|
|
22
|
+
openrouter: { label: "OpenRouter", credential: "OPENROUTER_API_KEY", harnesses: ["pi"] },
|
|
23
|
+
xai: { label: "xAI", credential: "XAI_API_KEY", harnesses: ["pi"] },
|
|
24
|
+
};
|
|
25
|
+
export const modelIdPattern = /^[A-Za-z0-9][A-Za-z0-9._:/-]{0,199}$/;
|
|
26
|
+
export function modelProvider(modelName) {
|
|
27
|
+
const prefix = modelName.split("/", 1)[0];
|
|
28
|
+
return providerIds.find((id) => id === prefix);
|
|
29
|
+
}
|
|
30
|
+
//# sourceMappingURL=providers.js.map
|
|
@@ -0,0 +1 @@
|
|
|
1
|
+
{"version":3,"file":"providers.js","sourceRoot":"","sources":["../../src/evaluation/providers.ts"],"names":[],"mappings":"AAEA,MAAM,CAAC,MAAM,WAAW,GAAG;IACzB,QAAQ;IACR,WAAW;IACX,QAAQ;IACR,MAAM;IACN,aAAa;IACb,SAAS;IACT,YAAY;IACZ,KAAK;CACG,CAAC;AAEX,MAAM,CAAC,MAAM,SAAS,GAGlB;IACF,MAAM,EAAE,EAAE,KAAK,EAAE,QAAQ,EAAE,UAAU,EAAE,gBAAgB,EAAE,SAAS,EAAE,CAAC,OAAO,EAAE,IAAI,CAAC,EAAE;IACrF,SAAS,EAAE;QACT,KAAK,EAAE,WAAW;QAClB,UAAU,EAAE,mBAAmB;QAC/B,SAAS,EAAE,CAAC,aAAa,EAAE,IAAI,CAAC;KACjC;IACD,MAAM,EAAE,EAAE,KAAK,EAAE,eAAe,EAAE,UAAU,EAAE,gBAAgB,EAAE,SAAS,EAAE,CAAC,IAAI,CAAC,EAAE;IACnF,IAAI,EAAE,EAAE,KAAK,EAAE,MAAM,EAAE,UAAU,EAAE,cAAc,EAAE,SAAS,EAAE,CAAC,IAAI,CAAC,EAAE;IACtE,WAAW,EAAE,EAAE,KAAK,EAAE,cAAc,EAAE,UAAU,EAAE,UAAU,EAAE,SAAS,EAAE,CAAC,IAAI,CAAC,EAAE;IACjF,OAAO,EAAE,EAAE,KAAK,EAAE,SAAS,EAAE,UAAU,EAAE,iBAAiB,EAAE,SAAS,EAAE,CAAC,IAAI,CAAC,EAAE;IAC/E,UAAU,EAAE,EAAE,KAAK,EAAE,YAAY,EAAE,UAAU,EAAE,oBAAoB,EAAE,SAAS,EAAE,CAAC,IAAI,CAAC,EAAE;IACxF,GAAG,EAAE,EAAE,KAAK,EAAE,KAAK,EAAE,UAAU,EAAE,aAAa,EAAE,SAAS,EAAE,CAAC,IAAI,CAAC,EAAE;CACpE,CAAC;AACF,MAAM,CAAC,MAAM,cAAc,GAAG,sCAAsC,CAAC;AACrE,MAAM,UAAU,aAAa,CAAC,SAAiB;IAC7C,MAAM,MAAM,GAAG,SAAS,CAAC,KAAK,CAAC,GAAG,EAAE,CAAC,CAAC,CAAC,CAAC,CAAC,CAAC;IAC1C,OAAO,WAAW,CAAC,IAAI,CAAC,CAAC,EAAE,EAAE,EAAE,CAAC,EAAE,KAAK,MAAM,CAAC,CAAC;AACjD,CAAC"}
|
|
@@ -0,0 +1,23 @@
|
|
|
1
|
+
import type { IncomingMessage, ServerResponse } from "node:http";
|
|
2
|
+
import type { ArtifactStore } from "../artifacts.js";
|
|
3
|
+
import type { User, UserStore } from "../auth/users.js";
|
|
4
|
+
import type { RepoStore } from "../site/repo-store.js";
|
|
5
|
+
import type { TaskStore } from "../site/task-store.js";
|
|
6
|
+
import type { CodexLogins } from "./codex-login.js";
|
|
7
|
+
import type { EncryptedRecordStore } from "./encrypted-records.js";
|
|
8
|
+
import type { EvaluationInput } from "./types.js";
|
|
9
|
+
export interface EvaluationRoutesOptions {
|
|
10
|
+
users: UserStore;
|
|
11
|
+
repos: RepoStore;
|
|
12
|
+
tasks: TaskStore;
|
|
13
|
+
artifacts: ArtifactStore;
|
|
14
|
+
publicUrl: string;
|
|
15
|
+
start(input: EvaluationInput): Promise<void>;
|
|
16
|
+
env?: NodeJS.ProcessEnv;
|
|
17
|
+
records?: EncryptedRecordStore;
|
|
18
|
+
codexLogins?: CodexLogins;
|
|
19
|
+
}
|
|
20
|
+
export declare function createEvaluationRoutes(options: EvaluationRoutesOptions): {
|
|
21
|
+
handle(request: IncomingMessage, url: URL, response: ServerResponse, user: User): Promise<boolean>;
|
|
22
|
+
};
|
|
23
|
+
//# sourceMappingURL=routes.d.ts.map
|
|
@@ -0,0 +1 @@
|
|
|
1
|
+
{"version":3,"file":"routes.d.ts","sourceRoot":"","sources":["../../src/evaluation/routes.ts"],"names":[],"mappings":"AAAA,OAAO,KAAK,EAAE,eAAe,EAAE,cAAc,EAAE,MAAM,WAAW,CAAC;AAGjE,OAAO,KAAK,EAAE,aAAa,EAAE,MAAM,iBAAiB,CAAC;AACrD,OAAO,KAAK,EAAE,IAAI,EAAE,SAAS,EAAE,MAAM,kBAAkB,CAAC;AACxD,OAAO,KAAK,EAAE,SAAS,EAAE,MAAM,uBAAuB,CAAC;AACvD,OAAO,KAAK,EAAE,SAAS,EAAE,MAAM,uBAAuB,CAAC;AAEvD,OAAO,KAAK,EAAE,WAAW,EAAE,MAAM,kBAAkB,CAAC;AAGpD,OAAO,KAAK,EAAE,oBAAoB,EAAE,MAAM,wBAAwB,CAAC;AAYnE,OAAO,KAAK,EAAE,eAAe,EAAE,MAAM,YAAY,CAAC;AAmBlD,MAAM,WAAW,uBAAuB;IACtC,KAAK,EAAE,SAAS,CAAC;IACjB,KAAK,EAAE,SAAS,CAAC;IACjB,KAAK,EAAE,SAAS,CAAC;IACjB,SAAS,EAAE,aAAa,CAAC;IACzB,SAAS,EAAE,MAAM,CAAC;IAClB,KAAK,CAAC,KAAK,EAAE,eAAe,GAAG,OAAO,CAAC,IAAI,CAAC,CAAC;IAC7C,GAAG,CAAC,EAAE,MAAM,CAAC,UAAU,CAAC;IACxB,OAAO,CAAC,EAAE,oBAAoB,CAAC;IAC/B,WAAW,CAAC,EAAE,WAAW,CAAC;CAC3B;AACD,wBAAgB,sBAAsB,CAAC,OAAO,EAAE,uBAAuB;oBAIxD,eAAe,OACnB,GAAG,YACE,cAAc,QAClB,IAAI,GACT,OAAO,CAAC,OAAO,CAAC;EA6MtB"}
|
|
@@ -0,0 +1,212 @@
|
|
|
1
|
+
import { z } from "zod";
|
|
2
|
+
import { readBody, sendJson, trustedMutation } from "../api/http.js";
|
|
3
|
+
import { tenantFor } from "../site/tenant.js";
|
|
4
|
+
import { codexLoginRoutes } from "./codex-login-routes.js";
|
|
5
|
+
import { evaluationSandboxes } from "./config.js";
|
|
6
|
+
import { harnessIds } from "./harnesses.js";
|
|
7
|
+
import { orgCredentialRoutes } from "./org-credential-routes.js";
|
|
8
|
+
import { platformRoutes } from "./platform-routes.js";
|
|
9
|
+
import { availableChoices, handleSetup } from "./profile-routes.js";
|
|
10
|
+
import { evaluationPrefix, getEvaluation, initialEvaluation, listEvaluations, saveEvaluation, } from "./store.js";
|
|
11
|
+
const route = /^\/api\/orgs\/([A-Za-z0-9_.-]+)\/repos\/([A-Za-z0-9_.-]+)\/([A-Za-z0-9_.-]+)\/evaluations(?:\/(options|profiles|[a-f0-9-]{36}))?(?:\/artifacts)?$/;
|
|
12
|
+
const requestSchema = z
|
|
13
|
+
.object({
|
|
14
|
+
id: z.uuid().transform((value) => value.toLowerCase()),
|
|
15
|
+
model: z.string().min(1).max(60),
|
|
16
|
+
harnesses: z.array(z.enum(harnessIds)).min(1).max(harnessIds.length),
|
|
17
|
+
sandbox: z.enum(evaluationSandboxes),
|
|
18
|
+
tasks: z
|
|
19
|
+
.array(z
|
|
20
|
+
.object({ runId: z.string().min(1).max(100), taskId: z.string().min(1).max(200) })
|
|
21
|
+
.strict())
|
|
22
|
+
.min(1),
|
|
23
|
+
})
|
|
24
|
+
.strict();
|
|
25
|
+
export function createEvaluationRoutes(options) {
|
|
26
|
+
const { users, repos, tasks, artifacts } = options;
|
|
27
|
+
return {
|
|
28
|
+
async handle(request, url, response, user) {
|
|
29
|
+
if (await codexLoginRoutes(options, request, url, response, user))
|
|
30
|
+
return true;
|
|
31
|
+
if (await orgCredentialRoutes(options, request, url, response, user))
|
|
32
|
+
return true;
|
|
33
|
+
if (await platformRoutes(options, request, url, response, user))
|
|
34
|
+
return true;
|
|
35
|
+
const match = route.exec(url.pathname);
|
|
36
|
+
if (!match?.[1] || !match[2] || !match[3])
|
|
37
|
+
return false;
|
|
38
|
+
response.setHeader("cache-control", "no-store");
|
|
39
|
+
const tenant = await tenantFor(users, user, match[1]);
|
|
40
|
+
const repo = tenant ? await repos.find(tenant.id, `${match[2]}/${match[3]}`) : undefined;
|
|
41
|
+
if (!tenant || !repo) {
|
|
42
|
+
sendJson(response, 404, { error: "Repository is not connected here" });
|
|
43
|
+
return true;
|
|
44
|
+
}
|
|
45
|
+
const env = options.env ?? process.env;
|
|
46
|
+
if (match[4] === "profiles") {
|
|
47
|
+
await handleSetup(request, response, artifacts, {
|
|
48
|
+
repoId: repo.id,
|
|
49
|
+
ownerId: user.githubId,
|
|
50
|
+
tenant: tenant.login,
|
|
51
|
+
publicUrl: options.publicUrl,
|
|
52
|
+
trusted: trustedMutation(request, options.publicUrl, user),
|
|
53
|
+
}, env);
|
|
54
|
+
return true;
|
|
55
|
+
}
|
|
56
|
+
if (request.method === "GET") {
|
|
57
|
+
if (match[4] === "options") {
|
|
58
|
+
const available = (await tasks.listForRepo(repo.id)).filter((task) => task.bundleKey &&
|
|
59
|
+
task.pipelineStatus === "accepted" &&
|
|
60
|
+
task.review?.decision === "approve");
|
|
61
|
+
sendJson(response, 200, {
|
|
62
|
+
...(await availableChoices(artifacts, repo.id, user.githubId, tenant.login, env)),
|
|
63
|
+
tasks: available.map((task) => ({
|
|
64
|
+
runId: task.runId,
|
|
65
|
+
taskId: task.taskId,
|
|
66
|
+
difficulty: task.difficulty,
|
|
67
|
+
})),
|
|
68
|
+
});
|
|
69
|
+
}
|
|
70
|
+
else if (match[4]) {
|
|
71
|
+
const run = await getEvaluation(artifacts, repo.id, match[4]);
|
|
72
|
+
if (!run) {
|
|
73
|
+
sendJson(response, 404, { error: "Evaluation not found" });
|
|
74
|
+
return true;
|
|
75
|
+
}
|
|
76
|
+
if (url.pathname.endsWith("/artifacts")) {
|
|
77
|
+
const name = url.searchParams.get("name") ?? "";
|
|
78
|
+
if (!run.trials.some((trial) => trial.artifacts.includes(name))) {
|
|
79
|
+
sendJson(response, 404, { error: "Artifact not found" });
|
|
80
|
+
return true;
|
|
81
|
+
}
|
|
82
|
+
const bytes = await artifacts.getByKey(`${evaluationPrefix(repo.id, run.id)}artifacts/${name}`);
|
|
83
|
+
if (!bytes) {
|
|
84
|
+
sendJson(response, 404, { error: "Artifact not found" });
|
|
85
|
+
return true;
|
|
86
|
+
}
|
|
87
|
+
response.writeHead(200, {
|
|
88
|
+
"content-type": "text/plain; charset=utf-8",
|
|
89
|
+
"x-content-type-options": "nosniff",
|
|
90
|
+
"content-disposition": "attachment; filename=solver-artifact.txt",
|
|
91
|
+
});
|
|
92
|
+
response.end(bytes);
|
|
93
|
+
}
|
|
94
|
+
else
|
|
95
|
+
sendJson(response, 200, run);
|
|
96
|
+
}
|
|
97
|
+
else {
|
|
98
|
+
sendJson(response, 200, {
|
|
99
|
+
runs: (await listEvaluations(artifacts, repo.id)).map((run) => ({
|
|
100
|
+
...run,
|
|
101
|
+
trials: run.trials.map((trial) => ({ ...trial, log: "", steps: [], artifacts: [] })),
|
|
102
|
+
})),
|
|
103
|
+
});
|
|
104
|
+
}
|
|
105
|
+
return true;
|
|
106
|
+
}
|
|
107
|
+
if (request.method !== "POST" || match[4]) {
|
|
108
|
+
sendJson(response, 405, { error: "Method not allowed" });
|
|
109
|
+
return true;
|
|
110
|
+
}
|
|
111
|
+
if (!trustedMutation(request, options.publicUrl, user)) {
|
|
112
|
+
sendJson(response, 403, { error: "Same-origin JSON request required" });
|
|
113
|
+
return true;
|
|
114
|
+
}
|
|
115
|
+
let body;
|
|
116
|
+
try {
|
|
117
|
+
body = requestSchema.parse(JSON.parse((await readBody(request, 16_384)).toString("utf8")));
|
|
118
|
+
}
|
|
119
|
+
catch {
|
|
120
|
+
sendJson(response, 400, { error: "Invalid evaluation selection" });
|
|
121
|
+
return true;
|
|
122
|
+
}
|
|
123
|
+
const choices = await availableChoices(artifacts, repo.id, user.githubId, tenant.login, env);
|
|
124
|
+
const model = choices.models.find((candidate) => candidate.id === body.model);
|
|
125
|
+
if (!model ||
|
|
126
|
+
(model.sandbox
|
|
127
|
+
? model.sandbox !== body.sandbox
|
|
128
|
+
: !choices.sandboxes.includes(body.sandbox)) ||
|
|
129
|
+
body.harnesses.some((harness) => !model.harnesses.includes(harness)) ||
|
|
130
|
+
new Set(body.harnesses).size !== body.harnesses.length ||
|
|
131
|
+
new Set(body.tasks.map((task) => `${task.runId}/${task.taskId}`)).size !== body.tasks.length) {
|
|
132
|
+
sendJson(response, 400, {
|
|
133
|
+
error: "Select configured models, harnesses, sandbox and unique tasks",
|
|
134
|
+
});
|
|
135
|
+
return true;
|
|
136
|
+
}
|
|
137
|
+
const selected = await Promise.all(body.tasks.map((task) => tasks.find(repo.id, task.runId, task.taskId)));
|
|
138
|
+
if (selected.some((task) => !task?.bundleKey ||
|
|
139
|
+
task.pipelineStatus !== "accepted" ||
|
|
140
|
+
task.review?.decision !== "approve")) {
|
|
141
|
+
sendJson(response, 400, {
|
|
142
|
+
error: "Every task must belong to this repository and have a human-approved Harbor bundle",
|
|
143
|
+
});
|
|
144
|
+
return true;
|
|
145
|
+
}
|
|
146
|
+
if (new Set(selected.map((task) => `${task?.runId}/${task?.taskId}`)).size !== selected.length) {
|
|
147
|
+
sendJson(response, 400, { error: "Select each task only once" });
|
|
148
|
+
return true;
|
|
149
|
+
}
|
|
150
|
+
const existing = await getEvaluation(artifacts, repo.id, body.id);
|
|
151
|
+
if (existing) {
|
|
152
|
+
if (existing.model !== body.model ||
|
|
153
|
+
existing.sandbox !== body.sandbox ||
|
|
154
|
+
JSON.stringify(existing.harnesses) !== JSON.stringify(body.harnesses) ||
|
|
155
|
+
JSON.stringify(existing.trials.map((trial) => `${trial.runId}/${trial.taskId}/${trial.harness}`)) !==
|
|
156
|
+
JSON.stringify(body.tasks.flatMap((task) => body.harnesses.map((harness) => `${task.runId}/${task.taskId}/${harness}`)))) {
|
|
157
|
+
sendJson(response, 409, { error: "This request ID belongs to a different selection" });
|
|
158
|
+
return true;
|
|
159
|
+
}
|
|
160
|
+
if (existing.status === "queued") {
|
|
161
|
+
const saved = await artifacts.getByKey(`${evaluationPrefix(repo.id, body.id)}request.json`);
|
|
162
|
+
if (!saved)
|
|
163
|
+
throw new Error("Evaluation request snapshot is missing");
|
|
164
|
+
await options.start(JSON.parse(Buffer.from(saved).toString("utf8")));
|
|
165
|
+
}
|
|
166
|
+
sendJson(response, 202, existing);
|
|
167
|
+
return true;
|
|
168
|
+
}
|
|
169
|
+
let input = {
|
|
170
|
+
...body,
|
|
171
|
+
modelName: model.model,
|
|
172
|
+
...(model.id.startsWith("saved-") ? { credentialOwnerId: user.githubId } : {}),
|
|
173
|
+
...(model.pricing ? { pricing: model.pricing } : {}),
|
|
174
|
+
repoId: repo.id,
|
|
175
|
+
tenant: tenant.login,
|
|
176
|
+
startedBy: user.login,
|
|
177
|
+
createdAt: new Date().toISOString(),
|
|
178
|
+
tasks: selected.map((task) => {
|
|
179
|
+
if (!task?.bundleKey)
|
|
180
|
+
throw new Error("Task bundle is missing");
|
|
181
|
+
return { runId: task.runId, taskId: task.taskId, bundleKey: task.bundleKey };
|
|
182
|
+
}),
|
|
183
|
+
};
|
|
184
|
+
const requestKey = `${evaluationPrefix(repo.id, body.id)}request.json`;
|
|
185
|
+
const savedRequest = await artifacts.getByKey(requestKey);
|
|
186
|
+
if (savedRequest) {
|
|
187
|
+
const previous = JSON.parse(Buffer.from(savedRequest).toString("utf8"));
|
|
188
|
+
const signature = (value) => JSON.stringify({
|
|
189
|
+
model: value.model,
|
|
190
|
+
modelName: value.modelName,
|
|
191
|
+
sandbox: value.sandbox,
|
|
192
|
+
harnesses: value.harnesses,
|
|
193
|
+
tasks: value.tasks.map(({ runId, taskId }) => ({ runId, taskId })),
|
|
194
|
+
});
|
|
195
|
+
if (signature(previous) !== signature(input)) {
|
|
196
|
+
sendJson(response, 409, { error: "This request ID belongs to a different selection" });
|
|
197
|
+
return true;
|
|
198
|
+
}
|
|
199
|
+
input = previous;
|
|
200
|
+
}
|
|
201
|
+
else {
|
|
202
|
+
await artifacts.put(requestKey, Buffer.from(JSON.stringify(input)), "application/json");
|
|
203
|
+
}
|
|
204
|
+
const run = initialEvaluation(input, model.label);
|
|
205
|
+
await saveEvaluation(artifacts, run);
|
|
206
|
+
await options.start(input);
|
|
207
|
+
sendJson(response, 202, run);
|
|
208
|
+
return true;
|
|
209
|
+
},
|
|
210
|
+
};
|
|
211
|
+
}
|
|
212
|
+
//# sourceMappingURL=routes.js.map
|
|
@@ -0,0 +1 @@
|
|
|
1
|
+
{"version":3,"file":"routes.js","sourceRoot":"","sources":["../../src/evaluation/routes.ts"],"names":[],"mappings":"AACA,OAAO,EAAE,CAAC,EAAE,MAAM,KAAK,CAAC;AACxB,OAAO,EAAE,QAAQ,EAAE,QAAQ,EAAE,eAAe,EAAE,MAAM,gBAAgB,CAAC;AAKrE,OAAO,EAAE,SAAS,EAAE,MAAM,mBAAmB,CAAC;AAE9C,OAAO,EAAE,gBAAgB,EAAE,MAAM,yBAAyB,CAAC;AAC3D,OAAO,EAAE,mBAAmB,EAAE,MAAM,aAAa,CAAC;AAElD,OAAO,EAAE,UAAU,EAAE,MAAM,gBAAgB,CAAC;AAC5C,OAAO,EAAE,mBAAmB,EAAE,MAAM,4BAA4B,CAAC;AACjE,OAAO,EAAE,cAAc,EAAE,MAAM,sBAAsB,CAAC;AACtD,OAAO,EAAE,gBAAgB,EAAE,WAAW,EAAE,MAAM,qBAAqB,CAAC;AACpE,OAAO,EACL,gBAAgB,EAChB,aAAa,EACb,iBAAiB,EACjB,eAAe,EACf,cAAc,GACf,MAAM,YAAY,CAAC;AAGpB,MAAM,KAAK,GACT,mJAAmJ,CAAC;AACtJ,MAAM,aAAa,GAAG,CAAC;KACpB,MAAM,CAAC;IACN,EAAE,EAAE,CAAC,CAAC,IAAI,EAAE,CAAC,SAAS,CAAC,CAAC,KAAK,EAAE,EAAE,CAAC,KAAK,CAAC,WAAW,EAAE,CAAC;IACtD,KAAK,EAAE,CAAC,CAAC,MAAM,EAAE,CAAC,GAAG,CAAC,CAAC,CAAC,CAAC,GAAG,CAAC,EAAE,CAAC;IAChC,SAAS,EAAE,CAAC,CAAC,KAAK,CAAC,CAAC,CAAC,IAAI,CAAC,UAAU,CAAC,CAAC,CAAC,GAAG,CAAC,CAAC,CAAC,CAAC,GAAG,CAAC,UAAU,CAAC,MAAM,CAAC;IACpE,OAAO,EAAE,CAAC,CAAC,IAAI,CAAC,mBAAmB,CAAC;IACpC,KAAK,EAAE,CAAC;SACL,KAAK,CACJ,CAAC;SACE,MAAM,CAAC,EAAE,KAAK,EAAE,CAAC,CAAC,MAAM,EAAE,CAAC,GAAG,CAAC,CAAC,CAAC,CAAC,GAAG,CAAC,GAAG,CAAC,EAAE,MAAM,EAAE,CAAC,CAAC,MAAM,EAAE,CAAC,GAAG,CAAC,CAAC,CAAC,CAAC,GAAG,CAAC,GAAG,CAAC,EAAE,CAAC;SACjF,MAAM,EAAE,CACZ;SACA,GAAG,CAAC,CAAC,CAAC;CACV,CAAC;KACD,MAAM,EAAE,CAAC;AAYZ,MAAM,UAAU,sBAAsB,CAAC,OAAgC;IACrE,MAAM,EAAE,KAAK,EAAE,KAAK,EAAE,KAAK,EAAE,SAAS,EAAE,GAAG,OAAO,CAAC;IACnD,OAAO;QACL,KAAK,CAAC,MAAM,CACV,OAAwB,EACxB,GAAQ,EACR,QAAwB,EACxB,IAAU;YAEV,IAAI,MAAM,gBAAgB,CAAC,OAAO,EAAE,OAAO,EAAE,GAAG,EAAE,QAAQ,EAAE,IAAI,CAAC;gBAAE,OAAO,IAAI,CAAC;YAC/E,IAAI,MAAM,mBAAmB,CAAC,OAAO,EAAE,OAAO,EAAE,GAAG,EAAE,QAAQ,EAAE,IAAI,CAAC;gBAAE,OAAO,IAAI,CAAC;YAClF,IAAI,MAAM,cAAc,CAAC,OAAO,EAAE,OAAO,EAAE,GAAG,EAAE,QAAQ,EAAE,IAAI,CAAC;gBAAE,OAAO,IAAI,CAAC;YAC7E,MAAM,KAAK,GAAG,KAAK,CAAC,IAAI,CAAC,GAAG,CAAC,QAAQ,CAAC,CAAC;YACvC,IAAI,CAAC,KAAK,EAAE,CAAC,CAAC,CAAC,IAAI,CAAC,KAAK,CAAC,CAAC,CAAC,IAAI,CAAC,KAAK,CAAC,CAAC,CAAC;gBAAE,OAAO,KAAK,CAAC;YACxD,QAAQ,CAAC,SAAS,CAAC,eAAe,EAAE,UAAU,CAAC,CAAC;YAChD,MAAM,MAAM,GAAG,MAAM,SAAS,CAAC,KAAK,EAAE,IAAI,EAAE,KAAK,CAAC,CAAC,CAAC,CAAC,CAAC;YACtD,MAAM,IAAI,GAAG,MAAM,CAAC,CAAC,CAAC,MAAM,KAAK,CAAC,IAAI,CAAC,MAAM,CAAC,EAAE,EAAE,GAAG,KAAK,CAAC,CAAC,CAAC,IAAI,KAAK,CAAC,CAAC,CAAC,EAAE,CAAC,CAAC,CAAC,CAAC,SAAS,CAAC;YACzF,IAAI,CAAC,MAAM,IAAI,CAAC,IAAI,EAAE,CAAC;gBACrB,QAAQ,CAAC,QAAQ,EAAE,GAAG,EAAE,EAAE,KAAK,EAAE,kCAAkC,EAAE,CAAC,CAAC;gBACvE,OAAO,IAAI,CAAC;YACd,CAAC;YACD,MAAM,GAAG,GAAG,OAAO,CAAC,GAAG,IAAI,OAAO,CAAC,GAAG,CAAC;YACvC,IAAI,KAAK,CAAC,CAAC,CAAC,KAAK,UAAU,EAAE,CAAC;gBAC5B,MAAM,WAAW,CACf,OAAO,EACP,QAAQ,EACR,SAAS,EACT;oBACE,MAAM,EAAE,IAAI,CAAC,EAAE;oBACf,OAAO,EAAE,IAAI,CAAC,QAAQ;oBACtB,MAAM,EAAE,MAAM,CAAC,KAAK;oBACpB,SAAS,EAAE,OAAO,CAAC,SAAS;oBAC5B,OAAO,EAAE,eAAe,CAAC,OAAO,EAAE,OAAO,CAAC,SAAS,EAAE,IAAI,CAAC;iBAC3D,EACD,GAAG,CACJ,CAAC;gBACF,OAAO,IAAI,CAAC;YACd,CAAC;YACD,IAAI,OAAO,CAAC,MAAM,KAAK,KAAK,EAAE,CAAC;gBAC7B,IAAI,KAAK,CAAC,CAAC,CAAC,KAAK,SAAS,EAAE,CAAC;oBAC3B,MAAM,SAAS,GAAG,CAAC,MAAM,KAAK,CAAC,WAAW,CAAC,IAAI,CAAC,EAAE,CAAC,CAAC,CAAC,MAAM,CACzD,CAAC,IAAI,EAAE,EAAE,CACP,IAAI,CAAC,SAAS;wBACd,IAAI,CAAC,cAAc,KAAK,UAAU;wBAClC,IAAI,CAAC,MAAM,EAAE,QAAQ,KAAK,SAAS,CACtC,CAAC;oBACF,QAAQ,CAAC,QAAQ,EAAE,GAAG,EAAE;wBACtB,GAAG,CAAC,MAAM,gBAAgB,CAAC,SAAS,EAAE,IAAI,CAAC,EAAE,EAAE,IAAI,CAAC,QAAQ,EAAE,MAAM,CAAC,KAAK,EAAE,GAAG,CAAC,CAAC;wBACjF,KAAK,EAAE,SAAS,CAAC,GAAG,CAAC,CAAC,IAAI,EAAE,EAAE,CAAC,CAAC;4BAC9B,KAAK,EAAE,IAAI,CAAC,KAAK;4BACjB,MAAM,EAAE,IAAI,CAAC,MAAM;4BACnB,UAAU,EAAE,IAAI,CAAC,UAAU;yBAC5B,CAAC,CAAC;qBACJ,CAAC,CAAC;gBACL,CAAC;qBAAM,IAAI,KAAK,CAAC,CAAC,CAAC,EAAE,CAAC;oBACpB,MAAM,GAAG,GAAG,MAAM,aAAa,CAAC,SAAS,EAAE,IAAI,CAAC,EAAE,EAAE,KAAK,CAAC,CAAC,CAAC,CAAC,CAAC;oBAC9D,IAAI,CAAC,GAAG,EAAE,CAAC;wBACT,QAAQ,CAAC,QAAQ,EAAE,GAAG,EAAE,EAAE,KAAK,EAAE,sBAAsB,EAAE,CAAC,CAAC;wBAC3D,OAAO,IAAI,CAAC;oBACd,CAAC;oBACD,IAAI,GAAG,CAAC,QAAQ,CAAC,QAAQ,CAAC,YAAY,CAAC,EAAE,CAAC;wBACxC,MAAM,IAAI,GAAG,GAAG,CAAC,YAAY,CAAC,GAAG,CAAC,MAAM,CAAC,IAAI,EAAE,CAAC;wBAChD,IAAI,CAAC,GAAG,CAAC,MAAM,CAAC,IAAI,CAAC,CAAC,KAAK,EAAE,EAAE,CAAC,KAAK,CAAC,SAAS,CAAC,QAAQ,CAAC,IAAI,CAAC,CAAC,EAAE,CAAC;4BAChE,QAAQ,CAAC,QAAQ,EAAE,GAAG,EAAE,EAAE,KAAK,EAAE,oBAAoB,EAAE,CAAC,CAAC;4BACzD,OAAO,IAAI,CAAC;wBACd,CAAC;wBACD,MAAM,KAAK,GAAG,MAAM,SAAS,CAAC,QAAQ,CACpC,GAAG,gBAAgB,CAAC,IAAI,CAAC,EAAE,EAAE,GAAG,CAAC,EAAE,CAAC,aAAa,IAAI,EAAE,CACxD,CAAC;wBACF,IAAI,CAAC,KAAK,EAAE,CAAC;4BACX,QAAQ,CAAC,QAAQ,EAAE,GAAG,EAAE,EAAE,KAAK,EAAE,oBAAoB,EAAE,CAAC,CAAC;4BACzD,OAAO,IAAI,CAAC;wBACd,CAAC;wBACD,QAAQ,CAAC,SAAS,CAAC,GAAG,EAAE;4BACtB,cAAc,EAAE,2BAA2B;4BAC3C,wBAAwB,EAAE,SAAS;4BACnC,qBAAqB,EAAE,0CAA0C;yBAClE,CAAC,CAAC;wBACH,QAAQ,CAAC,GAAG,CAAC,KAAK,CAAC,CAAC;oBACtB,CAAC;;wBAAM,QAAQ,CAAC,QAAQ,EAAE,GAAG,EAAE,GAAG,CAAC,CAAC;gBACtC,CAAC;qBAAM,CAAC;oBACN,QAAQ,CAAC,QAAQ,EAAE,GAAG,EAAE;wBACtB,IAAI,EAAE,CAAC,MAAM,eAAe,CAAC,SAAS,EAAE,IAAI,CAAC,EAAE,CAAC,CAAC,CAAC,GAAG,CAAC,CAAC,GAAG,EAAE,EAAE,CAAC,CAAC;4BAC9D,GAAG,GAAG;4BACN,MAAM,EAAE,GAAG,CAAC,MAAM,CAAC,GAAG,CAAC,CAAC,KAAK,EAAE,EAAE,CAAC,CAAC,EAAE,GAAG,KAAK,EAAE,GAAG,EAAE,EAAE,EAAE,KAAK,EAAE,EAAE,EAAE,SAAS,EAAE,EAAE,EAAE,CAAC,CAAC;yBACrF,CAAC,CAAC;qBACJ,CAAC,CAAC;gBACL,CAAC;gBACD,OAAO,IAAI,CAAC;YACd,CAAC;YACD,IAAI,OAAO,CAAC,MAAM,KAAK,MAAM,IAAI,KAAK,CAAC,CAAC,CAAC,EAAE,CAAC;gBAC1C,QAAQ,CAAC,QAAQ,EAAE,GAAG,EAAE,EAAE,KAAK,EAAE,oBAAoB,EAAE,CAAC,CAAC;gBACzD,OAAO,IAAI,CAAC;YACd,CAAC;YACD,IAAI,CAAC,eAAe,CAAC,OAAO,EAAE,OAAO,CAAC,SAAS,EAAE,IAAI,CAAC,EAAE,CAAC;gBACvD,QAAQ,CAAC,QAAQ,EAAE,GAAG,EAAE,EAAE,KAAK,EAAE,mCAAmC,EAAE,CAAC,CAAC;gBACxE,OAAO,IAAI,CAAC;YACd,CAAC;YACD,IAAI,IAAmC,CAAC;YACxC,IAAI,CAAC;gBACH,IAAI,GAAG,aAAa,CAAC,KAAK,CAAC,IAAI,CAAC,KAAK,CAAC,CAAC,MAAM,QAAQ,CAAC,OAAO,EAAE,MAAM,CAAC,CAAC,CAAC,QAAQ,CAAC,MAAM,CAAC,CAAC,CAAC,CAAC;YAC7F,CAAC;YAAC,MAAM,CAAC;gBACP,QAAQ,CAAC,QAAQ,EAAE,GAAG,EAAE,EAAE,KAAK,EAAE,8BAA8B,EAAE,CAAC,CAAC;gBACnE,OAAO,IAAI,CAAC;YACd,CAAC;YACD,MAAM,OAAO,GAAG,MAAM,gBAAgB,CAAC,SAAS,EAAE,IAAI,CAAC,EAAE,EAAE,IAAI,CAAC,QAAQ,EAAE,MAAM,CAAC,KAAK,EAAE,GAAG,CAAC,CAAC;YAC7F,MAAM,KAAK,GAAG,OAAO,CAAC,MAAM,CAAC,IAAI,CAAC,CAAC,SAAS,EAAE,EAAE,CAAC,SAAS,CAAC,EAAE,KAAK,IAAI,CAAC,KAAK,CAAC,CAAC;YAC9E,IACE,CAAC,KAAK;gBACN,CAAC,KAAK,CAAC,OAAO;oBACZ,CAAC,CAAC,KAAK,CAAC,OAAO,KAAK,IAAI,CAAC,OAAO;oBAChC,CAAC,CAAC,CAAC,OAAO,CAAC,SAAS,CAAC,QAAQ,CAAC,IAAI,CAAC,OAAO,CAAC,CAAC;gBAC9C,IAAI,CAAC,SAAS,CAAC,IAAI,CAAC,CAAC,OAAO,EAAE,EAAE,CAAC,CAAC,KAAK,CAAC,SAAS,CAAC,QAAQ,CAAC,OAAO,CAAC,CAAC;gBACpE,IAAI,GAAG,CAAC,IAAI,CAAC,SAAS,CAAC,CAAC,IAAI,KAAK,IAAI,CAAC,SAAS,CAAC,MAAM;gBACtD,IAAI,GAAG,CAAC,IAAI,CAAC,KAAK,CAAC,GAAG,CAAC,CAAC,IAAI,EAAE,EAAE,CAAC,GAAG,IAAI,CAAC,KAAK,IAAI,IAAI,CAAC,MAAM,EAAE,CAAC,CAAC,CAAC,IAAI,KAAK,IAAI,CAAC,KAAK,CAAC,MAAM,EAC5F,CAAC;gBACD,QAAQ,CAAC,QAAQ,EAAE,GAAG,EAAE;oBACtB,KAAK,EAAE,+DAA+D;iBACvE,CAAC,CAAC;gBACH,OAAO,IAAI,CAAC;YACd,CAAC;YACD,MAAM,QAAQ,GAAG,MAAM,OAAO,CAAC,GAAG,CAChC,IAAI,CAAC,KAAK,CAAC,GAAG,CAAC,CAAC,IAAI,EAAE,EAAE,CAAC,KAAK,CAAC,IAAI,CAAC,IAAI,CAAC,EAAE,EAAE,IAAI,CAAC,KAAK,EAAE,IAAI,CAAC,MAAM,CAAC,CAAC,CACvE,CAAC;YACF,IACE,QAAQ,CAAC,IAAI,CACX,CAAC,IAAI,EAAE,EAAE,CACP,CAAC,IAAI,EAAE,SAAS;gBAChB,IAAI,CAAC,cAAc,KAAK,UAAU;gBAClC,IAAI,CAAC,MAAM,EAAE,QAAQ,KAAK,SAAS,CACtC,EACD,CAAC;gBACD,QAAQ,CAAC,QAAQ,EAAE,GAAG,EAAE;oBACtB,KAAK,EACH,mFAAmF;iBACtF,CAAC,CAAC;gBACH,OAAO,IAAI,CAAC;YACd,CAAC;YACD,IACE,IAAI,GAAG,CAAC,QAAQ,CAAC,GAAG,CAAC,CAAC,IAAI,EAAE,EAAE,CAAC,GAAG,IAAI,EAAE,KAAK,IAAI,IAAI,EAAE,MAAM,EAAE,CAAC,CAAC,CAAC,IAAI,KAAK,QAAQ,CAAC,MAAM,EAC1F,CAAC;gBACD,QAAQ,CAAC,QAAQ,EAAE,GAAG,EAAE,EAAE,KAAK,EAAE,4BAA4B,EAAE,CAAC,CAAC;gBACjE,OAAO,IAAI,CAAC;YACd,CAAC;YACD,MAAM,QAAQ,GAAG,MAAM,aAAa,CAAC,SAAS,EAAE,IAAI,CAAC,EAAE,EAAE,IAAI,CAAC,EAAE,CAAC,CAAC;YAClE,IAAI,QAAQ,EAAE,CAAC;gBACb,IACE,QAAQ,CAAC,KAAK,KAAK,IAAI,CAAC,KAAK;oBAC7B,QAAQ,CAAC,OAAO,KAAK,IAAI,CAAC,OAAO;oBACjC,IAAI,CAAC,SAAS,CAAC,QAAQ,CAAC,SAAS,CAAC,KAAK,IAAI,CAAC,SAAS,CAAC,IAAI,CAAC,SAAS,CAAC;oBACrE,IAAI,CAAC,SAAS,CACZ,QAAQ,CAAC,MAAM,CAAC,GAAG,CAAC,CAAC,KAAK,EAAE,EAAE,CAAC,GAAG,KAAK,CAAC,KAAK,IAAI,KAAK,CAAC,MAAM,IAAI,KAAK,CAAC,OAAO,EAAE,CAAC,CAClF;wBACC,IAAI,CAAC,SAAS,CACZ,IAAI,CAAC,KAAK,CAAC,OAAO,CAAC,CAAC,IAAI,EAAE,EAAE,CAC1B,IAAI,CAAC,SAAS,CAAC,GAAG,CAAC,CAAC,OAAO,EAAE,EAAE,CAAC,GAAG,IAAI,CAAC,KAAK,IAAI,IAAI,CAAC,MAAM,IAAI,OAAO,EAAE,CAAC,CAC3E,CACF,EACH,CAAC;oBACD,QAAQ,CAAC,QAAQ,EAAE,GAAG,EAAE,EAAE,KAAK,EAAE,kDAAkD,EAAE,CAAC,CAAC;oBACvF,OAAO,IAAI,CAAC;gBACd,CAAC;gBACD,IAAI,QAAQ,CAAC,MAAM,KAAK,QAAQ,EAAE,CAAC;oBACjC,MAAM,KAAK,GAAG,MAAM,SAAS,CAAC,QAAQ,CACpC,GAAG,gBAAgB,CAAC,IAAI,CAAC,EAAE,EAAE,IAAI,CAAC,EAAE,CAAC,cAAc,CACpD,CAAC;oBACF,IAAI,CAAC,KAAK;wBAAE,MAAM,IAAI,KAAK,CAAC,wCAAwC,CAAC,CAAC;oBACtE,MAAM,OAAO,CAAC,KAAK,CAAC,IAAI,CAAC,KAAK,CAAC,MAAM,CAAC,IAAI,CAAC,KAAK,CAAC,CAAC,QAAQ,CAAC,MAAM,CAAC,CAAoB,CAAC,CAAC;gBAC1F,CAAC;gBACD,QAAQ,CAAC,QAAQ,EAAE,GAAG,EAAE,QAAQ,CAAC,CAAC;gBAClC,OAAO,IAAI,CAAC;YACd,CAAC;YACD,IAAI,KAAK,GAAoB;gBAC3B,GAAG,IAAI;gBACP,SAAS,EAAE,KAAK,CAAC,KAAK;gBACtB,GAAG,CAAC,KAAK,CAAC,EAAE,CAAC,UAAU,CAAC,QAAQ,CAAC,CAAC,CAAC,CAAC,EAAE,iBAAiB,EAAE,IAAI,CAAC,QAAQ,EAAE,CAAC,CAAC,CAAC,EAAE,CAAC;gBAC9E,GAAG,CAAC,KAAK,CAAC,OAAO,CAAC,CAAC,CAAC,EAAE,OAAO,EAAE,KAAK,CAAC,OAAO,EAAE,CAAC,CAAC,CAAC,EAAE,CAAC;gBACpD,MAAM,EAAE,IAAI,CAAC,EAAE;gBACf,MAAM,EAAE,MAAM,CAAC,KAAK;gBACpB,SAAS,EAAE,IAAI,CAAC,KAAK;gBACrB,SAAS,EAAE,IAAI,IAAI,EAAE,CAAC,WAAW,EAAE;gBACnC,KAAK,EAAE,QAAQ,CAAC,GAAG,CAAC,CAAC,IAAI,EAAE,EAAE;oBAC3B,IAAI,CAAC,IAAI,EAAE,SAAS;wBAAE,MAAM,IAAI,KAAK,CAAC,wBAAwB,CAAC,CAAC;oBAChE,OAAO,EAAE,KAAK,EAAE,IAAI,CAAC,KAAK,EAAE,MAAM,EAAE,IAAI,CAAC,MAAM,EAAE,SAAS,EAAE,IAAI,CAAC,SAAS,EAAE,CAAC;gBAC/E,CAAC,CAAC;aACH,CAAC;YACF,MAAM,UAAU,GAAG,GAAG,gBAAgB,CAAC,IAAI,CAAC,EAAE,EAAE,IAAI,CAAC,EAAE,CAAC,cAAc,CAAC;YACvE,MAAM,YAAY,GAAG,MAAM,SAAS,CAAC,QAAQ,CAAC,UAAU,CAAC,CAAC;YAC1D,IAAI,YAAY,EAAE,CAAC;gBACjB,MAAM,QAAQ,GAAG,IAAI,CAAC,KAAK,CAAC,MAAM,CAAC,IAAI,CAAC,YAAY,CAAC,CAAC,QAAQ,CAAC,MAAM,CAAC,CAAoB,CAAC;gBAC3F,MAAM,SAAS,GAAG,CAAC,KAAsB,EAAE,EAAE,CAC3C,IAAI,CAAC,SAAS,CAAC;oBACb,KAAK,EAAE,KAAK,CAAC,KAAK;oBAClB,SAAS,EAAE,KAAK,CAAC,SAAS;oBAC1B,OAAO,EAAE,KAAK,CAAC,OAAO;oBACtB,SAAS,EAAE,KAAK,CAAC,SAAS;oBAC1B,KAAK,EAAE,KAAK,CAAC,KAAK,CAAC,GAAG,CAAC,CAAC,EAAE,KAAK,EAAE,MAAM,EAAE,EAAE,EAAE,CAAC,CAAC,EAAE,KAAK,EAAE,MAAM,EAAE,CAAC,CAAC;iBACnE,CAAC,CAAC;gBACL,IAAI,SAAS,CAAC,QAAQ,CAAC,KAAK,SAAS,CAAC,KAAK,CAAC,EAAE,CAAC;oBAC7C,QAAQ,CAAC,QAAQ,EAAE,GAAG,EAAE,EAAE,KAAK,EAAE,kDAAkD,EAAE,CAAC,CAAC;oBACvF,OAAO,IAAI,CAAC;gBACd,CAAC;gBACD,KAAK,GAAG,QAAQ,CAAC;YACnB,CAAC;iBAAM,CAAC;gBACN,MAAM,SAAS,CAAC,GAAG,CAAC,UAAU,EAAE,MAAM,CAAC,IAAI,CAAC,IAAI,CAAC,SAAS,CAAC,KAAK,CAAC,CAAC,EAAE,kBAAkB,CAAC,CAAC;YAC1F,CAAC;YACD,MAAM,GAAG,GAAG,iBAAiB,CAAC,KAAK,EAAE,KAAK,CAAC,KAAK,CAAC,CAAC;YAClD,MAAM,cAAc,CAAC,SAAS,EAAE,GAAG,CAAC,CAAC;YACrC,MAAM,OAAO,CAAC,KAAK,CAAC,KAAK,CAAC,CAAC;YAC3B,QAAQ,CAAC,QAAQ,EAAE,GAAG,EAAE,GAAG,CAAC,CAAC;YAC7B,OAAO,IAAI,CAAC;QACd,CAAC;KACF,CAAC;AACJ,CAAC"}
|
|
@@ -0,0 +1,16 @@
|
|
|
1
|
+
import type { ArtifactStore } from "../artifacts.js";
|
|
2
|
+
import { runCommand } from "../process.js";
|
|
3
|
+
import type { EncryptedRecordStore } from "./encrypted-records.js";
|
|
4
|
+
import { type ThinkingLevel } from "./model-options.js";
|
|
5
|
+
import type { EvaluationInput, Harness } from "./types.js";
|
|
6
|
+
export declare function solverArguments(taskPath: string, jobs: string, harness: Harness, model: string, sandbox: string, thinking?: ThinkingLevel): string[];
|
|
7
|
+
export interface RunnerOptions {
|
|
8
|
+
command?: typeof runCommand;
|
|
9
|
+
env?: NodeJS.ProcessEnv;
|
|
10
|
+
signal?: AbortSignal;
|
|
11
|
+
heartbeat?: () => void;
|
|
12
|
+
pollMs?: number;
|
|
13
|
+
records?: EncryptedRecordStore;
|
|
14
|
+
}
|
|
15
|
+
export declare function executeEvaluation(store: ArtifactStore, input: EvaluationInput, options?: RunnerOptions): Promise<void>;
|
|
16
|
+
//# sourceMappingURL=runner.d.ts.map
|
|
@@ -0,0 +1 @@
|
|
|
1
|
+
{"version":3,"file":"runner.d.ts","sourceRoot":"","sources":["../../src/evaluation/runner.ts"],"names":[],"mappings":"AAKA,OAAO,KAAK,EAAE,aAAa,EAAE,MAAM,iBAAiB,CAAC;AACrD,OAAO,EAAE,UAAU,EAAE,MAAM,eAAe,CAAC;AAI3C,OAAO,KAAK,EAAE,oBAAoB,EAAE,MAAM,wBAAwB,CAAC;AAEnE,OAAO,EAAE,KAAK,aAAa,EAAqB,MAAM,oBAAoB,CAAC;AAY3E,OAAO,KAAK,EAAE,eAAe,EAAkC,OAAO,EAAE,MAAM,YAAY,CAAC;AAE3F,wBAAgB,eAAe,CAC7B,QAAQ,EAAE,MAAM,EAChB,IAAI,EAAE,MAAM,EACZ,OAAO,EAAE,OAAO,EAChB,KAAK,EAAE,MAAM,EACb,OAAO,EAAE,MAAM,EACf,QAAQ,CAAC,EAAE,aAAa,GACvB,MAAM,EAAE,CAyBV;AACD,MAAM,WAAW,aAAa;IAC5B,OAAO,CAAC,EAAE,OAAO,UAAU,CAAC;IAC5B,GAAG,CAAC,EAAE,MAAM,CAAC,UAAU,CAAC;IACxB,MAAM,CAAC,EAAE,WAAW,CAAC;IACrB,SAAS,CAAC,EAAE,MAAM,IAAI,CAAC;IACvB,MAAM,CAAC,EAAE,MAAM,CAAC;IAChB,OAAO,CAAC,EAAE,oBAAoB,CAAC;CAChC;AACD,wBAAsB,iBAAiB,CACrC,KAAK,EAAE,aAAa,EACpB,KAAK,EAAE,eAAe,EACtB,OAAO,GAAE,aAAkB,GAC1B,OAAO,CAAC,IAAI,CAAC,CAyGf"}
|
|
@@ -0,0 +1,241 @@
|
|
|
1
|
+
import { mkdir, mkdtemp, readFile, rm, writeFile } from "node:fs/promises";
|
|
2
|
+
import { tmpdir } from "node:os";
|
|
3
|
+
import { dirname, join } from "node:path";
|
|
4
|
+
import { fileURLToPath } from "node:url";
|
|
5
|
+
import { extractRegularArchive } from "../archive.js";
|
|
6
|
+
import { runCommand } from "../process.js";
|
|
7
|
+
import { HARBOR_VERSION, solverEnvironment } from "./config.js";
|
|
8
|
+
import { trialCost } from "./cost.js";
|
|
9
|
+
import { credentialExecution } from "./credential-execution.js";
|
|
10
|
+
import { gatewayTrial, solverAgent } from "./gateway-execution.js";
|
|
11
|
+
import { thinkingArguments } from "./model-options.js";
|
|
12
|
+
import { boundedSteps, collectOutput, completeLines, piSteps, record, redactOutput, trajectorySteps, } from "./output.js";
|
|
13
|
+
import { profileEnvironment } from "./profiles.js";
|
|
14
|
+
import { evaluationPrefix, getEvaluation, saveEvaluation } from "./store.js";
|
|
15
|
+
export function solverArguments(taskPath, jobs, harness, model, sandbox, thinking) {
|
|
16
|
+
return [
|
|
17
|
+
"run",
|
|
18
|
+
"--path",
|
|
19
|
+
taskPath,
|
|
20
|
+
"--agent",
|
|
21
|
+
solverAgent(harness, model),
|
|
22
|
+
"--model",
|
|
23
|
+
model,
|
|
24
|
+
"--env",
|
|
25
|
+
sandbox === "e2b" ? "harbor_e2b:SelfBenchE2BEnvironment" : sandbox,
|
|
26
|
+
"--jobs-dir",
|
|
27
|
+
jobs,
|
|
28
|
+
"--job-name",
|
|
29
|
+
"solver",
|
|
30
|
+
"--n-attempts",
|
|
31
|
+
"1",
|
|
32
|
+
"--n-concurrent",
|
|
33
|
+
"1",
|
|
34
|
+
"--max-retries",
|
|
35
|
+
"0",
|
|
36
|
+
"--delete",
|
|
37
|
+
"--yes",
|
|
38
|
+
...thinkingArguments(harness, thinking),
|
|
39
|
+
];
|
|
40
|
+
}
|
|
41
|
+
export async function executeEvaluation(store, input, options = {}) {
|
|
42
|
+
const run = await getEvaluation(store, input.repoId, input.id);
|
|
43
|
+
if (!run)
|
|
44
|
+
throw new Error("Evaluation record is missing");
|
|
45
|
+
if (run.status !== "queued")
|
|
46
|
+
throw new Error("Evaluation already attempted; refusing to repeat model spend");
|
|
47
|
+
const root = await mkdtemp(join(tmpdir(), "selfbench-evaluation-"));
|
|
48
|
+
const command = options.command ?? runCommand;
|
|
49
|
+
const environment = options.env ?? process.env;
|
|
50
|
+
const secrets = Object.entries(environment)
|
|
51
|
+
.filter(([name]) => /SECRET|TOKEN|PASSWORD|API_KEY/.test(name))
|
|
52
|
+
.map(([, value]) => value ?? "")
|
|
53
|
+
.filter(Boolean);
|
|
54
|
+
const redact = (text) => redactOutput(text, secrets);
|
|
55
|
+
run.status = "running";
|
|
56
|
+
try {
|
|
57
|
+
await saveEvaluation(store, run);
|
|
58
|
+
const home = join(root, "home");
|
|
59
|
+
await mkdir(home, { mode: 0o700 });
|
|
60
|
+
const resolved = input.credentials
|
|
61
|
+
? environment
|
|
62
|
+
: await profileEnvironment(store, input, environment);
|
|
63
|
+
secrets.push(...Object.entries(resolved)
|
|
64
|
+
.filter(([name]) => name.startsWith("SELFBENCH_EVAL_SECRET_"))
|
|
65
|
+
.map(([, value]) => value ?? "")
|
|
66
|
+
.filter(Boolean));
|
|
67
|
+
if (input.credentials && !options.records)
|
|
68
|
+
throw new Error("Credential storage unavailable on the worker");
|
|
69
|
+
const execution = input.credentials && options.records
|
|
70
|
+
? await credentialExecution(input, home, environment, options.records)
|
|
71
|
+
: { ...solverEnvironment(input, home, resolved), secrets: [] };
|
|
72
|
+
secrets.push(...execution.secrets);
|
|
73
|
+
const { profile, child } = execution;
|
|
74
|
+
child.PYTHONPATH = dirname(fileURLToPath(import.meta.url));
|
|
75
|
+
const version = await command("harbor", ["--version"], { env: child, timeoutMs: 15_000 });
|
|
76
|
+
if (version.stdout.trim() !== HARBOR_VERSION)
|
|
77
|
+
throw new Error("Worker Harbor version does not match the supported version");
|
|
78
|
+
for (const [index, trial] of run.trials.entries()) {
|
|
79
|
+
options.signal?.throwIfAborted();
|
|
80
|
+
const task = input.tasks.find((candidate) => candidate.runId === trial.runId && candidate.taskId === trial.taskId);
|
|
81
|
+
if (!task)
|
|
82
|
+
throw new Error("Evaluation task snapshot is missing");
|
|
83
|
+
const trialRoot = join(root, String(index));
|
|
84
|
+
await mkdir(trialRoot);
|
|
85
|
+
trial.status = "running";
|
|
86
|
+
trial.startedAt = new Date().toISOString();
|
|
87
|
+
await saveEvaluation(store, run);
|
|
88
|
+
try {
|
|
89
|
+
const bundle = await store.getByKey(task.bundleKey);
|
|
90
|
+
if (!bundle || bundle.byteLength > 100 * 1024 * 1024)
|
|
91
|
+
throw new Error("Task bundle is missing or exceeds 100 MiB");
|
|
92
|
+
const archive = join(trialRoot, "task.tar.gz");
|
|
93
|
+
await writeFile(archive, bundle, { mode: 0o600 });
|
|
94
|
+
const extracted = join(trialRoot, "task");
|
|
95
|
+
await mkdir(extracted);
|
|
96
|
+
await extractRegularArchive(archive, extracted, options.signal ? { signal: options.signal } : {});
|
|
97
|
+
const taskPath = await readFile(join(extracted, "harbor-task", "task.toml")).then(() => join(extracted, "harbor-task"), () => extracted);
|
|
98
|
+
await readFile(join(taskPath, "task.toml"));
|
|
99
|
+
await runTrial({
|
|
100
|
+
store,
|
|
101
|
+
run,
|
|
102
|
+
trial,
|
|
103
|
+
index,
|
|
104
|
+
taskPath,
|
|
105
|
+
jobs: join(trialRoot, "jobs"),
|
|
106
|
+
...gatewayTrial(input, trial.harness, profile.model, child),
|
|
107
|
+
command,
|
|
108
|
+
redact,
|
|
109
|
+
options,
|
|
110
|
+
});
|
|
111
|
+
}
|
|
112
|
+
catch (error) {
|
|
113
|
+
trial.status = "failed";
|
|
114
|
+
trial.error = redact(error instanceof Error ? error.message : "Solver failed");
|
|
115
|
+
}
|
|
116
|
+
trial.finishedAt = new Date().toISOString();
|
|
117
|
+
await saveEvaluation(store, run);
|
|
118
|
+
}
|
|
119
|
+
run.status = run.trials.some((trial) => trial.status === "failed") ? "failed" : "completed";
|
|
120
|
+
}
|
|
121
|
+
catch (error) {
|
|
122
|
+
run.status = "failed";
|
|
123
|
+
run.error = redact(error instanceof Error ? error.message : "Evaluation failed");
|
|
124
|
+
}
|
|
125
|
+
finally {
|
|
126
|
+
for (const trial of run.trials) {
|
|
127
|
+
if (trial.status === "queued" || trial.status === "running") {
|
|
128
|
+
trial.status = "failed";
|
|
129
|
+
trial.error = "Not completed because the evaluation stopped";
|
|
130
|
+
}
|
|
131
|
+
}
|
|
132
|
+
run.finishedAt = new Date().toISOString();
|
|
133
|
+
try {
|
|
134
|
+
await saveEvaluation(store, run);
|
|
135
|
+
}
|
|
136
|
+
finally {
|
|
137
|
+
await rm(root, { recursive: true, force: true });
|
|
138
|
+
}
|
|
139
|
+
}
|
|
140
|
+
}
|
|
141
|
+
async function runTrial(context) {
|
|
142
|
+
const { store, run, trial, index, taskPath, jobs, model, child, command, redact, options } = context;
|
|
143
|
+
let stdout = "";
|
|
144
|
+
let polling = Promise.resolve();
|
|
145
|
+
let refreshing = false;
|
|
146
|
+
let lastSnapshot = "";
|
|
147
|
+
const refresh = async (archive) => {
|
|
148
|
+
const files = await collectOutput(jobs);
|
|
149
|
+
const clean = new Map([...files].map(([name, text]) => [
|
|
150
|
+
name,
|
|
151
|
+
redact(!archive && !name.endsWith(".json") ? completeLines(text) : text),
|
|
152
|
+
]));
|
|
153
|
+
const trajectory = [...clean].find(([name]) => name.endsWith("/trajectory.json"));
|
|
154
|
+
const pi = [...clean].find(([name]) => name.endsWith("/pi.txt"));
|
|
155
|
+
if (trajectory) {
|
|
156
|
+
try {
|
|
157
|
+
trial.steps = trajectorySteps(JSON.parse(trajectory[1]));
|
|
158
|
+
}
|
|
159
|
+
catch {
|
|
160
|
+
trial.steps = [];
|
|
161
|
+
}
|
|
162
|
+
}
|
|
163
|
+
else if (pi)
|
|
164
|
+
trial.steps = piSteps(pi[1]);
|
|
165
|
+
trial.steps = boundedSteps(trial.steps);
|
|
166
|
+
trial.log = [
|
|
167
|
+
redact(archive ? stdout : completeLines(stdout)),
|
|
168
|
+
...[...clean]
|
|
169
|
+
.filter(([name]) => /\.(log|txt)$/.test(name))
|
|
170
|
+
.map(([name, text]) => `\n--- ${name} ---\n${text}`),
|
|
171
|
+
]
|
|
172
|
+
.join("\n")
|
|
173
|
+
.slice(-100_000);
|
|
174
|
+
if (archive) {
|
|
175
|
+
for (const [name, text] of clean) {
|
|
176
|
+
const artifactName = `${index}/${name}`;
|
|
177
|
+
await store.put(`${evaluationPrefix(run.repoId, run.id)}artifacts/${artifactName}`, Buffer.from(text), "text/plain");
|
|
178
|
+
trial.artifacts.push(artifactName);
|
|
179
|
+
}
|
|
180
|
+
const result = [...files].find(([name]) => /^solver\/[^/]+\/result\.json$/.test(name));
|
|
181
|
+
if (!result)
|
|
182
|
+
throw new Error("Harbor did not produce a trial result");
|
|
183
|
+
const parsed = record(JSON.parse(result[1]));
|
|
184
|
+
Object.assign(trial, trialCost(run, trial.harness, files, parsed));
|
|
185
|
+
const rewards = record(record(parsed.verifier_result).rewards);
|
|
186
|
+
trial.rewards = Object.fromEntries(Object.entries(rewards).filter((pair) => typeof pair[1] === "number" && Number.isFinite(pair[1])));
|
|
187
|
+
if (parsed.exception_info)
|
|
188
|
+
throw new Error(String(record(parsed.exception_info).exception_message ?? "Harbor trial failed"));
|
|
189
|
+
if (Object.keys(trial.rewards).length === 0)
|
|
190
|
+
throw new Error("Harbor returned no verifier scores");
|
|
191
|
+
}
|
|
192
|
+
const snapshot = JSON.stringify(trial);
|
|
193
|
+
if (snapshot !== lastSnapshot) {
|
|
194
|
+
await saveEvaluation(store, run);
|
|
195
|
+
lastSnapshot = snapshot;
|
|
196
|
+
}
|
|
197
|
+
};
|
|
198
|
+
const timer = setInterval(() => {
|
|
199
|
+
options.heartbeat?.();
|
|
200
|
+
if (refreshing)
|
|
201
|
+
return;
|
|
202
|
+
refreshing = true;
|
|
203
|
+
polling = polling
|
|
204
|
+
.then(() => refresh(false))
|
|
205
|
+
.catch(() => undefined)
|
|
206
|
+
.finally(() => {
|
|
207
|
+
refreshing = false;
|
|
208
|
+
});
|
|
209
|
+
}, options.pollMs ?? 3000);
|
|
210
|
+
try {
|
|
211
|
+
const result = await command("harbor", solverArguments(taskPath, jobs, trial.harness, model, run.sandbox, run.thinking), {
|
|
212
|
+
env: child,
|
|
213
|
+
cwd: taskPath,
|
|
214
|
+
timeoutMs: 2 * 60 * 60 * 1000,
|
|
215
|
+
allowFailure: true,
|
|
216
|
+
...(options.signal ? { signal: options.signal } : {}),
|
|
217
|
+
onOutput: (_stream, chunk) => {
|
|
218
|
+
stdout = `${stdout}${Buffer.from(chunk).toString("utf8")}`.slice(-200_000);
|
|
219
|
+
},
|
|
220
|
+
});
|
|
221
|
+
clearInterval(timer);
|
|
222
|
+
await polling;
|
|
223
|
+
await refresh(true);
|
|
224
|
+
if (result.exitCode !== 0)
|
|
225
|
+
throw new Error(`Harbor exited with status ${result.exitCode}`);
|
|
226
|
+
trial.status = "completed";
|
|
227
|
+
}
|
|
228
|
+
catch (error) {
|
|
229
|
+
clearInterval(timer);
|
|
230
|
+
await polling;
|
|
231
|
+
if (trial.artifacts.length === 0)
|
|
232
|
+
await refresh(true).catch(() => undefined);
|
|
233
|
+
throw error;
|
|
234
|
+
}
|
|
235
|
+
finally {
|
|
236
|
+
clearInterval(timer);
|
|
237
|
+
await polling;
|
|
238
|
+
trial.log = `${trial.log}\n${redact(stdout)}`.slice(-100_000);
|
|
239
|
+
}
|
|
240
|
+
}
|
|
241
|
+
//# sourceMappingURL=runner.js.map
|
|
@@ -0,0 +1 @@
|
|
|
1
|
+
{"version":3,"file":"runner.js","sourceRoot":"","sources":["../../src/evaluation/runner.ts"],"names":[],"mappings":"AAAA,OAAO,EAAE,KAAK,EAAE,OAAO,EAAE,QAAQ,EAAE,EAAE,EAAE,SAAS,EAAE,MAAM,kBAAkB,CAAC;AAC3E,OAAO,EAAE,MAAM,EAAE,MAAM,SAAS,CAAC;AACjC,OAAO,EAAE,OAAO,EAAE,IAAI,EAAE,MAAM,WAAW,CAAC;AAC1C,OAAO,EAAE,aAAa,EAAE,MAAM,UAAU,CAAC;AACzC,OAAO,EAAE,qBAAqB,EAAE,MAAM,eAAe,CAAC;AAEtD,OAAO,EAAE,UAAU,EAAE,MAAM,eAAe,CAAC;AAC3C,OAAO,EAAE,cAAc,EAAE,iBAAiB,EAAE,MAAM,aAAa,CAAC;AAChE,OAAO,EAAE,SAAS,EAAE,MAAM,WAAW,CAAC;AACtC,OAAO,EAAE,mBAAmB,EAAE,MAAM,2BAA2B,CAAC;AAEhE,OAAO,EAAE,YAAY,EAAE,WAAW,EAAE,MAAM,wBAAwB,CAAC;AACnE,OAAO,EAAsB,iBAAiB,EAAE,MAAM,oBAAoB,CAAC;AAC3E,OAAO,EACL,YAAY,EACZ,aAAa,EACb,aAAa,EACb,OAAO,EACP,MAAM,EACN,YAAY,EACZ,eAAe,GAChB,MAAM,aAAa,CAAC;AACrB,OAAO,EAAE,kBAAkB,EAAE,MAAM,eAAe,CAAC;AACnD,OAAO,EAAE,gBAAgB,EAAE,aAAa,EAAE,cAAc,EAAE,MAAM,YAAY,CAAC;AAG7E,MAAM,UAAU,eAAe,CAC7B,QAAgB,EAChB,IAAY,EACZ,OAAgB,EAChB,KAAa,EACb,OAAe,EACf,QAAwB;IAExB,OAAO;QACL,KAAK;QACL,QAAQ;QACR,QAAQ;QACR,SAAS;QACT,WAAW,CAAC,OAAO,EAAE,KAAK,CAAC;QAC3B,SAAS;QACT,KAAK;QACL,OAAO;QACP,OAAO,KAAK,KAAK,CAAC,CAAC,CAAC,oCAAoC,CAAC,CAAC,CAAC,OAAO;QAClE,YAAY;QACZ,IAAI;QACJ,YAAY;QACZ,QAAQ;QACR,cAAc;QACd,GAAG;QACH,gBAAgB;QAChB,GAAG;QACH,eAAe;QACf,GAAG;QACH,UAAU;QACV,OAAO;QACP,GAAG,iBAAiB,CAAC,OAAO,EAAE,QAAQ,CAAC;KACxC,CAAC;AACJ,CAAC;AASD,MAAM,CAAC,KAAK,UAAU,iBAAiB,CACrC,KAAoB,EACpB,KAAsB,EACtB,UAAyB,EAAE;IAE3B,MAAM,GAAG,GAAG,MAAM,aAAa,CAAC,KAAK,EAAE,KAAK,CAAC,MAAM,EAAE,KAAK,CAAC,EAAE,CAAC,CAAC;IAC/D,IAAI,CAAC,GAAG;QAAE,MAAM,IAAI,KAAK,CAAC,8BAA8B,CAAC,CAAC;IAC1D,IAAI,GAAG,CAAC,MAAM,KAAK,QAAQ;QACzB,MAAM,IAAI,KAAK,CAAC,8DAA8D,CAAC,CAAC;IAClF,MAAM,IAAI,GAAG,MAAM,OAAO,CAAC,IAAI,CAAC,MAAM,EAAE,EAAE,uBAAuB,CAAC,CAAC,CAAC;IACpE,MAAM,OAAO,GAAG,OAAO,CAAC,OAAO,IAAI,UAAU,CAAC;IAC9C,MAAM,WAAW,GAAG,OAAO,CAAC,GAAG,IAAI,OAAO,CAAC,GAAG,CAAC;IAC/C,MAAM,OAAO,GAAG,MAAM,CAAC,OAAO,CAAC,WAAW,CAAC;SACxC,MAAM,CAAC,CAAC,CAAC,IAAI,CAAC,EAAE,EAAE,CAAC,+BAA+B,CAAC,IAAI,CAAC,IAAI,CAAC,CAAC;SAC9D,GAAG,CAAC,CAAC,CAAC,EAAE,KAAK,CAAC,EAAE,EAAE,CAAC,KAAK,IAAI,EAAE,CAAC;SAC/B,MAAM,CAAC,OAAO,CAAC,CAAC;IACnB,MAAM,MAAM,GAAG,CAAC,IAAY,EAAE,EAAE,CAAC,YAAY,CAAC,IAAI,EAAE,OAAO,CAAC,CAAC;IAC7D,GAAG,CAAC,MAAM,GAAG,SAAS,CAAC;IACvB,IAAI,CAAC;QACH,MAAM,cAAc,CAAC,KAAK,EAAE,GAAG,CAAC,CAAC;QACjC,MAAM,IAAI,GAAG,IAAI,CAAC,IAAI,EAAE,MAAM,CAAC,CAAC;QAChC,MAAM,KAAK,CAAC,IAAI,EAAE,EAAE,IAAI,EAAE,KAAK,EAAE,CAAC,CAAC;QACnC,MAAM,QAAQ,GAAG,KAAK,CAAC,WAAW;YAChC,CAAC,CAAC,WAAW;YACb,CAAC,CAAC,MAAM,kBAAkB,CAAC,KAAK,EAAE,KAAK,EAAE,WAAW,CAAC,CAAC;QACxD,OAAO,CAAC,IAAI,CACV,GAAG,MAAM,CAAC,OAAO,CAAC,QAAQ,CAAC;aACxB,MAAM,CAAC,CAAC,CAAC,IAAI,CAAC,EAAE,EAAE,CAAC,IAAI,CAAC,UAAU,CAAC,wBAAwB,CAAC,CAAC;aAC7D,GAAG,CAAC,CAAC,CAAC,EAAE,KAAK,CAAC,EAAE,EAAE,CAAC,KAAK,IAAI,EAAE,CAAC;aAC/B,MAAM,CAAC,OAAO,CAAC,CACnB,CAAC;QACF,IAAI,KAAK,CAAC,WAAW,IAAI,CAAC,OAAO,CAAC,OAAO;YACvC,MAAM,IAAI,KAAK,CAAC,8CAA8C,CAAC,CAAC;QAClE,MAAM,SAAS,GACb,KAAK,CAAC,WAAW,IAAI,OAAO,CAAC,OAAO;YAClC,CAAC,CAAC,MAAM,mBAAmB,CAAC,KAAK,EAAE,IAAI,EAAE,WAAW,EAAE,OAAO,CAAC,OAAO,CAAC;YACtE,CAAC,CAAC,EAAE,GAAG,iBAAiB,CAAC,KAAK,EAAE,IAAI,EAAE,QAAQ,CAAC,EAAE,OAAO,EAAE,EAAE,EAAE,CAAC;QACnE,OAAO,CAAC,IAAI,CAAC,GAAG,SAAS,CAAC,OAAO,CAAC,CAAC;QACnC,MAAM,EAAE,OAAO,EAAE,KAAK,EAAE,GAAG,SAAS,CAAC;QACrC,KAAK,CAAC,UAAU,GAAG,OAAO,CAAC,aAAa,CAAC,MAAM,CAAC,IAAI,CAAC,GAAG,CAAC,CAAC,CAAC;QAC3D,MAAM,OAAO,GAAG,MAAM,OAAO,CAAC,QAAQ,EAAE,CAAC,WAAW,CAAC,EAAE,EAAE,GAAG,EAAE,KAAK,EAAE,SAAS,EAAE,MAAM,EAAE,CAAC,CAAC;QAC1F,IAAI,OAAO,CAAC,MAAM,CAAC,IAAI,EAAE,KAAK,cAAc;YAC1C,MAAM,IAAI,KAAK,CAAC,4DAA4D,CAAC,CAAC;QAChF,KAAK,MAAM,CAAC,KAAK,EAAE,KAAK,CAAC,IAAI,GAAG,CAAC,MAAM,CAAC,OAAO,EAAE,EAAE,CAAC;YAClD,OAAO,CAAC,MAAM,EAAE,cAAc,EAAE,CAAC;YACjC,MAAM,IAAI,GAAG,KAAK,CAAC,KAAK,CAAC,IAAI,CAC3B,CAAC,SAAS,EAAE,EAAE,CAAC,SAAS,CAAC,KAAK,KAAK,KAAK,CAAC,KAAK,IAAI,SAAS,CAAC,MAAM,KAAK,KAAK,CAAC,MAAM,CACpF,CAAC;YACF,IAAI,CAAC,IAAI;gBAAE,MAAM,IAAI,KAAK,CAAC,qCAAqC,CAAC,CAAC;YAClE,MAAM,SAAS,GAAG,IAAI,CAAC,IAAI,EAAE,MAAM,CAAC,KAAK,CAAC,CAAC,CAAC;YAC5C,MAAM,KAAK,CAAC,SAAS,CAAC,CAAC;YACvB,KAAK,CAAC,MAAM,GAAG,SAAS,CAAC;YACzB,KAAK,CAAC,SAAS,GAAG,IAAI,IAAI,EAAE,CAAC,WAAW,EAAE,CAAC;YAC3C,MAAM,cAAc,CAAC,KAAK,EAAE,GAAG,CAAC,CAAC;YACjC,IAAI,CAAC;gBACH,MAAM,MAAM,GAAG,MAAM,KAAK,CAAC,QAAQ,CAAC,IAAI,CAAC,SAAS,CAAC,CAAC;gBACpD,IAAI,CAAC,MAAM,IAAI,MAAM,CAAC,UAAU,GAAG,GAAG,GAAG,IAAI,GAAG,IAAI;oBAClD,MAAM,IAAI,KAAK,CAAC,2CAA2C,CAAC,CAAC;gBAC/D,MAAM,OAAO,GAAG,IAAI,CAAC,SAAS,EAAE,aAAa,CAAC,CAAC;gBAC/C,MAAM,SAAS,CAAC,OAAO,EAAE,MAAM,EAAE,EAAE,IAAI,EAAE,KAAK,EAAE,CAAC,CAAC;gBAClD,MAAM,SAAS,GAAG,IAAI,CAAC,SAAS,EAAE,MAAM,CAAC,CAAC;gBAC1C,MAAM,KAAK,CAAC,SAAS,CAAC,CAAC;gBACvB,MAAM,qBAAqB,CACzB,OAAO,EACP,SAAS,EACT,OAAO,CAAC,MAAM,CAAC,CAAC,CAAC,EAAE,MAAM,EAAE,OAAO,CAAC,MAAM,EAAE,CAAC,CAAC,CAAC,EAAE,CACjD,CAAC;gBACF,MAAM,QAAQ,GAAG,MAAM,QAAQ,CAAC,IAAI,CAAC,SAAS,EAAE,aAAa,EAAE,WAAW,CAAC,CAAC,CAAC,IAAI,CAC/E,GAAG,EAAE,CAAC,IAAI,CAAC,SAAS,EAAE,aAAa,CAAC,EACpC,GAAG,EAAE,CAAC,SAAS,CAChB,CAAC;gBACF,MAAM,QAAQ,CAAC,IAAI,CAAC,QAAQ,EAAE,WAAW,CAAC,CAAC,CAAC;gBAC5C,MAAM,QAAQ,CAAC;oBACb,KAAK;oBACL,GAAG;oBACH,KAAK;oBACL,KAAK;oBACL,QAAQ;oBACR,IAAI,EAAE,IAAI,CAAC,SAAS,EAAE,MAAM,CAAC;oBAC7B,GAAG,YAAY,CAAC,KAAK,EAAE,KAAK,CAAC,OAAO,EAAE,OAAO,CAAC,KAAK,EAAE,KAAK,CAAC;oBAC3D,OAAO;oBACP,MAAM;oBACN,OAAO;iBACR,CAAC,CAAC;YACL,CAAC;YAAC,OAAO,KAAK,EAAE,CAAC;gBACf,KAAK,CAAC,MAAM,GAAG,QAAQ,CAAC;gBACxB,KAAK,CAAC,KAAK,GAAG,MAAM,CAAC,KAAK,YAAY,KAAK,CAAC,CAAC,CAAC,KAAK,CAAC,OAAO,CAAC,CAAC,CAAC,eAAe,CAAC,CAAC;YACjF,CAAC;YACD,KAAK,CAAC,UAAU,GAAG,IAAI,IAAI,EAAE,CAAC,WAAW,EAAE,CAAC;YAC5C,MAAM,cAAc,CAAC,KAAK,EAAE,GAAG,CAAC,CAAC;QACnC,CAAC;QACD,GAAG,CAAC,MAAM,GAAG,GAAG,CAAC,MAAM,CAAC,IAAI,CAAC,CAAC,KAAK,EAAE,EAAE,CAAC,KAAK,CAAC,MAAM,KAAK,QAAQ,CAAC,CAAC,CAAC,CAAC,QAAQ,CAAC,CAAC,CAAC,WAAW,CAAC;IAC9F,CAAC;IAAC,OAAO,KAAK,EAAE,CAAC;QACf,GAAG,CAAC,MAAM,GAAG,QAAQ,CAAC;QACtB,GAAG,CAAC,KAAK,GAAG,MAAM,CAAC,KAAK,YAAY,KAAK,CAAC,CAAC,CAAC,KAAK,CAAC,OAAO,CAAC,CAAC,CAAC,mBAAmB,CAAC,CAAC;IACnF,CAAC;YAAS,CAAC;QACT,KAAK,MAAM,KAAK,IAAI,GAAG,CAAC,MAAM,EAAE,CAAC;YAC/B,IAAI,KAAK,CAAC,MAAM,KAAK,QAAQ,IAAI,KAAK,CAAC,MAAM,KAAK,SAAS,EAAE,CAAC;gBAC5D,KAAK,CAAC,MAAM,GAAG,QAAQ,CAAC;gBACxB,KAAK,CAAC,KAAK,GAAG,8CAA8C,CAAC;YAC/D,CAAC;QACH,CAAC;QACD,GAAG,CAAC,UAAU,GAAG,IAAI,IAAI,EAAE,CAAC,WAAW,EAAE,CAAC;QAC1C,IAAI,CAAC;YACH,MAAM,cAAc,CAAC,KAAK,EAAE,GAAG,CAAC,CAAC;QACnC,CAAC;gBAAS,CAAC;YACT,MAAM,EAAE,CAAC,IAAI,EAAE,EAAE,SAAS,EAAE,IAAI,EAAE,KAAK,EAAE,IAAI,EAAE,CAAC,CAAC;QACnD,CAAC;IACH,CAAC;AACH,CAAC;AACD,KAAK,UAAU,QAAQ,CAAC,OAYvB;IACC,MAAM,EAAE,KAAK,EAAE,GAAG,EAAE,KAAK,EAAE,KAAK,EAAE,QAAQ,EAAE,IAAI,EAAE,KAAK,EAAE,KAAK,EAAE,OAAO,EAAE,MAAM,EAAE,OAAO,EAAE,GACxF,OAAO,CAAC;IACV,IAAI,MAAM,GAAG,EAAE,CAAC;IAChB,IAAI,OAAO,GAAG,OAAO,CAAC,OAAO,EAAE,CAAC;IAChC,IAAI,UAAU,GAAG,KAAK,CAAC;IACvB,IAAI,YAAY,GAAG,EAAE,CAAC;IACtB,MAAM,OAAO,GAAG,KAAK,EAAE,OAAgB,EAAE,EAAE;QACzC,MAAM,KAAK,GAAG,MAAM,aAAa,CAAC,IAAI,CAAC,CAAC;QACxC,MAAM,KAAK,GAAG,IAAI,GAAG,CACnB,CAAC,GAAG,KAAK,CAAC,CAAC,GAAG,CAAC,CAAC,CAAC,IAAI,EAAE,IAAI,CAAC,EAAE,EAAE,CAAC;YAC/B,IAAI;YACJ,MAAM,CAAC,CAAC,OAAO,IAAI,CAAC,IAAI,CAAC,QAAQ,CAAC,OAAO,CAAC,CAAC,CAAC,CAAC,aAAa,CAAC,IAAI,CAAC,CAAC,CAAC,CAAC,IAAI,CAAC;SACzE,CAAC,CACH,CAAC;QACF,MAAM,UAAU,GAAG,CAAC,GAAG,KAAK,CAAC,CAAC,IAAI,CAAC,CAAC,CAAC,IAAI,CAAC,EAAE,EAAE,CAAC,IAAI,CAAC,QAAQ,CAAC,kBAAkB,CAAC,CAAC,CAAC;QAClF,MAAM,EAAE,GAAG,CAAC,GAAG,KAAK,CAAC,CAAC,IAAI,CAAC,CAAC,CAAC,IAAI,CAAC,EAAE,EAAE,CAAC,IAAI,CAAC,QAAQ,CAAC,SAAS,CAAC,CAAC,CAAC;QACjE,IAAI,UAAU,EAAE,CAAC;YACf,IAAI,CAAC;gBACH,KAAK,CAAC,KAAK,GAAG,eAAe,CAAC,IAAI,CAAC,KAAK,CAAC,UAAU,CAAC,CAAC,CAAC,CAAC,CAAC,CAAC;YAC3D,CAAC;YAAC,MAAM,CAAC;gBACP,KAAK,CAAC,KAAK,GAAG,EAAE,CAAC;YACnB,CAAC;QACH,CAAC;aAAM,IAAI,EAAE;YAAE,KAAK,CAAC,KAAK,GAAG,OAAO,CAAC,EAAE,CAAC,CAAC,CAAC,CAAC,CAAC;QAC5C,KAAK,CAAC,KAAK,GAAG,YAAY,CAAC,KAAK,CAAC,KAAK,CAAC,CAAC;QACxC,KAAK,CAAC,GAAG,GAAG;YACV,MAAM,CAAC,OAAO,CAAC,CAAC,CAAC,MAAM,CAAC,CAAC,CAAC,aAAa,CAAC,MAAM,CAAC,CAAC;YAChD,GAAG,CAAC,GAAG,KAAK,CAAC;iBACV,MAAM,CAAC,CAAC,CAAC,IAAI,CAAC,EAAE,EAAE,CAAC,cAAc,CAAC,IAAI,CAAC,IAAI,CAAC,CAAC;iBAC7C,GAAG,CAAC,CAAC,CAAC,IAAI,EAAE,IAAI,CAAC,EAAE,EAAE,CAAC,SAAS,IAAI,SAAS,IAAI,EAAE,CAAC;SACvD;aACE,IAAI,CAAC,IAAI,CAAC;aACV,KAAK,CAAC,CAAC,OAAO,CAAC,CAAC;QACnB,IAAI,OAAO,EAAE,CAAC;YACZ,KAAK,MAAM,CAAC,IAAI,EAAE,IAAI,CAAC,IAAI,KAAK,EAAE,CAAC;gBACjC,MAAM,YAAY,GAAG,GAAG,KAAK,IAAI,IAAI,EAAE,CAAC;gBACxC,MAAM,KAAK,CAAC,GAAG,CACb,GAAG,gBAAgB,CAAC,GAAG,CAAC,MAAM,EAAE,GAAG,CAAC,EAAE,CAAC,aAAa,YAAY,EAAE,EAClE,MAAM,CAAC,IAAI,CAAC,IAAI,CAAC,EACjB,YAAY,CACb,CAAC;gBACF,KAAK,CAAC,SAAS,CAAC,IAAI,CAAC,YAAY,CAAC,CAAC;YACrC,CAAC;YACD,MAAM,MAAM,GAAG,CAAC,GAAG,KAAK,CAAC,CAAC,IAAI,CAAC,CAAC,CAAC,IAAI,CAAC,EAAE,EAAE,CAAC,+BAA+B,CAAC,IAAI,CAAC,IAAI,CAAC,CAAC,CAAC;YACvF,IAAI,CAAC,MAAM;gBAAE,MAAM,IAAI,KAAK,CAAC,uCAAuC,CAAC,CAAC;YACtE,MAAM,MAAM,GAAG,MAAM,CAAC,IAAI,CAAC,KAAK,CAAC,MAAM,CAAC,CAAC,CAAC,CAAC,CAAC,CAAC;YAC7C,MAAM,CAAC,MAAM,CAAC,KAAK,EAAE,SAAS,CAAC,GAAG,EAAE,KAAK,CAAC,OAAO,EAAE,KAAK,EAAE,MAAM,CAAC,CAAC,CAAC;YACnE,MAAM,OAAO,GAAG,MAAM,CAAC,MAAM,CAAC,MAAM,CAAC,eAAe,CAAC,CAAC,OAAO,CAAC,CAAC;YAC/D,KAAK,CAAC,OAAO,GAAG,MAAM,CAAC,WAAW,CAChC,MAAM,CAAC,OAAO,CAAC,OAAO,CAAC,CAAC,MAAM,CAC5B,CAAC,IAAI,EAA4B,EAAE,CACjC,OAAO,IAAI,CAAC,CAAC,CAAC,KAAK,QAAQ,IAAI,MAAM,CAAC,QAAQ,CAAC,IAAI,CAAC,CAAC,CAAC,CAAC,CAC1D,CACF,CAAC;YACF,IAAI,MAAM,CAAC,cAAc;gBACvB,MAAM,IAAI,KAAK,CACb,MAAM,CAAC,MAAM,CAAC,MAAM,CAAC,cAAc,CAAC,CAAC,iBAAiB,IAAI,qBAAqB,CAAC,CACjF,CAAC;YACJ,IAAI,MAAM,CAAC,IAAI,CAAC,KAAK,CAAC,OAAO,CAAC,CAAC,MAAM,KAAK,CAAC;gBACzC,MAAM,IAAI,KAAK,CAAC,oCAAoC,CAAC,CAAC;QAC1D,CAAC;QACD,MAAM,QAAQ,GAAG,IAAI,CAAC,SAAS,CAAC,KAAK,CAAC,CAAC;QACvC,IAAI,QAAQ,KAAK,YAAY,EAAE,CAAC;YAC9B,MAAM,cAAc,CAAC,KAAK,EAAE,GAAG,CAAC,CAAC;YACjC,YAAY,GAAG,QAAQ,CAAC;QAC1B,CAAC;IACH,CAAC,CAAC;IACF,MAAM,KAAK,GAAG,WAAW,CAAC,GAAG,EAAE;QAC7B,OAAO,CAAC,SAAS,EAAE,EAAE,CAAC;QACtB,IAAI,UAAU;YAAE,OAAO;QACvB,UAAU,GAAG,IAAI,CAAC;QAClB,OAAO,GAAG,OAAO;aACd,IAAI,CAAC,GAAG,EAAE,CAAC,OAAO,CAAC,KAAK,CAAC,CAAC;aAC1B,KAAK,CAAC,GAAG,EAAE,CAAC,SAAS,CAAC;aACtB,OAAO,CAAC,GAAG,EAAE;YACZ,UAAU,GAAG,KAAK,CAAC;QACrB,CAAC,CAAC,CAAC;IACP,CAAC,EAAE,OAAO,CAAC,MAAM,IAAI,IAAI,CAAC,CAAC;IAC3B,IAAI,CAAC;QACH,MAAM,MAAM,GAAG,MAAM,OAAO,CAC1B,QAAQ,EACR,eAAe,CAAC,QAAQ,EAAE,IAAI,EAAE,KAAK,CAAC,OAAO,EAAE,KAAK,EAAE,GAAG,CAAC,OAAO,EAAE,GAAG,CAAC,QAAQ,CAAC,EAChF;YACE,GAAG,EAAE,KAAK;YACV,GAAG,EAAE,QAAQ;YACb,SAAS,EAAE,CAAC,GAAG,EAAE,GAAG,EAAE,GAAG,IAAI;YAC7B,YAAY,EAAE,IAAI;YAClB,GAAG,CAAC,OAAO,CAAC,MAAM,CAAC,CAAC,CAAC,EAAE,MAAM,EAAE,OAAO,CAAC,MAAM,EAAE,CAAC,CAAC,CAAC,EAAE,CAAC;YACrD,QAAQ,EAAE,CAAC,OAAO,EAAE,KAAK,EAAE,EAAE;gBAC3B,MAAM,GAAG,GAAG,MAAM,GAAG,MAAM,CAAC,IAAI,CAAC,KAAK,CAAC,CAAC,QAAQ,CAAC,MAAM,CAAC,EAAE,CAAC,KAAK,CAAC,CAAC,OAAO,CAAC,CAAC;YAC7E,CAAC;SACF,CACF,CAAC;QACF,aAAa,CAAC,KAAK,CAAC,CAAC;QACrB,MAAM,OAAO,CAAC;QACd,MAAM,OAAO,CAAC,IAAI,CAAC,CAAC;QACpB,IAAI,MAAM,CAAC,QAAQ,KAAK,CAAC;YAAE,MAAM,IAAI,KAAK,CAAC,6BAA6B,MAAM,CAAC,QAAQ,EAAE,CAAC,CAAC;QAC3F,KAAK,CAAC,MAAM,GAAG,WAAW,CAAC;IAC7B,CAAC;IAAC,OAAO,KAAK,EAAE,CAAC;QACf,aAAa,CAAC,KAAK,CAAC,CAAC;QACrB,MAAM,OAAO,CAAC;QACd,IAAI,KAAK,CAAC,SAAS,CAAC,MAAM,KAAK,CAAC;YAAE,MAAM,OAAO,CAAC,IAAI,CAAC,CAAC,KAAK,CAAC,GAAG,EAAE,CAAC,SAAS,CAAC,CAAC;QAC7E,MAAM,KAAK,CAAC;IACd,CAAC;YAAS,CAAC;QACT,aAAa,CAAC,KAAK,CAAC,CAAC;QACrB,MAAM,OAAO,CAAC;QACd,KAAK,CAAC,GAAG,GAAG,GAAG,KAAK,CAAC,GAAG,KAAK,MAAM,CAAC,MAAM,CAAC,EAAE,CAAC,KAAK,CAAC,CAAC,OAAO,CAAC,CAAC;IAChE,CAAC;AACH,CAAC"}
|
|
@@ -0,0 +1,8 @@
|
|
|
1
|
+
import type { ArtifactStore } from "../artifacts.js";
|
|
2
|
+
import type { EvaluationInput, EvaluationRun } from "./types.js";
|
|
3
|
+
export declare function evaluationPrefix(repoId: number, id?: string): string;
|
|
4
|
+
export declare function initialEvaluation(input: EvaluationInput, modelLabel: string): EvaluationRun;
|
|
5
|
+
export declare function saveEvaluation(store: ArtifactStore, run: EvaluationRun): Promise<void>;
|
|
6
|
+
export declare function getEvaluation(store: ArtifactStore, repoId: number, id: string): Promise<EvaluationRun | undefined>;
|
|
7
|
+
export declare function listEvaluations(store: ArtifactStore, repoId: number): Promise<EvaluationRun[]>;
|
|
8
|
+
//# sourceMappingURL=store.d.ts.map
|
|
@@ -0,0 +1 @@
|
|
|
1
|
+
{"version":3,"file":"store.d.ts","sourceRoot":"","sources":["../../src/evaluation/store.ts"],"names":[],"mappings":"AACA,OAAO,KAAK,EAAE,aAAa,EAAE,MAAM,iBAAiB,CAAC;AACrD,OAAO,KAAK,EAAE,eAAe,EAAE,aAAa,EAAE,MAAM,YAAY,CAAC;AAEjE,wBAAgB,gBAAgB,CAAC,MAAM,EAAE,MAAM,EAAE,EAAE,SAAK,GAAG,MAAM,CAIhE;AACD,wBAAgB,iBAAiB,CAAC,KAAK,EAAE,eAAe,EAAE,UAAU,EAAE,MAAM,GAAG,aAAa,CA6B3F;AACD,wBAAsB,cAAc,CAAC,KAAK,EAAE,aAAa,EAAE,GAAG,EAAE,aAAa,GAAG,OAAO,CAAC,IAAI,CAAC,CAO5F;AACD,wBAAsB,aAAa,CACjC,KAAK,EAAE,aAAa,EACpB,MAAM,EAAE,MAAM,EACd,EAAE,EAAE,MAAM,GACT,OAAO,CAAC,aAAa,GAAG,SAAS,CAAC,CAKpC;AACD,wBAAsB,eAAe,CACnC,KAAK,EAAE,aAAa,EACpB,MAAM,EAAE,MAAM,GACb,OAAO,CAAC,aAAa,EAAE,CAAC,CAoB1B"}
|