@nodetool-ai/agents 0.7.0-rc.22 → 0.8.1
This diff represents the content of publicly available package versions that have been released to one of the supported registries. The information contained in this diff is provided for informational purposes only and reflects changes between package versions as they appear in their respective public registries.
- package/README.md +35 -0
- package/dist/.nodetool-build-stamp +1 -0
- package/dist/analysis/audio-dsp.d.ts +223 -0
- package/dist/analysis/audio-dsp.d.ts.map +1 -0
- package/dist/analysis/audio-dsp.js +642 -0
- package/dist/analysis/audio-dsp.js.map +1 -0
- package/dist/analysis/media-decode.d.ts +85 -0
- package/dist/analysis/media-decode.d.ts.map +1 -0
- package/dist/analysis/media-decode.js +199 -0
- package/dist/analysis/media-decode.js.map +1 -0
- package/dist/analysis/mediabunny-runtime.d.ts +15 -0
- package/dist/analysis/mediabunny-runtime.d.ts.map +1 -0
- package/dist/analysis/mediabunny-runtime.js +30 -0
- package/dist/analysis/mediabunny-runtime.js.map +1 -0
- package/dist/analysis/video-frames.d.ts +120 -0
- package/dist/analysis/video-frames.d.ts.map +1 -0
- package/dist/analysis/video-frames.js +288 -0
- package/dist/analysis/video-frames.js.map +1 -0
- package/dist/apify/assets.d.ts +80 -0
- package/dist/apify/assets.d.ts.map +1 -0
- package/dist/apify/assets.js +192 -0
- package/dist/apify/assets.js.map +1 -0
- package/dist/apify/catalog.d.ts +44 -0
- package/dist/apify/catalog.d.ts.map +1 -0
- package/dist/apify/catalog.js +108 -0
- package/dist/apify/catalog.js.map +1 -0
- package/dist/apify/client.d.ts +137 -0
- package/dist/apify/client.d.ts.map +1 -0
- package/dist/apify/client.js +346 -0
- package/dist/apify/client.js.map +1 -0
- package/dist/apify/errors.d.ts +85 -0
- package/dist/apify/errors.d.ts.map +1 -0
- package/dist/apify/errors.js +90 -0
- package/dist/apify/errors.js.map +1 -0
- package/dist/apify/normalize.d.ts +131 -0
- package/dist/apify/normalize.d.ts.map +1 -0
- package/dist/apify/normalize.js +202 -0
- package/dist/apify/normalize.js.map +1 -0
- package/dist/apify/policy.d.ts +136 -0
- package/dist/apify/policy.d.ts.map +1 -0
- package/dist/apify/policy.js +223 -0
- package/dist/apify/policy.js.map +1 -0
- package/dist/apify/run.d.ts +86 -0
- package/dist/apify/run.d.ts.map +1 -0
- package/dist/apify/run.js +158 -0
- package/dist/apify/run.js.map +1 -0
- package/dist/app-build/author.d.ts +67 -0
- package/dist/app-build/author.d.ts.map +1 -0
- package/dist/app-build/author.js +231 -0
- package/dist/app-build/author.js.map +1 -0
- package/dist/app-build/bridge.d.ts +130 -0
- package/dist/app-build/bridge.d.ts.map +1 -0
- package/dist/app-build/bridge.js +647 -0
- package/dist/app-build/bridge.js.map +1 -0
- package/dist/app-build/build-service.d.ts +66 -0
- package/dist/app-build/build-service.d.ts.map +1 -0
- package/dist/app-build/build-service.js +249 -0
- package/dist/app-build/build-service.js.map +1 -0
- package/dist/app-build/build.d.ts +115 -0
- package/dist/app-build/build.d.ts.map +1 -0
- package/dist/app-build/build.js +901 -0
- package/dist/app-build/build.js.map +1 -0
- package/dist/app-build/interactions.d.ts +20 -0
- package/dist/app-build/interactions.d.ts.map +1 -0
- package/dist/app-build/interactions.js +121 -0
- package/dist/app-build/interactions.js.map +1 -0
- package/dist/app-build/judge.d.ts +113 -0
- package/dist/app-build/judge.d.ts.map +1 -0
- package/dist/app-build/judge.js +318 -0
- package/dist/app-build/judge.js.map +1 -0
- package/dist/app-build/markdown.d.ts +10 -0
- package/dist/app-build/markdown.d.ts.map +1 -0
- package/dist/app-build/markdown.js +65 -0
- package/dist/app-build/markdown.js.map +1 -0
- package/dist/app-build/spec.d.ts +228 -0
- package/dist/app-build/spec.d.ts.map +1 -0
- package/dist/app-build/spec.js +651 -0
- package/dist/app-build/spec.js.map +1 -0
- package/dist/app-build/tool-loop.d.ts +84 -0
- package/dist/app-build/tool-loop.d.ts.map +1 -0
- package/dist/app-build/tool-loop.js +113 -0
- package/dist/app-build/tool-loop.js.map +1 -0
- package/dist/app-build/types.d.ts +215 -0
- package/dist/app-build/types.d.ts.map +1 -0
- package/dist/app-build/types.js +28 -0
- package/dist/app-build/types.js.map +1 -0
- package/dist/author-graph.d.ts +91 -0
- package/dist/author-graph.d.ts.map +1 -0
- package/dist/author-graph.js +319 -0
- package/dist/author-graph.js.map +1 -0
- package/dist/background-subtasks.d.ts +79 -0
- package/dist/background-subtasks.d.ts.map +1 -0
- package/dist/background-subtasks.js +139 -0
- package/dist/background-subtasks.js.map +1 -0
- package/dist/blender-gate.d.ts +21 -0
- package/dist/blender-gate.d.ts.map +1 -0
- package/dist/blender-gate.js +27 -0
- package/dist/blender-gate.js.map +1 -0
- package/dist/browser-gate.d.ts +28 -0
- package/dist/browser-gate.d.ts.map +1 -0
- package/dist/browser-gate.js +34 -0
- package/dist/browser-gate.js.map +1 -0
- package/dist/capabilities/adapters.d.ts +36 -0
- package/dist/capabilities/adapters.d.ts.map +1 -0
- package/dist/capabilities/adapters.js +44 -0
- package/dist/capabilities/adapters.js.map +1 -0
- package/dist/capabilities/agents.d.ts +78 -0
- package/dist/capabilities/agents.d.ts.map +1 -0
- package/dist/capabilities/agents.js +366 -0
- package/dist/capabilities/agents.js.map +1 -0
- package/dist/capabilities/agents.specs.d.ts +27 -0
- package/dist/capabilities/agents.specs.d.ts.map +1 -0
- package/dist/capabilities/agents.specs.js +243 -0
- package/dist/capabilities/agents.specs.js.map +1 -0
- package/dist/capabilities/analysis.d.ts +93 -0
- package/dist/capabilities/analysis.d.ts.map +1 -0
- package/dist/capabilities/analysis.js +812 -0
- package/dist/capabilities/analysis.js.map +1 -0
- package/dist/capabilities/analysis.specs.d.ts +39 -0
- package/dist/capabilities/analysis.specs.d.ts.map +1 -0
- package/dist/capabilities/analysis.specs.js +217 -0
- package/dist/capabilities/analysis.specs.js.map +1 -0
- package/dist/capabilities/apify.d.ts +34 -0
- package/dist/capabilities/apify.d.ts.map +1 -0
- package/dist/capabilities/apify.js +449 -0
- package/dist/capabilities/apify.js.map +1 -0
- package/dist/capabilities/apify.specs.d.ts +28 -0
- package/dist/capabilities/apify.specs.d.ts.map +1 -0
- package/dist/capabilities/apify.specs.js +232 -0
- package/dist/capabilities/apify.specs.js.map +1 -0
- package/dist/capabilities/apps.d.ts +40 -0
- package/dist/capabilities/apps.d.ts.map +1 -0
- package/dist/capabilities/apps.js +413 -0
- package/dist/capabilities/apps.js.map +1 -0
- package/dist/capabilities/apps.specs.d.ts +23 -0
- package/dist/capabilities/apps.specs.d.ts.map +1 -0
- package/dist/capabilities/apps.specs.js +236 -0
- package/dist/capabilities/apps.specs.js.map +1 -0
- package/dist/capabilities/args.d.ts +65 -0
- package/dist/capabilities/args.d.ts.map +1 -0
- package/dist/capabilities/args.js +218 -0
- package/dist/capabilities/args.js.map +1 -0
- package/dist/capabilities/assets.d.ts +45 -0
- package/dist/capabilities/assets.d.ts.map +1 -0
- package/dist/capabilities/assets.js +952 -0
- package/dist/capabilities/assets.js.map +1 -0
- package/dist/capabilities/assets.specs.d.ts +54 -0
- package/dist/capabilities/assets.specs.d.ts.map +1 -0
- package/dist/capabilities/assets.specs.js +291 -0
- package/dist/capabilities/assets.specs.js.map +1 -0
- package/dist/capabilities/background-generation.d.ts +14 -0
- package/dist/capabilities/background-generation.d.ts.map +1 -0
- package/dist/capabilities/background-generation.js +29 -0
- package/dist/capabilities/background-generation.js.map +1 -0
- package/dist/capabilities/browser.d.ts +30 -0
- package/dist/capabilities/browser.d.ts.map +1 -0
- package/dist/capabilities/browser.js +250 -0
- package/dist/capabilities/browser.js.map +1 -0
- package/dist/capabilities/browser.specs.d.ts +37 -0
- package/dist/capabilities/browser.specs.d.ts.map +1 -0
- package/dist/capabilities/browser.specs.js +173 -0
- package/dist/capabilities/browser.specs.js.map +1 -0
- package/dist/capabilities/code-grading.d.ts +68 -0
- package/dist/capabilities/code-grading.d.ts.map +1 -0
- package/dist/capabilities/code-grading.js +75 -0
- package/dist/capabilities/code-grading.js.map +1 -0
- package/dist/capabilities/code.d.ts +56 -0
- package/dist/capabilities/code.d.ts.map +1 -0
- package/dist/capabilities/code.js +327 -0
- package/dist/capabilities/code.js.map +1 -0
- package/dist/capabilities/code.specs.d.ts +22 -0
- package/dist/capabilities/code.specs.d.ts.map +1 -0
- package/dist/capabilities/code.specs.js +175 -0
- package/dist/capabilities/code.specs.js.map +1 -0
- package/dist/capabilities/collections.d.ts +44 -0
- package/dist/capabilities/collections.d.ts.map +1 -0
- package/dist/capabilities/collections.js +432 -0
- package/dist/capabilities/collections.js.map +1 -0
- package/dist/capabilities/collections.specs.d.ts +25 -0
- package/dist/capabilities/collections.specs.d.ts.map +1 -0
- package/dist/capabilities/collections.specs.js +299 -0
- package/dist/capabilities/collections.specs.js.map +1 -0
- package/dist/capabilities/compositions.d.ts +57 -0
- package/dist/capabilities/compositions.d.ts.map +1 -0
- package/dist/capabilities/compositions.js +384 -0
- package/dist/capabilities/compositions.js.map +1 -0
- package/dist/capabilities/compositions.specs.d.ts +22 -0
- package/dist/capabilities/compositions.specs.d.ts.map +1 -0
- package/dist/capabilities/compositions.specs.js +126 -0
- package/dist/capabilities/compositions.specs.js.map +1 -0
- package/dist/capabilities/concurrency.d.ts +12 -0
- package/dist/capabilities/concurrency.d.ts.map +1 -0
- package/dist/capabilities/concurrency.js +30 -0
- package/dist/capabilities/concurrency.js.map +1 -0
- package/dist/capabilities/costs.d.ts +20 -0
- package/dist/capabilities/costs.d.ts.map +1 -0
- package/dist/capabilities/costs.js +40 -0
- package/dist/capabilities/costs.js.map +1 -0
- package/dist/capabilities/costs.specs.d.ts +12 -0
- package/dist/capabilities/costs.specs.d.ts.map +1 -0
- package/dist/capabilities/costs.specs.js +35 -0
- package/dist/capabilities/costs.specs.js.map +1 -0
- package/dist/capabilities/dispatcher.d.ts +57 -0
- package/dist/capabilities/dispatcher.d.ts.map +1 -0
- package/dist/capabilities/dispatcher.js +117 -0
- package/dist/capabilities/dispatcher.js.map +1 -0
- package/dist/capabilities/documents.d.ts +25 -0
- package/dist/capabilities/documents.d.ts.map +1 -0
- package/dist/capabilities/documents.js +278 -0
- package/dist/capabilities/documents.js.map +1 -0
- package/dist/capabilities/documents.specs.d.ts +18 -0
- package/dist/capabilities/documents.specs.d.ts.map +1 -0
- package/dist/capabilities/documents.specs.js +201 -0
- package/dist/capabilities/documents.specs.js.map +1 -0
- package/dist/capabilities/email.d.ts +23 -0
- package/dist/capabilities/email.d.ts.map +1 -0
- package/dist/capabilities/email.js +226 -0
- package/dist/capabilities/email.js.map +1 -0
- package/dist/capabilities/email.specs.d.ts +16 -0
- package/dist/capabilities/email.specs.d.ts.map +1 -0
- package/dist/capabilities/email.specs.js +102 -0
- package/dist/capabilities/email.specs.js.map +1 -0
- package/dist/capabilities/entities.d.ts +52 -0
- package/dist/capabilities/entities.d.ts.map +1 -0
- package/dist/capabilities/entities.js +478 -0
- package/dist/capabilities/entities.js.map +1 -0
- package/dist/capabilities/entities.specs.d.ts +25 -0
- package/dist/capabilities/entities.specs.d.ts.map +1 -0
- package/dist/capabilities/entities.specs.js +247 -0
- package/dist/capabilities/entities.specs.js.map +1 -0
- package/dist/capabilities/example-timelines.d.ts +3 -0
- package/dist/capabilities/example-timelines.d.ts.map +1 -0
- package/dist/capabilities/example-timelines.js +191 -0
- package/dist/capabilities/example-timelines.js.map +1 -0
- package/dist/capabilities/example-timelines.specs.d.ts +5 -0
- package/dist/capabilities/example-timelines.specs.d.ts.map +1 -0
- package/dist/capabilities/example-timelines.specs.js +53 -0
- package/dist/capabilities/example-timelines.specs.js.map +1 -0
- package/dist/capabilities/files.d.ts +53 -0
- package/dist/capabilities/files.d.ts.map +1 -0
- package/dist/capabilities/files.js +1172 -0
- package/dist/capabilities/files.js.map +1 -0
- package/dist/capabilities/files.specs.d.ts +28 -0
- package/dist/capabilities/files.specs.d.ts.map +1 -0
- package/dist/capabilities/files.specs.js +259 -0
- package/dist/capabilities/files.specs.js.map +1 -0
- package/dist/capabilities/flow.d.ts +93 -0
- package/dist/capabilities/flow.d.ts.map +1 -0
- package/dist/capabilities/flow.js +285 -0
- package/dist/capabilities/flow.js.map +1 -0
- package/dist/capabilities/flow.specs.d.ts +25 -0
- package/dist/capabilities/flow.specs.d.ts.map +1 -0
- package/dist/capabilities/flow.specs.js +114 -0
- package/dist/capabilities/flow.specs.js.map +1 -0
- package/dist/capabilities/game-asset-source.d.ts +31 -0
- package/dist/capabilities/game-asset-source.d.ts.map +1 -0
- package/dist/capabilities/game-asset-source.js +81 -0
- package/dist/capabilities/game-asset-source.js.map +1 -0
- package/dist/capabilities/game.d.ts +12 -0
- package/dist/capabilities/game.d.ts.map +1 -0
- package/dist/capabilities/game.js +1262 -0
- package/dist/capabilities/game.js.map +1 -0
- package/dist/capabilities/game.specs.d.ts +3 -0
- package/dist/capabilities/game.specs.d.ts.map +1 -0
- package/dist/capabilities/game.specs.js +96 -0
- package/dist/capabilities/game.specs.js.map +1 -0
- package/dist/capabilities/gate-from-context.d.ts +39 -0
- package/dist/capabilities/gate-from-context.d.ts.map +1 -0
- package/dist/capabilities/gate-from-context.js +97 -0
- package/dist/capabilities/gate-from-context.js.map +1 -0
- package/dist/capabilities/gate-tools.d.ts +18 -0
- package/dist/capabilities/gate-tools.d.ts.map +1 -0
- package/dist/capabilities/gate-tools.js +66 -0
- package/dist/capabilities/gate-tools.js.map +1 -0
- package/dist/capabilities/generations.d.ts +79 -0
- package/dist/capabilities/generations.d.ts.map +1 -0
- package/dist/capabilities/generations.js +382 -0
- package/dist/capabilities/generations.js.map +1 -0
- package/dist/capabilities/generations.specs.d.ts +62 -0
- package/dist/capabilities/generations.specs.d.ts.map +1 -0
- package/dist/capabilities/generations.specs.js +230 -0
- package/dist/capabilities/generations.specs.js.map +1 -0
- package/dist/capabilities/google.d.ts +50 -0
- package/dist/capabilities/google.d.ts.map +1 -0
- package/dist/capabilities/google.js +246 -0
- package/dist/capabilities/google.js.map +1 -0
- package/dist/capabilities/google.specs.d.ts +52 -0
- package/dist/capabilities/google.specs.d.ts.map +1 -0
- package/dist/capabilities/google.specs.js +391 -0
- package/dist/capabilities/google.specs.js.map +1 -0
- package/dist/capabilities/html-text.d.ts +28 -0
- package/dist/capabilities/html-text.d.ts.map +1 -0
- package/dist/capabilities/html-text.js +76 -0
- package/dist/capabilities/html-text.js.map +1 -0
- package/dist/capabilities/index.d.ts +19 -0
- package/dist/capabilities/index.d.ts.map +1 -0
- package/dist/capabilities/index.js +14 -0
- package/dist/capabilities/index.js.map +1 -0
- package/dist/capabilities/invoke.d.ts +89 -0
- package/dist/capabilities/invoke.d.ts.map +1 -0
- package/dist/capabilities/invoke.js +214 -0
- package/dist/capabilities/invoke.js.map +1 -0
- package/dist/capabilities/jobs.d.ts +32 -0
- package/dist/capabilities/jobs.d.ts.map +1 -0
- package/dist/capabilities/jobs.js +123 -0
- package/dist/capabilities/jobs.js.map +1 -0
- package/dist/capabilities/jobs.specs.d.ts +20 -0
- package/dist/capabilities/jobs.specs.d.ts.map +1 -0
- package/dist/capabilities/jobs.specs.js +103 -0
- package/dist/capabilities/jobs.specs.js.map +1 -0
- package/dist/capabilities/js-scripts.d.ts +90 -0
- package/dist/capabilities/js-scripts.d.ts.map +1 -0
- package/dist/capabilities/js-scripts.js +700 -0
- package/dist/capabilities/js-scripts.js.map +1 -0
- package/dist/capabilities/js-scripts.specs.d.ts +50 -0
- package/dist/capabilities/js-scripts.specs.d.ts.map +1 -0
- package/dist/capabilities/js-scripts.specs.js +344 -0
- package/dist/capabilities/js-scripts.specs.js.map +1 -0
- package/dist/capabilities/lazy-tool.d.ts +42 -0
- package/dist/capabilities/lazy-tool.d.ts.map +1 -0
- package/dist/capabilities/lazy-tool.js +88 -0
- package/dist/capabilities/lazy-tool.js.map +1 -0
- package/dist/capabilities/local-browser.d.ts +24 -0
- package/dist/capabilities/local-browser.d.ts.map +1 -0
- package/dist/capabilities/local-browser.js +105 -0
- package/dist/capabilities/local-browser.js.map +1 -0
- package/dist/capabilities/media.d.ts +121 -0
- package/dist/capabilities/media.d.ts.map +1 -0
- package/dist/capabilities/media.js +2217 -0
- package/dist/capabilities/media.js.map +1 -0
- package/dist/capabilities/media.specs.d.ts +65 -0
- package/dist/capabilities/media.specs.d.ts.map +1 -0
- package/dist/capabilities/media.specs.js +779 -0
- package/dist/capabilities/media.specs.js.map +1 -0
- package/dist/capabilities/memory.d.ts +31 -0
- package/dist/capabilities/memory.d.ts.map +1 -0
- package/dist/capabilities/memory.js +403 -0
- package/dist/capabilities/memory.js.map +1 -0
- package/dist/capabilities/memory.specs.d.ts +55 -0
- package/dist/capabilities/memory.specs.d.ts.map +1 -0
- package/dist/capabilities/memory.specs.js +208 -0
- package/dist/capabilities/memory.specs.js.map +1 -0
- package/dist/capabilities/model3d.d.ts +29 -0
- package/dist/capabilities/model3d.d.ts.map +1 -0
- package/dist/capabilities/model3d.js +482 -0
- package/dist/capabilities/model3d.js.map +1 -0
- package/dist/capabilities/model3d.specs.d.ts +28 -0
- package/dist/capabilities/model3d.specs.d.ts.map +1 -0
- package/dist/capabilities/model3d.specs.js +305 -0
- package/dist/capabilities/model3d.specs.js.map +1 -0
- package/dist/capabilities/models.d.ts +138 -0
- package/dist/capabilities/models.d.ts.map +1 -0
- package/dist/capabilities/models.js +803 -0
- package/dist/capabilities/models.js.map +1 -0
- package/dist/capabilities/models.specs.d.ts +21 -0
- package/dist/capabilities/models.specs.d.ts.map +1 -0
- package/dist/capabilities/models.specs.js +141 -0
- package/dist/capabilities/models.specs.js.map +1 -0
- package/dist/capabilities/nodes.d.ts +26 -0
- package/dist/capabilities/nodes.d.ts.map +1 -0
- package/dist/capabilities/nodes.js +266 -0
- package/dist/capabilities/nodes.js.map +1 -0
- package/dist/capabilities/nodes.specs.d.ts +20 -0
- package/dist/capabilities/nodes.specs.d.ts.map +1 -0
- package/dist/capabilities/nodes.specs.js +106 -0
- package/dist/capabilities/nodes.specs.js.map +1 -0
- package/dist/capabilities/packs.d.ts +69 -0
- package/dist/capabilities/packs.d.ts.map +1 -0
- package/dist/capabilities/packs.js +246 -0
- package/dist/capabilities/packs.js.map +1 -0
- package/dist/capabilities/packs.specs.d.ts +27 -0
- package/dist/capabilities/packs.specs.d.ts.map +1 -0
- package/dist/capabilities/packs.specs.js +64 -0
- package/dist/capabilities/packs.specs.js.map +1 -0
- package/dist/capabilities/project-scope.d.ts +12 -0
- package/dist/capabilities/project-scope.d.ts.map +1 -0
- package/dist/capabilities/project-scope.js +19 -0
- package/dist/capabilities/project-scope.js.map +1 -0
- package/dist/capabilities/projects.d.ts +23 -0
- package/dist/capabilities/projects.d.ts.map +1 -0
- package/dist/capabilities/projects.js +206 -0
- package/dist/capabilities/projects.js.map +1 -0
- package/dist/capabilities/projects.specs.d.ts +23 -0
- package/dist/capabilities/projects.specs.d.ts.map +1 -0
- package/dist/capabilities/projects.specs.js +118 -0
- package/dist/capabilities/projects.specs.js.map +1 -0
- package/dist/capabilities/registry.d.ts +109 -0
- package/dist/capabilities/registry.d.ts.map +1 -0
- package/dist/capabilities/registry.js +445 -0
- package/dist/capabilities/registry.js.map +1 -0
- package/dist/capabilities/script-link.d.ts +22 -0
- package/dist/capabilities/script-link.d.ts.map +1 -0
- package/dist/capabilities/script-link.js +35 -0
- package/dist/capabilities/script-link.js.map +1 -0
- package/dist/capabilities/scripts.d.ts +43 -0
- package/dist/capabilities/scripts.d.ts.map +1 -0
- package/dist/capabilities/scripts.js +1789 -0
- package/dist/capabilities/scripts.js.map +1 -0
- package/dist/capabilities/scripts.specs.d.ts +40 -0
- package/dist/capabilities/scripts.specs.d.ts.map +1 -0
- package/dist/capabilities/scripts.specs.js +319 -0
- package/dist/capabilities/scripts.specs.js.map +1 -0
- package/dist/capabilities/serpapi.d.ts +42 -0
- package/dist/capabilities/serpapi.d.ts.map +1 -0
- package/dist/capabilities/serpapi.js +262 -0
- package/dist/capabilities/serpapi.js.map +1 -0
- package/dist/capabilities/serpapi.specs.d.ts +24 -0
- package/dist/capabilities/serpapi.specs.d.ts.map +1 -0
- package/dist/capabilities/serpapi.specs.js +174 -0
- package/dist/capabilities/serpapi.specs.js.map +1 -0
- package/dist/capabilities/settings.d.ts +48 -0
- package/dist/capabilities/settings.d.ts.map +1 -0
- package/dist/capabilities/settings.js +226 -0
- package/dist/capabilities/settings.js.map +1 -0
- package/dist/capabilities/settings.specs.d.ts +42 -0
- package/dist/capabilities/settings.specs.d.ts.map +1 -0
- package/dist/capabilities/settings.specs.js +126 -0
- package/dist/capabilities/settings.specs.js.map +1 -0
- package/dist/capabilities/shared.d.ts +35 -0
- package/dist/capabilities/shared.d.ts.map +1 -0
- package/dist/capabilities/shared.js +165 -0
- package/dist/capabilities/shared.js.map +1 -0
- package/dist/capabilities/shared.specs.d.ts +19 -0
- package/dist/capabilities/shared.specs.d.ts.map +1 -0
- package/dist/capabilities/shared.specs.js +134 -0
- package/dist/capabilities/shared.specs.js.map +1 -0
- package/dist/capabilities/sketches.d.ts +52 -0
- package/dist/capabilities/sketches.d.ts.map +1 -0
- package/dist/capabilities/sketches.js +1163 -0
- package/dist/capabilities/sketches.js.map +1 -0
- package/dist/capabilities/sketches.specs.d.ts +44 -0
- package/dist/capabilities/sketches.specs.d.ts.map +1 -0
- package/dist/capabilities/sketches.specs.js +392 -0
- package/dist/capabilities/sketches.specs.js.map +1 -0
- package/dist/capabilities/skills.d.ts +28 -0
- package/dist/capabilities/skills.d.ts.map +1 -0
- package/dist/capabilities/skills.js +269 -0
- package/dist/capabilities/skills.js.map +1 -0
- package/dist/capabilities/skills.specs.d.ts +17 -0
- package/dist/capabilities/skills.specs.d.ts.map +1 -0
- package/dist/capabilities/skills.specs.js +108 -0
- package/dist/capabilities/skills.specs.js.map +1 -0
- package/dist/capabilities/storyboards.d.ts +68 -0
- package/dist/capabilities/storyboards.d.ts.map +1 -0
- package/dist/capabilities/storyboards.js +2246 -0
- package/dist/capabilities/storyboards.js.map +1 -0
- package/dist/capabilities/storyboards.specs.d.ts +42 -0
- package/dist/capabilities/storyboards.specs.d.ts.map +1 -0
- package/dist/capabilities/storyboards.specs.js +418 -0
- package/dist/capabilities/storyboards.specs.js.map +1 -0
- package/dist/capabilities/threads.d.ts +25 -0
- package/dist/capabilities/threads.d.ts.map +1 -0
- package/dist/capabilities/threads.js +240 -0
- package/dist/capabilities/threads.js.map +1 -0
- package/dist/capabilities/threads.specs.d.ts +26 -0
- package/dist/capabilities/threads.specs.d.ts.map +1 -0
- package/dist/capabilities/threads.specs.js +108 -0
- package/dist/capabilities/threads.specs.js.map +1 -0
- package/dist/capabilities/timeline-audio-bake.d.ts +114 -0
- package/dist/capabilities/timeline-audio-bake.d.ts.map +1 -0
- package/dist/capabilities/timeline-audio-bake.js +188 -0
- package/dist/capabilities/timeline-audio-bake.js.map +1 -0
- package/dist/capabilities/timeline-bake.d.ts +21 -0
- package/dist/capabilities/timeline-bake.d.ts.map +1 -0
- package/dist/capabilities/timeline-bake.js +118 -0
- package/dist/capabilities/timeline-bake.js.map +1 -0
- package/dist/capabilities/timeline-isolate-subject.d.ts +157 -0
- package/dist/capabilities/timeline-isolate-subject.d.ts.map +1 -0
- package/dist/capabilities/timeline-isolate-subject.js +293 -0
- package/dist/capabilities/timeline-isolate-subject.js.map +1 -0
- package/dist/capabilities/timeline-native-media.d.ts +4 -0
- package/dist/capabilities/timeline-native-media.d.ts.map +1 -0
- package/dist/capabilities/timeline-native-media.js +920 -0
- package/dist/capabilities/timeline-native-media.js.map +1 -0
- package/dist/capabilities/timeline-track-object.d.ts +83 -0
- package/dist/capabilities/timeline-track-object.d.ts.map +1 -0
- package/dist/capabilities/timeline-track-object.js +195 -0
- package/dist/capabilities/timeline-track-object.js.map +1 -0
- package/dist/capabilities/timelines.d.ts +190 -0
- package/dist/capabilities/timelines.d.ts.map +1 -0
- package/dist/capabilities/timelines.js +2326 -0
- package/dist/capabilities/timelines.js.map +1 -0
- package/dist/capabilities/timelines.specs.d.ts +134 -0
- package/dist/capabilities/timelines.specs.d.ts.map +1 -0
- package/dist/capabilities/timelines.specs.js +1362 -0
- package/dist/capabilities/timelines.specs.js.map +1 -0
- package/dist/capabilities/types.d.ts +197 -0
- package/dist/capabilities/types.d.ts.map +1 -0
- package/dist/capabilities/types.js +19 -0
- package/dist/capabilities/types.js.map +1 -0
- package/dist/capabilities/ui.d.ts +38 -0
- package/dist/capabilities/ui.d.ts.map +1 -0
- package/dist/capabilities/ui.js +128 -0
- package/dist/capabilities/ui.js.map +1 -0
- package/dist/capabilities/ui.specs.d.ts +22 -0
- package/dist/capabilities/ui.specs.d.ts.map +1 -0
- package/dist/capabilities/ui.specs.js +59 -0
- package/dist/capabilities/ui.specs.js.map +1 -0
- package/dist/capabilities/video-production-acceptance.d.ts +27 -0
- package/dist/capabilities/video-production-acceptance.d.ts.map +1 -0
- package/dist/capabilities/video-production-acceptance.js +41 -0
- package/dist/capabilities/video-production-acceptance.js.map +1 -0
- package/dist/capabilities/video-production.d.ts +179 -0
- package/dist/capabilities/video-production.d.ts.map +1 -0
- package/dist/capabilities/video-production.js +1492 -0
- package/dist/capabilities/video-production.js.map +1 -0
- package/dist/capabilities/video-production.specs.d.ts +13 -0
- package/dist/capabilities/video-production.specs.d.ts.map +1 -0
- package/dist/capabilities/video-production.specs.js +256 -0
- package/dist/capabilities/video-production.specs.js.map +1 -0
- package/dist/capabilities/web.d.ts +86 -0
- package/dist/capabilities/web.d.ts.map +1 -0
- package/dist/capabilities/web.js +850 -0
- package/dist/capabilities/web.js.map +1 -0
- package/dist/capabilities/web.specs.d.ts +33 -0
- package/dist/capabilities/web.specs.d.ts.map +1 -0
- package/dist/capabilities/web.specs.js +286 -0
- package/dist/capabilities/web.specs.js.map +1 -0
- package/dist/capabilities/workflows.d.ts +58 -0
- package/dist/capabilities/workflows.d.ts.map +1 -0
- package/dist/capabilities/workflows.js +1530 -0
- package/dist/capabilities/workflows.js.map +1 -0
- package/dist/capabilities/workflows.specs.d.ts +63 -0
- package/dist/capabilities/workflows.specs.d.ts.map +1 -0
- package/dist/capabilities/workflows.specs.js +852 -0
- package/dist/capabilities/workflows.specs.js.map +1 -0
- package/dist/code-gen/analyze.d.ts +26 -0
- package/dist/code-gen/analyze.d.ts.map +1 -0
- package/dist/code-gen/analyze.js +79 -0
- package/dist/code-gen/analyze.js.map +1 -0
- package/dist/code-gen/port-types.d.ts +58 -0
- package/dist/code-gen/port-types.d.ts.map +1 -0
- package/dist/code-gen/port-types.js +115 -0
- package/dist/code-gen/port-types.js.map +1 -0
- package/dist/code-gen/prompt.d.ts +37 -0
- package/dist/code-gen/prompt.d.ts.map +1 -0
- package/dist/code-gen/prompt.js +129 -0
- package/dist/code-gen/prompt.js.map +1 -0
- package/dist/code-gen/sandbox-manifest.d.ts +120 -0
- package/dist/code-gen/sandbox-manifest.d.ts.map +1 -0
- package/dist/code-gen/sandbox-manifest.js +1023 -0
- package/dist/code-gen/sandbox-manifest.js.map +1 -0
- package/dist/code-gen/sandbox-prompt.d.ts +26 -0
- package/dist/code-gen/sandbox-prompt.d.ts.map +1 -0
- package/dist/code-gen/sandbox-prompt.js +126 -0
- package/dist/code-gen/sandbox-prompt.js.map +1 -0
- package/dist/code-planner.d.ts +72 -0
- package/dist/code-planner.d.ts.map +1 -0
- package/dist/code-planner.js +271 -0
- package/dist/code-planner.js.map +1 -0
- package/dist/codeact/action-diagnostics.d.ts +35 -0
- package/dist/codeact/action-diagnostics.d.ts.map +1 -0
- package/dist/codeact/action-diagnostics.js +147 -0
- package/dist/codeact/action-diagnostics.js.map +1 -0
- package/dist/codeact/capability-modules.d.ts +66 -0
- package/dist/codeact/capability-modules.d.ts.map +1 -0
- package/dist/codeact/capability-modules.js +278 -0
- package/dist/codeact/capability-modules.js.map +1 -0
- package/dist/codeact/chat-codeact.d.ts +104 -0
- package/dist/codeact/chat-codeact.d.ts.map +1 -0
- package/dist/codeact/chat-codeact.js +424 -0
- package/dist/codeact/chat-codeact.js.map +1 -0
- package/dist/codeact/codeact-executor.d.ts +247 -0
- package/dist/codeact/codeact-executor.d.ts.map +1 -0
- package/dist/codeact/codeact-executor.js +1176 -0
- package/dist/codeact/codeact-executor.js.map +1 -0
- package/dist/codeact/compact-ids.d.ts +28 -0
- package/dist/codeact/compact-ids.d.ts.map +1 -0
- package/dist/codeact/compact-ids.js +77 -0
- package/dist/codeact/compact-ids.js.map +1 -0
- package/dist/codeact/execute-code-contract.d.ts +97 -0
- package/dist/codeact/execute-code-contract.d.ts.map +1 -0
- package/dist/codeact/execute-code-contract.js +225 -0
- package/dist/codeact/execute-code-contract.js.map +1 -0
- package/dist/codeact/fabric-package.d.ts +29 -0
- package/dist/codeact/fabric-package.d.ts.map +1 -0
- package/dist/codeact/fabric-package.js +62 -0
- package/dist/codeact/fabric-package.js.map +1 -0
- package/dist/codeact/flow-package.d.ts +31 -0
- package/dist/codeact/flow-package.d.ts.map +1 -0
- package/dist/codeact/flow-package.js +80 -0
- package/dist/codeact/flow-package.js.map +1 -0
- package/dist/codeact/graph-dsl-package.d.ts +42 -0
- package/dist/codeact/graph-dsl-package.d.ts.map +1 -0
- package/dist/codeact/graph-dsl-package.js +100 -0
- package/dist/codeact/graph-dsl-package.js.map +1 -0
- package/dist/codeact/graph-model.d.ts +30 -0
- package/dist/codeact/graph-model.d.ts.map +1 -0
- package/dist/codeact/graph-model.js +408 -0
- package/dist/codeact/graph-model.js.map +1 -0
- package/dist/codeact/mcp-guest-contract.d.ts +66 -0
- package/dist/codeact/mcp-guest-contract.d.ts.map +1 -0
- package/dist/codeact/mcp-guest-contract.js +108 -0
- package/dist/codeact/mcp-guest-contract.js.map +1 -0
- package/dist/codeact/nodetool-api.d.ts +60 -0
- package/dist/codeact/nodetool-api.d.ts.map +1 -0
- package/dist/codeact/nodetool-api.js +2056 -0
- package/dist/codeact/nodetool-api.js.map +1 -0
- package/dist/codeact/prompt.d.ts +88 -0
- package/dist/codeact/prompt.d.ts.map +1 -0
- package/dist/codeact/prompt.js +326 -0
- package/dist/codeact/prompt.js.map +1 -0
- package/dist/codeact/sandbox-package-docs.d.ts +48 -0
- package/dist/codeact/sandbox-package-docs.d.ts.map +1 -0
- package/dist/codeact/sandbox-package-docs.js +55 -0
- package/dist/codeact/sandbox-package-docs.js.map +1 -0
- package/dist/codeact/sandbox-package-listing.d.ts +81 -0
- package/dist/codeact/sandbox-package-listing.d.ts.map +1 -0
- package/dist/codeact/sandbox-package-listing.js +101 -0
- package/dist/codeact/sandbox-package-listing.js.map +1 -0
- package/dist/codeact/sandbox-packages.d.ts +72 -0
- package/dist/codeact/sandbox-packages.d.ts.map +1 -0
- package/dist/codeact/sandbox-packages.js +230 -0
- package/dist/codeact/sandbox-packages.js.map +1 -0
- package/dist/codeact/timeline-package.d.ts +5 -0
- package/dist/codeact/timeline-package.d.ts.map +1 -0
- package/dist/codeact/timeline-package.js +10 -0
- package/dist/codeact/timeline-package.js.map +1 -0
- package/dist/codeact/tool-api.d.ts +197 -0
- package/dist/codeact/tool-api.d.ts.map +1 -0
- package/dist/codeact/tool-api.js +445 -0
- package/dist/codeact/tool-api.js.map +1 -0
- package/dist/codeact/tools-prelude.d.ts +32 -0
- package/dist/codeact/tools-prelude.d.ts.map +1 -0
- package/dist/codeact/tools-prelude.js +65 -0
- package/dist/codeact/tools-prelude.js.map +1 -0
- package/dist/constants.d.ts +47 -2
- package/dist/constants.d.ts.map +1 -1
- package/dist/constants.js +64 -2
- package/dist/constants.js.map +1 -1
- package/dist/custom-animation-bake.d.ts +57 -0
- package/dist/custom-animation-bake.d.ts.map +1 -0
- package/dist/custom-animation-bake.js +88 -0
- package/dist/custom-animation-bake.js.map +1 -0
- package/dist/dynamic-slots.d.ts +72 -0
- package/dist/dynamic-slots.d.ts.map +1 -0
- package/dist/dynamic-slots.js +0 -0
- package/dist/dynamic-slots.js.map +1 -0
- package/dist/evals/app-build-cases.d.ts +35 -0
- package/dist/evals/app-build-cases.d.ts.map +1 -0
- package/dist/evals/app-build-cases.js +722 -0
- package/dist/evals/app-build-cases.js.map +1 -0
- package/dist/evals/app-build-eval.d.ts +208 -0
- package/dist/evals/app-build-eval.d.ts.map +1 -0
- package/dist/evals/app-build-eval.js +518 -0
- package/dist/evals/app-build-eval.js.map +1 -0
- package/dist/evals/app-build-plan.d.ts +76 -0
- package/dist/evals/app-build-plan.d.ts.map +1 -0
- package/dist/evals/app-build-plan.js +140 -0
- package/dist/evals/app-build-plan.js.map +1 -0
- package/dist/evals/code-gen-cases.d.ts +13 -0
- package/dist/evals/code-gen-cases.d.ts.map +1 -0
- package/dist/evals/code-gen-cases.js +124 -0
- package/dist/evals/code-gen-cases.js.map +1 -0
- package/dist/evals/code-gen-eval.d.ts +109 -0
- package/dist/evals/code-gen-eval.d.ts.map +1 -0
- package/dist/evals/code-gen-eval.js +299 -0
- package/dist/evals/code-gen-eval.js.map +1 -0
- package/dist/evals/codeact-api-cases.d.ts +14 -0
- package/dist/evals/codeact-api-cases.d.ts.map +1 -0
- package/dist/evals/codeact-api-cases.js +29 -0
- package/dist/evals/codeact-api-cases.js.map +1 -0
- package/dist/evals/codeact-api-core.d.ts +23 -0
- package/dist/evals/codeact-api-core.d.ts.map +1 -0
- package/dist/evals/codeact-api-core.js +1411 -0
- package/dist/evals/codeact-api-core.js.map +1 -0
- package/dist/evals/codeact-api-surfaces.d.ts +25 -0
- package/dist/evals/codeact-api-surfaces.d.ts.map +1 -0
- package/dist/evals/codeact-api-surfaces.js +2191 -0
- package/dist/evals/codeact-api-surfaces.js.map +1 -0
- package/dist/evals/codeact-cases.d.ts +76 -0
- package/dist/evals/codeact-cases.d.ts.map +1 -0
- package/dist/evals/codeact-cases.js +151 -0
- package/dist/evals/codeact-cases.js.map +1 -0
- package/dist/evals/codeact-eval.d.ts +66 -0
- package/dist/evals/codeact-eval.d.ts.map +1 -0
- package/dist/evals/codeact-eval.js +250 -0
- package/dist/evals/codeact-eval.js.map +1 -0
- package/dist/evals/codeact-sandbox-pack-cases.d.ts +31 -0
- package/dist/evals/codeact-sandbox-pack-cases.d.ts.map +1 -0
- package/dist/evals/codeact-sandbox-pack-cases.js +219 -0
- package/dist/evals/codeact-sandbox-pack-cases.js.map +1 -0
- package/dist/evals/escalation-cases.d.ts +22 -0
- package/dist/evals/escalation-cases.d.ts.map +1 -0
- package/dist/evals/escalation-cases.js +343 -0
- package/dist/evals/escalation-cases.js.map +1 -0
- package/dist/evals/escalation.d.ts +92 -0
- package/dist/evals/escalation.d.ts.map +1 -0
- package/dist/evals/escalation.js +152 -0
- package/dist/evals/escalation.js.map +1 -0
- package/dist/evals/goal-judge.d.ts +51 -0
- package/dist/evals/goal-judge.d.ts.map +1 -0
- package/dist/evals/goal-judge.js +152 -0
- package/dist/evals/goal-judge.js.map +1 -0
- package/dist/evals/graph-e2e-cases.d.ts +21 -0
- package/dist/evals/graph-e2e-cases.d.ts.map +1 -0
- package/dist/evals/graph-e2e-cases.js +262 -0
- package/dist/evals/graph-e2e-cases.js.map +1 -0
- package/dist/evals/graph-e2e-eval.d.ts +202 -0
- package/dist/evals/graph-e2e-eval.d.ts.map +1 -0
- package/dist/evals/graph-e2e-eval.js +438 -0
- package/dist/evals/graph-e2e-eval.js.map +1 -0
- package/dist/evals/graph-planner-cases.d.ts +12 -0
- package/dist/evals/graph-planner-cases.d.ts.map +1 -0
- package/dist/evals/graph-planner-cases.js +256 -0
- package/dist/evals/graph-planner-cases.js.map +1 -0
- package/dist/evals/graph-planner-eval.d.ts +178 -0
- package/dist/evals/graph-planner-eval.d.ts.map +1 -0
- package/dist/evals/graph-planner-eval.js +495 -0
- package/dist/evals/graph-planner-eval.js.map +1 -0
- package/dist/evals/planner-tools.d.ts +21 -0
- package/dist/evals/planner-tools.d.ts.map +1 -0
- package/dist/evals/planner-tools.js +66 -0
- package/dist/evals/planner-tools.js.map +1 -0
- package/dist/evals/subtask-cases.d.ts +88 -0
- package/dist/evals/subtask-cases.d.ts.map +1 -0
- package/dist/evals/subtask-cases.js +322 -0
- package/dist/evals/subtask-cases.js.map +1 -0
- package/dist/evals/subtask-eval.d.ts +129 -0
- package/dist/evals/subtask-eval.d.ts.map +1 -0
- package/dist/evals/subtask-eval.js +397 -0
- package/dist/evals/subtask-eval.js.map +1 -0
- package/dist/evals/surfaces/app.d.ts +19 -0
- package/dist/evals/surfaces/app.d.ts.map +1 -0
- package/dist/evals/surfaces/app.js +341 -0
- package/dist/evals/surfaces/app.js.map +1 -0
- package/dist/evals/surfaces/creative-pipeline.d.ts +203 -0
- package/dist/evals/surfaces/creative-pipeline.d.ts.map +1 -0
- package/dist/evals/surfaces/creative-pipeline.js +909 -0
- package/dist/evals/surfaces/creative-pipeline.js.map +1 -0
- package/dist/evals/surfaces/game.d.ts +5 -0
- package/dist/evals/surfaces/game.d.ts.map +1 -0
- package/dist/evals/surfaces/game.js +42 -0
- package/dist/evals/surfaces/game.js.map +1 -0
- package/dist/evals/surfaces/js-script.d.ts +66 -0
- package/dist/evals/surfaces/js-script.d.ts.map +1 -0
- package/dist/evals/surfaces/js-script.js +407 -0
- package/dist/evals/surfaces/js-script.js.map +1 -0
- package/dist/evals/surfaces/memory.d.ts +34 -0
- package/dist/evals/surfaces/memory.d.ts.map +1 -0
- package/dist/evals/surfaces/memory.js +264 -0
- package/dist/evals/surfaces/memory.js.map +1 -0
- package/dist/evals/surfaces/model3d.d.ts +45 -0
- package/dist/evals/surfaces/model3d.d.ts.map +1 -0
- package/dist/evals/surfaces/model3d.js +346 -0
- package/dist/evals/surfaces/model3d.js.map +1 -0
- package/dist/evals/surfaces/script.d.ts +96 -0
- package/dist/evals/surfaces/script.d.ts.map +1 -0
- package/dist/evals/surfaces/script.js +709 -0
- package/dist/evals/surfaces/script.js.map +1 -0
- package/dist/evals/surfaces/sketch.d.ts +124 -0
- package/dist/evals/surfaces/sketch.d.ts.map +1 -0
- package/dist/evals/surfaces/sketch.js +1298 -0
- package/dist/evals/surfaces/sketch.js.map +1 -0
- package/dist/evals/surfaces/storyboard.d.ts +97 -0
- package/dist/evals/surfaces/storyboard.d.ts.map +1 -0
- package/dist/evals/surfaces/storyboard.js +693 -0
- package/dist/evals/surfaces/storyboard.js.map +1 -0
- package/dist/evals/surfaces/timeline.d.ts +416 -0
- package/dist/evals/surfaces/timeline.d.ts.map +1 -0
- package/dist/evals/surfaces/timeline.js +3267 -0
- package/dist/evals/surfaces/timeline.js.map +1 -0
- package/dist/evals/task-planner-cases.d.ts +57 -0
- package/dist/evals/task-planner-cases.d.ts.map +1 -0
- package/dist/evals/task-planner-cases.js +117 -0
- package/dist/evals/task-planner-cases.js.map +1 -0
- package/dist/evals/task-planner-eval.d.ts +101 -0
- package/dist/evals/task-planner-eval.d.ts.map +1 -0
- package/dist/evals/task-planner-eval.js +380 -0
- package/dist/evals/task-planner-eval.js.map +1 -0
- package/dist/evals/tool-loop-bridge.d.ts +57 -0
- package/dist/evals/tool-loop-bridge.d.ts.map +1 -0
- package/dist/evals/tool-loop-bridge.js +129 -0
- package/dist/evals/tool-loop-bridge.js.map +1 -0
- package/dist/evals/tool-loop-cases.d.ts +21 -0
- package/dist/evals/tool-loop-cases.d.ts.map +1 -0
- package/dist/evals/tool-loop-cases.js +401 -0
- package/dist/evals/tool-loop-cases.js.map +1 -0
- package/dist/evals/tool-loop-eval.d.ts +234 -0
- package/dist/evals/tool-loop-eval.d.ts.map +1 -0
- package/dist/evals/tool-loop-eval.js +373 -0
- package/dist/evals/tool-loop-eval.js.map +1 -0
- package/dist/evals/tool-loop-permission.d.ts +43 -0
- package/dist/evals/tool-loop-permission.d.ts.map +1 -0
- package/dist/evals/tool-loop-permission.js +76 -0
- package/dist/evals/tool-loop-permission.js.map +1 -0
- package/dist/game-examples.d.ts +447 -0
- package/dist/game-examples.d.ts.map +1 -0
- package/dist/game-examples.js +187 -0
- package/dist/game-examples.js.map +1 -0
- package/dist/graph-builder.d.ts +55 -3
- package/dist/graph-builder.d.ts.map +1 -1
- package/dist/graph-builder.js +159 -4
- package/dist/graph-builder.js.map +1 -1
- package/dist/graph-dsl-core.d.ts +25 -0
- package/dist/graph-dsl-core.d.ts.map +1 -0
- package/dist/graph-dsl-core.js +312 -0
- package/dist/graph-dsl-core.js.map +1 -0
- package/dist/graph-dsl.d.ts +30 -0
- package/dist/graph-dsl.d.ts.map +1 -0
- package/dist/graph-dsl.js +116 -0
- package/dist/graph-dsl.js.map +1 -0
- package/dist/host-binaries.d.ts +6 -0
- package/dist/host-binaries.d.ts.map +1 -0
- package/dist/host-binaries.js +6 -0
- package/dist/host-binaries.js.map +1 -0
- package/dist/host-binary-guard.d.ts +100 -0
- package/dist/host-binary-guard.d.ts.map +1 -0
- package/dist/host-binary-guard.js +205 -0
- package/dist/host-binary-guard.js.map +1 -0
- package/dist/host-modules/aws.d.ts +57 -0
- package/dist/host-modules/aws.d.ts.map +1 -0
- package/dist/host-modules/aws.js +309 -0
- package/dist/host-modules/aws.js.map +1 -0
- package/dist/host-modules/chrono.d.ts +15 -0
- package/dist/host-modules/chrono.d.ts.map +1 -0
- package/dist/host-modules/chrono.js +35 -0
- package/dist/host-modules/chrono.js.map +1 -0
- package/dist/host-modules/csv.d.ts +17 -0
- package/dist/host-modules/csv.d.ts.map +1 -0
- package/dist/host-modules/csv.js +64 -0
- package/dist/host-modules/csv.js.map +1 -0
- package/dist/host-modules/diff.d.ts +9 -0
- package/dist/host-modules/diff.d.ts.map +1 -0
- package/dist/host-modules/diff.js +26 -0
- package/dist/host-modules/diff.js.map +1 -0
- package/dist/host-modules/dispatcher.d.ts +41 -0
- package/dist/host-modules/dispatcher.d.ts.map +1 -0
- package/dist/host-modules/dispatcher.js +99 -0
- package/dist/host-modules/dispatcher.js.map +1 -0
- package/dist/host-modules/docx.d.ts +19 -0
- package/dist/host-modules/docx.d.ts.map +1 -0
- package/dist/host-modules/docx.js +120 -0
- package/dist/host-modules/docx.js.map +1 -0
- package/dist/host-modules/epub.d.ts +48 -0
- package/dist/host-modules/epub.d.ts.map +1 -0
- package/dist/host-modules/epub.js +100 -0
- package/dist/host-modules/epub.js.map +1 -0
- package/dist/host-modules/exif.d.ts +8 -0
- package/dist/host-modules/exif.d.ts.map +1 -0
- package/dist/host-modules/exif.js +19 -0
- package/dist/host-modules/exif.js.map +1 -0
- package/dist/host-modules/expr.d.ts +16 -0
- package/dist/host-modules/expr.d.ts.map +1 -0
- package/dist/host-modules/expr.js +18 -0
- package/dist/host-modules/expr.js.map +1 -0
- package/dist/host-modules/fabric.d.ts +46 -0
- package/dist/host-modules/fabric.d.ts.map +1 -0
- package/dist/host-modules/fabric.js +178 -0
- package/dist/host-modules/fabric.js.map +1 -0
- package/dist/host-modules/html.d.ts +51 -0
- package/dist/host-modules/html.d.ts.map +1 -0
- package/dist/host-modules/html.js +198 -0
- package/dist/host-modules/html.js.map +1 -0
- package/dist/host-modules/ics.d.ts +9 -0
- package/dist/host-modules/ics.d.ts.map +1 -0
- package/dist/host-modules/ics.js +35 -0
- package/dist/host-modules/ics.js.map +1 -0
- package/dist/host-modules/limits.d.ts +41 -0
- package/dist/host-modules/limits.d.ts.map +1 -0
- package/dist/host-modules/limits.js +76 -0
- package/dist/host-modules/limits.js.map +1 -0
- package/dist/host-modules/mammoth.d.ts +11 -0
- package/dist/host-modules/mammoth.d.ts.map +1 -0
- package/dist/host-modules/mammoth.js +29 -0
- package/dist/host-modules/mammoth.js.map +1 -0
- package/dist/host-modules/notion.d.ts +45 -0
- package/dist/host-modules/notion.d.ts.map +1 -0
- package/dist/host-modules/notion.js +128 -0
- package/dist/host-modules/notion.js.map +1 -0
- package/dist/host-modules/ocr.d.ts +25 -0
- package/dist/host-modules/ocr.d.ts.map +1 -0
- package/dist/host-modules/ocr.js +74 -0
- package/dist/host-modules/ocr.js.map +1 -0
- package/dist/host-modules/pdf.d.ts +21 -0
- package/dist/host-modules/pdf.d.ts.map +1 -0
- package/dist/host-modules/pdf.js +53 -0
- package/dist/host-modules/pdf.js.map +1 -0
- package/dist/host-modules/pdflib.d.ts +17 -0
- package/dist/host-modules/pdflib.d.ts.map +1 -0
- package/dist/host-modules/pdflib.js +138 -0
- package/dist/host-modules/pdflib.js.map +1 -0
- package/dist/host-modules/pptx.d.ts +19 -0
- package/dist/host-modules/pptx.d.ts.map +1 -0
- package/dist/host-modules/pptx.js +67 -0
- package/dist/host-modules/pptx.js.map +1 -0
- package/dist/host-modules/pptxgen.d.ts +13 -0
- package/dist/host-modules/pptxgen.d.ts.map +1 -0
- package/dist/host-modules/pptxgen.js +98 -0
- package/dist/host-modules/pptxgen.js.map +1 -0
- package/dist/host-modules/prepared-request.d.ts +41 -0
- package/dist/host-modules/prepared-request.d.ts.map +1 -0
- package/dist/host-modules/prepared-request.js +104 -0
- package/dist/host-modules/prepared-request.js.map +1 -0
- package/dist/host-modules/registry.d.ts +31 -0
- package/dist/host-modules/registry.d.ts.map +1 -0
- package/dist/host-modules/registry.js +193 -0
- package/dist/host-modules/registry.js.map +1 -0
- package/dist/host-modules/subtitle.d.ts +15 -0
- package/dist/host-modules/subtitle.d.ts.map +1 -0
- package/dist/host-modules/subtitle.js +45 -0
- package/dist/host-modules/subtitle.js.map +1 -0
- package/dist/host-modules/supabase.d.ts +42 -0
- package/dist/host-modules/supabase.d.ts.map +1 -0
- package/dist/host-modules/supabase.js +141 -0
- package/dist/host-modules/supabase.js.map +1 -0
- package/dist/host-modules/tfjs.d.ts +42 -0
- package/dist/host-modules/tfjs.d.ts.map +1 -0
- package/dist/host-modules/tfjs.js +186 -0
- package/dist/host-modules/tfjs.js.map +1 -0
- package/dist/host-modules/tokens.d.ts +17 -0
- package/dist/host-modules/tokens.d.ts.map +1 -0
- package/dist/host-modules/tokens.js +74 -0
- package/dist/host-modules/tokens.js.map +1 -0
- package/dist/host-modules/twilio.d.ts +34 -0
- package/dist/host-modules/twilio.d.ts.map +1 -0
- package/dist/host-modules/twilio.js +67 -0
- package/dist/host-modules/twilio.js.map +1 -0
- package/dist/host-modules/xlsx.d.ts +26 -0
- package/dist/host-modules/xlsx.d.ts.map +1 -0
- package/dist/host-modules/xlsx.js +323 -0
- package/dist/host-modules/xlsx.js.map +1 -0
- package/dist/host-modules/xml.d.ts +23 -0
- package/dist/host-modules/xml.d.ts.map +1 -0
- package/dist/host-modules/xml.js +38 -0
- package/dist/host-modules/xml.js.map +1 -0
- package/dist/host-modules/zip.d.ts +24 -0
- package/dist/host-modules/zip.d.ts.map +1 -0
- package/dist/host-modules/zip.js +70 -0
- package/dist/host-modules/zip.js.map +1 -0
- package/dist/index.d.ts +197 -60
- package/dist/index.d.ts.map +1 -1
- package/dist/index.js +153 -60
- package/dist/index.js.map +1 -1
- package/dist/js-sandbox-worker/host.d.ts +136 -0
- package/dist/js-sandbox-worker/host.d.ts.map +1 -0
- package/dist/js-sandbox-worker/host.js +482 -0
- package/dist/js-sandbox-worker/host.js.map +1 -0
- package/dist/js-sandbox-worker/interpreter.d.ts +276 -0
- package/dist/js-sandbox-worker/interpreter.d.ts.map +1 -0
- package/dist/js-sandbox-worker/interpreter.js +1201 -0
- package/dist/js-sandbox-worker/interpreter.js.map +1 -0
- package/dist/js-sandbox-worker/protocol.d.ts +168 -0
- package/dist/js-sandbox-worker/protocol.d.ts.map +1 -0
- package/dist/js-sandbox-worker/protocol.js +217 -0
- package/dist/js-sandbox-worker/protocol.js.map +1 -0
- package/dist/js-sandbox-worker/worker-entry.d.ts +26 -0
- package/dist/js-sandbox-worker/worker-entry.d.ts.map +1 -0
- package/dist/js-sandbox-worker/worker-entry.js +307 -0
- package/dist/js-sandbox-worker/worker-entry.js.map +1 -0
- package/dist/js-sandbox.d.ts +455 -15
- package/dist/js-sandbox.d.ts.map +1 -1
- package/dist/js-sandbox.js +1961 -309
- package/dist/js-sandbox.js.map +1 -1
- package/dist/js-script-app-runner.d.ts +9 -0
- package/dist/js-script-app-runner.d.ts.map +1 -0
- package/dist/js-script-app-runner.js +49 -0
- package/dist/js-script-app-runner.js.map +1 -0
- package/dist/js-script-sandbox.d.ts +33 -0
- package/dist/js-script-sandbox.d.ts.map +1 -0
- package/dist/js-script-sandbox.js +82 -0
- package/dist/js-script-sandbox.js.map +1 -0
- package/dist/jtbd/friction.d.ts +26 -0
- package/dist/jtbd/friction.d.ts.map +1 -0
- package/dist/jtbd/friction.js +169 -0
- package/dist/jtbd/friction.js.map +1 -0
- package/dist/jtbd/index.d.ts +6 -0
- package/dist/jtbd/index.d.ts.map +1 -0
- package/dist/jtbd/index.js +6 -0
- package/dist/jtbd/index.js.map +1 -0
- package/dist/jtbd/optimize.d.ts +71 -0
- package/dist/jtbd/optimize.d.ts.map +1 -0
- package/dist/jtbd/optimize.js +176 -0
- package/dist/jtbd/optimize.js.map +1 -0
- package/dist/jtbd/registry.d.ts +29 -0
- package/dist/jtbd/registry.d.ts.map +1 -0
- package/dist/jtbd/registry.js +451 -0
- package/dist/jtbd/registry.js.map +1 -0
- package/dist/jtbd/run.d.ts +70 -0
- package/dist/jtbd/run.d.ts.map +1 -0
- package/dist/jtbd/run.js +177 -0
- package/dist/jtbd/run.js.map +1 -0
- package/dist/jtbd/types.d.ts +141 -0
- package/dist/jtbd/types.d.ts.map +1 -0
- package/dist/jtbd/types.js +25 -0
- package/dist/jtbd/types.js.map +1 -0
- package/dist/model-prompting-skills.d.ts +45 -0
- package/dist/model-prompting-skills.d.ts.map +1 -0
- package/dist/model-prompting-skills.js +116 -0
- package/dist/model-prompting-skills.js.map +1 -0
- package/dist/network-guard.d.ts +50 -0
- package/dist/network-guard.d.ts.map +1 -0
- package/dist/network-guard.js +106 -0
- package/dist/network-guard.js.map +1 -0
- package/dist/normalize-model-properties.d.ts +26 -0
- package/dist/normalize-model-properties.d.ts.map +1 -0
- package/dist/normalize-model-properties.js +58 -0
- package/dist/normalize-model-properties.js.map +1 -0
- package/dist/output-format.d.ts +5 -1
- package/dist/output-format.d.ts.map +1 -1
- package/dist/output-format.js.map +1 -1
- package/dist/parallel-task-executor.d.ts +68 -37
- package/dist/parallel-task-executor.d.ts.map +1 -1
- package/dist/parallel-task-executor.js +234 -203
- package/dist/parallel-task-executor.js.map +1 -1
- package/dist/prompts/background-subtask-prompt.d.ts +14 -0
- package/dist/prompts/background-subtask-prompt.d.ts.map +1 -0
- package/dist/prompts/background-subtask-prompt.js +64 -0
- package/dist/prompts/background-subtask-prompt.js.map +1 -0
- package/dist/prompts/product-knowledge.d.ts +3 -0
- package/dist/prompts/product-knowledge.d.ts.map +1 -0
- package/dist/prompts/product-knowledge.js +79 -0
- package/dist/prompts/product-knowledge.js.map +1 -0
- package/dist/prompts/read-only-search-prompt.d.ts +33 -0
- package/dist/prompts/read-only-search-prompt.d.ts.map +1 -0
- package/dist/prompts/read-only-search-prompt.js +109 -0
- package/dist/prompts/read-only-search-prompt.js.map +1 -0
- package/dist/prompts/sandbox-pack-catalog.d.ts +23 -0
- package/dist/prompts/sandbox-pack-catalog.d.ts.map +1 -0
- package/dist/prompts/sandbox-pack-catalog.js +40 -0
- package/dist/prompts/sandbox-pack-catalog.js.map +1 -0
- package/dist/prompts/security-monitor-prompt.d.ts +32 -0
- package/dist/prompts/security-monitor-prompt.d.ts.map +1 -0
- package/dist/prompts/security-monitor-prompt.js +99 -0
- package/dist/prompts/security-monitor-prompt.js.map +1 -0
- package/dist/prompts/workflow-authoring-knowledge.d.ts +77 -0
- package/dist/prompts/workflow-authoring-knowledge.d.ts.map +1 -0
- package/dist/prompts/workflow-authoring-knowledge.js +319 -0
- package/dist/prompts/workflow-authoring-knowledge.js.map +1 -0
- package/dist/run-policy.d.ts +25 -0
- package/dist/run-policy.d.ts.map +1 -0
- package/dist/run-policy.js +54 -0
- package/dist/run-policy.js.map +1 -0
- package/dist/sandbox-av-media.d.ts +33 -0
- package/dist/sandbox-av-media.d.ts.map +1 -0
- package/dist/sandbox-av-media.js +666 -0
- package/dist/sandbox-av-media.js.map +1 -0
- package/dist/sandbox-bootstrap-modules.d.ts +54 -0
- package/dist/sandbox-bootstrap-modules.d.ts.map +1 -0
- package/dist/sandbox-bootstrap-modules.js +104 -0
- package/dist/sandbox-bootstrap-modules.js.map +1 -0
- package/dist/sandbox-bytes.d.ts +43 -0
- package/dist/sandbox-bytes.d.ts.map +1 -0
- package/dist/sandbox-bytes.js +61 -0
- package/dist/sandbox-bytes.js.map +1 -0
- package/dist/sandbox-canvas-api.d.ts +30 -0
- package/dist/sandbox-canvas-api.d.ts.map +1 -0
- package/dist/sandbox-canvas-api.js +81 -0
- package/dist/sandbox-canvas-api.js.map +1 -0
- package/dist/sandbox-constants.d.ts +5 -0
- package/dist/sandbox-constants.d.ts.map +1 -0
- package/dist/sandbox-constants.js +13 -0
- package/dist/sandbox-constants.js.map +1 -0
- package/dist/sandbox-json-transport.d.ts +107 -0
- package/dist/sandbox-json-transport.d.ts.map +1 -0
- package/dist/sandbox-json-transport.js +324 -0
- package/dist/sandbox-json-transport.js.map +1 -0
- package/dist/sandbox-media-handle.d.ts +78 -0
- package/dist/sandbox-media-handle.d.ts.map +1 -0
- package/dist/sandbox-media-handle.js +114 -0
- package/dist/sandbox-media-handle.js.map +1 -0
- package/dist/sandbox-media-ref.d.ts +134 -0
- package/dist/sandbox-media-ref.d.ts.map +1 -0
- package/dist/sandbox-media-ref.js +442 -0
- package/dist/sandbox-media-ref.js.map +1 -0
- package/dist/sandbox-media.d.ts +177 -0
- package/dist/sandbox-media.d.ts.map +1 -0
- package/dist/sandbox-media.js +963 -0
- package/dist/sandbox-media.js.map +1 -0
- package/dist/sandbox-module-fixtures.d.ts +445 -0
- package/dist/sandbox-module-fixtures.d.ts.map +1 -0
- package/dist/sandbox-module-fixtures.js +214 -0
- package/dist/sandbox-module-fixtures.js.map +1 -0
- package/dist/sandbox-toolbelt.d.ts +13 -0
- package/dist/sandbox-toolbelt.d.ts.map +1 -0
- package/dist/sandbox-toolbelt.js +64 -0
- package/dist/sandbox-toolbelt.js.map +1 -0
- package/dist/security-monitor.d.ts +89 -0
- package/dist/security-monitor.d.ts.map +1 -0
- package/dist/security-monitor.js +307 -0
- package/dist/security-monitor.js.map +1 -0
- package/dist/serpapi/catalog.d.ts +101 -0
- package/dist/serpapi/catalog.d.ts.map +1 -0
- package/dist/serpapi/catalog.js +246 -0
- package/dist/serpapi/catalog.js.map +1 -0
- package/dist/serpapi/client.d.ts +81 -0
- package/dist/serpapi/client.d.ts.map +1 -0
- package/dist/serpapi/client.js +228 -0
- package/dist/serpapi/client.js.map +1 -0
- package/dist/serpapi/errors.d.ts +49 -0
- package/dist/serpapi/errors.d.ts.map +1 -0
- package/dist/serpapi/errors.js +64 -0
- package/dist/serpapi/errors.js.map +1 -0
- package/dist/serpapi/normalize.d.ts +58 -0
- package/dist/serpapi/normalize.d.ts.map +1 -0
- package/dist/serpapi/normalize.js +140 -0
- package/dist/serpapi/normalize.js.map +1 -0
- package/dist/skill-prompt.d.ts +27 -0
- package/dist/skill-prompt.d.ts.map +1 -0
- package/dist/skill-prompt.js +61 -0
- package/dist/skill-prompt.js.map +1 -0
- package/dist/step-executor.d.ts +90 -77
- package/dist/step-executor.d.ts.map +1 -1
- package/dist/step-executor.js +384 -690
- package/dist/step-executor.js.map +1 -1
- package/dist/subagent.d.ts +275 -0
- package/dist/subagent.d.ts.map +1 -0
- package/dist/subagent.js +444 -0
- package/dist/subagent.js.map +1 -0
- package/dist/supervisor/prompt.d.ts +12 -0
- package/dist/supervisor/prompt.d.ts.map +1 -0
- package/dist/supervisor/prompt.js +51 -0
- package/dist/supervisor/prompt.js.map +1 -0
- package/dist/supervisor/supervisor-agent.d.ts +72 -0
- package/dist/supervisor/supervisor-agent.d.ts.map +1 -0
- package/dist/supervisor/supervisor-agent.js +185 -0
- package/dist/supervisor/supervisor-agent.js.map +1 -0
- package/dist/supervisor/tools.d.ts +56 -0
- package/dist/supervisor/tools.d.ts.map +1 -0
- package/dist/supervisor/tools.js +83 -0
- package/dist/supervisor/tools.js.map +1 -0
- package/dist/supervisor/verdict-schema.d.ts +18 -0
- package/dist/supervisor/verdict-schema.d.ts.map +1 -0
- package/dist/supervisor/verdict-schema.js +58 -0
- package/dist/supervisor/verdict-schema.js.map +1 -0
- package/dist/system-skills.d.ts +68 -0
- package/dist/system-skills.d.ts.map +1 -0
- package/dist/system-skills.js +168 -0
- package/dist/system-skills.js.map +1 -0
- package/dist/task-executor.d.ts +90 -40
- package/dist/task-executor.d.ts.map +1 -1
- package/dist/task-executor.js +259 -282
- package/dist/task-executor.js.map +1 -1
- package/dist/task-planner.d.ts +18 -12
- package/dist/task-planner.d.ts.map +1 -1
- package/dist/task-planner.js +342 -328
- package/dist/task-planner.js.map +1 -1
- package/dist/timeline-preview/compare.d.ts +27 -0
- package/dist/timeline-preview/compare.d.ts.map +1 -0
- package/dist/timeline-preview/compare.js +71 -0
- package/dist/timeline-preview/compare.js.map +1 -0
- package/dist/timeline-preview/frames.d.ts +167 -0
- package/dist/timeline-preview/frames.d.ts.map +1 -0
- package/dist/timeline-preview/frames.js +646 -0
- package/dist/timeline-preview/frames.js.map +1 -0
- package/dist/timeline-preview/model3d.d.ts +67 -0
- package/dist/timeline-preview/model3d.d.ts.map +1 -0
- package/dist/timeline-preview/model3d.js +165 -0
- package/dist/timeline-preview/model3d.js.map +1 -0
- package/dist/timeline-preview/rasterize.d.ts +33 -0
- package/dist/timeline-preview/rasterize.d.ts.map +1 -0
- package/dist/timeline-preview/rasterize.js +83 -0
- package/dist/timeline-preview/rasterize.js.map +1 -0
- package/dist/timeline-preview/sheet.d.ts +38 -0
- package/dist/timeline-preview/sheet.d.ts.map +1 -0
- package/dist/timeline-preview/sheet.js +72 -0
- package/dist/timeline-preview/sheet.js.map +1 -0
- package/dist/tools/asset-persist.d.ts +29 -0
- package/dist/tools/asset-persist.d.ts.map +1 -0
- package/dist/tools/asset-persist.js +78 -0
- package/dist/tools/asset-persist.js.map +1 -0
- package/dist/tools/base-tool.d.ts +52 -5
- package/dist/tools/base-tool.d.ts.map +1 -1
- package/dist/tools/base-tool.js +180 -3
- package/dist/tools/base-tool.js.map +1 -1
- package/dist/tools/binary-output.d.ts +54 -0
- package/dist/tools/binary-output.d.ts.map +1 -0
- package/dist/tools/binary-output.js +127 -0
- package/dist/tools/binary-output.js.map +1 -0
- package/dist/tools/browser-tools.d.ts +5 -29
- package/dist/tools/browser-tools.d.ts.map +1 -1
- package/dist/tools/browser-tools.js +5 -205
- package/dist/tools/browser-tools.js.map +1 -1
- package/dist/tools/builtin-tools.d.ts +52 -0
- package/dist/tools/builtin-tools.d.ts.map +1 -0
- package/dist/tools/builtin-tools.js +276 -0
- package/dist/tools/builtin-tools.js.map +1 -0
- package/dist/tools/control-tool.d.ts +2 -5
- package/dist/tools/control-tool.d.ts.map +1 -1
- package/dist/tools/control-tool.js +18 -15
- package/dist/tools/control-tool.js.map +1 -1
- package/dist/tools/dataseo-tools.d.ts +34 -32
- package/dist/tools/dataseo-tools.d.ts.map +1 -1
- package/dist/tools/dataseo-tools.js +148 -233
- package/dist/tools/dataseo-tools.js.map +1 -1
- package/dist/tools/external-capability-tools.d.ts +36 -0
- package/dist/tools/external-capability-tools.d.ts.map +1 -0
- package/dist/tools/external-capability-tools.js +41 -0
- package/dist/tools/external-capability-tools.js.map +1 -0
- package/dist/tools/external-mcp-tools.d.ts +139 -0
- package/dist/tools/external-mcp-tools.d.ts.map +1 -0
- package/dist/tools/external-mcp-tools.js +552 -0
- package/dist/tools/external-mcp-tools.js.map +1 -0
- package/dist/tools/finish-step-tool.d.ts +1 -1
- package/dist/tools/finish-step-tool.d.ts.map +1 -1
- package/dist/tools/finish-step-tool.js +3 -3
- package/dist/tools/finish-step-tool.js.map +1 -1
- package/dist/tools/google-tools.d.ts +28 -39
- package/dist/tools/google-tools.d.ts.map +1 -1
- package/dist/tools/google-tools.js +51 -164
- package/dist/tools/google-tools.js.map +1 -1
- package/dist/tools/google-workspace-tools.d.ts +23 -0
- package/dist/tools/google-workspace-tools.d.ts.map +1 -0
- package/dist/tools/google-workspace-tools.js +26 -0
- package/dist/tools/google-workspace-tools.js.map +1 -0
- package/dist/tools/graph-validation-registry.d.ts +33 -0
- package/dist/tools/graph-validation-registry.d.ts.map +1 -0
- package/dist/tools/graph-validation-registry.js +54 -0
- package/dist/tools/graph-validation-registry.js.map +1 -0
- package/dist/tools/http-tools.d.ts +5 -53
- package/dist/tools/http-tools.d.ts.map +1 -1
- package/dist/tools/http-tools.js +5 -177
- package/dist/tools/http-tools.js.map +1 -1
- package/dist/tools/image-injection.d.ts +51 -0
- package/dist/tools/image-injection.d.ts.map +1 -0
- package/dist/tools/image-injection.js +85 -0
- package/dist/tools/image-injection.js.map +1 -0
- package/dist/tools/mcp-tool-support.d.ts +203 -0
- package/dist/tools/mcp-tool-support.d.ts.map +1 -0
- package/dist/tools/mcp-tool-support.js +536 -0
- package/dist/tools/mcp-tool-support.js.map +1 -0
- package/dist/tools/mcp-tools.d.ts +104 -390
- package/dist/tools/mcp-tools.d.ts.map +1 -1
- package/dist/tools/mcp-tools.js +128 -600
- package/dist/tools/mcp-tools.js.map +1 -1
- package/dist/tools/media-tools.d.ts +9 -272
- package/dist/tools/media-tools.d.ts.map +1 -1
- package/dist/tools/media-tools.js +11 -676
- package/dist/tools/media-tools.js.map +1 -1
- package/dist/tools/memory-tools.d.ts +36 -0
- package/dist/tools/memory-tools.d.ts.map +1 -0
- package/dist/tools/memory-tools.js +52 -0
- package/dist/tools/memory-tools.js.map +1 -0
- package/dist/tools/openai-tools.d.ts +28 -68
- package/dist/tools/openai-tools.d.ts.map +1 -1
- package/dist/tools/openai-tools.js +51 -181
- package/dist/tools/openai-tools.js.map +1 -1
- package/dist/tools/plan-builder-tools.d.ts +141 -7
- package/dist/tools/plan-builder-tools.d.ts.map +1 -1
- package/dist/tools/plan-builder-tools.js +159 -86
- package/dist/tools/plan-builder-tools.js.map +1 -1
- package/dist/tools/run-search-tool.d.ts +82 -0
- package/dist/tools/run-search-tool.d.ts.map +1 -0
- package/dist/tools/run-search-tool.js +121 -0
- package/dist/tools/run-search-tool.js.map +1 -0
- package/dist/tools/run-subtask-tool.d.ts +49 -0
- package/dist/tools/run-subtask-tool.d.ts.map +1 -0
- package/dist/tools/run-subtask-tool.js +88 -0
- package/dist/tools/run-subtask-tool.js.map +1 -0
- package/dist/tools/search-tools.d.ts +5 -36
- package/dist/tools/search-tools.d.ts.map +1 -1
- package/dist/tools/search-tools.js +5 -215
- package/dist/tools/search-tools.js.map +1 -1
- package/dist/tools/serp-providers/apify-provider.d.ts +18 -0
- package/dist/tools/serp-providers/apify-provider.d.ts.map +1 -0
- package/dist/tools/serp-providers/apify-provider.js +93 -0
- package/dist/tools/serp-providers/apify-provider.js.map +1 -0
- package/dist/tools/serp-providers/brave-provider.d.ts +34 -0
- package/dist/tools/serp-providers/brave-provider.d.ts.map +1 -0
- package/dist/tools/serp-providers/brave-provider.js +86 -0
- package/dist/tools/serp-providers/brave-provider.js.map +1 -0
- package/dist/tools/serp-providers/dataforseo-provider.d.ts +14 -1
- package/dist/tools/serp-providers/dataforseo-provider.d.ts.map +1 -1
- package/dist/tools/serp-providers/dataforseo-provider.js +0 -9
- package/dist/tools/serp-providers/dataforseo-provider.js.map +1 -1
- package/dist/tools/serp-providers/gemini-provider.d.ts +45 -0
- package/dist/tools/serp-providers/gemini-provider.d.ts.map +1 -0
- package/dist/tools/serp-providers/gemini-provider.js +148 -0
- package/dist/tools/serp-providers/gemini-provider.js.map +1 -0
- package/dist/tools/serp-providers/index.d.ts +61 -0
- package/dist/tools/serp-providers/index.d.ts.map +1 -1
- package/dist/tools/serp-providers/index.js +119 -3
- package/dist/tools/serp-providers/index.js.map +1 -1
- package/dist/tools/serp-providers/openai-provider.d.ts +58 -0
- package/dist/tools/serp-providers/openai-provider.d.ts.map +1 -0
- package/dist/tools/serp-providers/openai-provider.js +113 -0
- package/dist/tools/serp-providers/openai-provider.js.map +1 -0
- package/dist/tools/serp-providers/serpapi-provider.d.ts +9 -5
- package/dist/tools/serp-providers/serpapi-provider.d.ts.map +1 -1
- package/dist/tools/serp-providers/serpapi-provider.js +26 -44
- package/dist/tools/serp-providers/serpapi-provider.js.map +1 -1
- package/dist/tools/serp-tool-factory.d.ts +27 -0
- package/dist/tools/serp-tool-factory.d.ts.map +1 -0
- package/dist/tools/serp-tool-factory.js +56 -0
- package/dist/tools/serp-tool-factory.js.map +1 -0
- package/dist/tools/shared-tools.d.ts +33 -0
- package/dist/tools/shared-tools.d.ts.map +1 -0
- package/dist/tools/shared-tools.js +39 -0
- package/dist/tools/shared-tools.js.map +1 -0
- package/dist/tools/start-subtask-tool.d.ts +51 -0
- package/dist/tools/start-subtask-tool.d.ts.map +1 -0
- package/dist/tools/start-subtask-tool.js +189 -0
- package/dist/tools/start-subtask-tool.js.map +1 -0
- package/dist/tools/submit-code-tool.d.ts +159 -0
- package/dist/tools/submit-code-tool.d.ts.map +1 -0
- package/dist/tools/submit-code-tool.js +209 -0
- package/dist/tools/submit-code-tool.js.map +1 -0
- package/dist/tools/subtask-fields.d.ts +11 -0
- package/dist/tools/subtask-fields.d.ts.map +1 -0
- package/dist/tools/subtask-fields.js +11 -0
- package/dist/tools/subtask-fields.js.map +1 -0
- package/dist/tools/todo-tools.d.ts +15 -0
- package/dist/tools/todo-tools.d.ts.map +1 -0
- package/dist/tools/todo-tools.js +15 -0
- package/dist/tools/todo-tools.js.map +1 -0
- package/dist/tools/tool-permissions.d.ts +40 -0
- package/dist/tools/tool-permissions.d.ts.map +1 -0
- package/dist/tools/tool-permissions.js +52 -0
- package/dist/tools/tool-permissions.js.map +1 -0
- package/dist/tools/tool-search.d.ts +39 -0
- package/dist/tools/tool-search.d.ts.map +1 -0
- package/dist/tools/tool-search.js +108 -0
- package/dist/tools/tool-search.js.map +1 -0
- package/dist/tools/vector-tool-support.d.ts +15 -0
- package/dist/tools/vector-tool-support.d.ts.map +1 -0
- package/dist/tools/vector-tool-support.js +133 -0
- package/dist/tools/vector-tool-support.js.map +1 -0
- package/dist/tools/vector-tools.d.ts +13 -220
- package/dist/tools/vector-tools.d.ts.map +1 -1
- package/dist/tools/vector-tools.js +7 -578
- package/dist/tools/vector-tools.js.map +1 -1
- package/dist/tools/wait-subtasks-tool.d.ts +27 -0
- package/dist/tools/wait-subtasks-tool.d.ts.map +1 -0
- package/dist/tools/wait-subtasks-tool.js +66 -0
- package/dist/tools/wait-subtasks-tool.js.map +1 -0
- package/dist/tsconfig.tests.tmp.tsbuildinfo +1 -0
- package/dist/types.d.ts +16 -31
- package/dist/types.d.ts.map +1 -1
- package/dist/types.js +8 -1
- package/dist/types.js.map +1 -1
- package/dist/utils/dag-scheduler.d.ts +85 -0
- package/dist/utils/dag-scheduler.d.ts.map +1 -0
- package/dist/utils/dag-scheduler.js +180 -0
- package/dist/utils/dag-scheduler.js.map +1 -0
- package/dist/utils/json-parser.d.ts +32 -1
- package/dist/utils/json-parser.d.ts.map +1 -1
- package/dist/utils/json-parser.js +145 -10
- package/dist/utils/json-parser.js.map +1 -1
- package/dist/utils/json-schema-validate.d.ts +29 -0
- package/dist/utils/json-schema-validate.d.ts.map +1 -0
- package/dist/utils/json-schema-validate.js +228 -0
- package/dist/utils/json-schema-validate.js.map +1 -0
- package/dist/utils/link-abort.d.ts +13 -0
- package/dist/utils/link-abort.d.ts.map +1 -0
- package/dist/utils/link-abort.js +23 -0
- package/dist/utils/link-abort.js.map +1 -0
- package/dist/utils/merge-generators.d.ts +58 -0
- package/dist/utils/merge-generators.d.ts.map +1 -0
- package/dist/utils/merge-generators.js +148 -0
- package/dist/utils/merge-generators.js.map +1 -0
- package/dist/utils/remove-base64-images.d.ts +0 -5
- package/dist/utils/remove-base64-images.d.ts.map +1 -1
- package/dist/utils/remove-base64-images.js.map +1 -1
- package/dist/utils/step-failure.d.ts +20 -0
- package/dist/utils/step-failure.d.ts.map +1 -0
- package/dist/utils/step-failure.js +43 -0
- package/dist/utils/step-failure.js.map +1 -0
- package/dist/utils/think-tags.d.ts +16 -0
- package/dist/utils/think-tags.d.ts.map +1 -0
- package/dist/utils/think-tags.js +23 -0
- package/dist/utils/think-tags.js.map +1 -0
- package/dist/utils/type-guards.d.ts +26 -0
- package/dist/utils/type-guards.d.ts.map +1 -0
- package/dist/utils/type-guards.js +44 -0
- package/dist/utils/type-guards.js.map +1 -0
- package/dist/utils/ui-event-buffer.d.ts +18 -0
- package/dist/utils/ui-event-buffer.d.ts.map +1 -0
- package/dist/utils/ui-event-buffer.js +18 -0
- package/dist/utils/ui-event-buffer.js.map +1 -0
- package/dist/utils/wrap-generators-parallel.d.ts +0 -5
- package/dist/utils/wrap-generators-parallel.d.ts.map +1 -1
- package/dist/utils/wrap-generators-parallel.js +3 -6
- package/dist/utils/wrap-generators-parallel.js.map +1 -1
- package/dist/wasm-sandbox/host.d.ts +106 -0
- package/dist/wasm-sandbox/host.d.ts.map +1 -0
- package/dist/wasm-sandbox/host.js +409 -0
- package/dist/wasm-sandbox/host.js.map +1 -0
- package/dist/wasm-sandbox/workers.d.ts +43 -0
- package/dist/wasm-sandbox/workers.d.ts.map +1 -0
- package/dist/wasm-sandbox/workers.js +150 -0
- package/dist/wasm-sandbox/workers.js.map +1 -0
- package/dist/workspace-paths.d.ts +40 -0
- package/dist/workspace-paths.d.ts.map +1 -0
- package/dist/workspace-paths.js +102 -0
- package/dist/workspace-paths.js.map +1 -0
- package/dist/yt-dlp-gate.d.ts +19 -0
- package/dist/yt-dlp-gate.d.ts.map +1 -0
- package/dist/yt-dlp-gate.js +22 -0
- package/dist/yt-dlp-gate.js.map +1 -0
- package/package.json +112 -14
- package/dist/agent-executor.d.ts +0 -62
- package/dist/agent-executor.d.ts.map +0 -1
- package/dist/agent-executor.js +0 -324
- package/dist/agent-executor.js.map +0 -1
- package/dist/agent-step-executor.d.ts +0 -28
- package/dist/agent-step-executor.d.ts.map +0 -1
- package/dist/agent-step-executor.js +0 -99
- package/dist/agent-step-executor.js.map +0 -1
- package/dist/agent-workflow-runner.d.ts +0 -31
- package/dist/agent-workflow-runner.d.ts.map +0 -1
- package/dist/agent-workflow-runner.js +0 -124
- package/dist/agent-workflow-runner.js.map +0 -1
- package/dist/agent.d.ts +0 -115
- package/dist/agent.d.ts.map +0 -1
- package/dist/agent.js +0 -587
- package/dist/agent.js.map +0 -1
- package/dist/base-agent.d.ts +0 -42
- package/dist/base-agent.d.ts.map +0 -1
- package/dist/base-agent.js +0 -29
- package/dist/base-agent.js.map +0 -1
- package/dist/graph-planner.d.ts +0 -56
- package/dist/graph-planner.d.ts.map +0 -1
- package/dist/graph-planner.js +0 -311
- package/dist/graph-planner.js.map +0 -1
- package/dist/multi-mode-agent.d.ts +0 -86
- package/dist/multi-mode-agent.d.ts.map +0 -1
- package/dist/multi-mode-agent.js +0 -476
- package/dist/multi-mode-agent.js.map +0 -1
- package/dist/prompts/graph-planner-prompt.d.ts +0 -48
- package/dist/prompts/graph-planner-prompt.d.ts.map +0 -1
- package/dist/prompts/graph-planner-prompt.js +0 -180
- package/dist/prompts/graph-planner-prompt.js.map +0 -1
- package/dist/reject-agentic-provider.d.ts +0 -21
- package/dist/reject-agentic-provider.d.ts.map +0 -1
- package/dist/reject-agentic-provider.js +0 -32
- package/dist/reject-agentic-provider.js.map +0 -1
- package/dist/simple-agent.d.ts +0 -29
- package/dist/simple-agent.d.ts.map +0 -1
- package/dist/simple-agent.js +0 -54
- package/dist/simple-agent.js.map +0 -1
- package/dist/sub-agent-planner.d.ts +0 -33
- package/dist/sub-agent-planner.d.ts.map +0 -1
- package/dist/sub-agent-planner.js +0 -203
- package/dist/sub-agent-planner.js.map +0 -1
- package/dist/team/db-task-board.d.ts +0 -54
- package/dist/team/db-task-board.d.ts.map +0 -1
- package/dist/team/db-task-board.js +0 -361
- package/dist/team/db-task-board.js.map +0 -1
- package/dist/team/edge-message-bus.d.ts +0 -51
- package/dist/team/edge-message-bus.d.ts.map +0 -1
- package/dist/team/edge-message-bus.js +0 -169
- package/dist/team/edge-message-bus.js.map +0 -1
- package/dist/team/edge-task-board.d.ts +0 -77
- package/dist/team/edge-task-board.d.ts.map +0 -1
- package/dist/team/edge-task-board.js +0 -140
- package/dist/team/edge-task-board.js.map +0 -1
- package/dist/team/index.d.ts +0 -16
- package/dist/team/index.d.ts.map +0 -1
- package/dist/team/index.js +0 -17
- package/dist/team/index.js.map +0 -1
- package/dist/team/message-bus.d.ts +0 -58
- package/dist/team/message-bus.d.ts.map +0 -1
- package/dist/team/message-bus.js +0 -151
- package/dist/team/message-bus.js.map +0 -1
- package/dist/team/task-board.d.ts +0 -106
- package/dist/team/task-board.d.ts.map +0 -1
- package/dist/team/task-board.js +0 -370
- package/dist/team/task-board.js.map +0 -1
- package/dist/team/team-executor.d.ts +0 -72
- package/dist/team/team-executor.d.ts.map +0 -1
- package/dist/team/team-executor.js +0 -466
- package/dist/team/team-executor.js.map +0 -1
- package/dist/team/team-tools.d.ts +0 -252
- package/dist/team/team-tools.d.ts.map +0 -1
- package/dist/team/team-tools.js +0 -412
- package/dist/team/team-tools.js.map +0 -1
- package/dist/team/types.d.ts +0 -198
- package/dist/team/types.d.ts.map +0 -1
- package/dist/team/types.js +0 -8
- package/dist/team/types.js.map +0 -1
- package/dist/tools/add-edge-tool.d.ts +0 -18
- package/dist/tools/add-edge-tool.d.ts.map +0 -1
- package/dist/tools/add-edge-tool.js +0 -92
- package/dist/tools/add-edge-tool.js.map +0 -1
- package/dist/tools/add-node-tool.d.ts +0 -22
- package/dist/tools/add-node-tool.d.ts.map +0 -1
- package/dist/tools/add-node-tool.js +0 -128
- package/dist/tools/add-node-tool.js.map +0 -1
- package/dist/tools/asset-tools.d.ts +0 -58
- package/dist/tools/asset-tools.d.ts.map +0 -1
- package/dist/tools/asset-tools.js +0 -191
- package/dist/tools/asset-tools.js.map +0 -1
- package/dist/tools/calculator-tool.d.ts +0 -24
- package/dist/tools/calculator-tool.d.ts.map +0 -1
- package/dist/tools/calculator-tool.js +0 -64
- package/dist/tools/calculator-tool.js.map +0 -1
- package/dist/tools/claude-code-tools.d.ts +0 -96
- package/dist/tools/claude-code-tools.d.ts.map +0 -1
- package/dist/tools/claude-code-tools.js +0 -463
- package/dist/tools/claude-code-tools.js.map +0 -1
- package/dist/tools/code-tools.d.ts +0 -26
- package/dist/tools/code-tools.d.ts.map +0 -1
- package/dist/tools/code-tools.js +0 -82
- package/dist/tools/code-tools.js.map +0 -1
- package/dist/tools/create-plan-tool.d.ts +0 -29
- package/dist/tools/create-plan-tool.d.ts.map +0 -1
- package/dist/tools/create-plan-tool.js +0 -355
- package/dist/tools/create-plan-tool.js.map +0 -1
- package/dist/tools/create-task-tool.d.ts +0 -28
- package/dist/tools/create-task-tool.d.ts.map +0 -1
- package/dist/tools/create-task-tool.js +0 -254
- package/dist/tools/create-task-tool.js.map +0 -1
- package/dist/tools/email-tools.d.ts +0 -76
- package/dist/tools/email-tools.d.ts.map +0 -1
- package/dist/tools/email-tools.js +0 -259
- package/dist/tools/email-tools.js.map +0 -1
- package/dist/tools/filesystem-tools.d.ts +0 -77
- package/dist/tools/filesystem-tools.d.ts.map +0 -1
- package/dist/tools/filesystem-tools.js +0 -334
- package/dist/tools/filesystem-tools.js.map +0 -1
- package/dist/tools/find-model-tool.d.ts +0 -25
- package/dist/tools/find-model-tool.d.ts.map +0 -1
- package/dist/tools/find-model-tool.js +0 -236
- package/dist/tools/find-model-tool.js.map +0 -1
- package/dist/tools/finish-graph-tool.d.ts +0 -23
- package/dist/tools/finish-graph-tool.d.ts.map +0 -1
- package/dist/tools/finish-graph-tool.js +0 -46
- package/dist/tools/finish-graph-tool.js.map +0 -1
- package/dist/tools/js-code-tool.d.ts +0 -30
- package/dist/tools/js-code-tool.d.ts.map +0 -1
- package/dist/tools/js-code-tool.js +0 -98
- package/dist/tools/js-code-tool.js.map +0 -1
- package/dist/tools/local-get-node-info-tool.d.ts +0 -19
- package/dist/tools/local-get-node-info-tool.d.ts.map +0 -1
- package/dist/tools/local-get-node-info-tool.js +0 -75
- package/dist/tools/local-get-node-info-tool.js.map +0 -1
- package/dist/tools/local-list-nodes-tool.d.ts +0 -19
- package/dist/tools/local-list-nodes-tool.d.ts.map +0 -1
- package/dist/tools/local-list-nodes-tool.js +0 -69
- package/dist/tools/local-list-nodes-tool.js.map +0 -1
- package/dist/tools/local-search-nodes-tool.d.ts +0 -21
- package/dist/tools/local-search-nodes-tool.d.ts.map +0 -1
- package/dist/tools/local-search-nodes-tool.js +0 -111
- package/dist/tools/local-search-nodes-tool.js.map +0 -1
- package/dist/tools/math-tools.d.ts +0 -156
- package/dist/tools/math-tools.d.ts.map +0 -1
- package/dist/tools/math-tools.js +0 -422
- package/dist/tools/math-tools.js.map +0 -1
- package/dist/tools/model-tools.d.ts +0 -25
- package/dist/tools/model-tools.d.ts.map +0 -1
- package/dist/tools/model-tools.js +0 -48
- package/dist/tools/model-tools.js.map +0 -1
- package/dist/tools/pdf-tools.d.ts +0 -152
- package/dist/tools/pdf-tools.d.ts.map +0 -1
- package/dist/tools/pdf-tools.js +0 -339
- package/dist/tools/pdf-tools.js.map +0 -1
- package/dist/tools/tool-registry.d.ts +0 -9
- package/dist/tools/tool-registry.d.ts.map +0 -1
- package/dist/tools/tool-registry.js +0 -17
- package/dist/tools/tool-registry.js.map +0 -1
- package/dist/tools/workspace-tools.d.ts +0 -58
- package/dist/tools/workspace-tools.d.ts.map +0 -1
- package/dist/tools/workspace-tools.js +0 -152
- package/dist/tools/workspace-tools.js.map +0 -1
|
@@ -0,0 +1,250 @@
|
|
|
1
|
+
/**
|
|
2
|
+
* CodeAct evaluation harness — provider-agnostic.
|
|
3
|
+
*
|
|
4
|
+
* Drives a real {@link CodeActExecutor} over the instrumented offline
|
|
5
|
+
* toolbelt in `codeact-cases.ts` and scores each run structurally: required
|
|
6
|
+
* tools invoked through the bridge, action-round and tool-call bounds, and a
|
|
7
|
+
* predicate over the final result.
|
|
8
|
+
*/
|
|
9
|
+
import { randomUUID } from "node:crypto";
|
|
10
|
+
import { ProcessingContext } from "@nodetool-ai/runtime";
|
|
11
|
+
import { CodeActExecutor, EXECUTE_CODE_TOOL_NAME } from "../codeact/codeact-executor.js";
|
|
12
|
+
import { CODEACT_EVAL_CASES, createCodeActRecorder, createCodeActTools } from "./codeact-cases.js";
|
|
13
|
+
import { shippedPackCatalog } from "./codeact-sandbox-pack-cases.js";
|
|
14
|
+
const DEFAULT_MAX_ITERATIONS = 8;
|
|
15
|
+
/** Pure scorer over one run's observation. */
|
|
16
|
+
export function checkCodeActExpectations(obs, expect) {
|
|
17
|
+
const checks = [];
|
|
18
|
+
for (const name of expect.requiredTools ?? []) {
|
|
19
|
+
const pass = obs.toolsInvoked.has(name);
|
|
20
|
+
checks.push({
|
|
21
|
+
name: `tool:${name}`,
|
|
22
|
+
pass,
|
|
23
|
+
detail: pass ? undefined : `never invoked ${name}`
|
|
24
|
+
});
|
|
25
|
+
}
|
|
26
|
+
for (const name of expect.forbiddenTools ?? []) {
|
|
27
|
+
const hit = obs.toolsInvoked.has(name);
|
|
28
|
+
checks.push({
|
|
29
|
+
name: `not-tool:${name}`,
|
|
30
|
+
pass: !hit,
|
|
31
|
+
detail: hit ? `invoked forbidden tool ${name}` : undefined
|
|
32
|
+
});
|
|
33
|
+
}
|
|
34
|
+
for (const name of expect.requiredSessionTools ?? []) {
|
|
35
|
+
const pass = obs.sessionToolsInvoked.has(name);
|
|
36
|
+
checks.push({
|
|
37
|
+
name: `session-tool:${name}`,
|
|
38
|
+
pass,
|
|
39
|
+
detail: pass ? undefined : `never invoked ${name}`
|
|
40
|
+
});
|
|
41
|
+
}
|
|
42
|
+
if (expect.maxActions !== undefined) {
|
|
43
|
+
checks.push({
|
|
44
|
+
name: `actions<=${expect.maxActions}`,
|
|
45
|
+
pass: obs.actions <= expect.maxActions,
|
|
46
|
+
detail: `took ${obs.actions}`
|
|
47
|
+
});
|
|
48
|
+
}
|
|
49
|
+
if (expect.minToolCalls !== undefined) {
|
|
50
|
+
checks.push({
|
|
51
|
+
name: `toolCalls>=${expect.minToolCalls}`,
|
|
52
|
+
pass: obs.toolCalls >= expect.minToolCalls,
|
|
53
|
+
detail: `made ${obs.toolCalls}`
|
|
54
|
+
});
|
|
55
|
+
}
|
|
56
|
+
if (expect.maxToolCalls !== undefined) {
|
|
57
|
+
checks.push({
|
|
58
|
+
name: `toolCalls<=${expect.maxToolCalls}`,
|
|
59
|
+
pass: obs.toolCalls <= expect.maxToolCalls,
|
|
60
|
+
detail: `made ${obs.toolCalls}`
|
|
61
|
+
});
|
|
62
|
+
}
|
|
63
|
+
if (expect.resultCheck) {
|
|
64
|
+
const pass = expect.resultCheck(obs.result);
|
|
65
|
+
checks.push({
|
|
66
|
+
name: `result:${expect.resultCheckLabel ?? "predicate"}`,
|
|
67
|
+
pass,
|
|
68
|
+
detail: pass
|
|
69
|
+
? undefined
|
|
70
|
+
: `got ${JSON.stringify(obs.result)?.slice(0, 200)}`
|
|
71
|
+
});
|
|
72
|
+
}
|
|
73
|
+
return checks;
|
|
74
|
+
}
|
|
75
|
+
async function runCase(evalCase, opts) {
|
|
76
|
+
const recorder = createCodeActRecorder();
|
|
77
|
+
const tools = (evalCase.createTools ?? createCodeActTools)(recorder);
|
|
78
|
+
const maxIterations = evalCase.maxIterations ?? opts.maxIterations ?? DEFAULT_MAX_ITERATIONS;
|
|
79
|
+
// A case that allows packages needs a catalog to resolve them; one that does
|
|
80
|
+
// not stays on the process default, which is what every other case runs with.
|
|
81
|
+
const sandboxPackages = evalCase.sandboxPackages ?? [];
|
|
82
|
+
const contextInit = {
|
|
83
|
+
jobId: `codeact-eval-${randomUUID()}`,
|
|
84
|
+
userId: "eval-user"
|
|
85
|
+
};
|
|
86
|
+
if (sandboxPackages.length > 0) {
|
|
87
|
+
contextInit.sandboxModuleCatalog = shippedPackCatalog();
|
|
88
|
+
}
|
|
89
|
+
const context = new ProcessingContext(contextInit);
|
|
90
|
+
const step = {
|
|
91
|
+
id: randomUUID(),
|
|
92
|
+
instructions: evalCase.objective,
|
|
93
|
+
completed: false,
|
|
94
|
+
dependsOn: [],
|
|
95
|
+
logs: [],
|
|
96
|
+
outputSchema: evalCase.outputSchema
|
|
97
|
+
? JSON.stringify(evalCase.outputSchema)
|
|
98
|
+
: undefined
|
|
99
|
+
};
|
|
100
|
+
const task = {
|
|
101
|
+
id: randomUUID(),
|
|
102
|
+
title: evalCase.description,
|
|
103
|
+
steps: [step]
|
|
104
|
+
};
|
|
105
|
+
const executor = new CodeActExecutor({
|
|
106
|
+
task,
|
|
107
|
+
step,
|
|
108
|
+
context,
|
|
109
|
+
provider: opts.provider,
|
|
110
|
+
model: opts.model,
|
|
111
|
+
tools,
|
|
112
|
+
maxIterations,
|
|
113
|
+
signal: opts.signal,
|
|
114
|
+
sandboxPackages
|
|
115
|
+
});
|
|
116
|
+
const costBefore = opts.provider.getTotalCost();
|
|
117
|
+
const startedAt = Date.now();
|
|
118
|
+
let result = null;
|
|
119
|
+
let actions = 0;
|
|
120
|
+
let error;
|
|
121
|
+
const sessionToolsInvoked = new Set();
|
|
122
|
+
try {
|
|
123
|
+
for await (const item of executor.execute()) {
|
|
124
|
+
if (opts.signal?.aborted)
|
|
125
|
+
break;
|
|
126
|
+
if (item.type === "tool_call_update") {
|
|
127
|
+
const tc = item;
|
|
128
|
+
if (tc.name)
|
|
129
|
+
sessionToolsInvoked.add(tc.name);
|
|
130
|
+
// A "round" is one execute_code turn.
|
|
131
|
+
if (tc.name === EXECUTE_CODE_TOOL_NAME)
|
|
132
|
+
actions++;
|
|
133
|
+
}
|
|
134
|
+
if (item.type === "step_result") {
|
|
135
|
+
const sr = item;
|
|
136
|
+
if (sr.error)
|
|
137
|
+
error = sr.error;
|
|
138
|
+
else
|
|
139
|
+
result = sr.result;
|
|
140
|
+
}
|
|
141
|
+
}
|
|
142
|
+
}
|
|
143
|
+
catch (e) {
|
|
144
|
+
error = e instanceof Error ? e.message : String(e);
|
|
145
|
+
}
|
|
146
|
+
const durationMs = Date.now() - startedAt;
|
|
147
|
+
const costUsd = opts.provider.getTotalCost() - costBefore;
|
|
148
|
+
const accepted = error === undefined && step.completed === true;
|
|
149
|
+
const observation = {
|
|
150
|
+
toolsInvoked: new Set(recorder.invocations.map((i) => i.name)),
|
|
151
|
+
sessionToolsInvoked,
|
|
152
|
+
actions,
|
|
153
|
+
toolCalls: recorder.invocations.length,
|
|
154
|
+
result
|
|
155
|
+
};
|
|
156
|
+
const checks = [
|
|
157
|
+
{ name: "accepted", pass: accepted, detail: error }
|
|
158
|
+
];
|
|
159
|
+
if (accepted) {
|
|
160
|
+
checks.push(...checkCodeActExpectations(observation, evalCase.expect));
|
|
161
|
+
}
|
|
162
|
+
const score = accepted
|
|
163
|
+
? checks.filter((c) => c.pass).length / checks.length
|
|
164
|
+
: 0;
|
|
165
|
+
return {
|
|
166
|
+
caseId: evalCase.id,
|
|
167
|
+
description: evalCase.description,
|
|
168
|
+
accepted: accepted && checks.every((c) => c.pass),
|
|
169
|
+
score,
|
|
170
|
+
checks,
|
|
171
|
+
actions,
|
|
172
|
+
toolCalls: observation.toolCalls,
|
|
173
|
+
durationMs,
|
|
174
|
+
costUsd,
|
|
175
|
+
error
|
|
176
|
+
};
|
|
177
|
+
}
|
|
178
|
+
export async function runCodeActEval(opts) {
|
|
179
|
+
const cases = opts.cases ?? CODEACT_EVAL_CASES;
|
|
180
|
+
const results = [];
|
|
181
|
+
for (const evalCase of cases) {
|
|
182
|
+
if (opts.signal?.aborted)
|
|
183
|
+
break;
|
|
184
|
+
opts.onEvent?.(`▸ ${evalCase.id}`);
|
|
185
|
+
const result = await runCase(evalCase, opts);
|
|
186
|
+
results.push(result);
|
|
187
|
+
opts.onEvent?.(` ${result.accepted ? "pass" : "FAIL"} score=${result.score.toFixed(2)} ` +
|
|
188
|
+
`actions=${result.actions} toolCalls=${result.toolCalls}`);
|
|
189
|
+
}
|
|
190
|
+
const accepted = results.filter((r) => r.accepted).length;
|
|
191
|
+
const meanScore = results.length > 0
|
|
192
|
+
? results.reduce((sum, r) => sum + r.score, 0) / results.length
|
|
193
|
+
: 0;
|
|
194
|
+
return {
|
|
195
|
+
provider: opts.provider.provider,
|
|
196
|
+
model: opts.model,
|
|
197
|
+
startedAt: new Date().toISOString(),
|
|
198
|
+
cases: results,
|
|
199
|
+
summary: {
|
|
200
|
+
total: results.length,
|
|
201
|
+
accepted,
|
|
202
|
+
successRate: results.length > 0 ? accepted / results.length : 0,
|
|
203
|
+
meanScore,
|
|
204
|
+
avgActions: results.length > 0
|
|
205
|
+
? results.reduce((s, r) => s + r.actions, 0) / results.length
|
|
206
|
+
: 0,
|
|
207
|
+
avgToolCalls: results.length > 0
|
|
208
|
+
? results.reduce((s, r) => s + r.toolCalls, 0) / results.length
|
|
209
|
+
: 0,
|
|
210
|
+
totalCostUsd: results.reduce((s, r) => s + r.costUsd, 0)
|
|
211
|
+
}
|
|
212
|
+
};
|
|
213
|
+
}
|
|
214
|
+
export function formatCodeActReport(report) {
|
|
215
|
+
const lines = [];
|
|
216
|
+
lines.push(`CodeAct eval — provider=${report.provider} model=${report.model}`);
|
|
217
|
+
lines.push("");
|
|
218
|
+
const header = [
|
|
219
|
+
"case".padEnd(26),
|
|
220
|
+
"result".padEnd(7),
|
|
221
|
+
"score".padEnd(6),
|
|
222
|
+
"acts".padEnd(5),
|
|
223
|
+
"tools".padEnd(6),
|
|
224
|
+
"time".padEnd(7),
|
|
225
|
+
"cost"
|
|
226
|
+
].join("");
|
|
227
|
+
lines.push(header);
|
|
228
|
+
lines.push("-".repeat(header.length));
|
|
229
|
+
for (const r of report.cases) {
|
|
230
|
+
lines.push([
|
|
231
|
+
r.caseId.padEnd(26),
|
|
232
|
+
(r.accepted ? "pass" : "FAIL").padEnd(7),
|
|
233
|
+
r.score.toFixed(2).padEnd(6),
|
|
234
|
+
String(r.actions).padEnd(5),
|
|
235
|
+
String(r.toolCalls).padEnd(6),
|
|
236
|
+
`${Math.round(r.durationMs / 1000)}s`.padEnd(7),
|
|
237
|
+
r.costUsd > 0 ? `$${r.costUsd.toFixed(4)}` : "-"
|
|
238
|
+
].join(""));
|
|
239
|
+
for (const c of r.checks.filter((c) => !c.pass)) {
|
|
240
|
+
lines.push(` ✗ ${c.name}${c.detail ? ` — ${c.detail}` : ""}`);
|
|
241
|
+
}
|
|
242
|
+
}
|
|
243
|
+
lines.push("");
|
|
244
|
+
const s = report.summary;
|
|
245
|
+
lines.push(`success ${s.accepted}/${s.total} (${(s.successRate * 100).toFixed(0)}%) ` +
|
|
246
|
+
`meanScore ${s.meanScore.toFixed(2)} avgActions ${s.avgActions.toFixed(1)} ` +
|
|
247
|
+
`avgToolCalls ${s.avgToolCalls.toFixed(1)} cost $${s.totalCostUsd.toFixed(4)}`);
|
|
248
|
+
return lines.join("\n");
|
|
249
|
+
}
|
|
250
|
+
//# sourceMappingURL=codeact-eval.js.map
|
|
@@ -0,0 +1 @@
|
|
|
1
|
+
{"version":3,"file":"codeact-eval.js","sourceRoot":"","sources":["../../src/evals/codeact-eval.ts"],"names":[],"mappings":"AAAA;;;;;;;GAOG;AAEH,OAAO,EAAE,UAAU,EAAE,MAAM,aAAa,CAAC;AAEzC,OAAO,EAAE,iBAAiB,EAAE,MAAM,sBAAsB,CAAC;AAEzD,OAAO,EACL,eAAe,EACf,sBAAsB,EACvB,MAAM,gCAAgC,CAAC;AAExC,OAAO,EACL,kBAAkB,EAClB,qBAAqB,EACrB,kBAAkB,EAGnB,MAAM,oBAAoB,CAAC;AAC5B,OAAO,EAAE,kBAAkB,EAAE,MAAM,iCAAiC,CAAC;AAErE,MAAM,sBAAsB,GAAG,CAAC,CAAC;AAwDjC,8CAA8C;AAC9C,MAAM,UAAU,wBAAwB,CACtC,GAAuB,EACvB,MAA+B;IAE/B,MAAM,MAAM,GAAgB,EAAE,CAAC;IAC/B,KAAK,MAAM,IAAI,IAAI,MAAM,CAAC,aAAa,IAAI,EAAE,EAAE,CAAC;QAC9C,MAAM,IAAI,GAAG,GAAG,CAAC,YAAY,CAAC,GAAG,CAAC,IAAI,CAAC,CAAC;QACxC,MAAM,CAAC,IAAI,CAAC;YACV,IAAI,EAAE,QAAQ,IAAI,EAAE;YACpB,IAAI;YACJ,MAAM,EAAE,IAAI,CAAC,CAAC,CAAC,SAAS,CAAC,CAAC,CAAC,iBAAiB,IAAI,EAAE;SACnD,CAAC,CAAC;IACL,CAAC;IACD,KAAK,MAAM,IAAI,IAAI,MAAM,CAAC,cAAc,IAAI,EAAE,EAAE,CAAC;QAC/C,MAAM,GAAG,GAAG,GAAG,CAAC,YAAY,CAAC,GAAG,CAAC,IAAI,CAAC,CAAC;QACvC,MAAM,CAAC,IAAI,CAAC;YACV,IAAI,EAAE,YAAY,IAAI,EAAE;YACxB,IAAI,EAAE,CAAC,GAAG;YACV,MAAM,EAAE,GAAG,CAAC,CAAC,CAAC,0BAA0B,IAAI,EAAE,CAAC,CAAC,CAAC,SAAS;SAC3D,CAAC,CAAC;IACL,CAAC;IACD,KAAK,MAAM,IAAI,IAAI,MAAM,CAAC,oBAAoB,IAAI,EAAE,EAAE,CAAC;QACrD,MAAM,IAAI,GAAG,GAAG,CAAC,mBAAmB,CAAC,GAAG,CAAC,IAAI,CAAC,CAAC;QAC/C,MAAM,CAAC,IAAI,CAAC;YACV,IAAI,EAAE,gBAAgB,IAAI,EAAE;YAC5B,IAAI;YACJ,MAAM,EAAE,IAAI,CAAC,CAAC,CAAC,SAAS,CAAC,CAAC,CAAC,iBAAiB,IAAI,EAAE;SACnD,CAAC,CAAC;IACL,CAAC;IACD,IAAI,MAAM,CAAC,UAAU,KAAK,SAAS,EAAE,CAAC;QACpC,MAAM,CAAC,IAAI,CAAC;YACV,IAAI,EAAE,YAAY,MAAM,CAAC,UAAU,EAAE;YACrC,IAAI,EAAE,GAAG,CAAC,OAAO,IAAI,MAAM,CAAC,UAAU;YACtC,MAAM,EAAE,QAAQ,GAAG,CAAC,OAAO,EAAE;SAC9B,CAAC,CAAC;IACL,CAAC;IACD,IAAI,MAAM,CAAC,YAAY,KAAK,SAAS,EAAE,CAAC;QACtC,MAAM,CAAC,IAAI,CAAC;YACV,IAAI,EAAE,cAAc,MAAM,CAAC,YAAY,EAAE;YACzC,IAAI,EAAE,GAAG,CAAC,SAAS,IAAI,MAAM,CAAC,YAAY;YAC1C,MAAM,EAAE,QAAQ,GAAG,CAAC,SAAS,EAAE;SAChC,CAAC,CAAC;IACL,CAAC;IACD,IAAI,MAAM,CAAC,YAAY,KAAK,SAAS,EAAE,CAAC;QACtC,MAAM,CAAC,IAAI,CAAC;YACV,IAAI,EAAE,cAAc,MAAM,CAAC,YAAY,EAAE;YACzC,IAAI,EAAE,GAAG,CAAC,SAAS,IAAI,MAAM,CAAC,YAAY;YAC1C,MAAM,EAAE,QAAQ,GAAG,CAAC,SAAS,EAAE;SAChC,CAAC,CAAC;IACL,CAAC;IACD,IAAI,MAAM,CAAC,WAAW,EAAE,CAAC;QACvB,MAAM,IAAI,GAAG,MAAM,CAAC,WAAW,CAAC,GAAG,CAAC,MAAM,CAAC,CAAC;QAC5C,MAAM,CAAC,IAAI,CAAC;YACV,IAAI,EAAE,UAAU,MAAM,CAAC,gBAAgB,IAAI,WAAW,EAAE;YACxD,IAAI;YACJ,MAAM,EAAE,IAAI;gBACV,CAAC,CAAC,SAAS;gBACX,CAAC,CAAC,OAAO,IAAI,CAAC,SAAS,CAAC,GAAG,CAAC,MAAM,CAAC,EAAE,KAAK,CAAC,CAAC,EAAE,GAAG,CAAC,EAAE;SACvD,CAAC,CAAC;IACL,CAAC;IACD,OAAO,MAAM,CAAC;AAChB,CAAC;AAED,KAAK,UAAU,OAAO,CACpB,QAAyB,EACzB,IAA2B;IAE3B,MAAM,QAAQ,GAAG,qBAAqB,EAAE,CAAC;IACzC,MAAM,KAAK,GAAG,CAAC,QAAQ,CAAC,WAAW,IAAI,kBAAkB,CAAC,CAAC,QAAQ,CAAC,CAAC;IACrE,MAAM,aAAa,GACjB,QAAQ,CAAC,aAAa,IAAI,IAAI,CAAC,aAAa,IAAI,sBAAsB,CAAC;IAEzE,6EAA6E;IAC7E,8EAA8E;IAC9E,MAAM,eAAe,GAAG,QAAQ,CAAC,eAAe,IAAI,EAAE,CAAC;IACvD,MAAM,WAAW,GAAuD;QACtE,KAAK,EAAE,gBAAgB,UAAU,EAAE,EAAE;QACrC,MAAM,EAAE,WAAW;KACpB,CAAC;IACF,IAAI,eAAe,CAAC,MAAM,GAAG,CAAC,EAAE,CAAC;QAC/B,WAAW,CAAC,oBAAoB,GAAG,kBAAkB,EAAE,CAAC;IAC1D,CAAC;IACD,MAAM,OAAO,GAAG,IAAI,iBAAiB,CAAC,WAAW,CAAC,CAAC;IAEnD,MAAM,IAAI,GAAS;QACjB,EAAE,EAAE,UAAU,EAAE;QAChB,YAAY,EAAE,QAAQ,CAAC,SAAS;QAChC,SAAS,EAAE,KAAK;QAChB,SAAS,EAAE,EAAE;QACb,IAAI,EAAE,EAAE;QACR,YAAY,EAAE,QAAQ,CAAC,YAAY;YACjC,CAAC,CAAC,IAAI,CAAC,SAAS,CAAC,QAAQ,CAAC,YAAY,CAAC;YACvC,CAAC,CAAC,SAAS;KACd,CAAC;IACF,MAAM,IAAI,GAAS;QACjB,EAAE,EAAE,UAAU,EAAE;QAChB,KAAK,EAAE,QAAQ,CAAC,WAAW;QAC3B,KAAK,EAAE,CAAC,IAAI,CAAC;KACd,CAAC;IAEF,MAAM,QAAQ,GAAG,IAAI,eAAe,CAAC;QACnC,IAAI;QACJ,IAAI;QACJ,OAAO;QACP,QAAQ,EAAE,IAAI,CAAC,QAAQ;QACvB,KAAK,EAAE,IAAI,CAAC,KAAK;QACjB,KAAK;QACL,aAAa;QACb,MAAM,EAAE,IAAI,CAAC,MAAM;QACnB,eAAe;KAChB,CAAC,CAAC;IAEH,MAAM,UAAU,GAAG,IAAI,CAAC,QAAQ,CAAC,YAAY,EAAE,CAAC;IAChD,MAAM,SAAS,GAAG,IAAI,CAAC,GAAG,EAAE,CAAC;IAC7B,IAAI,MAAM,GAAY,IAAI,CAAC;IAC3B,IAAI,OAAO,GAAG,CAAC,CAAC;IAChB,IAAI,KAAyB,CAAC;IAC9B,MAAM,mBAAmB,GAAG,IAAI,GAAG,EAAU,CAAC;IAE9C,IAAI,CAAC;QACH,IAAI,KAAK,EAAE,MAAM,IAAI,IAAI,QAAQ,CAAC,OAAO,EAAE,EAAE,CAAC;YAC5C,IAAI,IAAI,CAAC,MAAM,EAAE,OAAO;gBAAE,MAAM;YAChC,IAAI,IAAI,CAAC,IAAI,KAAK,kBAAkB,EAAE,CAAC;gBACrC,MAAM,EAAE,GAAG,IAAI,CAAC;gBAChB,IAAI,EAAE,CAAC,IAAI;oBAAE,mBAAmB,CAAC,GAAG,CAAC,EAAE,CAAC,IAAI,CAAC,CAAC;gBAC9C,sCAAsC;gBACtC,IAAI,EAAE,CAAC,IAAI,KAAK,sBAAsB;oBAAE,OAAO,EAAE,CAAC;YACpD,CAAC;YACD,IAAI,IAAI,CAAC,IAAI,KAAK,aAAa,EAAE,CAAC;gBAChC,MAAM,EAAE,GAAG,IAAI,CAAC;gBAChB,IAAI,EAAE,CAAC,KAAK;oBAAE,KAAK,GAAG,EAAE,CAAC,KAAK,CAAC;;oBAC1B,MAAM,GAAG,EAAE,CAAC,MAAM,CAAC;YAC1B,CAAC;QACH,CAAC;IACH,CAAC;IAAC,OAAO,CAAC,EAAE,CAAC;QACX,KAAK,GAAG,CAAC,YAAY,KAAK,CAAC,CAAC,CAAC,CAAC,CAAC,OAAO,CAAC,CAAC,CAAC,MAAM,CAAC,CAAC,CAAC,CAAC;IACrD,CAAC;IAED,MAAM,UAAU,GAAG,IAAI,CAAC,GAAG,EAAE,GAAG,SAAS,CAAC;IAC1C,MAAM,OAAO,GAAG,IAAI,CAAC,QAAQ,CAAC,YAAY,EAAE,GAAG,UAAU,CAAC;IAC1D,MAAM,QAAQ,GAAG,KAAK,KAAK,SAAS,IAAI,IAAI,CAAC,SAAS,KAAK,IAAI,CAAC;IAEhE,MAAM,WAAW,GAAuB;QACtC,YAAY,EAAE,IAAI,GAAG,CAAC,QAAQ,CAAC,WAAW,CAAC,GAAG,CAAC,CAAC,CAAC,EAAE,EAAE,CAAC,CAAC,CAAC,IAAI,CAAC,CAAC;QAC9D,mBAAmB;QACnB,OAAO;QACP,SAAS,EAAE,QAAQ,CAAC,WAAW,CAAC,MAAM;QACtC,MAAM;KACP,CAAC;IAEF,MAAM,MAAM,GAAgB;QAC1B,EAAE,IAAI,EAAE,UAAU,EAAE,IAAI,EAAE,QAAQ,EAAE,MAAM,EAAE,KAAK,EAAE;KACpD,CAAC;IACF,IAAI,QAAQ,EAAE,CAAC;QACb,MAAM,CAAC,IAAI,CAAC,GAAG,wBAAwB,CAAC,WAAW,EAAE,QAAQ,CAAC,MAAM,CAAC,CAAC,CAAC;IACzE,CAAC;IACD,MAAM,KAAK,GAAG,QAAQ;QACpB,CAAC,CAAC,MAAM,CAAC,MAAM,CAAC,CAAC,CAAC,EAAE,EAAE,CAAC,CAAC,CAAC,IAAI,CAAC,CAAC,MAAM,GAAG,MAAM,CAAC,MAAM;QACrD,CAAC,CAAC,CAAC,CAAC;IAEN,OAAO;QACL,MAAM,EAAE,QAAQ,CAAC,EAAE;QACnB,WAAW,EAAE,QAAQ,CAAC,WAAW;QACjC,QAAQ,EAAE,QAAQ,IAAI,MAAM,CAAC,KAAK,CAAC,CAAC,CAAC,EAAE,EAAE,CAAC,CAAC,CAAC,IAAI,CAAC;QACjD,KAAK;QACL,MAAM;QACN,OAAO;QACP,SAAS,EAAE,WAAW,CAAC,SAAS;QAChC,UAAU;QACV,OAAO;QACP,KAAK;KACN,CAAC;AACJ,CAAC;AAED,MAAM,CAAC,KAAK,UAAU,cAAc,CAClC,IAA2B;IAE3B,MAAM,KAAK,GAAG,IAAI,CAAC,KAAK,IAAI,kBAAkB,CAAC;IAC/C,MAAM,OAAO,GAAwB,EAAE,CAAC;IAExC,KAAK,MAAM,QAAQ,IAAI,KAAK,EAAE,CAAC;QAC7B,IAAI,IAAI,CAAC,MAAM,EAAE,OAAO;YAAE,MAAM;QAChC,IAAI,CAAC,OAAO,EAAE,CAAC,KAAK,QAAQ,CAAC,EAAE,EAAE,CAAC,CAAC;QACnC,MAAM,MAAM,GAAG,MAAM,OAAO,CAAC,QAAQ,EAAE,IAAI,CAAC,CAAC;QAC7C,OAAO,CAAC,IAAI,CAAC,MAAM,CAAC,CAAC;QACrB,IAAI,CAAC,OAAO,EAAE,CACZ,KAAK,MAAM,CAAC,QAAQ,CAAC,CAAC,CAAC,MAAM,CAAC,CAAC,CAAC,MAAM,UAAU,MAAM,CAAC,KAAK,CAAC,OAAO,CAAC,CAAC,CAAC,GAAG;YACxE,WAAW,MAAM,CAAC,OAAO,cAAc,MAAM,CAAC,SAAS,EAAE,CAC5D,CAAC;IACJ,CAAC;IAED,MAAM,QAAQ,GAAG,OAAO,CAAC,MAAM,CAAC,CAAC,CAAC,EAAE,EAAE,CAAC,CAAC,CAAC,QAAQ,CAAC,CAAC,MAAM,CAAC;IAC1D,MAAM,SAAS,GACb,OAAO,CAAC,MAAM,GAAG,CAAC;QAChB,CAAC,CAAC,OAAO,CAAC,MAAM,CAAC,CAAC,GAAG,EAAE,CAAC,EAAE,EAAE,CAAC,GAAG,GAAG,CAAC,CAAC,KAAK,EAAE,CAAC,CAAC,GAAG,OAAO,CAAC,MAAM;QAC/D,CAAC,CAAC,CAAC,CAAC;IAER,OAAO;QACL,QAAQ,EAAE,IAAI,CAAC,QAAQ,CAAC,QAAQ;QAChC,KAAK,EAAE,IAAI,CAAC,KAAK;QACjB,SAAS,EAAE,IAAI,IAAI,EAAE,CAAC,WAAW,EAAE;QACnC,KAAK,EAAE,OAAO;QACd,OAAO,EAAE;YACP,KAAK,EAAE,OAAO,CAAC,MAAM;YACrB,QAAQ;YACR,WAAW,EAAE,OAAO,CAAC,MAAM,GAAG,CAAC,CAAC,CAAC,CAAC,QAAQ,GAAG,OAAO,CAAC,MAAM,CAAC,CAAC,CAAC,CAAC;YAC/D,SAAS;YACT,UAAU,EACR,OAAO,CAAC,MAAM,GAAG,CAAC;gBAChB,CAAC,CAAC,OAAO,CAAC,MAAM,CAAC,CAAC,CAAC,EAAE,CAAC,EAAE,EAAE,CAAC,CAAC,GAAG,CAAC,CAAC,OAAO,EAAE,CAAC,CAAC,GAAG,OAAO,CAAC,MAAM;gBAC7D,CAAC,CAAC,CAAC;YACP,YAAY,EACV,OAAO,CAAC,MAAM,GAAG,CAAC;gBAChB,CAAC,CAAC,OAAO,CAAC,MAAM,CAAC,CAAC,CAAC,EAAE,CAAC,EAAE,EAAE,CAAC,CAAC,GAAG,CAAC,CAAC,SAAS,EAAE,CAAC,CAAC,GAAG,OAAO,CAAC,MAAM;gBAC/D,CAAC,CAAC,CAAC;YACP,YAAY,EAAE,OAAO,CAAC,MAAM,CAAC,CAAC,CAAC,EAAE,CAAC,EAAE,EAAE,CAAC,CAAC,GAAG,CAAC,CAAC,OAAO,EAAE,CAAC,CAAC;SACzD;KACF,CAAC;AACJ,CAAC;AAED,MAAM,UAAU,mBAAmB,CAAC,MAAyB;IAC3D,MAAM,KAAK,GAAa,EAAE,CAAC;IAC3B,KAAK,CAAC,IAAI,CACR,2BAA2B,MAAM,CAAC,QAAQ,UAAU,MAAM,CAAC,KAAK,EAAE,CACnE,CAAC;IACF,KAAK,CAAC,IAAI,CAAC,EAAE,CAAC,CAAC;IACf,MAAM,MAAM,GAAG;QACb,MAAM,CAAC,MAAM,CAAC,EAAE,CAAC;QACjB,QAAQ,CAAC,MAAM,CAAC,CAAC,CAAC;QAClB,OAAO,CAAC,MAAM,CAAC,CAAC,CAAC;QACjB,MAAM,CAAC,MAAM,CAAC,CAAC,CAAC;QAChB,OAAO,CAAC,MAAM,CAAC,CAAC,CAAC;QACjB,MAAM,CAAC,MAAM,CAAC,CAAC,CAAC;QAChB,MAAM;KACP,CAAC,IAAI,CAAC,EAAE,CAAC,CAAC;IACX,KAAK,CAAC,IAAI,CAAC,MAAM,CAAC,CAAC;IACnB,KAAK,CAAC,IAAI,CAAC,GAAG,CAAC,MAAM,CAAC,MAAM,CAAC,MAAM,CAAC,CAAC,CAAC;IACtC,KAAK,MAAM,CAAC,IAAI,MAAM,CAAC,KAAK,EAAE,CAAC;QAC7B,KAAK,CAAC,IAAI,CACR;YACE,CAAC,CAAC,MAAM,CAAC,MAAM,CAAC,EAAE,CAAC;YACnB,CAAC,CAAC,CAAC,QAAQ,CAAC,CAAC,CAAC,MAAM,CAAC,CAAC,CAAC,MAAM,CAAC,CAAC,MAAM,CAAC,CAAC,CAAC;YACxC,CAAC,CAAC,KAAK,CAAC,OAAO,CAAC,CAAC,CAAC,CAAC,MAAM,CAAC,CAAC,CAAC;YAC5B,MAAM,CAAC,CAAC,CAAC,OAAO,CAAC,CAAC,MAAM,CAAC,CAAC,CAAC;YAC3B,MAAM,CAAC,CAAC,CAAC,SAAS,CAAC,CAAC,MAAM,CAAC,CAAC,CAAC;YAC7B,GAAG,IAAI,CAAC,KAAK,CAAC,CAAC,CAAC,UAAU,GAAG,IAAI,CAAC,GAAG,CAAC,MAAM,CAAC,CAAC,CAAC;YAC/C,CAAC,CAAC,OAAO,GAAG,CAAC,CAAC,CAAC,CAAC,IAAI,CAAC,CAAC,OAAO,CAAC,OAAO,CAAC,CAAC,CAAC,EAAE,CAAC,CAAC,CAAC,GAAG;SACjD,CAAC,IAAI,CAAC,EAAE,CAAC,CACX,CAAC;QACF,KAAK,MAAM,CAAC,IAAI,CAAC,CAAC,MAAM,CAAC,MAAM,CAAC,CAAC,CAAC,EAAE,EAAE,CAAC,CAAC,CAAC,CAAC,IAAI,CAAC,EAAE,CAAC;YAChD,KAAK,CAAC,IAAI,CAAC,OAAO,CAAC,CAAC,IAAI,GAAG,CAAC,CAAC,MAAM,CAAC,CAAC,CAAC,MAAM,CAAC,CAAC,MAAM,EAAE,CAAC,CAAC,CAAC,EAAE,EAAE,CAAC,CAAC;QACjE,CAAC;IACH,CAAC;IACD,KAAK,CAAC,IAAI,CAAC,EAAE,CAAC,CAAC;IACf,MAAM,CAAC,GAAG,MAAM,CAAC,OAAO,CAAC;IACzB,KAAK,CAAC,IAAI,CACR,WAAW,CAAC,CAAC,QAAQ,IAAI,CAAC,CAAC,KAAK,KAAK,CAAC,CAAC,CAAC,WAAW,GAAG,GAAG,CAAC,CAAC,OAAO,CAAC,CAAC,CAAC,MAAM;QACzE,aAAa,CAAC,CAAC,SAAS,CAAC,OAAO,CAAC,CAAC,CAAC,gBAAgB,CAAC,CAAC,UAAU,CAAC,OAAO,CAAC,CAAC,CAAC,IAAI;QAC9E,gBAAgB,CAAC,CAAC,YAAY,CAAC,OAAO,CAAC,CAAC,CAAC,WAAW,CAAC,CAAC,YAAY,CAAC,OAAO,CAAC,CAAC,CAAC,EAAE,CAClF,CAAC;IACF,OAAO,KAAK,CAAC,IAAI,CAAC,IAAI,CAAC,CAAC;AAC1B,CAAC"}
|
|
@@ -0,0 +1,31 @@
|
|
|
1
|
+
/**
|
|
2
|
+
* CodeAct eval cases for sandbox packages.
|
|
3
|
+
*
|
|
4
|
+
* A CodeAct action is code the model just wrote, so the packages it may import
|
|
5
|
+
* are the session's allowlist and nothing else (`codeact/sandbox-packages.ts`).
|
|
6
|
+
* Three behaviours follow from that, and none of them was measured against a
|
|
7
|
+
* real model before: importing an allowed pack and using it correctly, reading
|
|
8
|
+
* a pack's SKILL.md through `get_sandbox_package_docs` rather than guessing its
|
|
9
|
+
* API, and recovering when an import is refused.
|
|
10
|
+
*
|
|
11
|
+
* The packs here are the shipped ones under `packages/sandbox-packs/`, resolved
|
|
12
|
+
* through the real catalog. Only **host** packs appear: a guest pack must be
|
|
13
|
+
* compiled by `@nodetool-ai/sandbox-compiler` first, which this package does not
|
|
14
|
+
* depend on.
|
|
15
|
+
*
|
|
16
|
+
* Every objective carries its own data, so a case answers from the pack or not
|
|
17
|
+
* at all — the expected values cannot be reached by reasoning over the prompt.
|
|
18
|
+
*/
|
|
19
|
+
import type { SandboxModuleCatalog } from "@nodetool-ai/runtime";
|
|
20
|
+
import type { CodeActEvalCase } from "./codeact-cases.js";
|
|
21
|
+
export declare const SANDBOX_PACK_DOCS_TOOL = "get_sandbox_package_docs";
|
|
22
|
+
export declare const SANDBOX_PACK_LIST_TOOL = "list_sandbox_packages";
|
|
23
|
+
/**
|
|
24
|
+
* A catalog over the shipped packs, built once per process.
|
|
25
|
+
*
|
|
26
|
+
* Discovery reads the pack directories users install; nothing is stubbed, so a
|
|
27
|
+
* case that imports a pack runs the same library a Code node would.
|
|
28
|
+
*/
|
|
29
|
+
export declare function shippedPackCatalog(): SandboxModuleCatalog;
|
|
30
|
+
export declare const CODEACT_SANDBOX_PACK_EVAL_CASES: readonly CodeActEvalCase[];
|
|
31
|
+
//# sourceMappingURL=codeact-sandbox-pack-cases.d.ts.map
|
|
@@ -0,0 +1 @@
|
|
|
1
|
+
{"version":3,"file":"codeact-sandbox-pack-cases.d.ts","sourceRoot":"","sources":["../../src/evals/codeact-sandbox-pack-cases.ts"],"names":[],"mappings":"AAAA;;;;;;;;;;;;;;;;;GAiBG;AASH,OAAO,KAAK,EAAE,oBAAoB,EAAE,MAAM,sBAAsB,CAAC;AAEjE,OAAO,KAAK,EAAE,eAAe,EAAE,MAAM,oBAAoB,CAAC;AAM1D,eAAO,MAAM,sBAAsB,6BAA6B,CAAC;AACjE,eAAO,MAAM,sBAAsB,0BAA0B,CAAC;AAiB9D;;;;;GAKG;AACH,wBAAgB,kBAAkB,IAAI,oBAAoB,CASzD;AAyBD,eAAO,MAAM,+BAA+B,EAAE,SAAS,eAAe,EAwJrE,CAAC"}
|
|
@@ -0,0 +1,219 @@
|
|
|
1
|
+
/**
|
|
2
|
+
* CodeAct eval cases for sandbox packages.
|
|
3
|
+
*
|
|
4
|
+
* A CodeAct action is code the model just wrote, so the packages it may import
|
|
5
|
+
* are the session's allowlist and nothing else (`codeact/sandbox-packages.ts`).
|
|
6
|
+
* Three behaviours follow from that, and none of them was measured against a
|
|
7
|
+
* real model before: importing an allowed pack and using it correctly, reading
|
|
8
|
+
* a pack's SKILL.md through `get_sandbox_package_docs` rather than guessing its
|
|
9
|
+
* API, and recovering when an import is refused.
|
|
10
|
+
*
|
|
11
|
+
* The packs here are the shipped ones under `packages/sandbox-packs/`, resolved
|
|
12
|
+
* through the real catalog. Only **host** packs appear: a guest pack must be
|
|
13
|
+
* compiled by `@nodetool-ai/sandbox-compiler` first, which this package does not
|
|
14
|
+
* depend on.
|
|
15
|
+
*
|
|
16
|
+
* Every objective carries its own data, so a case answers from the pack or not
|
|
17
|
+
* at all — the expected values cannot be reached by reasoning over the prompt.
|
|
18
|
+
*/
|
|
19
|
+
import { dirname, join } from "node:path";
|
|
20
|
+
import { fileURLToPath } from "node:url";
|
|
21
|
+
import { createSandboxModuleCatalog, discoverSandboxPack } from "@nodetool-ai/node-sdk";
|
|
22
|
+
import { isObjectLike } from "../utils/type-guards.js";
|
|
23
|
+
const here = dirname(fileURLToPath(import.meta.url));
|
|
24
|
+
const PACKS_ROOT = join(here, "..", "..", "..", "sandbox-packs");
|
|
25
|
+
export const SANDBOX_PACK_DOCS_TOOL = "get_sandbox_package_docs";
|
|
26
|
+
export const SANDBOX_PACK_LIST_TOOL = "list_sandbox_packages";
|
|
27
|
+
/**
|
|
28
|
+
* The packs these cases allow. Three host packs plus the workflow DSL, whose
|
|
29
|
+
* modules are plain guest JavaScript shipped in the pack — no npm dependency,
|
|
30
|
+
* so no compiler either.
|
|
31
|
+
*/
|
|
32
|
+
const SHIPPED_PACKS = [
|
|
33
|
+
"sandbox-csv",
|
|
34
|
+
"sandbox-xml",
|
|
35
|
+
"sandbox-zip",
|
|
36
|
+
"sandbox-dsl",
|
|
37
|
+
"sandbox-flow"
|
|
38
|
+
];
|
|
39
|
+
let cached;
|
|
40
|
+
/**
|
|
41
|
+
* A catalog over the shipped packs, built once per process.
|
|
42
|
+
*
|
|
43
|
+
* Discovery reads the pack directories users install; nothing is stubbed, so a
|
|
44
|
+
* case that imports a pack runs the same library a Code node would.
|
|
45
|
+
*/
|
|
46
|
+
export function shippedPackCatalog() {
|
|
47
|
+
if (cached)
|
|
48
|
+
return cached;
|
|
49
|
+
const discoveries = SHIPPED_PACKS.map((dir) => {
|
|
50
|
+
const discovery = discoverSandboxPack(join(PACKS_ROOT, dir));
|
|
51
|
+
if (discovery === undefined)
|
|
52
|
+
throw new Error(`${dir} is not a sandbox pack`);
|
|
53
|
+
return discovery;
|
|
54
|
+
});
|
|
55
|
+
cached = createSandboxModuleCatalog(discoveries);
|
|
56
|
+
return cached;
|
|
57
|
+
}
|
|
58
|
+
const CSV = "@nodetool-ai/sandbox-csv";
|
|
59
|
+
const XML = "@nodetool-ai/sandbox-xml";
|
|
60
|
+
const ZIP = "@nodetool-ai/sandbox-zip";
|
|
61
|
+
const DIFF = "@nodetool-ai/sandbox-diff";
|
|
62
|
+
const ABSENT = "@acme/nope";
|
|
63
|
+
/** Six orders; EMEA totals 405.50, which is the answer only parsing produces. */
|
|
64
|
+
const ORDERS_CSV = [
|
|
65
|
+
"order_id,region,amount_eur",
|
|
66
|
+
"NT-1001,EMEA,120.00",
|
|
67
|
+
"NT-1002,AMER,80.50",
|
|
68
|
+
"NT-1003,EMEA,240.25",
|
|
69
|
+
"NT-1004,APAC,60.00",
|
|
70
|
+
"NT-1005,AMER,199.50",
|
|
71
|
+
"NT-1006,EMEA,45.25"
|
|
72
|
+
].join("\\n");
|
|
73
|
+
const TOTAL_SCHEMA = {
|
|
74
|
+
type: "object",
|
|
75
|
+
properties: { total: { type: "number" } },
|
|
76
|
+
required: ["total"]
|
|
77
|
+
};
|
|
78
|
+
export const CODEACT_SANDBOX_PACK_EVAL_CASES = [
|
|
79
|
+
{
|
|
80
|
+
id: "sandbox-pack-import",
|
|
81
|
+
description: "Import an allowed pack and compute from what it parses",
|
|
82
|
+
objective: "This CSV is your input:\\n\\n" +
|
|
83
|
+
ORDERS_CSV +
|
|
84
|
+
"\\n\\nParse it with the `" +
|
|
85
|
+
CSV +
|
|
86
|
+
"` sandbox package and finish with " +
|
|
87
|
+
"{total: <the summed amount_eur of the EMEA rows only>}. Parse it — do " +
|
|
88
|
+
"not add the numbers up yourself in your reply.",
|
|
89
|
+
sandboxPackages: [CSV],
|
|
90
|
+
outputSchema: TOTAL_SCHEMA,
|
|
91
|
+
expect: {
|
|
92
|
+
// 120.00 + 240.25 + 45.25. Wrong parsing gives a different sum, and the
|
|
93
|
+
// whole-file total (745.50) is the tell that the filter was skipped.
|
|
94
|
+
resultCheck: (r) => isObjectLike(r) &&
|
|
95
|
+
Math.abs(Number(r.total) - 405.5) < 0.001,
|
|
96
|
+
resultCheckLabel: "total=405.50 (EMEA only)",
|
|
97
|
+
maxActions: 4
|
|
98
|
+
}
|
|
99
|
+
},
|
|
100
|
+
{
|
|
101
|
+
id: "sandbox-pack-docs",
|
|
102
|
+
description: "Read a pack's SKILL.md instead of guessing its API",
|
|
103
|
+
objective: "You may import the `" +
|
|
104
|
+
XML +
|
|
105
|
+
"` sandbox package. Before you write any code, read that package's " +
|
|
106
|
+
"documentation to find out how attributes are represented in the parsed " +
|
|
107
|
+
"object. Then parse this feed:\\n\\n" +
|
|
108
|
+
'<?xml version="1.0"?><catalog><item sku="A-1"><name>Widget</name>' +
|
|
109
|
+
'</item></catalog>' +
|
|
110
|
+
"\\n\\nFinish with {prefix: <the exact key prefix the parser puts on " +
|
|
111
|
+
"attributes>}. It is two characters.",
|
|
112
|
+
sandboxPackages: [XML],
|
|
113
|
+
outputSchema: {
|
|
114
|
+
type: "object",
|
|
115
|
+
properties: { prefix: { type: "string" } },
|
|
116
|
+
required: ["prefix"]
|
|
117
|
+
},
|
|
118
|
+
expect: {
|
|
119
|
+
// The SKILL.md says attributes ride along prefixed `@_`. A model that
|
|
120
|
+
// guesses usually answers "@" or "$".
|
|
121
|
+
requiredSessionTools: [SANDBOX_PACK_DOCS_TOOL],
|
|
122
|
+
resultCheck: (r) => isObjectLike(r) &&
|
|
123
|
+
String(r.prefix).trim() === "@_",
|
|
124
|
+
resultCheckLabel: 'prefix="@_"',
|
|
125
|
+
maxActions: 4
|
|
126
|
+
}
|
|
127
|
+
},
|
|
128
|
+
{
|
|
129
|
+
id: "sandbox-pack-discover",
|
|
130
|
+
description: "Read the catalog instead of guessing what is installed",
|
|
131
|
+
objective: "Three sandbox packages might exist on this machine: `" +
|
|
132
|
+
ZIP +
|
|
133
|
+
"`, `" +
|
|
134
|
+
DIFF +
|
|
135
|
+
"` and `" +
|
|
136
|
+
ABSENT +
|
|
137
|
+
"`. Find out which of them are actually installed here — the answer is " +
|
|
138
|
+
"not in this prompt, and being off this session's allowlist is not the " +
|
|
139
|
+
"same as being absent. Finish with {installed: <the specifiers that are " +
|
|
140
|
+
"installed, as an array>}.",
|
|
141
|
+
sandboxPackages: [CSV],
|
|
142
|
+
namespaces: ["packs"],
|
|
143
|
+
outputSchema: {
|
|
144
|
+
type: "object",
|
|
145
|
+
properties: { installed: { type: "array", items: { type: "string" } } },
|
|
146
|
+
required: ["installed"]
|
|
147
|
+
},
|
|
148
|
+
expect: {
|
|
149
|
+
// Only zip is installed, and this session does not allow it — a listing
|
|
150
|
+
// that hid what it cannot import would answer with an empty array.
|
|
151
|
+
requiredSessionTools: [SANDBOX_PACK_LIST_TOOL],
|
|
152
|
+
resultCheck: (r) => {
|
|
153
|
+
const installed = isObjectLike(r)
|
|
154
|
+
? r.installed
|
|
155
|
+
: undefined;
|
|
156
|
+
return (Array.isArray(installed) &&
|
|
157
|
+
installed.length === 1 &&
|
|
158
|
+
String(installed[0]) === ZIP);
|
|
159
|
+
},
|
|
160
|
+
resultCheckLabel: `installed=[${ZIP}]`,
|
|
161
|
+
maxActions: 4
|
|
162
|
+
}
|
|
163
|
+
},
|
|
164
|
+
{
|
|
165
|
+
id: "sandbox-pack-denied",
|
|
166
|
+
// Graded on the outcome, not the route. A model reaches `available: false`
|
|
167
|
+
// either by attempting the import and reading the refusal, or by reading
|
|
168
|
+
// the allowlist the prompt advertises; both are correct, and live runs do
|
|
169
|
+
// both. What the case rules out is the failure that matters — hand-parsing
|
|
170
|
+
// the text and reporting success. The refusal mechanism itself is pinned
|
|
171
|
+
// deterministically in `tests/codeact-sandbox-packs.test.ts`.
|
|
172
|
+
description: "Report a pack as unavailable instead of working around it",
|
|
173
|
+
objective: "Try to parse this YAML with the `@nodetool-ai/sandbox-yaml` sandbox " +
|
|
174
|
+
"package:\\n\\nservice: nodetool-api\\nreplicas: 3\\n\\n" +
|
|
175
|
+
"If that package cannot be imported in this session, do not work around " +
|
|
176
|
+
"it and do not hand-parse the text. Finish with " +
|
|
177
|
+
"{available: false, reason: <a short reason>} instead. If it can be " +
|
|
178
|
+
"imported, finish with {available: true, reason: 'imported'}.",
|
|
179
|
+
// The session allows csv, so the yaml import is refused by the mount check.
|
|
180
|
+
sandboxPackages: [CSV],
|
|
181
|
+
outputSchema: {
|
|
182
|
+
type: "object",
|
|
183
|
+
properties: {
|
|
184
|
+
available: { type: "boolean" },
|
|
185
|
+
reason: { type: "string" }
|
|
186
|
+
},
|
|
187
|
+
required: ["available", "reason"]
|
|
188
|
+
},
|
|
189
|
+
expect: {
|
|
190
|
+
resultCheck: (r) => isObjectLike(r) &&
|
|
191
|
+
r.available === false,
|
|
192
|
+
resultCheckLabel: "available=false",
|
|
193
|
+
maxActions: 5
|
|
194
|
+
}
|
|
195
|
+
},
|
|
196
|
+
{
|
|
197
|
+
id: "sandbox-pack-two",
|
|
198
|
+
description: "Use two allowed packs in one action",
|
|
199
|
+
objective: "This CSV is your input:\\n\\n" +
|
|
200
|
+
ORDERS_CSV +
|
|
201
|
+
"\\n\\nUsing the `" +
|
|
202
|
+
ZIP +
|
|
203
|
+
"` and `" +
|
|
204
|
+
CSV +
|
|
205
|
+
"` sandbox packages: put the CSV into a zip archive under the entry " +
|
|
206
|
+
"name `orders.csv`, read that archive back, parse the entry you " +
|
|
207
|
+
"extracted, and finish with {total: <the number of parsed data rows>}. " +
|
|
208
|
+
"Go through the archive — do not count the lines in the prompt.",
|
|
209
|
+
sandboxPackages: [ZIP, CSV],
|
|
210
|
+
outputSchema: TOTAL_SCHEMA,
|
|
211
|
+
expect: {
|
|
212
|
+
resultCheck: (r) => isObjectLike(r) &&
|
|
213
|
+
Number(r.total) === 6,
|
|
214
|
+
resultCheckLabel: "total=6 rows",
|
|
215
|
+
maxActions: 5
|
|
216
|
+
}
|
|
217
|
+
}
|
|
218
|
+
];
|
|
219
|
+
//# sourceMappingURL=codeact-sandbox-pack-cases.js.map
|
|
@@ -0,0 +1 @@
|
|
|
1
|
+
{"version":3,"file":"codeact-sandbox-pack-cases.js","sourceRoot":"","sources":["../../src/evals/codeact-sandbox-pack-cases.ts"],"names":[],"mappings":"AAAA;;;;;;;;;;;;;;;;;GAiBG;AAEH,OAAO,EAAE,OAAO,EAAE,IAAI,EAAE,MAAM,WAAW,CAAC;AAC1C,OAAO,EAAE,aAAa,EAAE,MAAM,UAAU,CAAC;AAEzC,OAAO,EACL,0BAA0B,EAC1B,mBAAmB,EACpB,MAAM,uBAAuB,CAAC;AAI/B,OAAO,EAAE,YAAY,EAAE,MAAM,yBAAyB,CAAC;AAEvD,MAAM,IAAI,GAAG,OAAO,CAAC,aAAa,CAAC,OAAO,IAAI,CAAC,GAAG,CAAC,CAAC,CAAC;AACrD,MAAM,UAAU,GAAG,IAAI,CAAC,IAAI,EAAE,IAAI,EAAE,IAAI,EAAE,IAAI,EAAE,eAAe,CAAC,CAAC;AAEjE,MAAM,CAAC,MAAM,sBAAsB,GAAG,0BAA0B,CAAC;AACjE,MAAM,CAAC,MAAM,sBAAsB,GAAG,uBAAuB,CAAC;AAE9D;;;;GAIG;AACH,MAAM,aAAa,GAAG;IACpB,aAAa;IACb,aAAa;IACb,aAAa;IACb,aAAa;IACb,cAAc;CACf,CAAC;AAEF,IAAI,MAAwC,CAAC;AAE7C;;;;;GAKG;AACH,MAAM,UAAU,kBAAkB;IAChC,IAAI,MAAM;QAAE,OAAO,MAAM,CAAC;IAC1B,MAAM,WAAW,GAAG,aAAa,CAAC,GAAG,CAAC,CAAC,GAAG,EAAE,EAAE;QAC5C,MAAM,SAAS,GAAG,mBAAmB,CAAC,IAAI,CAAC,UAAU,EAAE,GAAG,CAAC,CAAC,CAAC;QAC7D,IAAI,SAAS,KAAK,SAAS;YAAE,MAAM,IAAI,KAAK,CAAC,GAAG,GAAG,wBAAwB,CAAC,CAAC;QAC7E,OAAO,SAAS,CAAC;IACnB,CAAC,CAAC,CAAC;IACH,MAAM,GAAG,0BAA0B,CAAC,WAAW,CAAC,CAAC;IACjD,OAAO,MAAM,CAAC;AAChB,CAAC;AAED,MAAM,GAAG,GAAG,0BAA0B,CAAC;AACvC,MAAM,GAAG,GAAG,0BAA0B,CAAC;AACvC,MAAM,GAAG,GAAG,0BAA0B,CAAC;AACvC,MAAM,IAAI,GAAG,2BAA2B,CAAC;AACzC,MAAM,MAAM,GAAG,YAAY,CAAC;AAE5B,iFAAiF;AACjF,MAAM,UAAU,GAAG;IACjB,4BAA4B;IAC5B,qBAAqB;IACrB,oBAAoB;IACpB,qBAAqB;IACrB,oBAAoB;IACpB,qBAAqB;IACrB,oBAAoB;CACrB,CAAC,IAAI,CAAC,KAAK,CAAC,CAAC;AAEd,MAAM,YAAY,GAAG;IACnB,IAAI,EAAE,QAAQ;IACd,UAAU,EAAE,EAAE,KAAK,EAAE,EAAE,IAAI,EAAE,QAAQ,EAAE,EAAE;IACzC,QAAQ,EAAE,CAAC,OAAO,CAAC;CACpB,CAAC;AAEF,MAAM,CAAC,MAAM,+BAA+B,GAA+B;IACzE;QACE,EAAE,EAAE,qBAAqB;QACzB,WAAW,EAAE,wDAAwD;QACrE,SAAS,EACP,+BAA+B;YAC/B,UAAU;YACV,2BAA2B;YAC3B,GAAG;YACH,oCAAoC;YACpC,wEAAwE;YACxE,gDAAgD;QAClD,eAAe,EAAE,CAAC,GAAG,CAAC;QACtB,YAAY,EAAE,YAAY;QAC1B,MAAM,EAAE;YACN,wEAAwE;YACxE,qEAAqE;YACrE,WAAW,EAAE,CAAC,CAAC,EAAE,EAAE,CACjB,YAAY,CAAC,CAAC,CAAC;gBACf,IAAI,CAAC,GAAG,CAAC,MAAM,CAAE,CAAyB,CAAC,KAAK,CAAC,GAAG,KAAK,CAAC,GAAG,KAAK;YACpE,gBAAgB,EAAE,0BAA0B;YAC5C,UAAU,EAAE,CAAC;SACd;KACF;IACD;QACE,EAAE,EAAE,mBAAmB;QACvB,WAAW,EAAE,oDAAoD;QACjE,SAAS,EACP,sBAAsB;YACtB,GAAG;YACH,oEAAoE;YACpE,yEAAyE;YACzE,qCAAqC;YACrC,mEAAmE;YACnE,mBAAmB;YACnB,sEAAsE;YACtE,qCAAqC;QACvC,eAAe,EAAE,CAAC,GAAG,CAAC;QACtB,YAAY,EAAE;YACZ,IAAI,EAAE,QAAQ;YACd,UAAU,EAAE,EAAE,MAAM,EAAE,EAAE,IAAI,EAAE,QAAQ,EAAE,EAAE;YAC1C,QAAQ,EAAE,CAAC,QAAQ,CAAC;SACrB;QACD,MAAM,EAAE;YACN,sEAAsE;YACtE,sCAAsC;YACtC,oBAAoB,EAAE,CAAC,sBAAsB,CAAC;YAC9C,WAAW,EAAE,CAAC,CAAC,EAAE,EAAE,CACjB,YAAY,CAAC,CAAC,CAAC;gBACf,MAAM,CAAE,CAA0B,CAAC,MAAM,CAAC,CAAC,IAAI,EAAE,KAAK,IAAI;YAC5D,gBAAgB,EAAE,aAAa;YAC/B,UAAU,EAAE,CAAC;SACd;KACF;IACD;QACE,EAAE,EAAE,uBAAuB;QAC3B,WAAW,EAAE,wDAAwD;QACrE,SAAS,EACP,uDAAuD;YACvD,GAAG;YACH,MAAM;YACN,IAAI;YACJ,SAAS;YACT,MAAM;YACN,wEAAwE;YACxE,wEAAwE;YACxE,yEAAyE;YACzE,2BAA2B;QAC7B,eAAe,EAAE,CAAC,GAAG,CAAC;QACtB,UAAU,EAAE,CAAC,OAAO,CAAC;QACrB,YAAY,EAAE;YACZ,IAAI,EAAE,QAAQ;YACd,UAAU,EAAE,EAAE,SAAS,EAAE,EAAE,IAAI,EAAE,OAAO,EAAE,KAAK,EAAE,EAAE,IAAI,EAAE,QAAQ,EAAE,EAAE,EAAE;YACvE,QAAQ,EAAE,CAAC,WAAW,CAAC;SACxB;QACD,MAAM,EAAE;YACN,wEAAwE;YACxE,mEAAmE;YACnE,oBAAoB,EAAE,CAAC,sBAAsB,CAAC;YAC9C,WAAW,EAAE,CAAC,CAAC,EAAE,EAAE;gBACjB,MAAM,SAAS,GACb,YAAY,CAAC,CAAC,CAAC;oBACb,CAAC,CAAE,CAA6B,CAAC,SAAS;oBAC1C,CAAC,CAAC,SAAS,CAAC;gBAChB,OAAO,CACL,KAAK,CAAC,OAAO,CAAC,SAAS,CAAC;oBACxB,SAAS,CAAC,MAAM,KAAK,CAAC;oBACtB,MAAM,CAAC,SAAS,CAAC,CAAC,CAAC,CAAC,KAAK,GAAG,CAC7B,CAAC;YACJ,CAAC;YACD,gBAAgB,EAAE,cAAc,GAAG,GAAG;YACtC,UAAU,EAAE,CAAC;SACd;KACF;IACD;QACE,EAAE,EAAE,qBAAqB;QACzB,2EAA2E;QAC3E,yEAAyE;QACzE,0EAA0E;QAC1E,2EAA2E;QAC3E,yEAAyE;QACzE,8DAA8D;QAC9D,WAAW,EAAE,2DAA2D;QACxE,SAAS,EACP,sEAAsE;YACtE,yDAAyD;YACzD,yEAAyE;YACzE,iDAAiD;YACjD,qEAAqE;YACrE,8DAA8D;QAChE,4EAA4E;QAC5E,eAAe,EAAE,CAAC,GAAG,CAAC;QACtB,YAAY,EAAE;YACZ,IAAI,EAAE,QAAQ;YACd,UAAU,EAAE;gBACV,SAAS,EAAE,EAAE,IAAI,EAAE,SAAS,EAAE;gBAC9B,MAAM,EAAE,EAAE,IAAI,EAAE,QAAQ,EAAE;aAC3B;YACD,QAAQ,EAAE,CAAC,WAAW,EAAE,QAAQ,CAAC;SAClC;QACD,MAAM,EAAE;YACN,WAAW,EAAE,CAAC,CAAC,EAAE,EAAE,CACjB,YAAY,CAAC,CAAC,CAAC;gBACd,CAA6B,CAAC,SAAS,KAAK,KAAK;YACpD,gBAAgB,EAAE,iBAAiB;YACnC,UAAU,EAAE,CAAC;SACd;KACF;IACD;QACE,EAAE,EAAE,kBAAkB;QACtB,WAAW,EAAE,qCAAqC;QAClD,SAAS,EACP,+BAA+B;YAC/B,UAAU;YACV,mBAAmB;YACnB,GAAG;YACH,SAAS;YACT,GAAG;YACH,qEAAqE;YACrE,iEAAiE;YACjE,wEAAwE;YACxE,gEAAgE;QAClE,eAAe,EAAE,CAAC,GAAG,EAAE,GAAG,CAAC;QAC3B,YAAY,EAAE,YAAY;QAC1B,MAAM,EAAE;YACN,WAAW,EAAE,CAAC,CAAC,EAAE,EAAE,CACjB,YAAY,CAAC,CAAC,CAAC;gBACf,MAAM,CAAE,CAAyB,CAAC,KAAK,CAAC,KAAK,CAAC;YAChD,gBAAgB,EAAE,cAAc;YAChC,UAAU,EAAE,CAAC;SACd;KACF;CACF,CAAC"}
|
|
@@ -0,0 +1,22 @@
|
|
|
1
|
+
/**
|
|
2
|
+
* Workflow-tool eval cases with interactive escalation.
|
|
3
|
+
*
|
|
4
|
+
* The graph-editor suite in `tool-loop-cases.ts` scores a model on objectives
|
|
5
|
+
* that are fully specified: everything it needs is in the prompt, so guessing
|
|
6
|
+
* is never required. These cases remove that guarantee. Each one is missing
|
|
7
|
+
* something only the user can supply — a name, permission to delete, a decision
|
|
8
|
+
* between two node types, a capability the catalog doesn't have — and hands the
|
|
9
|
+
* model an `ask_user` tool backed by a scripted user.
|
|
10
|
+
*
|
|
11
|
+
* What's being measured is the pair: escalate when (and only when) the request
|
|
12
|
+
* is underdetermined, then *act on the answer*. So every case scores both the
|
|
13
|
+
* question (`escalation.mustAsk`) and the resulting graph — a model that asks
|
|
14
|
+
* the right thing and then ignores the reply fails on state, and one that
|
|
15
|
+
* silently guesses fails on the ask. The `no-escalation-needed` case guards the
|
|
16
|
+
* other direction: a fully specified objective where reaching for the user is
|
|
17
|
+
* itself the failure.
|
|
18
|
+
*/
|
|
19
|
+
import type { ToolLoopEvalCase } from "./tool-loop-eval.js";
|
|
20
|
+
import { type ToolLoopFinalState } from "./tool-loop-bridge.js";
|
|
21
|
+
export declare const WORKFLOW_ESCALATION_TOOL_LOOP_CASES: readonly ToolLoopEvalCase<ToolLoopFinalState>[];
|
|
22
|
+
//# sourceMappingURL=escalation-cases.d.ts.map
|
|
@@ -0,0 +1 @@
|
|
|
1
|
+
{"version":3,"file":"escalation-cases.d.ts","sourceRoot":"","sources":["../../src/evals/escalation-cases.ts"],"names":[],"mappings":"AAAA;;;;;;;;;;;;;;;;;GAiBG;AAEH,OAAO,KAAK,EACV,gBAAgB,EAEjB,MAAM,qBAAqB,CAAC;AAC7B,OAAO,EAEL,KAAK,kBAAkB,EAGxB,MAAM,uBAAuB,CAAC;AAuH/B,eAAO,MAAM,mCAAmC,EAAE,SAAS,gBAAgB,CAAC,kBAAkB,CAAC,EAqQ5F,CAAC"}
|