@evalops/maestro 0.10.17 → 0.10.20
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 +56 -1
- package/dist/agent/action-approval.d.ts +12 -2
- package/dist/agent/action-approval.d.ts.map +1 -1
- package/dist/agent/action-approval.js +19 -32
- package/dist/agent/action-approval.js.map +1 -1
- package/dist/agent/agent.d.ts +1 -0
- package/dist/agent/agent.d.ts.map +1 -1
- package/dist/agent/agent.js +26 -0
- package/dist/agent/agent.js.map +1 -1
- package/dist/agent/compaction.d.ts +2 -0
- package/dist/agent/compaction.d.ts.map +1 -1
- package/dist/agent/compaction.js +4 -2
- package/dist/agent/compaction.js.map +1 -1
- package/dist/agent/index.d.ts +1 -1
- package/dist/agent/index.d.ts.map +1 -1
- package/dist/agent/index.js +1 -1
- package/dist/agent/index.js.map +1 -1
- package/dist/agent/modes.d.ts +56 -4
- package/dist/agent/modes.d.ts.map +1 -1
- package/dist/agent/modes.js +163 -2
- package/dist/agent/modes.js.map +1 -1
- package/dist/agent/prompt-recovery.d.ts.map +1 -1
- package/dist/agent/prompt-recovery.js +1 -3
- package/dist/agent/prompt-recovery.js.map +1 -1
- package/dist/agent/providers/anthropic.d.ts.map +1 -1
- package/dist/agent/providers/anthropic.js +9 -1
- package/dist/agent/providers/anthropic.js.map +1 -1
- package/dist/agent/providers/codex-app-server.d.ts +12 -0
- package/dist/agent/providers/codex-app-server.d.ts.map +1 -0
- package/dist/agent/providers/codex-app-server.js +788 -0
- package/dist/agent/providers/codex-app-server.js.map +1 -0
- package/dist/agent/providers/google-gemini-cli.js +14 -3
- package/dist/agent/providers/google.js +22 -3
- package/dist/agent/providers/scripted.d.ts +11 -0
- package/dist/agent/providers/scripted.d.ts.map +1 -0
- package/dist/agent/providers/scripted.js +399 -0
- package/dist/agent/providers/scripted.js.map +1 -0
- package/dist/agent/providers/vertex.js +14 -3
- package/dist/agent/scenario-source.d.ts +19 -0
- package/dist/agent/scenario-source.d.ts.map +1 -0
- package/dist/agent/scenario-source.js +112 -0
- package/dist/agent/scenario-source.js.map +1 -0
- package/dist/agent/subagent-specs.d.ts +1 -1
- package/dist/agent/subagent-specs.js +1 -1
- package/dist/agent/subagent-specs.js.map +1 -1
- package/dist/agent/swarm/executor.d.ts +25 -0
- package/dist/agent/swarm/executor.d.ts.map +1 -1
- package/dist/agent/swarm/executor.js +795 -20
- package/dist/agent/swarm/executor.js.map +1 -1
- package/dist/agent/swarm/runtime-events.d.ts +11 -0
- package/dist/agent/swarm/runtime-events.d.ts.map +1 -0
- package/dist/agent/swarm/runtime-events.js +26 -0
- package/dist/agent/swarm/runtime-events.js.map +1 -0
- package/dist/agent/swarm/types.d.ts +78 -0
- package/dist/agent/swarm/types.d.ts.map +1 -1
- package/dist/agent/swarm/types.js.map +1 -1
- package/dist/agent/tracing.d.ts +1 -1
- package/dist/agent/tracing.d.ts.map +1 -1
- package/dist/agent/tracing.js +17 -9
- package/dist/agent/tracing.js.map +1 -1
- package/dist/agent/transport/create-provider-stream.d.ts.map +1 -1
- package/dist/agent/transport/create-provider-stream.js +14 -0
- package/dist/agent/transport/create-provider-stream.js.map +1 -1
- package/dist/agent/transport/stable-stringify.d.ts +2 -0
- package/dist/agent/transport/stable-stringify.d.ts.map +1 -0
- package/dist/agent/transport/stable-stringify.js +14 -0
- package/dist/agent/transport/stable-stringify.js.map +1 -0
- package/dist/agent/transport/tool-execution-bridge.d.ts.map +1 -1
- package/dist/agent/transport/tool-execution-bridge.js +2 -15
- package/dist/agent/transport/tool-execution-bridge.js.map +1 -1
- package/dist/agent/transport/tool-safety-pipeline.d.ts +5 -1
- package/dist/agent/transport/tool-safety-pipeline.d.ts.map +1 -1
- package/dist/agent/transport/tool-safety-pipeline.js +184 -36
- package/dist/agent/transport/tool-safety-pipeline.js.map +1 -1
- package/dist/agent/transport.d.ts +8 -0
- package/dist/agent/transport.d.ts.map +1 -1
- package/dist/agent/transport.js +1613 -68
- package/dist/agent/transport.js.map +1 -1
- package/dist/agent/types.d.ts +170 -2
- package/dist/agent/types.d.ts.map +1 -1
- package/dist/agent/types.js.map +1 -1
- package/dist/api/enterprise-routes.d.ts.map +1 -1
- package/dist/api/enterprise-routes.js +34 -3
- package/dist/api/enterprise-routes.js.map +1 -1
- package/dist/app-server/in-process-client.d.ts +32 -0
- package/dist/app-server/in-process-client.d.ts.map +1 -0
- package/dist/app-server/in-process-client.js +73 -0
- package/dist/app-server/in-process-client.js.map +1 -0
- package/dist/app-server/session-api.d.ts +36 -0
- package/dist/app-server/session-api.d.ts.map +1 -0
- package/dist/app-server/session-api.js +713 -0
- package/dist/app-server/session-api.js.map +1 -0
- package/dist/approvals/platform-action-approval.d.ts +3 -1
- package/dist/approvals/platform-action-approval.d.ts.map +1 -1
- package/dist/approvals/platform-action-approval.js +16 -9
- package/dist/approvals/platform-action-approval.js.map +1 -1
- package/dist/approvals/service-client.d.ts.map +1 -1
- package/dist/approvals/service-client.js +7 -7
- package/dist/approvals/service-client.js.map +1 -1
- package/dist/audit/logger.d.ts +4 -4
- package/dist/bootstrap/agent-creation-setup.d.ts +4 -0
- package/dist/bootstrap/agent-creation-setup.d.ts.map +1 -1
- package/dist/bootstrap/agent-creation-setup.js +3 -1
- package/dist/bootstrap/agent-creation-setup.js.map +1 -1
- package/dist/bootstrap/auth-setup.js +1 -1
- package/dist/bootstrap/auth-setup.js.map +1 -1
- package/dist/bootstrap/event-subscriptions-setup.d.ts +6 -0
- package/dist/bootstrap/event-subscriptions-setup.d.ts.map +1 -1
- package/dist/bootstrap/event-subscriptions-setup.js +26 -2
- package/dist/bootstrap/event-subscriptions-setup.js.map +1 -1
- package/dist/bootstrap/model-resolution-setup.d.ts.map +1 -1
- package/dist/bootstrap/model-resolution-setup.js +4 -1
- package/dist/bootstrap/model-resolution-setup.js.map +1 -1
- package/dist/bootstrap/tools-setup.d.ts +2 -1
- package/dist/bootstrap/tools-setup.d.ts.map +1 -1
- package/dist/bootstrap/tools-setup.js +5 -1
- package/dist/bootstrap/tools-setup.js.map +1 -1
- package/dist/cli/args.d.ts +9 -1
- package/dist/cli/args.d.ts.map +1 -1
- package/dist/cli/args.js +130 -5
- package/dist/cli/args.js.map +1 -1
- package/dist/cli/commands/a2a.d.ts +9 -0
- package/dist/cli/commands/a2a.d.ts.map +1 -0
- package/dist/cli/commands/a2a.js +1714 -0
- package/dist/cli/commands/a2a.js.map +1 -0
- package/dist/cli/commands/agents.d.ts +16 -2
- package/dist/cli/commands/agents.d.ts.map +1 -1
- package/dist/cli/commands/agents.js +223 -35
- package/dist/cli/commands/agents.js.map +1 -1
- package/dist/cli/commands/codex.d.ts +1 -1
- package/dist/cli/commands/codex.d.ts.map +1 -1
- package/dist/cli/commands/codex.js +145 -39
- package/dist/cli/commands/codex.js.map +1 -1
- package/dist/cli/commands/context.d.ts +10 -0
- package/dist/cli/commands/context.d.ts.map +1 -0
- package/dist/cli/commands/context.js +176 -0
- package/dist/cli/commands/context.js.map +1 -0
- package/dist/cli/commands/hosted-runner.d.ts +1 -0
- package/dist/cli/commands/hosted-runner.d.ts.map +1 -1
- package/dist/cli/commands/hosted-runner.js +9 -1
- package/dist/cli/commands/hosted-runner.js.map +1 -1
- package/dist/cli/commands/init.d.ts.map +1 -1
- package/dist/cli/commands/init.js +6 -0
- package/dist/cli/commands/init.js.map +1 -1
- package/dist/cli/commands/modes.d.ts +8 -0
- package/dist/cli/commands/modes.d.ts.map +1 -0
- package/dist/cli/commands/modes.js +151 -0
- package/dist/cli/commands/modes.js.map +1 -0
- package/dist/cli/commands/operating-plane.d.ts +14 -0
- package/dist/cli/commands/operating-plane.d.ts.map +1 -0
- package/dist/cli/commands/operating-plane.js +160 -0
- package/dist/cli/commands/operating-plane.js.map +1 -0
- package/dist/cli/commands/run.d.ts +87 -0
- package/dist/cli/commands/run.d.ts.map +1 -0
- package/dist/cli/commands/run.js +356 -0
- package/dist/cli/commands/run.js.map +1 -0
- package/dist/cli/commands/scenario.d.ts +7 -0
- package/dist/cli/commands/scenario.d.ts.map +1 -0
- package/dist/cli/commands/scenario.js +195 -0
- package/dist/cli/commands/scenario.js.map +1 -0
- package/dist/cli/commands/skill.d.ts +7 -0
- package/dist/cli/commands/skill.d.ts.map +1 -0
- package/dist/cli/commands/skill.js +286 -0
- package/dist/cli/commands/skill.js.map +1 -0
- package/dist/cli/headless-protocol.d.ts +55 -1
- package/dist/cli/headless-protocol.d.ts.map +1 -1
- package/dist/cli/headless-protocol.js +446 -7
- package/dist/cli/headless-protocol.js.map +1 -1
- package/dist/cli/headless-runtime-selection.d.ts +38 -0
- package/dist/cli/headless-runtime-selection.d.ts.map +1 -0
- package/dist/cli/headless-runtime-selection.js +48 -0
- package/dist/cli/headless-runtime-selection.js.map +1 -0
- package/dist/cli/headless.d.ts +5 -1
- package/dist/cli/headless.d.ts.map +1 -1
- package/dist/cli/headless.js +11 -1
- package/dist/cli/headless.js.map +1 -1
- package/dist/cli/help.d.ts +3 -1
- package/dist/cli/help.d.ts.map +1 -1
- package/dist/cli/help.js +136 -30
- package/dist/cli/help.js.map +1 -1
- package/dist/cli/system-prompt.d.ts +22 -4
- package/dist/cli/system-prompt.d.ts.map +1 -1
- package/dist/cli/system-prompt.js +224 -127
- package/dist/cli/system-prompt.js.map +1 -1
- package/dist/cli-tui/commands/a2a-handlers.d.ts +7 -0
- package/dist/cli-tui/commands/a2a-handlers.d.ts.map +1 -0
- package/dist/cli-tui/commands/a2a-handlers.js +160 -0
- package/dist/cli-tui/commands/a2a-handlers.js.map +1 -0
- package/dist/cli-tui/commands/command-catalog.d.ts.map +1 -1
- package/dist/cli-tui/commands/command-catalog.js +15 -0
- package/dist/cli-tui/commands/command-catalog.js.map +1 -1
- package/dist/cli-tui/commands/command-registry-adapter.js +10 -1
- package/dist/cli-tui/commands/command-registry-adapter.js.map +1 -1
- package/dist/cli-tui/commands/handlers/mode-handler.d.ts.map +1 -1
- package/dist/cli-tui/commands/handlers/mode-handler.js +22 -10
- package/dist/cli-tui/commands/handlers/mode-handler.js.map +1 -1
- package/dist/cli-tui/commands/mcp-handlers.d.ts.map +1 -1
- package/dist/cli-tui/commands/mcp-handlers.js +6 -5
- package/dist/cli-tui/commands/mcp-handlers.js.map +1 -1
- package/dist/cli-tui/commands/types.d.ts +1 -0
- package/dist/cli-tui/commands/types.d.ts.map +1 -1
- package/dist/cli-tui/commands/types.js.map +1 -1
- package/dist/cli-tui/commands/utility-handlers.d.ts +4 -0
- package/dist/cli-tui/commands/utility-handlers.d.ts.map +1 -1
- package/dist/cli-tui/commands/utility-handlers.js +45 -6
- package/dist/cli-tui/commands/utility-handlers.js.map +1 -1
- package/dist/cli-tui/run/run-shell-command.d.ts.map +1 -1
- package/dist/cli-tui/run/run-shell-command.js +14 -5
- package/dist/cli-tui/run/run-shell-command.js.map +1 -1
- package/dist/cli-tui/run/streaming-shell-command.d.ts.map +1 -1
- package/dist/cli-tui/run/streaming-shell-command.js +14 -5
- package/dist/cli-tui/run/streaming-shell-command.js.map +1 -1
- package/dist/cli-tui/selectors/model-selector.d.ts +1 -0
- package/dist/cli-tui/selectors/model-selector.d.ts.map +1 -1
- package/dist/cli-tui/selectors/model-selector.js +6 -2
- package/dist/cli-tui/selectors/model-selector.js.map +1 -1
- package/dist/cli-tui/tui-renderer/command-registry-options.d.ts.map +1 -1
- package/dist/cli-tui/tui-renderer/command-registry-options.js +5 -0
- package/dist/cli-tui/tui-renderer/command-registry-options.js.map +1 -1
- package/dist/cli-tui/utils/commands/review-prompt.d.ts +1 -1
- package/dist/cli-tui/utils/commands/review-prompt.d.ts.map +1 -1
- package/dist/cli-tui/utils/commands/review-prompt.js +8 -3
- package/dist/cli-tui/utils/commands/review-prompt.js.map +1 -1
- package/dist/cli-tui/utils/runtime-badges.d.ts +1 -0
- package/dist/cli-tui/utils/runtime-badges.d.ts.map +1 -1
- package/dist/cli-tui/utils/runtime-badges.js +7 -0
- package/dist/cli-tui/utils/runtime-badges.js.map +1 -1
- package/dist/cli-tui/utils/text-formatting.d.ts +1 -0
- package/dist/cli-tui/utils/text-formatting.d.ts.map +1 -1
- package/dist/cli-tui/utils/text-formatting.js +5 -0
- package/dist/cli-tui/utils/text-formatting.js.map +1 -1
- package/dist/cli.js +49209 -26165
- package/dist/cli.js.map +1 -1
- package/dist/codex/app-server-client.d.ts +139 -0
- package/dist/codex/app-server-client.d.ts.map +1 -0
- package/dist/codex/app-server-client.js +352 -0
- package/dist/codex/app-server-client.js.map +1 -0
- package/dist/codex/compatibility.d.ts +48 -0
- package/dist/codex/compatibility.d.ts.map +1 -0
- package/dist/codex/compatibility.js +344 -0
- package/dist/codex/compatibility.js.map +1 -0
- package/dist/codex/subagent-dispatch-table.d.ts +66 -0
- package/dist/codex/subagent-dispatch-table.d.ts.map +1 -0
- package/dist/codex/subagent-dispatch-table.js +127 -0
- package/dist/codex/subagent-dispatch-table.js.map +1 -0
- package/dist/codex/subagent-workgraph.d.ts +56 -0
- package/dist/codex/subagent-workgraph.d.ts.map +1 -0
- package/dist/codex/subagent-workgraph.js +89 -0
- package/dist/codex/subagent-workgraph.js.map +1 -0
- package/dist/config/constants.d.ts +3 -1
- package/dist/config/constants.d.ts.map +1 -1
- package/dist/config/constants.js +9 -0
- package/dist/config/constants.js.map +1 -1
- package/dist/config/feature-flags.d.ts +10 -1
- package/dist/config/feature-flags.d.ts.map +1 -1
- package/dist/config/feature-flags.js +17 -6
- package/dist/config/feature-flags.js.map +1 -1
- package/dist/config/index.d.ts +1 -1
- package/dist/config/index.d.ts.map +1 -1
- package/dist/config/index.js +1 -1
- package/dist/config/index.js.map +1 -1
- package/dist/config/toml-config.d.ts +32 -0
- package/dist/config/toml-config.d.ts.map +1 -1
- package/dist/config/toml-config.js +152 -43
- package/dist/config/toml-config.js.map +1 -1
- package/dist/connectors/service-client.d.ts.map +1 -1
- package/dist/connectors/service-client.js +4 -7
- package/dist/connectors/service-client.js.map +1 -1
- package/dist/context/manifest-types.d.ts +59 -0
- package/dist/context/manifest-types.d.ts.map +1 -0
- package/dist/context/manifest-types.js +2 -0
- package/dist/context/manifest-types.js.map +1 -0
- package/dist/context/manifest.d.ts +16 -0
- package/dist/context/manifest.d.ts.map +1 -0
- package/dist/context/manifest.js +503 -0
- package/dist/context/manifest.js.map +1 -0
- package/dist/db/schema.d.ts +10 -1
- package/dist/db/schema.d.ts.map +1 -1
- package/dist/db/schema.js.map +1 -1
- package/dist/evalops/agent-bootstrap.d.ts +1 -0
- package/dist/evalops/agent-bootstrap.d.ts.map +1 -1
- package/dist/evalops/agent-bootstrap.js +41 -1
- package/dist/evalops/agent-bootstrap.js.map +1 -1
- package/dist/evalops/env-aliases.d.ts +13 -0
- package/dist/evalops/env-aliases.d.ts.map +1 -0
- package/dist/evalops/env-aliases.js +60 -0
- package/dist/evalops/env-aliases.js.map +1 -0
- package/dist/evalops/managed-context.d.ts.map +1 -1
- package/dist/evalops/managed-context.js +28 -62
- package/dist/evalops/managed-context.js.map +1 -1
- package/dist/guardian/config.d.ts.map +1 -1
- package/dist/guardian/config.js +1 -0
- package/dist/guardian/config.js.map +1 -1
- package/dist/guardian/runner.d.ts +2 -0
- package/dist/guardian/runner.d.ts.map +1 -1
- package/dist/guardian/runner.js +105 -2
- package/dist/guardian/runner.js.map +1 -1
- package/dist/guardian/types.d.ts +1 -0
- package/dist/guardian/types.d.ts.map +1 -1
- package/dist/guardian/types.js.map +1 -1
- package/dist/index.d.ts +6 -1
- package/dist/index.d.ts.map +1 -1
- package/dist/index.js +6 -1
- package/dist/index.js.map +1 -1
- package/dist/main.d.ts.map +1 -1
- package/dist/main.js +267 -55
- package/dist/main.js.map +1 -1
- package/dist/mcp/config.d.ts.map +1 -1
- package/dist/mcp/config.js +89 -7
- package/dist/mcp/config.js.map +1 -1
- package/dist/mcp/manager.d.ts +5 -1
- package/dist/mcp/manager.d.ts.map +1 -1
- package/dist/mcp/manager.js +159 -4
- package/dist/mcp/manager.js.map +1 -1
- package/dist/mcp/platform-plugin.d.ts.map +1 -1
- package/dist/mcp/platform-plugin.js +9 -26
- package/dist/mcp/platform-plugin.js.map +1 -1
- package/dist/mcp/schema.d.ts +15 -0
- package/dist/mcp/schema.d.ts.map +1 -1
- package/dist/mcp/schema.js +24 -0
- package/dist/mcp/schema.js.map +1 -1
- package/dist/mcp/tool-bridge.d.ts +39 -10
- package/dist/mcp/tool-bridge.d.ts.map +1 -1
- package/dist/mcp/tool-bridge.js +101 -24
- package/dist/mcp/tool-bridge.js.map +1 -1
- package/dist/mcp/types.d.ts +27 -0
- package/dist/mcp/types.d.ts.map +1 -1
- package/dist/mcp/types.js.map +1 -1
- package/dist/mcp/workspace-trust.d.ts +10 -0
- package/dist/mcp/workspace-trust.d.ts.map +1 -0
- package/dist/mcp/workspace-trust.js +295 -0
- package/dist/mcp/workspace-trust.js.map +1 -0
- package/dist/memory/service-client.d.ts +1 -0
- package/dist/memory/service-client.d.ts.map +1 -1
- package/dist/memory/service-client.js +194 -9
- package/dist/memory/service-client.js.map +1 -1
- package/dist/models/builtin.d.ts.map +1 -1
- package/dist/models/builtin.js +66 -25
- package/dist/models/builtin.js.map +1 -1
- package/dist/models/config-loader.d.ts +10 -10
- package/dist/models/config-loader.d.ts.map +1 -1
- package/dist/models/config-loader.js +2 -0
- package/dist/models/config-loader.js.map +1 -1
- package/dist/models/registry.d.ts +1 -0
- package/dist/models/registry.d.ts.map +1 -1
- package/dist/models/registry.js +1 -0
- package/dist/models/registry.js.map +1 -1
- package/dist/oauth/device-identity.d.ts.map +1 -1
- package/dist/oauth/device-identity.js +1 -3
- package/dist/oauth/device-identity.js.map +1 -1
- package/dist/oauth/evalops.d.ts.map +1 -1
- package/dist/oauth/evalops.js +2 -5
- package/dist/oauth/evalops.js.map +1 -1
- package/dist/opentelemetry.d.ts.map +1 -1
- package/dist/opentelemetry.js +12 -12
- package/dist/opentelemetry.js.map +1 -1
- package/dist/package-version.d.ts +2 -0
- package/dist/package-version.d.ts.map +1 -0
- package/dist/package-version.js +11 -0
- package/dist/package-version.js.map +1 -0
- package/dist/packages/constants.d.ts +3 -0
- package/dist/packages/constants.d.ts.map +1 -0
- package/dist/packages/constants.js +3 -0
- package/dist/packages/constants.js.map +1 -0
- package/dist/packages/inspection.d.ts.map +1 -1
- package/dist/packages/inspection.js +2 -1
- package/dist/packages/inspection.js.map +1 -1
- package/dist/packages/search.d.ts.map +1 -1
- package/dist/packages/search.js +1 -1
- package/dist/packages/search.js.map +1 -1
- package/dist/platform/a2a-capability-market.d.ts +21 -0
- package/dist/platform/a2a-capability-market.d.ts.map +1 -0
- package/dist/platform/a2a-capability-market.js +176 -0
- package/dist/platform/a2a-capability-market.js.map +1 -0
- package/dist/platform/a2a-client.d.ts +217 -0
- package/dist/platform/a2a-client.d.ts.map +1 -0
- package/dist/platform/a2a-client.js +589 -0
- package/dist/platform/a2a-client.js.map +1 -0
- package/dist/platform/a2a-fleet.d.ts +41 -0
- package/dist/platform/a2a-fleet.d.ts.map +1 -0
- package/dist/platform/a2a-fleet.js +109 -0
- package/dist/platform/a2a-fleet.js.map +1 -0
- package/dist/platform/a2a-maestro-peer.d.ts +20 -0
- package/dist/platform/a2a-maestro-peer.d.ts.map +1 -0
- package/dist/platform/a2a-maestro-peer.js +176 -0
- package/dist/platform/a2a-maestro-peer.js.map +1 -0
- package/dist/platform/a2a-peer-pairing.d.ts +107 -0
- package/dist/platform/a2a-peer-pairing.d.ts.map +1 -0
- package/dist/platform/a2a-peer-pairing.js +489 -0
- package/dist/platform/a2a-peer-pairing.js.map +1 -0
- package/dist/platform/a2a-peer-registry.d.ts +68 -0
- package/dist/platform/a2a-peer-registry.d.ts.map +1 -0
- package/dist/platform/a2a-peer-registry.js +320 -0
- package/dist/platform/a2a-peer-registry.js.map +1 -0
- package/dist/platform/a2a-task-ledger.d.ts +86 -0
- package/dist/platform/a2a-task-ledger.d.ts.map +1 -0
- package/dist/platform/a2a-task-ledger.js +546 -0
- package/dist/platform/a2a-task-ledger.js.map +1 -0
- package/dist/platform/a2a-work-graph.d.ts +43 -0
- package/dist/platform/a2a-work-graph.d.ts.map +1 -0
- package/dist/platform/a2a-work-graph.js +267 -0
- package/dist/platform/a2a-work-graph.js.map +1 -0
- package/dist/platform/agent-registry-client.d.ts +327 -0
- package/dist/platform/agent-registry-client.d.ts.map +1 -0
- package/dist/platform/agent-registry-client.js +803 -0
- package/dist/platform/agent-registry-client.js.map +1 -0
- package/dist/platform/agent-runtime-client.d.ts +135 -1
- package/dist/platform/agent-runtime-client.d.ts.map +1 -1
- package/dist/platform/agent-runtime-client.js +467 -10
- package/dist/platform/agent-runtime-client.js.map +1 -1
- package/dist/platform/cerebro-facts-client.d.ts.map +1 -1
- package/dist/platform/cerebro-facts-client.js +4 -7
- package/dist/platform/cerebro-facts-client.js.map +1 -1
- package/dist/platform/client.d.ts +2 -0
- package/dist/platform/client.d.ts.map +1 -1
- package/dist/platform/client.js +7 -14
- package/dist/platform/client.js.map +1 -1
- package/dist/platform/core-services.d.ts +68 -0
- package/dist/platform/core-services.d.ts.map +1 -1
- package/dist/platform/core-services.js +68 -0
- package/dist/platform/core-services.js.map +1 -1
- package/dist/platform/fermata-eval-client.d.ts +195 -0
- package/dist/platform/fermata-eval-client.d.ts.map +1 -0
- package/dist/platform/fermata-eval-client.js +146 -0
- package/dist/platform/fermata-eval-client.js.map +1 -0
- package/dist/platform/fermata-scenario-suite.d.ts +99 -0
- package/dist/platform/fermata-scenario-suite.d.ts.map +1 -0
- package/dist/platform/fermata-scenario-suite.js +459 -0
- package/dist/platform/fermata-scenario-suite.js.map +1 -0
- package/dist/platform/maestro-timeline-client.d.ts.map +1 -1
- package/dist/platform/maestro-timeline-client.js +5 -11
- package/dist/platform/maestro-timeline-client.js.map +1 -1
- package/dist/platform/operating-plane-client.d.ts +156 -0
- package/dist/platform/operating-plane-client.d.ts.map +1 -0
- package/dist/platform/operating-plane-client.js +131 -0
- package/dist/platform/operating-plane-client.js.map +1 -0
- package/dist/platform/operating-plane-summary.d.ts +47 -0
- package/dist/platform/operating-plane-summary.d.ts.map +1 -0
- package/dist/platform/operating-plane-summary.js +209 -0
- package/dist/platform/operating-plane-summary.js.map +1 -0
- package/dist/platform/tool-execution-client.d.ts.map +1 -1
- package/dist/platform/tool-execution-client.js +4 -11
- package/dist/platform/tool-execution-client.js.map +1 -1
- package/dist/prompts/service-client.d.ts.map +1 -1
- package/dist/prompts/service-client.js +3 -5
- package/dist/prompts/service-client.js.map +1 -1
- package/dist/prompts/system-prompt.d.ts +3 -0
- package/dist/prompts/system-prompt.d.ts.map +1 -1
- package/dist/prompts/system-prompt.js +13 -3
- package/dist/prompts/system-prompt.js.map +1 -1
- package/dist/prompts/types.d.ts +2 -0
- package/dist/prompts/types.d.ts.map +1 -1
- package/dist/prompts/types.js.map +1 -1
- package/dist/providers/api-keys.d.ts +1 -1
- package/dist/providers/api-keys.d.ts.map +1 -1
- package/dist/providers/api-keys.js +2 -1
- package/dist/providers/api-keys.js.map +1 -1
- package/dist/providers/auth.d.ts.map +1 -1
- package/dist/providers/auth.js +48 -25
- package/dist/providers/auth.js.map +1 -1
- package/dist/remote-runner/attach-client.d.ts.map +1 -1
- package/dist/remote-runner/attach-client.js +1 -0
- package/dist/remote-runner/attach-client.js.map +1 -1
- package/dist/remote-runner/client.d.ts.map +1 -1
- package/dist/remote-runner/client.js +4 -7
- package/dist/remote-runner/client.js.map +1 -1
- package/dist/rpc/rpc-types.d.ts.map +1 -1
- package/dist/rpc/rpc-types.js +1 -0
- package/dist/rpc/rpc-types.js.map +1 -1
- package/dist/safety/action-firewall.d.ts.map +1 -1
- package/dist/safety/action-firewall.js +69 -6
- package/dist/safety/action-firewall.js.map +1 -1
- package/dist/safety/attack-patterns.d.ts.map +1 -1
- package/dist/safety/attack-patterns.js +31 -14
- package/dist/safety/attack-patterns.js.map +1 -1
- package/dist/safety/bash-parser.d.ts.map +1 -1
- package/dist/safety/bash-parser.js +4 -3
- package/dist/safety/bash-parser.js.map +1 -1
- package/dist/safety/execpolicy.d.ts +15 -3
- package/dist/safety/execpolicy.d.ts.map +1 -1
- package/dist/safety/execpolicy.js +248 -37
- package/dist/safety/execpolicy.js.map +1 -1
- package/dist/safety/governance-service-client.d.ts.map +1 -1
- package/dist/safety/governance-service-client.js +4 -7
- package/dist/safety/governance-service-client.js.map +1 -1
- package/dist/safety/guarded-files.d.ts +34 -0
- package/dist/safety/guarded-files.d.ts.map +1 -0
- package/dist/safety/guarded-files.js +372 -0
- package/dist/safety/guarded-files.js.map +1 -0
- package/dist/safety/loop-detector.d.ts.map +1 -1
- package/dist/safety/loop-detector.js +6 -1
- package/dist/safety/loop-detector.js.map +1 -1
- package/dist/safety/permission-profile.d.ts +47 -0
- package/dist/safety/permission-profile.d.ts.map +1 -0
- package/dist/safety/permission-profile.js +349 -0
- package/dist/safety/permission-profile.js.map +1 -0
- package/dist/safety/safety-middleware.d.ts +5 -1
- package/dist/safety/safety-middleware.d.ts.map +1 -1
- package/dist/safety/safety-middleware.js +3 -2
- package/dist/safety/safety-middleware.js.map +1 -1
- package/dist/safety/suspicious-patterns.d.ts.map +1 -1
- package/dist/safety/suspicious-patterns.js +12 -7
- package/dist/safety/suspicious-patterns.js.map +1 -1
- package/dist/safety/tool-categorization.js +1 -1
- package/dist/safety/tool-categorization.js.map +1 -1
- package/dist/safety/validators/path-policy-validator.d.ts.map +1 -1
- package/dist/safety/validators/path-policy-validator.js +7 -0
- package/dist/safety/validators/path-policy-validator.js.map +1 -1
- package/dist/sandbox/index.d.ts +6 -1
- package/dist/sandbox/index.d.ts.map +1 -1
- package/dist/sandbox/index.js +30 -4
- package/dist/sandbox/index.js.map +1 -1
- package/dist/sandbox/native-sandbox.d.ts +3 -0
- package/dist/sandbox/native-sandbox.d.ts.map +1 -1
- package/dist/sandbox/native-sandbox.js +139 -18
- package/dist/sandbox/native-sandbox.js.map +1 -1
- package/dist/sentry.d.ts +17 -0
- package/dist/sentry.d.ts.map +1 -0
- package/dist/sentry.js +91 -0
- package/dist/sentry.js.map +1 -0
- package/dist/server/agent-runtime-ledger.d.ts +141 -0
- package/dist/server/agent-runtime-ledger.d.ts.map +1 -0
- package/dist/server/agent-runtime-ledger.js +334 -0
- package/dist/server/agent-runtime-ledger.js.map +1 -0
- package/dist/server/agent-trajectory-inspection.d.ts +104 -0
- package/dist/server/agent-trajectory-inspection.d.ts.map +1 -0
- package/dist/server/agent-trajectory-inspection.js +168 -0
- package/dist/server/agent-trajectory-inspection.js.map +1 -0
- package/dist/server/agent-trajectory-replay.d.ts +56 -0
- package/dist/server/agent-trajectory-replay.d.ts.map +1 -0
- package/dist/server/agent-trajectory-replay.js +214 -0
- package/dist/server/agent-trajectory-replay.js.map +1 -0
- package/dist/server/agent-trajectory-scenarios.d.ts +109 -0
- package/dist/server/agent-trajectory-scenarios.d.ts.map +1 -0
- package/dist/server/agent-trajectory-scenarios.js +822 -0
- package/dist/server/agent-trajectory-scenarios.js.map +1 -0
- package/dist/server/agent-trajectory-scorers.d.ts +61 -0
- package/dist/server/agent-trajectory-scorers.d.ts.map +1 -0
- package/dist/server/agent-trajectory-scorers.js +162 -0
- package/dist/server/agent-trajectory-scorers.js.map +1 -0
- package/dist/server/agent-trajectory-validation.d.ts +7 -0
- package/dist/server/agent-trajectory-validation.d.ts.map +1 -0
- package/dist/server/agent-trajectory-validation.js +101 -0
- package/dist/server/agent-trajectory-validation.js.map +1 -0
- package/dist/server/agent-trajectory.d.ts +50 -0
- package/dist/server/agent-trajectory.d.ts.map +1 -0
- package/dist/server/agent-trajectory.js +159 -0
- package/dist/server/agent-trajectory.js.map +1 -0
- package/dist/server/app-context.d.ts +18 -0
- package/dist/server/app-context.d.ts.map +1 -1
- package/dist/server/app-context.js.map +1 -1
- package/dist/server/handlers/headless-sessions.d.ts.map +1 -1
- package/dist/server/handlers/headless-sessions.js +64 -27
- package/dist/server/handlers/headless-sessions.js.map +1 -1
- package/dist/server/handlers/health.d.ts +6 -0
- package/dist/server/handlers/health.d.ts.map +1 -1
- package/dist/server/handlers/health.js +6 -0
- package/dist/server/handlers/health.js.map +1 -1
- package/dist/server/handlers/hosted-runner-drain.d.ts +69 -1
- package/dist/server/handlers/hosted-runner-drain.d.ts.map +1 -1
- package/dist/server/handlers/hosted-runner-drain.js +309 -18
- package/dist/server/handlers/hosted-runner-drain.js.map +1 -1
- package/dist/server/handlers/hosted-runner-identity.d.ts +16 -0
- package/dist/server/handlers/hosted-runner-identity.d.ts.map +1 -1
- package/dist/server/handlers/hosted-runner-identity.js +43 -0
- package/dist/server/handlers/hosted-runner-identity.js.map +1 -1
- package/dist/server/handlers/mode.d.ts.map +1 -1
- package/dist/server/handlers/mode.js +3 -1
- package/dist/server/handlers/mode.js.map +1 -1
- package/dist/server/handlers/pending-requests.d.ts.map +1 -1
- package/dist/server/handlers/pending-requests.js +336 -51
- package/dist/server/handlers/pending-requests.js.map +1 -1
- package/dist/server/handlers/platform-a2a-push.d.ts +6 -0
- package/dist/server/handlers/platform-a2a-push.d.ts.map +1 -0
- package/dist/server/handlers/platform-a2a-push.js +136 -0
- package/dist/server/handlers/platform-a2a-push.js.map +1 -0
- package/dist/server/handlers/sessions.d.ts.map +1 -1
- package/dist/server/handlers/sessions.js +21 -1
- package/dist/server/handlers/sessions.js.map +1 -1
- package/dist/server/headless-runtime-service.d.ts +9 -2
- package/dist/server/headless-runtime-service.d.ts.map +1 -1
- package/dist/server/headless-runtime-service.js +39 -0
- package/dist/server/headless-runtime-service.js.map +1 -1
- package/dist/server/hosted-agent-runtime-progress.d.ts +144 -0
- package/dist/server/hosted-agent-runtime-progress.d.ts.map +1 -0
- package/dist/server/hosted-agent-runtime-progress.js +1650 -0
- package/dist/server/hosted-agent-runtime-progress.js.map +1 -0
- package/dist/server/hosted-runner-lease.d.ts +41 -0
- package/dist/server/hosted-runner-lease.d.ts.map +1 -0
- package/dist/server/hosted-runner-lease.js +140 -0
- package/dist/server/hosted-runner-lease.js.map +1 -0
- package/dist/server/hosted-session-manager.d.ts +16 -8
- package/dist/server/hosted-session-manager.d.ts.map +1 -1
- package/dist/server/hosted-session-manager.js +98 -76
- package/dist/server/hosted-session-manager.js.map +1 -1
- package/dist/server/junit-xml.d.ts +2 -0
- package/dist/server/junit-xml.d.ts.map +1 -0
- package/dist/server/junit-xml.js +9 -0
- package/dist/server/junit-xml.js.map +1 -0
- package/dist/server/routes.d.ts.map +1 -1
- package/dist/server/routes.js +6 -0
- package/dist/server/routes.js.map +1 -1
- package/dist/server/scenario-recorder.d.ts +56 -0
- package/dist/server/scenario-recorder.d.ts.map +1 -0
- package/dist/server/scenario-recorder.js +109 -0
- package/dist/server/scenario-recorder.js.map +1 -0
- package/dist/server/scripted-scenario-runner.d.ts +46 -0
- package/dist/server/scripted-scenario-runner.d.ts.map +1 -0
- package/dist/server/scripted-scenario-runner.js +190 -0
- package/dist/server/scripted-scenario-runner.js.map +1 -0
- package/dist/server/server-middlewares.d.ts +6 -2
- package/dist/server/server-middlewares.d.ts.map +1 -1
- package/dist/server/server-middlewares.js +8 -2
- package/dist/server/server-middlewares.js.map +1 -1
- package/dist/server/server-request-manager.d.ts +9 -1
- package/dist/server/server-request-manager.d.ts.map +1 -1
- package/dist/server/server-request-manager.js +49 -4
- package/dist/server/server-request-manager.js.map +1 -1
- package/dist/services/intelligent-router/service.d.ts +2 -0
- package/dist/services/intelligent-router/service.d.ts.map +1 -1
- package/dist/services/intelligent-router/service.js +4 -0
- package/dist/services/intelligent-router/service.js.map +1 -1
- package/dist/services/service-authority.d.ts +25 -0
- package/dist/services/service-authority.d.ts.map +1 -0
- package/dist/services/service-authority.js +143 -0
- package/dist/services/service-authority.js.map +1 -0
- package/dist/services/workspace-config/service.d.ts +2 -0
- package/dist/services/workspace-config/service.d.ts.map +1 -1
- package/dist/services/workspace-config/service.js +4 -0
- package/dist/services/workspace-config/service.js.map +1 -1
- package/dist/session/manager.d.ts +8 -2
- package/dist/session/manager.d.ts.map +1 -1
- package/dist/session/manager.js +19 -4
- package/dist/session/manager.js.map +1 -1
- package/dist/session/session-branch.d.ts.map +1 -1
- package/dist/session/session-branch.js +5 -1
- package/dist/session/session-branch.js.map +1 -1
- package/dist/session/session-catalog.d.ts +7 -2
- package/dist/session/session-catalog.d.ts.map +1 -1
- package/dist/session/session-catalog.js +13 -2
- package/dist/session/session-catalog.js.map +1 -1
- package/dist/session/session-context.d.ts +6 -2
- package/dist/session/session-context.d.ts.map +1 -1
- package/dist/session/session-context.js +44 -4
- package/dist/session/session-context.js.map +1 -1
- package/dist/session/session-graph-projection.d.ts +32 -0
- package/dist/session/session-graph-projection.d.ts.map +1 -0
- package/dist/session/session-graph-projection.js +133 -0
- package/dist/session/session-graph-projection.js.map +1 -0
- package/dist/session/types.d.ts +18 -0
- package/dist/session/types.d.ts.map +1 -1
- package/dist/session/types.js +8 -0
- package/dist/session/types.js.map +1 -1
- package/dist/session/wire-format.generated.d.ts +3 -0
- package/dist/session/wire-format.generated.d.ts.map +1 -1
- package/dist/session/wire-format.generated.js +20 -0
- package/dist/session/wire-format.generated.js.map +1 -1
- package/dist/skills/eval-harness.d.ts +38 -0
- package/dist/skills/eval-harness.d.ts.map +1 -0
- package/dist/skills/eval-harness.js +112 -0
- package/dist/skills/eval-harness.js.map +1 -0
- package/dist/skills/index.d.ts +5 -1
- package/dist/skills/index.d.ts.map +1 -1
- package/dist/skills/index.js +5 -1
- package/dist/skills/index.js.map +1 -1
- package/dist/skills/linter.d.ts +36 -0
- package/dist/skills/linter.d.ts.map +1 -0
- package/dist/skills/linter.js +398 -0
- package/dist/skills/linter.js.map +1 -0
- package/dist/skills/loader.d.ts +36 -6
- package/dist/skills/loader.d.ts.map +1 -1
- package/dist/skills/loader.js +169 -86
- package/dist/skills/loader.js.map +1 -1
- package/dist/skills/package-contract.d.ts +45 -0
- package/dist/skills/package-contract.d.ts.map +1 -0
- package/dist/skills/package-contract.js +153 -0
- package/dist/skills/package-contract.js.map +1 -0
- package/dist/skills/runtime-activation.d.ts +68 -0
- package/dist/skills/runtime-activation.d.ts.map +1 -0
- package/dist/skills/runtime-activation.js +228 -0
- package/dist/skills/runtime-activation.js.map +1 -0
- package/dist/skills/service-client.d.ts.map +1 -1
- package/dist/skills/service-client.js +3 -4
- package/dist/skills/service-client.js.map +1 -1
- package/dist/skills/tool.d.ts.map +1 -1
- package/dist/skills/tool.js +2 -0
- package/dist/skills/tool.js.map +1 -1
- package/dist/skills/types.d.ts +14 -0
- package/dist/skills/types.d.ts.map +1 -1
- package/dist/skills/types.js.map +1 -1
- package/dist/telemetry/agent-operating-plane-context.d.ts +24 -0
- package/dist/telemetry/agent-operating-plane-context.d.ts.map +1 -0
- package/dist/telemetry/agent-operating-plane-context.js +54 -0
- package/dist/telemetry/agent-operating-plane-context.js.map +1 -0
- package/dist/telemetry/beacon.d.ts +24 -0
- package/dist/telemetry/beacon.d.ts.map +1 -0
- package/dist/telemetry/beacon.js +115 -0
- package/dist/telemetry/beacon.js.map +1 -0
- package/dist/telemetry/cli-command-aggregator.d.ts +36 -0
- package/dist/telemetry/cli-command-aggregator.d.ts.map +1 -0
- package/dist/telemetry/cli-command-aggregator.js +251 -0
- package/dist/telemetry/cli-command-aggregator.js.map +1 -0
- package/dist/telemetry/cli-startup.d.ts +21 -0
- package/dist/telemetry/cli-startup.d.ts.map +1 -0
- package/dist/telemetry/cli-startup.js +83 -0
- package/dist/telemetry/cli-startup.js.map +1 -0
- package/dist/telemetry/disablement.d.ts +13 -0
- package/dist/telemetry/disablement.d.ts.map +1 -0
- package/dist/telemetry/disablement.js +19 -0
- package/dist/telemetry/disablement.js.map +1 -0
- package/dist/telemetry/index.d.ts +6 -2
- package/dist/telemetry/index.d.ts.map +1 -1
- package/dist/telemetry/index.js +6 -2
- package/dist/telemetry/index.js.map +1 -1
- package/dist/telemetry/maestro-event-bus.d.ts +52 -0
- package/dist/telemetry/maestro-event-bus.d.ts.map +1 -1
- package/dist/telemetry/maestro-event-bus.js +179 -3
- package/dist/telemetry/maestro-event-bus.js.map +1 -1
- package/dist/telemetry/maestro-event-catalog.d.ts +3 -1
- package/dist/telemetry/maestro-event-catalog.d.ts.map +1 -1
- package/dist/telemetry/maestro-event-catalog.js +2 -0
- package/dist/telemetry/maestro-event-catalog.js.map +1 -1
- package/dist/telemetry/maestro-platform-replay-fixture.d.ts.map +1 -1
- package/dist/telemetry/maestro-platform-replay-fixture.js +74 -9
- package/dist/telemetry/maestro-platform-replay-fixture.js.map +1 -1
- package/dist/telemetry/maestro-publisher-conformance-fixture.d.ts +1 -1
- package/dist/telemetry/maestro-publisher-conformance-fixture.d.ts.map +1 -1
- package/dist/telemetry/maestro-publisher-conformance-fixture.js +46 -11
- package/dist/telemetry/maestro-publisher-conformance-fixture.js.map +1 -1
- package/dist/telemetry/mcp-beacon.d.ts +20 -0
- package/dist/telemetry/mcp-beacon.d.ts.map +1 -0
- package/dist/telemetry/mcp-beacon.js +54 -0
- package/dist/telemetry/mcp-beacon.js.map +1 -0
- package/dist/telemetry/metadata-normalization.d.ts +7 -0
- package/dist/telemetry/metadata-normalization.d.ts.map +1 -0
- package/dist/telemetry/metadata-normalization.js +189 -0
- package/dist/telemetry/metadata-normalization.js.map +1 -0
- package/dist/telemetry/meter-service-client.d.ts.map +1 -1
- package/dist/telemetry/meter-service-client.js +5 -10
- package/dist/telemetry/meter-service-client.js.map +1 -1
- package/dist/telemetry/metrics.d.ts +123 -0
- package/dist/telemetry/metrics.d.ts.map +1 -0
- package/dist/telemetry/metrics.js +184 -0
- package/dist/telemetry/metrics.js.map +1 -0
- package/dist/telemetry/turn-tracker.d.ts +1 -0
- package/dist/telemetry/turn-tracker.d.ts.map +1 -1
- package/dist/telemetry/turn-tracker.js +22 -1
- package/dist/telemetry/turn-tracker.js.map +1 -1
- package/dist/telemetry/wide-events.d.ts +31 -1
- package/dist/telemetry/wide-events.d.ts.map +1 -1
- package/dist/telemetry/wide-events.js +229 -1
- package/dist/telemetry/wide-events.js.map +1 -1
- package/dist/telemetry.d.ts +39 -3
- package/dist/telemetry.d.ts.map +1 -1
- package/dist/telemetry.js +160 -3
- package/dist/telemetry.js.map +1 -1
- package/dist/testing/test-verification-hook.d.ts.map +1 -1
- package/dist/testing/test-verification-hook.js +34 -21
- package/dist/testing/test-verification-hook.js.map +1 -1
- package/dist/tools/apply-patch-parser.d.ts +26 -0
- package/dist/tools/apply-patch-parser.d.ts.map +1 -0
- package/dist/tools/apply-patch-parser.js +146 -0
- package/dist/tools/apply-patch-parser.js.map +1 -0
- package/dist/tools/apply-patch.d.ts +42 -0
- package/dist/tools/apply-patch.d.ts.map +1 -0
- package/dist/tools/apply-patch.js +648 -0
- package/dist/tools/apply-patch.js.map +1 -0
- package/dist/tools/ask-user.d.ts +1 -0
- package/dist/tools/ask-user.d.ts.map +1 -1
- package/dist/tools/background/tool-handler.d.ts +1 -0
- package/dist/tools/background/tool-handler.d.ts.map +1 -1
- package/dist/tools/bash.d.ts +1 -0
- package/dist/tools/bash.d.ts.map +1 -1
- package/dist/tools/codesearch.d.ts +1 -0
- package/dist/tools/codesearch.d.ts.map +1 -1
- package/dist/tools/conductor-client.d.ts +1 -0
- package/dist/tools/conductor-client.d.ts.map +1 -1
- package/dist/tools/conductor-client.js +63 -0
- package/dist/tools/conductor-client.js.map +1 -1
- package/dist/tools/diff.d.ts +1 -0
- package/dist/tools/diff.d.ts.map +1 -1
- package/dist/tools/edit.d.ts +1 -0
- package/dist/tools/edit.d.ts.map +1 -1
- package/dist/tools/extract-document.d.ts +1 -0
- package/dist/tools/extract-document.d.ts.map +1 -1
- package/dist/tools/find.d.ts +10 -0
- package/dist/tools/find.d.ts.map +1 -1
- package/dist/tools/find.js +236 -27
- package/dist/tools/find.js.map +1 -1
- package/dist/tools/gh.d.ts +3 -0
- package/dist/tools/gh.d.ts.map +1 -1
- package/dist/tools/index.d.ts +50 -0
- package/dist/tools/index.d.ts.map +1 -1
- package/dist/tools/index.js +4 -0
- package/dist/tools/index.js.map +1 -1
- package/dist/tools/list.d.ts +1 -0
- package/dist/tools/list.d.ts.map +1 -1
- package/dist/tools/notebook.d.ts +1 -0
- package/dist/tools/notebook.d.ts.map +1 -1
- package/dist/tools/oracle.d.ts +1 -0
- package/dist/tools/oracle.d.ts.map +1 -1
- package/dist/tools/oracle.js +17 -32
- package/dist/tools/oracle.js.map +1 -1
- package/dist/tools/parallel-execution.d.ts +13 -2
- package/dist/tools/parallel-execution.d.ts.map +1 -1
- package/dist/tools/parallel-execution.js +178 -7
- package/dist/tools/parallel-execution.js.map +1 -1
- package/dist/tools/parallel-ripgrep.d.ts +1 -0
- package/dist/tools/parallel-ripgrep.d.ts.map +1 -1
- package/dist/tools/pipeline.d.ts +4 -0
- package/dist/tools/pipeline.d.ts.map +1 -1
- package/dist/tools/read.d.ts +1 -0
- package/dist/tools/read.d.ts.map +1 -1
- package/dist/tools/search.d.ts +1 -0
- package/dist/tools/search.d.ts.map +1 -1
- package/dist/tools/status.d.ts +1 -0
- package/dist/tools/status.d.ts.map +1 -1
- package/dist/tools/todo.d.ts +1 -0
- package/dist/tools/todo.d.ts.map +1 -1
- package/dist/tools/tool-dsl.d.ts +5 -1
- package/dist/tools/tool-dsl.d.ts.map +1 -1
- package/dist/tools/tool-dsl.js +1 -0
- package/dist/tools/tool-dsl.js.map +1 -1
- package/dist/tools/tool-result-cache.d.ts.map +1 -1
- package/dist/tools/tool-result-cache.js +1 -0
- package/dist/tools/tool-result-cache.js.map +1 -1
- package/dist/tools/typebox-tool.d.ts.map +1 -1
- package/dist/tools/typebox-tool.js +3 -0
- package/dist/tools/typebox-tool.js.map +1 -1
- package/dist/tools/webfetch.d.ts +1 -0
- package/dist/tools/webfetch.d.ts.map +1 -1
- package/dist/tools/websearch.d.ts +1 -0
- package/dist/tools/websearch.d.ts.map +1 -1
- package/dist/tools/write.d.ts +1 -0
- package/dist/tools/write.d.ts.map +1 -1
- package/dist/utils/abort-error.d.ts +2 -0
- package/dist/utils/abort-error.d.ts.map +1 -0
- package/dist/utils/abort-error.js +7 -0
- package/dist/utils/abort-error.js.map +1 -0
- package/dist/utils/downstream.d.ts.map +1 -1
- package/dist/utils/downstream.js +4 -0
- package/dist/utils/downstream.js.map +1 -1
- package/dist/utils/logger.d.ts +3 -0
- package/dist/utils/logger.d.ts.map +1 -1
- package/dist/utils/logger.js +14 -3
- package/dist/utils/logger.js.map +1 -1
- package/dist/utils/path-containment.d.ts +2 -0
- package/dist/utils/path-containment.d.ts.map +1 -0
- package/dist/utils/path-containment.js +14 -0
- package/dist/utils/path-containment.js.map +1 -0
- package/dist/utils/prompt-safe-text.d.ts +7 -0
- package/dist/utils/prompt-safe-text.d.ts.map +1 -0
- package/dist/utils/prompt-safe-text.js +44 -0
- package/dist/utils/prompt-safe-text.js.map +1 -0
- package/dist/version.json +2 -2
- package/dist/web-server.d.ts.map +1 -1
- package/dist/web-server.js +32 -7
- package/dist/web-server.js.map +1 -1
- package/dist/workspace/root-resolver.d.ts +4 -1
- package/dist/workspace/root-resolver.d.ts.map +1 -1
- package/dist/workspace/root-resolver.js +16 -6
- package/dist/workspace/root-resolver.js.map +1 -1
- package/node_modules/@evalops/contracts/dist/advisor-effort.d.ts +8 -0
- package/node_modules/@evalops/contracts/dist/advisor-effort.d.ts.map +1 -0
- package/node_modules/@evalops/contracts/dist/advisor-effort.js +35 -0
- package/node_modules/@evalops/contracts/dist/advisor-effort.js.map +1 -0
- package/node_modules/@evalops/contracts/dist/delegation-prompt.d.ts +10 -0
- package/node_modules/@evalops/contracts/dist/delegation-prompt.d.ts.map +1 -0
- package/node_modules/@evalops/contracts/dist/delegation-prompt.js +48 -0
- package/node_modules/@evalops/contracts/dist/delegation-prompt.js.map +1 -0
- package/node_modules/@evalops/contracts/dist/guarded-files-settings.d.ts +20 -0
- package/node_modules/@evalops/contracts/dist/guarded-files-settings.d.ts.map +1 -0
- package/node_modules/@evalops/contracts/dist/guarded-files-settings.js +147 -0
- package/node_modules/@evalops/contracts/dist/guarded-files-settings.js.map +1 -0
- package/node_modules/@evalops/contracts/dist/headless-protocol-generated.d.ts +2 -0
- package/node_modules/@evalops/contracts/dist/headless-protocol-generated.d.ts.map +1 -1
- package/node_modules/@evalops/contracts/dist/headless-protocol-generated.js +1 -0
- package/node_modules/@evalops/contracts/dist/headless-protocol-generated.js.map +1 -1
- package/node_modules/@evalops/contracts/dist/headless-protocol-schemas.generated.d.ts +280 -5
- package/node_modules/@evalops/contracts/dist/headless-protocol-schemas.generated.d.ts.map +1 -1
- package/node_modules/@evalops/contracts/dist/headless-protocol-schemas.generated.js +36 -1
- package/node_modules/@evalops/contracts/dist/headless-protocol-schemas.generated.js.map +1 -1
- package/node_modules/@evalops/contracts/dist/index.d.ts +15 -1
- package/node_modules/@evalops/contracts/dist/index.d.ts.map +1 -1
- package/node_modules/@evalops/contracts/dist/index.js +6 -0
- package/node_modules/@evalops/contracts/dist/index.js.map +1 -1
- package/node_modules/@evalops/contracts/dist/maestro-app-server.d.ts +593 -0
- package/node_modules/@evalops/contracts/dist/maestro-app-server.d.ts.map +1 -0
- package/node_modules/@evalops/contracts/dist/maestro-app-server.js +218 -0
- package/node_modules/@evalops/contracts/dist/maestro-app-server.js.map +1 -0
- package/node_modules/@evalops/contracts/dist/proto/maestro/v1/headless_pb.d.ts +91 -0
- package/node_modules/@evalops/contracts/dist/proto/maestro/v1/headless_pb.d.ts.map +1 -1
- package/node_modules/@evalops/contracts/dist/proto/maestro/v1/headless_pb.js +92 -61
- package/node_modules/@evalops/contracts/dist/proto/maestro/v1/headless_pb.js.map +1 -1
- package/node_modules/@evalops/contracts/dist/runtime-app-server.d.ts +3 -0
- package/node_modules/@evalops/contracts/dist/runtime-app-server.d.ts.map +1 -1
- package/node_modules/@evalops/contracts/dist/runtime-constraints.d.ts +28 -0
- package/node_modules/@evalops/contracts/dist/runtime-constraints.d.ts.map +1 -0
- package/node_modules/@evalops/contracts/dist/runtime-constraints.js +71 -0
- package/node_modules/@evalops/contracts/dist/runtime-constraints.js.map +1 -0
- package/node_modules/@evalops/contracts/dist/runtime-server-request.d.ts +7 -0
- package/node_modules/@evalops/contracts/dist/runtime-server-request.d.ts.map +1 -1
- package/node_modules/@evalops/contracts/dist/runtime-server-request.js +2 -0
- package/node_modules/@evalops/contracts/dist/runtime-server-request.js.map +1 -1
- package/node_modules/@evalops/contracts/dist/scenario.d.ts +191 -0
- package/node_modules/@evalops/contracts/dist/scenario.d.ts.map +1 -0
- package/node_modules/@evalops/contracts/dist/scenario.js +4 -0
- package/node_modules/@evalops/contracts/dist/scenario.js.map +1 -0
- package/node_modules/@evalops/contracts/dist/schemas.d.ts +17 -3
- package/node_modules/@evalops/contracts/dist/schemas.d.ts.map +1 -1
- package/node_modules/@evalops/contracts/dist/schemas.js +16 -0
- package/node_modules/@evalops/contracts/dist/schemas.js.map +1 -1
- package/node_modules/@evalops/contracts/package.json +1 -1
- package/node_modules/@evalops/tui/package.json +1 -1
- package/package.json +41 -7
- package/skills/incident-triage/SKILL.md +40 -0
- package/skills/incident-triage/mcp.json +14 -0
- package/skills/incident-triage/reference/triage.md +20 -0
- package/skills/incident-triage/toolbox/incident-timeline +22 -0
- package/skills/incident-triage/toolbox/incident-timeline.cmd +16 -0
- package/skills/pr-review/SKILL.md +40 -0
- package/skills/pr-review/mcp.json +14 -0
- package/skills/pr-review/reference/rubric.md +20 -0
- package/skills/pr-review/toolbox/review-summary +20 -0
- package/skills/pr-review/toolbox/review-summary.cmd +14 -0
- package/skills/release-verification/SKILL.md +40 -0
- package/skills/release-verification/mcp.json +15 -0
- package/skills/release-verification/reference/checklist.md +21 -0
- package/skills/release-verification/toolbox/release-readiness +18 -0
- package/skills/release-verification/toolbox/release-readiness.cmd +12 -0
|
@@ -0,0 +1,822 @@
|
|
|
1
|
+
import { readFileSync } from "node:fs";
|
|
2
|
+
import { resolve } from "node:path";
|
|
3
|
+
import { MAESTRO_SCENARIO_SCHEMA, MAESTRO_SCENARIO_WORKSPACE_MANIFEST_SCHEMA, } from "@evalops/contracts";
|
|
4
|
+
import { escapeXml } from "./junit-xml.js";
|
|
5
|
+
export const AGENT_TRAJECTORY_SCENARIO_RESULT_SCHEMA = "evalops.maestro.agent-trajectory-scenario-result.v1";
|
|
6
|
+
export { MAESTRO_SCENARIO_WORKSPACE_MANIFEST_SCHEMA, MAESTRO_SCENARIO_SCHEMA, } from "@evalops/contracts";
|
|
7
|
+
const AGENT_TRAJECTORY_SCENARIO_ASSERTION_KINDS = new Set([
|
|
8
|
+
"event.exists",
|
|
9
|
+
"event.forbidden",
|
|
10
|
+
"replay.deltas",
|
|
11
|
+
"score.finding",
|
|
12
|
+
"inspection.redaction",
|
|
13
|
+
"workspace.manifest",
|
|
14
|
+
"efficiency.budget",
|
|
15
|
+
"provenance.chain",
|
|
16
|
+
"human.review",
|
|
17
|
+
"external.refs",
|
|
18
|
+
"trajectory.diff",
|
|
19
|
+
]);
|
|
20
|
+
const EXTERNAL_REF_FIELDS = [
|
|
21
|
+
"ensembleTranscriptIds",
|
|
22
|
+
"platformTraceIds",
|
|
23
|
+
"platformWorkEnvelopeIds",
|
|
24
|
+
"slackThreadRefs",
|
|
25
|
+
"evidenceArtifactIds",
|
|
26
|
+
];
|
|
27
|
+
const RELEASE_GATE_TIERS = [
|
|
28
|
+
"smoke",
|
|
29
|
+
"regression",
|
|
30
|
+
"gauntlet",
|
|
31
|
+
];
|
|
32
|
+
const REQUIRED_ARTIFACTS = [
|
|
33
|
+
"trajectory",
|
|
34
|
+
"replay",
|
|
35
|
+
"score",
|
|
36
|
+
"inspection",
|
|
37
|
+
"workspace_manifest",
|
|
38
|
+
];
|
|
39
|
+
const HYDRATION_MODES = [
|
|
40
|
+
"manifest_only",
|
|
41
|
+
"fixture_workspace",
|
|
42
|
+
"frozen_archive",
|
|
43
|
+
];
|
|
44
|
+
const WORKSPACE_SOURCES = [
|
|
45
|
+
"production",
|
|
46
|
+
"canary",
|
|
47
|
+
"fixture",
|
|
48
|
+
"synthetic",
|
|
49
|
+
];
|
|
50
|
+
const TOOL_ADAPTER_MODES = [
|
|
51
|
+
"recorded",
|
|
52
|
+
"mocked",
|
|
53
|
+
"sandboxed",
|
|
54
|
+
"disabled",
|
|
55
|
+
];
|
|
56
|
+
function isObject(value) {
|
|
57
|
+
return typeof value === "object" && value !== null && !Array.isArray(value);
|
|
58
|
+
}
|
|
59
|
+
function readJson(path) {
|
|
60
|
+
return JSON.parse(readFileSync(path, "utf8"));
|
|
61
|
+
}
|
|
62
|
+
function requireString(value, name) {
|
|
63
|
+
if (typeof value !== "string" || value.length === 0) {
|
|
64
|
+
throw new Error(`${name} must be a non-empty string`);
|
|
65
|
+
}
|
|
66
|
+
return value;
|
|
67
|
+
}
|
|
68
|
+
function requireOptionalString(value, name) {
|
|
69
|
+
if (value !== undefined) {
|
|
70
|
+
requireString(value, name);
|
|
71
|
+
}
|
|
72
|
+
}
|
|
73
|
+
function requireStringArray(value, name) {
|
|
74
|
+
if (!Array.isArray(value) ||
|
|
75
|
+
value.some((item) => typeof item !== "string" || item.length === 0)) {
|
|
76
|
+
throw new Error(`${name} must contain non-empty strings`);
|
|
77
|
+
}
|
|
78
|
+
}
|
|
79
|
+
function requireOptionalStringArray(value, name) {
|
|
80
|
+
if (value !== undefined) {
|
|
81
|
+
requireStringArray(value, name);
|
|
82
|
+
}
|
|
83
|
+
}
|
|
84
|
+
function requireOptionalNonNegativeInteger(value, name) {
|
|
85
|
+
if (value !== undefined &&
|
|
86
|
+
(!Number.isInteger(value) || value < 0)) {
|
|
87
|
+
throw new Error(`${name} must be a non-negative integer`);
|
|
88
|
+
}
|
|
89
|
+
}
|
|
90
|
+
function loadTypedJson(path, schemaVersion, name) {
|
|
91
|
+
const value = readJson(path);
|
|
92
|
+
if (!isObject(value) ||
|
|
93
|
+
(value.schemaVersion !== schemaVersion &&
|
|
94
|
+
value.trajectorySchemaVersion !== schemaVersion)) {
|
|
95
|
+
throw new Error(`${name} at ${path} must use schemaVersion ${schemaVersion}`);
|
|
96
|
+
}
|
|
97
|
+
return value;
|
|
98
|
+
}
|
|
99
|
+
function validateWorkspaceManifest(manifest, label) {
|
|
100
|
+
requireString(manifest.id, `${label}.id`);
|
|
101
|
+
requireString(manifest.recordedAt, `${label}.recordedAt`);
|
|
102
|
+
if (!WORKSPACE_SOURCES.includes(manifest.source)) {
|
|
103
|
+
throw new Error(`${label}.source must be one of: ${WORKSPACE_SOURCES.join(", ")}`);
|
|
104
|
+
}
|
|
105
|
+
requireOptionalString(manifest.workspaceRoot, `${label}.workspaceRoot`);
|
|
106
|
+
if (!isObject(manifest.hydration)) {
|
|
107
|
+
throw new Error(`${label}.hydration must be an object`);
|
|
108
|
+
}
|
|
109
|
+
if (!HYDRATION_MODES.includes(manifest.hydration.mode)) {
|
|
110
|
+
throw new Error(`${label}.hydration.mode must be one of: ${HYDRATION_MODES.join(", ")}`);
|
|
111
|
+
}
|
|
112
|
+
requireOptionalString(manifest.hydration.archiveUri, `${label}.hydration.archiveUri`);
|
|
113
|
+
requireOptionalString(manifest.hydration.rootPath, `${label}.hydration.rootPath`);
|
|
114
|
+
if (!Array.isArray(manifest.files)) {
|
|
115
|
+
throw new Error(`${label}.files must be an array`);
|
|
116
|
+
}
|
|
117
|
+
for (const [index, file] of manifest.files.entries()) {
|
|
118
|
+
const fileLabel = `${label}.files[${index}]`;
|
|
119
|
+
if (!isObject(file)) {
|
|
120
|
+
throw new Error(`${fileLabel} must be an object`);
|
|
121
|
+
}
|
|
122
|
+
requireString(file.path, `${fileLabel}.path`);
|
|
123
|
+
requireOptionalString(file.sha256, `${fileLabel}.sha256`);
|
|
124
|
+
requireOptionalNonNegativeInteger(file.sizeBytes, `${fileLabel}.sizeBytes`);
|
|
125
|
+
requireOptionalString(file.purpose, `${fileLabel}.purpose`);
|
|
126
|
+
}
|
|
127
|
+
if (!Array.isArray(manifest.toolAdapters)) {
|
|
128
|
+
throw new Error(`${label}.toolAdapters must be an array`);
|
|
129
|
+
}
|
|
130
|
+
for (const [index, adapter] of manifest.toolAdapters.entries()) {
|
|
131
|
+
const adapterLabel = `${label}.toolAdapters[${index}]`;
|
|
132
|
+
if (!isObject(adapter)) {
|
|
133
|
+
throw new Error(`${adapterLabel} must be an object`);
|
|
134
|
+
}
|
|
135
|
+
requireString(adapter.tool, `${adapterLabel}.tool`);
|
|
136
|
+
if (!TOOL_ADAPTER_MODES.includes(adapter.mode)) {
|
|
137
|
+
throw new Error(`${adapterLabel}.mode must be one of: ${TOOL_ADAPTER_MODES.join(", ")}`);
|
|
138
|
+
}
|
|
139
|
+
requireOptionalString(adapter.fixturePath, `${adapterLabel}.fixturePath`);
|
|
140
|
+
requireOptionalString(adapter.rationale, `${adapterLabel}.rationale`);
|
|
141
|
+
}
|
|
142
|
+
if (!isObject(manifest.redaction)) {
|
|
143
|
+
throw new Error(`${label}.redaction must be an object`);
|
|
144
|
+
}
|
|
145
|
+
if (typeof manifest.redaction.secretsRemoved !== "boolean") {
|
|
146
|
+
throw new Error(`${label}.redaction.secretsRemoved must be a boolean`);
|
|
147
|
+
}
|
|
148
|
+
if (typeof manifest.redaction.rawPromptsIncluded !== "boolean") {
|
|
149
|
+
throw new Error(`${label}.redaction.rawPromptsIncluded must be a boolean`);
|
|
150
|
+
}
|
|
151
|
+
requireOptionalStringArray(manifest.redaction.notes, `${label}.redaction.notes`);
|
|
152
|
+
}
|
|
153
|
+
export function parseAgentTrajectoryScenario(value, label) {
|
|
154
|
+
if (!isObject(value)) {
|
|
155
|
+
throw new Error(`Scenario ${label} must be a JSON object`);
|
|
156
|
+
}
|
|
157
|
+
const schemaVersion = value.schemaVersion;
|
|
158
|
+
if (schemaVersion !== MAESTRO_SCENARIO_SCHEMA) {
|
|
159
|
+
throw new Error(`Scenario ${label} must use schemaVersion ${MAESTRO_SCENARIO_SCHEMA}`);
|
|
160
|
+
}
|
|
161
|
+
const scenario = value;
|
|
162
|
+
validateAgentTrajectoryScenario(scenario, label);
|
|
163
|
+
return scenario;
|
|
164
|
+
}
|
|
165
|
+
export function loadAgentTrajectoryScenario(path) {
|
|
166
|
+
return parseAgentTrajectoryScenario(readJson(path), path);
|
|
167
|
+
}
|
|
168
|
+
export function loadAgentTrajectoryScenarioInputs(scenario, baseDir) {
|
|
169
|
+
const resolvePath = (path) => resolve(baseDir, path);
|
|
170
|
+
const source = scenario.source;
|
|
171
|
+
const inputs = {
|
|
172
|
+
trajectory: loadTypedJson(resolvePath(source.trajectoryPath), "evalops.maestro.agent-trajectory.v1", "trajectory"),
|
|
173
|
+
replay: loadTypedJson(resolvePath(source.replayPath), "evalops.maestro.agent-trajectory-replay.v1", "trajectory replay"),
|
|
174
|
+
score: loadTypedJson(resolvePath(source.scorePath), "evalops.maestro.agent-trajectory-score.v1", "trajectory score"),
|
|
175
|
+
};
|
|
176
|
+
if (source.inspectionPath) {
|
|
177
|
+
inputs.inspection = loadTypedJson(resolvePath(source.inspectionPath), "evalops.maestro.agent-trajectory-inspection.v1", "trajectory inspection");
|
|
178
|
+
}
|
|
179
|
+
if (source.workspaceManifestPath) {
|
|
180
|
+
const manifestPath = resolvePath(source.workspaceManifestPath);
|
|
181
|
+
const workspaceManifest = loadTypedJson(manifestPath, MAESTRO_SCENARIO_WORKSPACE_MANIFEST_SCHEMA, "workspace manifest");
|
|
182
|
+
validateWorkspaceManifest(workspaceManifest, `workspace manifest at ${manifestPath}`);
|
|
183
|
+
inputs.workspaceManifest = workspaceManifest;
|
|
184
|
+
}
|
|
185
|
+
if (source.baselineTrajectoryPath) {
|
|
186
|
+
inputs.baselineTrajectory = loadTypedJson(resolvePath(source.baselineTrajectoryPath), "evalops.maestro.agent-trajectory.v1", "baseline trajectory");
|
|
187
|
+
}
|
|
188
|
+
if (source.candidateTrajectoryPath) {
|
|
189
|
+
inputs.candidateTrajectory = loadTypedJson(resolvePath(source.candidateTrajectoryPath), "evalops.maestro.agent-trajectory.v1", "candidate trajectory");
|
|
190
|
+
}
|
|
191
|
+
if (source.baselineScorePath) {
|
|
192
|
+
inputs.baselineScore = loadTypedJson(resolvePath(source.baselineScorePath), "evalops.maestro.agent-trajectory-score.v1", "baseline score");
|
|
193
|
+
}
|
|
194
|
+
if (source.candidateScorePath) {
|
|
195
|
+
inputs.candidateScore = loadTypedJson(resolvePath(source.candidateScorePath), "evalops.maestro.agent-trajectory-score.v1", "candidate score");
|
|
196
|
+
}
|
|
197
|
+
return inputs;
|
|
198
|
+
}
|
|
199
|
+
export function validateAgentTrajectoryScenario(scenario, label = scenario.id) {
|
|
200
|
+
requireString(scenario.id, `${label}.id`);
|
|
201
|
+
requireString(scenario.title, `${label}.title`);
|
|
202
|
+
requireString(scenario.description, `${label}.description`);
|
|
203
|
+
if (scenario.expectedOutcome !== undefined &&
|
|
204
|
+
scenario.expectedOutcome !== "pass" &&
|
|
205
|
+
scenario.expectedOutcome !== "fail") {
|
|
206
|
+
throw new Error(`${label}.expectedOutcome must be pass or fail`);
|
|
207
|
+
}
|
|
208
|
+
if (!isObject(scenario.source)) {
|
|
209
|
+
throw new Error(`${label}.source must be an object`);
|
|
210
|
+
}
|
|
211
|
+
requireString(scenario.source.trajectoryPath, `${label}.source.trajectoryPath`);
|
|
212
|
+
requireString(scenario.source.replayPath, `${label}.source.replayPath`);
|
|
213
|
+
requireString(scenario.source.scorePath, `${label}.source.scorePath`);
|
|
214
|
+
if (scenario.source.workspaceManifestPath !== undefined) {
|
|
215
|
+
requireString(scenario.source.workspaceManifestPath, `${label}.source.workspaceManifestPath`);
|
|
216
|
+
}
|
|
217
|
+
if (Boolean(scenario.source.baselineTrajectoryPath) !==
|
|
218
|
+
Boolean(scenario.source.candidateTrajectoryPath)) {
|
|
219
|
+
throw new Error(`${label}.source baselineTrajectoryPath and candidateTrajectoryPath must be provided together`);
|
|
220
|
+
}
|
|
221
|
+
if (Boolean(scenario.source.baselineScorePath) !==
|
|
222
|
+
Boolean(scenario.source.candidateScorePath)) {
|
|
223
|
+
throw new Error(`${label}.source baselineScorePath and candidateScorePath must be provided together`);
|
|
224
|
+
}
|
|
225
|
+
if (!Array.isArray(scenario.assertions) || scenario.assertions.length === 0) {
|
|
226
|
+
throw new Error(`${label}.assertions must contain at least one assertion`);
|
|
227
|
+
}
|
|
228
|
+
if (!Array.isArray(scenario.reviewLabels)) {
|
|
229
|
+
throw new Error(`${label}.reviewLabels must be an array`);
|
|
230
|
+
}
|
|
231
|
+
if (scenario.releaseGate !== undefined) {
|
|
232
|
+
if (!isObject(scenario.releaseGate)) {
|
|
233
|
+
throw new Error(`${label}.releaseGate must be an object`);
|
|
234
|
+
}
|
|
235
|
+
if (typeof scenario.releaseGate.releaseBlocking !== "boolean") {
|
|
236
|
+
throw new Error(`${label}.releaseGate.releaseBlocking must be a boolean`);
|
|
237
|
+
}
|
|
238
|
+
if (!RELEASE_GATE_TIERS.includes(scenario.releaseGate.tier)) {
|
|
239
|
+
throw new Error(`${label}.releaseGate.tier must be one of: ${RELEASE_GATE_TIERS.join(", ")}`);
|
|
240
|
+
}
|
|
241
|
+
if (!Array.isArray(scenario.releaseGate.requiredArtifacts) ||
|
|
242
|
+
scenario.releaseGate.requiredArtifacts.length === 0) {
|
|
243
|
+
throw new Error(`${label}.releaseGate.requiredArtifacts must not be empty`);
|
|
244
|
+
}
|
|
245
|
+
const unknownArtifacts = scenario.releaseGate.requiredArtifacts.filter((artifact) => !REQUIRED_ARTIFACTS.includes(artifact));
|
|
246
|
+
if (unknownArtifacts.length > 0) {
|
|
247
|
+
throw new Error(`${label}.releaseGate.requiredArtifacts contains unknown artifact(s): ${unknownArtifacts.join(", ")}`);
|
|
248
|
+
}
|
|
249
|
+
if (scenario.releaseGate.releaseBlocking &&
|
|
250
|
+
!scenario.releaseGate.requiredArtifacts.includes("workspace_manifest")) {
|
|
251
|
+
throw new Error(`${label}.releaseGate release-blocking scenarios must require workspace_manifest`);
|
|
252
|
+
}
|
|
253
|
+
if (scenario.releaseGate.requiredArtifacts.includes("inspection") &&
|
|
254
|
+
!scenario.source.inspectionPath) {
|
|
255
|
+
throw new Error(`${label}.releaseGate requires inspection but source.inspectionPath is missing`);
|
|
256
|
+
}
|
|
257
|
+
if (scenario.releaseGate.requiredArtifacts.includes("workspace_manifest") &&
|
|
258
|
+
!scenario.source.workspaceManifestPath) {
|
|
259
|
+
throw new Error(`${label}.releaseGate requires workspace_manifest but source.workspaceManifestPath is missing`);
|
|
260
|
+
}
|
|
261
|
+
requireOptionalNonNegativeInteger(scenario.releaseGate.maxEvents, `${label}.releaseGate.maxEvents`);
|
|
262
|
+
requireOptionalNonNegativeInteger(scenario.releaseGate.maxToolCalls, `${label}.releaseGate.maxToolCalls`);
|
|
263
|
+
requireOptionalNonNegativeInteger(scenario.releaseGate.maxReplayDeltas, `${label}.releaseGate.maxReplayDeltas`);
|
|
264
|
+
requireOptionalNonNegativeInteger(scenario.releaseGate.maxScoreFailures, `${label}.releaseGate.maxScoreFailures`);
|
|
265
|
+
requireOptionalNonNegativeInteger(scenario.releaseGate.maxScoreWarnings, `${label}.releaseGate.maxScoreWarnings`);
|
|
266
|
+
}
|
|
267
|
+
if (!isObject(scenario.platform)) {
|
|
268
|
+
throw new Error(`${label}.platform must be an object`);
|
|
269
|
+
}
|
|
270
|
+
if (!Array.isArray(scenario.platform.traceJoinKeys) ||
|
|
271
|
+
scenario.platform.traceJoinKeys.length === 0) {
|
|
272
|
+
throw new Error(`${label}.platform.traceJoinKeys must not be empty`);
|
|
273
|
+
}
|
|
274
|
+
if (scenario.externalRefs !== undefined) {
|
|
275
|
+
if (!isObject(scenario.externalRefs)) {
|
|
276
|
+
throw new Error(`${label}.externalRefs must be an object`);
|
|
277
|
+
}
|
|
278
|
+
let refs = 0;
|
|
279
|
+
for (const field of EXTERNAL_REF_FIELDS) {
|
|
280
|
+
const values = scenario.externalRefs[field];
|
|
281
|
+
if (values === undefined)
|
|
282
|
+
continue;
|
|
283
|
+
if (!Array.isArray(values) ||
|
|
284
|
+
values.some((value) => typeof value !== "string" || value.length === 0)) {
|
|
285
|
+
throw new Error(`${label}.externalRefs.${field} must contain non-empty strings`);
|
|
286
|
+
}
|
|
287
|
+
refs += values.length;
|
|
288
|
+
}
|
|
289
|
+
if (refs === 0) {
|
|
290
|
+
throw new Error(`${label}.externalRefs must contain at least one ref`);
|
|
291
|
+
}
|
|
292
|
+
}
|
|
293
|
+
if (!isObject(scenario.assumptions)) {
|
|
294
|
+
throw new Error(`${label}.assumptions must be an object`);
|
|
295
|
+
}
|
|
296
|
+
requireString(scenario.assumptions.workflow, `${label}.assumptions.workflow`);
|
|
297
|
+
requireString(scenario.assumptions.correctnessModel, `${label}.assumptions.correctnessModel`);
|
|
298
|
+
requireString(scenario.assumptions.threatModel, `${label}.assumptions.threatModel`);
|
|
299
|
+
if (!Array.isArray(scenario.assumptions.researchBasis) ||
|
|
300
|
+
scenario.assumptions.researchBasis.length === 0) {
|
|
301
|
+
throw new Error(`${label}.assumptions.researchBasis must not be empty`);
|
|
302
|
+
}
|
|
303
|
+
let hasWorkspaceManifestAssertion = false;
|
|
304
|
+
let hasWarningWorkspaceManifestAssertion = false;
|
|
305
|
+
for (const assertion of scenario.assertions) {
|
|
306
|
+
requireString(assertion.id, `${label}.assertions[].id`);
|
|
307
|
+
const kind = requireString(assertion.kind, `${label}.assertions[].kind`);
|
|
308
|
+
if (!AGENT_TRAJECTORY_SCENARIO_ASSERTION_KINDS.has(kind)) {
|
|
309
|
+
throw new Error(`${label}.assertions[].kind must be one of: ${Array.from(AGENT_TRAJECTORY_SCENARIO_ASSERTION_KINDS).join(", ")}`);
|
|
310
|
+
}
|
|
311
|
+
if (kind === "trajectory.diff" &&
|
|
312
|
+
assertion.maxAddedScoreFailures !== undefined &&
|
|
313
|
+
(!scenario.source.baselineScorePath ||
|
|
314
|
+
!scenario.source.candidateScorePath)) {
|
|
315
|
+
throw new Error(`${label}.assertions[].maxAddedScoreFailures requires baselineScorePath and candidateScorePath`);
|
|
316
|
+
}
|
|
317
|
+
if (kind === "workspace.manifest") {
|
|
318
|
+
hasWorkspaceManifestAssertion = true;
|
|
319
|
+
hasWarningWorkspaceManifestAssertion =
|
|
320
|
+
hasWarningWorkspaceManifestAssertion ||
|
|
321
|
+
assertion.severity === "warning";
|
|
322
|
+
if (!scenario.source.workspaceManifestPath) {
|
|
323
|
+
throw new Error(`${label}.assertions[].kind workspace.manifest requires source.workspaceManifestPath`);
|
|
324
|
+
}
|
|
325
|
+
requireOptionalStringArray(assertion.requiredWorkspaceFiles, `${label}.assertions[].requiredWorkspaceFiles`);
|
|
326
|
+
requireOptionalStringArray(assertion.requiredToolAdapters, `${label}.assertions[].requiredToolAdapters`);
|
|
327
|
+
if (assertion.requiredHydrationModes !== undefined &&
|
|
328
|
+
(!Array.isArray(assertion.requiredHydrationModes) ||
|
|
329
|
+
assertion.requiredHydrationModes.some((mode) => !HYDRATION_MODES.includes(mode)))) {
|
|
330
|
+
throw new Error(`${label}.assertions[].requiredHydrationModes must contain known hydration modes`);
|
|
331
|
+
}
|
|
332
|
+
if (assertion.requiredReleaseGateTier !== undefined &&
|
|
333
|
+
!RELEASE_GATE_TIERS.includes(assertion.requiredReleaseGateTier)) {
|
|
334
|
+
throw new Error(`${label}.assertions[].requiredReleaseGateTier must be one of: ${RELEASE_GATE_TIERS.join(", ")}`);
|
|
335
|
+
}
|
|
336
|
+
requireOptionalNonNegativeInteger(assertion.minWorkspaceFiles, `${label}.assertions[].minWorkspaceFiles`);
|
|
337
|
+
requireOptionalNonNegativeInteger(assertion.minToolAdapters, `${label}.assertions[].minToolAdapters`);
|
|
338
|
+
}
|
|
339
|
+
if (kind === "external.refs" &&
|
|
340
|
+
(!Array.isArray(assertion.requiredExternalRefKinds) ||
|
|
341
|
+
assertion.requiredExternalRefKinds.length === 0)) {
|
|
342
|
+
throw new Error(`${label}.assertions[].requiredExternalRefKinds must not be empty for external.refs`);
|
|
343
|
+
}
|
|
344
|
+
if (kind === "external.refs") {
|
|
345
|
+
const unknownKinds = (assertion.requiredExternalRefKinds ?? []).filter((refKind) => !EXTERNAL_REF_FIELDS.includes(refKind));
|
|
346
|
+
if (unknownKinds.length > 0) {
|
|
347
|
+
throw new Error(`${label}.assertions[].requiredExternalRefKinds contains unknown external ref kind(s): ${unknownKinds.join(", ")}`);
|
|
348
|
+
}
|
|
349
|
+
if (assertion.requiredExternalRefs !== undefined &&
|
|
350
|
+
(!Array.isArray(assertion.requiredExternalRefs) ||
|
|
351
|
+
assertion.requiredExternalRefs.some((ref) => typeof ref !== "string" || ref.length === 0))) {
|
|
352
|
+
throw new Error(`${label}.assertions[].requiredExternalRefs must contain non-empty strings for external.refs`);
|
|
353
|
+
}
|
|
354
|
+
}
|
|
355
|
+
}
|
|
356
|
+
if (scenario.releaseGate?.releaseBlocking === true &&
|
|
357
|
+
scenario.releaseGate.requiredArtifacts.includes("workspace_manifest") &&
|
|
358
|
+
!hasWorkspaceManifestAssertion) {
|
|
359
|
+
throw new Error(`${label}.releaseGate release-blocking workspace_manifest gates must include a workspace.manifest assertion`);
|
|
360
|
+
}
|
|
361
|
+
if (scenario.releaseGate?.releaseBlocking === true &&
|
|
362
|
+
scenario.releaseGate.requiredArtifacts.includes("workspace_manifest") &&
|
|
363
|
+
hasWarningWorkspaceManifestAssertion) {
|
|
364
|
+
throw new Error(`${label}.releaseGate release-blocking workspace_manifest assertions must use error severity`);
|
|
365
|
+
}
|
|
366
|
+
}
|
|
367
|
+
function eventMatches(event, selector) {
|
|
368
|
+
return ((selector.kind === undefined || event.kind === selector.kind) &&
|
|
369
|
+
(selector.phase === undefined || event.phase === selector.phase) &&
|
|
370
|
+
(selector.type === undefined || event.type === selector.type) &&
|
|
371
|
+
(selector.status === undefined || event.status === selector.status) &&
|
|
372
|
+
(selector.toolName === undefined || event.toolName === selector.toolName) &&
|
|
373
|
+
(selector.source === undefined || event.source === selector.source) &&
|
|
374
|
+
(selector.actor === undefined || event.actor === selector.actor));
|
|
375
|
+
}
|
|
376
|
+
function evidenceFromEvents(events) {
|
|
377
|
+
const seen = new Set();
|
|
378
|
+
const evidence = [];
|
|
379
|
+
for (const event of events) {
|
|
380
|
+
const eventKey = `trajectory:event:${event.id}`;
|
|
381
|
+
if (!seen.has(eventKey)) {
|
|
382
|
+
seen.add(eventKey);
|
|
383
|
+
evidence.push({
|
|
384
|
+
kind: "trajectory_event",
|
|
385
|
+
id: event.id,
|
|
386
|
+
source: "trajectory",
|
|
387
|
+
label: `${event.type}:${event.id}`,
|
|
388
|
+
});
|
|
389
|
+
}
|
|
390
|
+
for (const anchor of event.evidence) {
|
|
391
|
+
const key = `trajectory:${anchor.kind}:${anchor.id}`;
|
|
392
|
+
if (seen.has(key))
|
|
393
|
+
continue;
|
|
394
|
+
seen.add(key);
|
|
395
|
+
evidence.push({
|
|
396
|
+
kind: anchor.kind,
|
|
397
|
+
id: anchor.id,
|
|
398
|
+
source: "trajectory",
|
|
399
|
+
label: `${anchor.kind}:${anchor.id}`,
|
|
400
|
+
});
|
|
401
|
+
}
|
|
402
|
+
}
|
|
403
|
+
return evidence.sort((left, right) => left.label.localeCompare(right.label));
|
|
404
|
+
}
|
|
405
|
+
function scenarioEvidence(id, label) {
|
|
406
|
+
return [
|
|
407
|
+
{
|
|
408
|
+
kind: "scenario",
|
|
409
|
+
id,
|
|
410
|
+
source: "scenario",
|
|
411
|
+
label,
|
|
412
|
+
},
|
|
413
|
+
];
|
|
414
|
+
}
|
|
415
|
+
function result(assertion, status, message, evidence = []) {
|
|
416
|
+
return {
|
|
417
|
+
id: assertion.id,
|
|
418
|
+
kind: assertion.kind,
|
|
419
|
+
status,
|
|
420
|
+
severity: assertion.severity ?? "error",
|
|
421
|
+
message,
|
|
422
|
+
evidence,
|
|
423
|
+
};
|
|
424
|
+
}
|
|
425
|
+
function fail(assertion, message, evidence = []) {
|
|
426
|
+
const status = assertion.severity === "warning" ? "warn" : "fail";
|
|
427
|
+
return result(assertion, status, message, evidence);
|
|
428
|
+
}
|
|
429
|
+
function countToolCalls(report) {
|
|
430
|
+
return report.events.filter((event) => event.type === "tool.requested")
|
|
431
|
+
.length;
|
|
432
|
+
}
|
|
433
|
+
function scoreFailures(report) {
|
|
434
|
+
return report.findings.filter((finding) => finding.status === "fail").length;
|
|
435
|
+
}
|
|
436
|
+
function scoreWarnings(report) {
|
|
437
|
+
return report.findings.filter((finding) => finding.status === "warn").length;
|
|
438
|
+
}
|
|
439
|
+
function buildDiff(inputs) {
|
|
440
|
+
if (!inputs.baselineTrajectory || !inputs.candidateTrajectory)
|
|
441
|
+
return undefined;
|
|
442
|
+
return {
|
|
443
|
+
baselineRunId: inputs.baselineTrajectory.run.id,
|
|
444
|
+
candidateRunId: inputs.candidateTrajectory.run.id,
|
|
445
|
+
eventsDelta: inputs.candidateTrajectory.counts.events -
|
|
446
|
+
inputs.baselineTrajectory.counts.events,
|
|
447
|
+
toolCallsDelta: countToolCalls(inputs.candidateTrajectory) -
|
|
448
|
+
countToolCalls(inputs.baselineTrajectory),
|
|
449
|
+
scoreFailuresDelta: inputs.baselineScore && inputs.candidateScore
|
|
450
|
+
? scoreFailures(inputs.candidateScore) -
|
|
451
|
+
scoreFailures(inputs.baselineScore)
|
|
452
|
+
: 0,
|
|
453
|
+
};
|
|
454
|
+
}
|
|
455
|
+
function workspaceEvidence(manifest) {
|
|
456
|
+
return [
|
|
457
|
+
{
|
|
458
|
+
kind: "workspace_manifest",
|
|
459
|
+
id: manifest.id,
|
|
460
|
+
source: "scenario",
|
|
461
|
+
label: `workspace_manifest:${manifest.id}`,
|
|
462
|
+
},
|
|
463
|
+
...manifest.toolAdapters.map((adapter) => ({
|
|
464
|
+
kind: "tool_adapter",
|
|
465
|
+
id: adapter.tool,
|
|
466
|
+
source: "scenario",
|
|
467
|
+
label: `tool_adapter:${adapter.tool}:${adapter.mode}`,
|
|
468
|
+
})),
|
|
469
|
+
];
|
|
470
|
+
}
|
|
471
|
+
function buildWorkspaceSummary(manifest) {
|
|
472
|
+
if (!manifest)
|
|
473
|
+
return undefined;
|
|
474
|
+
return {
|
|
475
|
+
manifestId: manifest.id,
|
|
476
|
+
source: manifest.source,
|
|
477
|
+
recordedAt: manifest.recordedAt,
|
|
478
|
+
hydrationMode: manifest.hydration.mode,
|
|
479
|
+
files: manifest.files.length,
|
|
480
|
+
toolAdapters: manifest.toolAdapters.length,
|
|
481
|
+
};
|
|
482
|
+
}
|
|
483
|
+
function hasRequiredArtifact(artifact, inputs) {
|
|
484
|
+
switch (artifact) {
|
|
485
|
+
case "trajectory":
|
|
486
|
+
return true;
|
|
487
|
+
case "replay":
|
|
488
|
+
return true;
|
|
489
|
+
case "score":
|
|
490
|
+
return true;
|
|
491
|
+
case "inspection":
|
|
492
|
+
return inputs.inspection !== undefined;
|
|
493
|
+
case "workspace_manifest":
|
|
494
|
+
return inputs.workspaceManifest !== undefined;
|
|
495
|
+
}
|
|
496
|
+
}
|
|
497
|
+
function workspaceReleasePolicyViolations(gate, inputs) {
|
|
498
|
+
if (!gate.requiredArtifacts.includes("workspace_manifest")) {
|
|
499
|
+
return [];
|
|
500
|
+
}
|
|
501
|
+
const manifest = inputs.workspaceManifest;
|
|
502
|
+
if (!manifest) {
|
|
503
|
+
return [];
|
|
504
|
+
}
|
|
505
|
+
return [
|
|
506
|
+
manifest.redaction.secretsRemoved !== true
|
|
507
|
+
? "workspace manifest did not confirm secret redaction"
|
|
508
|
+
: undefined,
|
|
509
|
+
manifest.redaction.rawPromptsIncluded !== false
|
|
510
|
+
? "workspace manifest did not confirm raw prompts were excluded"
|
|
511
|
+
: undefined,
|
|
512
|
+
].filter((value) => value !== undefined);
|
|
513
|
+
}
|
|
514
|
+
function buildReleaseGateSummary(scenario, inputs) {
|
|
515
|
+
const gate = scenario.releaseGate;
|
|
516
|
+
if (!gate)
|
|
517
|
+
return undefined;
|
|
518
|
+
const observedToolCalls = countToolCalls(inputs.trajectory);
|
|
519
|
+
const observedFailures = scoreFailures(inputs.score);
|
|
520
|
+
const observedWarnings = scoreWarnings(inputs.score);
|
|
521
|
+
const missingArtifacts = gate.requiredArtifacts.filter((artifact) => !hasRequiredArtifact(artifact, inputs));
|
|
522
|
+
const budgetViolations = [
|
|
523
|
+
gate.maxEvents !== undefined &&
|
|
524
|
+
inputs.trajectory.counts.events > gate.maxEvents
|
|
525
|
+
? `events ${inputs.trajectory.counts.events}/${gate.maxEvents}`
|
|
526
|
+
: undefined,
|
|
527
|
+
gate.maxToolCalls !== undefined && observedToolCalls > gate.maxToolCalls
|
|
528
|
+
? `toolCalls ${observedToolCalls}/${gate.maxToolCalls}`
|
|
529
|
+
: undefined,
|
|
530
|
+
gate.maxReplayDeltas !== undefined &&
|
|
531
|
+
inputs.replay.counts.deltas > gate.maxReplayDeltas
|
|
532
|
+
? `replayDeltas ${inputs.replay.counts.deltas}/${gate.maxReplayDeltas}`
|
|
533
|
+
: undefined,
|
|
534
|
+
gate.maxScoreFailures !== undefined &&
|
|
535
|
+
observedFailures > gate.maxScoreFailures
|
|
536
|
+
? `scoreFailures ${observedFailures}/${gate.maxScoreFailures}`
|
|
537
|
+
: undefined,
|
|
538
|
+
gate.maxScoreWarnings !== undefined &&
|
|
539
|
+
observedWarnings > gate.maxScoreWarnings
|
|
540
|
+
? `scoreWarnings ${observedWarnings}/${gate.maxScoreWarnings}`
|
|
541
|
+
: undefined,
|
|
542
|
+
].filter((value) => value !== undefined);
|
|
543
|
+
const policyViolations = workspaceReleasePolicyViolations(gate, inputs);
|
|
544
|
+
return {
|
|
545
|
+
...gate,
|
|
546
|
+
satisfied: missingArtifacts.length === 0 &&
|
|
547
|
+
budgetViolations.length === 0 &&
|
|
548
|
+
policyViolations.length === 0,
|
|
549
|
+
missingArtifacts,
|
|
550
|
+
budgetViolations,
|
|
551
|
+
policyViolations,
|
|
552
|
+
};
|
|
553
|
+
}
|
|
554
|
+
function evaluateAssertion(assertion, scenario, inputs, diff) {
|
|
555
|
+
switch (assertion.kind) {
|
|
556
|
+
case "event.exists": {
|
|
557
|
+
if (!assertion.selector) {
|
|
558
|
+
return fail(assertion, "event.exists requires a selector.");
|
|
559
|
+
}
|
|
560
|
+
const matches = inputs.trajectory.events.filter((event) => eventMatches(event, assertion.selector ?? {}));
|
|
561
|
+
return matches.length > 0
|
|
562
|
+
? result(assertion, "pass", `Matched ${matches.length} trajectory event(s).`, evidenceFromEvents(matches))
|
|
563
|
+
: fail(assertion, "No trajectory event matched the selector.");
|
|
564
|
+
}
|
|
565
|
+
case "event.forbidden": {
|
|
566
|
+
if (!assertion.selector) {
|
|
567
|
+
return fail(assertion, "event.forbidden requires a selector.");
|
|
568
|
+
}
|
|
569
|
+
const matches = inputs.trajectory.events.filter((event) => eventMatches(event, assertion.selector ?? {}));
|
|
570
|
+
return matches.length === 0
|
|
571
|
+
? result(assertion, "pass", "No forbidden trajectory event matched.")
|
|
572
|
+
: fail(assertion, `Forbidden selector matched ${matches.length} trajectory event(s).`, evidenceFromEvents(matches));
|
|
573
|
+
}
|
|
574
|
+
case "replay.deltas": {
|
|
575
|
+
const maxDeltas = assertion.maxReplayDeltas ?? Number.POSITIVE_INFINITY;
|
|
576
|
+
const maxErrors = assertion.maxReplayErrors ?? Number.POSITIVE_INFINITY;
|
|
577
|
+
const failed = inputs.replay.counts.deltas > maxDeltas ||
|
|
578
|
+
inputs.replay.counts.errors > maxErrors;
|
|
579
|
+
return failed
|
|
580
|
+
? fail(assertion, `Replay deltas exceeded budget: ${inputs.replay.counts.deltas}/${maxDeltas}, errors ${inputs.replay.counts.errors}/${maxErrors}.`, inputs.replay.deltas.map((delta) => ({
|
|
581
|
+
kind: "replay_delta",
|
|
582
|
+
id: delta.id,
|
|
583
|
+
source: "replay",
|
|
584
|
+
label: `${delta.ruleId}:${delta.id}`,
|
|
585
|
+
})))
|
|
586
|
+
: result(assertion, "pass", `Replay stayed within delta and error budgets (${inputs.replay.counts.deltas} deltas, ${inputs.replay.counts.errors} errors).`);
|
|
587
|
+
}
|
|
588
|
+
case "score.finding": {
|
|
589
|
+
if (!assertion.ruleId) {
|
|
590
|
+
return fail(assertion, "score.finding requires ruleId.");
|
|
591
|
+
}
|
|
592
|
+
const finding = inputs.score.findings.find((item) => item.ruleId === assertion.ruleId);
|
|
593
|
+
if (!finding) {
|
|
594
|
+
return fail(assertion, `Missing score finding ${assertion.ruleId}.`);
|
|
595
|
+
}
|
|
596
|
+
if (assertion.status && finding.status !== assertion.status) {
|
|
597
|
+
return fail(assertion, `Score finding ${assertion.ruleId} was ${finding.status}; expected ${assertion.status}.`, evidenceFromEvents(inputs.trajectory.events.filter((event) => finding.eventIds.includes(event.id))));
|
|
598
|
+
}
|
|
599
|
+
return result(assertion, "pass", `Score finding ${assertion.ruleId} matched ${finding.status}.`, evidenceFromEvents(inputs.trajectory.events.filter((event) => finding.eventIds.includes(event.id))));
|
|
600
|
+
}
|
|
601
|
+
case "inspection.redaction": {
|
|
602
|
+
if (!inputs.inspection) {
|
|
603
|
+
return fail(assertion, "inspection.redaction requires inspectionPath.");
|
|
604
|
+
}
|
|
605
|
+
const inspectionJson = JSON.stringify(inputs.inspection);
|
|
606
|
+
const leaked = (assertion.forbiddenTerms ?? []).filter((term) => inspectionJson.includes(term));
|
|
607
|
+
const unredactedItem = inputs.inspection.timelineItems.find((item) => item.redacted !== true);
|
|
608
|
+
if (leaked.length > 0 || unredactedItem) {
|
|
609
|
+
return fail(assertion, `Inspection output was not fail-closed: ${leaked.length} forbidden term(s), unredacted item ${unredactedItem?.id ?? "none"}.`, scenarioEvidence(inputs.inspection.run.id, "inspection:redaction"));
|
|
610
|
+
}
|
|
611
|
+
return result(assertion, "pass", "Inspection artifact stayed redacted and omitted forbidden terms.", scenarioEvidence(inputs.inspection.run.id, "inspection:redaction"));
|
|
612
|
+
}
|
|
613
|
+
case "workspace.manifest": {
|
|
614
|
+
const manifest = inputs.workspaceManifest;
|
|
615
|
+
if (!manifest) {
|
|
616
|
+
return fail(assertion, "workspace.manifest requires source.workspaceManifestPath.");
|
|
617
|
+
}
|
|
618
|
+
const requiredFiles = assertion.requiredWorkspaceFiles ?? [];
|
|
619
|
+
const requiredAdapters = assertion.requiredToolAdapters ?? [];
|
|
620
|
+
const missingFiles = requiredFiles.filter((path) => !manifest.files.some((file) => file.path === path));
|
|
621
|
+
const missingAdapters = requiredAdapters.filter((tool) => !manifest.toolAdapters.some((adapter) => adapter.tool === tool));
|
|
622
|
+
const allowedHydrationModes = assertion.requiredHydrationModes ?? [];
|
|
623
|
+
const hydrationRejected = allowedHydrationModes.length > 0 &&
|
|
624
|
+
!allowedHydrationModes.includes(manifest.hydration.mode);
|
|
625
|
+
const releaseGateRejected = assertion.requiredReleaseGateTier !== undefined &&
|
|
626
|
+
scenario.releaseGate?.tier !== assertion.requiredReleaseGateTier;
|
|
627
|
+
const fileFloor = assertion.minWorkspaceFiles ?? 0;
|
|
628
|
+
const adapterFloor = assertion.minToolAdapters ?? 0;
|
|
629
|
+
const floorFailures = [
|
|
630
|
+
manifest.files.length < fileFloor
|
|
631
|
+
? `workspace files ${manifest.files.length}/${fileFloor}`
|
|
632
|
+
: undefined,
|
|
633
|
+
manifest.toolAdapters.length < adapterFloor
|
|
634
|
+
? `tool adapters ${manifest.toolAdapters.length}/${adapterFloor}`
|
|
635
|
+
: undefined,
|
|
636
|
+
].filter((value) => value !== undefined);
|
|
637
|
+
const failures = [
|
|
638
|
+
...missingFiles.map((path) => `missing file ${path}`),
|
|
639
|
+
...missingAdapters.map((tool) => `missing tool adapter ${tool}`),
|
|
640
|
+
hydrationRejected
|
|
641
|
+
? `hydration mode ${manifest.hydration.mode} not allowed`
|
|
642
|
+
: undefined,
|
|
643
|
+
releaseGateRejected
|
|
644
|
+
? `release tier ${scenario.releaseGate?.tier ?? "none"} did not match ${assertion.requiredReleaseGateTier}`
|
|
645
|
+
: undefined,
|
|
646
|
+
...floorFailures,
|
|
647
|
+
manifest.redaction.secretsRemoved !== true
|
|
648
|
+
? "workspace manifest did not confirm secret redaction"
|
|
649
|
+
: undefined,
|
|
650
|
+
manifest.redaction.rawPromptsIncluded !== false
|
|
651
|
+
? "workspace manifest did not confirm raw prompts were excluded"
|
|
652
|
+
: undefined,
|
|
653
|
+
].filter((value) => value !== undefined);
|
|
654
|
+
return failures.length === 0
|
|
655
|
+
? result(assertion, "pass", `Workspace manifest ${manifest.id} is release-gate ready (${manifest.files.length} file(s), ${manifest.toolAdapters.length} tool adapter(s), ${manifest.hydration.mode}).`, workspaceEvidence(manifest))
|
|
656
|
+
: fail(assertion, `Workspace manifest is not release-gate ready: ${failures.join("; ")}.`, workspaceEvidence(manifest));
|
|
657
|
+
}
|
|
658
|
+
case "efficiency.budget": {
|
|
659
|
+
const maxEvents = assertion.maxEvents ?? Number.POSITIVE_INFINITY;
|
|
660
|
+
const maxToolCalls = assertion.maxToolCalls ?? Number.POSITIVE_INFINITY;
|
|
661
|
+
const maxDeltas = assertion.maxReplayDeltas ?? Number.POSITIVE_INFINITY;
|
|
662
|
+
const maxFailures = assertion.maxScoreFailures ?? Number.POSITIVE_INFINITY;
|
|
663
|
+
const maxWarnings = assertion.maxScoreWarnings ?? Number.POSITIVE_INFINITY;
|
|
664
|
+
const observedToolCalls = countToolCalls(inputs.trajectory);
|
|
665
|
+
const observedFailures = scoreFailures(inputs.score);
|
|
666
|
+
const observedWarnings = scoreWarnings(inputs.score);
|
|
667
|
+
const exceeded = inputs.trajectory.counts.events > maxEvents ||
|
|
668
|
+
observedToolCalls > maxToolCalls ||
|
|
669
|
+
inputs.replay.counts.deltas > maxDeltas ||
|
|
670
|
+
observedFailures > maxFailures ||
|
|
671
|
+
observedWarnings > maxWarnings;
|
|
672
|
+
const message = `Observed events=${inputs.trajectory.counts.events}, toolCalls=${observedToolCalls}, replayDeltas=${inputs.replay.counts.deltas}, scoreFailures=${observedFailures}, scoreWarnings=${observedWarnings}.`;
|
|
673
|
+
return exceeded
|
|
674
|
+
? fail(assertion, `Efficiency budget exceeded. ${message}`)
|
|
675
|
+
: result(assertion, "pass", `Efficiency budget satisfied. ${message}`);
|
|
676
|
+
}
|
|
677
|
+
case "provenance.chain": {
|
|
678
|
+
if (!assertion.eventId) {
|
|
679
|
+
return fail(assertion, "provenance.chain requires eventId.");
|
|
680
|
+
}
|
|
681
|
+
const event = inputs.trajectory.events.find((item) => item.id === assertion.eventId);
|
|
682
|
+
if (!event) {
|
|
683
|
+
return fail(assertion, `Missing provenance event ${assertion.eventId}.`);
|
|
684
|
+
}
|
|
685
|
+
const kinds = new Set(event.evidence.map((anchor) => anchor.kind));
|
|
686
|
+
const missing = (assertion.requiredEvidenceKinds ?? []).filter((kind) => !kinds.has(kind));
|
|
687
|
+
return missing.length === 0
|
|
688
|
+
? result(assertion, "pass", `Event ${assertion.eventId} includes required provenance anchors.`, evidenceFromEvents([event]))
|
|
689
|
+
: fail(assertion, `Event ${assertion.eventId} is missing provenance anchors: ${missing.join(", ")}.`, evidenceFromEvents([event]));
|
|
690
|
+
}
|
|
691
|
+
case "human.review": {
|
|
692
|
+
const missing = (assertion.requiredLabels ?? []).filter((label) => !scenario.reviewLabels.includes(label));
|
|
693
|
+
return missing.length === 0
|
|
694
|
+
? result(assertion, "pass", `Human review labels present: ${(assertion.requiredLabels ?? []).join(", ")}.`, scenarioEvidence(scenario.id, "human-review:labels"))
|
|
695
|
+
: fail(assertion, `Missing human review labels: ${missing.join(", ")}.`, scenarioEvidence(scenario.id, "human-review:labels"));
|
|
696
|
+
}
|
|
697
|
+
case "external.refs": {
|
|
698
|
+
const externalRefs = scenario.externalRefs;
|
|
699
|
+
if (!externalRefs) {
|
|
700
|
+
return fail(assertion, "external.refs requires scenario.externalRefs.");
|
|
701
|
+
}
|
|
702
|
+
const missingKinds = (assertion.requiredExternalRefKinds ?? []).filter((kind) => {
|
|
703
|
+
const values = externalRefs[kind];
|
|
704
|
+
return !Array.isArray(values) || values.length === 0;
|
|
705
|
+
});
|
|
706
|
+
const flattenedRefs = new Set(EXTERNAL_REF_FIELDS.flatMap((field) => externalRefs[field] ?? []));
|
|
707
|
+
const missingRefs = (assertion.requiredExternalRefs ?? []).filter((ref) => !flattenedRefs.has(ref));
|
|
708
|
+
const missing = [...missingKinds, ...missingRefs];
|
|
709
|
+
return missing.length === 0
|
|
710
|
+
? result(assertion, "pass", `External refs present for ${(assertion.requiredExternalRefKinds ?? []).join(", ")}.`, scenarioEvidence(scenario.id, "external-refs"))
|
|
711
|
+
: fail(assertion, `Missing external refs: ${missing.join(", ")}.`, scenarioEvidence(scenario.id, "external-refs"));
|
|
712
|
+
}
|
|
713
|
+
case "trajectory.diff": {
|
|
714
|
+
if (!diff) {
|
|
715
|
+
return fail(assertion, "trajectory.diff requires baselineTrajectoryPath and candidateTrajectoryPath.");
|
|
716
|
+
}
|
|
717
|
+
const exceeded = diff.eventsDelta >
|
|
718
|
+
(assertion.maxAddedEvents ?? Number.POSITIVE_INFINITY) ||
|
|
719
|
+
diff.toolCallsDelta >
|
|
720
|
+
(assertion.maxAddedToolCalls ?? Number.POSITIVE_INFINITY) ||
|
|
721
|
+
diff.scoreFailuresDelta >
|
|
722
|
+
(assertion.maxAddedScoreFailures ?? Number.POSITIVE_INFINITY);
|
|
723
|
+
const message = `Diff eventsDelta=${diff.eventsDelta}, toolCallsDelta=${diff.toolCallsDelta}, scoreFailuresDelta=${diff.scoreFailuresDelta}.`;
|
|
724
|
+
return exceeded
|
|
725
|
+
? fail(assertion, `Trajectory diff budget exceeded. ${message}`)
|
|
726
|
+
: result(assertion, "pass", `Trajectory diff budget satisfied. ${message}`);
|
|
727
|
+
}
|
|
728
|
+
}
|
|
729
|
+
const _exhaustive = assertion.kind;
|
|
730
|
+
return fail(assertion, `Unsupported scenario assertion kind: ${_exhaustive}`);
|
|
731
|
+
}
|
|
732
|
+
function buildProvenance(report) {
|
|
733
|
+
return report.events
|
|
734
|
+
.filter((event) => event.evidence.length > 0)
|
|
735
|
+
.map((event) => ({
|
|
736
|
+
eventId: event.id,
|
|
737
|
+
eventType: event.type,
|
|
738
|
+
phase: event.phase,
|
|
739
|
+
actor: event.actor,
|
|
740
|
+
evidence: evidenceFromEvents([event]),
|
|
741
|
+
}));
|
|
742
|
+
}
|
|
743
|
+
export function evaluateAgentTrajectoryScenario(scenario, inputs) {
|
|
744
|
+
const diff = buildDiff(inputs);
|
|
745
|
+
const releaseGate = buildReleaseGateSummary(scenario, inputs);
|
|
746
|
+
const workspace = buildWorkspaceSummary(inputs.workspaceManifest);
|
|
747
|
+
const assertions = scenario.assertions.map((assertion) => evaluateAssertion(assertion, scenario, inputs, diff));
|
|
748
|
+
const failed = assertions.filter((assertion) => assertion.status === "fail");
|
|
749
|
+
const warnings = assertions.filter((assertion) => assertion.status === "warn");
|
|
750
|
+
const observedOutcome = failed.length > 0 ? "fail" : "pass";
|
|
751
|
+
return {
|
|
752
|
+
schemaVersion: AGENT_TRAJECTORY_SCENARIO_RESULT_SCHEMA,
|
|
753
|
+
scenarioSchemaVersion: MAESTRO_SCENARIO_SCHEMA,
|
|
754
|
+
scenario: {
|
|
755
|
+
id: scenario.id,
|
|
756
|
+
title: scenario.title,
|
|
757
|
+
expectedOutcome: scenario.expectedOutcome ?? "pass",
|
|
758
|
+
observedOutcome,
|
|
759
|
+
reviewLabels: scenario.reviewLabels,
|
|
760
|
+
},
|
|
761
|
+
...(scenario.externalRefs ? { externalRefs: scenario.externalRefs } : {}),
|
|
762
|
+
run: {
|
|
763
|
+
...inputs.trajectory.run,
|
|
764
|
+
scenarioId: scenario.id,
|
|
765
|
+
replay: true,
|
|
766
|
+
},
|
|
767
|
+
counts: {
|
|
768
|
+
assertions: assertions.length,
|
|
769
|
+
passed: assertions.filter((assertion) => assertion.status === "pass")
|
|
770
|
+
.length,
|
|
771
|
+
failed: failed.length,
|
|
772
|
+
warnings: warnings.length,
|
|
773
|
+
events: inputs.trajectory.counts.events,
|
|
774
|
+
toolCalls: countToolCalls(inputs.trajectory),
|
|
775
|
+
replayDeltas: inputs.replay.counts.deltas,
|
|
776
|
+
scoreFailures: scoreFailures(inputs.score),
|
|
777
|
+
scoreWarnings: scoreWarnings(inputs.score),
|
|
778
|
+
workspaceFiles: inputs.workspaceManifest?.files.length ?? 0,
|
|
779
|
+
toolAdapters: inputs.workspaceManifest?.toolAdapters.length ?? 0,
|
|
780
|
+
},
|
|
781
|
+
platform: {
|
|
782
|
+
...scenario.platform,
|
|
783
|
+
evidenceEventType: "maestro.events.eval.scored",
|
|
784
|
+
},
|
|
785
|
+
...(releaseGate ? { releaseGate } : {}),
|
|
786
|
+
...(workspace ? { workspace } : {}),
|
|
787
|
+
assumptions: scenario.assumptions,
|
|
788
|
+
assertions,
|
|
789
|
+
provenance: buildProvenance(inputs.trajectory),
|
|
790
|
+
...(diff ? { diff } : {}),
|
|
791
|
+
};
|
|
792
|
+
}
|
|
793
|
+
export function scenarioResultToJunit(result) {
|
|
794
|
+
const outcomeMatches = result.scenario.observedOutcome === result.scenario.expectedOutcome;
|
|
795
|
+
const failures = result.assertions.filter((assertion) => assertion.status === "fail");
|
|
796
|
+
const testcases = result.assertions
|
|
797
|
+
.map((assertion) => {
|
|
798
|
+
const failure = !outcomeMatches && assertion.status === "fail"
|
|
799
|
+
? `\n\t\t<failure message="${escapeXml(assertion.message)}">${escapeXml(JSON.stringify(assertion.evidence))}</failure>\n\t`
|
|
800
|
+
: "";
|
|
801
|
+
const expectedFailureOutput = outcomeMatches && assertion.status === "fail"
|
|
802
|
+
? `\n\t\t<system-out>${escapeXml([
|
|
803
|
+
`Expected failing assertion observed: ${assertion.message}`,
|
|
804
|
+
JSON.stringify(assertion.evidence),
|
|
805
|
+
].join("\n"))}</system-out>\n\t`
|
|
806
|
+
: "";
|
|
807
|
+
return `\t<testcase classname="${escapeXml(result.scenario.id)}" name="${escapeXml(assertion.id)}">${failure}${expectedFailureOutput}</testcase>`;
|
|
808
|
+
})
|
|
809
|
+
.join("\n");
|
|
810
|
+
const outcomeFailure = !outcomeMatches && failures.length === 0
|
|
811
|
+
? `\t<testcase classname="${escapeXml(result.scenario.id)}" name="scenario-outcome">\n\t\t<failure message="${escapeXml(`Observed outcome ${result.scenario.observedOutcome}; expected ${result.scenario.expectedOutcome}.`)}"></failure>\n\t</testcase>\n`
|
|
812
|
+
: "";
|
|
813
|
+
const failureCount = outcomeMatches ? 0 : Math.max(1, failures.length);
|
|
814
|
+
const testCount = result.counts.assertions + (outcomeFailure ? 1 : 0);
|
|
815
|
+
return `<?xml version="1.0" encoding="UTF-8"?>\n<testsuite name="${escapeXml(result.scenario.id)}" tests="${testCount}" failures="${failureCount}" warnings="${result.counts.warnings}">\n${outcomeFailure}${testcases}\n</testsuite>\n`;
|
|
816
|
+
}
|
|
817
|
+
export function runAgentTrajectoryScenarioFile(path, options = {}) {
|
|
818
|
+
const scenario = loadAgentTrajectoryScenario(path);
|
|
819
|
+
const inputs = loadAgentTrajectoryScenarioInputs(scenario, options.baseDir ?? process.cwd());
|
|
820
|
+
return evaluateAgentTrajectoryScenario(scenario, inputs);
|
|
821
|
+
}
|
|
822
|
+
//# sourceMappingURL=agent-trajectory-scenarios.js.map
|