stageflow 0.40.0 → 0.44.0
This diff represents the content of publicly available package versions that have been released to one of the supported registries. The information contained in this diff is provided for informational purposes only and reflects changes between package versions as they appear in their respective public registries.
- package/README.md +15 -4
- package/dist/agent/activity.d.ts +18 -0
- package/dist/agent/claudeAdapter.js +7 -0
- package/dist/agent/claudeTools.d.ts +27 -18
- package/dist/agent/claudeTools.js +54 -15
- package/dist/agent/compactionEvents.d.ts +34 -0
- package/dist/agent/compactionEvents.js +37 -0
- package/dist/agent/openRouterRouting.d.ts +7 -0
- package/dist/agent/openRouterRouting.js +23 -0
- package/dist/agent/piAdapter.d.ts +6 -0
- package/dist/agent/piAdapter.js +128 -27
- package/dist/agent/piSettings.d.ts +7 -0
- package/dist/agent/piSettings.js +9 -0
- package/dist/agent/sandboxPiTools.d.ts +3 -1
- package/dist/agent/sandboxPiTools.js +8 -2
- package/dist/agent/stageTurnDriver.d.ts +11 -2
- package/dist/agent/stageTurnDriver.js +45 -4
- package/dist/agent/toolOutputBudget.d.ts +15 -1
- package/dist/agent/toolOutputBudget.js +18 -0
- package/dist/agent/toolOutputBudgetExtension.d.ts +2 -0
- package/dist/agent/toolOutputBudgetExtension.js +40 -3
- package/dist/agents/authorChat.d.ts +5 -0
- package/dist/agents/authorChat.js +120 -12
- package/dist/agents/authorSessionStore.d.ts +38 -0
- package/dist/agents/authorSessionStore.js +167 -0
- package/dist/agents/chatTurn.d.ts +2 -0
- package/dist/agents/chatTurn.js +17 -1
- package/dist/agents/errors.d.ts +1 -1
- package/dist/agents/errors.js +1 -0
- package/dist/agents/evalGate.d.ts +61 -0
- package/dist/agents/evalGate.js +103 -0
- package/dist/agents/evalSession.d.ts +26 -0
- package/dist/agents/evalSession.js +29 -0
- package/dist/agents/hostContext.d.ts +4 -1
- package/dist/agents/hostContext.js +4 -1
- package/dist/agents/runtime/piRuntime.d.ts +6 -2
- package/dist/agents/runtime/piRuntime.js +50 -16
- package/dist/agents/runtime/types.d.ts +3 -1
- package/dist/agents/service.d.ts +27 -2
- package/dist/agents/service.js +91 -15
- package/dist/agents/sessions.d.ts +8 -0
- package/dist/agents/sessions.js +53 -11
- package/dist/agents/store.d.ts +14 -3
- package/dist/agents/store.js +45 -7
- package/dist/agents/toolRisk.d.ts +21 -0
- package/dist/agents/toolRisk.js +50 -0
- package/dist/cli/doctorCommand.js +26 -0
- package/dist/cli/evalCommand.d.ts +20 -0
- package/dist/cli/evalCommand.js +720 -0
- package/dist/cli/exportAllCommand.js +1 -1
- package/dist/cli/hostClient.d.ts +60 -0
- package/dist/cli/hostClient.js +71 -0
- package/dist/cli/validateOutput.d.ts +2 -2
- package/dist/cli/validateOutput.js +2 -2
- package/dist/cli.js +10 -0
- package/dist/compaction/directives.d.ts +18 -0
- package/dist/compaction/directives.js +119 -0
- package/dist/compaction/extension.d.ts +20 -0
- package/dist/compaction/extension.js +78 -0
- package/dist/compaction/mask.d.ts +18 -0
- package/dist/compaction/mask.js +155 -0
- package/dist/compaction/pin.d.ts +21 -0
- package/dist/compaction/pin.js +270 -0
- package/dist/compaction/scrub.d.ts +7 -0
- package/dist/compaction/scrub.js +26 -0
- package/dist/compaction/sessionWiring.d.ts +22 -0
- package/dist/compaction/sessionWiring.js +26 -0
- package/dist/compaction/summarize.d.ts +79 -0
- package/dist/compaction/summarize.js +171 -0
- package/dist/compaction/toolPolicy.d.ts +3 -0
- package/dist/compaction/toolPolicy.js +33 -0
- package/dist/compaction/types.d.ts +45 -0
- package/dist/compaction/types.js +7 -0
- package/dist/compaction/workingState.d.ts +56 -0
- package/dist/compaction/workingState.js +198 -0
- package/dist/config/atomicWrite.d.ts +15 -0
- package/dist/config/atomicWrite.js +43 -0
- package/dist/config/browseCatalog.d.ts +2 -0
- package/dist/config/browseCatalog.js +4 -2
- package/dist/config/createPipeline.js +2 -0
- package/dist/config/createTrigger.js +7 -1
- package/dist/config/csv.d.ts +9 -0
- package/dist/config/csv.js +81 -0
- package/dist/config/datasetAgentCheck.d.ts +18 -0
- package/dist/config/datasetAgentCheck.js +109 -0
- package/dist/config/datasetCandidates.d.ts +22 -0
- package/dist/config/datasetCandidates.js +86 -0
- package/dist/config/datasetFromSessions.d.ts +34 -0
- package/dist/config/datasetFromSessions.js +105 -0
- package/dist/config/datasetHash.d.ts +1 -0
- package/dist/config/datasetHash.js +4 -0
- package/dist/config/datasetImport.d.ts +21 -0
- package/dist/config/datasetImport.js +120 -0
- package/dist/config/datasetStageCheck.d.ts +9 -0
- package/dist/config/datasetStageCheck.js +17 -0
- package/dist/config/datasetTypes.d.ts +27 -0
- package/dist/config/datasetTypes.js +1 -0
- package/dist/config/ensureCatalogScanRoot.d.ts +4 -1
- package/dist/config/ensureCatalogScanRoot.js +8 -6
- package/dist/config/evalCatalog.d.ts +22 -0
- package/dist/config/evalCatalog.js +95 -0
- package/dist/config/evalWriteRoots.d.ts +17 -0
- package/dist/config/evalWriteRoots.js +109 -0
- package/dist/config/githubTriggerAction.d.ts +8 -0
- package/dist/config/githubTriggerAction.js +29 -0
- package/dist/config/hostConfig.js +3 -0
- package/dist/config/installEvalSample.d.ts +34 -0
- package/dist/config/installEvalSample.js +182 -0
- package/dist/config/loadDataset.d.ts +16 -0
- package/dist/config/loadDataset.js +318 -0
- package/dist/config/loadEval.d.ts +32 -0
- package/dist/config/loadEval.js +1103 -0
- package/dist/config/loadOutcome.d.ts +5 -0
- package/dist/config/loadStageflowManifest.js +64 -0
- package/dist/config/pipelineNeeds.d.ts +2 -0
- package/dist/config/pipelineNeeds.js +16 -3
- package/dist/config/pipelineRoute.js +5 -4
- package/dist/config/resolveAgentTarget.d.ts +32 -0
- package/dist/config/resolveAgentTarget.js +118 -0
- package/dist/config/resolveEvalTarget.d.ts +21 -0
- package/dist/config/resolveEvalTarget.js +89 -0
- package/dist/config/scanCatalogPaths.d.ts +1 -1
- package/dist/config/scanCatalogPaths.js +17 -5
- package/dist/config/validateCatalog.d.ts +2 -2
- package/dist/config/validateCatalog.js +21 -1
- package/dist/config/validateEval.d.ts +36 -0
- package/dist/config/validateEval.js +446 -0
- package/dist/config/writeDataset.d.ts +54 -0
- package/dist/config/writeDataset.js +186 -0
- package/dist/config/writeEval.d.ts +25 -0
- package/dist/config/writeEval.js +398 -0
- package/dist/diagnostics/checks.js +1 -1
- package/dist/envelope/assertEmittedArtifacts.d.ts +12 -0
- package/dist/envelope/assertEmittedArtifacts.js +43 -0
- package/dist/envelope/preEmitChecks.d.ts +3 -0
- package/dist/evals/agentDiff.d.ts +13 -0
- package/dist/evals/agentDiff.js +11 -0
- package/dist/evals/canonical.d.ts +3 -0
- package/dist/evals/canonical.js +24 -0
- package/dist/evals/caseTask.d.ts +3 -0
- package/dist/evals/caseTask.js +9 -0
- package/dist/evals/compare.d.ts +94 -0
- package/dist/evals/compare.js +201 -0
- package/dist/evals/constants.d.ts +62 -0
- package/dist/evals/constants.js +97 -0
- package/dist/evals/engine/FakeEngine.d.ts +33 -0
- package/dist/evals/engine/FakeEngine.js +177 -0
- package/dist/evals/engine/index.d.ts +9 -0
- package/dist/evals/engine/index.js +19 -0
- package/dist/evals/engine/promptfoo/PromptfooEngine.d.ts +52 -0
- package/dist/evals/engine/promptfoo/PromptfooEngine.js +312 -0
- package/dist/evals/engine/promptfoo/errors.d.ts +5 -0
- package/dist/evals/engine/promptfoo/errors.js +8 -0
- package/dist/evals/engine/promptfoo/generate.d.ts +30 -0
- package/dist/evals/engine/promptfoo/generate.js +345 -0
- package/dist/evals/engine/promptfoo/guard.d.ts +10 -0
- package/dist/evals/engine/promptfoo/guard.js +116 -0
- package/dist/evals/engine/promptfoo/install.d.ts +17 -0
- package/dist/evals/engine/promptfoo/install.js +73 -0
- package/dist/evals/engine/promptfoo/judgeProviders.d.ts +2 -0
- package/dist/evals/engine/promptfoo/judgeProviders.js +9 -0
- package/dist/evals/engine/promptfoo/judgeServices.d.ts +6 -0
- package/dist/evals/engine/promptfoo/judgeServices.js +41 -0
- package/dist/evals/engine/promptfoo/locate.d.ts +21 -0
- package/dist/evals/engine/promptfoo/locate.js +68 -0
- package/dist/evals/engine/promptfoo/nodeVersion.d.ts +7 -0
- package/dist/evals/engine/promptfoo/nodeVersion.js +14 -0
- package/dist/evals/engine/promptfoo/normalize.d.ts +66 -0
- package/dist/evals/engine/promptfoo/normalize.js +261 -0
- package/dist/evals/engine/promptfoo/pinned.d.ts +5 -0
- package/dist/evals/engine/promptfoo/pinned.js +5 -0
- package/dist/evals/engine/promptfoo/rubricPrompt.d.ts +14 -0
- package/dist/evals/engine/promptfoo/rubricPrompt.js +105 -0
- package/dist/evals/engine/promptfoo/runTokens.d.ts +18 -0
- package/dist/evals/engine/promptfoo/runTokens.js +31 -0
- package/dist/evals/engine/promptfoo/status.d.ts +13 -0
- package/dist/evals/engine/promptfoo/status.js +16 -0
- package/dist/evals/estimate.d.ts +56 -0
- package/dist/evals/estimate.js +66 -0
- package/dist/evals/graders/codeGrader.d.ts +6 -0
- package/dist/evals/graders/codeGrader.js +85 -0
- package/dist/evals/graders/conversationCheckGrader.d.ts +6 -0
- package/dist/evals/graders/conversationCheckGrader.js +62 -0
- package/dist/evals/graders/fieldGrader.d.ts +11 -0
- package/dist/evals/graders/fieldGrader.js +117 -0
- package/dist/evals/graders/goal.d.ts +3 -0
- package/dist/evals/graders/goal.js +14 -0
- package/dist/evals/graders/index.d.ts +18 -0
- package/dist/evals/graders/index.js +34 -0
- package/dist/evals/graders/notApplicable.d.ts +5 -0
- package/dist/evals/graders/notApplicable.js +9 -0
- package/dist/evals/graders/path.d.ts +26 -0
- package/dist/evals/graders/path.js +75 -0
- package/dist/evals/graders/referenceGrader.d.ts +5 -0
- package/dist/evals/graders/referenceGrader.js +59 -0
- package/dist/evals/graders/schemaGrader.d.ts +5 -0
- package/dist/evals/graders/schemaGrader.js +29 -0
- package/dist/evals/graders/subject.d.ts +6 -0
- package/dist/evals/graders/subject.js +33 -0
- package/dist/evals/graders/toolCallsGrader.d.ts +6 -0
- package/dist/evals/graders/toolCallsGrader.js +100 -0
- package/dist/evals/graders/transcript.d.ts +12 -0
- package/dist/evals/graders/transcript.js +57 -0
- package/dist/evals/graders/types.d.ts +20 -0
- package/dist/evals/graders/types.js +1 -0
- package/dist/evals/judgeProviders.d.ts +16 -0
- package/dist/evals/judgeProviders.js +27 -0
- package/dist/evals/lineDiff.d.ts +5 -0
- package/dist/evals/lineDiff.js +34 -0
- package/dist/evals/listStatus.d.ts +31 -0
- package/dist/evals/listStatus.js +48 -0
- package/dist/evals/manager.d.ts +187 -0
- package/dist/evals/manager.js +795 -0
- package/dist/evals/plan.d.ts +65 -0
- package/dist/evals/plan.js +621 -0
- package/dist/evals/planInternalAgent.d.ts +19 -0
- package/dist/evals/planInternalAgent.js +205 -0
- package/dist/evals/providerKeys.d.ts +1 -0
- package/dist/evals/providerKeys.js +21 -0
- package/dist/evals/read/agentTargets.d.ts +37 -0
- package/dist/evals/read/agentTargets.js +101 -0
- package/dist/evals/read/api.d.ts +208 -0
- package/dist/evals/read/api.js +363 -0
- package/dist/evals/read/compare.d.ts +4 -0
- package/dist/evals/read/compare.js +39 -0
- package/dist/evals/read/datasets.d.ts +105 -0
- package/dist/evals/read/datasets.js +318 -0
- package/dist/evals/read/list.d.ts +66 -0
- package/dist/evals/read/list.js +182 -0
- package/dist/evals/read/results.d.ts +46 -0
- package/dist/evals/read/results.js +107 -0
- package/dist/evals/read/runs.d.ts +154 -0
- package/dist/evals/read/runs.js +295 -0
- package/dist/evals/read/suggestions.d.ts +43 -0
- package/dist/evals/read/suggestions.js +73 -0
- package/dist/evals/runner/FakeAgentBackend.d.ts +40 -0
- package/dist/evals/runner/FakeAgentBackend.js +82 -0
- package/dist/evals/runner/agentCaseRunner.d.ts +24 -0
- package/dist/evals/runner/agentCaseRunner.js +388 -0
- package/dist/evals/runner/caseRunner.d.ts +8 -0
- package/dist/evals/runner/caseRunner.js +28 -0
- package/dist/evals/runner/conversationSubject.d.ts +18 -0
- package/dist/evals/runner/conversationSubject.js +123 -0
- package/dist/evals/runner/internalAgentCaseRunner.d.ts +6 -0
- package/dist/evals/runner/internalAgentCaseRunner.js +250 -0
- package/dist/evals/runner/simulatorModel.d.ts +51 -0
- package/dist/evals/runner/simulatorModel.js +58 -0
- package/dist/evals/runner/simulatorPrompt.d.ts +29 -0
- package/dist/evals/runner/simulatorPrompt.js +59 -0
- package/dist/evals/runner/stageCaseRunner.d.ts +30 -0
- package/dist/evals/runner/stageCaseRunner.js +279 -0
- package/dist/evals/runner/userDriver.d.ts +71 -0
- package/dist/evals/runner/userDriver.js +104 -0
- package/dist/evals/scoring.d.ts +34 -0
- package/dist/evals/scoring.js +67 -0
- package/dist/evals/snapshot.d.ts +191 -0
- package/dist/evals/snapshot.js +70 -0
- package/dist/evals/stageDiff.d.ts +8 -0
- package/dist/evals/stageDiff.js +48 -0
- package/dist/evals/stats.d.ts +95 -0
- package/dist/evals/stats.js +220 -0
- package/dist/evals/statsAdapter.d.ts +4 -0
- package/dist/evals/statsAdapter.js +30 -0
- package/dist/evals/statsTypes.d.ts +51 -0
- package/dist/evals/statsTypes.js +18 -0
- package/dist/evals/storeError.d.ts +5 -0
- package/dist/evals/storeError.js +8 -0
- package/dist/evals/storeTypes.d.ts +207 -0
- package/dist/evals/storeTypes.js +1 -0
- package/dist/evals/tTable.d.ts +3 -0
- package/dist/evals/tTable.js +15 -0
- package/dist/evals/triage/constants.d.ts +26 -0
- package/dist/evals/triage/constants.js +33 -0
- package/dist/evals/triage/credentials.d.ts +13 -0
- package/dist/evals/triage/credentials.js +20 -0
- package/dist/evals/triage/jevClassifier.d.ts +14 -0
- package/dist/evals/triage/jevClassifier.js +124 -0
- package/dist/evals/triage/judgeClassifier.d.ts +10 -0
- package/dist/evals/triage/judgeClassifier.js +76 -0
- package/dist/evals/triage/prompts.d.ts +16 -0
- package/dist/evals/triage/prompts.js +64 -0
- package/dist/evals/triage/service.d.ts +79 -0
- package/dist/evals/triage/service.js +217 -0
- package/dist/evals/triage/stateBuilder.d.ts +19 -0
- package/dist/evals/triage/stateBuilder.js +125 -0
- package/dist/evals/triage/types.d.ts +60 -0
- package/dist/evals/triage/types.js +8 -0
- package/dist/evals/types.d.ts +357 -0
- package/dist/evals/types.js +4 -0
- package/dist/guide/cardContract.d.ts +133 -0
- package/dist/guide/cardContract.js +156 -0
- package/dist/guide/chat.d.ts +45 -0
- package/dist/guide/chat.js +378 -0
- package/dist/guide/knowledge/index.d.ts +14 -0
- package/dist/guide/knowledge/index.js +109 -0
- package/dist/guide/knowledge/topics/agents.d.ts +2 -0
- package/dist/guide/knowledge/topics/agents.js +58 -0
- package/dist/guide/knowledge/topics/bestPractices.d.ts +2 -0
- package/dist/guide/knowledge/topics/bestPractices.js +67 -0
- package/dist/guide/knowledge/topics/connections.d.ts +2 -0
- package/dist/guide/knowledge/topics/connections.js +66 -0
- package/dist/guide/knowledge/topics/consoleMap.d.ts +2 -0
- package/dist/guide/knowledge/topics/consoleMap.js +55 -0
- package/dist/guide/knowledge/topics/evals.d.ts +2 -0
- package/dist/guide/knowledge/topics/evals.js +72 -0
- package/dist/guide/knowledge/topics/limitations.d.ts +2 -0
- package/dist/guide/knowledge/topics/limitations.js +84 -0
- package/dist/guide/knowledge/topics/mcpSkillsBrowser.d.ts +2 -0
- package/dist/guide/knowledge/topics/mcpSkillsBrowser.js +57 -0
- package/dist/guide/knowledge/topics/memories.d.ts +2 -0
- package/dist/guide/knowledge/topics/memories.js +63 -0
- package/dist/guide/knowledge/topics/modelsAndProviders.d.ts +2 -0
- package/dist/guide/knowledge/topics/modelsAndProviders.js +56 -0
- package/dist/guide/knowledge/topics/overview.d.ts +2 -0
- package/dist/guide/knowledge/topics/overview.js +56 -0
- package/dist/guide/knowledge/topics/pipelines.d.ts +2 -0
- package/dist/guide/knowledge/topics/pipelines.js +93 -0
- package/dist/guide/knowledge/topics/stages.d.ts +2 -0
- package/dist/guide/knowledge/topics/stages.js +54 -0
- package/dist/guide/knowledge/topics/tasks.d.ts +2 -0
- package/dist/guide/knowledge/topics/tasks.js +52 -0
- package/dist/guide/knowledge/topics/triggers.d.ts +2 -0
- package/dist/guide/knowledge/topics/triggers.js +68 -0
- package/dist/guide/knowledge/topics/troubleshooting.d.ts +2 -0
- package/dist/guide/knowledge/topics/troubleshooting.js +85 -0
- package/dist/guide/knowledgeTools.d.ts +2 -0
- package/dist/guide/knowledgeTools.js +56 -0
- package/dist/guide/showUi.d.ts +6 -0
- package/dist/guide/showUi.js +546 -0
- package/dist/guide/threadStore.d.ts +48 -0
- package/dist/guide/threadStore.js +189 -0
- package/dist/guide/writeRoot.d.ts +11 -0
- package/dist/guide/writeRoot.js +70 -0
- package/dist/hosted/translate.js +2 -0
- package/dist/mcp/agentTools.js +12 -0
- package/dist/mcp/catalogTools.js +18 -20
- package/dist/mcp/catalogWriteTools.d.ts +8 -0
- package/dist/mcp/catalogWriteTools.js +225 -0
- package/dist/mcp/connectionTools.d.ts +3 -0
- package/dist/mcp/connectionTools.js +85 -0
- package/dist/mcp/controlTools.js +34 -1
- package/dist/mcp/deps.d.ts +16 -0
- package/dist/mcp/evalTools.d.ts +3 -0
- package/dist/mcp/evalTools.js +943 -0
- package/dist/mcp/memoryHostTools.d.ts +3 -0
- package/dist/mcp/memoryHostTools.js +72 -0
- package/dist/mcp/pipelineAuthorTools.js +14 -1
- package/dist/mcp/server.js +6 -0
- package/dist/mcp/tools.js +8 -0
- package/dist/mcp/triggerTools.js +77 -31
- package/dist/models/response.d.ts +7 -2
- package/dist/models/response.js +6 -1
- package/dist/operatorAgent/agentAuthorContext.d.ts +1 -1
- package/dist/operatorAgent/agentAuthorContext.js +1 -1
- package/dist/operatorAgent/evalGate.d.ts +22 -0
- package/dist/operatorAgent/evalGate.js +77 -0
- package/dist/operatorAgent/fakeHost.js +10 -2
- package/dist/operatorAgent/host.js +18 -2
- package/dist/operatorAgent/index.d.ts +1 -0
- package/dist/operatorAgent/index.js +1 -0
- package/dist/operatorAgent/internalAgents.d.ts +11 -0
- package/dist/operatorAgent/internalAgents.js +13 -0
- package/dist/operatorAgent/piModel.d.ts +5 -1
- package/dist/operatorAgent/piModel.js +32 -14
- package/dist/operatorAgent/profiles/agentAuthor.d.ts +3 -1
- package/dist/operatorAgent/profiles/agentAuthor.js +201 -32
- package/dist/operatorAgent/profiles/agentAuthorPlaybook.d.ts +1 -1
- package/dist/operatorAgent/profiles/agentAuthorPlaybook.js +48 -15
- package/dist/operatorAgent/profiles/authorReadable.d.ts +2 -0
- package/dist/operatorAgent/profiles/authorReadable.js +26 -0
- package/dist/operatorAgent/profiles/guideConfirm.d.ts +21 -0
- package/dist/operatorAgent/profiles/guideConfirm.js +100 -0
- package/dist/operatorAgent/profiles/guideTools.d.ts +7 -0
- package/dist/operatorAgent/profiles/guideTools.js +123 -0
- package/dist/operatorAgent/profiles/stageflowGuide.d.ts +7 -0
- package/dist/operatorAgent/profiles/stageflowGuide.js +132 -0
- package/dist/operatorAgent/profiles/workshopAuthor.js +23 -0
- package/dist/operatorAgent/profiles/workshopAuthorPlaybook.d.ts +1 -1
- package/dist/operatorAgent/profiles/workshopAuthorPlaybook.js +9 -1
- package/dist/operatorAgent/profiles/workshopAuthorTools.d.ts +2 -0
- package/dist/operatorAgent/profiles/workshopAuthorTools.js +197 -9
- package/dist/operatorAgent/types.d.ts +23 -0
- package/dist/package-meta.d.ts +1 -1
- package/dist/package-meta.js +1 -1
- package/dist/runstore/port.d.ts +43 -0
- package/dist/runstore/retention.d.ts +2 -1
- package/dist/runstore/retention.js +3 -7
- package/dist/runstore/runStatusClock.d.ts +17 -0
- package/dist/runstore/runStatusClock.js +19 -0
- package/dist/runstore/sqlite/SqliteRunStore.d.ts +29 -0
- package/dist/runstore/sqlite/SqliteRunStore.js +144 -50
- package/dist/runstore/sqlite/evalStore.d.ts +20 -0
- package/dist/runstore/sqlite/evalStore.js +541 -0
- package/dist/runstore/sqlite/migrations/015-evals.d.ts +12 -0
- package/dist/runstore/sqlite/migrations/015-evals.js +157 -0
- package/dist/runstore/sqlite/migrations/016-eval-subject-ref.d.ts +12 -0
- package/dist/runstore/sqlite/migrations/016-eval-subject-ref.js +20 -0
- package/dist/runstore/sqlite/migrations/017-eval-result-suggestions.d.ts +11 -0
- package/dist/runstore/sqlite/migrations/017-eval-result-suggestions.js +36 -0
- package/dist/runstore/sqlite/migrations/index.js +6 -0
- package/dist/runstore/sqlite/suggestionStore.d.ts +14 -0
- package/dist/runstore/sqlite/suggestionStore.js +132 -0
- package/dist/runstore/workspaceLayout.js +6 -1
- package/dist/runtime/artifactDiskInspect.d.ts +14 -0
- package/dist/runtime/artifactDiskInspect.js +52 -0
- package/dist/runtime/artifactPath.d.ts +24 -0
- package/dist/runtime/artifactPath.js +84 -0
- package/dist/runtime/completionCheckRunner.js +15 -43
- package/dist/runtime/githubPollSource.d.ts +22 -4
- package/dist/runtime/githubPollSource.js +93 -8
- package/dist/runtime/pipelineRunner.d.ts +2 -1
- package/dist/runtime/pipelineRunner.js +2 -0
- package/dist/runtime/priorArtifactGrants.d.ts +8 -0
- package/dist/runtime/priorArtifactGrants.js +105 -0
- package/dist/runtime/runActionEligibility.d.ts +46 -0
- package/dist/runtime/runActionEligibility.js +123 -0
- package/dist/runtime/runManager.d.ts +6 -2
- package/dist/runtime/runManager.js +60 -16
- package/dist/runtime/runRetentionSweep.js +2 -2
- package/dist/runtime/runRetryCoordinator.d.ts +2 -14
- package/dist/runtime/runRetryCoordinator.js +3 -67
- package/dist/runtime/sandboxContainer.d.ts +7 -0
- package/dist/runtime/sandboxContainer.js +48 -6
- package/dist/runtime/sandboxDesk.d.ts +2 -1
- package/dist/runtime/sandboxDesk.js +11 -12
- package/dist/runtime/stageEnvironment.js +1 -0
- package/dist/runtime/stageRoots.d.ts +12 -0
- package/dist/runtime/stageSandbox.d.ts +2 -1
- package/dist/runtime/stageSandbox.js +18 -2
- package/dist/runtime/stageWorker.js +20 -0
- package/dist/server/agentAuthorRoutes.js +59 -1
- package/dist/server/bootstrap.d.ts +13 -0
- package/dist/server/bootstrap.js +69 -0
- package/dist/server/controlToken.js +3 -0
- package/dist/server/evalInternalRoutes.d.ts +14 -0
- package/dist/server/evalInternalRoutes.js +87 -0
- package/dist/server/evalRouteContext.d.ts +46 -0
- package/dist/server/evalRouteContext.js +43 -0
- package/dist/server/evalRoutes.d.ts +5 -0
- package/dist/server/evalRoutes.js +1456 -0
- package/dist/server/evalTriageRoutes.d.ts +2 -0
- package/dist/server/evalTriageRoutes.js +95 -0
- package/dist/server/guideRoutes.d.ts +5 -0
- package/dist/server/guideRoutes.js +166 -0
- package/dist/server/http.d.ts +15 -0
- package/dist/server/http.js +67 -3
- package/dist/server/internalAgentRoutes.d.ts +6 -0
- package/dist/server/internalAgentRoutes.js +12 -0
- package/dist/server/mcpHost.js +5 -0
- package/dist/toolkit/meta.d.ts +1 -1
- package/dist/toolkit/meta.js +53 -0
- package/dist/toolkit/profiles.d.ts +5 -0
- package/dist/toolkit/profiles.js +13 -1
- package/dist/tools/askOperator.js +2 -1
- package/dist/tools/emitStageEnvelope.d.ts +17 -16
- package/dist/tools/emitStageEnvelope.js +46 -32
- package/dist/tools/stageToolNames.d.ts +4 -0
- package/dist/tools/stageToolNames.js +4 -0
- package/dist/tools/workingState.d.ts +40 -0
- package/dist/tools/workingState.js +108 -0
- package/dist/tools/writeStageArtifact.d.ts +0 -9
- package/dist/tools/writeStageArtifact.js +2 -1
- package/dist/types/stageflowManifest.d.ts +6 -0
- package/dist/ui/assets/{MarkdownText-D7tYaIFO.js → MarkdownText-S5C_vemc.js} +2 -2
- package/dist/ui/assets/adapters-C0l4ruiR.js +17 -0
- package/dist/ui/assets/chat-CjnsmvqR.js +2 -0
- package/dist/ui/assets/chat-DFDd4UjS.css +2 -0
- package/dist/ui/assets/{main-BFo79krC.css → main-Del1GeLf.css} +1 -1
- package/dist/ui/assets/main-_yXVqcW9.js +277 -0
- package/dist/ui/chat.html +3 -4
- package/dist/ui/index.html +3 -4
- package/dist/workshop/chatTurn.js +1 -1
- package/dist/workshop/workshopPin.d.ts +3 -3
- package/dist/workshop/workshopPin.js +16 -15
- package/examples/README.md +2 -0
- package/examples/compaction-stress/README.md +30 -0
- package/examples/compaction-stress/check-run.mjs +53 -0
- package/examples/compaction-stress/collect-evidence.yaml +60 -0
- package/examples/compaction-stress/compaction-stress.pipeline.yaml +141 -0
- package/examples/compaction-stress/compaction-stress.task.yaml +15 -0
- package/examples/compaction-stress/generate-evidence.mjs +110 -0
- package/examples/compaction-stress/root-cause.yaml +26 -0
- package/examples/compaction-stress/stageflow.yaml +9 -0
- package/examples/compaction-stress/write-postmortem.yaml +40 -0
- package/examples/document-policy-review/README.md +35 -0
- package/examples/document-policy-review/agents/compliance.agent.yaml +42 -0
- package/examples/document-policy-review/documents/acme-vendor-sow.md +15 -0
- package/examples/document-policy-review/documents/support-runbook.md +13 -0
- package/examples/document-policy-review/policies/acceptable-use.md +13 -0
- package/examples/document-policy-review/policies/data-handling.md +17 -0
- package/examples/document-policy-review/policies/vendor-security.md +17 -0
- package/examples/evals/README.md +141 -0
- package/examples/evals/critique-sample.dataset.jsonl +8 -0
- package/examples/evals/critique-sample.eval.yaml +34 -0
- package/examples/evals/critique-sample.task.yaml +4 -0
- package/examples/evals/eval-sample.pipeline.yaml +35 -0
- package/examples/evals/guide-behavior.dataset.jsonl +15 -0
- package/examples/evals/guide-behavior.eval.yaml +35 -0
- package/examples/evals/guide-depth.dataset.jsonl +35 -0
- package/examples/evals/guide-depth.eval.yaml +169 -0
- package/examples/evals/support-sample.agent.yaml +20 -0
- package/examples/evals/support-sample.dataset.jsonl +6 -0
- package/examples/evals/support-sample.eval.yaml +38 -0
- package/examples/memory/evals/guide-behavior.dataset.jsonl +15 -0
- package/examples/memory/evals/guide-behavior.eval.yaml +35 -0
- package/examples/memory/evals/guide-depth.dataset.jsonl +35 -0
- package/examples/memory/evals/guide-depth.eval.yaml +169 -0
- package/examples/stageflow.yaml +6 -0
- package/package.json +1 -1
- package/skills/stageflow/SKILL.md +5 -2
- package/skills/stageflow/references/control-surface.md +4 -2
- package/skills/stageflow/references/evals.md +68 -0
- package/skills/stageflow-author/SKILL.md +4 -0
- package/skills/stageflow-run/SKILL.md +4 -0
- package/dist/ui/assets/AssistantRuntimeProvider-CsiPFhCT.js +0 -9
- package/dist/ui/assets/ReadonlyStore-BfXqHykm.js +0 -6
- package/dist/ui/assets/chat-BbLukq9W.js +0 -5
- package/dist/ui/assets/chat-zIZU_bP1.css +0 -2
- package/dist/ui/assets/main-CijYMeM0.js +0 -259
|
@@ -0,0 +1,546 @@
|
|
|
1
|
+
import path from "node:path";
|
|
2
|
+
import { Type } from "typebox";
|
|
3
|
+
import { z } from "zod";
|
|
4
|
+
import { isAgentError } from "../agents/errors.js";
|
|
5
|
+
import { defaultAgentProjectRoot } from "../agents/projectRoot.js";
|
|
6
|
+
import { loadDraftPackage, loadTaskArtifact, validateDraftPackage } from "../config/draftPackage.js";
|
|
7
|
+
import { listTasksMultiProject } from "../config/multiProjectCatalog.js";
|
|
8
|
+
import { loadCatalogTriggers } from "../config/triggerCatalog.js";
|
|
9
|
+
import { MemoryError } from "../memory/host.js";
|
|
10
|
+
import { buildEvalDetail, buildResultsPage, EvalApiError, locateEval, runByIdOr404 } from "../evals/read/api.js";
|
|
11
|
+
import { buildRunDetail } from "../evals/read/runs.js";
|
|
12
|
+
import { buildChoicesUi, buildLinksUi, buildPlanUi, GuideCardError, GUIDE_LINK_KINDS, GUIDE_RESOURCE_KINDS, GUIDE_UI_MAX_CHOICES, GUIDE_UI_MAX_FAILURES, GUIDE_UI_MAX_LINKS, GUIDE_UI_MAX_MESSAGES, GUIDE_UI_MAX_PLAN_STEPS, GUIDE_UI_MAX_TEXT, } from "./cardContract.js";
|
|
13
|
+
export const GUIDE_UI_TOOL_NAMES = ["show_ui"];
|
|
14
|
+
export { GUIDE_LINK_KINDS, GUIDE_RESOURCE_KINDS, } from "./cardContract.js";
|
|
15
|
+
const MAX_TEXT = GUIDE_UI_MAX_TEXT;
|
|
16
|
+
const MAX_PLAN_STEPS = GUIDE_UI_MAX_PLAN_STEPS;
|
|
17
|
+
const MAX_CHOICES = GUIDE_UI_MAX_CHOICES;
|
|
18
|
+
const MAX_LINKS = GUIDE_UI_MAX_LINKS;
|
|
19
|
+
const MAX_FAILURES = GUIDE_UI_MAX_FAILURES;
|
|
20
|
+
const MAX_MESSAGES = GUIDE_UI_MAX_MESSAGES;
|
|
21
|
+
const text = z.string().trim().min(1).max(MAX_TEXT);
|
|
22
|
+
const ref = z.string().trim().min(1).max(MAX_TEXT);
|
|
23
|
+
const pipelineArgs = z.object({
|
|
24
|
+
component: z.literal("pipeline"),
|
|
25
|
+
path: ref,
|
|
26
|
+
task_path: ref.optional(),
|
|
27
|
+
title: text.optional(),
|
|
28
|
+
});
|
|
29
|
+
const resourceArgs = z.object({
|
|
30
|
+
component: z.literal("resource"),
|
|
31
|
+
kind: z.string(),
|
|
32
|
+
id: ref,
|
|
33
|
+
project_root: ref.optional(),
|
|
34
|
+
title: text.optional(),
|
|
35
|
+
});
|
|
36
|
+
const planArgs = z.object({
|
|
37
|
+
component: z.literal("plan"),
|
|
38
|
+
title: text,
|
|
39
|
+
steps: z
|
|
40
|
+
.array(z.object({ title: text, detail: text.optional(), status: z.enum(["todo", "current", "done"]) }))
|
|
41
|
+
.min(1)
|
|
42
|
+
.max(MAX_PLAN_STEPS),
|
|
43
|
+
});
|
|
44
|
+
const choicesArgs = z.object({
|
|
45
|
+
component: z.literal("choices"),
|
|
46
|
+
question: text,
|
|
47
|
+
options: z
|
|
48
|
+
.array(z.object({ id: ref, label: text, description: text.optional() }))
|
|
49
|
+
.min(1)
|
|
50
|
+
.max(MAX_CHOICES),
|
|
51
|
+
allow_other: z.boolean().optional(),
|
|
52
|
+
});
|
|
53
|
+
const evalReportArgs = z.object({
|
|
54
|
+
component: z.literal("eval_report"),
|
|
55
|
+
eval_run_id: ref,
|
|
56
|
+
});
|
|
57
|
+
const linkArg = z.object({
|
|
58
|
+
kind: z.enum(GUIDE_LINK_KINDS),
|
|
59
|
+
id: ref,
|
|
60
|
+
path: ref.optional(),
|
|
61
|
+
project_root: ref.optional(),
|
|
62
|
+
label: text.optional(),
|
|
63
|
+
});
|
|
64
|
+
const linksArgs = z.object({
|
|
65
|
+
component: z.literal("links"),
|
|
66
|
+
title: text.optional(),
|
|
67
|
+
links: z.array(linkArg).min(1).max(MAX_LINKS),
|
|
68
|
+
});
|
|
69
|
+
class GuideUiError extends Error {
|
|
70
|
+
code;
|
|
71
|
+
constructor(code) {
|
|
72
|
+
super(code);
|
|
73
|
+
this.code = code;
|
|
74
|
+
}
|
|
75
|
+
}
|
|
76
|
+
function fail(code) {
|
|
77
|
+
return { ok: false, content: { ok: false, error: code }, error: code };
|
|
78
|
+
}
|
|
79
|
+
function parseArgs(schema, args) {
|
|
80
|
+
const parsed = schema.safeParse(args);
|
|
81
|
+
if (!parsed.success)
|
|
82
|
+
throw new GuideUiError("guide.ui_invalid_args");
|
|
83
|
+
return parsed.data;
|
|
84
|
+
}
|
|
85
|
+
function isRecord(value) {
|
|
86
|
+
return value !== null && typeof value === "object" && !Array.isArray(value);
|
|
87
|
+
}
|
|
88
|
+
function str(value) {
|
|
89
|
+
return typeof value === "string" && value.trim() ? value.trim() : undefined;
|
|
90
|
+
}
|
|
91
|
+
function clip(value) {
|
|
92
|
+
return value.length > MAX_TEXT ? `${value.slice(0, MAX_TEXT - 1)}…` : value;
|
|
93
|
+
}
|
|
94
|
+
function stringList(value) {
|
|
95
|
+
return Array.isArray(value) ? value.filter((item) => typeof item === "string") : [];
|
|
96
|
+
}
|
|
97
|
+
function stemOf(uses) {
|
|
98
|
+
return path.posix.basename(uses).replace(/(\.stage)?\.ya?ml$/i, "");
|
|
99
|
+
}
|
|
100
|
+
function channelNames(value) {
|
|
101
|
+
if (!Array.isArray(value))
|
|
102
|
+
return [];
|
|
103
|
+
const names = [];
|
|
104
|
+
for (const entry of value) {
|
|
105
|
+
if (!isRecord(entry))
|
|
106
|
+
continue;
|
|
107
|
+
const name = str(entry.alias) ?? str(entry.connection);
|
|
108
|
+
if (name)
|
|
109
|
+
names.push(name);
|
|
110
|
+
}
|
|
111
|
+
return names;
|
|
112
|
+
}
|
|
113
|
+
function hasEntries(value) {
|
|
114
|
+
if (Array.isArray(value))
|
|
115
|
+
return value.length > 0;
|
|
116
|
+
return isRecord(value) && Object.keys(value).length > 0;
|
|
117
|
+
}
|
|
118
|
+
function pipelineStages(draft) {
|
|
119
|
+
const files = new Map((draft.stages ?? []).map((stage) => [stage.path.trim().replace(/\\/g, "/"), stage.body]));
|
|
120
|
+
const stages = draft.pipeline.stages.map((entry) => {
|
|
121
|
+
const uses = str(entry.uses)?.replace(/\\/g, "/");
|
|
122
|
+
const body = uses ? (files.get(uses) ?? {}) : entry;
|
|
123
|
+
const id = str(entry.id) ?? str(body.id) ?? (uses ? stemOf(uses) : "");
|
|
124
|
+
const prompt = str(body.system_prompt)?.split("\n")[0]?.trim();
|
|
125
|
+
const model = str(body.model);
|
|
126
|
+
const channels = body.channels !== undefined ? body.channels : draft.pipeline.channels;
|
|
127
|
+
return {
|
|
128
|
+
id,
|
|
129
|
+
...(prompt ? { summary: clip(prompt) } : {}),
|
|
130
|
+
...(model ? { model } : {}),
|
|
131
|
+
needs: [],
|
|
132
|
+
gates: stringList(body.gate_kinds),
|
|
133
|
+
channels: channelNames(channels),
|
|
134
|
+
hasVerify: hasEntries(body.verify) || hasEntries(body.pre_emit_checks) || hasEntries(body.completion),
|
|
135
|
+
route: Array.isArray(entry.route) ? entry.route : [],
|
|
136
|
+
};
|
|
137
|
+
});
|
|
138
|
+
const byId = new Map(stages.map((stage) => [stage.id, stage]));
|
|
139
|
+
for (const stage of stages) {
|
|
140
|
+
for (const edge of stage.route) {
|
|
141
|
+
if (!isRecord(edge) || edge.type === "loop")
|
|
142
|
+
continue;
|
|
143
|
+
const target = byId.get(str(edge.to) ?? "");
|
|
144
|
+
if (target && !target.needs.includes(stage.id))
|
|
145
|
+
target.needs.push(stage.id);
|
|
146
|
+
}
|
|
147
|
+
}
|
|
148
|
+
return stages.map(({ route: _route, ...stage }) => stage);
|
|
149
|
+
}
|
|
150
|
+
async function pipelineUi(deps, args) {
|
|
151
|
+
const input = parseArgs(pipelineArgs, args);
|
|
152
|
+
const projectRoot = deps.projectRoot ?? deps.cwd;
|
|
153
|
+
const loaded = await loadDraftPackage(projectRoot, input.path, input.task_path ? { taskPath: input.task_path } : {});
|
|
154
|
+
if (!loaded.ok)
|
|
155
|
+
throw new GuideUiError(loaded.status === 404 ? "guide.ui_not_found" : "guide.ui_read_failed");
|
|
156
|
+
const validation = await validateDraftPackage(loaded.draft, { cwd: deps.cwd, projectRoot });
|
|
157
|
+
const { draft } = loaded;
|
|
158
|
+
const model = str(draft.pipeline.model);
|
|
159
|
+
const taskId = str(draft.task?.body.id);
|
|
160
|
+
const taskGoal = str(draft.task?.body.goal);
|
|
161
|
+
return {
|
|
162
|
+
component: "pipeline",
|
|
163
|
+
title: input.title ?? draft.pipeline.id,
|
|
164
|
+
pipeline: {
|
|
165
|
+
id: draft.pipeline.id,
|
|
166
|
+
path: loaded.pipelinePath,
|
|
167
|
+
...(model ? { model } : {}),
|
|
168
|
+
stages: pipelineStages(draft),
|
|
169
|
+
memories: stringList(draft.pipeline.memories),
|
|
170
|
+
...(taskId && taskGoal ? { task: { id: taskId, goal: clip(taskGoal) } } : {}),
|
|
171
|
+
},
|
|
172
|
+
validation: {
|
|
173
|
+
ok: validation.ok,
|
|
174
|
+
errors: validation.summary.errors,
|
|
175
|
+
warnings: validation.summary.warnings,
|
|
176
|
+
messages: validation.findings.slice(0, MAX_MESSAGES).map((finding) => clip(finding.message)),
|
|
177
|
+
},
|
|
178
|
+
link: { kind: "pipeline", id: draft.pipeline.id, path: loaded.pipelinePath, projectRoot },
|
|
179
|
+
};
|
|
180
|
+
}
|
|
181
|
+
function fields(entries) {
|
|
182
|
+
const out = [];
|
|
183
|
+
for (const [label, value] of entries) {
|
|
184
|
+
if (value === undefined || value === null)
|
|
185
|
+
continue;
|
|
186
|
+
const shown = Array.isArray(value) ? value.join(", ") : String(value);
|
|
187
|
+
if (shown.trim())
|
|
188
|
+
out.push({ label, value: clip(shown) });
|
|
189
|
+
}
|
|
190
|
+
return out;
|
|
191
|
+
}
|
|
192
|
+
async function agentResource(deps, id) {
|
|
193
|
+
if (!deps.agentService)
|
|
194
|
+
throw new GuideUiError("guide.ui_unavailable");
|
|
195
|
+
const { agent } = await deps.agentService.get(defaultAgentProjectRoot(deps.cwd), id);
|
|
196
|
+
return {
|
|
197
|
+
kind: "agent",
|
|
198
|
+
fields: fields([
|
|
199
|
+
["Name", agent.name],
|
|
200
|
+
["Description", agent.description],
|
|
201
|
+
["Model", agent.model?.id ?? "default"],
|
|
202
|
+
["Revision", agent.revision],
|
|
203
|
+
["Tools", agent.tools.builtin],
|
|
204
|
+
["MCP servers", agent.tools.mcpServers],
|
|
205
|
+
["Skills", agent.tools.skills],
|
|
206
|
+
["Schedule", agent.scheduleCron],
|
|
207
|
+
["Channel", agent.channelConnection],
|
|
208
|
+
["Sub-agents", agent.subAgents],
|
|
209
|
+
]),
|
|
210
|
+
status: agent.enabled ? "enabled" : "disabled",
|
|
211
|
+
link: { kind: "agent", id: agent.id, label: agent.name },
|
|
212
|
+
};
|
|
213
|
+
}
|
|
214
|
+
async function triggerResource(deps, id, projectRoot) {
|
|
215
|
+
const items = await loadCatalogTriggers({ store: deps.store, bootCwd: deps.cwd });
|
|
216
|
+
const matches = items.filter((item) => item.id === id && (projectRoot === undefined || item.project_root === projectRoot));
|
|
217
|
+
if (matches.length === 0)
|
|
218
|
+
throw new GuideUiError("guide.ui_not_found");
|
|
219
|
+
if (matches.length > 1)
|
|
220
|
+
throw new GuideUiError("guide.ui_ambiguous");
|
|
221
|
+
const trigger = matches[0];
|
|
222
|
+
return {
|
|
223
|
+
kind: "trigger",
|
|
224
|
+
fields: fields([
|
|
225
|
+
["Pipeline", trigger.pipeline],
|
|
226
|
+
["Task", trigger.task],
|
|
227
|
+
["Kind", trigger.kind],
|
|
228
|
+
["Schedule", trigger.schedule ? [trigger.schedule.cron, trigger.schedule.timezone].filter(Boolean).join(" ") : undefined],
|
|
229
|
+
["Event", trigger.event?.source],
|
|
230
|
+
["Last fired", trigger.last_fired_at],
|
|
231
|
+
["Last run", trigger.last_run_id],
|
|
232
|
+
["Next run", trigger.next_run_at],
|
|
233
|
+
]),
|
|
234
|
+
status: trigger.enabled ? "enabled" : "disabled",
|
|
235
|
+
link: { kind: "trigger", id: trigger.id, path: trigger.definition_ref, projectRoot: trigger.project_root },
|
|
236
|
+
};
|
|
237
|
+
}
|
|
238
|
+
async function evalResource(deps, id, projectRoot) {
|
|
239
|
+
const located = await locateEval({ store: deps.store, bootCwd: deps.cwd }, projectRoot, id);
|
|
240
|
+
const detail = await buildEvalDetail({
|
|
241
|
+
store: deps.store,
|
|
242
|
+
projectRoot: located.projectRoot,
|
|
243
|
+
absPath: located.ref.absPath,
|
|
244
|
+
...(deps.agentService !== undefined ? { agents: deps.agentService } : {}),
|
|
245
|
+
});
|
|
246
|
+
const errors = detail.findings.filter((finding) => finding.severity === "error").length;
|
|
247
|
+
const last = detail.last_run;
|
|
248
|
+
return {
|
|
249
|
+
kind: "eval",
|
|
250
|
+
fields: fields([
|
|
251
|
+
["Path", detail.path],
|
|
252
|
+
["Target model", detail.target_model],
|
|
253
|
+
["Judge", detail.judge_default],
|
|
254
|
+
["Runs", detail.runs_count],
|
|
255
|
+
["Last run", last ? `#${last.run_number} ${last.status}` : undefined],
|
|
256
|
+
["Last pass rate", typeof last?.pass_rate === "number" ? `${Math.round(last.pass_rate * 100)}%` : undefined],
|
|
257
|
+
["Problems", errors > 0 ? errors : undefined],
|
|
258
|
+
]),
|
|
259
|
+
status: detail.active_run?.status ?? last?.status ?? "not run",
|
|
260
|
+
link: { kind: "eval", id: detail.id, path: detail.path, projectRoot: located.projectRoot },
|
|
261
|
+
};
|
|
262
|
+
}
|
|
263
|
+
function connectionResource(deps, id) {
|
|
264
|
+
if (!deps.channels)
|
|
265
|
+
throw new GuideUiError("guide.ui_unavailable");
|
|
266
|
+
const listing = deps.channels.list();
|
|
267
|
+
const connection = listing.connections.find((item) => item.id === id);
|
|
268
|
+
if (!connection)
|
|
269
|
+
throw new GuideUiError(listing.error ? "guide.ui_read_failed" : "guide.ui_not_found");
|
|
270
|
+
const health = connection.health;
|
|
271
|
+
return {
|
|
272
|
+
kind: "connection",
|
|
273
|
+
fields: fields([
|
|
274
|
+
["Name", connection.displayName],
|
|
275
|
+
["Provider", connection.provider],
|
|
276
|
+
["Identity", connection.identity],
|
|
277
|
+
["Last check", health ? (health.state === "ok" ? "ok" : (health.error?.code ?? "failed")) : undefined],
|
|
278
|
+
]),
|
|
279
|
+
status: connection.enabled ? (health?.state === "failed" ? "failing" : "enabled") : "disabled",
|
|
280
|
+
link: { kind: "connection", id: connection.id, label: connection.displayName },
|
|
281
|
+
};
|
|
282
|
+
}
|
|
283
|
+
async function memoryResource(deps, id) {
|
|
284
|
+
if (!deps.memories)
|
|
285
|
+
throw new GuideUiError("guide.ui_unavailable");
|
|
286
|
+
const meta = await deps.memories.readMeta(id);
|
|
287
|
+
const cards = await deps.memories.listCards(id);
|
|
288
|
+
return {
|
|
289
|
+
kind: "memory",
|
|
290
|
+
fields: fields([
|
|
291
|
+
["Description", meta.description],
|
|
292
|
+
["Cards", cards.length],
|
|
293
|
+
]),
|
|
294
|
+
link: { kind: "memory", id: meta.id },
|
|
295
|
+
};
|
|
296
|
+
}
|
|
297
|
+
async function taskResource(deps, id, projectRoot) {
|
|
298
|
+
if (/\.ya?ml$/i.test(id) || id.includes("/")) {
|
|
299
|
+
const root = projectRoot ?? deps.projectRoot ?? deps.cwd;
|
|
300
|
+
const loaded = await loadTaskArtifact(root, id);
|
|
301
|
+
if (!loaded.ok)
|
|
302
|
+
throw new GuideUiError(loaded.status === 404 ? "guide.ui_not_found" : "guide.ui_read_failed");
|
|
303
|
+
const body = loaded.task.body;
|
|
304
|
+
const taskId = str(body.id) ?? stemOf(loaded.taskPath).replace(/\.task$/, "");
|
|
305
|
+
return {
|
|
306
|
+
kind: "task",
|
|
307
|
+
fields: fields([
|
|
308
|
+
["Goal", str(body.goal)],
|
|
309
|
+
["Context", str(body.context)],
|
|
310
|
+
["Constraints", str(body.constraints)],
|
|
311
|
+
["Checkout", str(body.checkout)],
|
|
312
|
+
["Repository", str(body.repository)],
|
|
313
|
+
["Ref", str(body.ref)],
|
|
314
|
+
["Path", loaded.taskPath],
|
|
315
|
+
]),
|
|
316
|
+
link: { kind: "task", id: taskId, path: loaded.taskPath, projectRoot: root },
|
|
317
|
+
};
|
|
318
|
+
}
|
|
319
|
+
const listed = await listTasksMultiProject({
|
|
320
|
+
store: deps.store,
|
|
321
|
+
bootCwd: deps.cwd,
|
|
322
|
+
...(projectRoot !== undefined ? { projectRootFilter: projectRoot } : {}),
|
|
323
|
+
});
|
|
324
|
+
const matches = listed.items.filter((item) => item.id === id);
|
|
325
|
+
if (matches.length === 0)
|
|
326
|
+
throw new GuideUiError("guide.ui_not_found");
|
|
327
|
+
if (matches.length > 1)
|
|
328
|
+
throw new GuideUiError("guide.ui_ambiguous");
|
|
329
|
+
const task = matches[0];
|
|
330
|
+
return {
|
|
331
|
+
kind: "task",
|
|
332
|
+
fields: fields([
|
|
333
|
+
["Goal", task.goal],
|
|
334
|
+
["Path", task.path],
|
|
335
|
+
]),
|
|
336
|
+
link: { kind: "task", id: task.id, path: task.path, projectRoot: task.project_root },
|
|
337
|
+
};
|
|
338
|
+
}
|
|
339
|
+
async function resourceUi(deps, args) {
|
|
340
|
+
const input = parseArgs(resourceArgs, args);
|
|
341
|
+
if (!GUIDE_RESOURCE_KINDS.includes(input.kind))
|
|
342
|
+
throw new GuideUiError("guide.ui_unsupported_kind");
|
|
343
|
+
const kind = input.kind;
|
|
344
|
+
const read = kind === "agent"
|
|
345
|
+
? await agentResource(deps, input.id)
|
|
346
|
+
: kind === "trigger"
|
|
347
|
+
? await triggerResource(deps, input.id, input.project_root)
|
|
348
|
+
: kind === "eval"
|
|
349
|
+
? await evalResource(deps, input.id, input.project_root)
|
|
350
|
+
: kind === "connection"
|
|
351
|
+
? connectionResource(deps, input.id)
|
|
352
|
+
: kind === "memory"
|
|
353
|
+
? await memoryResource(deps, input.id)
|
|
354
|
+
: await taskResource(deps, input.id, input.project_root);
|
|
355
|
+
return { component: "resource", title: input.title ?? read.link.label ?? read.link.id, ...read };
|
|
356
|
+
}
|
|
357
|
+
async function evalReportUi(deps, args) {
|
|
358
|
+
const input = parseArgs(evalReportArgs, args);
|
|
359
|
+
const row = await runByIdOr404(deps.store, input.eval_run_id);
|
|
360
|
+
const detail = await buildRunDetail({
|
|
361
|
+
store: deps.store,
|
|
362
|
+
row,
|
|
363
|
+
live: deps.evalRunManager?.getLive(row.evalRunId) ?? null,
|
|
364
|
+
...(deps.agentService !== undefined ? { agents: deps.agentService } : {}),
|
|
365
|
+
});
|
|
366
|
+
const failures = [];
|
|
367
|
+
for (const status of ["failed", "errored"]) {
|
|
368
|
+
if (failures.length >= MAX_FAILURES)
|
|
369
|
+
break;
|
|
370
|
+
const page = await buildResultsPage(deps.store, row.evalRunId, { status, limit: MAX_FAILURES - failures.length });
|
|
371
|
+
for (const item of page.items) {
|
|
372
|
+
failures.push({
|
|
373
|
+
caseId: item.case_id,
|
|
374
|
+
reason: clip(item.failed_graders.length > 0 ? `Failed ${item.failed_graders.join(", ")}` : status),
|
|
375
|
+
});
|
|
376
|
+
}
|
|
377
|
+
}
|
|
378
|
+
return {
|
|
379
|
+
component: "eval_report",
|
|
380
|
+
title: detail.label ?? `${detail.eval_id} run #${detail.run_number}`,
|
|
381
|
+
evalId: detail.eval_id,
|
|
382
|
+
runId: detail.eval_run_id,
|
|
383
|
+
status: detail.status,
|
|
384
|
+
passRate: detail.stats?.pass_rate ?? null,
|
|
385
|
+
passed: detail.passed,
|
|
386
|
+
cases: detail.progress.cases_total,
|
|
387
|
+
failures,
|
|
388
|
+
costUsd: detail.stats?.cost_total ?? detail.progress.cost_usd,
|
|
389
|
+
link: { kind: "eval_run", id: detail.eval_run_id, path: detail.eval_path },
|
|
390
|
+
};
|
|
391
|
+
}
|
|
392
|
+
function jsonList(value) {
|
|
393
|
+
if (typeof value !== "string")
|
|
394
|
+
return value;
|
|
395
|
+
const trimmed = value.trim();
|
|
396
|
+
if (!trimmed.startsWith("["))
|
|
397
|
+
return value;
|
|
398
|
+
try {
|
|
399
|
+
const parsed = JSON.parse(trimmed);
|
|
400
|
+
return Array.isArray(parsed) ? parsed : value;
|
|
401
|
+
}
|
|
402
|
+
catch {
|
|
403
|
+
return value;
|
|
404
|
+
}
|
|
405
|
+
}
|
|
406
|
+
function normalizePlainArgs(args) {
|
|
407
|
+
const next = { ...args };
|
|
408
|
+
for (const key of ["options", "steps", "links"]) {
|
|
409
|
+
if (key in next)
|
|
410
|
+
next[key] = jsonList(next[key]);
|
|
411
|
+
}
|
|
412
|
+
if (next.allow_other === "true")
|
|
413
|
+
next.allow_other = true;
|
|
414
|
+
if (next.allow_other === "false")
|
|
415
|
+
next.allow_other = false;
|
|
416
|
+
return next;
|
|
417
|
+
}
|
|
418
|
+
function plainUi(args, component) {
|
|
419
|
+
args = normalizePlainArgs(args);
|
|
420
|
+
if (component === "plan") {
|
|
421
|
+
const input = parseArgs(planArgs, args);
|
|
422
|
+
return buildPlanUi({
|
|
423
|
+
title: input.title,
|
|
424
|
+
steps: input.steps.map((step) => ({ title: step.title, ...(step.detail ? { detail: step.detail } : {}), status: step.status })),
|
|
425
|
+
});
|
|
426
|
+
}
|
|
427
|
+
if (component === "choices") {
|
|
428
|
+
const input = parseArgs(choicesArgs, args);
|
|
429
|
+
return buildChoicesUi({
|
|
430
|
+
question: input.question,
|
|
431
|
+
options: input.options.map((option) => ({
|
|
432
|
+
id: option.id,
|
|
433
|
+
label: option.label,
|
|
434
|
+
...(option.description ? { description: option.description } : {}),
|
|
435
|
+
})),
|
|
436
|
+
allowOther: input.allow_other ?? false,
|
|
437
|
+
});
|
|
438
|
+
}
|
|
439
|
+
const input = parseArgs(linksArgs, args);
|
|
440
|
+
return buildLinksUi({
|
|
441
|
+
...(input.title ? { title: input.title } : {}),
|
|
442
|
+
links: input.links.map((link) => ({
|
|
443
|
+
kind: link.kind,
|
|
444
|
+
id: link.id,
|
|
445
|
+
...(link.path ? { path: link.path } : {}),
|
|
446
|
+
...(link.project_root ? { projectRoot: link.project_root } : {}),
|
|
447
|
+
...(link.label ? { label: link.label } : {}),
|
|
448
|
+
})),
|
|
449
|
+
});
|
|
450
|
+
}
|
|
451
|
+
function errorCode(err) {
|
|
452
|
+
if (err instanceof GuideUiError || err instanceof GuideCardError)
|
|
453
|
+
return err.code;
|
|
454
|
+
if (err instanceof EvalApiError)
|
|
455
|
+
return err.status === 404 ? "guide.ui_not_found" : err.status === 409 ? "guide.ui_ambiguous" : "guide.ui_read_failed";
|
|
456
|
+
if (isAgentError(err) || err instanceof MemoryError)
|
|
457
|
+
return /not_found/i.test(err.code) ? "guide.ui_not_found" : "guide.ui_read_failed";
|
|
458
|
+
return "guide.ui_read_failed";
|
|
459
|
+
}
|
|
460
|
+
export async function buildGuideUi(deps, args) {
|
|
461
|
+
const component = args.component;
|
|
462
|
+
try {
|
|
463
|
+
let ui;
|
|
464
|
+
if (component === "pipeline")
|
|
465
|
+
ui = await pipelineUi(deps, args);
|
|
466
|
+
else if (component === "resource")
|
|
467
|
+
ui = await resourceUi(deps, args);
|
|
468
|
+
else if (component === "eval_report")
|
|
469
|
+
ui = await evalReportUi(deps, args);
|
|
470
|
+
else if (component === "plan" || component === "choices" || component === "links")
|
|
471
|
+
ui = plainUi(args, component);
|
|
472
|
+
else
|
|
473
|
+
return fail("guide.ui_unknown_component");
|
|
474
|
+
return { ok: true, content: { ui } };
|
|
475
|
+
}
|
|
476
|
+
catch (err) {
|
|
477
|
+
return fail(errorCode(err));
|
|
478
|
+
}
|
|
479
|
+
}
|
|
480
|
+
const DESCRIPTION = `Show a card in your reply. Pass component and its fields:
|
|
481
|
+
- pipeline: { component: "pipeline", path, task_path?, title? } shows a pipeline read from disk (path relative to the project root).
|
|
482
|
+
- resource: { component: "resource", kind: agent|trigger|eval|connection|memory|task, id, project_root?, title? } shows one saved thing. For a task, id is its path or its id.
|
|
483
|
+
- plan: { component: "plan", title, steps: [{ title, detail?, status: todo|current|done }] } (up to ${MAX_PLAN_STEPS} steps).
|
|
484
|
+
- choices: { component: "choices", question, options: [{ id, label, description? }], allow_other? } asks the person to pick (up to ${MAX_CHOICES} options). options, steps, and links are arrays of objects, never a JSON string.
|
|
485
|
+
- eval_report: { component: "eval_report", eval_run_id } shows an eval run's result.
|
|
486
|
+
- links: { component: "links", title?, links: [{ kind, id, path?, project_root?, label? }] } (up to ${MAX_LINKS}).
|
|
487
|
+
Text is at most ${MAX_TEXT} characters. The card shows what is on disk, so do not repeat it in prose.`;
|
|
488
|
+
const closed = { additionalProperties: false };
|
|
489
|
+
const uiText = Type.String({ minLength: 1, maxLength: MAX_TEXT });
|
|
490
|
+
const showUiParameters = Type.Union([
|
|
491
|
+
Type.Object({
|
|
492
|
+
component: Type.Literal("pipeline"),
|
|
493
|
+
path: uiText,
|
|
494
|
+
task_path: Type.Optional(uiText),
|
|
495
|
+
title: Type.Optional(uiText),
|
|
496
|
+
}, closed),
|
|
497
|
+
Type.Object({
|
|
498
|
+
component: Type.Literal("resource"),
|
|
499
|
+
kind: Type.String({ minLength: 1 }),
|
|
500
|
+
id: uiText,
|
|
501
|
+
project_root: Type.Optional(uiText),
|
|
502
|
+
title: Type.Optional(uiText),
|
|
503
|
+
}, closed),
|
|
504
|
+
Type.Object({
|
|
505
|
+
component: Type.Literal("plan"),
|
|
506
|
+
title: uiText,
|
|
507
|
+
steps: Type.Array(Type.Object({
|
|
508
|
+
title: uiText,
|
|
509
|
+
detail: Type.Optional(uiText),
|
|
510
|
+
status: Type.Union([Type.Literal("todo"), Type.Literal("current"), Type.Literal("done")]),
|
|
511
|
+
}, closed), { minItems: 1, maxItems: MAX_PLAN_STEPS }),
|
|
512
|
+
}, closed),
|
|
513
|
+
Type.Object({
|
|
514
|
+
component: Type.Literal("choices"),
|
|
515
|
+
question: uiText,
|
|
516
|
+
options: Type.Array(Type.Object({
|
|
517
|
+
id: uiText,
|
|
518
|
+
label: uiText,
|
|
519
|
+
description: Type.Optional(uiText),
|
|
520
|
+
}, closed), { minItems: 1, maxItems: MAX_CHOICES }),
|
|
521
|
+
allow_other: Type.Optional(Type.Boolean()),
|
|
522
|
+
}, closed),
|
|
523
|
+
Type.Object({
|
|
524
|
+
component: Type.Literal("eval_report"),
|
|
525
|
+
eval_run_id: uiText,
|
|
526
|
+
}, closed),
|
|
527
|
+
Type.Object({
|
|
528
|
+
component: Type.Literal("links"),
|
|
529
|
+
title: Type.Optional(uiText),
|
|
530
|
+
links: Type.Array(Type.Object({
|
|
531
|
+
kind: Type.Union(GUIDE_LINK_KINDS.map((kind) => Type.Literal(kind))),
|
|
532
|
+
id: uiText,
|
|
533
|
+
path: Type.Optional(uiText),
|
|
534
|
+
project_root: Type.Optional(uiText),
|
|
535
|
+
label: Type.Optional(uiText),
|
|
536
|
+
}, closed), { minItems: 1, maxItems: MAX_LINKS }),
|
|
537
|
+
}, closed),
|
|
538
|
+
]);
|
|
539
|
+
export function createShowUiTool(deps) {
|
|
540
|
+
return {
|
|
541
|
+
name: "show_ui",
|
|
542
|
+
description: DESCRIPTION,
|
|
543
|
+
parameters: showUiParameters,
|
|
544
|
+
handler: (args) => buildGuideUi(deps, args),
|
|
545
|
+
};
|
|
546
|
+
}
|
|
@@ -0,0 +1,48 @@
|
|
|
1
|
+
export type GuideAttachmentMeta = {
|
|
2
|
+
name: string;
|
|
3
|
+
size: number;
|
|
4
|
+
mediaType: string;
|
|
5
|
+
};
|
|
6
|
+
export type GuidePart = {
|
|
7
|
+
type: "text";
|
|
8
|
+
text: string;
|
|
9
|
+
} | {
|
|
10
|
+
type: "tool";
|
|
11
|
+
id: string;
|
|
12
|
+
name: string;
|
|
13
|
+
status: "complete" | "error";
|
|
14
|
+
args?: Record<string, unknown>;
|
|
15
|
+
result?: unknown;
|
|
16
|
+
error?: string;
|
|
17
|
+
};
|
|
18
|
+
export type GuideMessage = {
|
|
19
|
+
id: string;
|
|
20
|
+
role: "user" | "assistant";
|
|
21
|
+
text: string;
|
|
22
|
+
createdAt: string;
|
|
23
|
+
attachments?: GuideAttachmentMeta[];
|
|
24
|
+
parts?: GuidePart[];
|
|
25
|
+
};
|
|
26
|
+
export type GuideThreadRecord = {
|
|
27
|
+
version: 1;
|
|
28
|
+
id: string;
|
|
29
|
+
title: string;
|
|
30
|
+
createdAt: string;
|
|
31
|
+
updatedAt: string;
|
|
32
|
+
messages: GuideMessage[];
|
|
33
|
+
model?: string;
|
|
34
|
+
};
|
|
35
|
+
export type GuideThreadSummary = {
|
|
36
|
+
id: string;
|
|
37
|
+
title: string;
|
|
38
|
+
updatedAt: string;
|
|
39
|
+
};
|
|
40
|
+
export declare function guideThreadsDir(storeRoot?: string): string;
|
|
41
|
+
export declare function guideThreadFilePath(threadId: string, storeRoot?: string): string;
|
|
42
|
+
export declare function parseGuideThread(value: unknown): GuideThreadRecord | null;
|
|
43
|
+
export declare function createGuideThread(storeRoot?: string): GuideThreadRecord;
|
|
44
|
+
export declare function readGuideThread(threadId: string, storeRoot?: string): GuideThreadRecord | null;
|
|
45
|
+
export declare function saveGuideThread(record: GuideThreadRecord, storeRoot?: string): GuideThreadRecord;
|
|
46
|
+
export declare function setGuideThreadModel(threadId: string, model: string | null, storeRoot?: string): GuideThreadRecord | null;
|
|
47
|
+
export declare function listGuideThreads(storeRoot?: string): GuideThreadSummary[];
|
|
48
|
+
export declare function guideTitleFrom(text: string): string;
|