isis library

The TypeScript client SDK (layer:clients, libs/isis/client): IsisClient / createClient / createClientFromEnv with sub-clients such as generation (textToImage, waitForJob). The typed entry point external callers use to submit jobs and poll results.

The Isis persistence layer (layer:data, libs/isis/database): the Prisma client wrapper (createPrismaClient, transaction, executeRaw, loadGeneratedPrismaClientModule) plus stable schema-aligned domain types. The generated client makes this one of the larger libraries by line count.

The searchable 3D asset marketplace library (libs/isis/3d-asset-library, ~27k LOC): a module-blueprint manifest plus community, core, provenance, style, and variation sub-modules for discovery, contracts, and asset blueprints. The catalog/discovery infrastructure for the 3D marketplace.

Browser-first 3D viewing/splatting/editing/inference contracts (libs/isis/3d-browser, ~27k LOC): a module/blueprint manifest plus a runtime capability assessor (assessBrowser3DRuntime) that reads navigator/WebGPU adapter support snapshots and picks a runtime class with a fallback backend. The browser-side delivery surface for 3D.

A typed catalog of ComfyUI 3D nodes and workflow templates (libs/isis/3d-comfyui-nodes): node-family descriptors, conditioning modes, canonical workflow stage definitions, and a Hunyuan3D local runtime descriptor (ThreeDComfyUiNodeFamilyDescriptor, ThreeDComfyUiHunyuanLocalRuntime). The 3D counterpart to @isis/workflows/@isis/comfyui-sdk.

3D ControlNet / reference-control contracts (libs/isis/3d-control): IP-Adapter LoRA variant training (dataset manifests, training plans/gates, runIsis3dIpAdapterLoraVariantTraining) and a face-identity-preservation plan for multi-view generation. Plans conditioned/controlled multi-view 3D.

The text/image-to-3D generation core (libs/isis/3d-generation, ~56k LOC): request/result types and provider abstractions for text-to-3D and image-to-3D with post-processing (TextTo3DRequest, ImageTo3DRequest, Generated3DModel, QualityPreset, Model3DFormat). One of the substantial implementations, not a thin contract.

The 3D-generation benchmarking suite (libs/isis/3d-generation-benchmarks): standardized image and prompt suites with minimum counts and reference-image profiles, an automated benchmark runner with concurrency limits, and result storage versioning. Measures pipeline quality against fixed suites.

The self-hosted 3D inference kernel (libs/isis/3d-inference-local, ~76k LOC): config and runtime contracts for serving 3D models on local GPUs — server/device config, a GPU memory manager with pressure levels and allocation results, queue policy, model cache policy, and dispatch envelopes (createLocal3DGpuMemoryManagerConfig, Local3DDispatchEnvelope). The largest 3D library and the local execution target the contract libraries plan against.

Marketplace trust and IP-governance ops (libs/isis/3d-marketplace-ops, ~2.2k LOC over many small modules): seller/publisher verification workflows (planSellerVerificationWorkflow), license declaration, and listing controls. Bootstrap-manifest style — the seller-operations side of 3D commerce.

Unified 3D post-processing orchestration (libs/isis/3d-post-pipeline, ~31k LOC): a module-blueprint manifest plus Mawu-Studio asset import/attribution contracts (import receipts, mapped materials, pipeline diagnostics, eval fixtures/reports). Stitches post-processing stages after generation and bridges to Mawu Studio.

Prompting and product-parity primitives (libs/isis/3d-product-parity, ~2k LOC): a prompt helper with scored dimensions (create3DPromptHelper) and a structured prompt builder targeting topology/export targets (buildStructured3DPrompt), plus review-flow primitives. Input-quality assistance for 3D product workflows.

Release-gate taxonomy and review governance (libs/isis/3d-quality-gates): a failure-taxonomy with count3DQualityFailureDefinitions, a gold-standard benchmark corpus with coverage summaries and acceptance/approval records, and corpus/operations/taxonomy blueprints. The quality-gate vocabulary for 3D output.

Semantic scene assembly (libs/isis/3d-scene-assembly, ~1.9k LOC): a semantic scene graph with nodes/connectors/constraints, material bindings, and provenance (createThreeDSemanticSceneGraph, ThreeDSceneConstraintKind), backing kitbash assembly and placement validation. Bootstrap-manifest style with a real scene-graph core.

Semantic part decomposition and localized editing (libs/isis/3d-semantic-editing): semantic part nodes/assemblies with masks and confidence, cross-format part metadata read/write for glTF/USDA/blend payloads (readLocal3DSemanticMetadataFromGltfJson/…Usda/…BlendPayload), and assembly validation. Bridges semantic part data across DCC formats.

SV3D-style video-diffusion contracts for image-conditioned multi-view reconstruction (libs/isis/3d-video-diffusion): custom camera-trajectory construction with emphasis frames/regions, and a video-diffusion ↔ mesh-transformer hybrid generator (generateIsisVideoDiffusionMeshTransformerHybrid).

Advanced 3DGS production contracts (libs/isis/3dgs-advanced): GI-GS global illumination (G-buffer/decomposition plans), MiLo mesh-in-loop extraction, reflective-GS integration, SplatPainter interactive editing, and confidence-based mesh extraction, over a shared common module with scene/camera/light validators. Production-grade splatting techniques as typed plans.

Diffusion-guided 3DGS editing (libs/isis/3dgs-diffusion-editing): a GaussCtrl model-integration plan with depth conditioning, latent alignment, cross-view consistency, and propagation graphs (createIsis3DGSGaussCtrlModelIntegrationPlan). Plans diffusion-driven edits over a splat scene.

Material-decomposed 3DGS relighting (libs/isis/3dgs-relighting-pbr): a LumiGauss model loader with geometry/material/transfer-coefficient buffer layouts and integrity checks, plus a LumiGauss training pipeline (createIsisLumiGaussModelLoaderPlan, createIsisLumiGaussTrainingPipelinePlan). Decomposes splats into PBR materials for relighting.

The unified AI-provider layer and the area's largest library (~174k LOC). Its barrel (libs/isis/ai-providers/src/index.ts) exports typed adapters and factories for LLMs (Anthropic Claude, OpenAI GPT, Google Gemini, xAI Grok, Ollama via createClaude/createGPT/…), ElevenLabs TTS and Agents Platform 2.0 conversational AI, the Civitai model registry, RunComfy serverless execution, a usage/cost/budget tracker, and a caching AIProviderFactory. The provider spine the rest of Isis calls models through.

AI-native PBR texturing (libs/isis/ai-texturing, ~28k LOC): capability builders for AI-based texture baking, Hunyuan3D-Paint image-guided PBR, and PBR-aware super-resolution, denoising, detail synthesis, and compression (createLocal3DHunyuan3DPaintImageGuidedPbrCapabilities, …). Generates and refines material maps for generated meshes.

Anomaly detection and fraud prevention for the factory (libs/isis/anomaly-detection): an AnomalyDetectionService for suspicious activity, chargeback prediction, and account protection (RiskLevel, DetectionStatus, ActionType). A compact three-file service over rule/risk types.

AI audio generation (libs/isis/audio-generation, ~16k LOC): music synthesis, voice synthesis, and sound-effects generation with shared audio-format/metadata types (MusicGenerationRequest, VoiceID, audio format/sample-rate/bit-depth types). The general-purpose audio factory surface.

Joint audio-video narrative production for ISIS and Yemaya (libs/isis/av-narrative-studio, a single ~2.2k-line module): AV stems (dialogue/music/effects/ambient) sourced from Seedance joint-AV or editor generation, narrative shot/transition types, narrative templates, and ProRes/ DNxHR + broadcast-WAV export. Orchestrates synchronized AV storytelling.

The ComfyUI workflow factory (§24.3, libs/isis/comfyui-factory, ~430 LOC): an approved workflow-class catalog with guardrails, template-diff, portability-check, and a rehearsal-harness. Compact but real — it governs and diffs the approved class set rather than executing graphs.

A Python package (lang:python, libs/isis/comfyui-nodes/src/isis_comfyui_nodes) of real ComfyUI custom nodes backing Oshun's spiritual/sacred-art workflows — consciousness, sacred_geometry, lilith (asset I/O + metadata), spiritual_styles, and vfx_post node sets, with pytest suites (tests/test_vfx_post_nodes.py, tests/test_lilith_nodes.py) and example workflow JSON. The only Python node library in the area; it has no TS barrel.

The "ComfyUI Nodes 2.0" SDK (libs/isis/comfyui-sdk): a typed connection system with branded socket types (INT/FLOAT/IMAGE), node versioning with migration paths, node discovery/introspection, a validation system, and an auto-generating testing/benchmark framework (NodeSpec, createNodeValidator). For building ComfyUI nodes safely.

Curated creator generation cards (§24.6, libs/isis/curated-cards): card types, validators, a preflight checker, an entitlement gate, and a catalog (card-types/card-validators/preflight/entitlement-gate/catalog). Cards are bound to approved workflow classes so curated creators get safe, pre-vetted generation presets.

The generation-tier resolver and surface allowlist (§24.1, libs/isis/entitlements): generation-tier.ts deterministically maps an entitlement bundle to a tier (operator-admin / aaa-creator / curated-creator / contemplative) and a surface allowlist, and studio-boundary.ts enforces the studio edge. The documented single source of truth BFF/Admin/Studio middleware consult before rendering any generation surface.

AI face synthesis, de-aging, and identity-transfer orchestration (libs/isis/face-synthesis): notably an ethical-safeguards planner with consent records, watermark modes, safeguard alerts/decisions, usage contexts, and audit event plans (FaceSynthesisEthicalSafeguardsPlanner). Consent/safeguard-first by construction.

Fast-3D contracts for rectified-flow and consistency-model workflows (libs/isis/fast-generation): a consistency-mesh-transformer pipeline with a real runIsisConsistencyMeshTransformerPipeline, latency budgets, and a generation-speed benchmark with GPU-tier/method profiles. Plans and benchmarks low-latency 3D generation.

Foley and sound-design orchestration for ISIS and Yemaya (libs/isis/foley-studio, a single ~1.3k-line module): detected foley events typed by kind/material/intensity, provider options (foleycrafter, av-link, any2audio), DAW targets, and AAF/OMF/BWF-P export. Production foley as a typed workflow.

The end-to-end 3D Gaussian-Splatting pipeline (libs/isis/gaussian-splatting, ~31k LOC): real multi-camera capture (CaptureManager, FileSystemCameraProvider, SharpFrameQualityAnalyzer using sharp), preprocessing, training, rendering, and optimization. A substantial implemented pipeline, not just types.

Next-generation 3D-model integration contracts (libs/isis/gigascale-3d): typed profiles and readiness evaluation for large external models — a CHORD PBR pipeline (createIsisChordPbrGenerationPlan, evaluateIsisChordPbrReadiness) and Direct3D-S2 (DIRECT3D_S2_MODEL_ID), with license-policy and runtime device/OS modelling. Describes how gigascale models slot into Isis.

Multi-resolution hash-grid encoding for neural fields (libs/isis/hash-encoding): streaming level selection/update plans and an export path that serializes hash-encoded meshes/textures (e.g. serializeIsisHashEncodedMeshToObj, exportIsisHashEncodedMesh). The Instant-NGP-style encoding contracts plus real OBJ/texture export.

The LoRA training surface (§24.5, libs/isis/lora-training-surface, ~720 LOC): training-run, model-merging, quality-view, and lineage-tree modules — the training queue, model merging, tuning rehearsal, quality scoring, and lineage tracking for creator-trained LoRAs.

A thin but real registry-browser surface (libs/isis/managed-models, ~74 LOC): the barrel re-exports ./browser/index, exposing only post-intake approved models to the creator UI (§24.2). Small by design — it is a governed view over the model registry, not a model store itself.

Autoregressive mesh-generation transformer contracts (libs/isis/mesh-transformers): MeshGPT architecture (VQ-VAE tokenizer + decoder, face-token sequences), vocabulary learning, and the broader MeshGPT/PolyGen/MeshXL family (createIsisMeshGptArchitecturePlan, IsisMeshGptVqVaeTokenizerConfig). Typed model architectures and training objectives for mesh tokenization.

License/deployment/compliance governance for 3D models (libs/isis/model-governance-3d): an exception-approval workflow with audit entries (create3DModelGovernanceExceptionApproval, evaluate3DModelGovernanceExceptionApproval) and near-duplicate checking. The compliance gate before a 3D model is usable.

Interactive motion and camera-control orchestration for ISIS and Yemaya (libs/isis/motion-camera-control, a single ~835-line module): camera-trajectory presets (orbit/dolly/crane/…), easing curves, and motion validation that flags teleportation, interpenetration, impossible acceleration, and gravity violations (MotionValidationIssueKind).

Music + ambient generation (§24.8, libs/isis/music-generation, ~540 LOC): a provider abstraction, guardrails, provenance, and stems modules. Compact — the governed, watermark/provenance-aware music surface layered over a provider seam.

The unified output gallery surface (§24.10, libs/isis/output-gallery): output-record, lineage, branch-replay, compare-grid, and bulk-actions modules — filters, lineage, branch/replay, side-by-side compare, and bulk operations over generated outputs.

The output manifest and provenance layer (libs/isis/outputs): the OutputFile/ manifest schema, provenance records, and lineage-graph utilities (ManifestId, ProvenanceId, LineageNodeId, QualityTier, FileHash). Describes and tracks the artifacts a generation job produces.

Part-level, PartGen-style compositional 3D generation (libs/isis/part-level-3d): a part-quality metrics computer (computeIsisPartGenerationQualityMetrics) and a text-driven scene-composition planner (composeIsisSceneFromText, createIsisScenePartGenObjectGenerator). Generates and scores compositional multi-part assets.

Virtual-presenter and performance orchestration for ISIS and Yemaya (libs/isis/portrait-studio, a single ~1.2k-line module): inference providers (Hallo3/LivePortrait/MuseTalk/RAP), voice providers, presenter emotion/camera angles, lip-sync quality gates (syncnet/face-embedding), and multi-language dubbing types.

Per-production post-production intelligence (libs/isis/post-production-ai, ~31k LOC): a library manifest of analysis capabilities, a dailies-ingestion pipeline with staged timing and a media-probe seam (DailiesIngestionPipeline, NoopDailiesMediaProbe as the fail-safe default), and per-production model training/orchestration for AI-assisted finishing.

Quad-dominant mesh generation (libs/isis/quad-mesh-gen): a QuadGPT-style LOD-chain generator with real edge-collapse ordering (createIsisQuadMeshEdgeCollapseOrder, generateIsisQuadMeshLodChain) and a game-ready pipeline with procedural PBR synthesis, boundary-seam prediction, and engine exporters (runIsisQuadMeshGameReadyPipeline).

Generative relighting workflow orchestration for ISIS and Yemaya (libs/isis/relight-studio, a single ~2.4k-line module): light types/falloff, preview engines (ic-light/genlit/lightlab), lighting presets (golden-hour/rembrandt/butterfly/…), comparison modes, batch scopes, and NLE/comp export hosts.

The RunPod operator surface data model (§24.4, libs/isis/runpod-surface): endpoint-registry, cost-quota, queue-inspector, and secret-rotation modules — the typed dashboard model for managing serverless GPU endpoints.

Scene-from-single-image composition (block B of ISIS_GAPS, libs/isis/scene-from-image-composer, ~9.3k LOC across 66 non-test files). A substantial, fully-wired pipeline: the barrel exports working composeSceneFromImage / createSceneFromImagePipeline entry points plus a runPhases orchestrator over nine ordered phases (environment/grounding/detection/objects/align/bake/coherence/audio/manifest), alongside a real align/ module (alignMeshToSplatFrame, anchor resolution with UnknownAnchorError/UnresolvableAnchorsError), object cropping, detection, and a mesh-provider router.

Score Distillation Sampling for text-to-3D (libs/isis/score-distillation): genuine deterministic SDS math — computeIsisSdsLoss, addIsisSdsNoise, Gaussian-noise and linear noise-schedule builders, parameter-update and optimization-step helpers (IsisSdsLossResult, IsisSdsNoiseSchedule). Real loss utilities, not just descriptors.

The Seedance 2.0 provider (libs/isis/seedance-provider, ~22k LOC): multi-backend provider interfaces and exhaustive typed configuration — resolution/aspect/ duration/motion options, reference-asset conditioning (image/video), and a capability matrix (SeedanceCapabilityMatrix, SeedanceReferenceAssetInput). A deep single-vendor integration surface.

Text-guided mesh editing (libs/isis/text-mesh-editing): multi-view InstructPix2Pix/SDS edit transfer (transferIsisTextMeshEditToBatch) and a preview/undo workflow with an undo stack (previewIsisTextMeshEdit, undoIsisTextMeshEdit, createIsisTextMeshEditUndoStack). Plans localized, instruction-driven mesh edits.

The 3D pipeline-class catalog and provider abstraction (§24.9, libs/isis/three-d-pipelines): provider, pipeline-class, provenance, topology, a sky-event-card, and Mawu-Studio creation modules. The Mawu product path dispatches text-to-3d, text-to-texture, text-to-world, and text-to-npc requests through separately configured, fail-loud HTTP routes. It bounds and validates provider responses, recomputes tool-specific usability, and runs successful output through semantic grounding, attribution, provenance, content hashing, and prompt-output trust checks. Its local reference backend is explicitly procedural.

The V7 operational-evidence module inventories all ten headline execution boundaries: Crucible tabular-Q and match-meta, Loom reference vision and local terrain ML, the four distinct Mawu product routes, Themis dispatch, and full-engine parity. A catalog is ready only with exactly one fresh, digest-bound record per boundary. Each record must match its fixed implementation class and include exact implementation/model/dataset/output license identity, run/artifact provenance, sampled USD cost and latency, passing safety scenarios, threshold-derived quality metrics, and an exercised provider failover, deterministic retry, or engine restart inside its recovery objective with zero data loss. Model-free and dataset-free dispositions require an explicit reason; model-backed routes cannot use them. This is an admission contract, not live evidence: the V7 operational-evidence task remains open until real retained records pass it.

The bounded V7 flagship fixture composes all four governed reference artifacts into region:mawu-market-quarter:v1, a 128 m square procedural evidence region. Its independent evaluator derives completeness from exact terrain coverage, in-bounds source-linked assets, scheduled NPC roles, zero-sum economy transactions, an acyclic asset-targeted and reward-bound quest graph, complete English/Spanish/French message keys, and the V7 accessibility floor. Its deterministic simulator derives terrain reachability, asset navigation, quest completion, and ledger integrity from the sealed manifest. The first 32-session replay exposed one unreachable asset; an exact-replay policy moved it through a system-authored, full-auto common run and the after replay passed. Both the simulation and revision records are digest-bound and independently validate nested common-run and output evidence. This proves a coherent procedural manifest and governed headless revision, not a UE cook, product-model generation, player telemetry, playtest, human approval, or canary.

Phase 85–86 flywheel producer for generative media (libs/isis/training-data/src): IsisTrainingDataPipeline covers the widest signal surface of the flywheel — sixteen IsisTrainingKinds from prompt-quality, generation-preference, and lora-effectiveness to three-d-quality, gaussian-splat-quality, vfx-pair, and llm-orchestration. Every record is gated by an explicit IsisRightsGrant (rights-based rather than plain consent, reflecting creator content) before it reaches the IsisTrainingSink.

Topology-agnostic auto-rigging (libs/isis/universal-rigging, ~48k LOC and one of the largest): a broad export surface covering ML skeleton prediction, automatic skin-weight computation, animation clip core / blending-transition graphs / export / preview rendering, bone-hierarchy inference, and species/creature rig templates (bird, fish-marine, facial, eye, clothing, corrective blend shapes, cross-species retargeting). A substantial rigging foundation, not a contract stub.

Production video-editing orchestration for ISIS and Yemaya (libs/isis/video-edit-studio, a single ~970-line module): inpaint/outpaint/ restyle/extend/reference-to-video tasks with masks, reference kinds, aspect ratios, comparison modes, and versioning (VideoEditTaskMode, VideoEditVersionStatus).

AI video enhancement/restoration (libs/isis/video-enhancement, ~19k LOC): a capabilities manifest plus planners for super-resolution, frame interpolation, denoising, grain management, and cadence correction/deinterlacing (CadenceCorrectionPlanner, with pull-down patterns and field-dominance modes).

Production object-removal orchestration for ISIS and Yemaya (libs/isis/video-object-removal, a single ~3.2k-line module): selection tools (point/box/lasso/text-prompt), quality tiers, comparison modes, cross-shot subject tracking, and a set of domain pipelines (product-placement, safety-wire-rig, anachronism-removal, weather-element, …) with NLE-host targets.

Video-first 3D reconstruction (libs/isis/video-to-mesh, ~29k LOC): capability builders for monocular video depth estimation, camera-pose estimation, multi-frame depth fusion, point-cloud surface reconstruction, texture projection, and object segmentation (createLocal3DVideoToMesh*Capabilities). A full video→mesh pipeline contract set.

Visual dubbing and dialogue-editing orchestration (libs/isis/visual-dubbing, ~7k LOC): a voice-cloning visual-synchronization planner and an emotional-regrade planner with strategies/review-modes/quality-gates (VoiceCloningVisualSynchronizationPlanner, EmotionalRegradePlanner). Aligns re-voiced dialogue to on-screen performance.

Professional voice cloning and synthesis (libs/isis/voice-cloning): a voice-profile creation planner (dataset prep steps, source types, training strategies, quality gates) plus a capabilities/workflow manifest (VoiceProfileCreationPlanner, createVoiceCloningLibraryManifest). Plans the clone-and-synthesize pipeline.

The Living-Scene workflow-class catalog (§25.5, libs/isis/workflow-classes): five domain templates — tara, nyx, veritas, metis, arete — plus a template-catalog, locale-parity, and fixture eval gates. Curated, gated workflow blueprints rather than free-form graphs.

ComfyUI workflow definitions and tooling (libs/isis/workflows): the Workflow schema with validation, a versioning/migration pipeline (WorkflowSchemaMigration), a template DSL, and seed management. The typed representation of the ComfyUI graphs Isis runs.

Provider-registry infrastructure (libs/isis/llm-providers): model/provider registration, rate limiting, token counting, cost calculation, and health/ circuit-breaker tracking (ProviderRegistryConfig, RateLimiterConfig, TokenCounterConfig, CostCalculatorConfig, CircuitBreakerState). It governs how providers are selected and throttled, complementing @isis/ai-providers' concrete adapters.

Fine-tuning orchestration (libs/isis/model-fine-tuning): hyperparameter search trials, dataset curation/deduplication and quality stats, experiment tracking, model-merge jobs, benchmark runs, and metric snapshots (HpTrial, CuratedDataset, DeduplicationResult, MergeJob, MetricSnapshot).

The canonical Isis job envelope (libs/isis/job-envelope): the versioned queue schema plus exhaustive generation-type, ControlNet, IP-Adapter, and InstantID mode manifests with legacy-workflow alias maps (ISIS_JOB_ENVELOPE_VERSION, ISIS_GENERATION_TYPE_MANIFEST, ISIS_CONTROLNET_MAJOR_MODES, …). Every Isis generation job is shaped by this contract.

Multi-agent reasoning and consensus (libs/isis/agent-consensus): structured debate with claims/arguments/evidence, reasoning chains, fallacy detection, convergence sessions, and compromise proposals (Debate, Argument, DetectedFallacy, ConvergenceSession, CompromiseProposal, ReasoningQualityMetrics). Models how several agents argue toward an agreed answer.

Batch LLM pipeline (libs/isis/batch-llm-processing): batch request/job modelling, scheduling within time windows, result correlation, cost reporting, and throughput snapshots (BatchRequest, ScheduleDecision, BatchJobResult, CostReport, ThroughputSnapshot, plus ToolDefinition/ChatMessage). For high-volume, cost-aware LLM workloads.

The Isis event publisher (libs/isis/event-publisher): IsisEventPublisher plus a typed payload set for the job lifecycle (queued/started/progress/completed/ failed/cancelled) and asset/workflow/model events (IsisJobCompletedPayload, IsisAssetGeneratedPayload, IsisWorkflowRegisteredPayload). How Isis announces work onto the bus.

Structured-LLM orchestration (libs/isis/llm-orchestrator): typed extraction, schema validation, repair of malformed model output, budget framing, and an operation queue (ExtractionResult, SchemaValidationConfig, RepairConfig, BudgetTrackerConfig, OperationEntry, QueueStats). It turns raw model calls into validated, schema-conformant results.

Operation orchestration (libs/isis/operation-orchestrator): retry policies with failure classification, batch composition, dead-letter handling, execution chains/phases, and pipeline execution with checkpoints (RetryPolicy, FailureClassification, DeadLetterEntry, ExecutionChain, PipelineExecution). The reliability layer for multi-step generation jobs.

Prompt toolkit (libs/isis/prompt-engineering): versioned prompt templates with variables/sections, few-shot example selection, chain-of-thought construction and execution, role definitions, and prompt evaluation/scoring (PromptTemplate, ChainOfThought, FewShotExample, PromptEvaluation, PromptVersion).

A ReAct (reason+act) agent framework (libs/isis/react-framework): the thought→action→observation step loop with execution traces, tool definitions and results, and trace analysis/metrics (ReActStep, ExecutionTrace, ToolDefinition, ThoughtChainAnalysis, ActionPatternAnalysis). The agent-loop substrate for tool-using LLM agents inside Isis.

Token-budget engine (libs/isis/token-budget): per-tenant budgets with category allocations, cost prediction from ModelPricing, enforcement rules, alerts, rebalancing/optimization, and provider fallback decisions (TokenBudget, EnforcementCheckResult, CostPrediction, FallbackDecision). The accounting layer that keeps generation spend bounded.