version: 2 adapters: anthropic: implementation: kind: package package: '@moldea.ai/adapter-anthropic' distribution: public versionRange: '^2.0.0' implementationStatus: available supportedRepositoryFormatVersions: - 1 compatibleCoreRange: '^2.0.0' runtimeGuidance: expectation: optional notes: Project-local guidance is needed only for repository-specific wrappers or unsupported indirect integration patterns. targets: - id: typescript-messages-api-0-117 kind: package language: typescript packages: - ecosystem: npm name: '@anthropic-ai/sdk' role: primary versionRange: '>=0.117.1 <0.118.0' evidenceKinds: - instruction-loader - language - runtime-package - runtime-pattern - schema - tool-registration bindingSupport: runtime-agent: relationship: full symbol: full instruction-loader: relationship: full symbol: full tool-registration: relationship: full symbol: full tool-input-schema: relationship: full symbol: full patterns: - id: direct-messages-create kind: runtime support: full description: Direct Anthropic Messages API invocation through a module-local client in a directly exported TypeScript function. - id: direct-system-loader kind: instruction-loader support: full description: A directly bound instruction loader supplies the top-level system request property. - id: closed-client-tool-array kind: tool support: full description: Closed inline or immutable module-local arrays contain statically declared Anthropic client tools. - id: direct-tool-input-schema kind: schema support: full description: A bound tool input schema is referenced directly through the client tool input_schema property. - id: dynamic-request-construction kind: runtime support: ambiguous description: Dynamically assembled Messages requests cannot be mapped reliably without semantic analysis. - id: provider-server-tools kind: tool support: unsupported description: Anthropic provider or server tools are outside the initial client-tool target. providerLimits: - id: client-tool-name subject: tool-name kind: pattern value: '^[A-Za-z0-9_-]{1,64}$' description: Anthropic client-tool names match the complete ASCII letter, digit, underscore, and hyphen pattern. reference: Anthropic Messages API reference for client tools. knownLimitations: - Arbitrary compiler resolution, path aliases, directory indexes, package exports, and re-export graphs are not resolved. - Beta resources, client.messages.stream, parse helpers, and tool-runner abstractions are not interpreted; an exact stream property on direct messages.create requests is tolerated, but its semantics are not validated. - Client-tool input-schema contents, including the provider-required top-level type object, are not validated; the target establishes only direct schema wiring. - Source forms outside the verified TypeScript ESM target, dynamic factories, mutable requests, provider tools, output schemas, runtime variables, and handoffs are outside the initial target. lastVerifiedAt: '2026-08-17' lastVerifiedAt: '2026-08-17' claude-agent-sdk: implementation: kind: package package: '@moldea.ai/adapter-claude-agent-sdk' distribution: public versionRange: '^1.0.0' implementationStatus: available supportedRepositoryFormatVersions: - 1 compatibleCoreRange: '^2.0.0' runtimeGuidance: expectation: optional notes: Project-local guidance is needed only for repository-specific wrappers, main-thread agent selection, tool aliases, string-array prompts, filesystem-defined agents, dynamic agent construction, observer behavior, external MCP configuration, or other unsupported indirect integration patterns. targets: - id: typescript-query-subagents-0-3 kind: package language: typescript packages: - ecosystem: npm name: '@anthropic-ai/claude-agent-sdk' role: primary versionRange: '>=0.3.234 <0.4.0' evidenceKinds: - agent-definition - handoff-registration - instruction-loader - language - runtime-package - runtime-pattern - schema - tool-registration bindingSupport: runtime-agent: relationship: full symbol: full output-schema: relationship: partial symbol: partial instruction-loader: relationship: full symbol: full tool-implementation: relationship: full symbol: full tool-registration: relationship: full symbol: full tool-input-schema: relationship: full symbol: full patterns: - id: direct-query-wrapper kind: runtime support: full description: A directly exported TypeScript function contains one or more direct query calls in its own lexical body. - id: query-custom-system-prompt kind: instruction-loader support: full description: A query-wrapper agent wires the declared instruction loader through a direct custom systemPrompt call. - id: query-preset-append kind: instruction-loader support: full description: A query-wrapper agent uses the claude_code system-prompt preset and appends the declared canonical instruction loader directly. - id: query-system-prompt-block-array kind: instruction-loader support: unsupported description: String-array system prompts and dynamic prompt-cache boundaries do not establish canonical query-wrapper instruction-loader wiring in the initial target. - id: programmatic-agent-definition kind: agent support: full description: A directly exported immutable object-literal AgentDefinition supplies independently analyzable prompt, routing-description, and tool-restriction relationships. - id: closed-programmatic-agents-map kind: routing support: full description: A closed query agents map exposes supported programmatic subagent definitions under deterministic runtime names only when supported query-local rules classify query-configured Agent availability as available. - id: query-agent-delegation-availability kind: routing support: full description: Supported closed query tools and static bare disallowedTools patterns, including complete-name * globs, classify query-configured Agent availability; dynamic, scoped, legacy-alias, and unsupported non-* forms remain ambiguous. - id: effective-routing-description kind: routing support: full description: For an active programmatic subagent registration, AgentDefinition.description uses the target canonical handoff description when present and the canonical agent-description fallback otherwise. - id: query-json-schema-output kind: schema support: full description: A query-wrapper agent wires a bound output schema through outputFormat with the json_schema type. - id: sdk-mcp-tool-declaration kind: tool support: full description: A directly exported SDK tool uses a static name, direct implementation binding, and direct input-schema binding. - id: sdk-mcp-server-registration kind: tool support: full description: A closed SDK MCP server mounted under a canonical query mcpServers key matching ^[A-Za-z0-9_-]+$ exposes declared tools under fully qualified runtime names only when relationship-local query tool availability remains available after supported static deny-pattern analysis; optional server instructions remain outside canonical instruction validation. - id: sdk-mcp-server-key-normalization kind: tool support: unsupported description: An empty query mcpServers key cannot establish a supported canonical runtime-name segment, while a key containing characters outside [A-Za-z0-9_-] requires SDK normalization; either form produces the stable unsupported-key diagnostic and establishes no runtime-name or tool-availability conclusion for that mount. - id: inherited-subagent-tools kind: tool support: full description: An actively delegable programmatic subagent with omitted tools inherits a query-available SDK MCP tool when supported subagent deny-pattern analysis also leaves the exact tool available. - id: explicit-subagent-tools kind: tool support: full description: An actively delegable programmatic subagent has a closed AgentDefinition tools array containing the exact fully qualified SDK MCP tool name, and supported static deny-pattern analysis leaves the exact tool available. - id: dynamic-query-options kind: runtime support: ambiguous description: Query inputs or options assembled through variables, factories, spreads, mutation, or arbitrary wrappers cannot be mapped reliably without semantic analysis. - id: query-main-thread-agent-selection kind: agent support: unsupported description: Query-level agent selection can apply another definition's prompt and tool restrictions to the main thread and keeps affected instruction, delegation, and tool relationships unresolved. - id: query-tool-aliases kind: tool support: unsupported description: Query-level toolAliases can redirect model-emitted tool names and keep affected delegation and SDK MCP tool relationships unresolved. - id: query-built-in-tools-preset kind: routing support: unsupported description: The claude_code tools preset is not expanded to establish built-in Agent availability in the initial target. - id: dynamic-agent-delegation-availability kind: routing support: ambiguous description: Dynamic query tools or disallowedTools values, legacy Task aliases, scoped Agent or Task permission expressions, and unsupported non-* glob syntax do not establish whether a configured agents map is an active delegation surface. - id: dynamic-tool-availability kind: tool support: ambiguous description: Dynamic or unsupported query or AgentDefinition tool restrictions that could match an SDK MCP tool establish neither positive registration evidence nor a closed negative registration conclusion. - id: dynamic-agent-definition kind: agent support: ambiguous description: Runtime-generated, factory-produced, conditional, spread-based, or mutated programmatic agent definitions remain unestablished. - id: dynamic-routing-description kind: routing support: ambiguous description: Runtime-generated, transformed, or indirectly loaded AgentDefinition descriptions remain unestablished. - id: filesystem-subagents kind: agent support: unsupported description: Subagents defined through .claude/agents files are outside the initial repository-owned programmatic target. - id: built-in-subagents kind: routing support: unsupported description: The built-in general-purpose subagent and runtime Agent tool decisions are not mapped to registered moldea agents. - id: observer-agent-fields kind: routing support: unsupported description: AgentDefinition observer and observerMessage semantics do not create ordinary moldea agent, handoff, routing-description, or instruction-loader relationships in the initial target. - id: experimental-critical-system-reminder kind: instruction-loader support: unsupported description: AgentDefinition criticalSystemReminder_EXPERIMENTAL is additional model-facing content and is not interpreted as canonical subagent instruction wiring. - id: external-mcp-servers kind: tool support: unsupported description: Stdio, SSE, HTTP, remote, proxy, and other external MCP configurations do not establish repository-local manifest tool relationships in the initial target. - id: per-agent-mcp-servers kind: tool support: unsupported description: AgentDefinition-level MCP server configuration is outside the initial query-level SDK MCP target. - id: sdk-mcp-server-instructions kind: instruction-loader support: unsupported description: createSdkMcpServer instructions may coexist with supported tool relationships but are not canonical agent instruction-loader evidence or semantically validated content. - id: skills-plugins-and-hooks kind: runtime support: unsupported description: Skills, plugins, hooks, settings, CLAUDE.md loading, and other filesystem features are outside the initial deterministic relationship target. knownLimitations: - >- Arbitrary compiler resolution, path aliases, directory indexes, package exports, CommonJS, JavaScript, re-export graphs, and generated source are not resolved. - >- Query wrappers require a directly exported function and direct query calls with object-literal input and options forms. - >- Query-level agent selection, toolAliases, string-array system prompts, and built-in-tool preset expansion remain outside the initial target and keep only the relationships they can change unresolved. - >- Programmatic subagents require directly exported immutable object-literal AgentDefinitions; dynamic factories and filesystem-defined agents remain outside the target. - >- A configured agents map becomes active delegation evidence only when supported static query tools and bare tool-name deny-pattern analysis classify query-configured Agent availability as available; tools arrays that omit Agent, tools: [], and any supported deny pattern matching Agent make it unavailable. - >- Dynamic availability, scoped Agent or Task permission expressions, unsupported non-* glob syntax, and legacy Task aliases remain unresolved. Static bare * globs are matched against the complete tool name. allowedTools controls preapproval rather than availability and neither creates nor restores the Agent tool. - >- Routing-description validation supports only static inline, immutable module-local, and directly imported static strings in an active delegation context; loaders, file reads, transformations, and runtime-generated values remain unresolved. - >- AgentDefinition observer, observerMessage, and criticalSystemReminder_EXPERIMENTAL semantics are outside the target, although their presence does not erase independently proved prompt, description, tool, or active-registration relationships. - >- Agent output-schema support applies only to query-wrapper agents through outputFormat; programmatic AgentDefinitions have no initial output-schema relationship. - >- SDK MCP tool support is limited to repository-local tool and createSdkMcpServer definitions mounted through query-level mcpServers maps. - >- A supported query-level mcpServers key must match ^[A-Za-z0-9_-]+$ exactly. Empty or normalization-requiring keys produce CLAUDE_AGENT_SDK_MCP_SERVER_KEY_UNSUPPORTED and keep only the affected mount's runtime-name and tool-availability relationships unresolved; the adapter does not reproduce SDK normalization or infer normalized-key collisions. - >- Query and subagent tool-registration evidence requires an available relationship-local state after supported exact-name, server-selector, and complete-name * glob deny analysis. Dynamic or unsupported restrictions that could match the tool remain unresolved and produce neither optimistic evidence nor a false negative diagnostic. - >- createSdkMcpServer instructions are tolerated but are not canonical instruction-loader evidence or semantically validated model-facing content. - >- The fully qualified manifest tool name is derived from the canonical query mcpServers key and tool name; the createSdkMcpServer name does not replace that key. - >- Tool input and query output schema contents are not validated; the target establishes only direct schema wiring. - >- Tool output schemas, agent input schemas, external MCP tools, resources, prompts, skills, plugins, hooks, sessions, permission approval, sandboxing, workflows, model selection, and provider behavior are not interpreted. - >- Establishing query-configured Agent or SDK MCP tool availability does not prove that filesystem-loaded settings, managed policy, hooks, later session state, allowedTools, canUseTool, permission mode, or user approval preserve or permit a particular invocation, and no evidence claims that Claude will actually invoke the subagent or tool. lastVerifiedAt: '2026-08-19' lastVerifiedAt: '2026-08-19' cloudflare-agents: implementation: kind: package package: '@moldea.ai/adapter-cloudflare-agents' distribution: public versionRange: '^1.0.0' implementationStatus: available supportedRepositoryFormatVersions: - 1 compatibleCoreRange: '^2.0.0' runtimeGuidance: expectation: recommended notes: Project-local guidance should document Cloudflare bindings, Durable Object wiring, and deployment-specific behavior outside the verified static source boundary. targets: - id: typescript-think-0-16-ai-sdk-7 kind: package language: typescript packages: - ecosystem: npm name: '@cloudflare/think' role: primary versionRange: '>=0.16.0 <0.17.0' - ecosystem: npm name: agents role: companion versionRange: '>=0.21.0 <0.22.0' - ecosystem: npm name: ai role: companion versionRange: '>=7.0.0 <8.0.0' evidenceKinds: - agent-definition - handoff-registration - instruction-loader - language - runtime-package - schema - tool-registration bindingSupport: runtime-agent: relationship: partial symbol: partial instruction-loader: relationship: partial symbol: partial tool-implementation: relationship: partial symbol: partial tool-registration: relationship: partial symbol: partial tool-input-schema: relationship: partial symbol: partial tool-output-schema: relationship: partial symbol: partial patterns: - id: directly-exported-think-class kind: agent support: partial description: Directly exported TypeScript classes extending an exact named Think import with closed class initialization. - id: think-instruction-methods kind: instruction-loader support: partial description: Direct loader calls returned by getSystemPrompt or supported closed configureSession chaining. - id: closed-think-tools-map kind: tool support: partial description: Repository-local AI SDK function tools and Cloudflare agentTool helpers active in a closed getTools map. knownLimitations: - Bare Agent classes, factories, indirect subclasses, decorators, executable fields, static blocks, computed members, generators, and non-pass-through constructors are outside the target. - Dynamic session builders, onCompaction interpretation, runtime mutation, channel-provided tool replacement, and open tools maps are outside the target. - Agent input and output schemas are not supported for Think. lastVerifiedAt: '2026-08-19' - id: typescript-ai-chat-agent-0-10-ai-sdk-7 kind: package language: typescript packages: - ecosystem: npm name: '@cloudflare/ai-chat' role: primary versionRange: '>=0.10.2 <0.11.0' - ecosystem: npm name: agents role: companion versionRange: '>=0.21.0 <0.22.0' - ecosystem: npm name: ai role: companion versionRange: '>=7.0.0 <8.0.0' evidenceKinds: - agent-definition - handoff-registration - instruction-loader - language - runtime-package - runtime-pattern - schema - tool-registration bindingSupport: runtime-agent: relationship: partial symbol: partial instruction-loader: relationship: partial symbol: partial output-schema: relationship: partial symbol: partial tool-implementation: relationship: partial symbol: partial tool-registration: relationship: partial symbol: partial tool-input-schema: relationship: partial symbol: partial tool-output-schema: relationship: partial symbol: partial patterns: - id: directly-exported-ai-chat-agent-class kind: agent support: partial description: Directly exported TypeScript classes extending an exact named AIChatAgent import with the supported onChatMessage signature. - id: direct-ai-sdk-generation kind: runtime support: partial description: Direct generateText or streamText calls in the onChatMessage method's own lexical body. - id: ai-chat-structured-output-and-tools kind: tool support: partial description: Output.object agent schemas, repository-local AI SDK function tools, and Cloudflare agentTool helpers in closed generation-request tools maps. knownLimitations: - Nested or indirect generation, request variables, prepareStep instruction interpretation, and generation functions other than generateText and streamText are outside the target. - Output variants other than Output.object and all agent input schemas are outside the target. - Dynamic, provider, MCP-generated, or inline tools without exact repository registration identities are outside the target. lastVerifiedAt: '2026-08-19' lastVerifiedAt: '2026-08-19' custom: implementation: kind: built-in package: '@moldea.ai/core' distribution: public implementationStatus: available supportedRepositoryFormatVersions: - 1 compatibleCoreRange: '^2.0.0' runtimeGuidance: expectation: required notes: Project-local guidance defines the custom runtime integration. targets: - id: custom kind: custom language: any qualificationEvidence: url: 'https://skill.moldea.ai/evidence/qualification/custom/custom/' patterns: - id: explicit-repository-relationships kind: runtime support: full description: Universal Core validation of explicit repository relationships without runtime-specific inference. lastVerifiedAt: '2026-08-15' lastVerifiedAt: '2026-08-15' eve: implementation: kind: package package: '@moldea.ai/adapter-eve' distribution: public versionRange: '^1.0.0' implementationStatus: available supportedRepositoryFormatVersions: - 1 compatibleCoreRange: '^2.0.0' runtimeGuidance: expectation: optional notes: Project-local guidance is needed only for unsupported dynamic capabilities, extensions, remote agents, non-canonical or composed instructions, positive single-file subagent analysis, Markdown skill registration, framework-tool overrides, or other repository-specific Eve patterns outside the verified static filesystem target. targets: - id: typescript-filesystem-agent-0-39 kind: package language: typescript packages: - ecosystem: npm name: eve role: primary versionRange: '>=0.39.1 <0.40.0' evidenceKinds: - agent-definition - handoff-registration - instruction-loader - language - runtime-package - schema - skill-registration - tool-registration bindingSupport: runtime-agent: relationship: full symbol: partial output-schema: relationship: full symbol: full instruction-loader: relationship: partial symbol: partial tool-implementation: relationship: full symbol: partial tool-registration: relationship: full symbol: partial tool-input-schema: relationship: full symbol: full tool-output-schema: relationship: full symbol: full skill-implementation: relationship: partial symbol: partial skill-registration: relationship: partial symbol: partial patterns: - id: nested-root-agent kind: agent support: full description: An uncollided bound package-relative agent/agent.ts directly default-exports one defineAgent configuration using the positive key subset and a supported static model string. - id: flat-root-agent kind: agent support: full description: An uncollided bound package-root agent.ts directly default-exports one defineAgent configuration using the positive key subset and a supported static model string. - id: directory-local-subagent kind: routing support: full description: A unique direct subagents//agent.ts package registers one immediate local subagent only when both parent and target use the positive static agent-definition subset, the target has a statically proved non-empty string description, and the mapping and runtime tool namespace are unambiguous; an omitted or exact empty description produces no handoff evidence. - id: single-file-local-subagent kind: routing support: partial description: Single-file subagent candidates participate in name and namespace preflight but do not produce positive agent-definition or handoff-registration evidence in the initial target. - id: effective-routing-description kind: routing support: full description: A supported non-empty static local-subagent description is compared with the canonical handoff description when present and the canonical agent-description fallback otherwise; an omitted or exact empty value is classified as missing and suppresses handoff evidence. - id: exact-lowercase-markdown-instruction kind: instruction-loader support: full description: One uncollided exact-lowercase root instructions.md file is the exclusive modern Eve system-instruction slot and is identified by the declared instruction relationship. - id: exclusive-typescript-instruction-loader kind: instruction-loader support: full description: One uncollided root instructions.ts directly default-exports exact-shape defineInstructions whose system content comes from the declared loader call. - id: case-varied-markdown-instruction kind: instruction-loader support: partial description: Case-varied instructions.md and system.md candidates participate in Eve precedence and collision analysis but do not produce positive complete-instruction evidence. - id: legacy-system-instruction kind: instruction-loader support: partial description: Deprecated system.* fallback is recognized to suppress false absence conclusions but does not produce positive canonical instruction evidence. - id: directory-instruction-composition kind: instruction-loader support: ambiguous description: Root-plus-directory, directory-only, multiple-entry, user-role, or dynamic instruction composition is outside the complete initial instruction target. - id: recursive-filesystem-tool kind: tool support: full description: An uncollided direct or nested tools/**/*.ts module with valid path segments and one unique non-reserved flattened name default-exports one registration-eligible defineTool configuration with supported implementation and schema relationships. - id: flattened-tool-runtime-name kind: tool support: full description: Eve tool runtime names are derived by removing the authored extension and replacing relative path separators beneath tools/ with hyphens. - id: tool-runtime-name-collision kind: tool support: full description: Distinct authored tool paths that flatten to one runtime name are diagnosed and do not produce effective registration evidence. - id: local-subagent-tool-namespace kind: routing support: full description: Static local subagents produce handoff evidence only when their names are unique among the supported static candidates and do not collide with mechanically established prepared authored, active framework, or reserved load_skill tool names; unresolved same-name static candidates suppress the claim. - id: flat-markdown-skill kind: skill support: partial description: A direct skills/.md file may establish an implementation-path relationship, but Markdown acceptance and Eve registration are not claimed by the initial target. - id: packaged-skill kind: skill support: partial description: A direct skills//SKILL.md file may establish an implementation-path relationship, but frontmatter acceptance, sibling resources, and Eve registration are not claimed by the initial target. - id: typescript-skill kind: skill support: full description: An uncollided direct skills/.ts module default-exports one exact closed static-string defineSkill package and may establish registration. - id: filesystem-slot-collisions kind: runtime support: ambiguous description: Competing Eve-authored agent, instruction, tool, skill, or local-subagent sources and root contributions using an observably claimed mounted-extension namespace prefix prevent the initial target from selecting one effective source unless the focused target defines a deterministic collision diagnostic. - id: dynamic-capabilities kind: runtime support: ambiguous description: Dynamic or non-string model definitions and runtime-resolved agents, instructions, tools, and skills cannot be mapped to one static effective surface by the initial target. - id: remote-agents kind: routing support: unsupported description: Remote Eve subagents, remote authentication, URLs, and cross-repository target identity are outside the initial target. - id: extension-contributions kind: runtime support: unsupported description: Extension-mounted agents, tools, skills, connections, hooks, and namespaced overrides are outside positive target interpretation. - id: connections-and-framework-tools kind: tool support: partial description: Connection-provided, provider-managed, framework opt-in, disabled, and Workflow tools are not exposed as manifest capabilities. The target models the prepared static framework namespace needed for registration evidence but makes no turn-time dynamic-tool availability claim. providerLimits: - id: filesystem-tool-path-segment-pattern subject: tool-name kind: pattern value: '^[A-Za-z][A-Za-z0-9_-]{0,63}$' description: Every authored filesystem segment beneath tools/, including the module basename, must contain 1 to 64 ASCII letters, digits, underscores, or hyphens and start with an ASCII letter before Eve flattens nested paths. reference: 'eve@0.39.1 TOOL_SLUG_PATTERN' - id: filesystem-tool-reserved-name subject: tool-name kind: other value: Workflow description: Eve reserves exact Workflow for its runtime workflow tool and rejects an authored tool with that flattened name. reference: 'eve@0.39.1 WORKFLOW_TOOL_NAME' knownLimitations: - Positive agent-definition evidence requires the closed defineAgent object to use only model, optional description, and optional outputSchema, with a supported static model string. Every other verified Eve agent option remains present-unsupported until a future target validates its nested runtime shape; dynamic models and sibling overrides therefore cannot produce optimistic evidence. - The adapter analyzes only a root or directory-backed local subagent with an exact bound static agent.ts; a configuration-free Eve root or single-file local subagent produces no positive target-specific definition or handoff evidence. - Only direct TypeScript default exports, exact Eve helper imports, and limited relative named-import resolution are supported; path aliases, package exports, directory indexes, CommonJS, re-exports, wrappers, and arbitrary compiler resolution remain unresolved. - The adapter parses only .ts authored modules. It inspects .cts, .mts, .cjs, .mjs, .ts, and .js entry names only to prevent false positive evidence when an Eve filesystem slot or local-subagent identity is collided. - Complete positive instruction evidence requires one exclusive exact-lowercase modern Markdown or exact-shape TypeScript system source. Eve directory composition, case-varied Markdown, unsupported module extensions, deprecated system.* fallback, the deprecated markdown definition branch, user-role entries, and dynamic instruction sources remain outside positive target evidence. - Recursive static tools are supported. Static prepared-name effects from supported authored tools, framework defaults, and unresolved same-name override candidates participate in registration preflight; dynamic, extension, connection, provider, disable-sentinel, and Workflow execution semantics otherwise remain outside the target. - Extension declarations and contributions are not interpreted. Observable extension mount names participate only in conservative namespace-prefix preflight so root contributions cannot receive false registration evidence under a prefix Eve reserves for a mounted extension. - Tool and skill descriptions are required to be statically provable strings for positive registration evidence but are not compared byte for byte with manifest capability descriptions; semantic alignment remains with the skill, evaluate, and PR Assurance. - Tool approval and toModelOutput behavior is not interpreted beyond the static value shape required to avoid false registration claims. - TypeScript skill metadata and files are supported only as closed static string-valued records; Eve-valid Uint8Array skill files and dynamic package content remain outside positive registration evidence. - Agent output, tool input, and tool output schema contents are not validated; the target establishes only direct binding relationships. - Flat and packaged Markdown skill paths may establish implementation-path relationships, but the target emits no Markdown skill-registration evidence because it does not reproduce Eve frontmatter acceptance and collision resolution. - Dynamic capabilities, remote agents, extensions, connections, channels, schedules, hooks, sandboxes, approvals, auth, state, sessions, compaction, task orchestration, and runtime-variable providers are not interpreted. - Eve 0.39.x requires Node.js 24 or newer for the inspected application runtime. The adapter records but does not validate that application prerequisite and may itself run on another adapter-supported Node.js line. lastVerifiedAt: '2026-08-19' lastVerifiedAt: '2026-08-19' google-genai: implementation: kind: package package: '@moldea.ai/adapter-google-genai' distribution: public versionRange: '^1.0.3' implementationStatus: available supportedRepositoryFormatVersions: - 1 compatibleCoreRange: '^2.0.0' runtimeGuidance: expectation: optional notes: Project-local guidance is needed only for repository-specific wrappers or unsupported indirect integration patterns. targets: - id: typescript-models-generate-content-2 kind: package language: typescript packages: - ecosystem: npm name: '@google/genai' role: primary versionRange: '>=2.17.1 <3.0.0' evidenceKinds: - instruction-loader - language - runtime-package - runtime-pattern - schema - tool-registration bindingSupport: runtime-agent: relationship: full symbol: full instruction-loader: relationship: full symbol: full tool-registration: relationship: full symbol: full tool-input-schema: relationship: full symbol: full patterns: - id: direct-models-generate-content kind: runtime support: full description: Direct Google Gen AI models.generateContent invocation through a module-local client in a directly exported TypeScript function. - id: direct-config-system-instruction kind: instruction-loader support: full description: A directly bound instruction loader supplies config.systemInstruction in a closed generate-content request. - id: closed-function-declaration-tools kind: tool support: full description: Closed inline or immutable module-local collections expose statically declared functions through config.tools and functionDeclarations. - id: direct-parameters-json-schema kind: schema support: full description: A bound tool input schema is referenced directly through the function declaration parametersJsonSchema property. - id: dynamic-request-or-config kind: runtime support: ambiguous description: Dynamically assembled requests or configuration cannot be mapped reliably without semantic analysis. - id: alternative-parameters-schema kind: schema support: unsupported description: FunctionDeclaration.parameters and its OpenAPI-style Schema representation are outside the initial JSON-schema target. - id: callable-and-mcp-tools kind: tool support: unsupported description: Callable tools, MCP conversion helpers, and automatic tool execution are outside the initial static function-declaration target. - id: provider-server-tools kind: tool support: unsupported description: Google-hosted or provider/server tools do not establish version 1 repository-local manifest tool relationships. - id: streaming-chat-live-interactions kind: runtime support: unsupported description: Streaming generation, chat sessions, live sessions, and Interactions API calls are outside the initial direct generate-content target. providerLimits: - id: function-name-length subject: tool-name kind: max-unicode-scalars value: 128 description: The Google Gen AI SDK FunctionDeclaration contract limits function names to 128 Unicode scalar values. reference: Google Gen AI SDK FunctionDeclaration reference. - id: function-name-pattern subject: tool-name kind: pattern value: '^[A-Za-z_][A-Za-z0-9_.:-]*$' description: The Google Gen AI SDK FunctionDeclaration contract uses the documented ASCII leading and continuation character set. reference: Google Gen AI SDK FunctionDeclaration reference. - id: function-declaration-count subject: other kind: other value: 512 description: The Google Gen AI SDK Tool contract permits at most 512 function declarations in each closed functionDeclarations collection. reference: Google Gen AI SDK Tool reference. knownLimitations: - Arbitrary compiler resolution, path aliases, directory indexes, package exports, subpath imports, and re-export graphs are not resolved. - Constructor configuration, provider backend, API version, authentication mode, model selection, request contents, and response handling are not interpreted. - Backend-specific function-name restrictions are not validated; the published function-name rules cover only the version-matched SDK declaration contract. - Dynamic configuration, callable tools, MCP helpers, provider/server tools, automatic function execution, streaming, chats, live sessions, and Interactions API calls are outside the initial target. - Function input-schema contents, including top-level object shape and parameter-name restrictions, are not validated; the target establishes only direct parametersJsonSchema wiring. - Source forms outside the verified TypeScript ESM target, legacy @google/generative-ai, alternative parameters schemas, output schemas, runtime variables, and handoffs are outside the initial target. lastVerifiedAt: '2026-08-19' lastVerifiedAt: '2026-08-19' langchain: implementation: kind: package package: '@moldea.ai/adapter-langchain' distribution: public versionRange: '^1.0.0' implementationStatus: available supportedRepositoryFormatVersions: - 1 compatibleCoreRange: '^2.0.0' runtimeGuidance: expectation: optional notes: Project-local guidance is needed only for repository-specific wrappers, middleware-driven relationships, supervisor composition around a supported LangChain boundary, headless tools, dynamic tool collections, or other unsupported indirect LangChain integration patterns. targets: - id: typescript-create-agent-1-5 kind: package language: typescript packages: - ecosystem: npm name: langchain role: primary versionRange: '>=1.5.9 <1.6.0' - ecosystem: npm name: '@langchain/core' role: companion versionRange: '>=1.2.8 <1.3.0' evidenceKinds: - agent-definition - instruction-loader - language - runtime-package - schema - tool-registration bindingSupport: runtime-agent: relationship: partial symbol: partial output-schema: relationship: partial symbol: partial instruction-loader: relationship: partial symbol: partial tool-implementation: relationship: partial symbol: partial tool-registration: relationship: partial symbol: partial tool-input-schema: relationship: partial symbol: partial patterns: - id: direct-create-agent kind: agent support: full description: A directly exported TypeScript const initialized through the package-root createAgent helper with one closed object-literal configuration is recognized as the agent definition. - id: langchain-primary-boundary kind: runtime support: full description: A supported createAgent definition remains LangChain even though LangChain compiles the agent on top of LangGraph. - id: direct-system-prompt-loader kind: instruction-loader support: partial description: Direct loader calls and direct SystemMessage construction are supported when middleware is statically inactive. - id: direct-response-format-schema kind: schema support: partial description: One directly bound responseFormat schema establishes the agent output-schema relationship when middleware is statically inactive. - id: structured-output-strategies kind: schema support: partial description: Direct toolStrategy and providerStrategy wrappers around one bound schema are supported; the helper-produced array shape of a single-schema toolStrategy call is not treated as an authored multi-format collection. - id: normal-function-tools kind: tool support: partial description: Direct normal two-argument tool helper declarations support implementation, runtime-name, and input-schema relationships. - id: closed-tool-collections kind: tool support: partial description: Closed inline and immutable module-local tool arrays establish agent registration when middleware is statically inactive. - id: middleware-relationship-effects kind: agent support: ambiguous description: Non-empty or unresolved middleware may alter prompts, tools, and structured runtime behavior and is not interpreted by the initial target. - id: response-format-arrays kind: schema support: ambiguous description: Developer-authored or statically bound multiple response schemas or strategies cannot be mapped uniquely to the single Repository Format agent output-schema binding. - id: state-and-context-input-schema kind: schema support: unsupported description: stateSchema and contextSchema are not collapsed into the Repository Format agent input-schema relationship. - id: headless-client-tools kind: tool support: unsupported description: One-argument headless tools do not establish a repository-local server implementation relationship in the initial target. - id: direct-langgraph-graphs kind: agent support: unsupported description: Direct StateGraph, Functional API, graph, and compiled-graph applications belong to the separate LangGraph adapter. - id: legacy-langchain-agents kind: agent support: unsupported description: Legacy agent executors and deprecated agent helper surfaces are outside the initial target. - id: supervisor-routing kind: routing support: unsupported description: Agent description fields do not establish source-to-target routing or handoff relationships without a verified supervisor registration target. knownLimitations: - Only TypeScript ESM source and documented direct relative imports are interpreted. - Only directly exported package-root createAgent definitions are recognized. - Non-empty or unresolved middleware suppresses prompt, tool-registration, and output-schema relationship conclusions. - Developer-authored and statically resolved multi-format arrays are not mapped to the single canonical agent output-schema binding; a single-schema toolStrategy call remains supported despite its array-shaped helper return. - The target does not infer agent input schemas from stateSchema or contextSchema. - The target does not interpret headless tools, provider tools, server tools, toolkits, MCP tools, or tool output schemas. - The target does not infer supervisors, routing targets, handoffs, or subagent control transfer. - Direct LangGraph applications remain outside this adapter target. - Lockfiles and installed package versions are not inspected. lastVerifiedAt: '2026-08-20' lastVerifiedAt: '2026-08-20' langgraph: implementation: kind: package package: '@moldea.ai/adapter-langgraph' distribution: public versionRange: '^1.0.0' implementationStatus: available supportedRepositoryFormatVersions: - 1 compatibleCoreRange: '^2.0.0' runtimeGuidance: expectation: recommended notes: Project-local guidance is recommended for prompt ownership, model and tool boundaries, node responsibilities, dynamic routing, state semantics, subgraphs, persistence, interrupts, human control, supervisor composition, and other repository-specific LangGraph behavior not represented by the initial static targets. targets: - id: typescript-state-graph-1-4 kind: package language: typescript packages: - ecosystem: npm name: '@langchain/langgraph' role: primary versionRange: '>=1.4.12 <1.5.0' - ecosystem: npm name: '@langchain/core' role: companion versionRange: '>=1.2.9 <1.3.0' evidenceKinds: - agent-definition - language - runtime-package - runtime-pattern - schema bindingSupport: runtime-agent: relationship: partial symbol: partial input-schema: relationship: partial symbol: partial output-schema: relationship: partial symbol: partial patterns: - id: direct-compiled-state-graph kind: agent support: full description: A directly exported TypeScript const compiled from a supported package-root StateGraph builder with a target-viable version-matched constructor is recognized as the runtime-agent definition. - id: inline-state-graph-builder kind: runtime support: full description: One inline fluent StateGraph chain ending in compile is supported when every intermediate method belongs to the documented target boundary. - id: single-owner-state-graph-builder kind: runtime support: partial description: One module-local const builder with deterministic top-level ownership and one authoritative compile call is supported. - id: graph-input-output-schemas kind: schema support: partial description: Direct immutable non-object-literal schema value bindings in one closed modern object initializer are supported, including state fallback when distinct input or output schemas are absent. - id: direct-node-registration kind: runtime support: partial description: Direct named addNode calls produce positive node-pattern evidence only for static names outside the runtime-rejected sentinel and reserved-separator set, supported repository-local, inline callable, or opaque potentially runnable action sources, and an absent, explicitly omitted, or target-viable object-family options value. - id: direct-edge-registration kind: runtime support: partial description: Direct addEdge calls with role-valid START, END, admissible static node-name, or non-empty waiting-edge positions produce positive edge-pattern evidence, except for the version-matched rejected direct START-to-END pair. - id: conditional-edge-registration kind: routing support: partial description: Direct static-source addConditionalEdges calls with supported callable or opaque potentially runnable router sources and an absent, explicitly omitted, or target-viable path-map collection produce control-flow evidence without inferring path-map destinations, router destinations, or agent handoffs. - id: compile-runtime-name kind: agent support: partial description: A supported static compile name may establish the graph runtime identity used in agent-definition evidence only when it satisfies the common machine-string contract and the adapter's evidence-safe runtime-identity grammar. - id: compile-description-routing kind: routing support: unsupported description: Compiled-graph description is not validated against the effective routing description without a supported supervisor registration target. - id: dynamic-command-and-send-routing kind: routing support: ambiguous description: Command, Send, parent-graph navigation, and runtime-generated destinations cannot be reconstructed as a complete static topology. - id: graph-node-capabilities kind: tool support: unsupported description: Graph nodes and node-local model or tool calls do not establish manifest tool relationships in the initial target. - id: prebuilt-and-supervisor-agents kind: agent support: unsupported description: Prebuilt agents, ToolNode, Deep Agents, supervisor packages, and higher-level harnesses are outside the initial direct StateGraph target. knownLimitations: - Only TypeScript ESM source and documented direct relative imports are interpreted. - Only directly exported compiled StateGraph definitions with inline or single-owner builders are recognized. - StateGraph target selection requires a target-viable version-matched one- or two-argument constructor. Opaque first arguments and indirect second arguments preserve the target only through the closed direct immutable object-family and opaque potentially object-or-schema classifiers; arbitrary unknown expressions do not. Statically impossible first arguments, direct or indirectly bound object literals outside the direct closed modern form, raw state maps, legacy channels objects, and unsupported modern object initializers produce no agent-definition evidence. Conditional constructor and schema candidates are role-viable only when recursive branch classification leaves at least one admitted result; an all-incompatible conditional cannot preserve the target. - Graph schema relationships require direct immutable non-object-literal values in unambiguous state, stateSchema, input, or output properties of one closed modern object initializer. Direct and directly bound raw object maps are deliberately unverified; the overloaded direct constructor family may preserve target selection but produces no schema evidence or negative wiring diagnostics. - One-argument object-map and tuple-array addNode overloads and addSequence are not expanded into positive node or edge evidence. Target preservation requires a non-empty closed supported collection shape or an opaque value that could resolve to one; statically empty, malformed, or incompatible direct or directly bound collections cannot preserve the graph target. - Member-sensitive directly bound runnable maps, node collections, tuple collections, and path maps must satisfy the closed aggregate-binding use grammar. A const declaration alone proves only binding identity; rooted writes, aliases, separate export forms, returns, yields, spreads, container insertion, unsupported argument or receiver uses, computed or optional access, and other escapes prevent target preservation through member inspection. - Positive addNode evidence excludes the reserved __start__ and __end__ node names, names containing | or :, and action sources outside the supported callable or opaque potentially runnable forms. Closed recursive runnable-map node-action candidates and exact inline opaque runnable candidates may preserve the graph target without positive operation evidence. - Positive edge and conditional-edge evidence excludes separator-bearing static node endpoints. Only the closed opaque-potentially-string classifier may preserve unresolved string-valued operation roles. - Positive conditional-edge evidence excludes router sources outside the supported callable or opaque potentially runnable forms. Other router values preserve the graph target only through the closed target-viable conditional-router classifier. - Structurally open or incompatible runnable-like object maps, callable objects outside the classifier, generators, mutable bindings, aliases, re-exports, computed or optional accesses, and other excluded runnable forms remain unverified rather than being treated as invalid LangGraph usage. - Node metadata, node-option property values, node defaults, subgraphs, error handlers, reducers, checkpointers, stores, caches, stream transformers, path-map destination meaning, and runtime context are not semantically interpreted. Compile options, node options, and context or schema candidates preserve unresolved target roles only through their documented closed object- and schema-role classifiers. - 'Every recognized builder call requires a version-matched public runtime arity and explicit type-argument count. StateGraph construction admits one through ten explicit type arguments; positional addNode admits one through three; object-map addNode and addSequence admit exactly two; tuple-array or opaque-collection addNode and addSequence admit one through three; compile admits exactly one. Explicit type arguments on addEdge, addConditionalEdges, setNodeDefaults, setEntryPoint, or setFinishPoint produce no graph target. One-argument node collections, positional and tuple node-options values, conditional-edge path maps and object overloads, and recognized-but-uninterpreted methods also apply explicit target-preservation gates. A call without positive operation evidence preserves the graph target only through the closed object-or-schema, opaque-potentially-string, target-viable node-action, target-viable conditional-router, options, path-map, endpoint, and collection classifiers; statically non-object node options, non-collection path maps, role-invalid path-map destinations, reserved node identities, role-invalid sentinels, separator-bearing endpoints, target-excluded operation sources, and the rejected direct START-to-END pair produce no graph target.' - Runtime names and source, target, or task name details are omitted unless they satisfy the deliberately narrow evidence-safe runtime-identity grammar; locator-shaped or otherwise unsafe names may still participate in relationship analysis without being copied into evidence. - Runtime-pattern evidence is positive and does not represent a complete or validated graph topology. - Edge-pattern evidence classifies the version-matched START and END sentinel values before applying role-specific source and target rules; waiting-edge source arrays must be non-empty, and role-invalid sentinel positions or the rejected direct START-to-END pair produce no edge evidence or diagnostic and cannot preserve the graph target when statically established. - Command, Send, router-body destinations, and other dynamic routing remain unresolved. - Graph nodes do not become manifest tools, skills, subagents, or handoffs merely from their graph position. - Compiled-graph description is not validated without a supported supervisor registration. - Lockfiles and installed package versions are not inspected. lastVerifiedAt: '2026-08-21' - id: typescript-functional-api-1-4 kind: package language: typescript packages: - ecosystem: npm name: '@langchain/langgraph' role: primary versionRange: '>=1.4.12 <1.5.0' - ecosystem: npm name: '@langchain/core' role: companion versionRange: '>=1.2.9 <1.3.0' evidenceKinds: - agent-definition - language - runtime-package - runtime-pattern bindingSupport: runtime-agent: relationship: partial symbol: partial patterns: - id: direct-functional-entrypoint kind: agent support: full description: A directly exported TypeScript const initialized through the package-root entrypoint helper with a supported static string name or direct closed options object literal and a supported workflow function is recognized as the runtime-agent definition; the name is copied into runtimeName only when it satisfies the evidence-safe runtime-identity grammar. - id: direct-functional-tasks kind: runtime support: partial description: Direct local or relative-import package-root task declarations with supported non-generator functions produce positive task-pattern evidence when the non-generic returned task proxy is called directly, non-optionally, without explicit type arguments, from the entrypoint lexical body. - id: functional-interrupt kind: runtime support: partial description: An exact one-argument direct package-root interrupt call in the entrypoint body produces human-in-the-loop runtime-pattern evidence without interpreting approval semantics. - id: functional-previous-state kind: runtime support: partial description: An exact zero-argument direct getPreviousState call produces persistence-related runtime-pattern evidence without inferring a schema. - id: functional-final-state kind: runtime support: partial description: An exact one-argument direct entrypoint.final call with a closed value/save object produces saved-state separation evidence without inferring input, output, or persistent-state schemas. - id: functional-control-flow kind: runtime support: ambiguous description: Ordinary branches, loops, callbacks, helper calls, and dynamic task selection are not reconstructed as a static graph. - id: functional-agent-schemas kind: schema support: unsupported description: TypeScript parameter and return types are not treated as executable Repository Format agent schema bindings. - id: functional-task-capabilities kind: tool support: unsupported description: Functional API tasks do not establish model-visible manifest tool relationships in the initial target. - id: functional-routing kind: routing support: unsupported description: Entrypoint control flow and task calls do not establish source-to-target agent handoffs. knownLimitations: - Only TypeScript ESM source and documented direct relative imports are interpreted. - Only directly exported package-root entrypoint definitions with inline or directly resolved non-generator workflow functions are recognized. - Entrypoint and task options are interpreted only when authored as direct closed object literals in the corresponding helper call; indirect options bindings and expressions are unsupported. - Task declarations resolve only through local const values or direct relative named imports to exported const declarations; generator and async-generator task functions are unsupported. - Entrypoint and task names outside the evidence-safe runtime-identity grammar may still establish supported relationships but are omitted from runtimeName and name-detail fields. - Only non-optional direct calls to the exact returned task proxy, with no explicit type arguments, in the entrypoint function's own lexical body are attributed to the registered workflow; explicit type arguments or indirect callee forms at the task-use site produce no task-use evidence. - Entrypoint and task declarations require exactly two explicit type arguments when a list is present. Interrupt accepts one or two, getPreviousState accepts exactly one, and entrypoint.final accepts exactly two; another count produces no applicable target or optional evidence. Returned task proxies are non-generic, and final-state options must be a closed direct value/save object literal. - Nested callback, nested helper, and transitive task-call graphs are not followed. - Ordinary JavaScript control flow is not reconstructed as a graph topology. - TypeScript input and output types do not establish executable schema relationships. - Tasks do not become manifest tools, skills, or handoffs merely because they are exported or invoked. - Interrupt payloads, checkpoint behavior, replay determinism, idempotency, and human-approval semantics are not validated. - Lockfiles and installed package versions are not inspected. lastVerifiedAt: '2026-08-21' lastVerifiedAt: '2026-08-21' openai: implementation: kind: package package: '@moldea.ai/adapter-openai' distribution: public versionRange: '^2.0.0' implementationStatus: available supportedRepositoryFormatVersions: - 1 compatibleCoreRange: '^2.0.0' runtimeGuidance: expectation: recommended notes: Document project-specific model selection, tool execution, streaming, retry, and error behavior that static inspection cannot establish. targets: - id: typescript-responses-api-7 kind: package language: typescript packages: - ecosystem: npm name: openai role: primary versionRange: '>=7.4.0 <8.0.0' evidenceKinds: - runtime-package - language - runtime-pattern - instruction-loader - tool-registration - schema bindingSupport: runtime-agent: relationship: full symbol: full instruction-loader: relationship: full symbol: full tool-registration: relationship: full symbol: full tool-input-schema: relationship: full symbol: full patterns: - id: direct-responses-runtime-agent kind: runtime support: full description: A bound exported TypeScript function uses a module-local OpenAI client for one or more direct Responses API object-literal requests with relationship-specific closure. - id: direct-instruction-loader kind: instruction-loader support: full description: A bound loader is called directly, optionally through await, by a Responses request instructions property. - id: static-function-tools kind: tool support: full description: Bound static OpenAI function-tool objects with the supported exact fields are included in a closed inline or immutable module-local Responses tools array. - id: direct-tool-input-schema kind: schema support: full description: A bound tool input schema is referenced directly by function-tool parameters. - id: dynamic-source-indirection kind: runtime support: ambiguous description: Factories, relationship-affecting computed properties and spreads, mutable arrays, and indirect request values remain unresolved. - id: chat-completions kind: runtime support: ambiguous description: Chat Completions usage is outside this target and is not rejected merely because Responses is preferred. knownLimitations: - Only TypeScript ESM files with supported direct default and relative named imports are interpreted. - Package versions are classified from nearest package manifests; lockfiles and installed node_modules are not inspected. - Source forms outside the verified TypeScript ESM target, Realtime, Assistants, Agents SDK, streaming semantics, and provider-hosted configuration are not interpreted. - Agent input and output schemas, tool implementations and output schemas, skills, variables, and runtime-native routing do not produce evidence. lastVerifiedAt: '2026-08-17' lastVerifiedAt: '2026-08-17' openai-agents-sdk: implementation: kind: package package: '@moldea.ai/adapter-openai-agents-sdk' distribution: public versionRange: '^1.0.0' implementationStatus: available supportedRepositoryFormatVersions: - 1 compatibleCoreRange: '^2.0.0' runtimeGuidance: expectation: optional notes: Project-local guidance is needed only for repository-specific wrappers, dynamic graph construction, or unsupported indirect integration patterns. targets: - id: typescript-agent-handoffs-0-16 kind: package language: typescript packages: - ecosystem: npm name: '@openai/agents' role: primary versionRange: '>=0.16.1 <0.17.0' evidenceKinds: - agent-definition - handoff-registration - instruction-loader - language - runtime-package - schema - tool-registration bindingSupport: runtime-agent: relationship: full symbol: full output-schema: relationship: full symbol: full instruction-loader: relationship: full symbol: full tool-implementation: relationship: full symbol: full tool-registration: relationship: full symbol: full tool-input-schema: relationship: full symbol: full tool-output-schema: relationship: full symbol: full patterns: - id: direct-agent-construction kind: agent support: full description: A directly exported TypeScript const constructs an Agent through new Agent with one closed object-literal configuration. - id: agent-create-construction kind: agent support: full description: A directly exported TypeScript const constructs an Agent through Agent.create with one closed object-literal configuration. - id: direct-instruction-loader kind: instruction-loader support: full description: A declared instruction loader is used by direct call, direct reference, or one supported single-return dynamic-instruction wrapper. - id: direct-agent-output-schema kind: schema support: full description: A bound agent output schema is referenced directly through outputType. - id: closed-function-tool kind: tool support: full description: A directly exported function tool uses the root tool helper, an explicit normalized static name, direct implementation, and direct schema bindings. - id: closed-agent-tool-array kind: tool support: full description: Closed inline or immutable module-local arrays register supported function tools on an Agent. - id: direct-agent-handoff kind: routing support: full description: A source Agent registers a supported target Agent directly in its closed handoffs collection. - id: configured-handoff-helper kind: routing support: full description: A source Agent registers a supported target through handoff with optional closed name and description overrides. - id: effective-routing-description kind: routing support: full description: Target handoffDescription uses the canonical handoff description when present and the canonical agent-description fallback otherwise. - id: registration-description-override kind: routing support: full description: A non-empty static toolDescriptionOverride is authoritative for its handoff registration and must use the target effective routing description. - id: dynamic-agent-configuration kind: agent support: ambiguous description: Dynamically assembled Agent configurations cannot be mapped reliably without semantic analysis. - id: dynamic-routing-description kind: routing support: ambiguous description: Runtime-generated or transformed handoff descriptions and description overrides remain unestablished. - id: agents-as-tools kind: routing support: unsupported description: Agent-as-tool delegation retains manager control and is not interpreted as a handoff by the initial target. - id: hosted-and-mcp-tools kind: tool support: unsupported description: Hosted, MCP-generated, namespaced, and tool-search tools are outside the initial repository-local function-tool target. - id: realtime-and-sandbox-agents kind: agent support: unsupported description: Realtime and sandbox agent abstractions are outside the initial Agent target. knownLimitations: - Arbitrary compiler resolution, path aliases, directory indexes, package exports, CommonJS, and re-export graphs are not resolved. - Routing-description validation supports only static inline, immutable module-local, and directly imported static strings; loaders, file reads, transformations, and runtime-generated values remain unresolved. - Function tools require an explicit static name already in the initial normalized runtime subset; omitted names and names requiring SDK normalization are outside the target rather than invalid. - Handoff evidence reports a runtime name only for a supported static non-empty toolNameOverride that can be represented as a valid Runtime Adapter Contract machine string. SDK-generated default names and absent, empty, dynamic, unsupported, mutation-obscured, or evidence-unrepresentable overrides are reported as null. - Agent output, tool input, and tool output schema contents are not validated; the target establishes only direct schema wiring. - Handoff input schemas, callbacks, filters, enablement, runtime variables, guardrails, prompt templates, sessions, tracing, approvals, models, and provider behavior are not interpreted. - Custom Handoff objects, agents as tools, hosted tools, MCP tools, Realtime agents, sandbox agents, and dynamically assembled agent graphs are outside the initial target. lastVerifiedAt: '2026-08-19' lastVerifiedAt: '2026-08-19' vercel-ai-sdk: implementation: kind: package package: '@moldea.ai/adapter-vercel-ai-sdk' distribution: public versionRange: '^1.0.0' implementationStatus: available supportedRepositoryFormatVersions: - 1 compatibleCoreRange: '^2.0.0' runtimeGuidance: expectation: optional notes: Project-local guidance is needed only for repository-specific wrappers or unsupported dynamic integration patterns. targets: - id: typescript-tool-loop-agent-7 kind: package language: typescript packages: - ecosystem: npm name: ai role: primary versionRange: '>=7.0.66 <8.0.0' evidenceKinds: - agent-definition - instruction-loader - language - runtime-package - schema - tool-registration bindingSupport: runtime-agent: relationship: partial symbol: partial input-schema: relationship: partial symbol: partial output-schema: relationship: partial symbol: partial instruction-loader: relationship: partial symbol: partial tool-implementation: relationship: partial symbol: partial tool-registration: relationship: partial symbol: partial tool-input-schema: relationship: partial symbol: partial tool-output-schema: relationship: partial symbol: partial patterns: - id: direct-tool-loop-agent-construction kind: agent support: full description: Directly exported ToolLoopAgent construction through one closed object-literal settings value. - id: direct-agent-instruction-loader kind: instruction-loader support: partial description: Direct loader calls in instructions are supported when prepareCall and prepareStep cannot replace them. - id: call-options-input-schema kind: schema support: full description: Direct callOptionsSchema binding establishes the agent input-schema relationship. - id: object-output-schema kind: schema support: partial description: Direct Output.object schema binding establishes the agent output-schema relationship when no uninterpreted prepareCall can remove it. - id: closed-tools-map kind: tool support: partial description: Closed object-map registration supports repository-local function tools created through tool. - id: direct-function-tool-bindings kind: tool support: partial description: Direct execute, inputSchema, and outputSchema bindings are interpreted without executing the tool. - id: prepare-call-overrides kind: agent support: ambiguous description: prepareCall may replace instructions and tools or omit the construction-time output, and is not interpreted by the initial target. - id: prepare-step-instruction-overrides kind: instruction-loader support: ambiguous description: prepareStep may replace per-step instructions and is not interpreted by the initial target. - id: subagent-handoff-inference kind: routing support: unsupported description: A function tool that calls another agent does not establish a target or handoff relationship in the initial target. - id: workflow-agent kind: agent support: unsupported description: WorkflowAgent and @ai-sdk/workflow are outside the initial target. knownLimitations: - Only TypeScript ESM source and documented direct relative imports are interpreted. - The target does not infer providers, models, routing targets, handoffs, or subagent control transfer. - Only Output.object establishes an agent output-schema relationship. - prepareCall and prepareStep function bodies are not interpreted; prepareCall therefore leaves instruction, tool, and output-schema wiring unresolved. - Lockfiles and installed package versions are not inspected. lastVerifiedAt: '2026-08-19' - id: typescript-generate-stream-text-7 kind: package language: typescript packages: - ecosystem: npm name: ai role: primary versionRange: '>=7.0.66 <8.0.0' evidenceKinds: - instruction-loader - language - runtime-package - runtime-pattern - schema - tool-registration bindingSupport: runtime-agent: relationship: partial symbol: partial output-schema: relationship: partial symbol: partial instruction-loader: relationship: partial symbol: partial tool-implementation: relationship: partial symbol: partial tool-registration: relationship: partial symbol: partial tool-input-schema: relationship: partial symbol: partial tool-output-schema: relationship: partial symbol: partial patterns: - id: direct-generate-text-wrapper kind: agent support: full description: A directly exported function containing a direct generateText object-literal call is recognized as the runtime pattern. - id: direct-stream-text-wrapper kind: agent support: full description: A directly exported function containing a direct streamText object-literal call is recognized as the runtime pattern. - id: instructions-system-precedence kind: instruction-loader support: full description: instructions is authoritative and deprecated system is used only through the supported absence fallback. - id: direct-generation-instruction-loader kind: instruction-loader support: partial description: Direct loader calls are supported when prepareStep cannot replace the instructions. - id: object-output-schema kind: schema support: partial description: Direct Output.object schema binding establishes the agent output-schema relationship. - id: closed-tools-map kind: tool support: partial description: Closed object-map registration supports repository-local function tools created through tool. - id: direct-function-tool-bindings kind: tool support: partial description: Direct execute, inputSchema, and outputSchema bindings are interpreted without executing the tool. - id: prepare-step-instruction-overrides kind: instruction-loader support: ambiguous description: prepareStep may replace per-step instructions and is not interpreted by the initial target. - id: indirect-generation-wrapper kind: runtime support: unsupported description: Calls routed through arbitrary wrappers, factories, callbacks, or request builders are outside the initial target. - id: direct-agent-input-schema kind: schema support: unsupported description: The initial direct-generation target publishes no agent input-schema relationship. knownLimitations: - Only direct generateText and streamText calls in the bound function's own lexical body are interpreted. - Only TypeScript ESM source and documented direct relative imports are interpreted. - The target does not infer providers, models, routing targets, handoffs, or subagent control transfer. - Only Output.object establishes an agent output-schema relationship. - prepareStep function bodies are not interpreted. - Lockfiles and installed package versions are not inspected. lastVerifiedAt: '2026-08-19' lastVerifiedAt: '2026-08-19'