ai-capability-framework 1.0.0-rc.3 → 1.0.0

package/docs/openai-runtime.md

@@ -1,10 +1,10 @@
 # OpenAI Responses Runtime
 
-The OpenAI Responses runtime is an optional subpath for host applications that
-want AICF to run a bounded Responses tool loop. It builds runtime context,
-routes model-safe read and prepare capabilities, calls a caller-provided OpenAI
-Responses client, executes function calls through the R2 `AicfToolExecutor`,
-returns model-safe tool envelopes, and produces final text.
+The OpenAI Responses runtime is an optional subpath for host applications that want AICF
+to run a bounded Responses tool loop. It builds runtime context, routes model-safe read
+and prepare capabilities, calls a caller-provided OpenAI Responses client, executes
+function calls through the R2 `AicfToolExecutor`, returns model-safe tool envelopes, and
+produces final text.
 
 Import it separately:
 
@@ -15,20 +15,19 @@ import {
 } from "ai-capability-framework/openai";
 ```
 
-The package root and `ai-capability-framework/runtime` imports do not import the
-OpenAI runtime code.
+The package root and `ai-capability-framework/runtime` imports do not import the OpenAI
+runtime code.
 
 For a first pass through the support/refund example, start with the
-[OpenAI walkthrough](openai-walkthrough.md). This page is the runtime API
-reference for host applications that want to provide a real OpenAI-compatible
-client.
+[OpenAI walkthrough](openai-walkthrough.md). This page is the runtime API reference for
+host applications that want to provide a real OpenAI-compatible client.
 
 ## Boundary
 
-The runtime loop is non-streaming and Responses-only. It does not expose commit
-tools to the model, execute side effects through model tool calls, persist
-production state, collect approvals, write durable audit logs, add observability
-systems, or require the OpenAI SDK for Core or runtime imports.
+The runtime loop is non-streaming and Responses-only. It does not expose commit tools to
+the model, execute side effects through model tool calls, persist production state,
+collect approvals, write durable audit logs, add observability systems, or require the
+OpenAI SDK for Core or runtime imports.
 
 Host applications remain responsible for:
 
@@ -57,9 +56,8 @@ npm install openai
 const client = await createDefaultOpenAIResponsesClient();
 ```
 
-If the SDK is unavailable, `createDefaultOpenAIResponsesClient()` throws an
-actionable optional-dependency error. This is the only API that attempts to
-import `openai`.
+If the SDK is unavailable, `createDefaultOpenAIResponsesClient()` throws an actionable
+optional-dependency error. This is the only API that attempts to import `openai`.
 
 ## Running A Request
 
@@ -85,16 +83,16 @@ Defaults:
 - `parallel_tool_calls = false`
 - streaming is out of scope
 
-The runner builds context, routes capabilities, converts the routed slice into
-OpenAI function tools, sends a non-streaming `responses.create` request, executes
-returned `function_call` items through `AicfToolExecutor`, and continues with
-`function_call_output` items. Tool output is serialized with runtime
-model-safe envelope helpers.
+The runner builds context, routes capabilities, converts the routed slice into OpenAI
+function tools, sends a non-streaming `responses.create` request, executes returned
+`function_call` items through `AicfToolExecutor`, and continues with
+`function_call_output` items. Tool output is serialized with runtime model-safe envelope
+helpers.
 
-If an OpenAI function call omits `call_id`, the runner fails safely rather than
-guessing how to pair output. Unknown tool names and invalid arguments become
-model-safe unavailable or validation-error envelopes. Provider failures return
-`provider_error` without raw provider payloads.
+If an OpenAI function call omits `call_id`, the runner fails safely rather than guessing
+how to pair output. Unknown tool names and invalid arguments become model-safe
+unavailable or validation-error envelopes. Provider failures return `provider_error`
+without raw provider payloads.
 
 ## Result Shape
 
@@ -111,8 +109,8 @@ model-safe unavailable or validation-error envelopes. Provider failures return
 - lightweight runtime events
 - safe errors
 
-Runtime events are summaries only. They must not be treated as trace storage and
-do not include raw provider payloads, prompts, secrets, or diagnostics.
+Runtime events are summaries only. They must not be treated as trace storage and do not
+include raw provider payloads, prompts, secrets, or diagnostics.
 
 Pass `traceSink` and `contentCapture` to emit sanitized events to the optional
 observability subpath:
@@ -125,29 +123,26 @@ const result = await runOpenAIResponses({
 });
 ```
 
-See [observability runtime](observability-runtime.md) for trace sinks and
-redaction behavior.
+See [observability runtime](observability-runtime.md) for trace sinks and redaction
+behavior.
 
 ## Agents SDK Bridge
 
 The OpenAI subpath also exports `buildAgentsSdkTools()` and
-`createDefaultAgentsSdkBridgeFactory()` for host applications that already use
-the OpenAI Agents SDK.
+`createDefaultAgentsSdkBridgeFactory()` for host applications that already use the
+OpenAI Agents SDK.
 
-The bridge creates function-tool definitions backed by `AicfToolExecutor`. It
-does not create an Agent, start an Agents SDK run, call a model, or make the
-Agents SDK approval system the source of truth. Approval-required AICF prepare
-results are returned as model-safe envelopes with prepared-action references.
-Commit tools are not generated.
+The bridge creates function-tool definitions backed by `AicfToolExecutor`. It does not
+create an Agent, start an Agents SDK run, call a model, or make the Agents SDK approval
+system the source of truth. Approval-required AICF prepare results are returned as
+model-safe envelopes with prepared-action references. Commit tools are not generated.
 
 ## Testing
 
 Mock tests use `MockOpenAIResponsesClient`:
 
 ```ts
-const client = new MockOpenAIResponsesClient([
-  mockTextResponse("Done.")
-]);
+const client = new MockOpenAIResponsesClient([mockTextResponse("Done.")]);
 ```
 
 Run mock tests without API keys:
@@ -170,14 +165,14 @@ Then run:
 npm run test:openai:live
 ```
 
-Live tests should stay harmless and synthetic. Do not store raw provider
-payloads, traces, prompts, or customer data in the repository.
+Live tests should stay harmless and synthetic. Do not store raw provider payloads,
+traces, prompts, or customer data in the repository.
 
 Live eval suites use this same runtime path with mock clients by default. See
 [live evals](live-evals.md).
 
-For a no-model runtime flow that uses the same context, router, executor, and
-action lifecycle primitives, run:
+For a no-model runtime flow that uses the same context, router, executor, and action
+lifecycle primitives, run:
 
 ```bash
 npm run build
@@ -186,8 +181,8 @@ node examples/runtime-support-billing/run-mock.mjs
 
 ## Reference Basis
 
-The runtime uses the Responses `responses.create` request shape and function
-calling continuation pattern documented by OpenAI:
+The runtime uses the Responses `responses.create` request shape and function calling
+continuation pattern documented by OpenAI:
 
 - [Responses create API](https://platform.openai.com/docs/api-reference/responses/create?api-mode=responses)
 - [Function calling guide](https://platform.openai.com/docs/guides/function-calling?api-mode=responses)

package/docs/openai-walkthrough.md

@@ -1,9 +1,9 @@
 # OpenAI Walkthrough
 
-This walkthrough uses the public support/refund example to show how AICF works
-with OpenAI. The default path is API-key-free: it exports tools, shows the tool
-call and repair shapes, and uses the mock runtime example. Live OpenAI usage is
-optional and requires a caller-provided client.
+This walkthrough uses the public support/refund example to show how AICF works with
+OpenAI. The default path is API-key-free: it exports tools, shows the tool call and
+repair shapes, and uses the mock runtime example. Live OpenAI usage is optional and
+requires a caller-provided client.
 
 ## The Scenario
 
@@ -19,14 +19,13 @@ The support example has three relevant capabilities:
 - `support.refund.prepare_case`: prepare a refund case for approval.
 - `support.refund.commit_case`: commit the refund after host approval.
 
-The first two can be model-facing tools. The commit capability is host
-controlled and is not exposed to the model by default.
+The first two can be model-facing tools. The commit capability is host controlled and is
+not exposed to the model by default.
 
 ## 1. Capability Manifests Define The Contract
 
-A capability manifest describes the model-facing purpose, the input and output
-schemas, the risk tier, required permissions, policy gates, lifecycle flags, and
-eval links.
+A capability manifest describes the model-facing purpose, the input and output schemas,
+the risk tier, required permissions, policy gates, lifecycle flags, and eval links.
 
 Small excerpt:
 
@@ -41,8 +40,7 @@ policy:
   approval_required: true
 ```
 
-AICF validates these manifests before any provider tool export or runtime
-execution.
+AICF validates these manifests before any provider tool export or runtime execution.
 
 ## 2. AICF Exports OpenAI Function Tools
 
@@ -77,8 +75,8 @@ Small excerpt:
 }
 ```
 
-The binding map is important. AICF never guesses a capability ID from a provider
-tool name; it maps tool calls back through generated bindings.
+The binding map is important. AICF never guesses a capability ID from a provider tool
+name; it maps tool calls back through generated bindings.
 
 ## 3. OpenAI Returns A Function Call
 
@@ -93,14 +91,13 @@ A model might return a Responses function call like this:
 }
 ```
 
-AICF parses the provider call, resolves the binding to `support.ticket.get`,
-parses JSON arguments, and validates those arguments against the original AICF
-`input_schema`.
+AICF parses the provider call, resolves the binding to `support.ticket.get`, parses JSON
+arguments, and validates those arguments against the original AICF `input_schema`.
 
 ## 4. Invalid Calls Become Repair Envelopes
 
-If the model calls an unknown tool or sends incorrect arguments, AICF fails
-closed. The app can return a model-safe repair envelope as the tool output.
+If the model calls an unknown tool or sends incorrect arguments, AICF fails closed. The
+app can return a model-safe repair envelope as the tool output.
 
 Example validation failure:
 
@@ -128,9 +125,9 @@ OpenAI continuation shape:
 }
 ```
 
-The model can then repair its next call. The app may also decide to stop, ask
-the user for clarification, or escalate to a stronger model. AICF only provides
-the safe envelope and deterministic validation result.
+The model can then repair its next call. The app may also decide to stop, ask the user
+for clarification, or escalate to a stronger model. AICF only provides the safe envelope
+and deterministic validation result.
 
 ## 5. Prepare Can Require Approval
 
@@ -149,8 +146,7 @@ envelope:
 }
 ```
 
-This tells the model that work has paused. It must not claim the refund was
-issued.
+This tells the model that work has paused. It must not claim the refund was issued.
 
 ## 6. Commit Is Host-Controlled
 
@@ -180,16 +176,16 @@ Expected excerpt:
 
 ## 7. Optional Live OpenAI Runtime
 
-The optional `ai-capability-framework/openai` subpath can run a bounded
-non-streaming Responses loop with a caller-provided compatible client:
+The optional `ai-capability-framework/openai` subpath can run a bounded non-streaming
+Responses loop with a caller-provided compatible client:
 
 ```ts
 import { runOpenAIResponses } from "ai-capability-framework/openai";
 ```
 
-AICF does not require the OpenAI SDK for root or runtime imports. Live usage is
-opt-in, and host applications remain responsible for provider credentials,
-model selection, auth, side effects, approvals, storage, and audit.
+AICF does not require the OpenAI SDK for root or runtime imports. Live usage is opt-in,
+and host applications remain responsible for provider credentials, model selection,
+auth, side effects, approvals, storage, and audit.
 
 ## What To Read Next
 

package/docs/policy-broker.md

@@ -1,7 +1,7 @@
 # Policy Broker
 
-`DefaultPolicyBroker` combines Core manifest decisions with runtime context. It
-is deterministic and fail-closed.
+`DefaultPolicyBroker` combines Core manifest decisions with runtime context. It is
+deterministic and fail-closed.
 
 Import it from:
 
@@ -15,13 +15,12 @@ Policy evaluation receives:
 
 - the loaded capability manifest;
 - operation: `select`, `prepare`, or `commit`;
-- runtime context with subject, account, tenant, autonomy, risk, facts, and
-  metadata;
+- runtime context with subject, account, tenant, autonomy, risk, facts, and metadata;
 - args, approval, prepared-action, idempotency, and built context when relevant.
 
 Host applications remain responsible for real authorization, account state,
-entitlements, facts, and approval verification. AICF evaluates the information
-the host supplies.
+entitlements, facts, and approval verification. AICF evaluates the information the host
+supplies.
 
 ## Default Behavior
 
@@ -34,8 +33,8 @@ The broker wraps Core `decideCapability()` and adds runtime checks for:
 - missing, rejected, expired, or mismatched approvals;
 - missing idempotency keys when required by the capability.
 
-Ambiguous state denies or pauses. The broker does not guess missing facts or
-repair incomplete context.
+Ambiguous state denies or pauses. The broker does not guess missing facts or repair
+incomplete context.
 
 ## Host Hooks
 
@@ -45,5 +44,5 @@ An optional host policy hook may return a stricter decision:
 - `approval_required` can override `allowed`;
 - `allowed` cannot override an AICF denial.
 
-Thrown hooks fail closed. Hook diagnostics must stay safe for logs and must not
-include secrets, raw prompts, provider payloads, or customer data.
+Thrown hooks fail closed. Hook diagnostics must stay safe for logs and must not include
+secrets, raw prompts, provider payloads, or customer data.

package/docs/provenance.md

@@ -1,12 +1,11 @@
 # Content Provenance Hooks
 
-The provenance subpath creates public-safe metadata for generated content that
-may become customer-facing text, documents, images, audio, video, or other
-media.
+The provenance subpath creates public-safe metadata for generated content that may
+become customer-facing text, documents, images, audio, video, or other media.
 
-It does not sign content, implement C2PA, process files, call providers, store
-records, or claim authenticity. Host applications own real signing, document
-embedding, CMS/media pipelines, approval identity, retention, and legal review.
+It does not sign content, implement C2PA, process files, call providers, store records,
+or claim authenticity. Host applications own real signing, document embedding, CMS/media
+pipelines, approval identity, retention, and legal review.
 
 ## TypeScript
 
@@ -27,12 +26,15 @@ const provenance = createGeneratedContentProvenance({
     { capabilityId: "support.ticket.get", operation: "read" },
     { capabilityId: "support.refund.prepare_case", operation: "prepare" }
   ],
-  sourceRefs: [{
-    contentHash: "sha256:1111111111111111111111111111111111111111111111111111111111111111",
-    sourceId: "support.ticket.synthetic.ref",
-    sourceType: "app_record",
-    trust: "app_data"
-  }],
+  sourceRefs: [
+    {
+      contentHash:
+        "sha256:1111111111111111111111111111111111111111111111111111111111111111",
+      sourceId: "support.ticket.synthetic.ref",
+      sourceType: "app_record",
+      trust: "app_data"
+    }
+  ],
   approvalRefs: ["approval.support.refund.synthetic"],
   createdAt: new Date().toISOString()
 });
@@ -44,20 +46,20 @@ const validation = validateGeneratedContentProvenance(provenance);
 
 `GeneratedContentProvenance` records:
 
-- content ID, content type, creation time, and whether content was model,
-  model-assisted human, or human-approved model output;
+- content ID, content type, creation time, and whether content was model, model-assisted
+  human, or human-approved model output;
 - provider, model, capability, approval, and trace references;
 - source references with hashes or public-safe IDs.
 
-The record is intentionally a sidecar contract. It contains refs, hashes,
-statuses, and metadata only. It must not include report body text, raw prompts,
-provider request/response payloads, transcripts, secrets, stack traces,
-unredacted tenant/account/user IDs, or sensitive source content.
+The record is intentionally a sidecar contract. It contains refs, hashes, statuses, and
+metadata only. It must not include report body text, raw prompts, provider
+request/response payloads, transcripts, secrets, stack traces, unredacted
+tenant/account/user IDs, or sensitive source content.
 
 ## Adapter Hooks
 
-`runProvenanceAdapterHook()` lets a host pass validated metadata to a signing,
-CMS, document, or media pipeline adapter:
+`runProvenanceAdapterHook()` lets a host pass validated metadata to a signing, CMS,
+document, or media pipeline adapter:
 
 ```ts
 const result = await runProvenanceAdapterHook(provenance, async (record) => ({
@@ -71,20 +73,19 @@ const result = await runProvenanceAdapterHook(provenance, async (record) => ({
 }));
 ```
 
-The hook receives only validated, public-safe provenance metadata. If a hook
-throws or returns unsafe metadata, AICF returns a safe failed result without
-stack traces or raw error text.
+The hook receives only validated, public-safe provenance metadata. If a hook throws or
+returns unsafe metadata, AICF returns a safe failed result without stack traces or raw
+error text.
 
 ## Validation
 
-Validation is fail-closed. It rejects missing required references, invalid
-timestamps, unknown content types, unsafe raw-content-looking fields, provider
-payload markers, credential-looking keys, and source refs that should include a
-hash but do not.
+Validation is fail-closed. It rejects missing required references, invalid timestamps,
+unknown content types, unsafe raw-content-looking fields, provider payload markers,
+credential-looking keys, and source refs that should include a hash but do not.
 
 The public example at
-`examples/support/provenance/support.refund.summary.provenance.json` shows a
-synthetic support/refund document sidecar with hashes and refs only.
+`examples/support/provenance/support.refund.summary.provenance.json` shows a synthetic
+support/refund document sidecar with hashes and refs only.
 
 ## Relationship To Other AICF Surfaces
 

package/docs/providers/anthropic.md

@@ -1,12 +1,12 @@
 # Anthropic Claude
 
-The Anthropic provider subpath exposes routed AICF read/prepare capabilities as
-Claude Messages tools and can run a bounded mock-tested tool-use loop with a
-caller-provided client.
+The Anthropic provider subpath exposes routed AICF read/prepare capabilities as Claude
+Messages tools and can run a bounded mock-tested tool-use loop with a caller-provided
+client.
 
 Read:
 
 - [Anthropic Claude runtime](../anthropic-runtime.md)
 
-Default tests use mock clients. Live Claude tests are opt-in and require
-explicit environment variables.
+Default tests use mock clients. Live Claude tests are opt-in and require explicit
+environment variables.

package/docs/providers/choose-a-runtime.md

@@ -0,0 +1,37 @@
+# Choose a Runtime
+
+Start with provider-neutral AICF contracts, then pick the smallest adapter layer your
+host application needs.
+
+| Path                | Use When                                                                                | Default Test Mode                                      |
+| ------------------- | --------------------------------------------------------------------------------------- | ------------------------------------------------------ |
+| Core contracts      | You need manifests, validation, deterministic decisions, and evals without model calls. | `npm run validate`, `npm run conformance`              |
+| OpenAI Responses    | You want a bounded OpenAI tool loop over routed read/prepare capabilities.              | `npm run test:openai:mock`                             |
+| Anthropic Claude    | You want Claude Messages tool use over routed read/prepare capabilities.                | `npm run test:anthropic:mock`                          |
+| Google Gemini       | You want GenerateContent function calling over routed read/prepare capabilities.        | `npm run test:gemini:mock`                             |
+| Vercel AI SDK       | You want AI SDK tools around host-supplied `generateText` or `streamText`.              | `npm run test:ai-sdk:mock`                             |
+| LangChain/LangGraph | You want LangChain tools or a host-supplied LangGraph `ToolNode`.                       | `npm run test:langchain:mock`                          |
+| MCP                 | You want MCP descriptors or an AICF-backed MCP server surface.                          | `npm run test:mcp-provider`, `npm run test:mcp-server` |
+| Semantic Kernel     | You want MCP guidance or OpenAPI plugin metadata for Semantic Kernel hosts.             | `npm run test:semantic-kernel`                         |
+
+## Selection Rules
+
+- If you are learning AICF, start with the
+  [provider-neutral quickstart](../getting-started/provider-neutral-quickstart.md).
+- If you already have an OpenAI app, use the
+  [OpenAI quickstart](../getting-started/openai-quickstart.md).
+- If you need a non-OpenAI direct runtime, use the
+  [Anthropic quickstart](../getting-started/anthropic-quickstart.md) or
+  [Gemini quickstart](../getting-started/gemini-quickstart.md).
+- If an agent framework owns the model loop, use the AI SDK, LangChain, MCP, or Semantic
+  Kernel bridge and keep AICF as the validation and lifecycle authority.
+
+## Boundary
+
+All runtime choices share the same safety boundary:
+
+- provider descriptors expose routed read/prepare capabilities only;
+- commit remains host-controlled through the action lifecycle;
+- provider SDK validation does not replace AICF validation;
+- raw provider payloads and raw prompts are not stored by default;
+- live calls require explicit environment variables and host-owned credentials.

package/docs/providers/conformance.md

@@ -4,9 +4,9 @@ AICF conformance checks prove that one routed capability slice can be exported
 consistently across provider and framework targets. The canonical import path is
 `ai-capability-framework/conformance`.
 
-The conformance harness is descriptor/mock-only. It does not call live models,
-use credentials, execute provider SDKs, store provider payloads, or expose
-commit capabilities to models.
+The conformance harness is descriptor/mock-only. It does not call live models, use
+credentials, execute provider SDKs, store provider payloads, or expose commit
+capabilities to models.
 
 ## Install And Import
 
@@ -22,8 +22,8 @@ import {
 } from "ai-capability-framework/conformance";
 ```
 
-The older `ai-capability-framework/providers/conformance` path remains available
-as a compatibility alias.
+The older `ai-capability-framework/providers/conformance` path remains available as a
+compatibility alias.
 
 ## Targets
 
@@ -38,8 +38,8 @@ The canonical matrix covers:
 - Semantic Kernel through MCP compatibility;
 - Semantic Kernel through OpenAPI plugin metadata.
 
-`semantic-kernel` is accepted as an alias for `semantic-kernel-openapi`.
-`vercel-ai-sdk` and `vercel_ai_sdk` are accepted as aliases for `ai-sdk`.
+`semantic-kernel` is accepted as an alias for `semantic-kernel-openapi`. `vercel-ai-sdk`
+and `vercel_ai_sdk` are accepted as aliases for `ai-sdk`.
 
 ## What Gets Checked
 
@@ -56,9 +56,8 @@ The report scores each target across these dimensions:
 - provider error normalization;
 - loop or streaming boundary metadata.
 
-Failures are structured by provider, capability, case, and dimension. The report
-also keeps the older `passed`, `counts`, and `results` fields for existing
-callers.
+Failures are structured by provider, capability, case, and dimension. The report also
+keeps the older `passed`, `counts`, and `results` fields for existing callers.
 
 ## CLI
 
@@ -84,12 +83,12 @@ aicf providers conformance examples --format text
 
 ## Boundary
 
-Host applications remain responsible for model calls, provider SDK setup,
-runtime context, real authorization, approval collection, idempotency, side
-effects, audit persistence, and optional live tests.
+Host applications remain responsible for model calls, provider SDK setup, runtime
+context, real authorization, approval collection, idempotency, side effects, audit
+persistence, and optional live tests.
 
-Conformance uses public manifests, synthetic contexts, and mock/canonical calls.
-Live provider behavior belongs in provider-specific opt-in tests.
+Conformance uses public manifests, synthetic contexts, and mock/canonical calls. Live
+provider behavior belongs in provider-specific opt-in tests.
 
 ## Checks
 

package/docs/providers/gemini.md

@@ -1,12 +1,12 @@
 # Google Gemini
 
-The Gemini provider subpath exposes routed AICF read/prepare capabilities as
-Gemini function declarations and can run a bounded mock-tested GenerateContent
-loop with a caller-provided client.
+The Gemini provider subpath exposes routed AICF read/prepare capabilities as Gemini
+function declarations and can run a bounded mock-tested GenerateContent loop with a
+caller-provided client.
 
 Read:
 
 - [Google Gemini runtime](../gemini-runtime.md)
 
-Default tests use mock clients. Live Gemini tests are opt-in and require
-explicit environment variables.
+Default tests use mock clients. Live Gemini tests are opt-in and require explicit
+environment variables.

package/docs/providers/mcp.md

@@ -1,8 +1,7 @@
 # Model Context Protocol
 
-AICF supports MCP descriptors and an optional MCP server runtime. MCP clients
-receive routed read/prepare tools backed by AICF validation and executor
-boundaries.
+AICF supports MCP descriptors and an optional MCP server runtime. MCP clients receive
+routed read/prepare tools backed by AICF validation and executor boundaries.
 
 Read:
 

package/docs/providers/openai.md

@@ -11,5 +11,5 @@ Start here:
 - [OpenAI Responses descriptor adapter](../openai-responses.md)
 - [OpenAI Responses runtime](../openai-runtime.md)
 
-Default examples do not call OpenAI. Live use requires a caller-provided client
-or optional SDK setup.
+Default examples do not call OpenAI. Live use requires a caller-provided client or
+optional SDK setup.

package/docs/providers/semantic-kernel.md

@@ -1,7 +1,7 @@
 # Semantic Kernel
 
-AICF supports Semantic Kernel-compatible hosts through MCP guidance and OpenAPI
-plugin metadata export. It does not provide a native Semantic Kernel runtime.
+AICF supports Semantic Kernel-compatible hosts through MCP guidance and OpenAPI plugin
+metadata export. It does not provide a native Semantic Kernel runtime.
 
 Read: