docs: document LLM request and tool execution outcomes

docs/build-plugins/code-examples.mdx

@@ -37,11 +37,14 @@ class HeaderPlugin:
             name: str,
             request: nemo_relay.LLMRequest,
             annotated: nemo_relay.AnnotatedLLMRequest | None
-        ) -> tuple[nemo_relay.LLMRequest, nemo_relay.AnnotatedLLMRequest | None]:
+        ) -> nemo_relay.LLMRequestInterceptOutcome:
             # The request object is immutable, however we can return a new instance with updated headers.
             headers = request.headers.copy()
             headers[plugin_config["header_name"]] = plugin_config["value"]
-            return nemo_relay.LLMRequest(headers=headers, content=request.content), annotated
+            return nemo_relay.LLMRequestInterceptOutcome(
+                nemo_relay.LLMRequest(headers=headers, content=request.content),
+                annotated,
+            )
 
         context.register_llm_request_intercept("inject-header", 100, False, add_header)
 

docs/build-plugins/register-behavior.mdx

@@ -51,10 +51,13 @@ class HeaderPlugin:
             name: str,
             request: nemo_relay.LLMRequest,
             annotated: nemo_relay.AnnotatedLLMRequest | None
-        ) -> tuple[nemo_relay.LLMRequest, nemo_relay.AnnotatedLLMRequest | None]:
+        ) -> nemo_relay.LLMRequestInterceptOutcome:
             headers = request.headers.copy()
             headers[plugin_config["header_name"]] = plugin_config["value"]
-            return nemo_relay.LLMRequest(headers=headers, content=request.content), annotated
+            return nemo_relay.LLMRequestInterceptOutcome(
+                nemo_relay.LLMRequest(headers=headers, content=request.content),
+                annotated,
+            )
 
         context.register_llm_request_intercept("inject-header", 100, False, add_header)
 
@@ -99,6 +102,7 @@ plugin.register('header-plugin', headerPlugin);
 
 <Tab title="Rust" language="rust">
 ```rust
+use nemo_relay::api::llm::LlmRequestInterceptOutcome;
 use nemo_relay::plugin::{
     register_plugin, ConfigDiagnostic, DiagnosticLevel, Plugin, PluginRegistrationContext,
     Result as PluginResult,
@@ -166,7 +170,7 @@ impl Plugin for HeaderPlugin {
                     request
                         .headers
                         .insert(header_name.clone(), header_value.clone().into());
-                    Ok((request, annotated))
+                    Ok(LlmRequestInterceptOutcome::new(request, annotated))
                 }),
             )?;
             Ok(())

docs/instrument-applications/advanced-guide.mdx

@@ -65,7 +65,8 @@ def require_query(tool_name, args):
 async def measure_tool(tool_name, args, next_call):
     started = time.perf_counter()
     try:
-        return await next_call(args)
+        result = await next_call(args)
+        return nemo_relay.ToolExecutionInterceptOutcome(result)
     finally:
         elapsed_ms = round((time.perf_counter() - started) * 1000, 2)
         print(f"{tool_name} completed in {elapsed_ms} ms")
@@ -98,7 +99,8 @@ registerToolConditionalExecutionGuardrail("search.require_query", 20, (_toolName
 registerToolExecutionIntercept("search.measure", 30, async (args, next) => {
   const started = performance.now();
   try {
-    return await next(args);
+    const result = await next(args);
+    return { result };
   } finally {
     console.log(`search completed in ${Math.round(performance.now() - started)} ms`);
   }
@@ -113,6 +115,7 @@ use nemo_relay::api::registry::{
     register_tool_execution_intercept,
     register_tool_sanitize_request_guardrail,
 };
+use nemo_relay::api::tool::ToolExecutionInterceptOutcome;
 use serde_json::json;
 use std::sync::Arc;
 use std::time::Instant;
@@ -145,11 +148,12 @@ register_tool_execution_intercept(
     "search.measure",
     30,
     Arc::new(|name, args, next| {
+        let name = name.to_owned();
         Box::pin(async move {
             let started = Instant::now();
-            let result = next(name.clone(), args).await;
+            let result = next(args).await;
             println!("{name} completed in {:?}", started.elapsed());
-            result
+            Ok(ToolExecutionInterceptOutcome::new(result?))
         })
     }),
 )?;
@@ -158,6 +162,10 @@ register_tool_execution_intercept(
 
 </Tabs>
 
+Tool execution intercepts return an outcome even when they only forward the
+result from `next`. Refer to [Tool Execution Intercept Outcomes](/reference/tool-execution-intercept-outcomes)
+for pending marks, lifecycle ordering, and binding-specific return types.
+
 ## Scope Middleware to One Request
 
 Use scope-local middleware when a policy applies only to one request, tenant, experiment, or agent run.

docs/instrument-applications/code-examples.mdx

@@ -260,8 +260,8 @@ tool_args = nemo_relay.tools.request_intercepts("search", {"query": "weather"})
 nemo_relay.tools.conditional_execution("search", tool_args)
 
 llm_request = LLMRequest({}, {"messages": [{"role": "user", "content": "hello"}]})
-llm_request = nemo_relay.llm.request_intercepts("demo-provider", llm_request)
-nemo_relay.llm.conditional_execution(llm_request)
+outcome = nemo_relay.llm.request_intercepts("demo-provider", llm_request)
+nemo_relay.llm.conditional_execution(outcome.request)
 ```
 </Tab>
 
@@ -279,8 +279,8 @@ const toolArgs = await toolRequestIntercepts('search', { query: 'weather' });
 await toolConditionalExecution('search', toolArgs);
 
 const request = new LlmRequest({}, { messages: [{ role: 'user', content: 'hello' }] });
-const rewritten = await llmRequestIntercepts('demo-provider', request);
-await llmConditionalExecution(rewritten);
+const outcome = await llmRequestIntercepts('demo-provider', request);
+await llmConditionalExecution(outcome.request);
 ```
 </Tab>
 
@@ -297,8 +297,8 @@ let request = LlmRequest {
     headers: Default::default(),
     content: json!({"messages": [{"role": "user", "content": "hello"}]}),
 };
-let rewritten = llm_request_intercepts("demo-provider", request)?;
-llm_conditional_execution(&rewritten)?;
+let outcome = llm_request_intercepts("demo-provider", request)?;
+llm_conditional_execution(&outcome.request)?;
 ```
 </Tab>
 

docs/integrate-into-frameworks/code-examples.mdx

@@ -184,10 +184,11 @@ import nemo_relay
 from nemo_relay import LLMRequest
 
 rewritten_args = nemo_relay.tools.request_intercepts("search", {"query": "weather"})
-rewritten_request = nemo_relay.llm.request_intercepts(
+outcome = nemo_relay.llm.request_intercepts(
     "demo-provider",
     LLMRequest({}, {"messages": []}),
 )
+rewritten_request = outcome.request
 ```
 </Tab>
 
@@ -196,7 +197,8 @@ rewritten_request = nemo_relay.llm.request_intercepts(
 import { LlmRequest, llmRequestIntercepts, toolRequestIntercepts } from 'nemo-relay-node';
 
 const rewrittenArgs = await toolRequestIntercepts('search', { query: 'weather' });
-const rewrittenRequest = await llmRequestIntercepts('demo-provider', new LlmRequest({}, { messages: [] }));
+const outcome = await llmRequestIntercepts('demo-provider', new LlmRequest({}, { messages: [] }));
+const rewrittenRequest = outcome.request;
 ```
 </Tab>
 
@@ -208,7 +210,8 @@ use serde_json::json;
 
 let rewritten_args = tool_request_intercepts("search", json!({"query": "weather"}))?;
 let request = LlmRequest { headers: Default::default(), content: json!({"messages": []}) };
-let rewritten_request = llm_request_intercepts("demo-provider", request)?;
+let outcome = llm_request_intercepts("demo-provider", request)?;
+let rewritten_request = outcome.request;
 ```
 </Tab>
 

docs/integrate-into-frameworks/provider-codecs.mdx

@@ -41,10 +41,17 @@ When a managed LLM call has a request codec:
 
 1. NeMo Relay calls `decode` before LLM request intercepts run.
 2. Request intercepts receive both the raw request and the annotated request.
-3. Intercepts may edit the raw request, the annotated request, or both.
+3. Intercepts edit provider-body fields through the annotated request and may
+   edit transport headers through the raw request. Raw `request.content` is
+   read-only while the codec is active.
 4. NeMo Relay calls `encode` to merge the annotated request back into the original raw request.
 5. Execution intercepts and the provider callback receive the encoded provider request.
 
+If a codec-aware intercept changes raw `request.content` or omits the returned
+annotation, Relay rejects the outcome before creating the LLM lifecycle. When
+no request codec is active, the raw request remains fully writable and is the
+provider-visible source of truth.
+
 When a managed LLM call has a response codec, NeMo Relay decodes the raw provider response for observability and attaches the result to the emitted LLM end event. The response codec does not rewrite the value returned to the application. Use [Provider Response Codecs](/integrate-into-frameworks/provider-response-codecs) for response-only behavior and custom response codec examples.
 
 Codec implementations should preserve fields they do not understand. Treat `encode` as a merge operation over the original provider payload, not as a full replacement.
@@ -87,7 +94,7 @@ from nemo_relay.codecs import OpenAIChatCodec
 
 def add_system_message(_name, request, annotated):
     if annotated is None:
-        return request, annotated
+        return nemo_relay.LLMRequestInterceptOutcome(request)
 
     # Attributes of the annotated request can be re-assigned, but cannot be modified in-place.
     # For example `annotated.messages.append(...)` would not work, but re-assigning
@@ -96,7 +103,7 @@ def add_system_message(_name, request, annotated):
         {"role": "system", "content": "Answer with concise technical detail."},
         *annotated.messages,
     ]
-    return request, annotated
+    return nemo_relay.LLMRequestInterceptOutcome(request, annotated)
 
 nemo_relay.intercepts.register_llm_request(
     "framework.add_system_message",

docs/reference/llm-request-intercept-outcomes.mdx

@@ -0,0 +1,118 @@
+---
+title: "LLM Request Intercept Outcomes"
+description: "Canonical result returned by LLM request intercepts and its managed lifecycle behavior."
+---
+{/* SPDX-FileCopyrightText: Copyright (c) 2026, NVIDIA CORPORATION & AFFILIATES. All rights reserved.
+SPDX-License-Identifier: Apache-2.0 */}
+
+An LLM request intercept rewrites a request before managed execution. This page
+describes the canonical outcome returned by each intercept, including how Relay
+uses it to resolve the provider request and schedule lifecycle marks.
+
+A canonical outcome serialization looks like this:
+
+```json
+{
+  "request": {"headers": {}, "content": {}},
+  "annotated_request": null,
+  "pending_marks": []
+}
+```
+
+`request` is required. `annotated_request` defaults to `null` when omitted on
+input, and `pending_marks` defaults to an empty list. Canonical serialization
+includes all three fields. A pending mark only contains `name`, optional
+`category` and `category_profile`, and optional `data` and `metadata`. Relay
+owns event UUIDs, parent UUIDs, and timestamps.
+
+## Request Authority
+
+The provider-body source of truth only depends on whether a request codec is
+active:
+
+Request codecs translate provider-specific request payloads into Relay's
+normalized annotated request for intercepts, then encode accepted annotated
+edits back into the provider request before execution. They normalize the
+payload shape rather than translating between providers; response codecs are a
+separate response-side path used to attach normalized data to lifecycle events.
+
+| Request codec | Provider body source | Header source |
+| --- | --- | --- |
+| No codec | `outcome.request.content` | `outcome.request.headers` |
+| Active codec | `outcome.annotated_request` | `outcome.request.headers` |
+
+With an active codec, `request.content` is read-only context. Every intercept
+must return an annotation and make provider-body changes through that
+annotation, including its flattened `extra` fields for provider-specific data.
+Relay rejects a changed raw body or missing annotation at the offending
+intercept before invoking later middleware or creating an LLM lifecycle.
+
+The following diagram shows how Relay resolves an intercept outcome before
+managed execution.
+
+```mermaid
+flowchart TD
+    INPUT["Original LlmRequest"] --> CODEC{"Request codec active?"}
+
+    CODEC -->|No| RAWCHAIN["Run intercept chain"]
+    RAWCHAIN --> RAWPROVIDER["Provider receives outcome.request"]
+
+    CODEC -->|Yes| DECODE["Decode content into annotated_request"]
+    DECODE --> INTERCEPT["Invoke next intercept"]
+    INTERCEPT --> CHECKANN{"Annotation returned?"}
+    CHECKANN -->|No| FAIL["Fail before lifecycle"]
+    CHECKANN -->|Yes| CHECKRAW{"request.content unchanged?"}
+    CHECKRAW -->|No| FAIL
+    CHECKRAW -->|Yes| MORE{"More intercepts?"}
+    MORE -->|Yes| INTERCEPT
+    MORE -->|No| ENCODE["Encode final annotated_request"]
+    ENCODE --> HEADERS["Apply final request.headers"]
+    HEADERS --> PROVIDER["Provider receives one resolved LlmRequest"]
+```
+
+## Binding Contract
+
+The following callbacks return the same logical outcome in their native type
+or object shape:
+
+- Python callbacks return `LLMRequestInterceptOutcome`.
+- Rust callbacks return `LlmRequestInterceptOutcome`.
+- Go callbacks return `LLMRequestInterceptOutcome`.
+- Node.js and WebAssembly callbacks return `{ request, annotated?, pendingMarks? }`.
+  JavaScript pending-mark DTOs use `categoryProfile`; canonical JSON retains
+  `pending_marks` and `category_profile`.
+- Public C callbacks return one owned canonical outcome JSON string, and native
+  ABI v1 callbacks return one host-owned outcome JSON string.
+- Rust and Python `grpc-v1` worker SDKs return their canonical outcome in a
+  `JsonEnvelope` with schema `nemo.relay.LlmRequestInterceptOutcome@1`.
+
+The standalone request-intercept helper returns the complete outcome but does
+not emit its pending marks because it does not own an LLM lifecycle.
+
+## Managed Lifecycle
+
+Managed execution runs all effective global and scope-local intercepts before
+creating the LLM handle. Each accepted request and annotation pair feeds the
+next intercept under the authority rules above, while pending marks append in
+middleware order. A breaking intercept retains the marks it returned. If any
+intercept fails or its boundary result is malformed, Relay discards all
+accumulated marks and creates no LLM lifecycle.
+
+After successful interception, Relay creates the handle and captures one
+subscriber snapshot. It emits the LLM start at `T`, every pending mark at
+`T + 1µs` in returned order with the LLM UUID as parent, and the LLM end no
+earlier than `T + 1µs`. Streaming and non-streaming calls use the same rules.
+Pending marks are never added to the provider request, annotated request,
+codec input, sanitizer input, or start payload.
+
+## Migration
+
+This finalizes unpublished native ABI v1 and `grpc-v1` contracts. Rebuild all
+development native plugins and workers. Replace tuple results, split C/Go
+outputs, metadata envelopes, and parallel mark-aware registrations with the
+canonical outcome and the existing `register_llm_request_intercept`
+registration name.
+
+## Related Topics
+
+- [Tool Execution Intercept Outcomes](/reference/tool-execution-intercept-outcomes)