Merge pull request #936 from NVIDIA/release/0.8

Forward-merge release/0.8 into main

Author: GPUtester

Cargo.lock

@@ -68,6 +68,12 @@ shared persistence:
 cargo add nemo-relay-adaptive --features redis-backend
 ```
 
+The response cache is opt-in. Its required `namespace` identifies one trusted
+cache-sharing domain and must not be reused across mutually untrusted tenants
+or upstream authorities. When tenants share a backend, include every
+answer-affecting tenant or routing header in `header_allowlist`; Relay cannot
+infer application-owned tenant identity.
+
 For local source development:
 
 ```bash

crates/adaptive/README.md

@@ -1099,7 +1099,7 @@ fn cache_error_policy_partitions_tool_keys() {
 
 #[test]
 fn header_allowlist_policy_partitions_keys_and_normalizes_case() {
-    let request = request(json!({
+    let mut request = request(json!({
         "model": "gpt-4o",
         "messages": [{"role": "user", "content": "hi"}],
         "temperature": 0.0,
@@ -1124,6 +1124,19 @@ fn header_allowlist_policy_partitions_keys_and_normalizes_case() {
         key_of("openai", &request, &duplicate_spelling),
         "case-only and duplicate policy spellings are equivalent"
     );
+
+    request
+        .headers
+        .insert("x-tenant".to_string(), json!("tenant-a"));
+    let tenant_a = key_of("openai", &request, &tenant_partitioned);
+    request
+        .headers
+        .insert("x-tenant".to_string(), json!("tenant-b"));
+    assert_ne!(
+        tenant_a,
+        key_of("openai", &request, &tenant_partitioned),
+        "different allowlisted tenant identities must not share entries"
+    );
 }
 
 #[test]

crates/adaptive/tests/unit/response_cache/key_tests.rs

@@ -190,6 +190,13 @@ the dynamic plugin references in the selected physical `plugins.toml`. Dynamic
 plugins with a manifest-declared JSON Schema provide structured field controls.
 Other dynamic plugins use a raw JSON object editor.
 
+At runtime, dynamic plugins that do not have an explicit host-policy override
+are classified as required and must carry a valid signature from a configured
+trusted public key. Installation and inspection remain available for unsigned
+plugins, but Relay refuses to activate them. Worker processes inherit only the
+small environment allowlist needed for process startup and TLS. Native plugins
+run in-process and must be treated as trusted host code after verification.
+
 The canonical plugin file is `plugins.toml`; user config lives at
 `~/.config/nemo-relay/plugins.toml` or
 `$XDG_CONFIG_HOME/nemo-relay/plugins.toml`. Use

crates/cli/README.md

@@ -6,16 +6,16 @@ use serde_json::{Value, json};
 
 use crate::agents::shared::adapters::{
     AdapterOutcome, CLAUDE_CODE_PAYLOAD_EXTRACTOR, ClassificationRules, classify,
+    permission_request,
 };
-use crate::events::{AgentKind, NormalizedEvent};
+use crate::events::AgentKind;
 
 /// Normalizes Claude Code hook payloads and returns the hook response Claude expects.
 ///
-/// Claude Code uses permission-bearing tool hooks, so pre-tool events are explicitly allowed
-/// instead of returning the generic `{ continue: true }` shape. All other hooks acknowledge with
-/// `{ continue: true }` so the gateway remains observational and never blocks Claude's lifecycle
-/// by default. Note: Claude's hook output schema rejects `null` for optional string fields like
-/// `stopReason`; omit them entirely instead.
+/// Claude Code uses permission-bearing tool hooks. Pre-tool events acknowledge guardrail success
+/// without granting host permission; the later `PermissionRequest` receives the final decision.
+/// Note: Claude's hook output schema rejects `null` for optional string fields like `stopReason`;
+/// omit them entirely instead.
 pub(crate) fn adapt(payload: Value, headers: &HeaderMap) -> AdapterOutcome {
     let events = classify(
         &payload,
@@ -40,17 +40,15 @@ pub(crate) fn adapt(payload: Value, headers: &HeaderMap) -> AdapterOutcome {
             ],
         },
     );
-    // Response shape is decided by the primary event (first in the vec); secondary events like
-    // `TurnEnded` are observability-only and don't influence the hook response Claude gets back.
-    let response = match events.first() {
-        Some(NormalizedEvent::ToolStarted(_)) => json!({
-            "continue": true,
-            "hookSpecificOutput": {
-                "hookEventName": "PreToolUse",
-                "permissionDecision": "allow"
-            }
-        }),
-        _ => json!({ "continue": true }),
-    };
-    AdapterOutcome { events, response }
+    let response = json!({ "continue": true });
+    AdapterOutcome {
+        events,
+        response,
+        permission: permission_request(
+            &payload,
+            headers,
+            AgentKind::ClaudeCode,
+            &CLAUDE_CODE_PAYLOAD_EXTRACTOR,
+        ),
+    }
 }

crates/cli/src/agents/claude/adapter.rs

@@ -36,10 +36,13 @@ pub(crate) fn prepare(
             [
                 "--plugin-dir".into(),
                 "<temporary-claude-plugin-dir>".into(),
-                "--settings".into(),
-                "<temporary-claude-settings>".into(),
             ],
         );
+        insert_before_argument_boundary(
+            &mut launch.argv,
+            launch.host_index,
+            ["--settings".into(), "<temporary-claude-settings>".into()],
+        );
         launch
             .env
             .push(("ANTHROPIC_BASE_URL".into(), gateway_url.to_string()));
@@ -80,12 +83,12 @@ pub(crate) fn prepare(
     insert_after_host(
         &mut launch.argv,
         launch.host_index,
-        [
-            "--plugin-dir".into(),
-            root.display().to_string(),
-            "--settings".into(),
-            settings_path.display().to_string(),
-        ],
+        ["--plugin-dir".into(), root.display().to_string()],
+    );
+    insert_before_argument_boundary(
+        &mut launch.argv,
+        launch.host_index,
+        ["--settings".into(), settings_path.display().to_string()],
     );
     launch
         .env
@@ -94,6 +97,19 @@ pub(crate) fn prepare(
     Ok(())
 }
 
+fn insert_before_argument_boundary(
+    argv: &mut Vec<String>,
+    host_index: usize,
+    values: impl IntoIterator<Item = String>,
+) {
+    let boundary = argv
+        .iter()
+        .skip(host_index + 1)
+        .position(|argument| argument == "--")
+        .map_or(argv.len(), |offset| host_index + 1 + offset);
+    argv.splice(boundary..boundary, values);
+}
+
 fn replace_custom_header(existing: &str, replacement: &str) -> String {
     let replacement_name = replacement
         .split_once(':')

crates/cli/src/agents/claude/launch.rs

@@ -5,7 +5,7 @@ use axum::http::HeaderMap;
 use serde_json::{Value, json};
 
 use crate::agents::shared::adapters::{
-    AdapterOutcome, CODEX_PAYLOAD_EXTRACTOR, ClassificationRules, classify,
+    AdapterOutcome, CODEX_PAYLOAD_EXTRACTOR, ClassificationRules, classify, permission_request,
 };
 use crate::events::AgentKind;
 
@@ -32,5 +32,11 @@ pub(crate) fn adapt(payload: Value, headers: &HeaderMap) -> AdapterOutcome {
     AdapterOutcome {
         events,
         response: json!({}),
+        permission: permission_request(
+            &payload,
+            headers,
+            AgentKind::Codex,
+            &CODEX_PAYLOAD_EXTRACTOR,
+        ),
     }
 }

crates/cli/src/agents/codex/adapter.rs

@@ -28,6 +28,8 @@ pub(crate) struct AdapterOutcome {
     pub(crate) events: Vec<NormalizedEvent>,
     /// Hook response body returned to the invoking agent process.
     pub(crate) response: Value,
+    /// Final permission request evaluated separately so its observable hook event stays unchanged.
+    pub(crate) permission: Option<Result<ToolEvent, String>>,
 }
 
 pub(super) struct ClassificationRules<'a> {
@@ -221,6 +223,57 @@ pub(crate) struct ToolPathSet {
     status: &'static [&'static [&'static str]],
 }
 
+pub(super) fn permission_request(
+    payload: &Value,
+    headers: &HeaderMap,
+    kind: AgentKind,
+    extractor: &dyn AgentPayloadExtractor,
+) -> Option<Result<ToolEvent, String>> {
+    let event_name = extractor.event_name(payload)?;
+    if normalize_name(&event_name) != "permissionrequest" {
+        return None;
+    }
+    let session_id = match extractor.session_id(payload, headers) {
+        Some(value) if !value.trim().is_empty() => value,
+        _ => {
+            return Some(Err(
+                "permission request is missing a session identifier".into()
+            ));
+        }
+    };
+    let tool = extractor.tool_call(payload, headers, &event_name);
+    let tool_call_id = match tool.tool_call_id {
+        Some(value) if !value.trim().is_empty() => value,
+        None if kind == AgentKind::ClaudeCode => String::new(),
+        _ => {
+            return Some(Err(
+                "permission request is missing a tool-call identifier".into()
+            ));
+        }
+    };
+    let tool_name = match tool.tool_name {
+        Some(value) if !value.trim().is_empty() => value,
+        _ => return Some(Err("permission request is missing a tool name".into())),
+    };
+    let arguments = match tool.arguments {
+        Some(value) => value,
+        None => return Some(Err("permission request is missing tool arguments".into())),
+    };
+    Some(Ok(ToolEvent {
+        session_id,
+        agent_kind: kind,
+        event_name: event_name.clone(),
+        tool_call_id,
+        tool_name,
+        subagent_id: tool.subagent_id,
+        arguments,
+        result: Value::Null,
+        status: tool.status,
+        payload: payload.clone(),
+        metadata: extractor.metadata(payload, headers, kind, &event_name),
+    }))
+}
+
 /// Whether an extractor accepts the Claude installed-mode session header.
 #[derive(Clone, Copy)]
 pub(crate) enum SessionHeaderPolicy {

crates/cli/src/agents/shared/adapters.rs

@@ -216,6 +216,7 @@ impl ProviderRequestExtractor for AnthropicCountTokensRequestExtractor {
 pub(crate) struct SessionAlias {
     pub(crate) parent_session_id: String,
     pub(crate) subagent_id: String,
+    authenticated_owner: Option<String>,
     // Metadata explains why this alias exists and is stamped on rewritten events. Phoenix traces
     // then stay filterable/debuggable even after the event has been moved under its parent scope.
     metadata: Value,
@@ -228,6 +229,7 @@ impl SessionAlias {
         Self {
             parent_session_id,
             subagent_id,
+            authenticated_owner: None,
             metadata,
         }
     }
@@ -237,6 +239,14 @@ impl SessionAlias {
     pub(crate) fn metadata(&self) -> Value {
         self.metadata.clone()
     }
+
+    pub(crate) fn set_authenticated_owner(&mut self, owner: Option<String>) {
+        self.authenticated_owner = owner;
+    }
+
+    pub(crate) fn authenticated_owner(&self) -> Option<&str> {
+        self.authenticated_owner.as_deref()
+    }
 }
 
 #[derive(Debug, Clone)]
@@ -245,6 +255,7 @@ pub(crate) struct PendingSubagentStart {
     // hook or gateway request, after this hook request has already returned.
     pub(crate) event: SessionEvent,
     context: SubagentSessionContext,
+    authenticated_owner: Option<String>,
 }
 
 impl PendingSubagentStart {
@@ -259,6 +270,14 @@ impl PendingSubagentStart {
     pub(crate) fn alias_for_child_session(&self, child_session_id: String) -> SessionAlias {
         alias_for_child_session(child_session_id, &self.context)
     }
+
+    pub(crate) fn set_authenticated_owner(&mut self, owner: Option<String>) {
+        self.authenticated_owner = owner;
+    }
+
+    pub(crate) fn authenticated_owner(&self) -> Option<&str> {
+        self.authenticated_owner.as_deref()
+    }
 }
 
 // Owns all cross-session correlation state used by the session manager. Keeping aliases and
@@ -551,6 +570,7 @@ pub(crate) async fn pending_subagent_start(
         PendingSubagentStart {
             event: session_event.clone(),
             context,
+            authenticated_owner: None,
         },
     ))
 }

crates/cli/src/agents/shared/alignment.rs

@@ -475,12 +475,14 @@ const BOOTSTRAP_HMAC_KEY_BYTES: usize = 32;
 const BOOTSTRAP_HMAC_LOCK_TIMEOUT: Duration = Duration::from_secs(5);
 const BOOTSTRAP_CHALLENGE_DOMAIN: &[u8] = b"nemo-relay/bootstrap-health/v1\0";
 const BOOTSTRAP_CLIENT_TOKEN_DOMAIN: &[u8] = b"nemo-relay/bootstrap-client/v1\0";
+const HOOK_CLIENT_TOKEN_DOMAIN: &[u8] = b"nemo-relay/hook-client/v1\0";
 const TRANSPARENT_GATEWAY_DOMAIN: &[u8] = b"nemo-relay/transparent-gateway/v1\0";
 const PYTHON_ENVIRONMENT_ATTESTATION_DOMAIN: &[u8] =
     b"nemo-relay/python-environment-attestation/v1\0";
 
 /// Private proof installed into supported coding-agent provider configuration.
 pub(crate) const BOOTSTRAP_CLIENT_TOKEN_HEADER: &str = "x-nemo-relay-client-token";
+pub(crate) const HOOK_CLIENT_TOKEN_HEADER: &str = "x-nemo-relay-hook-client";
 
 /// Stable health-proof context shared by a transparent wrapper and plugin-owned MCP client.
 pub(crate) fn transparent_gateway_fingerprint(gateway_url: &str) -> String {
@@ -558,6 +560,33 @@ impl BootstrapChallengeKey {
         hmac::verify(&self.0, BOOTSTRAP_CLIENT_TOKEN_DOMAIN, &tag).is_ok()
     }
 
+    pub(crate) fn hook_client_token(&self, identity: &str) -> String {
+        let identity = digest::digest(&digest::SHA256, identity.as_bytes())
+            .as_ref()
+            .iter()
+            .map(|byte| format!("{byte:02x}"))
+            .collect::<String>();
+        let mut context = hmac::Context::with_key(&self.0);
+        context.update(HOOK_CLIENT_TOKEN_DOMAIN);
+        context.update(identity.as_bytes());
+        format!("{identity}.{}", encode_hmac_tag(context.sign()))
+    }
+
+    pub(crate) fn verify_hook_client_token(&self, token: &str) -> Option<String> {
+        let (identity, signature) = token.split_once('.')?;
+        if identity.len() != 64 || !identity.bytes().all(|byte| byte.is_ascii_hexdigit()) {
+            return None;
+        }
+        let encoded = signature.strip_prefix("hmac-sha256:")?;
+        let tag = decode_fixed_hex::<32>(encoded)?;
+        let mut message = Vec::with_capacity(HOOK_CLIENT_TOKEN_DOMAIN.len() + identity.len());
+        message.extend_from_slice(HOOK_CLIENT_TOKEN_DOMAIN);
+        message.extend_from_slice(identity.as_bytes());
+        hmac::verify(&self.0, &message, &tag)
+            .is_ok()
+            .then(|| format!("hook-client:{identity}"))
+    }
+
     #[cfg(test)]
     pub(crate) fn from_bytes(bytes: &[u8]) -> Self {
         Self(hmac::Key::new(hmac::HMAC_SHA256, bytes))