Support advanced prompt voice agent settings

cli/azd/extensions/azure.ai.agents/cspell.yaml

@@ -11,6 +11,9 @@ words:
   # Voice (prompt-voice) agents
   - BYOM
   - Nanami
+  - pcma
+  - pcmu
+  - webrtc
   # Azure region names
   - australiaeast
   - brazilsouth

cli/azd/extensions/azure.ai.agents/internal/cmd/doctor/checks_agent_status.go

@@ -704,9 +704,10 @@ func readAgentServices(prior []Result) []string {
 
 // filterHostedAgentServices removes prompt-voice services from the hosted-agent
 // status probe. The remote.agent-status check targets hosted agent versions
-// (`/agents/{name}/versions/{version}`) and relies on AGENT_<KEY>_VERSION; a
-// prompt-voice deploy writes NAME+ENDPOINT only and should be covered by a
-// voice-specific doctor check in a future PR.
+// (`/agents/{name}/versions/{version}`). Unified prompt-voice deploys record a
+// VERSION, but voice readiness follows its WebSocket endpoint rather than the
+// hosted-agent version lifecycle and should be covered by a voice-specific
+// doctor check in a future PR.
 func filterHostedAgentServices(ctx context.Context, azdClient *azdext.AzdClient, services []string) []string {
 	if len(services) == 0 || azdClient == nil {
 		return services

cli/azd/extensions/azure.ai.agents/internal/cmd/nextstep/state.go

@@ -36,9 +36,9 @@ const (
 	agentVersionVarFormat = "AGENT_%s_VERSION"
 
 	// agentEndpointVarFormat is the base endpoint env-var written for every
-	// deployed agent. Voice agents (kind: prompt-voice) are created
-	// synchronously with no agent-version object, so this base endpoint is the
-	// only deployment marker they set — isDeployed falls back to it.
+	// deployed agent. Voice agents (kind: prompt-voice) use it as the deploy
+	// completion marker. Unified voice deploys also set VERSION, while legacy
+	// voice environments may have only this endpoint marker.
 	agentEndpointVarFormat = "AGENT_%s_ENDPOINT"
 
 	// projectEndpointVar is the env-var that carries the Foundry project
@@ -822,23 +822,22 @@ func isDeployed(
 		*errs = append(*errs, fmt.Errorf("read %s: %w", key, err))
 		return false
 	}
-	if value != "" {
-		return true
-	}
-
-	// Voice agents (kind: prompt-voice) deploy without an agent-version object,
-	// so they never set AGENT_<KEY>_VERSION. Fall back to the base endpoint
-	// marker, which every voice deploy writes, so a successfully created voice
-	// agent is not reported as undeployed. Gate this on the service's actual
-	// declared kind: a hosted agent whose deploy partially failed can also
-	// present an empty VERSION with a lingering ENDPOINT, and must stay reported
-	// as not-deployed. This mirrors the kind gate in
+	if !isVoice {
+		return value != ""
+	}
+
+	// Voice deploys use the base ENDPOINT env var as the completion marker.
+	// Unified voice deploys write VERSION before ENDPOINT to keep ENDPOINT as the
+	// final marker.
+	// Require ENDPOINT for voice even when VERSION is present, otherwise a partial
+	// env write could be reported as deployed before the callable endpoint was
+	// persisted. Gate this on the service's actual declared kind: a hosted agent
+	// whose deploy partially failed can also present an empty VERSION with a
+	// lingering ENDPOINT, and must stay reported as not-deployed. This mirrors the
+	// kind gate in
 	// AgentServiceTargetProvider.Endpoints (project package); the two live in
 	// separate packages because project imports nextstep, so a literally shared
 	// helper would create an import cycle.
-	if !isVoice {
-		return false
-	}
 	endpointKey := fmt.Sprintf(agentEndpointVarFormat, serviceKey(serviceName))
 	endpointValue, err := src.EnvValue(ctx, envName, endpointKey)
 	if err != nil {

cli/azd/extensions/azure.ai.agents/internal/cmd/nextstep/state_test.go

@@ -739,10 +739,10 @@ func TestServiceKey(t *testing.T) {
 	}
 }
 
-// TestIsDeployed_VoiceEndpointFallback verifies that a voice agent — which sets
-// only AGENT_<KEY>_NAME and AGENT_<KEY>_ENDPOINT, never AGENT_<KEY>_VERSION — is
-// still reported as deployed via the base endpoint marker, while an agent with
-// neither version nor endpoint is reported undeployed.
+// TestIsDeployed_VoiceEndpointFallback verifies that voice readiness is based
+// on the base endpoint marker. Legacy voice agents never set VERSION, while
+// unified voice agents set VERSION before ENDPOINT; in both cases ENDPOINT is
+// the deploy completion marker.
 func TestIsDeployed_VoiceEndpointFallback(t *testing.T) {
 	t.Parallel()
 
@@ -757,6 +757,21 @@ func TestIsDeployed_VoiceEndpointFallback(t *testing.T) {
 			values: map[string]string{"env1/AGENT_VOICE_SVC_VERSION": "1"},
 			want:   true,
 		},
+		{
+			name:    "version set but endpoint missing: undeployed (voice agent partial write)",
+			values:  map[string]string{"env1/AGENT_VOICE_SVC_VERSION": "1"},
+			isVoice: true,
+			want:    false,
+		},
+		{
+			name: "version and endpoint set: deployed (unified voice agent)",
+			values: map[string]string{
+				"env1/AGENT_VOICE_SVC_VERSION":  "1",
+				"env1/AGENT_VOICE_SVC_ENDPOINT": "wss://x/agents/a/endpoint/protocols/voice?api-version=v1",
+			},
+			isVoice: true,
+			want:    true,
+		},
 		{
 			name:    "no version but base endpoint set: deployed (voice agent)",
 			values:  map[string]string{"env1/AGENT_VOICE_SVC_ENDPOINT": "https://x/voice_agents/a"},

cli/azd/extensions/azure.ai.agents/internal/pkg/agents/agent_api/models.go

@@ -69,10 +69,9 @@ const (
 	AgentKindHosted   AgentKind = "hosted"
 	AgentKindWorkflow AgentKind = "workflow"
 	// AgentKindVoice is the data-plane (service) kind for a declarative voice
-	// (speech-to-speech) agent. Note: this is the wire value posted to the
-	// /voice_agents collection. The azd manifest authoring kind is
-	// "prompt-voice" (agent_yaml.AgentKindPromptVoice), which the map layer
-	// translates to this value.
+	// (speech-to-speech) agent. The azd manifest authoring kind is "prompt-voice"
+	// (agent_yaml.AgentKindPromptVoice), which the map layer translates to this
+	// value before posting through the unified /agents API.
 	AgentKindVoice AgentKind = "voice"
 )
 
@@ -360,7 +359,7 @@ const (
 // VoiceAudioFormat describes a PCM audio stream format (e.g. audio/pcm @ 24 kHz).
 type VoiceAudioFormat struct {
 	Type string `json:"type"`
-	Rate int    `json:"rate"`
+	Rate *int   `json:"rate,omitempty"`
 }
 
 // VoiceTurnDetection configures server-side voice-activity detection so the
@@ -370,32 +369,69 @@ type VoiceTurnDetection struct {
 	Threshold         *float64 `json:"threshold,omitempty"`
 	PrefixPaddingMs   *int     `json:"prefix_padding_ms,omitempty"`
 	SilenceDurationMs *int     `json:"silence_duration_ms,omitempty"`
+	CreateResponse    *bool    `json:"create_response,omitempty"`
+	Eagerness         *string  `json:"eagerness,omitempty"`
+	SpeechDurationMs  *int     `json:"speech_duration_ms,omitempty"`
+	RemoveFillerWords *bool    `json:"remove_filler_words,omitempty"`
+	InterruptResponse *bool    `json:"interrupt_response,omitempty"`
+	Languages         []string `json:"languages,omitempty"`
+	AutoTruncate      *bool    `json:"auto_truncate,omitempty"`
 }
 
 // VoiceTranscription enables user-speech transcription events on the input stream.
 type VoiceTranscription struct {
-	Model string `json:"model,omitempty"`
+	Model    string  `json:"model,omitempty"`
+	Language *string `json:"language,omitempty"`
+	Prompt   *string `json:"prompt,omitempty"`
+}
+
+// VoiceNoiseReduction configures input audio noise reduction.
+type VoiceNoiseReduction struct {
+	Type string `json:"type"`
 }
 
 // VoiceInputConfig is the input (caller -> agent) audio configuration.
 type VoiceInputConfig struct {
-	Format        *VoiceAudioFormat   `json:"format,omitempty"`
-	TurnDetection *VoiceTurnDetection `json:"turn_detection,omitempty"`
-	Transcription *VoiceTranscription `json:"transcription,omitempty"`
+	Format           *VoiceAudioFormat    `json:"format,omitempty"`
+	NoiseReduction   *VoiceNoiseReduction `json:"noise_reduction,omitempty"`
+	EchoCancellation map[string]any       `json:"echo_cancellation,omitempty"`
+	TurnDetection    *VoiceTurnDetection  `json:"turn_detection,omitempty"`
+	Transcription    *VoiceTranscription  `json:"transcription,omitempty"`
 }
 
 // VoiceConfig selects the output voice. Type is "openai" for realtime voices
 // (single lowercase word, e.g. "alloy") or "azure_standard" for Azure Neural
 // voices (e.g. "en-US-Ava:DragonHDLatestNeural").
 type VoiceConfig struct {
-	Type string `json:"type"`
-	Name string `json:"name"`
+	Type   string  `json:"type"`
+	Name   string  `json:"name"`
+	Style  *string `json:"style,omitempty"`
+	Pitch  *string `json:"pitch,omitempty"`
+	Rate   *string `json:"rate,omitempty"`
+	Locale *string `json:"locale,omitempty"`
+	Volume *string `json:"volume,omitempty"`
 }
 
 // VoiceOutputConfig is the output (agent -> caller) audio configuration.
 type VoiceOutputConfig struct {
 	Format *VoiceAudioFormat `json:"format,omitempty"`
 	Voice  *VoiceConfig      `json:"voice,omitempty"`
+	Speed  *float64          `json:"speed,omitempty"`
+}
+
+// VoiceOutputConfigFlat is the newer Voice Live output shape used by the
+// unified /agents voice API in TiP. Older regions still accept/return the
+// object-shaped VoiceOutputConfig above.
+type VoiceOutputConfigFlat struct {
+	Format      *VoiceAudioFormat `json:"format,omitempty"`
+	Voice       string            `json:"voice,omitempty"`
+	VoiceType   string            `json:"voice_type,omitempty"`
+	VoiceLocale string            `json:"voice_locale,omitempty"`
+	Style       *string           `json:"style,omitempty"`
+	Pitch       *string           `json:"pitch,omitempty"`
+	Rate        *string           `json:"rate,omitempty"`
+	Volume      *string           `json:"volume,omitempty"`
+	Speed       *float64          `json:"speed,omitempty"`
 }
 
 // VoiceAudioConfig bundles the input and output audio configuration.
@@ -404,17 +440,54 @@ type VoiceAudioConfig struct {
 	Output *VoiceOutputConfig `json:"output,omitempty"`
 }
 
-// VoiceAgentDefinition is the data-plane definition body POSTed to the
-// /voice_agents collection for a declarative (managed) voice agent. Its Kind
-// is always AgentKindVoice ("voice").
+// VoiceAudioConfigFlat bundles voice audio config with the flat output shape.
+type VoiceAudioConfigFlat struct {
+	Input  *VoiceInputConfig      `json:"input,omitempty"`
+	Output *VoiceOutputConfigFlat `json:"output,omitempty"`
+}
+
+// VoiceAgentDefinition is retained for compatibility with object-shaped voice
+// definitions returned by older services. New prompt voice deploys use
+// VoiceAgentDefinitionFlat.
 type VoiceAgentDefinition struct {
 	AgentDefinition
-	ModelType        VoiceModelType    `json:"model_type"`
-	Model            string            `json:"model"`
-	Instructions     string            `json:"instructions,omitempty"`
-	Audio            *VoiceAudioConfig `json:"audio,omitempty"`
-	OutputModalities []string          `json:"output_modalities,omitempty"`
-	Store            *bool             `json:"store,omitempty"`
+	ModelType         VoiceModelType    `json:"model_type"`
+	Model             string            `json:"model"`
+	Instructions      string            `json:"instructions,omitempty"`
+	StructuredInputs  map[string]any    `json:"structured_inputs,omitempty"`
+	Audio             *VoiceAudioConfig `json:"audio,omitempty"`
+	OutputModalities  []string          `json:"output_modalities,omitempty"`
+	Store             *bool             `json:"store,omitempty"`
+	Tools             []map[string]any  `json:"tools,omitempty"`
+	Avatar            map[string]any    `json:"avatar,omitempty"`
+	Greeting          map[string]any    `json:"greeting,omitempty"`
+	Handoff           map[string]any    `json:"handoff,omitempty"`
+	ToolChoice        any               `json:"tool_choice,omitempty"`
+	ParallelToolCalls *bool             `json:"parallel_tool_calls,omitempty"`
+	MaxOutputTokens   any               `json:"max_output_tokens,omitempty"`
+	Include           []string          `json:"include,omitempty"`
+}
+
+// VoiceAgentDefinitionFlat is the voice definition shape aligned with the
+// unified /agents TiP API, where audio.output.voice is a string and the voice
+// provider details are sibling fields.
+type VoiceAgentDefinitionFlat struct {
+	AgentDefinition
+	ModelType         VoiceModelType        `json:"model_type"`
+	Model             string                `json:"model"`
+	Instructions      string                `json:"instructions,omitempty"`
+	StructuredInputs  map[string]any        `json:"structured_inputs,omitempty"`
+	Audio             *VoiceAudioConfigFlat `json:"audio,omitempty"`
+	OutputModalities  []string              `json:"output_modalities,omitempty"`
+	Store             *bool                 `json:"store,omitempty"`
+	Tools             []map[string]any      `json:"tools,omitempty"`
+	Avatar            map[string]any        `json:"avatar,omitempty"`
+	Greeting          map[string]any        `json:"greeting,omitempty"`
+	Handoff           map[string]any        `json:"handoff,omitempty"`
+	ToolChoice        any                   `json:"tool_choice,omitempty"`
+	ParallelToolCalls *bool                 `json:"parallel_tool_calls,omitempty"`
+	MaxOutputTokens   any                   `json:"max_output_tokens,omitempty"`
+	Include           []string              `json:"include,omitempty"`
 }
 
 // CreateAgentVersionRequest represents a request to create an agent version