rasa-pro 3.20.0.dev2__tar.gz → 3.20.0.dev3__tar.gz

{rasa_pro-3.20.0.dev2 → rasa_pro-3.20.0.dev3}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.3
 Name: rasa-pro
-Version: 3.20.0.dev2
+Version: 3.20.0.dev3
 Summary: State-of-the-art open-core Conversational AI framework for Enterprises that natively leverages generative AI for effortless assistant development.
 Keywords: nlp,machine-learning,machine-learning-library,bot,bots,botkit,rasa conversational-agents,conversational-ai,chatbot,chatbot-framework,bot-framework
 Author: Rasa Technologies GmbH

{rasa_pro-3.20.0.dev2 → rasa_pro-3.20.0.dev3}/pyproject.toml

@@ -4,7 +4,7 @@ build-backend = "poetry.core.masonry.api"
 
 [tool.poetry]
 name = "rasa-pro"
-version = "3.20.0.dev2"
+version = "3.20.0.dev3"
 description = "State-of-the-art open-core Conversational AI framework for Enterprises that natively leverages generative AI for effortless assistant development."
 authors = [ "Rasa Technologies GmbH <hi@rasa.com>",]
 maintainers = [ "Tom Bocklisch <tom@rasa.com>",]

{rasa_pro-3.20.0.dev2 → rasa_pro-3.20.0.dev3}/rasa/cli/tools/skills_data/mantle/mantle-building-from-conversations/SKILL.md

@@ -12,7 +12,7 @@ engine: mantle
 rasa_version: ">=3.18"
 metadata:
   author: rasa
-  version: "0.1.1"
+  version: "0.1.2"
   docs-url: https://rasa-2f7eb63d.mintlify.site
 ---
 
@@ -95,8 +95,11 @@ account?", "verify identity" — is not part of any one goal. Extract it as a
   memory key. **Declare every key a tool writes in `memory.yml`** — `rasa train`
   rejects an undeclared `context.memory.set()` (`undeclared_memory_write`).
 - **Verbatim responses only for wording that must be exact.** Recording notices,
-  legal disclosures, regulated confirmations → `responses.yml`. Everything else stays
-  prose so the agent sounds natural, not scripted.
+  legal disclosures, regulated confirmations → `responses.yml`. When the **same**
+  mandated line must change with a fact already in memory (fee copy for lost vs
+  stolen, tier-specific disclaimers), use **conditional variants** — multiple
+  entries under one response name with `condition:` — rather than separate response
+  names. Everything else stays prose so the agent sounds natural, not scripted.
 - **Control levers only for branching the business requires.** Default to prose and
   let the LLM sequence. Escalate to a lever only when a transcript shows behavior that
   MUST be guaranteed: a tool that fired only after inputs existed (`tool_constraints`
@@ -145,8 +148,10 @@ don't recognize" — become two skills, `replace_card` and `dispute_charge`, plu
 shared `select_card` sub-skill for the "which card?" decision both make. `replace_card`
 gets prose for the happy path, a `get_customer_cards` + `order_replacement` tool pair,
 a `replacement_reason` categorical memory key (the transcripts branch on
-lost/stolen/damaged), and a verbatim lock disclosure for the stolen path. Propose that
-breakdown, wait for the nod, then generate.
+lost/stolen/damaged), a verbatim lock disclosure for the stolen path, and — if fee
+wording differs by reason — one `responses.yml` response with a `condition` per
+case instead of separate response names. Propose that breakdown, wait for the nod,
+then generate.
 
 ## Bridge to evaluation
 
@@ -182,3 +187,4 @@ Most relevant here:
 - [skill.md reference](https://rasa-2f7eb63d.mintlify.site/reference/skill-md): the full frontmatter and body contract
 - [Progressive control guides](https://rasa-2f7eb63d.mintlify.site/build-guide/tool-constraints): choosing the narrowest lever for an observed drift
 - [Conditions](https://rasa-2f7eb63d.mintlify.site/reference/conditions): the expression grammar
+- [responses.yml reference](https://rasa-2f7eb63d.mintlify.site/reference/responses-yml): verbatim templates and conditional variants

{rasa_pro-3.20.0.dev2 → rasa_pro-3.20.0.dev3}/rasa/cli/tools/skills_data/mantle/mantle-building-skills/SKILL.md

@@ -10,7 +10,7 @@ engine: mantle
 rasa_version: ">=3.18"
 metadata:
   author: rasa
-  version: "0.1.3"
+  version: "0.1.4"
   docs-url: https://rasa-2f7eb63d.mintlify.site
 ---
 
@@ -166,6 +166,7 @@ don't add levers speculatively.
 | Calls a tool before its inputs exist | `tool_constraints` → `requires:` | frontmatter |
 | Executes an irreversible action without asking | In-flow prose confirmation + `requires:` on a confirmation memory key (e.g. `order_confirmed`), or `requires_confirmation:` on the tool in `tool_constraints` | body + frontmatter |
 | Fixed wording after a gated tool succeeds or fails | `on_success:` / `on_failure:` on the tool + templates in `responses.yml` | frontmatter |
+| Same response name, exact wording must change with memory | `condition` on variants in `responses.yml` | `responses.yml` |
 | Follows the wrong branch / mixes workflows | one `if:` marker per case | body |
 | Paraphrases wording that must be exact (legal, compliance) | `utter:` triggers + `responses.yml` | frontmatter |
 | Skips or reorders steps where order IS the requirement | ordered block | body |
@@ -318,6 +319,49 @@ utter:
 Tool outcome utters (`on_success:` / `on_failure:`) are separate: they attach to a
 gated tool in `tool_constraints`, not to skill `utter:` frontmatter.
 
+### Conditional response variants
+
+When one response **name** must deliver different exact wording depending on memory
+(for example lost vs stolen replacement-fee copy), declare **multiple variants**
+under that name in `responses.yml` instead of splitting into separate response names
+or duplicating `utter:` triggers.
+
+```yaml skills/card_replace/responses.yml
+responses:
+  utter_card_replacement_fee:
+    - text: Standard replacement is free.
+    - text: A fee applies for lost-card replacement.
+      condition: 'session.card_replace.replacement_reason == "lost"'
+    - text: No fee for cards reported stolen.
+      condition: 'session.card_replace.replacement_reason == "stolen"'
+```
+
+**Selection at delivery** (YAML order):
+
+1. Evaluate conditioned variants; the first whose `condition` is true wins.
+2. If none match, deliver the single **default** — the one variant that omits
+   `condition` (it may sit first, last, or between conditioned variants).
+3. If every variant is conditioned and none match, delivery fails closed (same as a
+   missing response).
+
+Use a skill `utter:` **`when:`** when a **whole response** should fire once the
+first time a condition becomes true. Use a variant **`condition`** when any
+reference to that response name (action step, collect `utterance:`, tool
+`on_success:` / confirmation ask, `on: activate`, etc.) must pick wording from
+memory at that moment. Same
+[condition grammar](https://rasa-2f7eb63d.mintlify.site/reference/conditions) as
+`requires:` and prose `if:`.
+
+Rules:
+
+- When any variant has a `condition`, exactly **one** variant must omit it.
+- Conditions reference readable memory as `session.<skill_id>.<entry>` or
+  `session.project.<entry>`.
+- Only `text`, `metadata`, and `condition` on each variant — no `channel:` or
+  `buttons:`.
+- `metadata.rephrase: true` applies to the **selected** variant only; keep it off
+  for compliance/legal text.
+
 ### Ordered blocks — last resort, not first
 
 Most skills never need one. Reach for a block only when the *sequence itself* is the
@@ -405,4 +449,5 @@ Most relevant here:
 - [Ordered block steps](https://rasa-2f7eb63d.mintlify.site/reference/ordered-block-steps): step kinds, step fields, branching
 - [Tools reference](https://rasa-2f7eb63d.mintlify.site/reference/tools): ToolContext, ToolResult, gating, post-write hooks, cancellation
 - [memory.yml reference](https://rasa-2f7eb63d.mintlify.site/reference/memory-yml): scopes, types, visibility, access control
+- [responses.yml reference](https://rasa-2f7eb63d.mintlify.site/reference/responses-yml): verbatim templates and conditional variants
 - [Progressive control guides](https://rasa-2f7eb63d.mintlify.site/build-guide/tool-constraints): one lever at a time, with worked examples

{rasa_pro-3.20.0.dev2 → rasa_pro-3.20.0.dev3}/rasa/cli/tools/skills_data/mantle/mantle-configuring-agent/SKILL.md

@@ -11,7 +11,7 @@ engine: mantle
 rasa_version: ">=3.18"
 metadata:
   author: rasa
-  version: "0.1.5"
+  version: "0.1.6"
   docs-url: https://rasa-2f7eb63d.mintlify.site
 ---
 
@@ -249,12 +249,54 @@ responses:
     - text: >-
         This interaction may be recorded for quality assurance
         and training purposes.
+  utter_card_replacement_fee:
+    - text: Standard replacement is free.
+    - text: A fee applies for lost-card replacement.
+      condition: 'session.card_replace.replacement_reason == "lost"'
   utter_replacement_failed:
     - text: >-
         We were unable to process your replacement. A support ticket has
         been created. Reference: {session.card_replace.ticket_id}.
 ```
 
+Each response name maps to a **list of variants**. At delivery the engine picks
+**one** variant: the first whose `condition` is true (YAML order), or the single
+**default** variant with no `condition` when none match.
+
+### Conditional variants
+
+Add `condition` on a variant when the **wording** of a response must change with
+memory — for example, different replacement-fee copy for lost vs stolen cards.
+When any variant is conditioned, exactly **one** variant must omit `condition`
+(the fallback). That fallback may sit anywhere in the list.
+
+```yaml
+responses:
+  utter_card_replacement_fee:
+    - text: Standard replacement is free.
+    - text: A fee applies for lost-card replacement.
+      condition: 'session.card_replace.replacement_reason == "lost"'
+    - text: No fee for cards reported stolen.
+      condition: 'session.card_replace.replacement_reason == "stolen"'
+```
+
+This is different from a skill `utter:` trigger with `when:` in `skill.md`:
+`when:` fires a **whole response** once when memory first becomes true;
+`condition` on a variant picks **which wording** to deliver each time something
+(action step, collect ask, tool outcome, confirmation prompt, `on: activate`,
+etc.) resolves that response name. Both use the same
+[condition expression grammar](https://rasa-2f7eb63d.mintlify.site/reference/conditions).
+
+Only `text`, `metadata`, and `condition` are permitted on a variant.
+`channel:` and `buttons:` are not supported. `metadata.rephrase: true` applies to
+**whichever variant is selected**, not always the first variant — leave rephrase
+off for compliance-sensitive wording.
+
+`rasa train` / `rasa data validate` check every project `responses.yml`: non-empty
+`text`, valid conditions, readable memory references, exactly one default when any
+variant is conditioned, and no unknown variant keys. Findings accumulate across
+all skill files and the project-root file.
+
 Use verbatim responses for wording that must be exact (legal, compliance,
 brand-mandated). Four things reference a response by name:
 
@@ -297,4 +339,4 @@ Most relevant here:
 - [agent.yml reference](https://rasa-2f7eb63d.mintlify.site/reference/agent-yml): every key, and which are top-level rather than nested
 - [integrations.yml reference](https://rasa-2f7eb63d.mintlify.site/reference/integrations-yml): LLM providers, channels, voice ASR/TTS, model groups
 - [memory.yml reference](https://rasa-2f7eb63d.mintlify.site/reference/memory-yml): skill schema vs project memory, types, access control
-- [responses.yml reference](https://rasa-2f7eb63d.mintlify.site/reference/responses-yml): templates, interpolation, and the four ways a response fires
+- [responses.yml reference](https://rasa-2f7eb63d.mintlify.site/reference/responses-yml): templates, conditional variants, interpolation, and the four ways a response fires

{rasa_pro-3.20.0.dev2 → rasa_pro-3.20.0.dev3}/rasa/cli/tools/skills_data/mantle/mantle-testing-debugging/SKILL.md

@@ -10,7 +10,7 @@ engine: mantle
 rasa_version: ">=3.18"
 metadata:
   author: rasa
-  version: "0.1.3"
+  version: "0.1.4"
   docs-url: https://rasa-2f7eb63d.mintlify.site
 ---
 
@@ -63,6 +63,7 @@ by rewriting prose louder. Find the symptom, apply the narrowest fix:
 | Tool ran but mandated success/failure wording missing | Outcome utter not wired | `on_success:` / `on_failure:` on the tool + `responses.yml` templates |
 | Wrong workflow branch followed / branches blended | All branches visible in prompt | one `if:` marker per case, each stating its own condition |
 | Compliance/legal wording paraphrased | LLM generating what must be verbatim | `utter:` trigger + `responses.yml` |
+| Mandated wording wrong for the situation (e.g. lost fee when reason is stolen) | Single-variant response or prose instead of memory-gated copy | `condition` on variants under one response name in `responses.yml`; ensure a default variant when any variant is conditioned |
 | Steps skipped/reordered where order is the requirement | Prose can't enforce sequence | ordered block for that section (`@block.<id>`), keep the rest prose |
 | Skill activates when it shouldn't | Routing on description alone | sharpen `description`; add skill-level `requires:` |
 | Skill doesn't activate when it should | Description doesn't match user phrasing | add trigger phrasings to `description` |
@@ -110,6 +111,10 @@ Frequent causes:
   prose (use `@memory.…` or move the reference into an `if:` condition)
 - a tool writing a memory key not declared in any `memory.yml`
   (`undeclared_memory_write`)
+- conditional response variants: empty `text`, invalid or empty `condition`,
+  unreadable memory in a `condition`, more than one default variant, conditioned
+  variants with no default, or unknown variant keys such as `channel:` / `buttons:`
+  (codes under `mantle.validation.response.*`)
 - `else:` or `elif:` in prose, which is rejected; write one `if:` paragraph per case
 - an ordered block whose id is a YAML key in the body rather than a fence attribute
   (`:::ordered_block id=pick_card`), or a tab-indented block body
@@ -150,3 +155,4 @@ Most relevant here:
 - [System prompt](https://rasa-2f7eb63d.mintlify.site/reference/system-prompt): which sections the model sees, and what each agent.yml prompts key overrides
 - [Prompt templates](https://rasa-2f7eb63d.mintlify.site/reference/prompt-templates): the assembled prompt and the framework tools offered per situation
 - [Tools reference](https://rasa-2f7eb63d.mintlify.site/reference/tools): why a tool may not appear in the schema
+- [responses.yml reference](https://rasa-2f7eb63d.mintlify.site/reference/responses-yml): conditional variants and train-time validation

rasa_pro-3.20.0.dev3/rasa/core/channels/voice_stream/tts/_client_session.py

@@ -0,0 +1,10 @@
+import ssl
+
+import aiohttp
+from aiohttp import ClientTimeout
+
+
+def create_ssl_client_session(timeout: ClientTimeout) -> aiohttp.ClientSession:
+    """Create an HTTP client session using the current certificate environment."""
+    connector = aiohttp.TCPConnector(ssl=ssl.create_default_context())
+    return aiohttp.ClientSession(timeout=timeout, connector=connector)

{rasa_pro-3.20.0.dev2 → rasa_pro-3.20.0.dev3}/rasa/core/channels/voice_stream/tts/azure.py

@@ -15,6 +15,9 @@ from rasa.core.channels.voice_stream.audio_bytes import (
     CurrentLanguageConfig,
     RasaAudioBytes,
 )
+from rasa.core.channels.voice_stream.tts._client_session import (
+    create_ssl_client_session,
+)
 from rasa.core.channels.voice_stream.tts.config import StreamingConfig
 from rasa.core.channels.voice_stream.tts.tts_engine import (
     TTSEngine,
@@ -102,7 +105,7 @@ class AzureTTS(TTSEngine[AzureTTSConfig]):
         super().__init__(rasa_language, format, config, additional_languages)
         timeout = ClientTimeout(total=self.config.timeout)
         if self.__class__.session is None or self.__class__.session.closed:
-            self.__class__.session = aiohttp.ClientSession(timeout=timeout)
+            self.__class__.session = create_ssl_client_session(timeout)
 
         self._use_streaming: bool = False
         self._text_buffer: List[str] = []

{rasa_pro-3.20.0.dev2 → rasa_pro-3.20.0.dev3}/rasa/core/channels/voice_stream/tts/cartesia.py

@@ -14,6 +14,9 @@ from rasa.core.channels.voice_stream.audio_bytes import (
     AudioFormat,
     RasaAudioBytes,
 )
+from rasa.core.channels.voice_stream.tts._client_session import (
+    create_ssl_client_session,
+)
 from rasa.core.channels.voice_stream.tts.tts_engine import (
     TTSEngine,
     TTSEngineConfig,
@@ -75,10 +78,9 @@ class CartesiaTTS(TTSEngine[CartesiaTTSConfig]):
     @voice_tracing.traced_tts_connect
     async def connect(self, config: Optional[CartesiaTTSConfig] = None) -> None:
         """Establish WebSocket connection to Cartesia TTS."""
-        # Create session if it doesn't exist or is closed
         if self.session is None or self.session.closed:
             timeout = ClientTimeout(total=self.config.timeout)
-            self.session = aiohttp.ClientSession(timeout=timeout)
+            self.session = create_ssl_client_session(timeout)
 
         headers = self.get_request_headers(self.config)
         ws_url = self.config.endpoint

{rasa_pro-3.20.0.dev2 → rasa_pro-3.20.0.dev3}/rasa/core/channels/voice_stream/tts/deepgram.py

@@ -14,6 +14,9 @@ from rasa.core.channels.voice_stream.audio_bytes import (
     AudioFormat,
     RasaAudioBytes,
 )
+from rasa.core.channels.voice_stream.tts._client_session import (
+    create_ssl_client_session,
+)
 from rasa.core.channels.voice_stream.tts.tts_engine import (
     TTSEngine,
     TTSEngineConfig,
@@ -66,10 +69,9 @@ class DeepgramTTS(TTSEngine[DeepgramTTSConfig]):
 
     @voice_tracing.traced_tts_connect
     async def connect(self, config: Optional[DeepgramTTSConfig] = None) -> None:
-        # Create session if it doesn't exist or is closed
         if self.session is None or self.session.closed:
             timeout = ClientTimeout(total=self.config.timeout)
-            self.session = aiohttp.ClientSession(timeout=timeout)
+            self.session = create_ssl_client_session(timeout)
 
         headers = self.get_request_headers(self.config)
         ws_url = self.get_websocket_url(self.config)

{rasa_pro-3.20.0.dev2 → rasa_pro-3.20.0.dev3}/rasa/core/channels/voice_stream/tts/rime.py

@@ -16,6 +16,9 @@ from rasa.core.channels.voice_stream.audio_bytes import (
     AudioFormat,
     RasaAudioBytes,
 )
+from rasa.core.channels.voice_stream.tts._client_session import (
+    create_ssl_client_session,
+)
 from rasa.core.channels.voice_stream.tts.tts_engine import (
     TTSEngine,
     TTSEngineConfig,
@@ -146,10 +149,9 @@ class RimeTTS(TTSEngine[RimeTTSConfig]):
     @voice_tracing.traced_tts_connect
     async def connect(self, config: Optional[RimeTTSConfig] = None) -> None:
         """Establish WebSocket connection to Rime TTS."""
-        # Create session if it doesn't exist or is closed
         if self.session is None or self.session.closed:
             timeout = ClientTimeout(total=self.config.timeout)
-            self.session = aiohttp.ClientSession(timeout=timeout)
+            self.session = create_ssl_client_session(timeout)
 
         headers = self.get_request_headers()
         ws_url = self.get_websocket_url()

{rasa_pro-3.20.0.dev2 → rasa_pro-3.20.0.dev3}/rasa/core/channels/voice_stream/voice_channel.py

@@ -952,6 +952,7 @@ class VoiceOutputChannel(OutputChannel):
         await super().send_response_chunk(recipient_id, chunk, **kwargs)
 
         chunk = remove_emojis(chunk)
+        voice_tracing.get_current_synthesize_speech_recorder().append_text(chunk)
         await self.tts_engine.send_text_chunk(chunk)
 
     async def send_response_chunk_end(self, recipient_id: Text, **kwargs: Any) -> None:

{rasa_pro-3.20.0.dev2 → rasa_pro-3.20.0.dev3}/rasa/core/voice/event_composition/handlers/barge_in/barge_in_handler.py

@@ -111,8 +111,8 @@ class BargeInHandler:
 
         # Capture playback flags before accept clears them so the decision log records
         # the gate state that allowed classification.
-        was_bot_speaking = call_state.is_bot_speaking
-        was_pending_bot_audio = call_state.pending_bot_audio
+        audio_was_playing = call_state.is_bot_speaking
+        audio_was_buffered = call_state.pending_bot_audio
 
         with voice_tracing.handle_barge_in_span(transcript=event.text) as barge_in_span:
             barge_in_verdict = await self._classify_barge_in(event)
@@ -122,6 +122,8 @@ class BargeInHandler:
                 barge_in_span.record_accepted(
                     reason=barge_in_verdict.reason,
                     transcript=event.text,
+                    audio_was_playing=audio_was_playing,
+                    audio_was_buffered=audio_was_buffered,
                 )
                 # No transcript here: user text must stay out of INFO-level
                 # logs; the DEBUG log below and content-gated span carry it.
@@ -131,8 +133,8 @@ class BargeInHandler:
                     event_type=type(event).__name__,
                     reason=barge_in_verdict.reason,
                     verdict_metadata=barge_in_verdict.metadata,
-                    was_bot_speaking=was_bot_speaking,
-                    was_pending_bot_audio=was_pending_bot_audio,
+                    audio_was_playing=audio_was_playing,
+                    audio_was_buffered=audio_was_buffered,
                 )
                 structlogger.debug(
                     "barge_in_handler.accepted_transcript",
@@ -143,6 +145,8 @@ class BargeInHandler:
                 barge_in_span.record_ignored(
                     reason=barge_in_verdict.reason,
                     transcript=event.text,
+                    audio_was_playing=audio_was_playing,
+                    audio_was_buffered=audio_was_buffered,
                 )
                 structlogger.debug(
                     "barge_in_handler.ignored",
@@ -151,8 +155,8 @@ class BargeInHandler:
                     event_type=type(event).__name__,
                     reason=barge_in_verdict.reason,
                     verdict_metadata=barge_in_verdict.metadata,
-                    was_bot_speaking=was_bot_speaking,
-                    was_pending_bot_audio=was_pending_bot_audio,
+                    audio_was_playing=audio_was_playing,
+                    audio_was_buffered=audio_was_buffered,
                 )
 
     async def _classify_barge_in(

{rasa_pro-3.20.0.dev2 → rasa_pro-3.20.0.dev3}/rasa/mantle/model_archive/bootstrap.py

@@ -16,7 +16,7 @@ from rasa.mantle.config.agent_spec import AgentSpec, agent_spec_from_config
 from rasa.mantle.config.session_spec import SessionSpec, session_spec_from_config
 from rasa.mantle.memory.schema import ProjectMemorySchema
 from rasa.mantle.skills.catalog import SkillCatalog
-from rasa.mantle.skills.responses import ResponseRegistry
+from rasa.mantle.skills.responses import RESPONSES_FILENAME, ResponseRegistry
 from rasa.shared.engine_version import MANTLE_ENGINE_VERSION
 from rasa.shared.project_structure import (
     V2_AGENT_FILE,
@@ -27,7 +27,7 @@ from rasa.shared.utils.yaml import read_yaml_file
 
 logger = structlog.get_logger()
 
-V2_RESPONSES_ARTIFACT_PATH = f"{V2_SNAPSHOT_DIR}/responses.yml"
+V2_RESPONSES_ARTIFACT_PATH = f"{V2_SNAPSHOT_DIR}/{RESPONSES_FILENAME}"
 V2_PROJECT_MEMORY_ARTIFACT_PATH = f"{V2_SNAPSHOT_DIR}/memory.yml"
 V2_AGENT_ARTIFACT_PATH = f"{V2_SNAPSHOT_DIR}/{V2_AGENT_FILE}"
 
@@ -236,6 +236,26 @@ def load_default_system_responses() -> ResponseRegistry:
     return ResponseRegistry.from_yaml(path)
 
 
+def load_bundled_response_registry() -> ResponseRegistry:
+    """Load packaged system and default-skill responses (no project files).
+
+    Used as the base layer before merging project ``skills/*/responses.yml``
+    and top-level ``responses.yml`` at model load or during validation.
+    """
+    registry = load_default_system_responses()
+
+    default_skills_dir = Path(__file__).parent.parent / "default_skills"
+    if default_skills_dir.is_dir():
+        for skill_dir in sorted(default_skills_dir.iterdir()):
+            if not skill_dir.is_dir() or skill_dir.name.startswith("_"):
+                continue
+            skill_responses = skill_dir / RESPONSES_FILENAME
+            if skill_responses.is_file():
+                registry = ResponseRegistry.from_yaml(skill_responses).merge(registry)
+
+    return registry
+
+
 def load_response_registry(
     snapshot_skills_root: Path,
     top_level_path: Path,
@@ -259,28 +279,14 @@ def load_response_registry(
         top_level_path: The top-level ``responses.yml`` file, which may not
             exist.
     """
-    # 0. Packaged system defaults (lowest priority — overridable by every source
-    # below, including user-defined responses).
-    registry = load_default_system_responses()
-
-    # 1. Default skills (bundled package source). This module lives in the
-    # ``model_archive`` subpackage, so the bundled skills sit one level up under
-    # ``rasa/mantle/default_skills``.
-    default_skills_dir = Path(__file__).parent.parent / "default_skills"
-    if default_skills_dir.is_dir():
-        for skill_dir in sorted(default_skills_dir.iterdir()):
-            if not skill_dir.is_dir() or skill_dir.name.startswith("_"):
-                continue
-            skill_responses = skill_dir / "responses.yml"
-            if skill_responses.is_file():
-                registry = ResponseRegistry.from_yaml(skill_responses).merge(registry)
+    registry = load_bundled_response_registry()
 
     # 2. Per-project-skill responses from snapshot
     if snapshot_skills_root.is_dir():
         for skill_dir in sorted(snapshot_skills_root.iterdir()):
             if not skill_dir.is_dir() or skill_dir.name.startswith("_"):
                 continue
-            skill_responses = skill_dir / "responses.yml"
+            skill_responses = skill_dir / RESPONSES_FILENAME
             if skill_responses.is_file():
                 registry = ResponseRegistry.from_yaml(skill_responses).merge(registry)
 

{rasa_pro-3.20.0.dev2 → rasa_pro-3.20.0.dev3}/rasa/mantle/model_archive/snapshot.py

@@ -19,6 +19,7 @@ from rasa.mantle.model_archive.references_index import build_references_index
 from rasa.mantle.model_archive.training_telemetry import (
     build_mantle_training_started_attributes,
 )
+from rasa.mantle.skills.responses import RESPONSES_FILENAME
 from rasa.mantle.tools.loader import TOOLS_DIR_NAME
 from rasa.shared.constants import CONFIG_ADDITIONAL_LANGUAGES_KEY
 from rasa.shared.core.domain import Domain
@@ -37,7 +38,7 @@ structlogger = structlog.get_logger(__name__)
 V2_SNAPSHOT_SOURCE_FILES = (
     "integrations.yml",
     "memory.yml",
-    "responses.yml",
+    RESPONSES_FILENAME,
 )
 
 # Project source directories copied verbatim into the snapshot. ``skills/``

{rasa_pro-3.20.0.dev2 → rasa_pro-3.20.0.dev3}/rasa/mantle/model_archive/training_telemetry.py

@@ -38,7 +38,7 @@ from rasa.mantle.skills.catalog import (
     load_skill_catalog_result,
     ordered_block_ids_for_skill,
 )
-from rasa.mantle.skills.responses import ResponseRegistry
+from rasa.mantle.skills.responses import RESPONSES_FILENAME, ResponseRegistry
 from rasa.mantle.tools.loader import TOOLS_DIR_NAME
 from rasa.shared.constants import CONFIG_ADDITIONAL_LANGUAGES_KEY
 from rasa.shared.project_structure import (
@@ -50,7 +50,6 @@ from rasa.shared.utils.yaml import read_yaml_file
 
 structlogger = structlog.get_logger()
 
-_RESPONSES_FILE = "responses.yml"
 _SKILLS_DIR = "skills"
 _LIB_DIR = "lib"
 _EXTERNAL_VECTOR_STORE_TYPES = frozenset({"qdrant", "milvus"})
@@ -429,9 +428,9 @@ def _filesystem_attributes(project_root: Path) -> dict[str, Any]:
     )
     return {
         "num_shared_tools": _count_tool_modules(root / TOOLS_DIR_NAME),
-        "num_project_responses": _count_responses(root / _RESPONSES_FILE),
+        "num_project_responses": _count_responses(root / RESPONSES_FILENAME),
         "num_skill_responses": sum(
-            _count_responses(d / _RESPONSES_FILE) for d in skill_dirs
+            _count_responses(d / RESPONSES_FILENAME) for d in skill_dirs
         ),
         "num_project_reference_files": _count_files(root / V2_REFERENCES_DIR),
         "num_skill_reference_files": sum(

{rasa_pro-3.20.0.dev2 → rasa_pro-3.20.0.dev3}/rasa/mantle/orchestration/skill_executor.py

@@ -2187,15 +2187,22 @@ class SkillExecutor:
     ) -> tuple[str, Optional[Rephrase]]:
         """Resolve a response utterance to text and an optional rephrase payload.
 
-        Looks up *response_name* in the :class:`ResponseRegistry`, substitutes
+        Looks up *response_name* in the :class:`ResponseRegistry`, selects the
+        matching variant via :meth:`ResponseRegistry.select_variant`, substitutes
         memory placeholders via :func:`substitute_memory_placeholders`, and
         returns the rendered text with whether the variant requests LLM
         rephrasing.
 
-        Variant selection uses the **first YAML variant only** (the primary
-        variant). Channel- and condition-based filtering is not implemented yet.
+        When ``select_variant`` returns ``None`` — missing name, empty
+        variants, or fail-closed CRV selection (every variant conditioned and
+        none match / every condition raises
+        :class:`~rasa.mantle.skills.conditions.ConditionEvaluationError`) —
+        delivery uses the same pre-existing missing-name fallback: utter
+        *response_name* as the template text with no rephrase.
         """
-        definition = self._response_registry.get_definition(response_name)
+        definition = self._response_registry.select_variant(
+            response_name, self.current_memory_values(tracker_handle)
+        )
         if definition is None:
             structlogger.debug(
                 "mantle.skill_executor.response.missing_definition",

{rasa_pro-3.20.0.dev2 → rasa_pro-3.20.0.dev3}/rasa/mantle/orchestration/turn_context.py

@@ -191,6 +191,9 @@ class MessageTurnContext:
         """
         if self.is_cancelled:
             structlogger.debug("turn_context.send_skipped_cancelled")
+            voice_tracing.get_current_handle_turn_recorder().record_response_dropped(
+                text=text, is_filler=is_filler
+            )
             return
         metadata: dict[str, object] = {
             RESPONSE_SOURCE_METADATA_KEY: self._resolve_response_source(
@@ -247,6 +250,9 @@ class MessageTurnContext:
         """
         if self.is_cancelled:
             structlogger.debug("turn_context.record_streamed_message_skipped_cancelled")
+            voice_tracing.get_current_handle_turn_recorder().record_response_dropped(
+                text=text, is_filler=is_filler
+            )
             # Preserve TTS timing for audio played before barge-in.
             self._record_tts_finished()
             return