Changelog

​

Improvements

• #7471: rasa init now asks which coding agents to set up (Claude Code, Cursor, VS Code, or OpenAI Codex), with Claude Code and Cursor pre-selected, and installs the Mantle skills, registers the Rasa MCP server, and downloads the Mantle documentation only for the ones you pick. Pass --ides (for example --ides claude,vscode) to choose without the prompt, or --ides '' to skip. With --no-prompt and no --ides, Claude Code and Cursor are set up as before.
• #7528: Improved default tool-call acknowledgements (fillers). They are less likely to promise or predict the result, such as “I’ll place the order now” or “Sure, I can help with that”, and are phrased to stay true if a tool fails, finds nothing, or needs the user’s confirmation. The reply after the tools is more likely to continue from them instead of repeating or contradicting them. Agents that set their own fillers.instructions keep their wording.

​

Documentation

• #7557: The Anthropic integrations.yml example recommends claude-sonnet-4-6.

​

Features

• #7539: Project memory fields can opt into mid-session updates with mutable: true and restrict which skills may write them with writable_by.
• #7483: A modify_tool_call hook can attach metadata sent only as MCP _meta on that call (not in tool arguments or model history). When tool-call hooks run, preflight rejects duplicate tool-call ids in one batch. When an MCP tool requires confirmation, that metadata is stored on system.pending_tool_confirmation and replayed on resume. The write stays in tracker history after the slot is cleared, so do not put secrets in metadata.
• #7552: Ordered blocks can record a decided human handoff with execute_tool: record_handoff. Studio analytics use that call for containment and escalation, and the model is never offered the tool.

​

Improvements

• #7565: Scenarios can now set an optional type to label the kind of case they cover, such as happy_path, edge_case, or escalation.
• #7581: A Copilot build from an approved plan now survives a page reload and a service restart. Studio can follow the build again, and a build that a restart stopped ends as built or failed instead of staying in progress. Approving a plan twice starts only one build.
• #7590: The builder service’s GET /api/eval/scenarios now returns each scenario’s parsed fields (type, simulation_context, criteria, initial_slots, assertions, and expected_tools) alongside its raw contents.
• #7606: Copilot keeps each plan in the conversation that made it, so other users never see or change it. A Copilot chat can no longer change an agent while a build runs, and approving a plan never replaces an agent that already exists.

​

Bugfixes

• #7550: When a user repeats a request for a skill that is waiting on an open question and the agent answers in its own words, it now follows that reply with the authored question and its buttons instead of leaving the question unasked.

​

Features

• #7536: Mantle CLI commands accept --integrations to load runtime configuration from a file outside the project root.

​

Bugfixes

• #7461: A rephrased reply from a Mantle assistant called over A2A is no longer stored only as token artifacts. Non-streaming callers receive one agent message per utterance and no token artifacts. Streaming callers still receive the tokens, and each finished utterance is also a history message whose metadata streamed_artifact_id names that artifact. An interrupted stream is not stored as a finished message.
• #7537: The message that pauses a task for a side question can no longer also answer whether to resume it. The agent asks whether to continue instead.

​

Breaking Changes

• #7521: Use name: deepgram_flux for Flux ASR configurations. name: deepgram now selects Nova, so update existing Flux configurations before upgrading.

​

Improvements

• #7450: The voice backchannel detection model is now downloaded from the Rasa CDN instead of Hugging Face.
• #7494: Voice agents now mistake fewer real interruptions for backchannels like “uh-huh”, so callers who cut in are heard more reliably. The default backchannel_threshold is now 0.7. If you set it explicitly, re-check your value.

​

Bugfixes

• #7553: Starting a conversation with /session_start (for example from Inspector) no longer routes the trigger as user input, so the bundled default_session_start skill runs without an agent message before the customer speaks. A project override of that skill that includes an utterance step can still speak on bootstrap. Inspector no longer shows the typing indicator after a silent bootstrap.
• #7453: A short “thank you” while the agent is still waiting for an answer no longer ends the active skill or gets a silent reply; the agent asks its question again.
• #7521: Deepgram Flux no longer loses an utterance’s final transcript when it differs from the last partial transcript.
• #7526: Mantle agents without privacy configuration publish conversation events to Conversation Review again.
• #7546: When a user asks for a skill that has to wait for a question the running skill is still asking, the agent repeats that question with its buttons instead of rewording it, declining, or filling in an answer the user never gave.
• #7549: When the agent declines an off-topic request while a skill is waiting on a question, it repeats that question with its buttons right after the decline instead of rewording it or filling in an answer the user never gave.

​

Features

• #7510: Voice channels can use ElevenLabs for speech recognition with Scribe v2 Realtime and speech synthesis with Eleven Multilingual v2, including word-level timestamps. Speech synthesis requires a voice ID configured for each language.
• #7475: Voice agents can play ambient background audio, such as office noise, during calls. Add a background_audio block to a voice channel in integrations.yml to set the volume or use your own audio file.

​

Improvements

• #7492: Voice channels can set the pause between consecutive agent messages with pause_between_messages in integrations.yml. The default stays at 1 second, and 0 adds no pause.
• #7490: Voice channels now wait 0.8 seconds after a filler before the next message. Set pause_after_filler on the channel in integrations.yml to change it, or set it to 0 for no pause.

​

Bugfixes

• #7509: Twilio Media Streams can now use a custom sender ID to select an existing conversation while keeping call audio and playback controls working.
• #7500: The recordings configuration no longer accepts unknown or misspelled keys. provider is required when recordings are enabled, and s3 is no longer the default provider.

​

Breaking Changes

• #7425: An endpoints.yml file at the project root fails startup, and passing --endpoints fails too. Mantle only accepts integrations.yml. Delete endpoints.yml. The silence timeout belongs in agent.yml as voice.silence_timeout. Call recordings are read only from integrations.yml. Move a recordings: block to recordings: before upgrading. Omit it to leave recording off. Studio trains and runs a Mantle assistant without --endpoints, so those launches succeed. An endpoints.yml in another working directory is not this project’s file and does not fail startup.

​

Features

• #7428: Copilot now searches the Rasa docs and answers Mantle questions from the Mantle docs without extra configuration, because the rasa-pro package and the rasa/rasa-pro Docker image include both documentation indexes.

​

Improvements

• #7446: Copilot for Mantle agents now refers to Rasa Studio instead of Hello Rasa, answers how-to questions with the agent’s own skills, memory, and responses, and offers to make the change.
• #7392: rasa tools init now sets up Mantle projects: it detects them, downloads the Mantle documentation, and offers to install the Mantle agent skills. Use --engine mantle to choose Mantle explicitly. For Mantle projects, the Rasa MCP server no longer describes CALM-only tools it does not offer, and train_rasa_assistant also works when your IDE starts the server outside the project folder.
• #7320: Voice now uses word-level timestamps as intermediate markers. On barge-in LLM is now aware of the playback state based on the last acknowledged marker. This is available for selected TTS channels: Cartesia, Rime and Azure.

​

Bugfixes

• #7447: After a digression that finishes or answers a knowledge question in the same turn, the agent now asks whether to resume the paused task instead of sometimes dropping it without asking.
• #7443: The model answers from relevant tool results before it finishes a skill, so a lookup is not replaced by the wrap-up question.
• #7423: When a turn hits the orchestration iteration limit, the agent now speaks utter_model_call_error instead of staying silent. Set metadata.rephrase: true on your override to have it reworded.
• #7417: rasa train validation logs, the error for a removed collect tool: key, and the text returned when a write is rejected now refer to run_after_setting_<entry> tools as memory validation tools.

​

Documentation

• #7445: Docs for ActivateSkill now say a blocks-only skill starts at its entry block, and that a confirmation in the same response holds every handoff until the user answers.

​

Breaking Changes

• #7409: Omitting skill_retrieval leaves retrieval off, so the full routing catalogue stays and turns do not call the embedder. Set active: true to opt in. Retrieval then runs only when there are more startable skills than num_skills (at least 1). A smaller catalogue is not indexed or embedded, and train logs a warning with both counts.
• #7398: Answering from documents, the pre-tool line, and channel style move in agent.yml. Use knowledge_answering (including instructions and top_k), fillers, and orchestrator.text_rules / orchestrator.voice_rules. A skill sets knowledge_answering.instructions in its frontmatter. prompts stays for routing_no_active_skill. Filler and channel-style keys belong on fillers and orchestrator. Omitting the new keys keeps the previous behavior, including four chunks and voice-on, text-off fillers. Retrain after you move the keys.
• #7369: Fully guided skills start with activate and the skill id. Routing and topic change list that skill id and the skill description. An ordered block no longer has a description; training fails if one is still set. A fully guided skill has one entry block, and other blocks are reached with call: or link:.
• #7369: Silence hangup is its own skill, default_silence_timeout_hangup. default_silence_timeout only checks in. An override of default_silence_timeout that still contains the hangup block fails training.
• #7380: rasa train rejects a description on a memory field the model cannot set, including project fields and bundled default skills. Keep a description only on a field marked llm_settable or filled by a collect step. Remove it from every other field.
• #7301: Lifecycle tools are now complete and cancel instead of complete_skill and cancel_skill. Update skill prose that tells the model to call the old names, and rename orchestrator.tool_constraints keys from complete_skill and cancel_skill to complete and cancel. A custom @tool cannot be named complete or cancel.
• #7372: rasa train now rejects unknown top-level keys in integrations.yml. CALM-only keys (nlg, nlu, action_endpoint, vector_store, interaction_handling) were silently ignored before; each error names what to use instead. A recordings: section logs a warning, since it still belongs in endpoints.yml. See Keys this file does not accept.
• #7337: Training rejects a memory field the model can set when it has no description. Put the description on the memory field or on the collect step that fills it. When both are set, training warns and the collect-step description is the one shown to the model.

​

Features

• #7370: You can mark memory fields pii: true and set an anonymization method on that field. agent.yml privacy holds tracker-store cron settings only. Published events are anonymized when integrations.yml sets stream_pii: false. Optional GLiNER-based detection (when the pii install extra and model are available) masks additional entities in user and bot message text.
• #7363: You can set orchestrator.tool_constraints in agent.yml to turn off cancel, complete, search_knowledge, or cannot_help, and to replace a built-in tool’s when-to-call guidance with invoke_when. Omitting the section keeps today’s prompt and offer rules. A retrain is required to pick up tool_constraints from agent.yml.
• #7365: A top-level a2a_server block in integrations.yml configures the inbound A2A server. It accepts include_default_skills, a2a_message_cache_ttl_seconds, and the push notification keys; ${ENV_VAR} references are expanded, unknown keys (including enabled) and a literal JWT secret are rejected, and channels.a2a is not an input channel. With the block present, session_config in agent.yml must set start_session_after_expiry: false. When push_notifications_enabled is true, task updates are posted to callback URLs and the AgentCard advertises them.
• #7403: A Mantle assistant called over A2A reports a finished skill as completed together with that skill’s public memory, stays on input required while a skill or a tool confirmation is still open, and rejects a decline that leaves nothing else to do. Button titles travel on the working update. Cancelling the task marks the skill cancelled.
• #7376: The inbound A2A AgentCard is generated from the skill catalog and named with agent.id from agent.yml. Routable skills are advertised with their skill.md name and description; engine_managed, disabled, and hidden skills stay off the card, and a default skill appears only when include_default_skills is true and a project override made it routable. A static agent_card_path must be A2A 1.0 JSON with supportedInterfaces. The card is generated when the model loads at startup.
• #7381: PUT /model and DELETE /model on the Rasa HTTP API point the inbound A2A server at the swapped model and rebuild the AgentCard. The response waits until a turn already in progress finishes on the model it started with, and that model is retired only after the turn finishes. DELETE /model publishes a card with no generated skills. A model that fails the A2A engine or session check is rejected and the current model stays loaded. PUT /model of a model trained for the other engine fails the same way, so a running CALM server is not migrated to Mantle through these endpoints; restart rasa run with the retrained model.
• #7385: You can require confirmation before cancel runs by setting orchestrator.tool_constraints.cancel.requires_confirmation in agent.yml. If the user confirms, the skill is cancelled as it is today; if they decline, the skill stays in place. Omitting the setting keeps immediate cancel. See tool_constraints.

​

Improvements

• #7393: Coding agents that simulate a Mantle conversation write scenarios with the current checks and wait for you to choose a fix when an evaluation fails.
• #7308: The system prompt no longer includes the skills glossary.
• #7388: Coding agents that author skills now apply the decision rules for when a sentence in the skill body is not enough.
• #7344: On a voice call, during a collect step, the agent waits silently when the user has started answering but not finished (for example, halfway through a card number) instead of talking over them.

​

Bugfixes

• #7407: When loading a trained model fails, the error log names the underlying problem and refers to the model as Mantle. Training logs that the Mantle model package was saved.
• #7405: rasa run and rasa inspect load channels, endpoints, and the secrets manager from integrations.yml in the project directory when --model is a Mantle archive, even with no skills/ folder beside it. Tracker, lock, and timer values that remain only in endpoints.yml fall back to memory, and an event broker or model server left only there is ignored, so move those blocks into integrations.yml before serving this way. If the archive cannot be read as Mantle, those commands keep credentials.yml and endpoints.yml and log a warning.
• #7386: rasa init --engine mantle registers the rasa-tools MCP server for Claude Code and Cursor, so simulation and evaluation no longer require a separate setup step.
• #7366: On voice, the silence timeout now also starts after a turn where the agent waits with listen, so a caller who stays quiet gets the silence check-in instead of dead air.
• #7360: A button that stores a memory value now counts as the user’s answer, so the agent continues the conversation instead of treating the question as unanswered.

​

Documentation

• #7370: Documented privacy in agent.yml and the Mantle PII pipeline in memory.yml. The mantle_privacy_e2e agent in maestro-benchmarking-agents includes local Kafka and GLiNER verification steps.
• #7363: The tools reference defines each built-in tool, including what it does and when the model is offered it.

​

Breaking Changes

• #7361: The tracker store, lock store, and secrets manager are read only from integrations.yml. tracker_store:, lock_store:, and secrets_manager: blocks in endpoints.yml are ignored, and startup logs a warning. Move them to tracker_store:, lock_store:, and secrets_manager: before upgrading. A tracker_store: block left only in endpoints.yml stores conversations in memory, so they are lost when the process restarts.
• #7296: The event broker, timer store, and remote model server are read only from integrations.yml. event_broker:, timer_store:, and models: blocks in endpoints.yml are now ignored, and startup logs a warning. Move them to event_broker:, timer_store:, and rasa_model_server: in integrations.yml.
• #7289: The orchestrator model group is named with orchestrator.model_group in agent.yml. Optional orchestrator.max_prompt_tokens defaults to 8000. integrations.yml must not contain an llm: section; define the group under model_groups. Existing projects need that name added and a retrain.
• #7217: Unknown top-level keys in skill.md frontmatter are rejected at train. Remove the extra key.

​

Features

• #7286: Tools can stream a message to the user and attach buttons to a message they send.
• #7361: You can configure the tracker store, lock store, and Vault secrets manager in integrations.yml. VAULT_URL, VAULT_TOKEN, and VAULT_RASA_SECRETS_PATH still override secrets_manager: and still work when that section is omitted. See Tracker store, Lock store, and Secrets manager.
• #7296: You can configure the event broker, timer store, and remote model server under integrations.yml instead of endpoints.yml. Use rasa_model_server: for the remote model server, not the models list inside a model group. With no local model, rasa run points at integrations.yml. See Event broker, Timer store, and Remote model server.
• #7224: A tool can now hand off to another skill by returning ToolResult(next=[ActivateSkill(skill_id=...)]). The skill activates after the tool’s result is recorded, following the same rules as the LLM’s activate tool. See ActivateSkill.
• #7217: JSON Schema for Mantle authoring files is published under Reference → Schemas, generated from the models that load those files, including skill frontmatter and preconditions.

​

Improvements

• #7375: Voice calls that use Deepgram speech recognition now include the Deepgram request ID in the metadata of each user message (metadata.asr.request_id), so you can match a conversation to Deepgram’s own logs when debugging transcription issues with Deepgram support. If you subclass VoiceInputChannel and override handle_asr_event, add the new optional asr_metadata keyword argument to your override.
• #7335: Voice agents now keep work already underway when a user speaks over a response, then respond to their latest input once that work finishes.
• #7354: A missing model group named by references.embeddings or orchestrator.model_group now reports that id. The references.embeddings error says to clear the field to use the built-in OpenAI embedder. A missing orchestrator group stops the server instead of starting with no model loaded.
• #7336: Turns no longer wait on an extra model call after a skill switch to store facts the conversation does not use.
• #7145: on_model_request / modify_model_request and on_model_response / modify_model_response also run on rephrase, confirm-rephrase, and the post-activate set_fields fill. A request-hook crash does not send that prompt. RetryModel on those calls discards the completion instead of starting another main-loop iteration.
• #7227: rasa export publishes Mantle ProtoJSON when --integrations points at an existing integrations file (default integrations.yml). Classic projects keep flat events.
• #7257: Training or validating a Mantle project that has no skills reports that no skills/<id>/skill.md files were found.
• #7284: When training fails on a Mantle project, Copilot’s analysis now searches the Mantle documentation, as Copilot chat already does.
• #7292: Mantle trackers now record skill, block, and memory changes only as skill_*, ordered_block_*, memory_set, and memory_cleared events; the classic flow_* and slot_set copies are no longer written. Conversations recorded with those copies still load.

​

Bugfixes

• #7361: Sanic workers now load the secrets manager. A Vault-backed tracker store works with more than one worker for Mantle and for classic projects.
• #7250: Resuming an older paused skill with activate pauses the task in progress and keeps newer paused tasks on the stack. Returning to that skill or to its block also keeps an earlier block of that skill that was parked by a block switch, and finishing the resumed block continues that earlier block. When the resumed task finishes, the user is offered to continue the nearest paused task.
• #7145: When activating a skill lands on an ordered block that would immediately wait for the user, a collect or settable memory value already stated in that same message is recorded before the skill asks for it again.

​

Documentation

• #7355: Mantle docs now explain how to simulate and evaluate an agent, including the assertion keys a Mantle scenario can use.
• #7275: A best-practices page covers what to put in memory, responses, constraints, and ordered blocks when a sentence in the skill body is not enough.
• #7272: Skill-writing guidance now says when to name another skill with @skill.<id> so a related request stays part of the current task, and when to leave a topic change alone.
• #7364: A new Voice section covers how Mantle runs skills in a voice conversation, a build-a-voice-agent tutorial, designing for voice, connecting a phone line, and testing and monitoring. Getting Started now lets you talk to your agent in the Inspector. The reference gains Speech providers and Custom speech providers pages, documents language_map, interruptions, and silence_timeout for voice channels, and documents keypad input with dtmf: on a collect step and the user_input_suppressed event.

​

Deprecations and Removals

• #7284: rasa tools run no longer offers the search_rasa_documentation tool, which searched the CALM docs. Your coding agent reads the Mantle docs in .rasa/docs/mantle/ through the mantle-docs skill; add them to an existing project with rasa tools init docs mantle.

​

Breaking Changes

• #7242: Tracing and metrics are read only from integrations.yml. tracing: (including Langfuse) and metrics: blocks in endpoints.yml are now ignored, and startup logs a warning. Move them to tracing: and metrics: in integrations.yml.

​

Features

• #7242: You can configure OpenTelemetry tracing (otlp or jaeger) and OTLP metrics under integrations.yml instead of endpoints.yml. List Langfuse and an OTLP or Jaeger tracer together under tracing: to use both. A Jaeger entry with username/password must set insecure (false exports over TLS). See Tracing and Metrics.

​

Breaking Changes

• #7156: Models trained before skill retrieval that still have independently startable skills must be retrained before they will load (or set skill_retrieval.active: false). Omitting skill_retrieval leaves retrieval on.
• #7189: rasa train requires a non-empty skill description, rejects unknown skill.md frontmatter keys and non-string name/description, and fails on incomplete @skill / @block / @tool tokens. Bundled default skills are checked the same way. See skill.md.

​

Features

• #7156: Mantle agents can now retrieve a subset of independently startable skills for routing. skill_retrieval is on by default; set active: false to keep the full routing catalogue.
• #7157: Skill retrieval can pin skills with always_include_in_prompt and keep recently started skills in the routing catalogue. A search outage fails the turn instead of listing every skill. This is not knowledge search (search_knowledge).
• #7282: A collect step can accept keypad input via dtmf, using length or finish_on_key. Set allow_audio_input: false to ignore speech while digits are expected.

​

Improvements

• #7223: Inspector NextGen now shows Mantle conversation events in nested ProtoJSON form while still accepting classic flat event JSON. Mantle inspect tracker dumps and downloaded logs preserve the ProtoJSON envelopes for replay.

​

Bugfixes

• #7268: The first message of a new conversation no longer activates a skill before that message is visible.
• #7183: Skills can import an MCP tool when an unused shared tool has the same name without causing model loading to fail, including when that MCP tool is gated with tool_constraints or ordered-block tools.

​

Documentation

• #7260: Builders can look up Mantle’s default skills and the wording of the built-in messages.
• #7262: Docs call run_after_setting_<entry> tools memory validation tools.

​

Deprecations and Removals

• #7260: The unused default response utter_ask_how_to_help is removed. Completed still asks with utter_ask_wants_to_continue.

​

Improvements

• ENG-3360: When modify_model_request or modify_tool_call crashes or times out, or the main model call fails, the user hears utter_model_request_hook_error, utter_tool_call_hook_error, or utter_model_call_error instead of silence. Override those names in responses.yml. The exception text is not spoken, and rephrase on those responses is ignored.

​

Bugfixes

• #7225: The hangup tool is now offered only when an ordered block explicitly uses action: hangup, so it is not selected in other contexts.

​

Breaking Changes

• #7185: Mantle model credentials must use api_key: ${ENV_VAR_NAME} (and the same ${…} form for other sensitive model keys). Literal secrets are rejected. See Secrets.
• #7104: Skill-level requires is rejected at train. Use precondition plus agent.yml orchestrator.preconditions to run a resolver first, or hidden: true to omit a skill from routing. Tool-level requires under tool_constraints is unchanged. See precondition, preconditions, and hidden.
• #7056: Mantle tracker persistence and the event broker now use nested ProtoJSON for conversation events instead of flat "event" JSON. Classic (non-Mantle) projects are unchanged. See Conversation events.
• #7170: The default voice silence timeout is now 30 seconds (was 7). Set silence_timeout on the voice channel in integrations.yml if you need a different wait.

​

Features

• #7104: Skills can declare a semantic precondition, and projects bind it in agent.yml orchestrator.preconditions so the engine parks the consumer, runs a resolver skill, then continues or abandons the request without hiding the skill from routing. If the resolver is already the active skill, the consumer is parked under that run instead of interrupting it. Skills can also set hidden: true to stay off the routing catalogue while remaining startable via call, link or resolve_with; prose @skill mentions of a hidden skill fail validation. Inspector labels a parked consumer as pending while the resolver is the active skill.
• #7035: on_incoming_message / modify_incoming_message now run on final user text before it is stored. Modified text is what gets stored and what the turn runs on. Rejected text and fail-closed hook errors stay on the tracker as discarded, are omitted from model and completion-judge history, skip the user-message orchestrator turn, and are answered with response templates: HookRejection.response_message may select a bot-defined response, while failures use utter_incoming_message_rejected. When that first payload also opens a session, default_session_start still runs so session-start memory writes apply, and named reject replies are resolved after that turn. See Hooks.
• #7107: on_outgoing_text / modify_outgoing_text now run once on each completed agent utterance, including fillers. Modified text is stored and sent; rejected text is dropped without being stored and replaced with a response template (HookRejection.response_message or utter_outgoing_text_rejected), so the tracker only ever records what the user actually received. modify_outgoing_text buffers streamed output until the whole utterance passes the hook; observe-only hooks do not. Incoming-message rejection replies also pass these outgoing hooks. See Hooks.
• #7187: Train and validate fail when skill.md prose references @tool.<name> for a tool that does not exist or is not visible to that skill.
• #7170: After three silence-timeout check-ins, the next consecutive silence ends the call with a goodbye. A skill can do the same with an ordered-block action: hangup after an authored goodbye response.

​

Improvements

• #7105: Mantle conversation events now include stable event_id and sequence fields so consumers can order and deduplicate reliably.
• #7130: Mantle ProtoJSON now covers the remaining conversation event types (session, voice, skills, ordered blocks, and related engine events).
• #7159: Voice channels now use a classifier and backchannel_threshold to distinguish short backchannels from interruptions, using the current agent utterance as context. Accepted interruptions enter the active turn, while ignored backchannels remain visible as user_input_suppressed events; the classifier loads at channel startup and downloads from Rasa’s CDN if it is not cached. See voice interruption settings.

​

Bugfixes

• #7162: Paged collects can now clear a button choice and return to the same collect to show refreshed dynamic buttons without an explicit listen step or an LLM turn.
• #7202: Memory saved during a conversation stays available when the assistant uses Postgres, Redis, or another tracker store, so collect steps continue and tools see the values they require.
• #7191: Agents no longer receive set_fields prompt guidance when the tool is unavailable for the current turn.

​

Documentation

• #7130: Documented Mantle conversation event ProtoJSON under Conversation events, including the dual-write skip list and golden fixtures for consumers.
• #7188: Mantle documentation now uses Rasa’s rebrand colors and typography for a consistent visual experience.

​

Breaking Changes

• #7123: The bundled default_session_start skill no longer greets. Override that skill and send /session_start when the agent should speak first. See Start a conversation.

​

Features

• #7025: on_turn_started / on_turn_finished now run around each Mantle turn. Builders can observe turn start and the final completed / cancelled / error status without blocking the turn. See Hooks.
• #7024: on_knowledge_query / modify_knowledge_query and on_knowledge_results / modify_knowledge_results now run around search_knowledge. Builders can rewrite the query or filter/reorder retrieved passages; neither point accepts HookRejection or RetryModel. Query-hook crashes are fail-closed (search skipped); results-hook crashes are fail-open (original passages kept). See Hooks.
• #6972: Dynamic button source tools are engine-only and validated by rasa train; missing, MCP, and confirmation-gated sources are rejected, while invalid runtime button references are omitted without dropping valid siblings.
• #6969: A response variant can name a tool with buttons: {source: <tool_name>}, and that tool supplies the buttons for the message with context.set_buttons([...]). Buttons built with a ref write the chosen value to memory on tap, exactly like a static set:session.<skill>.<entry>=<value> button; the message is sent without buttons when the tool returns none or fails.
• #6707: The model can now wait for the next user message without sending a reply, by calling the reserved listen tool after a question already asked this turn. Guided collect steps, post-search answer calls, and pending tool confirmation do not offer this tool. Mixed listen plus text or other tools ignores listen.
• #6760: on_model_request / modify_model_request now run on each main LLM loop iteration. Builders can rewrite provider-bound messages and allowlisted per-call model_settings (temperature, top_p, max_tokens, presence_penalty, frequency_penalty, stop, seed). Empty-completion retries reuse that hooked request without re-running hooks. Credentials, provider/model, and tools stay engine-owned; this point has no intentional veto (HookRejection is not accepted). See Hooks.
• #6771: on_model_response / modify_model_response now run after each main LLM completion and before any user-visible emit or tool dispatch. Builders can rewrite response text and tool calls, or raise RetryModel(feedback=...) to discard the response and steer the next iteration with an engine-only system note. HookRejection is not accepted at this point; unexpected crashes, timeouts, or invalid returns are fail-open. A registered response modifier buffers the main completion (no live token stream) so rewrites and retries match what the user hears. See Hooks.
• #7023: on_tool_call / modify_tool_call and on_tool_result / modify_tool_result now run on the LLM tool-call batch path. The engine prefights every proposed call before any tool runs; RetryModel discards the whole batch (including accompanying text) and steers the next iteration. A registered tool-call modifier buffers the main completion (no live token stream) so retries and fail-closed crashes cannot leak speech. Modified arguments reach dispatch; modified results are what McpToolExecuted / ToolExecuted store and the next prompt replays. HookRejection is not accepted on tool calls; tool results have no control signals. Confirmation resume, settable writes, and correction replay do not invoke these hooks. See Hooks.

​

Improvements

• #6987: Session-start channel metadata on Mantle conversations is now recorded as a memory_set event (with write provenance), so Studio and other consumers can read it without relying on a legacy slot event for that value.
• #6895: context.memory.get accepts an optional default. Unset, None, or unreadable fields return that fallback (get("card_list", []) is []) instead of None or a TypeError. Local tool failures log error_type at ERROR and the exception message at DEBUG.
• #6831: Inspector bot-turn latency details now include time spent in earlier LLM calls, making multi-step and tool-driven responses easier to diagnose.
• #6824: complete_skill is offered when the active skill has prose instructions, including while a hybrid skill is in an ordered block. Ordered-block-only skills omit the tool and finish when the block ends; cancel_skill is still available.
• #6773: Voice conversations now process overlapping transcripts normally and preserve rejected barge-ins or audio blocked during DTMF collection in the new user_input_suppressed tracker event.
• #7144: Evaluation scenarios can now assert skill and ordered-block lifecycles, memory writes and clears, and tool runs, including several tools in one user turn, and can check those facts in order. Inspector downloads include these assertions. A Mantle run rejects slot_was_set, slot_was_not_set, generative_response_is_relevant, and generative_response_is_grounded. Use the memory assertions for written state.

​

Bugfixes

• #7346: Deepgram Flux TTS now accepts the provider’s full speed range and supports models that require the default speed setting.
• #7161: Hybrid skills no longer offer complete_skill while an ordered block is active, preventing repeated refused completion calls before the block finishes.
• #7123: When a new or expired session starts on a real user message, that message is now handled in the same request instead of being answered only after the next turn.
• #7060: complete_skill no longer intermittently refuses to complete a skill whose exit criteria were actually met, which could otherwise re-engage the caller with an unnecessary step after the conversation had already wrapped up.
• #7057: Copilot documentation search now finds pages on the current docs site and omits unpublished drafts.
• #7052: Mantle voice calls now keep processing spans in a single call trace, making individual calls easier to debug.
• #6968: Voice agents now release LLM requests as soon as response generation completes while speech continues in order and remains interruptible.
• #6863: context.memory.set("project.<entry>", …) and context.memory.set("<active_skill>.<entry>", …) now write the named field instead of failing as undeclared, matching how get already resolves those names. An own-skill qualified name does not fall through to project memory. Cross-skill writes stay rejected.
• #6906: Skills with complete_when criteria no longer refuse exit when an early path detail that exempts a requirement (for example a damaged-card reason) was pushed out of the completion check by multi-message bot replies.
• #6826: After a canned ordered-block reply, leftover hybrid prose can continue on the same turn instead of waiting for the next user message.
• #6847: Custom default skills now start from their configured entry and are discarded cleanly when the customer moves to another task.
• #6846: After a customer answers a tool confirmation, repeated same-turn LLM calls no longer ask for confirmation again; unanswered confirmations are still re-asked after a digression. A later ordered execute_tool step of the same tool can still ask again in that turn.

​

Documentation

• #7080: New Writing prose page covers how to word a skill’s description and body. Name the subject before the action in a description, and separate them with a dash, so routing matches on the thing being asked about.
• #6850: Skills that use requires_confirmation must not also ask for confirmation in prose — pick one mechanism so the customer is not prompted twice.
• #6762: Clarified hook control-flow semantics, including RetryModel feedback as a non-persisted system note, sequential multi-tool modify_tool_call preflight, discarded text handling, and unavailable audio hooks.

​

Deprecations and Removals

• #6912: Skill deny_read / deny_write lists are removed. Memory visibility and writes follow public, private, and project scopes only. Leftover access: / deny_read / deny_write keys in skill memory.yml are ignored and are not enforced.

​

Breaking Changes

• #6729: Skill-level complete_when must be a mapping with condition and/or criteria (bare strings are rejected). Use it on skills that have prose instructions; ordered-block-only skills cannot declare it (they finish via block end without this gate). When authored, complete_skill is hard-gated on those requirements and is refused while an ordered block is active. See skill.md.

​

Features

• #6783: The live conversation stack now distinguishes skills from ordered blocks. Guided skills show the skill together with a main block. Inspector and Copilot display this stack shape. Pre-existing Mantle conversations that still use the previous flow-based live stack still load and replay, but are not upgraded in place — restart the conversation (including any in-flight Inspector session) before continuing under the new stack shape. Block-level flow_* events are still written as well.
• #6732: Mantle voice agents can now respond when a user remains silent for the configured timeout through the customizable default_silence_timeout skill. Silence timeouts and their configured duration are recorded in the conversation tracker and included in the agent’s conversation context.
• #6729: Optional complete_when.criteria (sentences in natural language) are checked on complete_skill. The check fails closed: timeouts, errors, or malformed judge answers refuse completion (skill stays active) instead of treating the skill as done. With criteria present, condition-only auto-complete is disabled.

​

Improvements

• #6800: Large FAQ files under references/ now train without a manual split, and search_knowledge returns the matching section. Builders can set character chunk size and overlap under references.chunking.

​

Bugfixes

• #6729: complete_skill is refused while a tool confirmation is pending for the active skill. Resolve or decline the confirmation first; cancel_skill still abandons the skill.
• #6787: When a user revises a value the agent already captured, a refused memory write is now reported as an error and the agent is told to use correct instead of silently keeping the old value.
• #6759: Invalid LLM-issued local and MCP tool calls now return a retryable result to the model instead of running the tool or sending its on_failure response. Ordered execute_tool steps with invalid arguments stay on the step without running the tool or on_failure.

​

Features

• #6765: Conversations now record ordered-block lifecycle (ordered_block_entered / completed / cancelled / interrupted / resumed). Block-level flow_* events are still written as well. Inspector and Copilot map these types. See Tracker events.
• #6699: Tapping a button whose payload is a set: memory reference now writes the declared memory field directly, without an LLM set_fields call (llm_settable is not required). Memory validation tools run on these taps the same way as on LLM writes. The payload must match a button on the latest agent message — typing a schema-valid set: string that was not offered there is rejected. If the tap targets a field the active skill cannot write (for example a memory entry from an earlier skill), or the payload is stale/invalid, the agent replies with a fixed notice that the selection is no longer available or an invalid value — naming it when the inbound message metadata includes button_title — and asks the user to choose an available option or type their message. A second tap on a write-once project field is rejected the same way and leaves the first value. Override the packaged notice by declaring utter_set_memory_button_skipped in your own responses.yml.
• #6699: Response variant metadata from responses.yml is now stored on bot messages in the conversation tracker, not only the internal rephrase flag.

​

Improvements

• #6789: Conversation-scoped structured logs now include conversation_id across engine and voice processing. Voice TTS debug events no longer also emit a duplicate recipient_id.

​

Bugfixes

• #6746: Updated the prompt to instruct the agent to answer in-skill follow-ups instead of calling cannot_help.
• #6745: Clarification questions and task switches no longer cancel a pending tool confirmation.

​

Features

• #6669: You can now add static buttons to named responses in responses.yml; the engine delivers them on collect and action asks in chat and web channels, including after streamed rephrase, with memory placeholders resolved in button titles and payloads. Taps with a free-text payload (or an omitted payload that defaults to the title) are handled like typed input — see Plain-text payloads.
• Mantle hooks authoring surface (preview): import on_* and modify_* decorators, typed payloads, and HookRejection / RetryModel from rasa.mantle.hooks. Place modules in hooks.py or hooks/**/*.py; train-time validation runs before packaging. The runtime does not invoke hooks yet — turn-path wiring lands in a follow-up release.

​

Improvements

• #6587: rasa init --engine mantle now bundles the Mantle documentation into your project at .rasa/docs/mantle/, and the installed agent skills tell your coding agent to read it rather than guess. Add it to an existing project with rasa tools init docs mantle. See Getting started.

​

Bugfixes

• #6695: Rasa CLI commands now work on Python 3.14. Previously every command, including rasa init and rasa --version, failed immediately on startup.

​

Breaking Changes

• #6610: The orchestrator LLM now uses llm.model_group and a matching model_groups entry. Move inline provider, model, and credential settings from llm: into that group:
  llm:
    model_group: orchestrator
  
  model_groups:
    - id: orchestrator
      models:
        - provider: openai
          model: gpt-5.1
          api_key_env: OPENAI_API_KEY
  
  See integrations.yml.
• #6733: rasa data validate now requires every conditional next to include an else fallback:
  next:
    - if: "session.card_replace.selected_card_id"
      then: check_card_eligible
    - else: END
  
  See Conditional branching.

​

Bugfixes

• #6733: Called skills no longer restart when no conditional next branch matches.

​

Breaking Changes

• #6670: rasa data validate now fails when a tool reaches the internal memory manager or clears memory entries. Clear fields this skill owns with context.memory.set. Fields another skill owns must be cleared from that skill. See Tools.

​

Features

• #6654: You can now declare static buttons on response variants in responses.yml (each with a title and optional payload). This PR adds parsing and train-time validation only — rasa train checks button shape and set: memory payloads against the same readable and settable memory rules the engine will use when those buttons write memory on tap; button delivery at runtime is not included yet.
• #6650: Conversations now record skill lifecycle (skill_activated / completed / cancelled / interrupted / resumed), tool runs (tool_executed, local or MCP), and memory writes (memory_set / memory_cleared). Skill-level flow_*, mcp_tool_executed, and slot_set events are still written as well. Inspector and Copilot map these types. See Tracker events.

​

Improvements

• #6670: When a collect step is skipped because the value is already known, the agent is told not to ask for that value again in the same turn. See Prompt templates.

​

Bugfixes

• #6715: Voice responses no longer play stale audio from an interrupted response after a barge-in.

​

Breaking Changes

• #6678: Starting the same skill again in a session now resets that skill’s public and private memory to each field’s initial_value (null when unset). Values still stay after complete, cancel, and interrupt until that next run. Put session-long facts in project memory. See Lifetime.
• #6637: Project memory is no longer writable by the LLM (set_fields, collect, or correct). Train rejects llm_settable: true and collect: on a project field. Collect user-facing values in skill memory, then write project.* from a tool or post-write hook. After the first successful write, further writes fail for the rest of the session. See Skill vs project memory.
• #6637: Memory fields no longer support immutable. Skill public and private memory stays writable. Project fields remain write-once after the first successful write. See memory.yml.
• #6608: Tool calls now time out after 10 seconds by default. Set the top-level tool_timeout in agent.yml for slow local @tool functions. Imported MCP tools use the same default unless their server sets its own tool_timeout in integrations.yml.

​

Features

• #6608: You can connect agents to MCP servers. Declare servers under mcp_servers in integrations.yml. Skills can import remote MCP tools with import_tools: [mcp/<server>:<tool>]. See Import MCP tools.

​

Improvements

• #6718: OpenTelemetry metrics now show the time between consecutive agent responses in a Mantle turn, including waits after fillers and before final responses.
• #6637: Set project memory now stays visible in a dedicated ### Project Memory prompt section, including while routing. That section tells the model those facts are already set for the session and will remain so. See System Prompt and Prompt templates.
• #6637: Session-start metadata can populate project memory for fields marked seed: true. A metadata value that does not match the declared field type is skipped (the field stays unset). Train rejects seed: true on skill public or private fields. See Seeding from session start.
• #6608: Langfuse tool-call traces now record the tool name and argument keys as the span input, not argument values, so account identifiers and other inputs stay out of the trace.
• #6693: When call recording is enabled, voice traces now include a span for storing the recording, with its storage location and whether it was saved, skipped, or failed.

​

Features

• #6645: You can now declare conditional response variants in responses.yml: each variant may include a memory condition, the engine delivers the first match in YAML order or a required default, and rasa train rejects empty text, invalid conditions, and malformed default rules with a located error.

​

Improvements

• #6685: Voice traces now show whether a barge-in happened while audio was playing or still queued, and they include dropped agent responses when content capture is enabled.

​

Features

• #6697: Added Deepgram Flux TTS. Voice agents can now use Deepgram Flux for streaming text-to-speech by setting the TTS provider name to deepgram_flux.
• #6343: Voice interruption events in Inspector now show how far the agent response played before the user interrupted, including confirmed playback progress and the full response duration when available. The engine also tells the model that the interrupted response may not have been heard in full.

​

Bugfixes

• #6679: Installing Rasa Pro on macOS no longer fails while building LiteLLM. Previously LiteLLM had no prebuilt macOS wheel, so it was compiled from source and the install failed unless a recent Rust toolchain (rustc 1.94.1 or newer) was already available.
• #6674: Voice agents can now connect to speech providers on systems that require custom CA certificates configured through SSL_CERT_FILE.

​

Breaking Changes

• #6603: The engine internals are now named Mantle throughout (the calm_v2 package became mantle, and the persisted engine_version value changed from calm_v2 to mantle). Custom tool modules must update their imports from rasa.calm_v2.tools.* to rasa.mantle.tools.* — e.g. from rasa.mantle.tools.decorator import tool, ToolContext and from rasa.mantle.tools.result import ToolResult. Agents whose tools.py files import the old rasa.calm_v2.* path fail to load until updated. Models trained on earlier .dev builds will not load and must be retrained.
• #6652: Python 3.10 is no longer supported. Use Python 3.11 through 3.14.

​

Improvements

• #6652: Mantle now includes the latest Rasa Pro main line after 3.19.0 (Sanic 25, Python 3.14, and optional voice call recording), and pins rasa-sdk to 3.20.0.dev1.

​

Bugfixes

• #6595: Setting rephrase: true on a response now also rewords confirmation questions, messages when the user says no, and success or failure messages.

​

Breaking Changes

• #6576: The CLI engine value is now mantle instead of maestro. Use rasa init --engine mantle and rasa tools skills install mantle. The old value is rejected rather than deprecated, so update any scripts that pass maestro. See Getting started.

​

Bugfixes

• #6600: Users can now correct a previously collected skill-scoped memory value.

​

Improvements

• #6531: Tool-call acknowledgement fillers (prompts.ack_enabled) are off by default on text channels and on by default on voice. Set ack_enabled: true or false in agent.yml to force either way for both modalities. Removed redundant inline-text guidance from cancel_skill lifecycle and tool schema; generic acknowledgement guidance still applies when ack is enabled.
• #6512: The default system prompt now places deploy-stable and episode-stable sections before the current date and time so provider prompt-prefix caches stay warmer across turns. The System Prompt and Prompt templates reference pages match the new order and wording.

​

Features

• #6360: Voice agents now expose OpenTelemetry traces for calls, provider connections, transcription, input decisions, speech synthesis, and audio playback. See Observability.
• #6504: Skill instruction prose can reference live memory with @memory.project.<field> and @memory.<skill_id>.<field> (prompt substitution and train validation). Unreadable @memory refs, incomplete @memory lookalikes, and bare session.* in free prose fail train. See skill.md — Memory.

​

Bugfixes

• #6552: Voice-ready and custom channels can now mark incoming turns as voice so the engine applies voice-specific prompt guidance.
• #6546: Voice calls now reject spoken input while a DTMF collect step disables audio input, with the rejection recorded in tracing.
• #6558: The tool-call acknowledgement section (and its examples) is no longer repeated in the system prompt on every agent-loop iteration. The agent acknowledges once per turn, so the guidance is now included only on the first LLM iteration and dropped on later iterations, trimming prompt tokens. See the System Prompt reference.

​

Breaking Changes

• #6401: Ordered-block steps now reject unknown YAML properties. The removed collect-step keys force_collection and tool: are no longer accepted — use ask_before_filling: true instead. The agent also fails to start when a skill directory cannot be parsed or when two skill directories in the same skills/ tree declare the same skill id.

​

Improvements

• #6428: Mantle projects without an endpoints.yml no longer log errors about the missing file when running CLI commands such as rasa train. Endpoint-backed tracing, metrics, and secrets configuration is skipped instead. CALM projects are unaffected.
• #6423: Tools can read another skill’s public memory (and project.*) via fully-qualified names on context.memory.get, with the same visibility as conditions / read_all (private fields and deny_read still blocked). Writes stay scoped to the active skill and project.
• #6401: Step and skill load validation now surfaces clearer, actionable error messages (including migration hints for the removed collect-step keys), and rasa data validate reports unloadable skill directories instead of relying on silent skips at runtime.
• #6405: Langfuse tracing is configured under optional tracing: in integrations.yml (type: langfuse, ${VAR} keys, optional knobs such as environment and timeout). Example agents no longer rely on endpoints.yml for tracing; Langfuse declared only in endpoints.yml still works with a deprecation warning until you migrate the block.

​

Documentation

• #6423: The tools and memory.yml references document fully-qualified context.memory.get reads for sibling public fields.
• #6401: The skill.md reference documents ordered-block step validation, the removed-key migration table, and collect-step authoring without per-step tool: overrides.