Breaking Changes
- Ordered blocks
now wait with
execute_tool: listenand end the conversation withexecute_tool: hangup, and anaction:step accepts only anutter_*response. Update skills that still sayaction: action_listenoraction: hangup. New turns record the wait and hangup as tool runs; events already stored stay unchanged. - The outgoing agent
text hook is
modify_outgoing_message/on_outgoing_messagewithOutgoingMessagePayloadand insertion pointoutgoing_message(replacingmodify_outgoing_text,on_outgoing_text, andOutgoingTextPayload). The packaged reject template isutter_outgoing_message_rejected(replacingutter_outgoing_text_rejected). - You configure built-in tools with
tool_constraints.<tool>.description, which fully replaces that tool’s offered description on turns where it is offered.invoke_whenis rejected; migrate by copying the published default description, merging in formerinvoke_whenwording, and setting the result asdescription— not by renaming the key alone. Engine workflow bullets in the system prompt stay; align them with your tool text so the model does not see conflicting guidance. Onelisten.descriptionapplies on voice and text turns.tool_constraints.enteracceptsdescriptiononly. - A skill
memory.ymlno longer ignores leftoveraccess:,deny_read, ordeny_writekeys.responses.ymlallows only a top-levelresponseskey, andintegrations.ymlallows only the keys that file defines. A response whose value is not a list of variants also fails to load, instead of being skipped. Remove the extra keys and give each response a list of variants; the agent fails to load until you do. - Session-start
metadata is stored as
system.session_started_metadata. Read that key on new sessions. Events already stored assession_started_metadataremain on the tracker. - A revision of a value the user already gave is recorded with
set_memory, one earlier value per call. When a call tries to change more than one earlier value, only the first such argument in call order is revised; each other is deferred and the tool result asks for a separateset_memorycall, unless that revision’s rewind cleared the memory item, in which case the later argument is written as an unset item in the same call. A value the call repeats unchanged is kept, including when that call revises a different earlier value. Passingnullto revise a memory item is rejected; it does not clear the stored value. Update prompts, hooks, and eval scenarios to nameset_memoryfor revisions.invoke_whentext for revisions belongs undertool_constraints.set_memory. - A conversation that was started under CALM is no longer continued by a Mantle agent. Start it on a new tracker.
Features
- Every Mantle agent
ships pass-through hook defaults at every insertion point, so you do not need
hooks.pyuntil you override a point; the hooks reference explains what runs on the live path versus register-only defaults. - Voice agents can now use Gradium for real-time speech recognition and speech synthesis with word-level timing events.
- Inworld is now available for speech recognition and synthesis in voice channels, with speech recognition buffering short audio frames and logging provider error details.
- Voice calls now
recover when a speech provider’s connection drops mid-call. Speech
recognition reconnects and keeps the words already recognized, instead of
ending the call. If speech synthesis fails, the caller hears the audio
produced so far, or silence, and the next response reconnects. This also
applies to custom speech providers.
Configure it with
reconnect, and setmax_attempts: 0to turn it off.
Improvements
- Training logs that an agent package was saved, instead of a model package.
- The bundled customer-satisfaction score question is sent exactly as written, with its Satisfied and Unsatisfied buttons.
- Successful built-in tool calls from earlier turns no longer fill conversation history, so longer chats spend fewer tokens on receipts that memory and skill state already show.
Bugfixes
- When an ordered block finishes a skill with a connecting or wrap-up line and nothing is left to run, the turn ends on that line instead of asking another follow-up on the same user message.
- Matching skills now activate before asking for skill-specific details, so their setup instructions run first.
- Activating a skill that is waiting under a different unfinished skill is refused, and the reply names the skill that is still in charge.
- A correction runs the steps up to a waiting confirmation again and asks that confirmation again.
- A builder tool the model was not offered this turn is refused instead of running.
Documentation
- You can follow a documented example to write a custom lock store, timer store, tracker store, event broker, knowledge store, or channel, and the custom speech engine guide uses the published speech imports.
- The Schemas page
names
@rasahq/mantle-authoring-schemasas the versioned contract for authoring files. The package version matches the Rasa release you run, and the JSON files on that page stay the currentmaincontract.
Breaking Changes
enteris now a reserved framework tool name. Rename any builder tool or MCP tool calledenter;rasa trainfails until you do. If yourinvoke_when.activatetext tells the model to useactivatefor blocks of the running skill, update it to useenter.activateof{skill_id}__{block_id}starts that skill and leaves the named block closed. A fully guided skill still opens its entry block because the skill started. Useenterto open a block.- The built-in
set_fieldstool is renamed toset_memory. Rename it inorchestrator.tool_constraintsinagent.ymland in skill instructions;rasa trainfails on the old name until you do.
Features
- Voice channels can now use AssemblyAI for real-time speech recognition and speech synthesis.
- A custom ASR
engine subclasses
ASREnginefromrasa.api.voice.asr.v1, and a custom TTS engine subclassesTTSEnginefromrasa.api.voice.tts.v1. Shared audio types come fromrasa.api.voice.v1. Channelprovidervalues are unchanged, and a custom class path still loads with the same beta warning. - A custom tracker
store subclasses
BaseTrackerStorefromrasa.api.tracker.v1and implementsconnect,save_tracker,retrieve_tracker, andtracker_keys. The built-inin_memory,redis,mongod,sql, anddynamostores use that contract. Rasa owns broker streaming, session slicing, creation, and deserialisation. Configure the store withtracker_store.typeinintegrations.yml. - Builders can read
the names, parameters, and injected arguments of
@tooland the hook decorators from the Python decorator contract. - A custom input
channel subclasses
InputChannelfromrasa.api.channels.v1, or a published builtin such asTwilioInputorSlackInputfrom the same package, and implementsconnectplusblueprint. Set the channel key inintegrations.ymlto a builtin short name such asrestortwilio, or to that class. A class imported fromrasa.core.channelsis rejected. A custom output channel subclassesOutputChannel,CollectingOutputChannel, or a published output builtin from the same package. - A custom event
broker subclasses
EventBrokerfromrasa.api.brokers.v1and implementsconnectandpublish. Brokers may also overridepublish_batchto send one conversation’s new events together; the default publishes them one at a time. Omittingevent_brokerpublishes nothing. Starting a Mantle agent withintegrations.ymluses that event broker. Pre-existing broker behaviour is preserved in the migration. - A custom knowledge
store subclasses
KnowledgeStorefromrasa.api.knowledge.v1and implementsconnectandsearch. Setknowledge_stores[].typeinintegrations.ymltofaiss,qdrant,milvus, or that class. - The Studio HTTP
event broker (
type: http) is published underrasa.api.brokers.v1with the other builtins. ImportHttpEventBrokerfrom there; configure it in Mantleintegrations.yml. CALMendpoints.ymldoes not construct this broker. - A custom timer
store subclasses
SessionTimerStorefromrasa.api.timers.v1and implementsconnect,store_timer,delete_timer,get_timer, andget_expired_timers. Settimer_store.typeinintegrations.ymlto a builtin or that class. A custom store recovers persisted timers after a restart. Omittingtimer_storekeeps the in-memory store. - Starting a skill
can now record memory values the user already stated, so they are not asked
again. See
activate.
Improvements
- When to call a built-in tool is now in that tool’s description. The system prompt keeps the search trigger, the instruction to answer from tool results before
complete(), the line that callscancel()when the user abandons the whole task, and theset_memoryresume bullet.invoke_whenadds your text for most tools, and still replaces the search trigger and the resume bullet.activaterecords values the user already stated inmemory_items. The routing list labels those namesmemory items. - When call recording is enabled, the event broker receives an event when a recording starts, is stored, fails, or captures no audio. These events are not added to the conversation, and voice traces no longer include a span for storing the recording.
- A custom lock store
subclasses
BaseLockStorefromrasa.api.locks.v1and implementsconnect,get_lock,save_lock, anddelete_lock. Conversation ordering is included. - Pending tool confirmation now tells the model to take consent or denial only from the reply to the confirmation question, so an earlier “I already confirmed” is not treated as an answer.
- A skill that waits
on its
resolve_withskill is reset once, when it is parked. It is not reset again when it starts, so values saved to it while it waited are not wiped. - A number stated for a
text field is stored as text, and a whole number stated for an integer field
is stored as an integer.
correctuses the same conversion. A whole number such as4109.0is stored as4109. - Intermediate markers are now generated from any-level timestamps that TTS providers send. They are no longer processed to carry only word-level timestamp. On barge-in, heard text is reconstructed from acknowledged markers and sent to LLM.
- Deepgram Nova and Flux ASR
endpointvalues now accept full WebSocket URLs, while existing host-only values continue to work. - Mantle startup no longer loads the CALM validator, so Mantle installs can omit scikit-learn and SciPy.
- Coding agents and the Rasa Copilot now read the best practices pages before writing a skill’s description or instructions, or adding a control lever, so new skills follow them from the first draft.
- Scenarios can now set an optional
description, one line saying what the scenario checks. Studio shows it under the scenario name.
Bugfixes
- A titled memory button tap now keeps its label before that turn’s tool results in the model prompt, so the agent continues with the next step.
- Voice channels
loaded by module path now use
voice.asrandvoice.ttsfromagent.yml, like built-in voice channels do. - Calling
enterfor the ordered block that is already running continues that block. If it is still waiting on a collect answer, the result names the field to record withset_memory. - A collect question that was already asked now comes back with its buttons after a side question, and after the user confirms a return from another skill. A message that arrives before the question was asked is not followed by a repeat, and a value given in that same turn still fills the field.
- A keypad answer during a silence check-in now counts as a response, so the check-in ends and the silence count starts over.
Breaking Changes
- Bundled skill ids now start with
builtin_instead ofdefault_. Rename a project override such asdefault_completedtobuiltin_completed. Training fails when any other project skill id starts withbuiltin_. - The bundled
builtin_customer_satisfactionskill now asks whether the user was satisfied or unsatisfied, and stores that outcome incsat_score, instead of collecting a 1–5 score. Projects that read a numeric score, or that override the skill and still expect a float, need to follow the new categorical values. - Buttons on a
confirmation question must use
set:session.system.confirmation=trueorset:session.system.confirmation=false. Any other button on that question, or that key on any other response, failsrasa train. A question with no buttons is unchanged. - When the A2A
server has
authset, an orchestrator JWT that names none ofsub,preferred_username, orclient_idis now rejected with401instead of being accepted as a shared anonymous caller. Add one of those claims to the orchestrator’s token. - The Copilot
plan write routes are gone:
POST /agent-plans,POST /agent-plans/<id>/refine,PATCH /agent-plans/<id>,DELETE /agent-plans/<id>andPOST /agent-plans/<id>/approvenow answer 404. A client plans in the chat instead: every chat run for an empty agent plans it, and the client approves, discards or edits the plan with theapprove_plan,discard_planandsave_plan_textchat actions. A refused click answers 400 for a malformed action, 404 for a plan outside the chat, and 409 with acodesuch asstale_revision,job_runningoragent_not_emptywhen the plan’s state refuses it. The read routesGET /agent-plansandGET /agent-plans/<id>stay. - Copilot now treats an agent with no commits as a Mantle agent in every chat.
Features
- A skill can import
mcp/<server>:*ormcp/<server>:<prefix>*so tools that match the server’s list are available after a restart or model reload, without naming each tool. - You can add a
customJSON object to a response variant and have it delivered to the channel. - A confirmation
question can carry
set:session.system.confirmation=trueand=falsebuttons so a tap agrees or declines without a model call. - Voice channels can
temporarily ignore further interruptions after too many occur in a short
time. Set
max_interruptions,window, anddurationtogether underinterruptions.cooldowninintegrations.yml, or undervoice.interruptions.cooldowninagent.yml. A channel block replaces the agent block. The voice prompt now tells the agent to answer briefly after an interruption and not restart a long response that was cut off. - Voice channels now
take ASR and TTS from a model group id on the channel in
integrations.yml, or fromvoice.asrandvoice.ttsinagent.yml, with the channel value winning. Inspector still uses Deepgram when a side is unset; other voice channels fail until that side is set. An engine config usesprovider(namestill works and warns), andapi_keymust be${ENV_VAR_NAME}or omitted so the engine uses its usual environment variable; a literal key is rejected. A Flux group setsprovidertodeepgram_flux;provider: deepgramstays on Nova. A speech group lists one model. A speech group cannot be used asorchestrator.model_groupor as an embeddings model. - Voice agents can
set interruption handling once for every voice channel with
voice.interruptionsinagent.yml. It takes the sameenabled,min_words, andbackchannel_thresholdkeys as a channel’sinterruptionsblock, and a channel block overrides only the keys it sets. - For a new agent, Copilot now asks a few questions, plans the agent and revises the plan in the chat, and it does not change any files while it plans. Planning runs in the background, so it continues after a page reload. Plans list the agent’s integrations and the tools each skill uses. Approving the plan, starting over and editing the plan text all happen in the chat, and only the Approve button starts a build.
- Eval results now include how many prompt tokens were served from cache and how many were written into the cache, beside the existing conversation token totals. Both counts are already included in the prompt total.
- Eval results now also list the conversation token totals for each model call and each model. The calls are the main reply, a rephrase, a confirmation of that rephrase, a silent field update, the completion judge, and embedding lookups for skill retrieval and knowledge.
- Eval
run_N.jsonnow adds US list-price estimates to the existingengine_tokensobject for the dialogue engine’s LLM calls only (not the user simulator or the eval judge). If LiteLLM cannot price a model, the token counts are still recorded andcostisnullon the conversation and on any bucket that includes that call.run_N.txtstill omits tokens and cost. - Eval
run_N.jsonnow addsevaluation_tokenswith US list-price estimates for the user simulator and the eval judge, separate from the agent spend inengine_tokens.summary.jsonadds per-scenarioagent_costandevaluation_cost, andsummary.txtsums them intoAgent cost:andEvaluation cost:lines. If LiteLLM cannot price a model, the token counts are still recorded andcostisnull, so an affected total is incomplete.run_N.txtstill omits tokens and cost. - Voice eval
run_N.jsonfiles now include how many seconds of audio were sent to speech recognition and how many characters were sent to speech synthesis, per vendor and model. These counts are not a price, they stay out ofrun_N.txt, and a text run omits them. - With
RASA_RECORD_COMMANDS_AND_PROMPTS=true, rephrased responses published to the event broker carry their rephrase prompt and their text before rephrasing, so Rasa Studio can show them. Apromptsordomain_ground_truthkey in response variantmetadatais dropped with a warning.
Improvements
- Studio and the
Inspector now show each MCP import pattern you wrote. Studio names it like
card_*or*, with the server, and marks the row as a pattern rather than a tool. The Inspector lists the full pattern, such asmcp/banking:card_*, under Available tools. The tools that match still appear only after the agent starts or reloads. - When you omit
reasoning_effortin your model config, Rasa now defaults gpt-6-luna and gpt-6-sol toreasoning_effort: "none"(the lowest supported setting). gpt-6-astra and gpt-6.1-sol are unchanged: Rasa still omitsreasoning_effortso the provider default applies unless you set it explicitly. - The inbound A2A
server rejects an
httppush callback that sets a token or authentication. A callback with neither still works, andrasa runstill starts. Get and list of a push config no longer return the token or credentials; send them again to update a config. Whenauthis set, the caller that first uses acontext_idowns that conversation, and a different caller receives a context-not-found error. The first message may still choose its owncontext_id. Rotating the JWT secret or public key invalidates existing A2A contexts: the same caller no longer matches conversations they already own, and follow-ups are rejected as context not found. The caller identity now includes whether it came fromsub,preferred_username, orclient_id, so changing that claim also invalidates existing contexts. - Voice channels now
check their
interruptionssettings at startup. An unknown key, a value of the wrong type, a negativemin_words, or abackchannel_thresholdoutside0to1stops startup with an error that names the channel. - When the model repeats a tool call that was already refused, or repeats a successful call that changed nothing, the turn ends with
utter_could_not_finishinstead of running until the iteration limit. An identical read that writes nothing will not run twice in a row. - When Copilot builds a new agent from an approved plan, it now shows each build step in the chat, stops a build that takes longer than 15 minutes, and tells you in the chat whether the build worked.
- Eval run result
files (
run_N.json) now include the judge’s explanation for each criterion, quality metric and task completion, the judges’ overall summary, the detail of each assertion, and the conversation transcript. - The
builder service’s
GET /api/eval/batches/<batch_id>now also returnsstarted, the runs that have started and have no result file yet, so Studio can show which runs of a running batch are in progress. - Eval results now show a criterion as not reached when the conversation never got to it, instead of failed. A criterion that was not reached still counts as not passed.
Bugfixes
-
Sending
/session_endon a text channel ends the Mantle conversation without a model call. - A hybrid skill now finishes when one of its blocks links to another skill, even if another block of that skill was still open.
-
After a knowledge
search, activating the skill that is already running continues that skill.
The agent no longer keeps calling
activateuntil the turn stops. -
An
interruptionsblock on the Inspector channel that sets only some keys, such asmin_words, no longer turns interruptions off. - A skill that starts again after it has finished or been cancelled asks a collect question again when that field is empty; a previous run’s answer no longer counts, and a value already set is skipped. A link or call that stays inside the same skill is still one run.
- A fixed reply followed by another model step in the same turn no longer fails on Claude. A model group that includes Claude uses that same request, so the other models in the group see the note too.
-
A
set_fieldscall with no fields is skipped. The turn continues so the model can still reply, and repeating that empty call ends the turn.
Documentation
- The docs now describe interruptions correctly: playback stops, a tool that is already running finishes, then the agent answers what the caller said. Only a session timeout, a hangup, or a dropped connection cancels a tool.
Features
- You can set
orchestrator.tool_constraints.listeninagent.ymlto turn offlistenor replace when the model should call it. Thelistendescription only says to wait without a message and to call it with no text. When to call it stays in the prompt, including not saving a partial answer. Omitting the key keeps the default offer rules. A retrain is required. - Mantle agents can use speech to speech models and APIs, including OpenAI Realtime, OpenAI GPT-Live, and PolyAI Dialog-RSN-1, to handle live voice conversations while the Rasa agent carries out background work.
Improvements
- ElevenLabs speech
synthesis now uses the
eleven_v4_turbomodel by default and supports theeleven_v3andeleven_v4models. Earlier models such aseleven_multilingual_v2are no longer supported, so setmodel_id(or the per-languagemodel) to aneleven_v3oreleven_v4model if you configured one explicitly. Thesimilarity_boostsetting is ignored. - License validation warnings and errors now include the last four characters of the license key, making it easier to identify the configured license without exposing the full key.
rasa initnow 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-promptand no--ides, Claude Code and Cursor are set up as before.- 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.instructionskeep their wording. - Copilot can show cards with buttons in the chat, and a button click goes through the same conversation, so Copilot sees what the user clicked. A click never starts a model run.
- Eval run result
files (
run_N.json) now record the scenario name and type, the run’serrorwhen it failed, whether that error came from the evaluation itself (error_is_harness), and the quality-metrics judge’smetrics_error. - You can ask Copilot to train or retrain your agent without changing it, for example after a failed training, and Studio then trains the agent as it is. Copilot no longer tells you to use Apply Changes or claims that training succeeded, since Studio trains the agent after every change and shows its status.
Bugfixes
- Conditions in the
nextof acallstep now see values the called skill set, so you can branch on the result of the call. - A hybrid skill that links to another skill now ends, so the conversation can start that skill again.
Bugfixes
- Voice hangups now close the Mantle conversation without starting an engine turn.
Documentation
- The Anthropic
integrations.ymlexample recommendsclaude-sonnet-4-6.
Features
- Mantle voice channels can now use Speechmatics Agent STT for turn-based transcription with automatic turn detection.
meta_map.from_memorysends skill or project memory values as MCP_metaon each tool call. An unset field is sent as null. Metadata from amodify_tool_callhook still replaces those keys.- Project memory
fields can opt into mid-session updates with
mutable: trueand restrict which skills may write them withwritable_by. - A
modify_tool_callhook can attachmetadatasent only as MCP_metaon 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 onsystem.pending_tool_confirmationand replayed on resume. The write stays in tracker history after the slot is cleared, so do not put secrets inmetadata. - 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
- Scenarios can now set an optional
typeto label the kind of case they cover, such ashappy_path,edge_case, orescalation. - 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.
- The builder service’s
GET /api/eval/scenariosnow returns each scenario’s parsed fields (type,simulation_context,criteria,initial_slots,assertions, andexpected_tools) alongside its raw contents. - 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
- 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.
First official open beta. Rasa Pro 3.21.0 is the first official open beta
release of Mantle. Start a project with
rasa init --engine mantle, then follow
the getting started guide.Features
- Mantle CLI commands
accept
--integrationsto load runtime configuration from a file outside the project root.
Bugfixes
- 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_idnames that artifact. An interrupted stream is not stored as a finished message. - 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
- Use
name: deepgram_fluxfor Flux ASR configurations.name: deepgramnow selects Nova, so update existing Flux configurations before upgrading.
Improvements
- The voice backchannel detection model is now downloaded from the Rasa CDN instead of Hugging Face.
- Voice agents now mistake fewer real interruptions for backchannels like “uh-huh”, so callers who cut in are heard more reliably.
The default
backchannel_thresholdis now0.7. If you set it explicitly, re-check your value.
Bugfixes
- Starting a conversation with
/session_start(for example from Inspector) no longer routes the trigger as user input, so the bundledbuiltin_session_startskill 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. - 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.
- Deepgram Flux no longer loses an utterance’s final transcript when it differs from the last partial transcript.
- Mantle agents without privacy configuration publish conversation events to Conversation Review again.
- 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.
- 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
- 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.
- Voice agents can play
ambient background audio, such as office noise, during calls. Add a
background_audioblock to a voice channel inintegrations.ymlto set the volume or use your own audio file.
Improvements
- Voice channels can
set the pause between consecutive agent messages with
pause_between_messagesinintegrations.yml. The default stays at 1 second, and0adds no pause. - Voice channels now
wait 0.8 seconds after a filler before the next message. Set
pause_after_filleron the channel inintegrations.ymlto change it, or set it to0for no pause.
Bugfixes
- Twilio Media Streams can now use a custom sender ID to select an existing conversation while keeping call audio and playback controls working.
- The
recordingsconfiguration no longer accepts unknown or misspelled keys.provideris required when recordings are enabled, ands3is no longer the default provider.
Breaking Changes
- An
endpoints.ymlfile at the project root fails startup, and passing--endpointsfails too. Mantle only acceptsintegrations.yml. Deleteendpoints.yml. The silence timeout belongs inagent.ymlasvoice.silence_timeout. Call recordings are read only fromintegrations.yml. Move arecordings:block torecordings:before upgrading. Omit it to leave recording off. Studio trains and runs a Mantle assistant without--endpoints, so those launches succeed. Anendpoints.ymlin another working directory is not this project’s file and does not fail startup.
Features
- Copilot now searches the Rasa docs and answers Mantle questions from the Mantle docs without extra configuration, because the
rasa-propackage and therasa/rasa-proDocker image include both documentation indexes.
Improvements
- 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.
rasa tools initnow sets up Mantle projects: it detects them, downloads the Mantle documentation, and offers to install the Mantle agent skills. Use--engine mantleto choose Mantle explicitly. For Mantle projects, the Rasa MCP server no longer describes CALM-only tools it does not offer, andtrain_rasa_assistantalso works when your IDE starts the server outside the project folder.- 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
- 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.
- The model answers from relevant tool results before it finishes a skill, so a lookup is not replaced by the wrap-up question.
- When a turn hits the orchestration iteration limit, the agent now speaks
utter_model_call_errorinstead of staying silent. Setmetadata.rephrase: trueon your override to have it reworded. rasa trainvalidation logs, the error for a removed collecttool:key, and the text returned when a write is rejected now refer torun_after_setting_<entry>tools as memory validation tools.
Documentation
- Docs for
ActivateSkillnow 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
- Omitting
skill_retrievalleaves retrieval off, so the full routing catalogue stays and turns do not call the embedder. Setactive: trueto opt in. Retrieval then runs only when there are more startable skills thannum_skills(at least 1). A smaller catalogue is not indexed or embedded, and train logs a warning with both counts. - Answering from documents, the pre-tool line, and channel style move in
agent.yml. Useknowledge_answering(includinginstructionsandtop_k),fillers, andorchestrator.text_rules/orchestrator.voice_rules. A skill setsknowledge_answering.instructionsin its frontmatter.promptsstays forrouting_no_active_skill. Filler and channel-style keys belong onfillersandorchestrator. Omitting the new keys keeps the previous behavior, including four chunks and voice-on, text-off fillers. Retrain after you move the keys. - Fully guided skills start with
activateand the skill id. Routing and topic change list that skill id and the skill description. An ordered block no longer has adescription; training fails if one is still set. A fully guided skill has one entry block, and other blocks are reached withcall:orlink:. - Silence hangup is its own skill,
builtin_silence_timeout_hangup.builtin_silence_timeoutonly checks in. An override ofbuiltin_silence_timeoutthat still contains the hangup block fails training. rasa trainrejects adescriptionon a memory field the model cannot set, including project fields and bundled default skills. Keep a description only on a field markedllm_settableor filled by a collect step. Remove it from every other field.- Lifecycle tools are now
completeandcancelinstead ofcomplete_skillandcancel_skill. Update skill prose that tells the model to call the old names, and renameorchestrator.tool_constraintskeys fromcomplete_skillandcancel_skilltocompleteandcancel. A custom@toolcannot be namedcompleteorcancel. rasa trainnow rejects unknown top-level keys inintegrations.yml. CALM-only keys (nlg,nlu,action_endpoint,vector_store,interaction_handling) were silently ignored before; each error names what to use instead. Arecordings:section logs a warning, since it still belongs inendpoints.yml. See Keys this file does not accept.- 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
-
You can
mark memory fields
pii: trueand set ananonymizationmethod on that field.agent.ymlprivacyholds tracker-store cron settings only. Published events are anonymized whenintegrations.ymlsetsstream_pii: false. Optional GLiNER-based detection (when thepiiinstall extra and model are available) masks additional entities in user and bot message text. -
You can set
orchestrator.tool_constraintsinagent.ymlto turn offcancel,complete,search_knowledge, orcannot_help, and to replace a built-in tool’s when-to-call guidance withinvoke_when. Omitting the section keeps today’s prompt and offer rules. A retrain is required to pick uptool_constraintsfromagent.yml. -
A top-level
a2a_serverblock inintegrations.ymlconfigures the inbound A2A server. It acceptsinclude_default_skills,a2a_message_cache_ttl_seconds, and the push notification keys;${ENV_VAR}references are expanded, unknown keys (includingenabled) and a literal JWTsecretare rejected, andchannels.a2ais not an input channel. With the block present,session_configinagent.ymlmust setstart_session_after_expiry: false. Whenpush_notifications_enabledis true, task updates are posted to callback URLs and the AgentCard advertises them. - 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.
-
The inbound A2A AgentCard is generated from the skill catalog and named with
agent.idfromagent.yml. Routable skills are advertised with theirskill.mdname and description;engine_managed,disabled, andhiddenskills stay off the card, and a default skill appears only wheninclude_default_skillsistrueand a project override made it routable. A staticagent_card_pathmust be A2A 1.0 JSON withsupportedInterfaces. The card is generated when the model loads at startup. -
PUT /modelandDELETE /modelon 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 /modelpublishes 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 /modelof a model trained for the other engine fails the same way, so a running CALM server is not migrated to Mantle through these endpoints; restartrasa runwith the retrained model. -
You can require confirmation before
cancelruns by settingorchestrator.tool_constraints.cancel.requires_confirmationinagent.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. Seetool_constraints.
Improvements
- 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.
- The system prompt no longer includes the skills glossary.
- Coding agents that author skills now apply the decision rules for when a sentence in the skill body is not enough.
- 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
- 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.
rasa runandrasa inspectload channels, endpoints, and the secrets manager fromintegrations.ymlin the project directory when--modelis a Mantle archive, even with noskills/folder beside it. Tracker, lock, and timer values that remain only inendpoints.ymlfall back to memory, and an event broker or model server left only there is ignored, so move those blocks intointegrations.ymlbefore serving this way. If the archive cannot be read as Mantle, those commands keepcredentials.ymlandendpoints.ymland log a warning.rasa init --engine mantleregisters the rasa-tools MCP server for Claude Code and Cursor, so simulation and evaluation no longer require a separate setup step.- 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. - 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
-
Documented
privacyinagent.ymland the Mantle PII pipeline inmemory.yml. Themantle_privacy_e2eagent in maestro-benchmarking-agents includes local Kafka and GLiNER verification steps. - The tools reference defines each built-in tool, including what it does and when the model is offered it.
Breaking Changes
-
The tracker store, lock store, and secrets manager are read only from
integrations.yml.tracker_store:,lock_store:, andsecrets_manager:blocks inendpoints.ymlare ignored, and startup logs a warning. Move them totracker_store:,lock_store:, andsecrets_manager:before upgrading. Atracker_store:block left only inendpoints.ymlstores conversations in memory, so they are lost when the process restarts. -
The event
broker, timer store, and remote model server are read only from
integrations.yml.event_broker:,timer_store:, andmodels:blocks inendpoints.ymlare now ignored, and startup logs a warning. Move them toevent_broker:,timer_store:, andrasa_model_server:inintegrations.yml. -
The orchestrator model group is named with
orchestrator.model_groupinagent.yml. Optionalorchestrator.max_prompt_tokensdefaults to 8000.integrations.ymlmust not contain anllm:section; define the group undermodel_groups. Existing projects need that name added and a retrain. -
Unknown top-level keys in
skill.mdfrontmatter are rejected at train. Remove the extra key.
Features
- Tools can stream a message to the user and attach buttons to a message they send.
-
You can configure the tracker store, lock store, and Vault secrets manager
in
integrations.yml.VAULT_URL,VAULT_TOKEN, andVAULT_RASA_SECRETS_PATHstill overridesecrets_manager:and still work when that section is omitted. See Tracker store, Lock store, and Secrets manager. -
You can
configure the event broker, timer store, and remote model server under
integrations.ymlinstead ofendpoints.yml. Userasa_model_server:for the remote model server, not themodelslist inside a model group. With no local model,rasa runpoints atintegrations.yml. See Event broker, Timer store, and Remote model server. -
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’sactivatetool. SeeActivateSkill. - 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
-
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 subclassVoiceInputChanneland overridehandle_asr_event, add the new optionalasr_metadatakeyword argument to your override. - Voice agents now keep work already underway when a user speaks over a response, then respond to their latest input once that work finishes.
-
A missing model group named by
references.embeddingsororchestrator.model_groupnow reports that id. Thereferences.embeddingserror 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. - Turns no longer wait on an extra model call after a skill switch to store facts the conversation does not use.
-
on_model_request/modify_model_requestandon_model_response/modify_model_responsealso run on rephrase, confirm-rephrase, and the post-activateset_fieldsfill. A request-hook crash does not send that prompt.RetryModelon those calls discards the completion instead of starting another main-loop iteration. -
rasa exportpublishes Mantle ProtoJSON when--integrationspoints at an existing integrations file (defaultintegrations.yml). Classic projects keep flat events. -
Training or validating a Mantle project that has no skills reports that no
skills/<id>/skill.mdfiles were found. - When training fails on a Mantle project, Copilot’s analysis now searches the Mantle documentation, as Copilot chat already does.
-
Mantle trackers now
record skill, block, and memory changes only as
skill_*,ordered_block_*,memory_set, andmemory_clearedevents; the classicflow_*andslot_setcopies are no longer written. Conversations recorded with those copies still load.
Bugfixes
- Sanic workers now load the secrets manager. A Vault-backed tracker store works with more than one worker for Mantle and for classic projects.
- Resuming an older
paused skill with
activatepauses 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. - 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
- Mantle docs now explain how to simulate and evaluate an agent, including the assertion keys a Mantle scenario can use.
- 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.
- 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. - 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, andsilence_timeoutfor voice channels, and documents keypad input withdtmf:on a collect step and theuser_input_suppressedevent.
Deprecations and Removals
rasa tools runno longer offers thesearch_rasa_documentationtool, which searched the CALM docs. Your coding agent reads the Mantle docs in.rasa/docs/mantle/through themantle-docsskill; add them to an existing project withrasa tools init docs mantle.
Breaking Changes
- Tracing and
metrics are read only from
integrations.yml.tracing:(including Langfuse) andmetrics:blocks inendpoints.ymlare now ignored, and startup logs a warning. Move them totracing:andmetrics:inintegrations.yml.
Features
- You can configure
OpenTelemetry tracing (
otlporjaeger) and OTLP metrics underintegrations.ymlinstead ofendpoints.yml. List Langfuse and an OTLP or Jaeger tracer together undertracing:to use both. A Jaeger entry withusername/passwordmust setinsecure(falseexports over TLS). See Tracing and Metrics.
Breaking Changes
- Models trained before skill retrieval that still have independently startable skills must be retrained before they will load (or set
skill_retrieval.active: false). Omittingskill_retrievalleaves retrieval on. rasa trainrequires a non-empty skilldescription, rejects unknownskill.mdfrontmatter keys and non-stringname/description, and fails on incomplete@skill/@block/@tooltokens. Bundled default skills are checked the same way. Seeskill.md.
Features
- Mantle agents can now retrieve a subset of independently startable skills for routing.
skill_retrievalis on by default; setactive: falseto keep the full routing catalogue. - Skill retrieval can pin skills with
always_include_in_promptand 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). - A collect step can accept keypad input via
dtmf, usinglengthorfinish_on_key. Setallow_audio_input: falseto ignore speech while digits are expected.
Improvements
- 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
- The first message of a new conversation no longer activates a skill before that message is visible.
- 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_constraintsor ordered-blocktools.
Documentation
- Builders can look up Mantle’s default skills and the wording of the built-in messages.
- Docs call
run_after_setting_<entry>tools memory validation tools.
Deprecations and Removals
- The unused default response
utter_ask_how_to_helpis removed. Completed still asks withutter_ask_wants_to_continue.
Improvements
- ENG-3360: When
modify_model_requestormodify_tool_callcrashes or times out, or the main model call fails, the user hearsutter_model_request_hook_error,utter_tool_call_hook_error, orutter_model_call_errorinstead of silence. Override those names inresponses.yml. The exception text is not spoken, andrephraseon those responses is ignored.
Breaking Changes
-
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. -
Skill-level
requiresis rejected at train. Usepreconditionplusagent.ymlorchestrator.preconditionsto run a resolver first, orhidden: trueto omit a skill from routing. Tool-levelrequiresundertool_constraintsis unchanged. Seeprecondition,preconditions, andhidden. -
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. -
The default voice
silence timeout is now 30 seconds (was 7). Set
silence_timeouton the voice channel inintegrations.ymlif you need a different wait.
Features
-
Skills can declare a semantic
precondition, and projects bind it inagent.ymlorchestrator.preconditionsso 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 sethidden: trueto stay off the routing catalogue while remaining startable viacall,linkorresolve_with; prose@skillmentions of a hidden skill fail validation. Inspector labels a parked consumer as pending while the resolver is the active skill. -
on_incoming_message/modify_incoming_messagenow 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_messagemay select a bot-defined response, while failures useutter_incoming_message_rejected. When that first payload also opens a session,builtin_session_startstill runs so session-start memory writes apply, and named reject replies are resolved after that turn. See Hooks. -
on_outgoing_message/modify_outgoing_messagenow 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_messageorutter_outgoing_message_rejected), so the tracker only ever records what the user actually received.modify_outgoing_messagebuffers 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. -
Train and validate fail when
skill.mdprose references@tool.<name>for a tool that does not exist or is not visible to that skill. -
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: hangupafter an authored goodbye response.
Improvements
- Mantle conversation
events now include stable
event_idandsequencefields so consumers can order and deduplicate reliably. - Mantle ProtoJSON now covers the remaining conversation event types (session, voice, skills, ordered blocks, and related engine events).
- Voice channels now
use a classifier and
backchannel_thresholdto distinguish short backchannels from interruptions, using the current agent utterance as context. Accepted interruptions enter the active turn, while ignored backchannels remain visible asuser_input_suppressedevents; the classifier loads at channel startup and downloads from Rasa’s CDN if it is not cached. See voice interruption settings.
Bugfixes
- 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.
- 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.
- Agents no longer receive
set_fieldsprompt guidance when the tool is unavailable for the current turn.
Documentation
- Documented Mantle conversation event ProtoJSON under Conversation events, including the dual-write skip list and golden fixtures for consumers.
- Mantle documentation now uses Rasa’s rebrand colors and typography for a consistent visual experience.
Breaking Changes
- The bundled
builtin_session_startskill no longer greets. Override that skill and send/session_startwhen the agent should speak first. See Start a conversation.
Features
-
on_turn_started/on_turn_finishednow run around each Mantle turn. Builders can observe turn start and the finalcompleted/cancelled/errorstatus without blocking the turn. See Hooks. -
on_knowledge_query/modify_knowledge_queryandon_knowledge_results/modify_knowledge_resultsnow run aroundsearch_knowledge. Builders can rewrite the query or filter/reorder retrieved passages; neither point acceptsHookRejectionorRetryModel. Query-hook crashes are fail-closed (search skipped); results-hook crashes are fail-open (original passages kept). See Hooks. -
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. -
A response variant can
name a tool with
buttons: {source: <tool_name>}, and that tool supplies the buttons for the message withcontext.set_buttons([...]). Buttons built with arefwrite the chosen value to memory on tap, exactly like a staticset:session.<skill>.<entry>=<value>button; the message is sent without buttons when the tool returns none or fails. -
The model can now wait for the next user message without sending a reply,
by calling the reserved
listentool after a question already asked this turn. Guided collect steps, post-search answer calls, and pending tool confirmation do not offer this tool. Mixedlistenplus text or other tools ignoreslisten. -
on_model_request/modify_model_requestnow run on each main LLM loop iteration. Builders can rewrite provider-bound messages and allowlisted per-callmodel_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 (HookRejectionis not accepted). See Hooks. -
on_model_response/modify_model_responsenow run after each main LLM completion and before any user-visible emit or tool dispatch. Builders can rewrite response text and tool calls, or raiseRetryModel(feedback=...)to discard the response and steer the next iteration with an engine-only system note.HookRejectionis 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. -
on_tool_call/modify_tool_callandon_tool_result/modify_tool_resultnow run on the LLM tool-call batch path. The engine prefights every proposed call before any tool runs;RetryModeldiscards 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 whatMcpToolExecuted/ToolExecutedstore and the next prompt replays.HookRejectionis 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
-
Session-start
channel metadata on Mantle conversations is now recorded as a
memory_setevent (with write provenance), so Studio and other consumers can read it without relying on a legacyslotevent for that value. -
context.memory.getaccepts an optional default. Unset,None, or unreadable fields return that fallback (get("card_list", [])is[]) instead ofNoneor aTypeError. Local tool failures logerror_typeat ERROR and the exception message at DEBUG. - Inspector bot-turn latency details now include time spent in earlier LLM calls, making multi-step and tool-driven responses easier to diagnose.
-
complete_skillis 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_skillis still available. -
Voice conversations now
process overlapping transcripts normally and preserve rejected barge-ins or audio
blocked during DTMF collection in the new
user_input_suppressedtracker event. -
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, andgenerative_response_is_grounded. Use the memory assertions for written state.
Bugfixes
- Deepgram Flux TTS now accepts the provider’s full speed range and supports models that require the default speed setting.
-
Hybrid skills no
longer offer
complete_skillwhile an ordered block is active, preventing repeated refused completion calls before the block finishes. - 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.
-
complete_skillno 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. - Copilot documentation search now finds pages on the current docs site and omits unpublished drafts.
- Mantle voice calls now keep processing spans in a single call trace, making individual calls easier to debug.
- Voice agents now release LLM requests as soon as response generation completes while speech continues in order and remains interruptible.
-
context.memory.set("project.<entry>", …)andcontext.memory.set("<active_skill>.<entry>", …)now write the named field instead of failing as undeclared, matching howgetalready resolves those names. An own-skill qualified name does not fall through to project memory. Cross-skill writes stay rejected. -
Skills with
complete_whencriteria 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. - After a canned ordered-block reply, leftover hybrid prose can continue on the same turn instead of waiting for the next user message.
- Custom default skills now start from their configured entry and are discarded cleanly when the customer moves to another task.
-
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_toolstep of the same tool can still ask again in that turn.
Documentation
- New
Writing prose page covers how to word a skill’s
descriptionand 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. - Skills that use
requires_confirmationmust not also ask for confirmation in prose — pick one mechanism so the customer is not prompted twice. - Clarified hook
control-flow semantics, including
RetryModelfeedback as a non-persisted system note, sequential multi-toolmodify_tool_callpreflight, discarded text handling, and unavailable audio hooks.
Deprecations and Removals
- Skill
deny_read/deny_writelists are removed. Memory visibility and writes follow public, private, and project scopes only. Leftoveraccess:/deny_read/deny_writekeys in skillmemory.ymlare ignored and are not enforced.
Breaking Changes
- Skill-level
complete_whenmust be a mapping withconditionand/orcriteria(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_skillis hard-gated on those requirements and is refused while an ordered block is active. See skill.md.
Features
- The live conversation
stack now distinguishes skills from ordered blocks. Guided skills show the
skill together with a
mainblock. 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-levelflow_*events are still written as well. - Mantle voice agents can now
respond when a user remains silent for the configured timeout through the customizable
builtin_silence_timeoutskill. Silence timeouts and their configured duration are recorded in the conversation tracker and included in the agent’s conversation context. - Optional
complete_when.criteria(sentences in natural language) are checked oncomplete_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
- Large FAQ files under
references/now train without a manual split, andsearch_knowledgereturns the matching section. Builders can set character chunk size and overlap underreferences.chunking.
Bugfixes
complete_skillis refused while a tool confirmation is pending for the active skill. Resolve or decline the confirmation first;cancel_skillstill abandons the skill.- 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
correctinstead of silently keeping the old value. - 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_failureresponse. Orderedexecute_toolsteps with invalid arguments stay on the step without running the tool oron_failure.
Features
- Conversations now record ordered-block lifecycle
(
ordered_block_entered/completed/cancelled/interrupted/resumed). Block-levelflow_*events are still written as well. Inspector and Copilot map these types. See Tracker events. - Tapping a button whose
payloadis aset:memory reference now writes the declared memory field directly, without an LLMset_fieldscall (llm_settableis 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-validset: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 includesbutton_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 declaringutter_set_memory_button_skippedin your ownresponses.yml. - Response variant
metadatafromresponses.ymlis now stored on bot messages in the conversation tracker, not only the internalrephraseflag.
Improvements
- Conversation-scoped
structured logs now include
conversation_idacross engine and voice processing. Voice TTS debug events no longer also emit a duplicaterecipient_id.
Bugfixes
- Updated the prompt to
instruct the agent to answer in-skill follow-ups instead of calling
cannot_help. - Clarification questions and task switches no longer cancel a pending tool confirmation.
Features
- You can now add static
buttonsto named responses inresponses.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-textpayload(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_*andmodify_*decorators, typed payloads, andHookRejection/RetryModelfromrasa.mantle.hooks. Place modules inhooks.pyorhooks/**/*.py; train-time validation runs before packaging. The runtime does not invoke hooks yet — turn-path wiring lands in a follow-up release.
Improvements
rasa init --engine mantlenow 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 withrasa tools init docs mantle. See Getting started.
Bugfixes
- Rasa CLI commands now work on Python 3.14.
Previously every command, including
rasa initandrasa --version, failed immediately on startup.
Breaking Changes
-
The orchestrator LLM now
uses
llm.model_groupand a matchingmodel_groupsentry. Move inline provider, model, and credential settings fromllm:into that group:Seeintegrations.yml. -
rasa data validatenow requires every conditionalnextto include anelsefallback:See Conditional branching.
Bugfixes
- Called skills no longer
restart when no conditional
nextbranch matches.
Breaking Changes
rasa data validatenow fails when a tool reaches the internal memory manager or clears memory entries. Clear fields this skill owns withcontext.memory.set. Fields another skill owns must be cleared from that skill. See Tools.
Features
- You can now declare static
buttonson response variants inresponses.yml(each with atitleand optionalpayload). This PR adds parsing and train-time validation only —rasa trainchecks button shape andset:memory payloads against the readable and button-writable memory rules the engine will use when those buttons write memory on tap; button delivery at runtime is not included yet. - 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-levelflow_*,mcp_tool_executed, andslot_setevents are still written as well. Inspector and Copilot map these types. See Tracker events.
Improvements
- 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
- Voice responses no longer play stale audio from an interrupted response after a barge-in.
Breaking Changes
- Starting the same
skill again in a session now resets that skill’s public and private memory
to each field’s
initial_value(nullwhen unset). Values still stay after complete, cancel, and interrupt until that next run. Put session-long facts in project memory. See Lifetime. - Project memory is no longer writable by the LLM (
set_fields,collect, orcorrect). Train rejectsllm_settable: trueandcollect:on a project field. Collect user-facing values in skill memory, then writeproject.*from a tool or memory validation tool. After the first successful write, further writes fail for the rest of the session. See Skill vs project memory. - 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. - Tool calls now
time out after
10seconds by default. Set the top-leveltool_timeoutinagent.ymlfor slow local@toolfunctions. Imported MCP tools use the same default unless their server sets its owntool_timeoutinintegrations.yml.
Features
- You can connect
agents to MCP servers. Declare servers under
mcp_serversinintegrations.yml. Skills can import remote MCP tools withimport_tools: [mcp/<server>:<tool>]. See Import MCP tools.
Improvements
- OpenTelemetry metrics now show the time between consecutive agent responses in a Mantle turn, including waits after fillers and before final responses.
- Set project memory now stays visible in a dedicated
### Project Memoryprompt 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. - 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 rejectsseed: trueon skill public or private fields. See Seeding from session start. - 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.
- 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
- You can now declare conditional
response variants in
responses.yml: each variant may include a memorycondition, the engine delivers the first match in YAML order or a required default, andrasa trainrejects empty text, invalid conditions, and malformed default rules with a located error.
Improvements
- 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
-
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. - 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
- 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.
-
Voice agents can now
connect to speech providers on systems that require custom CA certificates configured
through
SSL_CERT_FILE.
Breaking Changes
- The engine internals are now
named Mantle throughout (the
calm_v2package becamemantle, and the persistedengine_versionvalue changed fromcalm_v2tomantle). Custom tool modules must update their imports fromrasa.calm_v2.tools.*torasa.mantle.tools.*— e.g.from rasa.mantle.tools.decorator import tool, ToolContextandfrom rasa.mantle.tools.result import ToolResult. Agents whosetools.pyfiles import the oldrasa.calm_v2.*path fail to load until updated. Models trained on earlier.devbuilds will not load and must be retrained. - Python 3.10 is no longer supported. Use Python 3.11 through 3.14.
Improvements
- Mantle now includes the
latest Rasa Pro
mainline after 3.19.0 (Sanic 25, Python 3.14, and optional voice call recording), and pinsrasa-sdkto3.20.0.dev1.
Bugfixes
- Setting
rephrase: trueon a response now also rewords confirmation questions, messages when the user says no, and success or failure messages.
Breaking Changes
- The CLI engine value is now
mantleinstead ofmaestro. Userasa init --engine mantleandrasa tools skills install mantle. The old value is rejected rather than deprecated, so update any scripts that passmaestro. See Getting started.
Bugfixes
- Users can now correct a previously collected skill-scoped memory value.
Improvements
- Tool-call acknowledgement fillers (
prompts.ack_enabled) are off by default on text channels and on by default on voice. Setack_enabled: trueorfalseinagent.ymlto force either way for both modalities. Removed redundant inline-text guidance fromcancel_skilllifecycle and tool schema; generic acknowledgement guidance still applies when ack is enabled. - 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
- Voice agents now expose OpenTelemetry traces for calls, provider connections, transcription, input decisions, speech synthesis, and audio playback. See Observability.
- Skill instruction prose can
reference live memory with
@memory.project.<field>and@memory.<skill_id>.<field>(prompt substitution and train validation). Unreadable@memoryrefs, incomplete@memorylookalikes, and baresession.*in free prose fail train. See skill.md — Memory.
Bugfixes
- Voice-ready and custom channels can now mark incoming turns as voice so the engine applies voice-specific prompt guidance.
- Voice calls now reject spoken input while a DTMF collect step disables audio input, with the rejection recorded in tracing.
- 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
- Ordered-block steps now
reject unknown YAML properties. The removed collect-step keys
force_collectionandtool:are no longer accepted — useask_before_filling: trueinstead. The agent also fails to start when a skill directory cannot be parsed or when two skill directories in the sameskills/tree declare the same skill id.
Improvements
- Mantle projects without an
endpoints.ymlno longer log errors about the missing file when running CLI commands such asrasa train. Endpoint-backed tracing, metrics, and secrets configuration is skipped instead. CALM projects are unaffected. - Tools can read
another skill’s
publicmemory (andproject.*) via fully-qualified names oncontext.memory.get, with the same visibility as conditions /read_all(private fields anddeny_readstill blocked). Writes stay scoped to the active skill and project. - Step and skill load
validation now surfaces clearer, actionable error messages (including migration
hints for the removed collect-step keys), and
rasa data validatereports unloadable skill directories instead of relying on silent skips at runtime. - Langfuse tracing is configured under optional
tracing:inintegrations.yml(type: langfuse,${VAR}keys, optional knobs such asenvironmentandtimeout). Example agents no longer rely onendpoints.ymlfor tracing; Langfuse declared only inendpoints.ymlstill works with a deprecation warning until you migrate the block.
Documentation
- The
tools and memory.yml references
document fully-qualified
context.memory.getreads for sibling public fields. - The
skill.md reference documents ordered-block step
validation, the removed-key migration table, and collect-step authoring without
per-step
tool:overrides.