v4.0.62026-11-16
[4.0.6] - 2026-11-16
Added
- parlant lint lints a Parlant agent project through your coding agent. It
prints instructions for the coding agent: which lint chapters to run, each in
its own subagent (parlant lint --section SECTION --chapter CHAPTER prints a
chapter's rules), and how to handle findings by severity. HIGH findings are
fixed right away, at most 3 MEDIUM issue types are proposed per round, and
LOW ones only with your OK. The first sections are Behavior (observations
that combine several situations, whether alternatives or ones that must all
hold) and Wording and Tone of Voice (canned
responses rather than wording in policies or instructions). Every rule has an
ID, such as BHV001, that stays the same across releases.
- parlant init now installs a git pre-push hook (.githooks/pre-push, set
through core.hooksPath) that stops a push until the commit has been linted,
and the scaffolded AGENTS.md tells the coding agent to run parlant lint
before every push. After linting and committing the fixes, parlant lint
--done records the commit so the push goes through; git push --no-verify
skips the hook. A clone of the project turns the hook on with
git config core.hooksPath .githooks. --no-git installs no hook.
- parlant dashboard agent cards AGENT_ID DESTINATION writes every file your
site serves so personal AI agents find your agent, into DESTINATION/.well-known:
agent-card.json for A2A, and poppy.json and poppy-0.4.json for Poppy, the
Personal Agent Protocol that Meta Muse and Shopify use. --no-a2a and
--no-poppy leave a family out. The Poppy documents need the agent's site set
in the Dashboard, since Poppy identifies a business by its domain; without it the
command writes the card and says so. parlant dashboard agent view shows the
agent's POPPY ADDRESS, and parlant widget init points at the new command.
parlant dashboard agent card still prints the A2A card alone for older
instructions.
- The chat widget's onPathChanged(path, notifyAgent) tells the agent which page
the visitor is on. It's called when the widget loads and when the path changes
without a reload; notifyAgent(path, data?) adds a custom event
({ type: "pageview", page: path, ...data }, source customerui) that doesn't start
a reply. A new conversation gets it before anything else, so the agent can tell
a question asked on your pricing page from one asked on your home page.
Changed
- At high agent effort, pruning previously discovered rules now defaults to
low model reasoning, matching rule ranking. Explicit model-route reasoning
settings still take precedence.
- When one policy entails another, the entailed policy is now judged with the
policies that entail it in view. This helps when its wording relies on
theirs, such as "the add-on" that another policy defines. It still matches on
its own merits, and entailment works as before: when an entailing policy
matches, the entailed one is activated too.
- Guests and signed-in customers may add custom events to their own sessions,
with source customer_ui, up to 4 KB of data and 30 a minute. Until now only
the backend could, so a site had to relay page views through its own server
with the admin key. Other sources still need the admin key; a custom permission
gate that doesn't grant CREATECUSTOMEVENT keeps refusing them.
PARLANTCUSTOMERUIEVENTMAX_BYTES changes the size limit (4096 by default).
- A chat widget with a video avatar now moves the video into a circle beside
the agent's name once the conversation scrolls, and the space it leaves
closes up. Scrolling up past the top brings it back. The header keeps its
height and fades out over the messages below the circle. The name and
subtitle fade out while the video moves and back in beside the circle, the
subtitle under the name. A photo avatar is now the same size as the docked
circle, and its header fades out over the messages below it too. Scrolling up
in the widget no longer gets pulled back down in a short conversation, and
sending a message brings the view to it.
On phones the widget only shows the circle, which is cropped onto the face
by the avatar's launcher setting. Your project keeps the copy
parlant widget init made. To pick this up, run
parlant widget init --force, which replaces that copy and any changes you
made to it.
- With Gemini as your provider, replies at medium, high and max effort and both
review checks now run on Gemini 4 Hydrogen instead of Gemini 3.8 Flash. Medium
effort keeps low reasoning. On a website agent this cut replies from 15–30
seconds to 6–8, and Hydrogen kept its reasoning out of replies where 3.8 Flash
at low reasoning had written it into one. Minimal and low effort stay on
Flash-Lite. A react or review.* route you set yourself is unchanged. To
keep the old model, route it explicitly, e.g.
"react": p.ModelConfig("gemini:gemini-3.8-flash", reasoning="low").
- LiteLLM's default for both review checks is now gemini/gemini-4-hydrogen
instead of gemini/gemini-3.8-flash.
Fixed
- An observation with a title but no note no longer appears in the agent's
instructions as an empty entry (its title followed by "Details: [None]",
without its condition). A rule now reaches the agent's prompt only when it
says something: its content (an observation's note, a policy's action), or
the tools it makes relevant, in which case its condition is shown even when
it has a title. Rules without content are also left out of the note that
tells the agent which instructions lost priority this turn. Observations
without a note keep working for relationships, canned-response scoping and
matching; they simply cost no prompt tokens.
- Logging scopes now unwind after exceptions and cancellation, so later logs
in the same task do not retain stale component labels.
- The chat widget centers its disclaimer and "Powered by Parlant" under the input on phones.
- The chat widget no longer shows "↓ New messages" in an empty or short chat that the
visitor scrolled up in. The pill now appears only when there is something below to jump to.
- The chat widget no longer ends a finished reply with a stray *. A reply with
an odd number of * bullets used to get one after its last word. Half-typed
bold and italics still close while a reply streams. To pick up this and the
other widget fixes, run parlant widget init --force, which replaces your
project's copy of the widget and any changes you made to it.
- A reasoning level set on a Gemini 4 model in a schematic route, such as
"review.*": p.ModelConfig("gemini:gemini-4-hydrogen", reasoning="high"), is now
sent to Gemini. It used to be dropped, so the model thought at its own default.
- Canned-response relevance now considers matches from every query chunk before
selecting the closest responses, retaining each response's best matching score.
v4.0.52026-11-12
[4.0.5] - 2026-11-12
Added
- Gemini 4 Hydrogen works through the Gemini provider, including streamed replies,
tool calls, saved conversation replay and token accounting. Select it as
gemini:gemini-4-hydrogen; existing default models are unchanged.
- The chat widget shows Markdown tables in an agent's reply as tables. They
used to show as lines of pipes. Column alignment, inline Markdown in cells
and escaped pipes work. A streamed table appears once its header and divider
rows arrive and fills in cell by cell. A wide table scrolls sideways inside
the message, and your site's own table styles don't change it. Your project
keeps the copy parlant widget init made. To pick this up, run
parlant widget init --force, which replaces that copy and any changes you
made to it.
- Event streams (sse=true) now carry an id on every frame and resume from
it. A client that reconnects with Last-Event-ID picks up after the frame it
last saw, instead of replaying the session from the beginning. Streams ask a
reconnecting browser to wait three seconds (retry: 3000), and an idle
stream sends a keep-alive comment every 15 seconds, so a proxy no longer
closes a quiet connection before waitfordata expires. When a request
carries both min_offset and Last-Event-ID, the later of the two wins: a
reconnecting client usually still has the original min_offset in its URL,
and honoring it would replay the session. The delay and the keep-alive
interval can be set with PARLANTSSERETRY_MS and
PARLANTSSEHEARTBEAT_SECONDS; 0 disables the keep-alive, and 0 for the
delay omits the retry: line.
- Gemini serves Google's Gemini Decision 1 (gemini-decision-1), which answers
many yes/no, choice and scale questions about one conversation in a single
request. A {"*": "gemini"} route now uses it for rule ranking at the
efforts that rank rules in one batch, as OpenAI and OpenRouter routes already
do with their decision models. Name it on a route with
"matching.rank[capability=decision]": "gemini:gemini-decision-1". It uses
your existing GEMINIAPIKEY. Gemini reports no cost for it, so add a
gemini-decision-1 entry to server.set_pricing(...) (Google lists $0.10 per
million input tokens and nothing for output) to have it priced.
Changed
- Gemini models count tokens on your server instead of asking the Gemini API,
which took about 300ms per count. A step after a tool call made two such
calls in a row while choosing canned responses, so replies after a tool call
now start about 0.4–0.6s sooner. Counts use the open Gemma 3 tokenizer and
are within a token or two of Google's. The gemini extra now installs
sentencepiece. The tokenizer downloads once from raw.githubusercontent.com,
in the background, on the first count. Until it's ready, or if your server
can't reach GitHub, Parlant counts with the API as before and logs one warning
saying so.
- Agents no longer repeat what the customer just said back to them before
replying ("Lead generation for a logistics startup, so…"). A reaction now
responds to what the message means, and when the customer is answering the
agent's own question, the agent goes straight to its next step. The prompt no
longer bans stock acknowledgments such as "makes sense" by name.
- Agents created by parlant init no longer learn to repeat the visitor's words
back. The warm-host and upbeat personalities now react to what a message
means, and the warm host goes straight to its next step when the visitor is
answering its question. Existing projects keep their own text; copy the new
wording from a fresh parlant init if your agent echoes.
Fixed
- On managed (Dashboard-hosted) servers, session.metadata.get(), .set()
and .delete(), session.update() and server.get_session() failed with
"Session not found" for live conversations, because they looked in the store
that holds startup declarations instead of the one that holds sessions. They
now use the same session store as the engine and the API, as
session.labels already did, so a retriever or tool that reads or writes
session metadata works the same hosted as it does locally.
- Pricing checks no longer warn about unused provider-default models when a
more specific model route overrides them, such as Gemini embeddings on an
Anthropic server. Models still reachable through other routes, and models
actually observed in use, remain included in pricing coverage.
- MCP tools that return structured results, such as a pydantic model, now give
the agent the result as plain JSON. Before, the call ran but the agent was
told "Tool call error", so it could retry a call that had already taken
effect or tell the customer it failed.
- Gemini agents under concurrent load no longer queue behind each other while
a prompt cache is being created. Requests for other caches, and requests whose
cache already exists, go ahead at once; requests for the same cache still
share one creation, so it is created and its storage charged once.
- Servers with MongoDB stores start faster as session and event history grows.
When a store's recorded version already matches the running one, startup no
longer reads every document back through its loader. When a migration is
needed, documents are loaded and migrated as before.
v4.0.42026-11-08
[4.0.4] - 2026-11-08
Added
- Agent.queue_response() combines background follow-ups with other ready results or a customer turn. Pass after=coroutine to register work before it starts; only the deferred reply waits for the registered batch. Coordination lives in one process and requires session routing affinity; pending work is not recovered after a restart. Queued replies tolerate history edits and terminate interrupted streams without appending stale text.
Changed
- When tool preambles are encouraged, the agent writes them, and the progress updates that follow, in the customer's language.
- Agents acknowledge recurring information gaps when different questions cannot be answered, without treating a new question as one already answered.
- Agents ask which record or request the customer means when it's ambiguous, and which of several unfinished requests to handle first, before collecting task details.
Fixed
- Gemini conversations can resume from background results after history is cleared, without an invalid tool-response ordering error.
- The Parlant dashboard now shows your agent's reasoning under Agent activity when the agent streams its messages, the default for create_agent. Each step's reasoning is now recorded as a loop.reasoning trace event. Before, it reached the dashboard only as a log line, which the dashboard read only from agents in block output mode.
parlant[anthropic] works with the current Anthropic SDK. A fresh install picked anthropic 1.0 or later, whose client refused Parlant's request timeout, so no Claude model could start. The extra now requires anthropic>=1.0.0; if your project capped anthropic below 1.0 to work around this, drop the cap.
- The chat widget lays out a right-to-left reply, such as Hebrew, right to left even when it opens with an English word, a URL or code. It used to take the direction from the first letter, so such a reply read left to right with its punctuation on the wrong side. Each paragraph, the message box and the suggestion buttons now follow the script most of their words are in.
[4.0.3] - 2026-11-05
Changed
- parlant dashboard commands now tell the Parlant dashboard which command sent each request. Every request carries User-Agent: parlant-cli/<version>, the command's name in X-Parlant-CLI-Command (for example parlant dashboard session list, never its arguments or option values), and an X-Parlant-CLI-Invocation id shared by every request from one run. The dashboard uses them to show your CLI activity as the commands you ran.
v4.0.22026-11-04
[4.0.2] - 2026-11-05
Added
- The chat widget lets the AI agent in a visitor's browser talk to your agent
over WebMCP (document.modelContext; in Chrome behind a flag today). On
every page it's on, the widget registers an ask_agent tool described with
the agent's name. A call starts the widget's conversation if there is none
yet, as one an agent holds for the customer (customerrepresentedbyanagent),
sends the question through the chat, so it shows there with the reply, opens
the chat, and answers the browser agent once the turn has ended. Follow-up
calls continue the same conversation. The message carries
metadata.parlantwidget.askedby = "customer_agent", so the chat and your
tools can tell the visitor's agent from the visitor. Nothing happens in a
browser without WebMCP; webmcp={false} turns it off. ParlantTransport's
sendMessage takes an optional metadata argument and ensureSession an
option for the flag.
- Server(sessionstore=..., customerstore=...) accepts an s3://bucket
URL to keep sessions and customers in Amazon S3. Give Parlant a bucket of
its own. Credentials come from the AWS credential chain; add ?region= or,
for S3-compatible storage, ?endpoint=. The storage must support
conditional writes, which keep concurrent servers from overwriting each
other's updates. Empty collections survive restarts, and a failed migration
write leaves the original data in place. Any query other than a lookup by ID,
such as listing sessions, reads every object in the collection, so use
MongoDB if you need indexed queries at scale.
- The chat widget (parlant widget init) takes the agent's picture in one
avatar prop: { type: "logo", src }, { type: "photo", src, live? } or
{ type: "video", src?, poster?, live?, launcher? }. A video avatar plays a
short muted loop under the header, scrolls away with the conversation once
it starts, and keeps its opened size when the chat is expanded. On the closed
launcher, launcher crops it onto the face. Leave out src to use the
widget's own video of an agent. live adds the green dot, on the video's
corner too, red while the server can't be reached. Use an H.264 MP4 saved
with -movflags +faststart; iPhone Safari stalls on WebM.
Changed
- The chat widget's persistent suggestions stand out from its opening ones:
outlined and lettered in the theme's primary color on the chat's background,
instead of the grey fill.
- On phones and tablets, the chat widget closes the on-screen keyboard once a
message is sent, whether typed or picked from the suggestions, so the reply
isn't hidden behind it.
- With the LiteLLM provider, review routes now default to
gemini/gemini-3.8-flash instead of gemini/gemini-3.6-flash, the same
model the Gemini provider already uses for them.
- The lead-gen scaffold's agent no longer guesses when a visitor asks about
the business, its products or their project and its product knowledge has no
answer. It says it's a great question but hard to answer properly in the
chat, offers to connect them with the team, asks for an email or phone
number, and passes the question on with requestteamfollow_up. Unrelated
questions, existing customers' order questions and prices are handled as
before. To make that possible, record_lead now accepts a lead with just a
goal and a contact: buying_for and role are optional, and an unknown one
scores nothing.
- parlant dashboard project confirm finishes a local or self-hosted project
even when its server already holds sessions, such as from trying the agent.
It imports them into the dashboard first, as the dashboard's Sync does, and
then completes the setup, so the project no longer stays at "Finish setup".
An import that takes longer than about 90 seconds stops with
SYNCINPROGRESS; run confirm again to keep waiting for the same import.
A failed import stops with SYNC_FAILED and its reason.
- parlant init --dashboard --json and parlant dashboard project create --json
for a local or self-hosted project now list nextSteps, ending with
project confirm. The scaffolded AGENTS.md tells coding agents to confirm
the project once its server is connected.
- After creating a managed project whose setup check found how to run it,
parlant dashboard project create now points to
parlant dashboard --project ID deploy: the first live deploy completes the
setup. The dashboard's setup page is optional for reviewing the settings.
Deprecated
- The chat widget's avatarUrl, avatarStyle and liveIndicator props, in
favor of avatar. They still work when avatar isn't set.
Fixed
- Gemini Argon can replay saved tool calls with their original signed thought
blocks instead of unnecessarily falling back to result-only history.
- Filtering sessions by labels now returns complete pages and working
continuation cursors, with consistent total counts in either sort order.
- MongoDB cursor pagination now preserves $or filters on later pages, so
filtered listings no longer include records outside the requested set.
- Canned responses now retain separate search vectors for their text and
every signal, so signals no longer collide in the relevance index.
Applies when responses are created or updated; existing indexes are not
rebuilt automatically.
- On iPhone, focusing the chat widget's message box no longer zooms the page
in, which pushed the open chat off the right edge of the screen. The box
uses a 16px font on touch screens, the size below which iOS zooms.
- Local and MCP tools now accept comma-separated list arguments without
brackets while preserving quoted commas and empty lists.
- The chat widget's latest message no longer slips under the suggestions and
message box when something takes the transcript's space, such as the
"Connection interrupted" banner appearing just after a send. The transcript
stayed pinned to the bottom only as messages grew, not as its view shrank.
- Invalid context-variable freshness schedules are rejected before saving,
so a failed schedule update leaves the variable's existing data intact.
- A customer message that arrives while the agent is still working on the
previous one now replaces that turn, as intended, even when the engine was in
the middle of rejecting a step: a tool call or message turned down by a
review hook, by the reviewer, or by moderation. Before, the rejection restarted
the step and dropped the pending cancellation, so the agent finished replying
to the first message and then answered the second one in a separate turn.
Customers who send two messages in quick succession now get one reply that
takes both into account.
Security
- Raise the minimum python-multipart version to 0.0.30 for both uv
and pip installations, preventing installation of versions affected by
CVE-2026-53539.
[4.0.1] - 2026-11-02
Changed
- parlant init --scaffold lead-gen now creates its agent with
preamble_config=p.PreambleConfiguration.encourage(). Before its first tool
call the agent says one short line, such as "Let me check that", so the
visitor isn't left waiting in silence. Without it, preambles were unmanaged:
whatever the model wrote before a tool call went straight through.
- Every tool in the lead-gen and retail-store scaffolds now has a
narration, such as "Checking which days are free" or "Tracking your
shipment". The chat shows it while the tool runs, instead of the generic
"Running tools".
Fixed
- With the Anthropic provider, indexing no longer fails on Claude Opus 4.7 and
later or any Claude 5 model, including the default routes (Sonnet 5.5 and
Haiku 5.5). Parlant sent a temperature, which those models reject with
"temperature is
v3.3.22026-04-28
[3.3.2] - 2026-04-28
Added
- Add PARLANTTOOLCALL_TIMEOUT environment variable to control the plugin tool-call timeout
Changed
- Include guideline description (in addition to the condition) in matcher prompts
- Render active journeys' descriptions in the canned response draft prompt
- Improve error messages surfaced by EmcieService to be more diagnostic
- Relax the aiopenapi3 pin to >=0.8.1,<0.9.0
Fixed
- Fix redundant batch creation in ToolCaller across guideline matches, which in some agents caused a lot of latency unnecessarily
- Fix WebSocketLogger event loop starvation which progressively blocked the async event loop, causing increasing latency over time
- Fix p.Server.current becoming unavailable in the resolvetool plugin endpoint under Uvicorn 0.39+. Context vars are now restored at the start of resolvetool, mirroring the existing pattern in call_tool, so choice providers depending on p.Server.current work again
- Fix MCP tool clients getting stuck after sessions go stale. Clients now reconnect automatically instead of failing subsequent calls
Security
- Upgrade dependencies to address known CVEs: authlib (>=1.6.11), requests (>=2.33.0), fastmcp (>=3.2.0), litellm (>=1.83.0), pytest (>=9.0.3), pyjwt (>=2.11.1), and constrain transitive deps — aiohttp, cryptography, pillow, pyopenssl, werkzeug, Mako, pyasn1, python-multipart, orjson, Pygments, diskcache
- Upgrade chat frontend: vite (>=7.3.2) and override transitive deps: picomatch, lodash, flatted, brace-expansion, immutable, yaml
v3.3.12026-04-14
[3.3.1] - 2026-04-14
Added
- Allow passing ToolId when attaching tools throughout the SDK
- Add AnyOf(tag) and AllOf(tag) modifiers for explicit control over tag dependency semantics in depend_on() — AnyOf requires at least one tagged member to be active, AllOf requires all of them (bare Tag defaults to AllOf)
- Add dependonany() to Guideline, Tag, and Journey for OR dependency relationships — at least one target must be active. Multiple dependonany() calls create independent OR groups that are AND'd together
- Add event loop health monitoring to /healthz endpoint — measures callback latency and reports healthy, degraded, or unhealthy status with peak latency over a configurable window
- Add resolution tracking to the relational resolver — every entity that enters resolution gets a Resolution with a ResolutionKind (NONE, DEPRIORITIZED, UNMETDEPENDENCYALL, UNMETDEPENDENCYANY, ENTAILED) and structured ResolutionDetails (relationship ID, target IDs) explaining why
Changed
- Split RelationshipEntityKind.TAG into TAGALL and TAGANY to support explicit tag dependency semantics at the core level (existing TAG entries treated as TAG_ALL for backwards compatibility)
Fixed
- Fix priority and dependency relationships propagating through inactive intermediaries — only direct relationships now affect resolution, consistent with the reinstatement principle from argumentation theory
- Fix entailment recording only the highest-scoring source when multiple guidelines entail the same target — all entailing relationships are now recorded in resolution details
- Fix dep-failed guidelines not recovering when entailment satisfies their dependency target in a later iteration
- Fix SDK startup appearing stuck at 100% after evaluations — add "Applying evaluations" progress bar for the metadata-writing phase
- Fix Variable.get_value() returning None when called from a retriever, caused by retrievers starting before context variables were loaded
- Fix journey tool-state auto-advancing even when the tool did not run
v3.3.02026-03-15
[3.3.0] - 2026-03-15
Added
- Add per-agent planners via Server.create_agent(planner=...), allowing each agent to use a custom Planner implementation
- Accept Tag as a target in dependon(), exclude(), and prioritizeover() on both Guideline and Tag, enabling relationships that target all guidelines sharing a custom tag
- Add Tag.dependon(), Tag.exclude(), and Tag.prioritizeover() methods to the SDK, enabling tag-based dependency and priority relationships with guidelines and journeys
- Support custom TAG as source for DEPENDENCY relationships in the relational resolver
- Add tags parameter to createguideline, createobservation, and create_journey on both Agent and Journey, allowing custom tags to be attached to entities at creation time
- Add Tag.reevaluate_after() method to the SDK, enabling tag-based reevaluation relationships with tools
- Add tag-based reevaluation support in the engine: when a tool fires, all guidelines carrying a tag that has a reevaluation relationship with that tool are now re-evaluated
- Add staged_events to GuidelineMatchingContext in SDK
- Add priority property to guidelines and journeys for priority-based filtering in the relational resolver
- Add transient guidelines (renamed from tool-provided guidelines), allowing tools to dynamically inject behavioral guidelines into the agent's context
- Add Agent.utter() to the SDK, enabling programmatic agent message generation with transient guidelines
- Add Customer.update() and CustomerMetadata to the SDK, allowing tools to update customer name and metadata
- Add Session.update(), SessionMetadata, and SessionLabels to the SDK, allowing tools to update session properties, metadata, and labels
- Add customer, agent, mode, and title properties to SDK Session class
- Add Server.get_tag() to the SDK, supporting lookup by either id or name
- Add name-based filtering to TagStore.list_tags() and the GET /tags API endpoint via an optional name query parameter
- Enforce tag name uniqueness in TagStore, raising an error when creating a tag with a duplicate name
Changed
- Made extended thinking indicator optional in perceived performance policy
- Change reevaluateafter() on Tag and Guideline to accept multiple tools (tools) and return Sequence[Relationship]
- Change tags field type from Sequence[TagId] to Sequence[Tag] on Guideline, Journey, Capability, Term, Variable, Customer, and Agent in the SDK
- Change Tag.preamble() to return a full Tag object instead of a TagId
- Upgrade MCP service and bump dependency versions to resolve security vulnerabilities
Deprecated
- OpenAPI tool services are now deprecated; please migrate to SDK tool services
Fixed
- Fix deadlock when sending a new message right after a preamble
- Fix transitive filtering in relational resolver for custom tag dependency targets (guidelines depending on a custom tag are now correctly deactivated when a tagged member is deprioritized)
- Fix SSE read_event endpoint stalling after first streaming chunk until full completion
- Fix response analysis logs not always reaching the integrated UI
- Fix guideline formatting in canned response and streaming modes when condition is absent
- Fix AzureService small text embedding dimension size
- Fix onnxruntime compatibility with Python 3.10 and transformers 5.x type changes
- Fix agent intention proposer prompt clarification
- Fix embedding LRU cache eviction corrupting the length index when entries share the same text length
- Fix LiteLLMEmbedder failing to resolve via lagom container when LITELLMEMBEDDINGMODEL_NAME is set
- Fix non-consequential tool calls being rejected when optional parameters are missing
Removed
- Remove stale parlant-test entry point and testing framework documentation from README
v3.2.12026-02-17
[3.2.1] - 2026-02-17
Added
- Add optional dependencies parameter to guideline, observation, and journey creation methods
- Add exclude() as an alias for prioritize_over() on guidelines and journeys
- Add tools parameter to create_observation methods
Changed
- Deprecate attachtool() in favor of createguideline()/create_observation() with tools parameter
Fixed
- Preserve draft message language during canned response recomposition
- Fix server hang when an exception occurs during setup
- Fix canned response field extraction to handle falsy values
v3.2.02026-02-09
[3.2.0] - 2026-02-09
Added
- Add labels to Guidelines, Journeys, JourneyNodes, and Sessions for categorization and filtering
- Add automatic session label propagation from matched entities (guidelines, observations, journeys)
- Add track parameter to guidelines to control "previously applied" tracking
- Support multiple targets in prioritizeover() and dependon() methods
- Add field_dependencies to canned responses for explicit field availability requirements
- Add attach_retriever() to Guideline, Journey, and JourneyState for conditional data retrieval
- Add onmatch and onmessage hooks to journeys for lifecycle callbacks
- Add per-agent preamble configuration (custom examples and instructions)
- Add separate default greeting responses for first agent message in fluid mode
- Add streaming message output mode
- Allow specifying custom journey node ID
- Add matched guidelines/journey states to completion ready event
Changed
- Make condition optional for SDK guidelines
- Tweak default preamble examples
- Soften log levels for relational guideline resolver
- Add activated/skipped logs to custom guideline matcher batches
Fixed
- Fix websocket warning upon startup
- Fix agent intention proposer (guidelines were getting rewritten incorrectly)
- Fix multiple customer guideline matchers not working
- Fix bug with context variable access in SDK
v3.1.02026-01-05
[3.1.0] - 2026-01-05
Added
- Add .current property for Server, Agent, and Customer in SDK
- Add /healthz endpoint
- Add API for CRUD operations on session metadata
- Add EmcieService
- Add GLM service
- Add Mistral service
- Add OpenRouter service
- Add OpenTelemetry integration for Meter, Logger and Tracer
- Add Qdrant VectorDatabase adapter
- Add Snowflake Cortex service
- Add ability to configure and extend the FastAPI app object
- Add deferred retrievers
- Add dynamic composition mode
- Add follow-up canned responses
- Add guideline criticality level
- Add guideline on_match() hooks
- Add persistence option for context variable values (variable store)
- Added guideline descriptions
- Allow bailing out of canned response selection and utilize the draft directly, using a hook
- Allow controlling max tool result payload via environment variable
- Allow controlling perceived performance policy per agent
- Allow journey transitions from one tool state to another
- Allow specifying custom IDs when creating agents via SDK and API
- Allow specifying custom IDs when creating customers via SDK and API
- Allow specifying custom IDs when creating guidelines, journeys, and glossary terms via SDK and API
- Expose IoC container in server object
- Support adding custom canrep fields to matched guidelines and journey states
- Support code-based, custom guideline matchers
Changed
- Changed default NLPService to EmcieService
- Improved efficiency of journey state matching when first state is a tool state
- Rename ContextualCorrelator to Tracer
- Rename LoadedContext to EngineContext
- Support proxy URL for LiteLLM
Fixed
- Fix critical bug with cancellation during response analysis
- Fix critical similarity calculation error in TransientVectorDatabase
- Fix unnecessary extra evaluation of journeys and tools in some edge cases
- Improved Gemini Flash 2.5 output consistency by using function call trick instead of structured outputs