Skip to content

STELLA v2.0.0

Choose a tag to compare

@DatMoser DatMoser released this 10 Nov 19:55
· 25 commits to main since this release
36dc851

The agent platform. Anyone can now write an agent with the Agent SDK, describe
what it needs in its agent.yaml, upload it and run it in a session. STELLA
holds no code for any one agent. Agents can use outside tools through
connectors, which each user sets up in Settings, and the credentials never
reach the agent. Participants can hold a whole conversation in writing on the
new Text face. STELLA Jev joins as a third built-in agent, and the ELIAS
interviewer is the first uploaded one. If you upgrade, read "Upgrade notes"
first: agents uploaded before this release need one rebuild.

Upgrade notes

  • Back up the database before you upgrade. This release applies seven database migrations on its own during the deploy (connectors, connector icons and descriptions, the agent manifest, the agent's SDK version, sharing uploaded agents with projects with a backfill, and a column marking which configuration templates have been checked). No template data is changed, and nothing backs up for you first
  • Rebuild every uploaded agent once after upgrading. STELLA now records which SDK an agent's image contains and refuses to deploy an image whose SDK it doesn't know, so agents built before this release can't be deployed until they are rebuilt from Settings → Agent Library. Agent SDK 0.10.0 is on PyPI before this release reaches production, so the rebuild works straight away. On production this includes ELIAS. Built-in agents are unaffected
  • Uploaded agents need Agent SDK 0.10.0 (MIN_SUPPORTED_SDK_VERSION is 0.10.0). Built-in agents are built with it from the repository
  • A deploy without a required env var is now refused up front with a 400 that names it, instead of starting a pod that then fails. Check that each built-in agent's required keys (for example OPENAI_API_KEY) are set
  • "Review uploaded agents" is on after the upgrade (Settings → Platform Defaults), so a system admin approves each upload before it can run
  • No new required setting for the platform itself

Added

  • ELIAS-Interviewer, the first uploaded agent (custom-agents/elias-interviewer, Agents → ELIAS-Interviewer). It runs an ELIAS interview guide in a voice session: it asks every question word for word, asks an unclear answer again up to three times, and writes each turn to ELIAS as it happens. It reaches ELIAS only through a connector, so the ELIAS token never reaches the agent. It is uploaded and deployed like any custom agent

  • Uploaded agents can be deployed into a session. Before, any agent type other than the built-in STELLA V2 and STELLA Light was turned away at deploy time. Now an uploaded agent deploys once a system admin approves it, and its pod runs the build's own start command. A rejected agent never deploys, including on restart or wake after auto-pause.

  • Admin switch: "Review uploaded agents" in Settings → Platform Defaults. It is on by default, because building an upload runs its Dockerfile on the server and the agent's code runs in sessions unreviewed. A system admin can turn it off; then a user can deploy their own upload as soon as it is built, and other users still need an admin's approval.

  • Uploaded agents belong to their owner and are shared per project. In Settings → Agent Library the owner picks the projects an upload is shared with; members of those projects see it marked "Shared by …" and can deploy it there only. Only the owner can build it, upload a new version, change its sharing or delete it. The upgrade shares each existing upload with the projects that use it as their public agent or have it running or paused; every other upload starts private to its owner

  • A new version replaces your agent in place. Uploading the same metadata.slug with a higher metadata.version keeps its env var templates, configurations, shares and public projects pointing at it; the new version must be built (and approved, if review is on) before it deploys

  • Delete an upload from the Agent Library. It is refused while the agent still runs in another user's session or a public project uses it, and the message says how many

  • Participants and researchers see when STELLA is still getting ready to listen. After a quiet spell, speech recognition can take up to about 30 seconds to warm up at the start of a session. The participant now sees "Getting ready to listen…" (in German or English), and the researcher sees an "STT warming up" chip with a timer. After 30 seconds it says it is taking longer than usual, and if warming up fails the participant is asked to tell the study team. Nothing shows when the model is already warm (#564)

  • STELLA Jev (stella-jev-agent), a third built-in agent (#849). STELLA V2's pipeline without the bridge: one Jev call decides the experts and barge-in, and the agent asks back when what it heard isn't believable (the common-sense expert). The reply no longer waits for task extraction, which saves the answer in the background. Take-back: if someone carries on speaking after their turn was closed and before any of the reply was heard, the reply is dropped and the joined-up words are answered, and nothing from the dropped reply is saved (take_back_window_ms on the new Turn Taking node, 0 is off). It sends conversation text to TypeSafe and needs TYPESAFE_API_KEY. It is seeded and built from agents/stella-jev-agent/ like the other built-ins, with no agent-specific code in STELLA.

  • Connectors in Settings as a grid of cards, like personas and plans. Each connector has an emoji and a description (new nullable Connector.icon and Connector.description columns, also on the MCP connectors_create and connectors_update tools). Its detail page lists the tools from the last test with their parameters, has Try it to call a tool for real with the owner's token (audited like an agent's call), and shows the connection, sharing and recent calls. New endpoints: POST /me/connectors/:id/tools/:tool/call and GET /me/connectors/:id/calls, both owner-only. See Agent SDK → Connectors.

  • Text face for participants (#92). Text is one of the faces in the participant's face gallery: the whole conversation in writing. Replies show as the agent sends them: streamed if it streams, whole if it doesn't. While it shows, the agent answers in text only, speaks nothing and does not wait for speech, and the microphone is off. Picking any other face switches back to voice, and the microphone comes back if the switch took it. Speech-to-text and text-to-speech stay connected, so switching is instant. An invitation locked to the Text face is a text-only session: the join page skips the microphone setup. The chat drawer on the voice faces stays as a fallback for typing next to speech. See Guides → Text sessions.

  • A declared channel between the screen and the agent. agent.yaml lists the events the screen may send (client.events) and the commands the agent may send (client.commands), with a JSON Schema for each one's data. Anything else is refused. The screen can ask for an answer and gets the agent's reply. Every event, reply and command is kept in the session history. See Agent SDK → The platform contract.

  • Agent manifest 2.0 (agent.yaml). An agent now declares what it needs in its own blocks: features (persona, plans and their modes, configurator), env (required and optional env vars) and connectors. Anything a deployer chooses is an env variable; configSchema only describes the agent's fixed defaultConfig. Manifest 1.0 still works and is upgraded on read, with a deprecation warning. A published JSON Schema (docs-site/static/schemas/agent-manifest.v2.json) gives editors completion and validation. See Agent SDK → The agent manifest.

  • Deploy enforces what an agent's manifest declares. One check runs on every deploy, restart and wake after auto-pause, and refuses the deployment with a 400 that lists every problem:

    • A required env var is missing or empty.
    • The agent runs plans and has neither a plan nor a companion session.
    • A plan, companion session or Configurator configuration is given to an agent that doesn't declare it.
    • The agent needs a configuration (features.configurator: { required: true }) and has none.
    • The agent needs a newer SDK than this STELLA ships.
    • A deployment tries to set agent settings beyond the platform keys (persona, mode, plan, configuration); those are env vars now.
    • An env var value doesn't fit its declared type, range or options.

    Before, only a stella-v2 rule hard-coded in the backend was checked. The agent's defaultConfig now reaches the pod, where it used to be dropped. Pods get the manifest's resource requests as well as its limits.

  • GET /agent-types returns each type's requirements (persona, plans and modes, configurator, env vars, connectors, settings schema and defaults, resources, SDK minimum), so the deploy wizard can stop reading x-stella-* flags. GET /agent-types/sdk-support returns the SDK versions this STELLA runs (minSupported, current).

  • The build records the SDK an image really has. After each build of an uploaded agent, STELLA reads the installed stella-ai-agent-sdk version from the image (new AgentType.sdkVersion column). A build with no SDK, or one below the supported minimum, fails. Deploy, restart and wake refuse an upload whose build is outside the support window. My Agents returns sdkProblem for those. Uploads built before this have no recorded version and need one rebuild.

  • image.command in the manifest sets a pod's start command. Without it the image's own CMD runs.

  • Agent platform standards (Contributing → Coding standards → Agent platform): the rules that keep STELLA free of agent-specific code, and which module owns which concern.

  • MCP: platform_settings_get and platform_settings_update for system admins. An assistant can read the platform switches and turn "Review uploaded agents" on or off. Nobody else sees the two tools, and a call by name answers 403. The MCP switch and the TTS playback default stay UI-only.

  • Settings → Agent Review, where system admins approve or reject uploaded agents (#897). Before, approving an upload was only possible through an API call, so shared uploads stayed stuck. The page has Pending, Approved and Rejected tabs. Each upload shows its latest build first, then the owner, version, upload date, SDK and what its agent.yaml declares. Approve takes optional notes and Reject needs a reason, and the uploader sees either one on their agent card. A decision can be reversed. The same actions are on MCP for system admins (agent_reviews_list, _get, _approve, _reject)

  • An upload, or uploading a new version, builds on its own. Before, a build only started when the owner pressed Build after the upload finished. The button is now Rebuild. Only one build of an agent type runs at a time: a second upload while one is already building queues behind it, and a further upload while one is queued supersedes it (shown as "Superseded by a newer upload"), so only the latest upload is ever built next

  • Build status and log, live, on the agent card (Settings → Agent Library and the deploy wizard's "My Agents"). Queued, Building, Built or Build Failed shows on the card itself and updates on its own as the build runs, with a link to the log — no manual refresh needed

Fixed

  • An agent that has been stopped no longer reports its last health as ready over MCP (#807)
  • Dependency updates from Dependabot for npm, pip and the Docker base images (#826)
  • STELLA V2 no longer cuts off its bridge when a safety rule replaces the reply. The bridge sentence stayed on screen and was spoken to the end, and the fixed safety reply follows it at the same speaking speed (#827)
  • Over MCP, sessions_get and sessions_list count only what was said in messageCount, the same lines session_transcript_get returns, not internal bookkeeping rows (#808)
  • The MCP agent-configuration tools point clients to agent_types_get for an agent's settings schema, which agent_types_list no longer returns (#820)
  • Deploys remove untagged image digests that K3s keeps after each tag moves, one at a time and never with --prune, so the image disk stops filling up (#799)
  • An agent that speaks first started as soon as an organiser opened the session page. The SDK's wait_for_participant() counted anyone who is not an agent. It now returns when a session participant joins, or when an organiser writes or speaks (so organisers can still test it), and never just because someone is watching. Needs stella-ai-agent-sdk 0.10.0.
  • Agent cards in the deploy dialogs spilled their tags and deprecation note outside the card. Cards now grow to fit.
  • Chat history shown twice on the participant screen. Opening a chat after the agent had spoken showed the greeting and every message so far twice; the stored history is now matched to what the screen already shows.
  • A restart or wake that STELLA refuses (agent type rejected, under review, missing a required env var, or built with an unsupported SDK) now leaves the running agent alone. Before, the pod was deleted first and the agent was left in STARTING with no pod.
  • An agent whose manifest sets a memory or CPU limit below 512Mi / 250m now starts. Pod requests were fixed at those values, so Kubernetes refused the pod (request above limit) and the agent went FAILED without a pod. Requests are now lowered to the limit when the limit is smaller.
  • An uploaded agent rejected in review was labelled "Failed" in the Agent Library. It now reads "Rejected"
  • Live build logs never reached the browser. The backend's log stream resolved each poll's status but sent the unresolved promise as the event instead, so every update went out with no data; separately, the browser only listened for the stream's default message event, but the backend sends each update as a named build event, which that listener never receives. Both are fixed — live status and log streaming now work everywhere they're used, including the existing build-log viewer
  • An agent built in Settings still showed as not deployable in the deploy wizard. The wizard kept the "My Agents" list and the review flag it last loaded; it now reloads both when it opens and after a build finishes

Changed

  • Agent settings in tabs. Deploying an agent and editing a configuration template show each setting the agent declares as its own control (switch, slider, list, number or text), in tabs by topic, with anything undeclared under "Custom". A deploy starts from the agent's defaults or a template, shows what changed, and can save the result as a new template. The agents' manifests now give every optional setting a type, label and tab, declare BARGE_IN_DUCK_GAIN, and state the right default for BARGE_IN_MIN_SPEECH_MS (600 ms) (#585, #548)
  • Agent settings take two screens in both wizards. Deploying an agent into a session and setting up a public project first ask where the settings start from (the agent's defaults or one of your templates), then show the settings in tabs, which always open on Keys. Back keeps what you changed unless you pick a different template. An agent with no saved templates goes straight to the tabs. The public setup now uses the same tabs instead of a plain list of variables (#892)
  • The deploy wizard's Connectors step is one list to tick, with no categories. STELLA fills the agent's declared slots from the ticked connectors (those covering a slot's tools first, matched so as many slots as possible are covered, then untested ones; never one connector for two slots) and binds the rest as extras.
  • The device check appears only when voice needs it. A participant who joins on the Text face, locked or not, skips it. Switching from Text to a voice face opens it in the session, unless the browser already allows the microphone, with Stay in writing to go back. Where a browser cannot report the microphone permission, a microphone the page can already see counts as allowed.
  • The participant's end screens no longer show a code. The completion code and the time-limit failure code (from the Prolific study flow) are gone; both screens thank the participant and say they can close the window. The Session complete screen also waits until the agent's last message has been heard and read (3 to 10 s by length, 25 s at most) instead of covering it at once.
  • One look for the wizards. Selectable cards in the deploy, project and invite wizards share one style (one outline that does not shift the card, one check badge, no stray violet or blue), and every wizard uses the same step bar, black in light mode and purple in dark mode. The Text face icon has purple bubbles. Settings pages have room at the bottom.
  • The deploy wizards ("Deploy agent" and the public project setup) are sized by the screen, not by the step: the same box on every step, a bit larger on full HD and 4K, with the step scrolling inside it.
  • Agent Library: an agent someone else shared now says which of your projects you can deploy it in ("Shared by Olga · in Study A"). GET /agent-types returns sharedInProjects for it.
  • The deploy dialogs no longer upload agent packages. Upload in Settings → Agent Library; the dialogs' Custom section links there.
  • Built-in and uploaded agents are stored by one mapping. Before, the seed and the upload endpoint each turned agent.yaml into an AgentType their own way. Each had its own expert reader, and companion experts were handled differently in the two. Now both use one mapping, and each AgentType keeps its full manifest (new manifest column; the migration adds it and the next seed or re-upload fills it).
  • Manifest mistakes are refused instead of ignored. Unknown keys, unknown capabilities, a resource request above its limit, and a defaultConfig that doesn't fit configSchema now fail validation. Before, these were dropped or let through silently. Manifest 1.0 files keep the old leniency for unknown capabilities (a warning).
  • One default agent type. Every path that deploys without naming a type now asks one place (DEFAULT_AGENT_TYPE, default stella-v2-agent). The deploy scripts defaulted to stella-light-agent. A spec fails the build when an agent slug appears in platform code.
  • Built-in agent images come from their manifests. The hard-coded image registry and the python -m <slug> start command convention are gone. A built-in is built from agents/<slug>/ with its manifest's Dockerfile, and starts with its manifest's image.command or the image's CMD. The deploy scripts find agents by their agent.yaml. The seed refuses a built-in whose folder isn't named after its slug.
  • Removed AgentsService.createStandaloneAgent and updateGrpcAddress. Nothing called them, and the first skipped deploy validation.
  • A pod whose agent type has no valid resource limit now gets the manifest default (512Mi, 250m) instead of a separate 2Gi / 2000m fallback.
  • The agent card opens on Build. Opening an uploaded agent in Settings → Agent Library now shows its build status and log first. The Overview tab is renamed Capabilities, since that is what it lists, and the button reads Rebuild. In the deploy wizard, a selected uploaded agent gets the same outline as a built-in one, and its build badge opens the log
  • One Agent SDK section in the docs. The SDK pages were in two overlapping folders and showed code that could not run: a stella_sdk module, an AudioPipeline the agent built itself, and classes the SDK never had (VADConfig, @tool, TodoItem, the message classes). Agent SDK now holds them all, rewritten against SDK 0.10.0, and CI checks that every example imports against the SDK. Old /docs/sdk/... links no longer work; the pages are under /docs/agent-sdk/... (#887)

Security

  • Connectors can't reach STELLA's own network. Connector URLs that resolve to loopback, private, link-local, CGNAT, multicast or cluster-internal addresses are refused, in every IPv6 spelling too (IPv4-mapped, NAT64, 6to4). The address is checked again at each connection's own DNS lookup, so a host that answers publicly first and privately later is still refused. CONNECTORS_ALLOW_PRIVATE_HOSTS=true allows them for local development only
  • An agent's connector token works only while that agent runs. Each pod gets a fresh token when it is created, restarted or woken, and the backend keeps only its hash. It stops working when the agent stops or is paused or its session closes, so a token copied out of a pod is useless afterwards
  • Approving or rejecting an uploaded agent (/admin/agent-types) now needs a system admin. Before, any signed-in user could approve their own upload.
  • Uploads now record who uploaded them. Before, every upload was stored without an owner, so My Agents listed everyone's uploads and anyone could build or delete any upload. Now only the uploader (or a system admin) can build or delete an upload, and My Agents shows only your own. Uploads stored before this fix have no owner and only a system admin can manage them.
  • An uploaded agent can't point at a prebuilt image (image.imageUrl) any more; STELLA builds it from the package's Dockerfile. Before, such an upload counted as built straight away and skipped the base-image check.
  • An uploaded configSchema may not use regular expressions (pattern, patternProperties, propertyNames patterns), wherever they sit, including behind a $ref or in draft-7 dependencies: STELLA's schema checker refuses to compile any expression. STELLA now checks settings against the schema on upload and deploy, and a backtracking expression could freeze the server (ReDoS). A schema Ajv can't compile, or one with a $schema it doesn't know, is refused with a 400 instead of failing with a 500.
  • The probe that reads an image's SDK version runs without network or capabilities, with capped memory and processes, and its container is removed even when it times out.
  • Building an uploaded agent no longer runs its Dockerfile path, name or version through a shell. The Dockerfile must be a plain relative path inside the package.
  • An uploaded agent no longer receives the platform's LiveKit key and secret. It gets a token for its own session's room and identity only, so it can't join or listen to other sessions. Built-in agents are unchanged. Because only SDK 0.10.0 and later read the token, STELLA now supports SDK 0.10.0 or later: an uploaded agent built with an older SDK is refused at deploy, restart and wake, and My Agents marks it as needing a rebuild (#856)
  • Connectors. Each user can store outside MCP servers (URL + bearer token) in Settings → Connectors, test them, and see their tools. Agents reach a connector only through the backend, so the token never reaches the agent pod. The owner can share a connector with projects they belong to; members' agents there may then use it, without seeing or editing the token, until it is unshared. Agents declare the connectors they need in the manifest's connectors block, and users may offer extra ones; the deploy wizard's Connectors step binds them (connectorBindings on the deploy request). Connectors can also be managed from an AI assistant over Stella's MCP server (connectors_list, _get, _create, _update, _test, connector_shares_set; no delete), and agents_deploy takes connectorBindings. One database migration (Connector, ConnectorProjectShare, ConnectorAuditLog, AgentInstance.connectorBindings/deployedByUserId/connectorTokenHash)