Polyant v1.2.0

Polyant v1.2.0

Polyant v1.2.0 moves product behaviour to the scope that owns it: installation
policies are editable by platform admins, agent behaviour is configurable per
agent, and attachment storage uses each agent's bucket. It makes the core
runtime leaner by moving integration-heavy tool families into build-time
plugins, receives Telegram and Slack through webhooks, and runs on Node 24.
That cleanup is deliberately breaking: the environment variables that set
product behaviour are gone, extracted tools need their plugins installed and
take namespaced names, storing inbound attachments becomes a per-agent opt-in
with the agent's own access keys, and a new agent starts with its Web channel
off and requires its API key.

Read docs/UPGRADING.md before starting. Extracted
tools need their plugins, agents that should keep users' files need access
keys and Store attachments turned on, Telegram and Slack need a public
address, scripts that create agents must enable their Web channel, and
non-default values of the removed variables have to be entered in the panel.

Highlights

• A guided first access. A new user gets a guided setup (provider and key,
  model, agent, a test chat in the Playground, channels and the agent key,
  optional knowledge) and then an interface tour; the first visit to an agent's
  page offers a tour of its sections. Each step can be skipped, the Help menu
  reopens them, and the state is stored per user so it follows the person
  across devices. Accounts that exist at the upgrade start with it dismissed.
• Settings live at the right tier. Settings → General holds the
  installation's operational policies: analytics retention, the engine's public
  address, live-stream caps, rate limits, agent-to-agent and MCP timeouts, and
  scheduler deadlines. Each agent can set its timezone, locale, memory
  deduplication and message timing, and knowledge-document caps can vary by
  organization. The shipped defaults are unchanged.
• A smaller core, with build-time plugins. GitHub, Render, HubSpot and
  Markdown-to-PDF tools moved to plugins, so the base image no longer carries
  their optional binaries and libraries. A plugin can declare its Alpine, global
  npm and environment requirements, and the Docker build installs their union.
  fileUpload follows them into the extra plugin. HubSpot lives in
  polyant-hubspot-plugin
  (v1.1.1), Markdown-to-PDF and fileUpload in
  polyant-extra-plugin
  (v1.1.0); no plugin is shipped for GitHub and Render. Manifests can also name the
  plugin and describe the parameters its tools ask for, and the panel shows
  them.
• Telegram and Slack through webhooks. Both channels now arrive like
  WhatsApp, on /webhooks/*, and the Telegram webhook survives rolling
  deploys. A failed Telegram registration is retried instead of disabling the
  channel, redelivered updates and events are dropped, and each bot's webhooks
  are rate limited on their own. The CDK stack routes webhook and HTTP-channel
  traffic to the engine without the panel's sign-in.
• Web chat context. chat/stream and /v1/chat/completions accept a
  context object projected onto the conversation state through a per-agent
  field mapping, so a web or voice front end can hand tools and hooks the
  caller's identity. An agent that maps a context field requires its key on
  every web turn. On both routes, a client that disconnects before the turn
  produced anything aborts it, and a turn that has produced output runs to its
  end and is saved.
• A clearer agent page. Tools and hooks open in side sheets with the
  parameters they declare, conversations can be followed live on any channel,
  activity reads as a timeline and now includes tool calls, and the page says
  what a capability still needs before it works.
• Secure defaults for new agents. Web is a channel: an agent's public HTTP
  endpoints answer only while its Web channel is on, and a new agent starts
  with it off and requires its API key. Existing agents keep their endpoint and
  their authentication setting. The panel's Playground has its own
  session-based route and the agent.playground:use permission, held by every
  built-in role, so it works with Web off and needs no key. A fresh install no
  longer seeds a demo-agent without authentication.
• API keys from the panel. An organization Owner issues and revokes
  management API keys under API keys, so polyant-cli and scripts can call
  the management API of an OSS deployment with X-Polyant-Key. A key holds the
  Owner permission set, can expire, is shown once, and cannot issue or revoke
  keys itself.
• Updated models and gateway. OpenAI gpt-6, Claude Opus 5, Opus 5.5,
  Sonnet 5.5, Haiku 5.5 and Fable 5.1, and new Nebius models are in the
  catalog, with corrected prices. Claude Haiku 5.5, GPT-5.4, GPT-5.6 and GPT-6
  are priced by prompt length, so a call with a long prompt is recorded at the
  provider's higher rates. Providers and embedders can be registered at boot,
  and switching thinking off now really switches it off.
• Bedrock works without Anthropic approval. The default standard tier uses
  Amazon Nova Pro and the heavy tier uses OpenAI gpt-oss 120B. Nova retains
  system-prompt caching while avoiding unsupported cache markers on tool
  messages.
• One bucket per agent, and attachments kept only by choice. Attachment
  persistence and extra:fileUpload share one per-agent S3 configuration,
  set on the agent's Advanced page and reached with the agent's own access
  keys. Each agent decides whether to keep the files its users send, and
  stored files are deleted with their conversation and with the agent.
  Filenames with URL-significant characters round-trip correctly.
• Scheduled work recovers. Runs left behind by a dead process are recovered
  at startup; live runs that exceed their configured deadline are reaped
  without touching a newer claim of the same task, and a scheduler health
  endpoint exposes stuck-run and capacity counts.
• Faster and cleaner on Node 24. Under load the platform's overhead per
  turn drops sharply on Node 24, and the engine now finishes shutting down
  while a panel holds an activity stream open, instead of waiting for the
  platform to kill it and losing the last buffered audit rows.
• Faster lists and dashboards. Conversations carry their own message
  counters, so lists and analytics no longer count message history, and
  analytics reads run on a separate, smaller connection pool.
• Recorded spend matches what the providers bill. An audit against live
  OpenAI, Anthropic and Bedrock calls found per-call tokens and prices exact
  and four places where billed spend went missing or was mispriced: Amazon
  Nova carried the US price instead of the eu-south-1 one, a turn preempted by
  a follow-up message left no row, embedding calls were never logged, and
  long-period totals lost precision in the sum. All four are fixed; Nova costs
  more than 1.1.x reported.

Breaking changes

• The engine and the panel require Node 24. The published images carry it;
  source deployments and custom images must move.
• Extracted tools are namespaced, and fileUpload is now extra:fileUpload in
  the extra plugin. Migrations 0087_rename_extracted_tools and
  0089_rename_extracted_tools_file_upload rename their rows in place, so
  agents keep them enabled and skills keep their links; install the relevant
  plugins before building (HubSpot from polyant-hubspot-plugin v1.1.1,
  Markdown-to-PDF and fileUpload from polyant-extra-plugin v1.1.0). GitHub
  and Render tools are removed and no replacement is shipped. verifyDocument has no replacement and is removed
  from agents at the first boot.
• Telegram and Slack need a public HTTPS address with /webhooks/* routed to
  the engine. Each Slack app needs Socket Mode off and its Request URL set to
  the agent's webhook; the Slack app token is no longer used. Each Telegram
  channel gets its own random webhook secret, which the engine mints and
  registers when it first starts on 1.2.0; Telegram's updates arrive once it
  has.
• A new agent starts with its Web channel off: /v1/chat/completions and
  chat/stream answer 404 until the channel is enabled, which needs the
  agent's API key unless authentication is turned off. Migration
  0091_web_channel switches Web on for every existing agent, and bundles move
  to format 1.3.
• A new agent requires its API key: migration 0090_new_agents_require_auth
  makes auth_enabled default to true. Existing agents and imported bundles
  keep their setting.
• Migration 0091_web_channel deletes the seeded demo-agent when it was never
  used. A demo-agent with conversations, memories, knowledge, scheduled tasks,
  secrets or edits stays.
• AUTH_MODE, DEFAULT_INSTANCE_ID, deployment-wide AWS_REGION,
  PLATFORM_ADMIN_EMAIL, WORKSPACES_ROOT, DEBUG_LLM_PAYLOAD and the unused
  LANGSMITH_* variables are no longer product configuration.
• The environment variables that set agent, organization or installation
  behaviour are read by nothing; the panel holds those values now. Their
  defaults are unchanged, and a deployment that set a different value must enter
  it in the panel before removing the variable. THROTTLE_ENABLED stays an
  environment variable.
• Attachment storage no longer reads the four PLATFORM_S3_* variables. On
  each agent that needs persisted attachments, set the bucket, region and the
  agent's own aws_access_key_id and aws_secret_access_key on its Advanced
  page, and turn on its Store attachments switch: migration 0088 adds it
  off for every agent. Deleting stored files needs s3:DeleteObject on the
  attachments/ prefix. Files earlier versions stored in the platform bucket
  stay there.
• Unpinned OpenAI agents move from gpt-4o-mini/gpt-4o to the gpt-6 tiers,
  and unpinned Bedrock agents to Nova Pro and gpt-oss 120B. The Anthropic
  fast tier moves from Claude Haiku 4.5 to Claude Haiku 5.5, so the background
  jobs of every Anthropic agent, pinned or not, change model and price.
• The images run as the unprivileged node user (uid 1000). A volume mounted
  on the engine's /app/logs or /app/packages/engine/workspaces must be
  writable by that user.

Database migrations

The release migrations, 0078 to 0093, are:

• 0078_scheduler_crash_safety: a per-task run deadline and the index the
  scheduler reaper scans.
• 0079_org_knowledge_doc_cap: an organization's own cap on the knowledge
  documents one agent may hold.
• 0080_env_vars_to_product and 0082_platform_settings_expand: per-agent
  behaviour and installation policy columns that replace deployment variables;
  an unset column keeps the shipped default.
• 0081_drop_s3_endpoint_secret: removes a storage secret that only
  pre-release builds of 1.2.0 wrote; a 1.1.x database has none.
• 0083_agent_capability_defaults: new agents start with memory off and speech
  transcription disabled until configured; existing agents keep their settings.
• 0084_add_trace_cache_write_5m_tokens: records five-minute cache-write tokens
  on traces, so Anthropic cache writes are priced by TTL.
• 0085_add_web_context_field_mapping: the per-agent web context field mapping.
• 0086_conversation_message_counters: per-conversation message counters,
  backfilled from every conversation message.
• 0087_rename_extracted_tools: renames the tools moved into plugins to their
  namespaced names, keeping agent enablement and skill links.
• 0088_attachment_storage_opt_in: the per-agent Store attachments switch,
  off for every agent.
• 0089_rename_extracted_tools_file_upload: renames fileUpload to
  extra:fileUpload, keeping agent enablement and skill links.
• 0090_new_agents_require_auth: auth_enabled defaults to true for new
  agents; existing rows keep their value.
• 0091_web_channel: deletes the seeded demo-agent when it was never used,
  and switches the Web channel on for every remaining agent.
• 0092_playground_use: grants agent.playground:use to every built-in role.
• 0093_user_onboarding: the per-user onboarding state, set to dismissed for
  accounts that already exist.

Operational notes

• Telegram and Slack redelivery checks are kept in memory in each process. With
  several replicas, a retry that reaches another replica is answered again and
  the contact can get two replies. Run one replica, or accept the occasional
  duplicate.

Security

• Plugin artifacts are limited to 10 MB each, 10 minutes of lifetime, 20
  handles or 25 MB per conversation, and 100 MB or 1,000 live handles per
  engine process. A handle is bound to its agent as well as its conversation.
• Tenant scope is required and branded across the affected stores, closing
  cross-tenant keyword search and mutation paths and refusing unresolved request
  tenancy rather than dropping a predicate.
• Disabled skills no longer resolve retained credential placeholders. MCP OAuth
  metadata uses the shared private-network denylist, rotating bearer strings no
  longer evade throttling indefinitely, and provider response bodies are not
  logged raw.
• Management keys are verified against a SHA-256 digest instead of bcrypt,
  which held the event loop for about 250 ms on every authenticated request. A
  bcrypt key keeps working and is rewritten on first use, so a rollback to
  1.1.x needs the keys used on 1.2.0 issued again.
• Changing an MCP server's address to another host no longer reuses its stored
  credential until the credential is entered again.
• undici was updated to 8.11.2 and 7.30.0, and axios to 1.20.0, for their open
  security advisories.

Upgrading

1. Back up PostgreSQL; migration 0091 deletes an unused seeded demo-agent,
   and migration 0086 reads every conversation message, so schedule it for a
   quiet moment or build its indexes beforehand as the upgrade guide shows.
2. Move to Node 24 if you run from source or build your own images, and make
   any volume mounted on the engine's /app/logs or
   /app/packages/engine/workspaces writable by uid 1000.
3. Add the plugins of the extracted tools your agents use to the image:
   polyant-hubspot-plugin
   v1.1.1 for HubSpot and
   polyant-extra-plugin
   v1.1.0 for extra:markdownToPdf and extra:fileUpload. Each README gives
   the install command. The GitHub and Render tools have no plugin.
4. On each agent that should keep files, set its bucket, region and its own
   access keys on the Advanced page and turn on Store attachments.
5. Copy any non-default value of a removed variable into the panel, then remove
   the obsolete environment variables listed in the upgrade guide.
6. Route /webhooks/* to the engine, reconfigure Slack apps for webhooks, and
   start the new engine promptly so Telegram's webhook is registered.
7. Review agents without a pinned model, and Anthropic agents' background jobs,
   which now run on Claude Haiku 5.5.
8. Update scripts that create agents and then call them: save the agent's API
   key and enable its Web channel.
9. Run npm ci, apply migrations, build, restart, then smoke-test sign-in, one
   affected plugin tool, attachment upload/read, a Telegram or Slack message if
   you use them, and a scheduled task.

Compatible clients

These clients are released alongside Polyant v1.2.0:

• polyant-cli v0.9.0, the
  command-line client. Install it with
  npm install -g https://github.com/polyant-ai/polyant-cli/releases/download/v0.9.0/polyant-ai-cli-0.9.0.tgz.
  It authenticates with a management key issued from the panel's API keys
  page.
• polyant-dev-kit v1.2.0.
• Plugin SDK v1.9.0, the version
  this engine pins.

Links

• Changelog
• Upgrade guide
• Plugin guide
• Documentation
• Plugin SDK (v1.9.0)
• polyant-hubspot-plugin (v1.1.1)
• polyant-extra-plugin (v1.1.0)
• polyant-cli (v0.9.0)
• polyant-dev-kit (v1.2.0)
• GitHub repository
• Commit history