# Release notes

What changed in TicketIQ, month by month, taken from the merge history.

What shipped, newest first. Earlier months are taken from the repository's merge history each time the docs are built.

## 23 October – 5 November 2026

### Build agents

- **Workflow canvas.** Design an agent's workflow by drag and drop: ports, undo and redo, keyboard moves and autosave. Draft a workflow from a description, then dry run it with stubbed tools before it acts. Workflows have stages and case types, each stage with its own SLA, and the runs list shows where each case is. See [Workflow canvas](/docs/agents/workflow-canvas).
- **Solution templates you author.** Package agents, workflows and settings as a versioned template, publish it to your gallery, and install, upgrade or uninstall it whole. An upgrade does not undo a change made while it was being prepared. See [Solution templates](/docs/agents/solution-templates).
- **Apps around agents.** Tenant pages, forms and dashboards built around agents, with a branded web-chat widget.
- **Personas and Home.** Developer, AI engineer, data scientist and business user roles; Home is drawn for the signed-in persona, with an onboarding checklist.
- **Next steps after an answer**, in the workspace and in Teams, and self-service `raise_ticket`, filed as the asker and held for approval.
- **Answers stream as they are written**, released only as their guardrails pass them, with live steps while an agent works.
- **SDKs for Python and TypeScript** (`liyagent` on PyPI, `@liyagent/sdk` on npm, Apache-2.0): typed models, retries, an async client and typed errors, signing in with a service account. See [SDKs](/docs/reference/sdks).
- **The `liya` CLI** signs in through the browser (OAuth device flow), opens your sessions and approvals in the terminal, chats with your own agents, and ships as standalone binaries per OS.

### Data and knowledge

- **Ontology.** Object types (including from SAP and Salesforce metadata), objects, typed links, bound actions and lineage, with agent tools to describe an object, get it and follow its links. See [Ontology](/docs/connect/ontology).
- **Kafka and Kinesis consumers.** Stream sources join webhooks and queues, with batch-quality alerts and lineage. See [Stream consumers](/docs/connect/stream-consumers).
- **Deletion propagation.** A record deleted at its source is removed here too, unless a hold keeps it. See [Deletion propagation](/docs/connect/deletion-propagation).
- **New knowledge connectors**: SharePoint, Confluence, Google Drive and Salesforce Knowledge. Long records are split into cited passages; citations quote the passage that matched.
- **Retrieval**: hybrid fusion by reciprocal rank and an optional reranker, held to tenant data residency.
- **Knowledge freshness**: review dates, owners, expiry and re-verification in the console.
- **ServiceNow** syncs incrementally by its update watermark.

### Models and evaluation

- **Cost per agent action.** A ledger of what each agent action costs, with alerts, and per-stage latency on every request (`Server-Timing` and `/api/perf/stages`). See [Cost per action](/docs/monitor/cost-per-action).
- **Open-weight route.** Eligible steps can go to an open-weight model, with the frontier model kept for the rest and for any step the open-weight answer looks wrong on.
- **Output drift.** A nightly check of whether each agent still answers as it did: refusals, answer length, tool mix, escalations, confidence and grounding, beside the input-drift check. See [Output drift](/docs/monitor/output-drift).
- **Model router**: pick a binding's model per call by cost and quality; a model registry with versions, stages and an eval-gated promotion; experiment tracking; fine-tuning jobs on Bedrock and SageMaker from approved corrections.
- **Gateway**: governed embeddings, images and speech beside chat.
- **Evals**: faithfulness, judge calibration against human review, slice metrics and a bias check, adaptive red-team runs.

### Govern

- **Policy drafts.** Describe a rule in plain language and get a Cedar policy draft to review; activating it takes a second approver, or a recorded reason on a one-admin tenant. See [Policy drafts](/docs/govern/policy-drafts).
- **High-risk agent gate.** An agent assessed as high risk cannot run until its AI impact assessment is signed off, and not by whoever wrote that version. See [High-risk agent gate](/docs/govern/high-risk-agent-gate).
- **MFA for every local account**, on by default; lifting the requirement is a weakening that takes a reason. The environment administrator enrols like everyone else; automation signs in with service accounts.
- **Security fixes from an internal penetration test**: tenant checks on source routes, CSRF protection on cookie writes, and turning on MFA ends existing password sessions.
- **Support access**: approved, time-boxed, read-only use of one person's data.
- **Tenant isolation**: row-level security on every table that holds a tenant's data.
- **Evidence export**: the control crosswalk (now with SOC 2 and ISO/IEC 27001 Annex A) as OSCAL and GRC-tool CSV, signed.
- **Supply chain**: SBOM, image scan and keyless signing for release images.
- An optional OS-level sandbox around code mode.

### Operate

- **Crash resume for every run kind.** Managed runs, cron, triggers, co-pilot turns and Teams turns resume on another process from their last step, and a re-executed step applies once at the vendor. See [Crash resume and high availability](/docs/operate/crash-resume-ha).
- **Two app processes in production**, with zero-downtime blue/green deploys; N replicas on one database, and a Helm chart with an autoscaler and a disruption budget.
- **Public status page** at `/status`: live component state, uptime measured over the days actually recorded, the SLA and incidents. See [Status page](/docs/operate/status-page).
- **Backups**: an S3 copy with Object Lock, WAL archiving and point-in-time recovery, drilled against a real bucket.
- **Trace export** to your OpenTelemetry collector.
- **Alert email** that says what happened and how bad, sent through the outbound relay.

## October 2026

- **2026-10-22**: Three home page options for review (/home-a, /home-b, /home-c)
- **2026-10-07**: Business Central live-environment fixes
- **2026-10-06**: Salesforce org context: org model, personas, playbooks, eval gate
- **2026-10-06**: Dynamics 365 family support
- **2026-10-06**: The red-team release gate and signed audit bundles
- **2026-10-06**: Secrets for ticket sources and the service accounts console
- **2026-10-06**: The Finance agents
- **2026-10-05**: Salesforce Agentforce-style actions
- **2026-10-05**: The Salesforce solution: Overview and Setup pages, Liya for Salesforce
- **2026-10-04**: Console sign-in session limits: inactivity, lifetime, Access › Sessions
- **2026-10-04**: Data policy fix: overlapping allow-list keeps are unwrapped
- **2026-10-04**: Login sessions in Postgres: a deploy no longer signs everybody out
- **2026-10-04**: Solutions per tenant: owner write path, tagging, enforcement switch (shadow), legacy adoption
- **2026-10-04**: Provider bindings details, in-page confirmations for destructive actions, and sightings registration
- **2026-10-04**: The console look-and-feel standard (S1-S13), row menus and the MCP page
- **2026-10-04**: Main into the provider and usage work
- **2026-10-04**: The provider page: needs attention, one range, traffic chart, leaderboard tab
- **2026-10-04**: The Cost & Usage selection-versus-total and shared crosshair work
- **2026-10-03**: signed SAP buttons and per-person FSM identity, with routines
- **2026-10-01**: Connect SAP
- **2026-10-01**: SAP SuccessFactors connector
- **2026-10-01**: The privacy policy, DPA, terms and privacy-rights requests
- **2026-10-01**: Main (citation original links) into the privacy branch
- **2026-10-01**: Agent finder (/start) and Liya site chat widget
- **2026-10-01**: HR agents: Liya HR and Liya Onboarding on Workday/BambooHR, synthetic HR system, Teams routing, HR evals
- **2026-10-01**: Per-client subagent switches: built-ins permanent, each tenant turns its own on or off
- **2026-10-01**: Unit-test regression fixes: gateway route permissions, console contract, Liya inheritance tests, SQL checker, CLI imports
- **2026-10-01**: tab URLs, trigger icon tiles, cleaner audit toolbar, provider marks, login copy, rail badges, Access More menu
- **2026-10-01**: Agents list, Agent network and Access redesign
- **2026-10-01**: Main (numpy pin; main's own close of the two open media blocks)

## September 2026

- **2026-09-30**: Main (Audit log, LLM providers, Guardrails, Docs, audit-coverage entries)
- **2026-09-30**: Docs site and Labs; keep the CLI worker's liya pages and rebuild
- **2026-09-30**: Guardrails page
- **2026-09-30**: LLM providers page
- **2026-09-30**: Main (liya CLI, Secrets store) into the agents/network/access branch
- **2026-09-30**: Secrets store redesign
- **2026-09-30**: Liya CLI
- **2026-09-30**: Agent page redesign and Liya orchestrator hierarchy
- **2026-09-30**: Semantic answer cache, known-issue flow and repeat protection
- **2026-09-30**: Liya agent names and avatars
- **2026-09-30**: Durable task handles for long-running tool calls
- **2026-09-30**: Substantial-modification re-review, prohibited-practice screening, high-risk tracking
- **2026-09-30**: LLM simulated users with hidden goals for multi-turn testing
- **2026-09-30**: Ticket-event triggers with filters, exactly-once runs and declared run-as
- **2026-09-30**: Decision API (PDP) for framework-local tools, retrieval and memory writes
- **2026-09-30**: Session failure-pattern clustering with lifecycle, root cause and proposed fixes
- **2026-09-30**: Email as a conversational agent channel with threaded replies
- **2026-09-30**: tiq CLI lifecycle verbs with CI exit codes and scoped service token
- **2026-09-29**: Registered agents as subagents, shared context modes, bounded parallel tool calls, depth cap
- **2026-09-29**: Embedding-model migration: dual-write, shadow read, gated cutover, rollback
- **2026-09-29**: Interactive and batch lanes on the vector store lock
- **2026-09-29**: Per-run cancellation for triggers, gateway disconnects and A2A
- **2026-09-29**: Sender trust for mail intake; exclude external resolutions from grounding
- **2026-09-29**: LLM-judge evaluators, idle-conversation scoring, online sampling, OTLP eval events
- **2026-09-29**: Cedar AgentKillSwitch and KillSwitchPolicy child entities
- **2026-09-29**: Real parent billing for delegation, chain on every row, depth and loop guard
- **2026-09-29**: General alert rules with levels, windows, no-data, preview, digests, PagerDuty/email
- **2026-09-29**: Governed intake, change requests and offboarding for agents
- **2026-09-29**: Sessions pin the agent version they started on
- **2026-09-29**: Remote A2A proxy streaming passthrough, signed-card verification, outbound streaming
- **2026-09-29**: Eval-validated optimizer for instructions, tool descriptions and model choice
- **2026-09-28**: Automated red teaming: attack corpus, transforms, indirect injection, ASR reports
- **2026-09-28**: Playground variant compare, dataset runs, replay with edits, simulate-as
- **2026-09-28**: Per-tenant DEKs with crypto-shred and optional content sealing
- **2026-09-28**: Egress rule engine with audit/enforce modes, proxy support, LLM and stdio coverage
- **2026-09-28**: Decision replay, candidate settings/versions, eval-gated promotion with sign-off
- **2026-09-28**: Output-side system-prompt/context leak detection with canary tokens
- **2026-09-28**: A2A message/stream and tasks/resubscribe with ordered status/artifact events
- **2026-09-28**: r6-well-known-logos
- **2026-09-28**: r6-transcript-otlp-enduser-timerange