What We Shipped in June 2026

May was about confidence before you ship. June is about control after you ship: who can see and do what while agents are running, and where everything they produce ends up. Fixed admin and member roles gave way to customizable permission groups, dashboards can now be shared with exactly the people you choose or published read-only, and channels and drains route notifications, approvals, and logs to your own systems. We also sharpened the built-in prompt-injection and data-exfiltration guardrails, added AI-assisted filtering to the database and runs table, and taught deployments to drain in-flight runs before they roll over.

Customizable Permission Groups

Members used to be either an admin or a member, each with a fixed set of permissions. Now you build your own permission groups from granular, action-level permissions and assign them to members and pending invites. Each permission maps to a specific action, so a group can deploy and activate builds without ever touching billing or team membership.

# Composed from granular, action-level permissions
deployments.activate      # promote a build to live
deployments.override      # override deployment checks
connectors.view
budgets.view
# billing.manage and team.update left out on purpose

• •Action-level access: Grant specific actions like connectors.create or approvals.decide instead of a blunt role
• •Applies to invites: Groups attach to pending invites, so a new teammate lands with the right access on day one
• •Ownership transfer: Hand a project over to another owner from team settings

See team and permissions for the full list of actions.

Shareable Dashboards with Access Controls

Observability dashboards now carry their own visibility and edit controls. Keep one private, open it to specific permission groups, share it with the whole project, or publish it. A public dashboard gets a secure share link and an environment picker, so you can hand a live view to someone who does not have a Connic account.

private   only you, plus editors and view-all admins
groups    members of specific permission groups
project   everyone on the project
public    anyone with the secure share link

• •View vs edit: Separate who can see a dashboard from who can change it
• •Public links: Share a read-only, tokenized link with an environment picker for people outside the project
• •Tied to groups: New dashboards.viewAll and dashboards.update permissions decide who manages what

More in the observability docs.

Channels and Drains

Connic can now push project activity out to your own systems. Create a channel that POSTs JSON to an HTTPS endpoint you control, then route notifications and approval requests to it. Drains do the same continuously for two high-volume streams: project logs and audit events, batched and forwarded as they happen. Webhook delivery is self-service, with Slack, Teams, and PagerDuty available on request. Channels are an Enterprise feature.

POST https://hooks.acme.com/connic
Content-Type: application/json
X-Connic-Signature: <HMAC-SHA256 over "{timestamp}.{body}">
X-Connic-Timestamp: 1782864000

• •Notifications and approvals: Route project notifications and approval requests to a channel, down to the individual agent
• •Log and audit drains: Stream project logs and audit events to an endpoint, batched for volume
• •Signed and verifiable: Set a signing secret and every request carries an HMAC-SHA256 signature you can check before you trust it

Sharper Guardrails

Two of the built-in guardrails got harder to slip past. The prompt-injection guardrail now normalizes confusable Unicode characters and strips invisible formatting marks before it runs its checks, closing the Unicode-smuggling trick where lookalike characters hide an injection. The data-exfiltration guardrail now flags Base64-encoded data blocks in the output, catching payloads that try to slip out in encoded form.

guardrails:
  input:
    - type: prompt_injection
      mode: block
      config:
        sensitivity: high
  output:
    - type: data_exfiltration
      mode: block
      config:
        allowed_domains: [example.com]

• •Unicode smuggling blocked: Confusable characters are normalized and format marks stripped before the heuristics run
• •Base64 exfiltration caught: Base64-encoded data blocks in the output are flagged, so payloads cannot slip out encoded
• •Skip after-block middleware: Set run_after_on_block: false so a blocked input skips your after middleware

Full reference in the guardrails docs.

AI-Assisted Filtering

Describe what you are looking for and Connic writes the filter. The runs table and the database browser both take a plain-language prompt and turn it into a real filter expression, status, and time range you can still tweak by hand. The database browser also swapped its raw-JSON document editor for a visual key/value editor.

Prompt:  orders over $100 that failed this week

Generates:
  statuses     failed
  time range   7d
  expression   input.amount > 100

• •Runs and database: Natural-language filtering across both the runs table and the database browser
• •Real expressions: Prompts compile to the same expression language you can write by hand, like context.tier == 'enterprise'
• •Visual document editor: Edit stored documents as key/value pairs instead of raw JSON, with knowledge-base column sorting alongside

See the database docs for the query and filter reference.

Safer Deploys and Searchable Logs

A new deployment no longer cuts off the old one. Deployment draining waits for in-flight runs on the previous deployment to finish before it retires, while new and re-triggered runs go straight to the new one, so a rollout never kills a run mid-flight. Project logs also gained full-text search: filter by log message, source, or agent name, backed by a partial index that keeps the log feed fast on busy environments.

• •Graceful rollouts: In-flight runs finish on the old deployment while new runs move to the new one
• •Log search: Find a log line by message, source, or agent name
• •Built for scale: A supporting partial index keeps the log feed responsive on large environments

More Improvements

• •Testing, continued: expected_tool_call_order asserts tools fire in a given relative order, and file-level default mocks cut repetition across a test file
• •The connic skill CLI: Run connic skill (or connic init --skill) to drop the Connic skill into a project so your AI coding agent follows Connic conventions
• •Wider public API: Environments, deployments, connectors, channels, audit logs, judges, knowledge bases, and approvals are now reachable with an API key
• •Budget grouping: Group budget usage by agent or by model to see where spend concentrates