WEEKLY UPDATE October 19 to 25, 2026 This week was about running Roark from your own code and your own tools, instead of clicking through ours. Five things landed in that direction, and I think the shape of the platform is finally where I wanted it. |
Nº 01 Config as Code in TypeScriptYou can now author your entire Roark config in TypeScript, with typed references and autocomplete, and ship it with the same CLI you use for YAML.

What's new:
- Author agents, tools, flows and plans in typed TypeScript, with inline prompts and autocomplete across the whole config.
- Reference other entities by variable instead of string names, so a rename or a typo surfaces before you deploy.
- Apply from the CLI the same way you do YAML, with a diff preview before anything changes in your project.
👉 Read the TypeScript authoring guide
Nº 02 Analytics and Dashboards over the APIYour agents and your CI can now read the same numbers your dashboards show, define dashboards in code, and have a widget's query round-trip cleanly between the API and the platform.

What you can do:
POST /v1/analytics/query for ad-hoc analytics, with GET /v1/analytics/schema so an agent or script can build a valid query on the first try (metrics, filters, breakdowns, intervals, display types).
/v1/dashboard CRUD plus /data and per-widget /data to create dashboards, edit single widgets by name, and run every widget server-side.
kind: dashboard in Config as Code, so a dashboard defined in git shows a Managed by config pill in the platform and stays in sync on every apply.
👉 Read the analytics API guide
Nº 03 Email and Slack delivery for simulation plan reportsSimulation plans can now send each run's Debrief by email (PDF attached) and post it to Slack (PDF uploaded into the thread), using the same shape saved groups already use.

What's new:
- Define recipients once on the plan, in config or on the Report delivery panel in the dashboard, with up to 50 email addresses (no Roark account required) and 10 Slack channels.
- Run a saved plan by its config key, not just by its UUID, so your CI step and your YAML stay name-based end to end.
👉 See the report delivery docs
Nº 04 Project-scoped spend access, plus cost per call and per runYou can now pull Roark spend into your own dashboards with a project-scoped credential that reads only its project's cost, and break spend down to a single call or a single simulation run.

What's new:
GET /v1/usage with a project API key that holds the new billing:read grant: scoped to that project, with a 403 (never a silent retarget) if a caller names another project.
GET /v1/call/{callId}/cost returns a call's billable spend by source and surface, with the cost lines the Cost tab already shows.
GET /v1/simulation/plan/job/{jobId}/cost returns a run's total spend plus per-simulated-call cost.
👉 Read the usage reporting guide
Nº 05 Voice Consistency by SentenceA new grouped metric that catches voice problems only visible across calls: one sentence of your prompt read in a different voice in most calls, while any single call looks fine.

How it works:
- Add Voice Consistency by Sentence (or its check) to any metric collection job of 6 or more calls, and Roark groups lines by normalised sentence, flags any that break far more often than the rest of the group, and writes a per-turn explanation pointing at the sentence as the likely source.
- Calls are compared only within the same agent and endpoint, so an A/B test on two numbers doesn't pollute the baseline.
- Deterministic, so it's free to run alongside your existing metrics.
|
|