Tokens by Provider

Raw Groups

Every provider × model × agentType group as returned by the API.

AI Provider Defaults

These override the server's built-in fallback order for every user. Leave a field on "Automatic" to use the server's normal logic (per-agent preference, then whatever provider is configured).

Project Memory

What the Flexible (beta) engine above knows about this specific project beyond the conversation itself — architecture, past incidents, conventions, decisions. Entries you add here are active immediately. Entries the agent records itself (via its remember_lesson tool) always land as a Draft first — review and confirm one below before it can affect any future command.

Strategies

The Core AI Orchestrator's Evaluation Agent may propose a short, general instruction after a complex request — it has ZERO effect on anything until approved here. Approving one adds it as guidance for every future request of that agent type (or every type, if unscoped); rejecting — or later revoking an already approved one — keeps it inert.

Recent Runs

Every COMPLEX-path request runs a bounded debate: two independent solutions, an adversarial Critic, one reconciliation round only on a real disagreement, then Verification. Most recent first — click a run to expand its full trace. This is a transparency record, not an editable one. "Verified" here means only "internally consistent, nothing here needed external confirmation this system can't perform" — never "independently fact-checked against reality."

Agent Performance

Aggregated, read-only metrics per agent type, provider, and model — success rate, verified rate, how often the two solvers disagreed, average latency, self-assessed quality scores, and token cost. Use this to decide, by hand, whether to change a default provider in Settings — nothing here changes routing automatically.

Speak naturally to control the platform — "check why video generation is failing", "have the coding agent fix it", "run all tests", "deploy the update to staging". This talks to the same Dev Pipeline and Core AI agents as everything else in this dashboard, just triggered by voice or text instead of clicking through screens. Voice control carries no more authority than a typed admin command — same ADMIN-only login as every tab here.