Quota radar
Provider-reported windows only. Projections are linear, from recorded samples.
Providers
Agent sessions
Start here
THUS is a shared, permission-aware memory for engineering changes. Before a person or a coding agent changes something, Max checks the company's own evidence (code, decisions, policies, past incidents, chat) and freezes what it found as an Assurance Case: every claim cites its source, unknowns are listed, nothing is marked verified until a person checks it. Claude Code and Codex read and hand off the same Case.
What is real here, and what is not
- Synthetic: the company (Northwind), its people, documents and chat. No real data.
- Real: the engine, permissions, evidence citations, frozen Cases and review flow, the same code our private pilot runs.
- Not in this sandbox: repositories, coding agents, model calls. Those run in the pilot and are shown in the video.
- Measured honestly: on 64 held-out real-repository changes, Max matched a whole-repository Claude review on precision (0.72 vs 0.75) at about a quarter of the model cost, but found fewer of the files that changed (recall 0.41 vs 0.53).
Max change preflight
Freezes an immutable Assurance Case from evidence the acting persona may read at the cutoff.
Activity
What people and agents are working on, tied to Cases. "Verified" counts only what a person checked; everything else is agent-reported.
Connect Claude Code or Codex to this workspace
Creates a token for one agent on one machine. The agent can read Cases, run preflight and report work as you, in this workspace only. It cannot verify, review or administer. Valid 90 days; revoke it here at any time.
Work Sessions
Structured continuity across agents. No transcripts.
Authorized sources
Read-only. Add, revoke or purge with python -m thus_core sources ... (see docs/ACTIVATE.md), then restart.