Launching a run
The agent calls the run tool after the case has been materialized and validated against its solver pack's gates (mesh quality, BC completeness, solver settings). The compute plane receives the Input Package, hashes it, and either provisions a container or attaches to a warm pool slot.
Live monitor sidecar
A sidecar (see the live-monitor-sidecar runbook) tails solver logs and emits typed LiveMonitorFragment records — residuals per equation, iteration counters, monitor fields, wall time, warnings. The case view renders them as residual plots and status badges in real time.
Pause and resume
Runs can be paused (the next checkpoint commits, the container hibernates) and resumed without losing iteration state. Useful for cost control or when you want to inspect intermediate state before deciding to continue.
Cost envelope
Each run carries a typed cost envelope — CPU budget, GPU budget, wall-clock cap, dollar cap. The agent will refuse to exceed the envelope without re-approval. Spend metering hits packages/billing so admins can audit per-tenant.
The run record
When the run finishes (success or failure), the platform freezes a RunRecord: content-addressed input hashes, command transcript, exit status, residual history, monitor fields, artifact paths, cost. This record is what the report cites — never the chat log.