Context and scope
For a workspace conversation, the request can include the active folder catalog, relevant source profiles, the immediate user message, recent conversation context, workspace context, and—when the conversation belongs to an Investigation—the selected Question, source scope, and saved definitions or assumptions. The harness records the workspace revision and Investigation context used by each run. Later edits do not rewrite the saved run. Without a mounted workspace, data-analysis tools are unavailable. The model can still answer general questions, but those answers have no workspace evidence.Tool loop
- The selected model returns text or a request to use one or more registered tools.
- The harness validates tool names and arguments, then sends approved calls to the Rust engine.
- The engine returns bounded results or an explicit error.
- The harness records the result as evidence and supplies it to the next model turn.
- When the model finishes, deterministic verification checks the answer against available evidence and execution records.
Presentation composition
When accepted analysis results support a useful structured view, the model composes a JSON-render document using the registered component catalog. It can choose a prose, metric, table, chart, or mixed layout that fits the question. It is not required to produce the same visualization for every result. Analytical values reference saved result IDs and fields. The validator rejects unknown components, invalid props, missing results, and unsupported bindings. The renderer does not execute model-authored code and cannot independently access workspace files. Presentation failure leaves the canonical answer and analysis evidence available. Saved presentation snapshots allow the interface to restore an answer without calling the model or recomputing the analysis merely to reopen the conversation.Guidance and enforcement
Limits
- The model may misunderstand a question or choose a semantically wrong query.
- Tool success is not proof that a query answered the intended question.
- Verification can identify concrete inconsistencies but cannot guarantee correctness.
- JSON-render validates what is shown; it does not validate the meaning of the model’s prose by itself.
- No model output can add a shell, filesystem write operation, web tool, or connector.
Built-in tools
See the local operations the harness can call.
Verification
Understand deterministic checks and their limits.