Workspace Intelligence
Ask evidence-grounded questions, inspect reconciled context, prioritize recommendations, and understand limits and cost.
Workspace Intelligence turns the latest architectural truth in a workspace into a shared, reviewable conversation. It combines three surfaces: Recommendations, Ask Hyperoru, and Context.
Recommendations
Recommendations turn existing audit artifacts into prioritized follow-up work. Every recommendation has a category, priority, confidence, lifecycle state, and evidence references.
Teams can:
- acknowledge a recommendation;
- assign it to a workspace member;
- snooze it until a specific time;
- dismiss it with a recorded reason;
- discuss it in a shared conversation;
- prepare a remediation proposal.
Preparing an action does not mutate the workspace. Hyperoru returns inert action data first. The existing audit, report, and remediation permissions are checked again when someone explicitly confirms an action.
Ask Hyperoru
Every workspace member can ask questions about the evidence Hyperoru has already collected. Conversations are shared with the workspace and retained until their creator—or a workspace Admin or Owner—deletes them.
Typical questions include:
- Which services can reach the customer database?
- Why does Hyperoru believe this trust boundary exists?
- Which findings affect the deployment path?
- What changed between the latest two audits?
- Which recommendation should the team address first, and why?
The answer workflow has visible states: queued, gathering context, routing, specialist review, citation verification, completed, cached, limited, cancelled, or failed. The application streams these states without exposing hidden reasoning or model tool transcripts.
Grounding contract
Hyperoru only makes factual workspace claims when it can cite accessible evidence IDs from the same tenant and selected source scope. Unsupported conclusions are labelled as hypotheses or coverage gaps.
The assistant cannot use:
- a shell or unrestricted source access;
- arbitrary SQL;
- external browsing;
- outbound MCP tool execution;
- another tenant's context or cache;
- a prepared action as authorization to change the workspace.
Repository and source text is always treated as untrusted data, including text that resembles instructions to an agent.
Context inspector
The Context view exposes the snapshot used for answers:
| Signal | What it tells you |
|---|---|
| Context hash | The exact reconciled truth version used for retrieval and caching |
| Generated time | When the workspace snapshot was rebuilt |
| Source freshness | When each repository, upload, or MCP source was last observed |
| Coverage | Which expected evidence signals were present or missing |
| Entities and relationships | The structural facts currently believed to exist |
| Contradictions | Evidence that disagrees and still needs review |
| Superseded facts | Historical truth replaced by newer evidence |
The latest valid snapshot for each repository, upload, and MCP source forms the default context. Historical snapshots are retrieved only when the question explicitly asks for a comparison.
Limits and cost
The default launch allowance is:
- 20 questions per user per day;
- 100 questions per workspace per day;
- one active run per user and two per workspace;
- 1,200 final-answer tokens;
- a $10 monthly assistant cap nested inside the workspace model quota.
Limit checks happen before a model call and return a typed error with the reset time. Usage records distinguish chat, proactive recommendation enrichment, cached and uncached work. Cost is attributable to the message, conversation, workspace, day, and month.
Operator governance
Platform Operators can inspect metadata-only context health, run status, cache efficiency, recommendations, and cost. They can change workspace assistant allowances or cancel an active run. Every control action creates an immutable audit event; prompts, answers, source, and evidence content are not exposed in the operator dashboard.
API flow
Create a shared conversation, submit a question with an idempotency key, then follow the durable run event stream:
POST /v1/workspaces/{workspace_id}/assistant/conversations
POST /v1/workspaces/{workspace_id}/assistant/conversations/{conversation_id}/messages
Idempotency-Key: question-unique-123
GET /v1/workspaces/{workspace_id}/assistant/runs/{run_id}/eventsUse the generated Workspace Intelligence reference for request and response schemas.