Background company research

Workspace creation persists independent quick and deep intentions in PostgreSQL. Discoverybeat repairs ordinary task records and completed outputs; it never calls a public research provider. These tasks join the deployment’s shared FCFS worker queue. Background classification does not reserve or limit worker slots.

oblivectl organization research requires the scoped capability of an active research run. The backend verifies the run, owner, task version, company generation, pass, cycle, and scope. Quick research retains at most four documents and twenty claims, with two 120-second attempts. Deep research retains at most sixteen documents and eighty organization claims, with three 300-second attempts. Each department may submit forty claims using the same retained deep evidence.

Public providers run concurrently. Partial evidence remains usable. Missing deployment research credentials return unavailable coverage; they do not prevent workspace creation or trigger an unbounded research loop. An integration’s research API key is independent of deployment credentials.

Once deep evidence is retained and the owner has continued or skipped optional integration setup, three department tasks become eligible without waiting for the organization’s synthesis. New workspaces persist integrationSetup: "pending" in discovery durability. The owner-facing completion endpoint changes it to complete without requiring a connection or provider call. The scheduler recovers this intent after a lost response or restart. Company research, ordinary tasks, and integration exploration are independent of this marker. Existing workspaces without it remain unchanged. Pending setup retains the initial deep evidence rather than starting a routine refresh. Quick and deep tasks have no dependency on each other. Normal task/run records own execution state. Collection leases own only evidence collection. Task dedupe keys include identity generation, pass, cycle, and scope. Heartbeats cannot extend the backend-issued execution deadline; the worker also aborts execution at that deadline.

Each completed outputs/organization-research.json packet is validated and retained under an immutable content hash. Individual claims require direct, credible, scoped evidence and an explicit enrichment-eligible fact definition. Only unknown fields without an owner edit can be filled. False, zero, and empty collections remain known values. Operating authority, preferences, and workflows are never changed by research. Conflicts remain suggestions; a conflicting research-owned fact is marked disputed without replacing its value.

The immutable Context pointer and the task’s processing receipt commit together. Concurrent publishers rebase against the latest revision. Deep refresh repeats on a thirty-day cadence; the initial quick pass is not repeated by routine refresh.

GET /organizations/:organizationId/discovery returns setup status, each pass’s progress, and ordinary task links. The workspace shows a dismissible notice opening a research sheet with exact task links and human-readable statuses. Failures and human waits remain visible while other research runs. Dismissal is scoped to the displayed task states; new progress or attention resurfaces the notice. Titles truncate within the available width, with their full text available on hover or focus. Published complete deep research supersedes an earlier quick-pass failure for the workspace notice; the original quick task and its history remain unchanged. Name and same-host URL changes preserve research. A new website host increments the identity generation and clears only unchanged facts owned by the prior research generation. Owner edits, submitted settings, and operating defaults survive.