Teams

A team shares one context the same way your own tools already do: what one teammate’s tool learns, every teammate’s tools can draw on. The team context is a separate, shared store that members copy selected learnings into — it is never a merge of anyone’s personal context.

​

Create a team and invite your teammates

In the OneLamp app, use the team switcher in the sidebar:

1. Create team — name it; you become its owner. “Personal” stays the default context until you switch.
2. Settings → Invite by email — invite teammates as member or admin. They get an email link that drops them straight into the team.
3. Switch the active context any time between Personal and any team you belong to.

Free includes 1 team; Pro lifts it to unlimited. See pricing.

​

Promote a learning

Promotion is the one place a personal learning crosses into the shared store. It is always explicit — there is no automatic copy. Promotion is a web action in the OneLamp app’s Teams UI — there is no chat tool for it. On the Library page, use Promote to <team> on any row (find the entry the way your tools do, via get_context / list_context) — or bulk-promote a whole filtered set. Categories like Projects / Work / People are nudged as team-worthy; Preferences / Style stay personal by default. What happens on promote:

• Secrets are redacted at the boundary — API keys, tokens, and .env-style values are scrubbed before the learning reaches the team.
• You’re recorded as the contributor. If two teammates promote the same (or a similar) learning, it dedupes to one entry and credits both.
• It enters the trust gate (below) — live immediately, but unverified until a teammate independently corroborates it, and never silently overwriting a conflict.

​

Trust & corroboration

The same mechanism that makes the good case great — one teammate teaches, everyone’s tools learn — could let one incorrect fix propagate just as fast. So a promoted learning has to earn its standing. Every team entry carries a trust state: OneLamp decides this at promote time: it finds the nearest existing team learnings and, when one is close, classifies whether the new promotion agrees (corroborate → trust rises) or contradicts (both flip to contested). A contradiction is never allowed to silently overwrite the other side — instead both are frozen and surface in the team’s Needs review list, where an owner/admin keeps one, removes one, or dismisses the conflict. (This judgment is the one model call on the path, run at promote time; retrieval stays generation-free.)

​

Retrieve across scopes

get_context takes an optional scope so a tool can read from your personal store, a team’s store, or both: When more than one store is in scope, OneLamp fuses each store’s ranked results with a final scope-weighted Reciprocal Rank Fusion — personal context wins ties. This is pure rank math in the Worker: still retrieval, never generation. Every returned chunk carries a scope marker (personal / team:<id>), the contributed_by member, and its trust — so the tool (and you) can see provenance. Team results are also trust-weighted: verified learnings rank above unverified above contested (which is surfaced with a ⚠). By default everything is returned (so a brand-new team isn’t empty); pass verified_only: true for high-stakes reads to drop anything not yet corroborated. Your personal context is always included.

​

Provenance, governance & privacy

• Who taught what — the team view groups the shared context by contributor and category, so the team can see exactly what each member taught OneLamp.
• Curation — owners and admins can resolve contested pairs (keep one, remove one, or dismiss), prune any team entry (it’s removed for everyone), and export the whole team context as portable JSON. Members can promote and read, but not prune or resolve.
• Privacy invariants — joining a team never exposes your un-promoted entries; promotion is opt-in, redacted, and reversible; the team owns its data, the same way you own yours.