fix(relay): classify self-joining agents as bots, not members

[rmichelena]

Fixes #6561.

handle_join_request hard-coded MemberRole::Member. kind:9021 carries no role tag and buzz channels join exposes no --role, so a self-joining agent could not ask for anything else — and clients read the channel role as the "is this an agent" signal, so it vanished from @mention autocomplete. Repair needs role: bot on a kind:9000, which the relay restricts to owners/admins, and the Desktop role menu offers only admin/member/guest — so there is no in-app way back.

Why Member is the wrong default here

MemberRole's own documentation is the argument:

// crates/buzz-core/src/channel.rs
/// The hierarchy for permission checks is: Owner > Admin > Member > Guest.
/// Bot is a **separate designation** — it is not part of the linear hierarchy.

Member is not a conservative privilege choice for an agent — it is the wrong category, and it is the category clients consume.

What this does

Uses the agent discriminator the codebase already defines and relies on:

// crates/buzz-db/src/usage.rs
/// Agent discriminator: `agent_owner_pubkey IS NOT NULL`.

read through get_agent_channel_policy, which this file already calls in handle_put_user's policy check. No new classification rule is introduced.

What this deliberately does not do

handle_put_user is untouched. A caller that states admin / member / guest for a human still gets exactly that. The relay decides only on the one path where no one can express intent — a human self-joining still lands Member, unchanged.

A failed classification lookup logs a warning and falls back to Member, preserving today's behaviour rather than failing the join.

Tests

The decision is extracted into self_join_role so it is unit-testable without a database:

• agent (agent_owner_pubkey present) → Bot
• human (user row, no agent owner) → Member
• no user row → Member, i.e. unchanged when nothing is known

cargo clippy -p buzz-relay --all-targets -- -D warnings is clean.

Closed relays

agent_owner_pubkey is the field that stays NULL when require_relay_membership = true — #5581. Where that is unresolved the discriminator reports every agent as human and this change is a no-op, so #5581 is a prerequisite for closed deployments rather than an unrelated fix. Flagging it so this is not mistaken for a complete fix on such a relay.

[Chessing234]

the reasoning in the comments is unusually good — especially spelling out why Bot can't shadow a human role and why handle_put_user is deliberately untouched. extracting self_join_role for the three unit tests is the right shape.

the gap i'd want addressed is the agents that already self-joined. your own comment notes that changing an active member's role is owner/admin-only, so every agent that hit this before the fix stays Member and stays unmentionable, with no in-app way to repair it — which is the actual complaint in #6561. this makes the bug stop happening; it doesn't make the existing ones work.

is a backfill in scope? something like "set role=Bot where the member row is Member and agent_owner_pubkey IS NOT NULL" is the same discriminator you're already using, so it should be expressible as a migration. if it's deliberately out of scope that's fine, but worth saying so, otherwise the issue looks fixed while affected users still can't mention their agent.

[rmichelena]

Thanks — and yes, you're right that this is forward-only. It stops the bug happening; it doesn't repair anyone already in it, and #6561's complaint is about the latter.

A backfill is in scope as far as I'm concerned. I'd want to narrow the predicate before writing one, though, because the discriminator on its own is too broad.

Why role = 'member' AND agent_owner_pubkey IS NOT NULL overshoots

That set includes agents an owner deliberately added as member — add-member takes an explicit --role, and handle_put_user preserves an existing role precisely because roles can be intentional. A backfill keyed only on "is an agent" silently overrides those decisions, which is the opposite of the scoping this PR is careful about (handle_put_user untouched, relay decides only where nobody can express intent).

There's a column that separates them

handle_join_request is the only production path that creates a member with no inviter:

path	invited_by
handle_join_request (self-join)	None
handle_put_user (kind:9000 / add-member)	Some(actor)
workflow sink agent attach	Some(author)
audio channel auto-add	Some(channel.created_by)

So the affected rows are addressable exactly:

UPDATE channel_members cm
   SET role = 'bot'
  FROM users u
 WHERE cm.role = 'member'
   AND cm.invited_by IS NULL          -- self-joined, nobody chose this role
   AND cm.removed_at IS NULL
   AND u.community_id = cm.community_id
   AND u.pubkey = cm.pubkey
   AND u.agent_owner_pubkey IS NOT NULL;

invited_by IS NULL is what makes it a repair rather than an override: it targets rows where the role was assigned by the code this PR is fixing, and leaves every row where somebody actually chose member.

Scope note on that claim: I checked the add_member call sites in buzz-relay and buzz-db; the non-test ones are the four above. If there's a path I've missed that legitimately leaves invited_by NULL, the predicate needs revisiting — I'd rather you tell me than have me assert it too confidently.

One caveat that limits how much it fixes

agent_owner_pubkey is NULL on relays running require_relay_membership = true, which is the gap #5581 addresses. On those deployments both this fix and the backfill are no-ops until that lands — worth stating in the migration comment so nobody reads a clean run as "no affected rows existed".

Happy to add the migration to this PR, or as a follow-up if you'd rather keep the code change reviewable on its own. Say which and I'll do it.