Designing a Polymorphic Assignee Model for Humans and AI Agents

Most bug trackers were designed when the only thing that could own a ticket was a person. Now AI coding agents triage, analyze, and fix bugs. If your data model treats an agent as a special case — a flag here, a side table there — every query, every API response, and every UI dropdown grows a branch. BugMojo took the opposite bet: an agent is a peer of a human at the row level. The mechanism is a polymorphic assignee, and it is small enough to fit on one screen of schema. This post walks the actual implementation: the shared enum, the inline columns, the discriminated-union resolver, and the loop where assigning to an agent queues real work.

The naive alternative is two nullable foreign keys, member_id and agent_id, with the convention that only one is set. The database cannot enforce that convention. It will happily store a row with both set, or neither, and every read has to COALESCE across the columns to figure out who owns the thing. Add a third actor type and you add a third column and a third partial index. The pair model collapses all of that into one discriminant plus one id.

Prisma has no native union type, so it documents this shape as Single-Table Inheritance: one table, a discriminator column, and variant-specific fields marked optional with ?. BugMojo applies that pattern directly. A shared ActorType enum names the kinds of owner; assignee_id and assignee_type live inline on the Issue row; a composite index makes ownership queries a single B-tree lookup.

The honest cost is in the comment: assignee_id is not a foreign key Postgres can validate, because it points at different tables depending on assignee_type. Referential integrity for the member-or-agent link moves into application code. We accept that because the read path — list a project's bugs and who owns each — is the hot path, and inline columns keep it join-free.

On the read side, the discriminant becomes a TypeScript discriminated union. The Handbook describes the rule precisely: when every member of a union shares a common property with literal types, the compiler treats it as a discriminated union and narrows the members on a switch. assignee_type is that discriminant. Better still, you can lean on never for exhaustiveness — add a new ActorType and the build breaks until you handle it.

Doing one lookup per issue is the obvious N+1 trap. Batch it instead: collect every MEMBER id and every AGENT id from a page of issues, run two IN queries, then stitch the results back by id. A 50-row board costs two queries, not fifty.

Here is where the model pays off. The Model Context Protocol defines tools as model-controlled: per the spec, "the language model can discover and invoke tools automatically based on its contextual understanding and the user's prompts." Each tool is a JSON-RPC 2.0 message with a name, a description, and an inputSchema, discovered via tools/list and invoked via tools/call. So assign_bug is just another tool the model selects — and its inputSchema carries the very same assignee_type enum the database uses.

The architecture spec builds every interaction on JSON-RPC 2.0 with capability negotiation at initialization, so the same tool surface (assign_bug, list_team, get_bug) is exposed identically to every client — no per-client assignee logic. An agent calling assign_bug with assignee_type: 'AGENT' is doing exactly what a human does clicking a dropdown. And BugMojo closes the loop: an AGENT assignment auto-spawns an AgentTask (FIX_BUG or BUG_ANALYSIS), so the assignment queues work rather than just painting a label.

• Define a shared ActorType enum (MEMBER, AGENT, optionally SYSTEM) once, in the schema package.
• Add assignee_id String? and assignee_type ActorType? inline on the owned model; add @@index([project_id, assignee_id]).
• Backfill: set assignee_type = 'MEMBER' for every existing member_id, copy the id, then drop the old FK columns.
• Write one hydrateAssignee resolver with a switch and a never default for exhaustiveness.
• Batch hydration — two IN queries per page, never one per row.
• Expose assign_bug with an inputSchema whose enum matches ActorType; spawn an AgentTask on AGENT assignment.