Problem

OpenCodeWorkerd.create({ storage }) puts OpenCode's tables in the Durable Object's SQLite database without a prefix: session_v2, session_message, event, kv, project, permission, workspace, and so on. That causes two problems when the object also stores its own data:

1. Bootstrap fails. DatabaseMigration.apply stops with Database is not empty and has no session table if any table exists whose name doesn't start with _. Most objects create their tables before OpenCode boots. The Cloudflare Agents SDK, for example, keeps its scheduler and streams in cf_agents_* tables. So in practice, embedding OpenCode means it has to boot first, or the embedder has to hide its tables from OpenCode's discovery query.
2. Names can collide. Names like session, message, event, kv and project are common, so an object's own tables can clash with OpenCode's.

Workerd is where this shows up most, because the database belongs to the Durable Object, not to OpenCode. The same applies to anyone sharing a database through database.path.

Proposal

Add a table prefix option to the database layer, and expose it from OpenCodeWorkerd.create and OpenCode.create:

OpenCodeWorkerd.create({ storage, database: { prefix: "opencode_" } })

When it is set:

• every table and index in schema.gen.ts and the migrations is created as <prefix><name>
• the bootstrap check only looks at tables that start with the prefix, so tables outside it are ignored, as underscore-prefixed tables are today
• the migration journal is prefixed too

Drizzle table definitions already go through sqliteTable(name, ...), so a prefixed table builder plus prefixed DDL in the generated schema and migrations would cover most of it.

What we do for now

cloudflare/agents is adding an OpenCodeHarness built on @opencode/sdk/workerd. Until there is a prefix option, it hands OpenCode a view of DurableObjectStorage that rewrites the SQL OpenCode sends: table names get an opencode_ prefix, and the bootstrap discovery query only sees opencode_* tables. This relies on the shape of OpenCode's SQL, which is why we'd like to replace it with a supported option.

Related: #53221, the same bootstrap check failing on a shared database.