serve: InstanceStore.boot eagerly initializes every registered project directory - startup is O(projects), minute-scale

Summary

opencode serve runs InstanceStore.boot at listen time, which eagerly initializes an instance for every row in project/project_directory. Each instance pays full init cost (project config load, plugin init, remote MCP client connects, LSP enumeration) — measured ~2–20s per directory. With ~19 projects / ~80 registered directories, the server isn't ready for several minutes. Downstream supervisors that spawn opencode serve and wait for readiness (e.g., Paseo) time out and kill it.

Environment

• opencode-ai 1.18.30 (npm), Windows 11
• ~/.local/share/opencode/opencode.db (~3.3GB) accumulated over months of worktree-based sessions

Evidence

Log excerpt from a single run= of opencode serve — creating instance/fromDirectory/bootstrapping repeating per directory, each followed by config loads, plugin init, MCP connects, and the ~35-entry "enabled LSP servers" enumeration:

message="creating instance" directory="C:\\Docker\\regmais"
message=fromDirectory directory="C:\\Docker\\regmais"
message=bootstrapping directory="C:\\Docker\\project-bulldog\\placeholder-folder"
message=bootstrapping directory="C:\\Docker\\fns-inteligencia-operacional"
message=bootstrapping directory="C:\\Docker\\project-bulldog"
message=bootstrapping directory="C:\\Docker\\dispensamais"
message="enabled LSP servers" serverIds="zls, yaml-ls, vue, typescript, ..."   (per instance)
message=init                                                                 (per instance)

Directories bootstrapped include long-dead paths (%TEMP% dirs, removed worktrees) — those rows are never skipped or pruned, they just add boot cost.

In my case the server was still bootstrapping minutes after listen, while a supervisor's readiness probe timed out and SIGTERM'd it.

Suggestions

• Lazy instance creation: initialize an instance on first request for that directory, not at listen time.
• Or parallelize InstanceStore.boot and skip directories that no longer exist on disk.
• Related: #36840 reports the same eager bootstrap deadlocking under concurrent triggers on WSL2 — lazy init would help there too.

Also encountered on the same install

• #42170 — SQLiteError: no such column: project_id from Project.migrateProjectId inside this same boot path; resolved manually with the workspace table drop/recreate posted in that thread.

Repro

1. Accumulate projects/dirs in opencode.db (or insert project_directory rows directly).
2. opencode serve --port <port> and watch log/opencode.log — one creating instance/bootstrapping/init cycle per registered directory before the server is fully ready.
3. Time to readiness scales roughly linearly with registered directory count.