Creating workspace failed

[b0tm1nd]

Steps to Reproduce

1. Clone repo, run cd plugin && npm install && npm run build
2. Add path-based entry in opencode.jsonc pointing to the build:
   ["/path/to/plugin", { "providers": { "default": { "hosts": [...] } } }]
3. Run OPENCODE_EXPERIMENTAL_WORKSPACES=1 opencode
4. Press /warp — you see "New workspace: SSH Provider" option
5. Select it — error appears immediately: "Creating Workspace Failed"

Root Cause (initial)

The resolveProvider function in src/index.ts throws immediately if extra.provider is not set:

if (!workspace.extra || typeof ...provider !== "string") {
  throw new Error("Workspace extra.provider must be configured");
}

OpenCode TUI’s workspace creation does not pass extra.provider, so it fails instantly.

After Patching

We added a fallback to the first available provider when extra is empty. After that:

• The plugin proceeds to bootstrap (SCP stub, kill old process, start stub, SSH tunnel)
• But bootstrap itself has bugs (see below)
• After fixing those, bootstrap succeeds (~10s), health check passes, workspace is saved to SQLite DB
• But the TUI still shows "Creating Workspace Failed" (now after 10 seconds timeout) — this is a separate issue

Additional Bugs Found in the Plugin Code

1. SCP fails when stub is running — SCP happens before pkill, so if stub exists, the file is locked
2. pkill -f kills the SSH session — the command itself contains the pattern it’s searching for
3. pkill without -f silently fails — Linux truncates process names to 15 chars
4. No python3 fallback — buildRemoteStartCommand uses python2 or python, not python3
5. No start_new_session=True — child process may die when SSH closes
6. State not persisted — RuntimeState is in-memory only; after TUI restart, getTarget fails
7. Bootstrap is slow — each step (mkdir, SCP, pkill, start) is a separate SSH connection (~12s)

Tested with OpenCode 1.14.51 and 1.15.7 — same behavior.

Missing npm Package

The opencode-remote-provider npm package does not exist on npmjs.com.
This means the documented package-based installation ("opencode-remote-provider" in config) cannot work.
Only path-based local development loading is possible.

Config Format

Scripts in ./scripts/ only support opencode.json (JSON), not opencode.jsonc (JSON with comments). Had to use OPENCODE_CONFIG env var to point to the config file.

[b0tm1nd]

Update: Progress after investigation

Phase 1 — Workspace creation fixed

• Fixed stub: SSE flush, auto-create sessions (ensureSession()), added missing endpoints (/sync/history, /session/{id}/message, /session/{id}/command)
• Fixed plugin: pkill -f → fuser -k PORT/tcp, unique ports per workspace, per-workspace state dir
• Result: workspace creation succeeds, bootstrap works

Phase 2 — Session routing issue

• Workspace appears in TUI, but first shell/command call fails with "workspace not found"
• Root cause: OpenCode TUI does not pass workspaceID in SDK calls (shell/prompt/command)
• Stub falls back to last registered workspace — this overwrites the previous one
• Fixed stub to properly track sessions, previous workspace no longer gets overwritten

Phase 3 — Fundamental limitation

• Nothing actually goes remote without OpenCode core changes. Not even shell commands.
• The workspace routing middleware cannot match requests without ?workspace= in SDK calls
• File tools (read/write/search/glob) AND shell/prompt/command all operate locally
• The stub runs on the remote server, SSH tunnel works, but routing is broken at the OpenCode level

Required OpenCode Core Changes (PR ready in fork)

1. Add workspace: workspaceID to shell/prompt/command SDK calls in prompt/index.tsx (3 lines)
2. Route POST /session locally (workspace-routing.ts)
3. Read workspace from query in session handler

Fork with all changes: https://github.com/b0tm1nd/opencode-remote-ssh