Troubleshooting

← English docs · 中文

Everything here is written as symptom → cause → fix. If a check fails, dsh-codespaces doctor usually tells you which of these applies.

The GitHub port-forwarding URL does not work

Symptom — you set port 3080 to Public/Private and open https://xxx-3080.app.github.dev, and see:

dsh web authentication required; reopen the URL printed by dsh web.

Cause — dsh binds its session cookie to the authority 127.0.0.1:3080; a different hostname never matches.

Fix — forward the port locally and open the URL dsh printed:

gh codespace ports forward 3080:3080 -c <codespace-name>     # keep this window open
gh codespace ssh -c <codespace-name> -- -i <key> "bash /workspaces/<repo>/.dsh-cloud/start.sh"
# then open http://127.0.0.1:3080/?token=...

Load key ...: Permission denied

Cause — the private key was created by another user (or an agent sandbox), so Windows OpenSSH refuses to load it.

Fix — generate it as the person using the machine:

ssh-keygen -t ed25519 -N "" -f "%USERPROFILE%\.ssh\dsh_cs_key" -C dsh-codespace
icacls "%USERPROFILE%\.ssh\dsh_cs_key" /inheritance:r /grant:r "%USERNAME%:R"

Chinese text becomes ???? in .bat files

Symptom

FINDSTR: cannot open C:\Users\<you>\...\dsh????.txt

Cause — a .bat saved as ASCII mangles non-ASCII filenames; a %USERPROFILE% containing non-ASCII characters makes redirection fail outright.

Fix — keep .bat files pure ASCII (comments, log names, everything), always build paths from %USERPROFILE% / %LOCALAPPDATA%, and add chcp 65001 >nul when you need to display UTF-8 output.

.bat says “The system cannot find the batch label specified”

Cause — the file has LF line endings; cmd’s goto/label handling needs CRLF.

Fix — save the .bat with CRLF endings (and keep it ASCII).

“dsh can’t see images” — it’s the model, not the app

Model Input modality Result
deepseek-flash text + image can read images
deepseek-v4-pro text text only

Switch models. Only custom/proxy vision models need Settings → Models → custom → model options → input types (stored as input / inputModalities).

The container lost its environment after a rebuild

Cause — /workspaces is persistent, $HOME is not. Rebuilding the container wipes dsh, ~/.dsh history and the deploy key.

Fix — keep scripts in /workspaces/<repo>/.dsh-cloud/, push your work, and re-run dsh-codespaces repair (or install/cloud-setup.sh) after a rebuild.

pkill -f "dsh web" kills your own shell

Cause — the command string itself contains dsh web, so pkill -f matches its own process (you see shell closed: exit status 0xffffffff).

Fix — find the PID by port and kill that:

PID=$(ss -ltnp 2>/dev/null | grep ':3080' | grep -o 'pid=[0-9]*' | head -1 | cut -d= -f2)
[ -n "$PID" ] && kill "$PID"

A long command gets truncated over SSH

Symptom — syntax error near unexpected token '(' when pasting a big script.

Cause — a PTY connection truncates very long one-shot commands.

Fix — write the script to a file (base64 works well) and run it, without a TTY.

Empty folders never show up in the repo

Cause — git does not track empty directories.

Fix — the sync script drops a .gitkeep into empty directories (skipping ignored paths). Manual sync: bash .dsh-cloud/sync.sh --now.

~/.nvm/nvm.sh does not exist in Codespaces

Cause — Codespaces ships nvm under /usr/local/share/nvm.

Fix — use absolute paths, or probe several locations:

for d in "$HOME/.nvm/versions/node"/*/bin /usr/local/share/nvm/versions/node/*/bin; do
  [ -x "$d/dsh" ] && { export PATH="$d:$PATH"; break; }
done

The sidebar button for a plugin never appears

  1. Is the plugin in dsh.profile.bundles in ~/.dsh/profiles/web/package.json?
  2. Did you restart dsh (tools/restart-dsh.sh)?
  3. Hard-refresh the page (Ctrl+Shift+R).
  4. Check dsh’s startup log for a loader error — a row that fails to activate can also break other plugins’ settings forms until the page/host restarts.