Thanks for raising this. Could we first separate two questions: can the sandbox run Git at all, and can Git reach the remote repository? Those are controlled at different layers.

For the first question, would you try this from the host?

nemoclaw <sandbox-name> exec -- git --version

Does that print a Git version? The current NemoClaw sandbox image includes /usr/bin/git, and git --version does not use the network, so it needs no egress policy. If the command succeeds, that tells us the sandbox can run Git and the chat agent is having trouble discovering or invoking its shell tool. In that case, could you share:

• nemoclaw --version
• the selected agent, model, and provider
• whether you are using the dashboard or TUI
• the exact response from the agent

If you are on v0.0.74 and the host-side command succeeds, one useful diagnostic is:

nemoclaw <sandbox-name> rebuild --tool-disclosure direct --yes

That changes which registered tools the model sees; it does not bypass sandbox policy, credentials, or approvals. The exec and tool-disclosure commands are documented here.

For the second question, which service are you cloning from?

• For GitHub over HTTPS, you can add the maintained preset:

  nemoclaw <sandbox-name> policy-add github --yes

• For Bitbucket Cloud over HTTPS, NemoClaw does not currently ship a built-in preset. You can open openshell term, retry the clone, and inspect or approve the exact bitbucket.org:443 request. For persistent access, use a reviewed custom preset restricted to bitbucket.org:443 and /usr/bin/git.

One more question: is the repository private? A network policy permits the connection, but it does not provide repository credentials. Please do not paste a token into chat or embed one in the clone URL.

Since you are willing to perform the Git operation yourself, the simplest safe path for a private repository may be:

git clone <private-repository-url>
nemoclaw <sandbox-name> upload ./REPO /sandbox/.openclaw/workspace/

Then ask the agent to analyze /sandbox/.openclaw/workspace/REPO.

Could you start with the exec -- git --version check and let us know whether it prints a version or returns an error? That result should tell us whether to investigate tool visibility or the sandbox runtime next.