claude-code: boundary silently crashes if user is not logged into Coder

[matifali]

Problem

When enable_boundary = true, coder boundary requires an active Coder session. If the workspace owner is not logged in (e.g. session token expired, or first-time setup), boundary crashes silently — Claude Code either fails to start or runs without any network restrictions, with no clear error surfaced to the user.

This is particularly bad because:

• The failure is silent, so the user doesn't know boundary isn't enforcing anything
• There's no mention in the module README that an active Coder login is a prerequisite
• A first-time user enabling boundary has no indication of why things aren't working

Proposed fixes

Two possible directions (not mutually exclusive):

1. Document the requirement — Add a note to the README that enable_boundary = true requires the workspace owner to have an active Coder session (coder login), and link to agent boundaries docs.

2. Fail gracefully — If coder boundary can't authenticate, it should either:

   • Fall back to running Claude Code without boundary but emit a visible warning to the workspace terminal, so the user knows restrictions are not active
   • Or hard-fail with a clear error message rather than a silent crash

The preferred behaviour is debatable — failing open (warn + run) is friendlier but less secure; failing closed (hard error) is safer but may frustrate users. At minimum, the failure should never be silent.

Steps to reproduce

1. Create a workspace from a template using claude-code module with enable_boundary = true
2. Ensure the workspace owner is not logged into Coder (or session is expired)
3. Start the workspace — boundary crashes silently, no clear error surfaced

[DevelopmentCats]

I Tend to lean towards a hard failure here, since most people that are using boundary might prefer that the workspace fail when their preferred security levels aren't met

[matifali]

I Tend to lean towards a hard failure here, since most people that are using boundary might prefer that the workspace fail when their preferred security levels aren't met

That's a good point, but today we are not surfacing that error nicely. The agentAPI just fails to start after 150 attempts, and you need to download the logs to see why.
This is also partly dependent on coder/coder#21946.