[BUG] codex and opencode providers fail through OpenShell while claude succeeds

[nam685]

[BUG] codex and opencode providers fail through OpenShell while claude succeeds

Agent Diagnostic

Investigated before filing. Reproduced with a teammate. Verified the same CLIs succeed locally outside OpenShell, then fail only when launched through OpenShell sandboxes. Also verified claude is a working control case in the same environment.

Description

We are seeing a selective provider brokering/injection failure in OpenShell:

• claude: local works, OpenShell works
• codex: local works, OpenShell fails
• opencode: local works, OpenShell fails

This does not look like a general sandbox/runtime/network issue. It appears limited to the codex and opencode provider paths.

Expected

• If the local CLI is already authenticated and can run successfully outside OpenShell, the equivalent openshell sandbox create ... --provider ... -- <cli> invocation should also work.
• Provider credential brokering should behave consistently across claude, codex, and opencode.

Actual

• claude works locally and through OpenShell.
• codex works locally, but through OpenShell it repeatedly reconnects and ultimately fails against https://api.openai.com/v1/responses.
• opencode works locally, but through OpenShell it hangs after its one-time database migration step.

Minimal Reproduction

Control case (claude, works):

claude --dangerously-skip-permissions -p "say hello"

openshell sandbox create --name claude-probe \
  --provider claude \
  --auto-providers \
  --policy configs/policies/default-deny.yaml \
  --no-keep --no-tty -- \
  claude --dangerously-skip-permissions -p "say hello"

Failing case 1 (codex, local works, OpenShell fails):

codex exec --dangerously-bypass-approvals-and-sandbox "say hello"

openshell sandbox create --name codex-probe \
  --provider codex \
  --auto-providers \
  --policy configs/policies/default-deny.yaml \
  --no-keep --no-tty -- \
  codex exec --dangerously-bypass-approvals-and-sandbox "say hello"

Failing case 2 (opencode, local works, OpenShell fails):

opencode auth login
opencode run --model anthropic/claude-sonnet-4-20250514 "say hello"

openshell sandbox create --name opencode-probe \
  --provider opencode \
  --auto-providers \
  --policy configs/policies/default-deny.yaml \
  --no-keep --no-tty -- \
  opencode run --model anthropic/claude-sonnet-4-20250514 "say hello"

Observed Failure Modes

codex through OpenShell:

Reconnecting... 1/5 (unexpected status 401 Unauthorized: Missing bearer or basic authentication in header, url: https://api.openai.com/v1/responses, ...)

and in a later repro:

Reconnecting... 1/5 (stream disconnected before completion: error sending request for url (https://api.openai.com/v1/responses))

opencode through OpenShell:

Created sandbox: opencode-probe

Performing one time database migration, may take a few minutes...
sqlite-migration:done
Database migration complete.

and then it hangs indefinitely.

Additional Diagnostics

• openshell provider get codex shows Credential keys: <none>, but that field alone is probably not diagnostic because claude can still work while also showing <none>.
• In a direct env probe inside a sandbox launched with --provider codex, the sandbox had:
  • proxy vars set (for example HTTP_PROXY=http://10.200.0.1:3128)
  • OPENAI_API_KEY=openshell:resolve:env:OPENAI_API_KEY
• In a direct Anthropic-backed opencode repro inside the sandbox, the environment still appeared OpenAI-shaped:
  • OPENAI_API_KEY=openshell:resolve:env:OPENAI_API_KEY
  • ANTHROPIC_API_KEY was absent

That suggests the opencode provider path may not be switching credential injection based on model family, and the codex provider path may not be brokering OpenAI auth correctly.

Impact

OpenShell currently works for claude in this setup, but not for codex or opencode, even when both CLIs are already known-good locally. That blocks multi-agent/runtime integrations that rely on OpenShell to launch different coding agents interchangeably.

Environment

• OS: macOS 26.3.1 (build 25D771280a), Apple Silicon
• Kernel: Darwin 25.3.0 arm64
• OpenShell: 0.0.11

Notes

• We separately fixed a local policy issue requiring the internal proxy endpoint 10.200.0.1:3128 to be allowed in our bundled policies. That was necessary for sandbox egress, but it does not resolve this provider-specific behavior.
• After that proxy fix, claude remains healthy, while codex and opencode still fail only through OpenShell.

Agent-First Checklist

• I pointed my agent at the repo and had it investigate this issue
• I loaded relevant local context and repro logs before filing
• My agent could not resolve this — the diagnostic above explains why

[johntmyers]

Can you provide the full steps for one of the scenarios? For example you have:

opencode auth login
opencode run --model anthropic/claude-sonnet-4-20250514 "say hello"

openshell sandbox create --name opencode-probe \
  --provider opencode \
  --auto-providers \
  --policy configs/policies/default-deny.yaml \
  --no-keep --no-tty -- \
  opencode run --model anthropic/claude-sonnet-4-20250514 "say hello"

You reference --provider opencode but you are also setting an anthropic model? It would be helpful to know exactly what env you set for the opencode provider here.

Additionally, have you checked the sandbox logs for network allowances that need to be made? You can also do that by running openshell term and navigating to the sandbox and seeing if there are network rules that need to be allowed.

[johntmyers]

Hi since we haven't heard back I am going to close this issue.