[Bug]: Remote delegate_task ignores target.driverKind and runs the requested model on the caller's provider instance #34778

Before submitting

• I searched existing issues and did not find a duplicate.
• I included enough detail to reproduce or investigate the problem.

Area

apps/server

Steps to reproduce

1. Have two linked environments, A and B. Both have a Claude provider instance (claudeAgent) and a Codex provider instance (codex).

2. In a Claude thread on A, call delegate_task aimed at B, naming the provider by driver kind only:

   {
     "task": "Reply with the single word: ok. Do not run any tools.",
     "target": {
       "environmentId": "<B>",
       "projectId": "<a project on B>",
       "driverKind": "codex",
       "model": "gpt-6.1-sol",
       "options": { "reasoningEffort": "low" }
     }
   }

3. Look at providerInstanceId in the tool result, or at the child thread on B.

4. Repeat with "providerInstanceId": "codex" in place of "driverKind": "codex".

Expected behavior

driverKind: "codex" resolves to B's Codex instance, as it does for a local delegate, or the call is rejected because the model does not belong to the resolved instance.

Actual behavior

The child thread on B is created on the caller's instance, claudeAgent, with the Codex model and Codex options kept:

step 2 (driverKind: "codex"):          "providerInstanceId":"claudeAgent","model":"gpt-6.1-sol"
step 4 (providerInstanceId: "codex"):  "providerInstanceId":"codex","model":"gpt-6.1-sol"

The stored thread on B has modelSelection: {"instanceId":"claudeAgent","model":"gpt-6.1-sol","options":[{"id":"reasoningEffort","value":"medium"}]}, and the Claude Agent SDK session is opened with model: "gpt-6.1-sol".

Nothing reports the mismatch. With a stock Claude setup the run would presumably fail on an unknown model. Behind an LLM gateway that serves both model families it runs: the GPT model is driven through the Claude Code harness instead of Codex. That is how this was noticed, because Claude Code then warns that auto mode's server-side classifier is unavailable for the session.

The agent that hit this had called orchestrator_capabilities first and passed a provider and model taken from its output, so the call followed the documented usage.

Impact

Major degradation or frequent failure

Version or commit

Caller 0.1.8 → target 0.1.6 (original occurrence, five threads); caller 0.1.6 → another target (minimal repro above)

Environment

macOS on all machines, environments linked over Tailscale HTTPS. Claude Code 2.1.322. Providers: claudeAgent and codex (existing CLI) on both ends.

Logs or stack traces

Provider event log on the target for the child thread:

{"direction":"outgoing","payload":{"type":"query.open","options":{"model":"gpt-6.1-sol","tools":{"type":"preset","preset":"claude_code"},"permissionMode":"auto", ...}}}