[BUG] All subagents fail on Cowork (3P/Vertex AI) — model routing hardcoded to Anthropic API format

Preflight Checklist

• [x] I have searched existing issues and this hasn't been reported yet
• [x] This is a single bug report (please file separate reports for different bugs)
• [x] I am using the latest version of Claude Code

What's Wrong?

All subagent types fail on Cowork when deployed via Vertex AI (3P). Every subagent—regardless of type or model override passed—attempts to use claude-haiku-4-5-20251001 (Anthropic API format) and is rejected before consuming any tokens. The main session (Opus 4.6) works fine.

This appears to be a Cowork-specific issue, separate from the CLI bug in #16594. In the CLI, only Explore and claude-code-guide hardcode Haiku, while Plan and general-purpose inherit from the parent. In Cowork, all four agent types fail with Haiku, suggesting Cowork's subagent layer routes everything through Haiku as a first step—and that routing does not respect the Vertex AI provider's model ID format.

We also tried manually overriding the model name in the 3P config to both:

• claude-haiku-4-5-20251001 (Anthropic API format)

• claude-haiku-4-5@20251001 (Vertex AI format)

Neither worked. The override is not passed through to the subagent layer.

Test Results

| # | Agent Type | Model Attempted | Result | Error |
|---|-------------------|-------------------------------|--------|-------|
| 1 | Explore | claude-haiku-4-5-20251001 | FAILED | "There's an issue with the selected model (claude-haiku-4-5-20251001). It may not exist or you may not have access to it." |
| 2 | claude-code-guide | claude-haiku-4-5-20251001 | FAILED | Same error. 0 tokens, 321ms. |
| 3 | general-purpose | claude-haiku-4-5-20251001 | FAILED | Same error. 0 tokens, 377ms. |
| 4 | Plan | claude-haiku-4-5-20251001 | FAILED | Same error. |

All four failed identically. Zero tokens consumed—failure occurs at model resolution and never reaches the LLM.

What Should Happen?

Subagents should use the same model routing as the main session. If the main session successfully routes to Vertex AI, subagents should too.

Error Messages/Logs

"There's an issue with the selected model (claude-haiku-4-5-20251001). It may not exist or you may not have access to it."

Identical error for all four subagent types. Zero tokens consumed in every attempt — the request fails at model resolution before reaching the LLM.

Steps to Reproduce

1. Deploy Cowork on Vertex AI (3P)
2. Confirm Haiku 4.5 is enabled in the GCP project's Model Garden
3. Ask Claude to spawn any subagent (e.g. "spawn a subagent to answer: what is 2+2?")
4. All subagent types fail with model resolution error

Claude Model

Other

Is this a regression?

No, this never worked

Last Working Version

_No response_

Claude Code Version

Cowork (3P Vertex AI deployment)

Platform

Google Vertex AI

Operating System

macOS

Terminal/Shell

Other

Additional Information

_No response_