Setting ANTHROPIC_BASE_URL (even a pass-through proxy) drops dispatch + message threads and nearly disables parallel tool calls

Summary

When ANTHROPIC_BASE_URL points anywhere other than the default (even a byte-for-byte
pass-through proxy to api.anthropic.com), Claude Code 2.1.282 stops sending two
first-party features, and the model almost stops making parallel tool calls. Setting
_CLAUDE_CODE_ASSUME_FIRST_PARTY_BASE_URL=1 does not restore either feature.

What changes on the wire

With a custom base URL, Claude Code no longer sends:

• the anthropic-dispatch-id: v2d header
• message threads: the thread: {create|continue} request field and the betas

message-threads-2026-08-12 and mid-conversation-system-clear-at-2026-08-21

System prompt, tools, thinking and effort settings are otherwise identical.

Effect (headless claude -p, Sonnet 5, subscription OAuth, same task and repo)

| Condition | Sessions with any parallel tool call | Tool calls per turn |
|---|---|---|
| Default base URL | 6/6 | 1.56 |
| Pass-through proxy base URL + _CLAUDE_CODE_ASSUME_FIRST_PARTY_BASE_URL=1 | 1/7 | 1.03 |
| Same proxy, both features forced back via internal env switches | 4/4 | 1.67 |
| Same proxy, threads only forced back | 0/4 | 1.00 |
| Same proxy, dispatch only forced back | 1/4 | 1.09 |

On a task that reads several files and fetches a URL, the serial sessions need roughly
twice as many requests, so gateway and proxy users pay more for the same work.

Repro

1. Run a local HTTP proxy that forwards every request unchanged to https://api.anthropic.com.
2. Run the same claude -p "<multi-file task>" twice: once with no base URL, once with

ANTHROPIC_BASE_URL=http://127.0.0.1:<port> and _CLAUDE_CODE_ASSUME_FIRST_PARTY_BASE_URL=1.

1. Count assistant messages with more than one tool_use block, and compare the

request headers and bodies captured at the proxy.

Ask

• Gate these features on something other than the base URL, or honour

_CLAUDE_CODE_ASSUME_FIRST_PARTY_BASE_URL=1 for them, so compliant pass-through
proxies and gateways get the same model behaviour.

• For message threads specifically, a gateway that needs full history per request

would benefit from a documented way to keep full-history requests while retaining
the parallel-call behaviour.

Environment

• Claude Code 2.1.282, headless claude -p
• Model: claude-sonnet-5, subscription OAuth
• macOS