v0.8.56: Spike DeepSeek V4 over Anthropic Messages API and compare web-search/tool behavior

[Hmbown]

Goal

Add an experimental DeepSeek V4 transport that uses DeepSeek's Anthropic-compatible Messages API instead of the current OpenAI Chat Completions path, then compare behavior, token accounting, tool support, and web-search availability.

This is a spike for a possible simpler provider architecture: DeepSeek V4 through Anthropic Messages format, OpenAI Codex through Responses API, and less reliance on OpenAI Chat Completions if the alternate routes are comparable.

Current evidence

• CodeWhale docs currently say shipped providers use Chat Completions today (docs/PROVIDERS.md, docs/ARCHITECTURE.md).
• DeepSeek's official docs say the Anthropic API format is supported at https://api.deepseek.com/anthropic.
• DeepSeek's V4 update says V4-Pro and V4-Flash are available through both OpenAI ChatCompletions and the Anthropic interface.
• Anthropic's official API docs include server tools such as web_search, where the provider runs web search on its infrastructure.
• It is not yet verified whether DeepSeek's Anthropic-compatible endpoint supports Anthropic server tools like web search, or only the Messages API request/response/tool-call shape.

References:

• DeepSeek Anthropic API docs: https://api-docs.deepseek.com/guides/anthropic_api
• DeepSeek V4 update: https://api-docs.deepseek.com/updates
• Anthropic web search tool docs: https://docs.anthropic.com/en/docs/build-with-claude/tool-use/web-search-tool
• Anthropic tool-use overview: https://docs.anthropic.com/en/docs/build-with-claude/tool-use/overview

Scope

• Add an opt-in provider/request mode for DeepSeek V4 using the Anthropic Messages wire format.
• Map CodeWhale messages, system prompt, tool definitions, tool results, stop reasons, streaming deltas, thinking/reasoning fields, and usage fields to/from Anthropic Messages format.
• Smoke test deepseek-v4-pro and deepseek-v4-flash over both Chat Completions and Anthropic Messages against the same small tasks.
• Explicitly test whether Anthropic server tools such as web_search are accepted, ignored, rejected, or behave differently through DeepSeek's Anthropic endpoint.
• Compare the Anthropic path against current Chat Completions for correctness, latency, token usage, cache/reasoning telemetry, and tool reliability.
• Document the result before changing defaults.

Non-goals

• Do not make Anthropic Messages the default route until smoke tests show it is at least comparable.
• Do not remove Chat Completions support in this issue.
• Do not promise web search support unless DeepSeek's endpoint proves it supports Anthropic server tools in live testing.
• Do not conflate Anthropic's Claude API capabilities with DeepSeek's Anthropic-compatible endpoint; they may differ.

Acceptance criteria

• A config flag or provider setting can route DeepSeek V4 through Anthropic Messages without breaking the existing Chat Completions route.
• A small comparison report records Chat Completions vs Anthropic Messages for deepseek-v4-pro and deepseek-v4-flash.
• The report states whether Anthropic-style web_search works with DeepSeek's Anthropic endpoint and includes the exact request/response evidence.
• Usage telemetry is mapped into the normalized fields from v0.8.56: Align OpenAI Codex provider usage telemetry with Codex CLI #2955/v0.8.57: Normalize cached and reasoning token telemetry across all providers #2961 where possible, with explicit null for unsupported fields.
• The conclusion is documented: keep experimental, promote as preferred, or reject/defer.

Related: #2952, #2955, #2961, #964, #755, #2915