v0.8.65: DeepSeek Anthropic-compatible endpoint wire-protocol spike

[Hmbown]

Goal

Spike DeepSeek V4 through DeepSeek's Anthropic-compatible endpoint and compare it with the existing OpenAI Chat Completions path for correctness, token accounting, tool behavior, and web-search/server-tool availability.

This issue is now a focused wire-protocol conformance spike under #3084. It is not a prompt-mode switch and it should not change provider/model routing by inspecting user prompts.

Current Evidence

• Existing CodeWhale provider paths historically leaned on Chat Completions-style request handling.
• DeepSeek's official docs have described an Anthropic-compatible endpoint at https://api.deepseek.com/anthropic.
• DeepSeek V4 docs have described V4-Pro and V4-Flash availability through both OpenAI Chat Completions and Anthropic-compatible interfaces.
• Anthropic's official API includes server tools such as web search, but that does not prove DeepSeek's Anthropic-compatible endpoint supports the same server tools.

References from the original spike:

• 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

Architecture Contract

Endpoint wire protocol is selected by the resolved route, not by prompt content.

The spike should exercise:

• provider descriptor endpoint selection;
• selected-route wire protocol request encoding;
• response/stream decoding;
• tool-call and tool-result mapping;
• reasoning/thinking field mapping;
• stop reason and usage telemetry mapping;
• route explanations and conformance reporting.

Do not conflate Anthropic Claude API capabilities with DeepSeek's Anthropic-compatible endpoint. A compatible request shape does not guarantee identical server tools or telemetry.

Scope

1. Add an opt-in route/endpoint mode for DeepSeek V4 using the Anthropic-compatible wire protocol.
2. Map CodeWhale messages, system prompt, tool definitions, tool results, stop reasons, streaming deltas, thinking/reasoning fields, and usage fields.
3. Smoke test the same small tasks over Chat Completions and Anthropic-compatible endpoints.
4. Explicitly test whether Anthropic-style server tools such as web search are accepted, ignored, rejected, or behave differently through DeepSeek's endpoint.
5. Compare correctness, latency, token usage, cache/reasoning telemetry, and tool reliability.
6. Document the result before changing defaults.

Non-Goals

• Do not make the Anthropic-compatible endpoint 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 server tools in live testing.
• Do not add model-family or provider-string branches outside the descriptor/route/protocol layer.

Acceptance Criteria

• A config flag or provider setting can route DeepSeek V4 through the Anthropic-compatible endpoint without breaking the existing Chat Completions route.
• A comparison report records both paths for the same DeepSeek V4 models available to the test account.
• The report states whether server-side web search works and includes request/response evidence.
• Usage telemetry maps into normalized v0.8.65: Normalize provider usage telemetry for tokens, cache, reasoning, and quota #2961/v0.8.65: Provider/offering usage and PricingSku engine with provenance #3085 fields where possible, with explicit unknown/null for unsupported fields.
• The conclusion is documented: keep experimental, promote as preferred, or reject/defer.

Related

• v0.8.65: Provider descriptors, wire protocol adapters, and conformance tests #3084 provider descriptors and wire protocol adapters.
• v0.8.65: Selected-route reasoning and token-limit wire-protocol mapping #3024 selected-route reasoning and token-limit mapping.
• v0.8.65: Normalize provider usage telemetry for tokens, cache, reasoning, and quota #2961 provider usage telemetry normalization.
• v0.8.65: Make web_search reliable with a SearXNG JSON backend, health checks, and visible agent status #3079 web search reliability.
• v0.8.65 EPIC: Separate provider facts, model facts, offerings, and route resolution #2608 provider/model/offering route architecture.