docs(2963): DeepSeek Anthropic-compatible comparison report (decision pending live numbers)

Adds benchmark_results/deepseek-anthropic-comparison-2026-06-24.md for #2963. The deepseek-anthropic / deepseek-claude route already landed (5b8a5ac0b / #3449); this is the reporting deliverable, not code. The report: - documents what landed (route, x-api-key + anthropic-version auth, AnthropicMessages wire format, body/SSE/usage parsing) with file:line cites; - records code-derived findings without live calls: server tools / web search are filtered out on encode (anthropic.rs:361) so that capability is not exercised via this route today, and reasoning_tokens / server_tool_use are always null on the Anthropic usage path vs the Chat-Completions parser; - gives the comparison methodology and a copy-pasteable live checklist for a human with DEEPSEEK_API_KEY to fill in latency/token/correctness numbers; - states the decision honestly: keep as Experimental, keep-vs-promote PENDING the live numbers. No verified verdict is fabricated.
2026-09-03 06:50:13 +08:00 · 2026-06-24 01:28:23 -07:00
parent 6ab9f72954
commit 35d50878bf
1 changed files with 524 additions and 0 deletions
--- a/benchmark_results/deepseek-anthropic-comparison-2026-06-24.md
+++ b/benchmark_results/deepseek-anthropic-comparison-2026-06-24.md
@@ -0,0 +1,524 @@
+# DeepSeek Anthropic-Compatible Endpoint — Comparison Report & Decision (#2963)
+
+- **Issue:** [#2963](https://github.com/Hmbown/CodeWhale/issues/2963) — *v0.8.65: DeepSeek Anthropic-compatible endpoint wire-protocol spike*
+- **Release lane:** v0.8.65
+- **Date:** 2026-06-24
+- **Status:** Implementation **landed and Experimental**. Keep-vs-promote decision **PENDING live numbers** (Section 4).
+- **Scope of this document:** A *report*. It changes no Rust code and makes no live API calls — no DeepSeek credentials are available in this environment, so all live figures below are left as a checklist for a human operator to fill in.
+
+> Do **not** reimplement the route. It already exists on `main` (commit
+> `5b8a5ac0b2c478261740f49756d29c4a7f83d89c`, PR
+> [#3449](https://github.com/Hmbown/CodeWhale/pull/3449)). This document
+> verifies what landed, derives what can be concluded from the code without a
+> network, and specifies the exact live procedure to settle the open question.
+
+All file:line citations below are against the tree at this report's commit
+(verified ancestry: `5b8a5ac0b` is an ancestor of `HEAD`).
+
+---
+
+## 1. What's landed
+
+The opt-in DeepSeek route that speaks the **Anthropic Messages** wire protocol
+is implemented end to end. **It is already in `main`; do not re-implement it.**
+
+### 1.1 Provider descriptor / route selection
+
+- `crates/config/src/provider.rs:140-178` — `DeepseekAnthropic` provider:
+  - id `deepseek-anthropic` (`provider.rs:143-145`)
+  - display name `DeepSeek (Anthropic-compatible)` (`provider.rs:151-153`)
+  - aliases `deepseek_anthropic`, `deepseek-claude`, `deepseek_claude`
+    (`provider.rs:171-173`)
+  - **wire format `WireFormat::AnthropicMessages`** (`provider.rs:175-177`)
+  - API-key env var: **`DEEPSEEK_API_KEY` only** (`provider.rs:163-165`) — it
+    does **not** fall back to `ANTHROPIC_API_KEY`.
+- `crates/config/src/provider.rs:31-38` — `WireFormat` enum
+  (`ChatCompletions` / `Responses` / `AnthropicMessages`).
+- Registry wiring: static entry `provider.rs:544`, registered at
+  `provider.rs:573`.
+- Defaults (`crates/config/src/provider_defaults.rs`):
+  - base URL `https://api.deepseek.com/anthropic`
+    (`provider_defaults.rs:14`)
+  - default model `deepseek-v4-pro` — `DEFAULT_DEEPSEEK_ANTHROPIC_MODEL`
+    aliases `DEFAULT_DEEPSEEK_MODEL` (`provider_defaults.rs:8-9`)
+- The Chat-Completions DeepSeek route, for contrast, defaults to base URL
+  `https://api.deepseek.com/beta` (`provider_defaults.rs:13`) with the same
+  default model `deepseek-v4-pro`.
+
+### 1.2 Dispatch
+
+- `crates/tui/src/client.rs:1331-1339` (`create_message`) and
+  `client.rs:1341-1352` (`create_message_stream`) route to the Anthropic
+  adapter when `api_provider_uses_anthropic_messages(self.api_provider)` is
+  true.
+- `client.rs:864-869` — `api_provider_uses_anthropic_messages` returns true for
+  `ApiProvider::Anthropic | ApiProvider::DeepseekAnthropic`.
+- Request payload mode is selected by route, not prompt:
+  `crates/tui/src/config.rs:526-530` sets
+  `RequestPayloadMode::AnthropicMessages` for `DeepseekAnthropic`, else
+  `ChatCompletions`.
+
+### 1.3 Auth dialect
+
+- `crates/tui/src/client.rs:805-838` builds headers:
+  - injects `anthropic-version: 2023-06-01` for Anthropic-wire providers
+    (`client.rs:808-815`)
+  - uses **`x-api-key`** (never `Authorization: Bearer`) for the API key
+    (`client.rs:817-819`, applied `client.rs:831-837`)
+- `client.rs:846-862` strips any caller-supplied `Authorization` / `api-key` /
+  `x-api-key` extra headers so a stale OpenAI-style auth header cannot leak onto
+  the Anthropic wire (`is_auth_dialect_header`, `client.rs:858-862`).
+- Tests: `deepseek_anthropic_uses_anthropic_header_dialect`
+  (`client.rs:2216`+) asserts `x-api-key` + `anthropic-version` are present and
+  that Bearer / MiMo headers are absent.
+
+### 1.4 Request encoding (Messages body)
+
+- `crates/tui/src/client/anthropic.rs:40-143` — `build_anthropic_body`:
+  - `model` / `max_tokens` / `stream` (`anthropic.rs:41-45`)
+  - `system` as text or cache-aware blocks (`anthropic.rs:47-66`)
+  - `messages` via `message_to_anthropic` (`anthropic.rs:68-74`,
+    `anthropic.rs:291-301`)
+  - `tools` with `strict` + `cache_control` (`anthropic.rs:76-98`)
+  - `tool_choice` mapped from OpenAI-style string/object to Anthropic object
+    form (`anthropic.rs:100-102`, `anthropic.rs:279-287`)
+  - reasoning → `thinking: {type: adaptive}` + `output_config.effort`
+    (low/medium/high/max), gated on `model_supports_reasoning`
+    (`anthropic.rs:104-128`)
+  - sampling-parameter rules: send at most one of temperature/top_p, or neither
+    for models that reject them (`anthropic.rs:130-139`,
+    `anthropic.rs:269-275`)
+  - `cache_control` breakpoint placement, capped at 4
+    (`anthropic.rs:141`, `anthropic.rs:367-446`)
+- Endpoint URL builder tolerates a `/v1` suffix
+  (`anthropic.rs:259-266`); `https://api.deepseek.com/anthropic` →
+  `…/anthropic/v1/messages`.
+
+### 1.5 Response & stream decoding
+
+- Non-streaming: `anthropic.rs:240-254` (`handle_anthropic_message`) parses the
+  JSON body and normalizes `usage`.
+- Streaming: `anthropic.rs:170-237` (`handle_anthropic_stream`) is an SSE
+  pass-through; `convert_anthropic_sse_data` (`anthropic.rs:450-494`) decodes
+  `message_start` / `content_block_*` / `message_delta` / `message_stop` /
+  `ping` / `error`, tolerates unknown event types, and normalizes usage on
+  `message_start` / `message_delta`.
+- Send/error path: `anthropic.rs:145-167` (`send_anthropic_request`) sets
+  `Accept: text/event-stream`, maps non-2xx into a typed error via
+  `parse_anthropic_error_envelope` (`anthropic.rs:528-548`).
+
+### 1.6 Usage / cache normalization (#2961 convention)
+
+- `anthropic.rs:499-523` (`parse_anthropic_usage`):
+  - `prompt_cache_hit_tokens = cache_read_input_tokens`
+  - `prompt_cache_miss_tokens = input_tokens + cache_creation_input_tokens`
+  - normalized `input_tokens = input_tokens + cache_creation + cache_read`
+    (total prompt — the DeepSeek convention)
+
+### 1.7 Operational guardrails added with the route
+
+- Health check **skips the `/anthropic/v1/models` probe** for this route
+  (`client.rs:871-873`, `api_provider_skips_models_probe`); test
+  `deepseek_anthropic_health_check_skips_models_probe` (`client.rs:2301`+).
+- **FIM is unsupported** on this route and fails locally with a clear message
+  (`client.rs:1722-1727`); test `deepseek_anthropic_fim_fails_without_http_request`
+  (`client.rs:2314`+).
+- Base-URL env override is route-aware: `CODEWHALE_BASE_URL` / `DEEPSEEK_BASE_URL`
+  writes into `providers.deepseek_anthropic.base_url`
+  (`crates/tui/src/config.rs:3928-3939`).
+- Translation helper uses the Messages endpoint for this provider
+  (`client.rs:974-977`); test
+  `deepseek_anthropic_translate_uses_messages_endpoint` (`client.rs:2251`+).
+
+### 1.8 Docs framing
+
+- `docs/PROVIDERS.md:48-51`, `:81`, `:111-112`, `:237` document the route as
+  **Anthropic *wire-protocol* compatibility** (not Anthropic model/provider
+  semantics), list the aliases, and state "Keep `provider = "deepseek"` for the
+  default Chat Completions path."
+
+### 1.9 Test coverage already present (no live calls)
+
+In `crates/tui/src/client/anthropic.rs` `#[cfg(test)]` (from `anthropic.rs:550`):
+body cache-control placement, reasoning→effort mapping, sampling-param dropping,
+signed/unsigned thinking replay, breakpoint cap, full SSE fixture decode
+(text + thinking + signature + tool_use + usage), error/unknown-event handling,
+usage mapping with missing cache fields, error-envelope parsing, URL `/v1`
+tolerance. In `crates/tui/src/client.rs`: the auth-dialect, models-probe-skip,
+translate-endpoint, and FIM-unsupported tests cited above.
+
+---
+
+## 2. Code-derived findings (no live calls needed)
+
+These are behavioral facts that can be stated **from the code today**, before
+any live comparison. They are the deltas a reviewer most needs to know.
+
+### 2.1 Server tools / web search are NOT exercised via this route today
+
+`content_block_to_anthropic` **drops** the server-tool block types on encode:
+
+```
+crates/tui/src/client/anthropic.rs:359-364
+    // Server-tool block types are DeepSeek/internal concepts with no
+    // Anthropic client-side wire equivalent.
+    ContentBlock::ServerToolUse { .. }
+    | ContentBlock::ToolSearchToolResult { .. }
+    | ContentBlock::CodeExecutionToolResult { .. } => None,
+```
+
+Consequence: any server-tool / web-search content the engine holds is filtered
+out before the request is sent on this route. There is also no encode-side path
+that *injects* an Anthropic-style server-tool definition (e.g. a `web_search`
+tool) into the outbound body — `build_anthropic_body` only forwards
+caller-supplied client tools (`anthropic.rs:76-98`). So **server-side web
+search / code execution is not exercised through the DeepSeek Anthropic route as
+implemented.** Whether DeepSeek's endpoint would *accept* such a tool is a
+separate, still-open question that only live testing (Section 4, Test E) can
+answer; the code neither offers nor depends on it.
+
+### 2.2 Usage telemetry: two real deltas vs the Chat-Completions path
+
+Compare the two usage parsers:
+
+| Field | Anthropic route (`anthropic.rs:499-523`) | Chat-Completions route (`client.rs:1643-1711`) |
+|---|---|---|
+| `input_tokens` (normalized) | `input + cache_creation + cache_read` | `input_tokens`/`prompt_tokens` as-is |
+| `prompt_cache_hit_tokens` | `cache_read_input_tokens` | `prompt_cache_hit_tokens`, else `prompt_tokens_details.cached_tokens` |
+| `prompt_cache_miss_tokens` | `input + cache_creation` | `prompt_cache_miss_tokens`, else `input − hit` |
+| `reasoning_tokens` | **always `None`** (`anthropic.rs:519`) | parsed from `completion_tokens_details.reasoning_tokens` (`client.rs:1658-1685`) |
+| `reasoning_replay_tokens` | `None` (`anthropic.rs:520`) | `None` (`client.rs:1708`) |
+| `server_tool_use` | **always `None`** (`anthropic.rs:521`) | parsed from `server_tool_use.{code_execution,tool_search}_requests` (`client.rs:1687-1700`) |
+| `output_tokens` | Anthropic `output_tokens` | `output_tokens`/`completion_tokens`, with fallbacks to reasoning or `total − input` (`client.rs:1648-1670`) |
+
+Two concrete deltas to record honestly in any telemetry comparison:
+
+1. **`reasoning_tokens` is never populated on the Anthropic route.** Reasoning
+   *content* still flows (thinking blocks decode and signed blocks replay —
+   `anthropic.rs:315-330`, `anthropic.rs:822-868` fixture), but the **count**
+   is dropped. On the Chat-Completions route the count is read from
+   `completion_tokens_details.reasoning_tokens`. This is per the #2961/#3085
+   "explicit unknown/null for unsupported fields" rule, but it means
+   reasoning-token *accounting parity* between the two routes cannot be
+   expected — confirm in Test C.
+2. **`server_tool_use` is never populated on the Anthropic route** (consistent
+   with §2.1: the route doesn't drive server tools).
+
+### 2.3 Thinking / reasoning request shaping differs by design
+
+The Anthropic route maps `reasoning_effort` tiers to
+`thinking: {type: adaptive}` + `output_config.effort`
+(`anthropic.rs:104-128`), gated on `model_supports_reasoning`. The
+Chat-Completions DeepSeek path uses its own reasoning-split / payload
+conventions (`config.rs:526-530` selects the payload mode; DeepSeek-family
+reasoning handling lives on the Chat path). Equivalent *output* is the bar to
+test (Section 3/4), not byte-identical requests.
+
+### 2.4 Caching model differs in shape
+
+The Anthropic route places explicit `cache_control` breakpoints (max 4) on the
+prefix and latest user turn (`anthropic.rs:367-446`) and reports cache
+hit/miss from Anthropic's `cache_read` / `cache_creation` fields. The
+Chat-Completions route relies on DeepSeek's automatic prefix caching and reads
+`prompt_cache_hit_tokens` / `prompt_cache_miss_tokens` (or
+`prompt_tokens_details.cached_tokens`). Both normalize into the same #2961
+fields, so cache *telemetry* is comparable even though the *mechanism* differs.
+
+### 2.5 Capability/operational deltas (route-level, from code)
+
+- **FIM**: supported on Chat-Completions DeepSeek; **unsupported** on the
+  Anthropic route (`client.rs:1722-1727`).
+- **Models probe**: skipped on the Anthropic route (`client.rs:871-873`); the
+  Chat path probes `/models`.
+- **Auth**: `x-api-key` + `anthropic-version` (Anthropic route) vs
+  `Authorization: Bearer` (Chat route) — `client.rs:817-827`.
+- **Endpoint**: `…/anthropic/v1/messages` vs `…/beta` chat completions.
+
+### 2.6 What is *equivalent* by construction
+
+Tool-call and tool-result mapping, image blocks, system prompt, and stop
+reasons all have direct encoders (`anthropic.rs:303-358`) and the SSE decoder
+reconstructs tool-use input JSON (fixture `anthropic.rs:816-897`). So for an
+ordinary "prompt → text / tool_use" exchange, the two routes are expected to be
+functionally equivalent; the open questions are the *quantitative* ones
+(latency, token counts) and the *server-tool* one.
+
+---
+
+## 3. Comparison methodology
+
+Compare DeepSeek's **Chat-Completions** route (`provider = "deepseek"`) against
+its **Anthropic-Messages** route (`provider = "deepseek-anthropic"`) for the
+**same model** (`deepseek-v4-pro`, and `deepseek-v4-flash` if the account has
+it). Hold everything else constant (same prompt, same `max_tokens`, same
+reasoning effort, same temperature where accepted).
+
+Dimensions:
+
+1. **Correctness / output equivalence** — same prompt → semantically equivalent
+   answer; same tool selection and arguments for a tool-use prompt; valid JSON
+   for a structured prompt.
+2. **Latency** — wall-clock total and (for streaming) time-to-first-token, over
+   N≥5 runs each; report median + spread, not a single sample.
+3. **Token / usage accounting parity** — compare `input_tokens` (normalized),
+   `output_tokens`, `prompt_cache_hit_tokens`, `prompt_cache_miss_tokens`,
+   `reasoning_tokens`. **Expect `reasoning_tokens` to be null on the Anthropic
+   route** (§2.2) — record it, don't treat it as a bug.
+4. **Telemetry fields** — which of the #2961/#3085 normalized fields are
+   populated vs null on each route; note `server_tool_use` is null on the
+   Anthropic route by construction.
+5. **Server-tool / web-search support** — does DeepSeek's Anthropic endpoint
+   *accept*, *ignore*, or *reject* an Anthropic-style server tool (e.g.
+   `web_search`)? Capture the raw request/response. (Recall the engine does not
+   send such a tool today — §2.1 — so this is an endpoint-capability probe with
+   a hand-built request, not a test of CodeWhale's encoder.)
+6. **Error envelopes & rate limiting** — confirm 4xx/5xx map cleanly
+   (`anthropic.rs:528-548`) and that the route honors the same retry/backoff.
+
+Pass bar for "comparable" (issue Acceptance Criteria): equivalent correctness on
+the smoke tasks, latency within a reasonable band, and usage telemetry that maps
+into the normalized fields (with explicit nulls where unsupported).
+
+---
+
+## 4. Runnable live checklist (human, with `DEEPSEEK_API_KEY` set)
+
+All commands are copy-pasteable. They assume the repo root and a DeepSeek key.
+**No credentials exist in this environment; these are for a human to run.**
+
+### 4.0 One-time setup
+
+```bash
+export DEEPSEEK_API_KEY="sk-..."           # your DeepSeek key
+MODEL="deepseek-v4-pro"                      # also repeat with deepseek-v4-flash if available
+CHAT_BASE="https://api.deepseek.com"        # Chat Completions (OpenAI-compatible)
+ANTH_BASE="https://api.deepseek.com/anthropic"  # Anthropic Messages
+mkdir -p benchmark_results/2963-live && cd "$(git rev-parse --show-toplevel)"
+```
+
+### Test A — correctness, single turn (text)
+
+Chat Completions:
+
+```bash
+curl -sS -w '\n[http %{http_code} | total %{time_total}s | ttfb %{time_starttransfer}s]\n' \
+  -X POST "$CHAT_BASE/v1/chat/completions" \
+  -H "Authorization: Bearer $DEEPSEEK_API_KEY" \
+  -H "Content-Type: application/json" \
+  -d "{\"model\":\"$MODEL\",\"max_tokens\":64,\"stream\":false,
+       \"messages\":[{\"role\":\"user\",\"content\":\"Reply with exactly the word: PONG\"}]}" \
+  | tee benchmark_results/2963-live/A_chat.json
+```
+
+Anthropic Messages (note `x-api-key` + `anthropic-version`, no Bearer):
+
+```bash
+curl -sS -w '\n[http %{http_code} | total %{time_total}s | ttfb %{time_starttransfer}s]\n' \
+  -X POST "$ANTH_BASE/v1/messages" \
+  -H "x-api-key: $DEEPSEEK_API_KEY" \
+  -H "anthropic-version: 2023-06-01" \
+  -H "Content-Type: application/json" \
+  -d "{\"model\":\"$MODEL\",\"max_tokens\":64,\"stream\":false,
+       \"messages\":[{\"role\":\"user\",\"content\":\"Reply with exactly the word: PONG\"}]}" \
+  | tee benchmark_results/2963-live/A_anthropic.json
+```
+
+Record: does each return "PONG"? HTTP status, total time.
+
+### Test B — usage / token accounting (read the `usage` object on both)
+
+```bash
+echo "Chat usage:";      jq '.usage'  benchmark_results/2963-live/A_chat.json
+echo "Anthropic usage:"; jq '.usage'  benchmark_results/2963-live/A_anthropic.json
+```
+
+Fill in the table:
+
+| Field | Chat Completions | Anthropic Messages |
+|---|---|---|
+| prompt/input tokens | | |
+| completion/output tokens | | |
+| cache hit (`prompt_cache_hit_tokens` / `cache_read_input_tokens`) | | |
+| cache miss (`prompt_cache_miss_tokens` / `cache_creation_input_tokens`) | | |
+| reasoning tokens (`completion_tokens_details.reasoning_tokens`) | | (expected absent) |
+
+### Test C — reasoning / thinking
+
+Chat Completions (DeepSeek reasoner-style):
+
+```bash
+curl -sS -X POST "$CHAT_BASE/v1/chat/completions" \
+  -H "Authorization: Bearer $DEEPSEEK_API_KEY" -H "Content-Type: application/json" \
+  -d "{\"model\":\"$MODEL\",\"max_tokens\":512,\"stream\":false,
+       \"messages\":[{\"role\":\"user\",\"content\":\"A bat and ball cost \$1.10. The bat costs \$1 more than the ball. How much is the ball? Think, then answer.\"}]}" \
+  | tee benchmark_results/2963-live/C_chat.json | jq '{content:.choices[0].message, usage:.usage}'
+```
+
+Anthropic Messages with adaptive thinking:
+
+```bash
+curl -sS -X POST "$ANTH_BASE/v1/messages" \
+  -H "x-api-key: $DEEPSEEK_API_KEY" -H "anthropic-version: 2023-06-01" -H "Content-Type: application/json" \
+  -d "{\"model\":\"$MODEL\",\"max_tokens\":512,\"stream\":false,
+       \"thinking\":{\"type\":\"adaptive\"},\"output_config\":{\"effort\":\"high\"},
+       \"messages\":[{\"role\":\"user\",\"content\":\"A bat and ball cost \$1.10. The bat costs \$1 more than the ball. How much is the ball? Think, then answer.\"}]}" \
+  | tee benchmark_results/2963-live/C_anthropic.json | jq '{content:.content, usage:.usage}'
+```
+
+Record: both should answer **\$0.05**. Note whether a `thinking` block is
+returned by the Anthropic route and whether reasoning tokens appear anywhere.
+
+### Test D — tool use (same tool both routes)
+
+Chat Completions:
+
+```bash
+curl -sS -X POST "$CHAT_BASE/v1/chat/completions" \
+  -H "Authorization: Bearer $DEEPSEEK_API_KEY" -H "Content-Type: application/json" \
+  -d "{\"model\":\"$MODEL\",\"max_tokens\":256,
+       \"tools\":[{\"type\":\"function\",\"function\":{\"name\":\"get_weather\",
+         \"description\":\"Get weather for a city\",
+         \"parameters\":{\"type\":\"object\",\"properties\":{\"city\":{\"type\":\"string\"}},\"required\":[\"city\"]}}}],
+       \"messages\":[{\"role\":\"user\",\"content\":\"What's the weather in Paris? Use the tool.\"}]}" \
+  | tee benchmark_results/2963-live/D_chat.json | jq '.choices[0].message.tool_calls'
+```
+
+Anthropic Messages:
+
+```bash
+curl -sS -X POST "$ANTH_BASE/v1/messages" \
+  -H "x-api-key: $DEEPSEEK_API_KEY" -H "anthropic-version: 2023-06-01" -H "Content-Type: application/json" \
+  -d "{\"model\":\"$MODEL\",\"max_tokens\":256,
+       \"tools\":[{\"name\":\"get_weather\",\"description\":\"Get weather for a city\",
+         \"input_schema\":{\"type\":\"object\",\"properties\":{\"city\":{\"type\":\"string\"}},\"required\":[\"city\"]}}],
+       \"messages\":[{\"role\":\"user\",\"content\":\"What's the weather in Paris? Use the tool.\"}]}" \
+  | tee benchmark_results/2963-live/D_anthropic.json | jq '.content'
+```
+
+Record: does each emit a `get_weather` call with `city = "Paris"`?
+
+### Test E — server-tool / web-search capability probe (the open question)
+
+Send an Anthropic-style server tool and **record whether DeepSeek accepts,
+ignores, or rejects it** (capture the full body). The engine does not send this
+today (§2.1); this is a raw endpoint probe.
+
+```bash
+curl -sS -w '\n[http %{http_code}]\n' -X POST "$ANTH_BASE/v1/messages" \
+  -H "x-api-key: $DEEPSEEK_API_KEY" -H "anthropic-version: 2023-06-01" -H "Content-Type: application/json" \
+  -d "{\"model\":\"$MODEL\",\"max_tokens\":256,
+       \"tools\":[{\"type\":\"web_search_20250305\",\"name\":\"web_search\",\"max_uses\":2}],
+       \"messages\":[{\"role\":\"user\",\"content\":\"Search the web: what is the latest stable Rust version? Cite a source.\"}]}" \
+  | tee benchmark_results/2963-live/E_websearch.json
+```
+
+Classify the outcome:
+- **Accepted + worked** — response contains server-tool-use / search results.
+- **Ignored** — 200 OK, plain answer, no tool activity.
+- **Rejected** — 4xx with an error envelope (record `error.type` / message).
+
+### Test F — streaming smoke (both routes)
+
+```bash
+# Anthropic SSE
+curl -N -sS -X POST "$ANTH_BASE/v1/messages" \
+  -H "x-api-key: $DEEPSEEK_API_KEY" -H "anthropic-version: 2023-06-01" \
+  -H "Accept: text/event-stream" -H "Content-Type: application/json" \
+  -d "{\"model\":\"$MODEL\",\"max_tokens\":64,\"stream\":true,
+       \"messages\":[{\"role\":\"user\",\"content\":\"Count: one two three\"}]}" \
+  | tee benchmark_results/2963-live/F_anthropic.sse | head -40
+```
+
+Confirm `message_start` → `content_block_*` → `message_delta` → `message_stop`
+arrive (the shapes `convert_anthropic_sse_data` decodes, `anthropic.rs:450-494`).
+
+### Test G — end-to-end through CodeWhale (optional, exercises the real adapter)
+
+```bash
+# Anthropic route through the built binary
+cargo run -q -p codewhale -- --provider deepseek-anthropic --model "$MODEL" \
+  --print "Reply with exactly: PONG"
+# Chat route for comparison
+cargo run -q -p codewhale -- --provider deepseek --model "$MODEL" \
+  --print "Reply with exactly: PONG"
+```
+
+(Adjust the binary/flag names to the project's actual non-interactive entry
+point if different; the point is to run one prompt through each resolved route.)
+
+### 4.1 Results table to fill in
+
+| Dimension | Chat Completions | Anthropic Messages | Verdict |
+|---|---|---|---|
+| Correctness (A/C/D) | | | |
+| Latency median (N=…) | | | |
+| TTFT (streaming) | | | |
+| Token accounting (B) | | | |
+| reasoning_tokens present | | (expected no) | |
+| Tool use (D) | | | |
+| Web search (E) | n/a | accept / ignore / reject | |
+| Streaming (F) | | | |
+
+---
+
+## 5. Decision
+
+**Recommendation: KEEP as Experimental. The keep-vs-promote decision is PENDING
+the live numbers in Section 4. This report does not assert a "verified" verdict
+because no live calls were made.**
+
+Rationale, grounded in code:
+
+- **Keep (not reject):** the route is fully implemented, isolated behind opt-in
+  provider selection (`deepseek-anthropic` / `deepseek-claude`), guarded
+  (FIM-unsupported message, models-probe skip, auth-header hygiene), and covered
+  by unit + SSE-fixture tests. It does not touch or regress the default
+  Chat-Completions DeepSeek path (separate dispatch at `client.rs:1331-1352`;
+  docs say keep `provider = "deepseek"` for the default). Nothing in the code
+  argues for ripping it out.
+- **Do not promote yet:** the issue's promotion bar requires the Anthropic route
+  to be *at least comparable* on a live A/B, plus explicit server-tool evidence.
+  That evidence does not exist here. Two code-derived caveats that promotion
+  must weigh: (a) `reasoning_tokens` accounting is dropped on this route
+  (§2.2 #1), and (b) server tools / web search are not exercised through it
+  (§2.1) — so if web search is a requirement for "preferred," this route does
+  not satisfy it today regardless of what Test E shows about the endpoint.
+- **Gate to flip the decision:** complete Section 4 (especially Tests A–E),
+  fill the §4.1 table, and confirm equivalent correctness + comparable latency +
+  clean telemetry mapping. If all green and web search is not a blocker →
+  candidate to promote to preferred for DeepSeek V4. Otherwise → remain
+  Experimental, or reject the *promotion* (not the route) if telemetry/latency
+  regress.
+
+### Suggested issue note (after live numbers are in)
+
+> Implementation verified landed (#3449 / `5b8a5ac0b`); see
+> `benchmark_results/deepseek-anthropic-comparison-2026-06-24.md`. Live A/B
+> results: [fill in]. Server-tool/web-search probe (Test E): [accept/ignore/
+> reject + evidence]. Decision: [keep experimental | promote to preferred].
+
+---
+
+## Appendix — citation index
+
+| Topic | Location |
+|---|---|
+| `WireFormat` enum | `crates/config/src/provider.rs:31-38` |
+| `DeepseekAnthropic` descriptor | `crates/config/src/provider.rs:140-178` |
+| Registry entry | `crates/config/src/provider.rs:544`, `:573` |
+| Base URL / model defaults | `crates/config/src/provider_defaults.rs:8-9,13-14` |
+| Dispatch to Anthropic adapter | `crates/tui/src/client.rs:1331-1352` |
+| `api_provider_uses_anthropic_messages` | `crates/tui/src/client.rs:864-869` |
+| Auth header build (`x-api-key`/`anthropic-version`) | `crates/tui/src/client.rs:805-862` |
+| Models-probe skip | `crates/tui/src/client.rs:871-873` |
+| FIM unsupported | `crates/tui/src/client.rs:1722-1727` |
+| Chat-Completions usage parser | `crates/tui/src/client.rs:1643-1711` |
+| Base-URL env override (route-aware) | `crates/tui/src/config.rs:3928-3939` |
+| Payload-mode selection | `crates/tui/src/config.rs:526-530` |
+| `build_anthropic_body` | `crates/tui/src/client/anthropic.rs:40-143` |
+| Messages URL builder | `crates/tui/src/client/anthropic.rs:259-266` |
+| **Server-tool blocks dropped on encode** | `crates/tui/src/client/anthropic.rs:359-364` |
+| Anthropic usage normalizer | `crates/tui/src/client/anthropic.rs:499-523` |
+| Error-envelope parser | `crates/tui/src/client/anthropic.rs:528-548` |
+| Docs framing | `docs/PROVIDERS.md:48-51,81,111-112,237` |
+| Landed commit / PR | `5b8a5ac0b2c478261740f49756d29c4a7f83d89c` / [#3449](https://github.com/Hmbown/CodeWhale/pull/3449) |