Claude finished @Hmbown's task in 4m 27s —— View job

Code Review — DeepSeek Anthropic-compatible route

• Gather context and read changed files
• Review crates/config/src/provider.rs — provider metadata, env vars, aliases
• Review crates/config/src/lib.rs — EnvRuntimeOverrides
• Review crates/tui/src/client.rs — routing, header dialect
• Review crates/tui/src/config.rs — subagent key matching, capability
• Review remaining files (tests, docs, scripts)
• Summarize findings

Summary

The core routing is well-structured: api_provider_uses_anthropic_messages predicate correctly gates both create_message / create_message_stream, the x-api-key + anthropic-version header dialect, and the reasoning-effort no-op. URL construction for the messages endpoint (https://api.deepseek.com/anthropic/v1/messages) is correct. Three Gemini findings are false positives in the current code. Two correctness bugs and one footgun need attention.

Correctness Bugs

1. translate() uses the wrong endpoint for DeepseekAnthropic — HIGH

crates/tui/src/client.rs — translate() always calls api_url_with_suffix(..., "chat/completions", ...) regardless of provider. For DeepseekAnthropic with base URL https://api.deepseek.com/anthropic, this resolves to https://api.deepseek.com/anthropic/v1/chat/completions, which does not exist on DeepSeek's Anthropic-compatible path. Compare create_message and create_message_stream, which both guard with:

if api_provider_uses_anthropic_messages(self.api_provider) {
    return self.handle_anthropic_message(request).await;
}

translate is missing an equivalent guard. A user who enables translation with deepseek-anthropic as the active provider will hit an HTTP error on every translation call. Fix this →

2. health_check() and maybe_probe_recovery() probe a non-existent endpoint — MEDIUM

Both methods call api_url(&self.base_url, "models"). For DeepseekAnthropic with base URL https://api.deepseek.com/anthropic, this resolves to https://api.deepseek.com/anthropic/v1/models — a path not documented in DeepSeek's Anthropic-compatible API. If this returns a non-2xx status (which is likely), mark_request_failure is called and the connection enters a degraded state. Once degraded, maybe_probe_recovery fires on every request and keeps probing the same wrong URL, keeping the connection permanently degraded.

The existing Anthropic provider has the same exposure (Anthropic's API does expose GET /v1/models, and DeepSeek may or may not mirror it on the /anthropic path). Worth adding an explicit health-check path for api_provider_uses_anthropic_messages providers that either skips the check or hits /v1/messages with a minimal probe body. Fix this →

3. fim_completion() has no guard for DeepseekAnthropic — LOW

fim_completion calls the beta/completions FIM endpoint:

let url = api_url_with_suffix(&self.base_url, "beta/completions", None);

For DeepseekAnthropic this hits https://api.deepseek.com/beta/completions, not the Anthropic path at all. The method should bail! early with a clear message if api_provider_uses_anthropic_messages(self.api_provider).

Minor Issues

4. TUI ProvidersConfig missing deepseekAnthropic camelCase alias

crates/config/src/lib.rs:325-330:

alias = "deepseek-anthropic",
alias = "deepseekAnthropic",   // ← present in crates/config
alias = "deepseek-claude",
alias = "deepseek_claude"

crates/tui/src/config.rs:2663-2668:

alias = "deepseek-anthropic",
// ← deepseekAnthropic is missing in crates/tui
alias = "deepseek-claude",
alias = "deepseek_claude"

Low impact (camelCase TOML keys are uncommon), but the two config structs should stay in sync.

Gemini False Positives

All three findings flagged by Gemini are not present in the current code:

• ANTHROPIC_API_KEY in env_vars — DeepseekAnthropic.env_vars() is already &["DEEPSEEK_API_KEY"] only (provider.rs:163-165). No leakage risk.

• Missing hyphenated aliases in subagent_provider_key_matches — normalize_subagent_provider_key maps both - and _ to _, so "deepseek-anthropic" normalizes to "deepseek_anthropic" before the matches! check. The hyphenated forms are already handled via the normalization path at config.rs:488-496.

• EnvRuntimeOverrides base URL gap — deepseek_anthropic_base_url is a struct field (lib.rs:4307) loaded from DEEPSEEK_ANTHROPIC_BASE_URL / DEEPSEEK_CLAUDE_BASE_URL (lib.rs:4429-4432) and wired into base_url_for (lib.rs:4583). The override path is fully implemented.

Good Choices

• Re-using handle_anthropic_message / handle_anthropic_stream through api_provider_uses_anthropic_messages is clean and avoids protocol-specific hacks in the shared dispatch path.
• anthropic_messages_url("https://api.deepseek.com/anthropic") correctly produces https://api.deepseek.com/anthropic/v1/messages, matching DeepSeek's documented Anthropic-compatible path.
• Auth header routing — x-api-key + anthropic-version applied via the existing predicate, and the ANTHROPIC_API_KEY / DEEPSEEK_API_KEY separation in env_vars is correct.
• wire_model_for_provider for DeepseekAnthropic strips the deepseek-ai/ or deepseek/ namespace prefix, so the wire model sent to DeepSeek is the bare deepseek-v4-pro the Anthropic-compatible endpoint expects.
• Scoping DeepSeek-native cache telemetry to existing routes while marking DeepseekAnthropic as AnthropicMessages in the capability matrix is the right call.

Claude finished @Hmbown's task in 3m 3s —— View job

Code Review — DeepSeek Anthropic-compatible route (updated)

• Gather context and understand the PR
• Read crates/tui/src/client.rs — check translate(), health_check(), fim_completion() guards
• Read crates/tui/src/config.rs — check deepseekAnthropic alias
• Read crates/config/src/provider.rs — verify env_vars, aliases
• Summarize findings

Summary

All four correctness bugs from the first review pass are fixed. The implementation is clean and ready for merge. Details below.

All Prior Findings Resolved

1. translate() endpoint guard — FIXED

client.rs:949-954 now delegates to handle_anthropic_message when api_provider_uses_anthropic_messages is true, before falling through to the chat/completions path.

2. health_check() / maybe_probe_recovery() /models probe — FIXED

A new factored predicate api_provider_skips_models_probe (line 845-847) returns true for DeepseekAnthropic. Both health_check (line 1278) and maybe_probe_recovery (line 1143) guard on it, mark success, and return early — no HTTP call to the non-existent /anthropic/v1/models path.

3. fim_completion() guard — FIXED

client.rs:1696-1700 now bails with a clear diagnostic:

bail!("FIM completion is not supported for {} because it uses the Anthropic Messages protocol", ...)

4. deepseekAnthropic camelCase alias — FIXED

config.rs:2663-2666 now has all four aliases in sync with crates/config:

alias = "deepseek-anthropic",
alias = "deepseekAnthropic",
alias = "deepseek-claude",
alias = "deepseek_claude"

Gemini False Positives — Confirmed No-ops

• ANTHROPIC_API_KEY in env_vars — provider.rs:163-165 is &["DEEPSEEK_API_KEY"] only. No leakage risk.
• Missing hyphenated aliases in subagent_provider_key_matches — normalize_subagent_provider_key maps - to _ before the matches! check, so "deepseek-anthropic" → "deepseek_anthropic" which matches. No change needed.
• EnvRuntimeOverrides base URL gap — deepseek_anthropic_base_url field at lib.rs:4307 is loaded from DEEPSEEK_ANTHROPIC_BASE_URL / DEEPSEEK_CLAUDE_BASE_URL (line 4429-4432) and wired into base_url_for. Fully implemented.

Good Choices

• Factoring api_provider_skips_models_probe as a separate predicate from api_provider_uses_anthropic_messages keeps the semantics clean — skipping the /models probe is specific to DeepseekAnthropic (Anthropic itself exposes GET /v1/models), while the Anthropic Messages wire protocol is shared by both.
• translation_message_request passes reasoning_effort: Some("off") which the Anthropic adapter no-ops correctly (line 113: "off" | "disabled" | "none" | "false" => {}) — thinking block stays absent for translation.
• model_completion_names_for_provider grouping DeepseekAnthropic with native DeepSeek variants (config.rs:1226-1228) is correct — same model catalog, different wire dialect.
• wire_model_for_provider strips deepseek-ai/ / deepseek/ namespace prefixes for DeepseekAnthropic via the shared canonical_official_deepseek_model_id path (config.rs:1177-1191), so the wire model sent to DeepSeek is the bare deepseek-v4-pro the Anthropic-compatible endpoint expects.