feat: add Volcengine provider with DeepSeek-V4-Pro/Flash support

[dzyuan]

Add Volcengine (Volcano Engine Ark) provider for DeepSeek-V4-Pro/Flash via Coding API.

[gemini-code-assist]

Code Review

This pull request introduces support for the Volcengine (Volcano Engine Ark) provider across the agent, CLI, and TUI crates. Key changes include adding Volcengine to the model registry, configuration structures, and environment variable mappings, as well as updating the TUI to support Volcengine authentication and model selection. Feedback highlights a critical issue where Volcengine was incorrectly categorized as a provider not requiring authentication, which would lead to authorization failures. Additionally, reviewers noted potential issues with model aliasing for reasoning models, unnecessary cross-contamination of environment variables between providers, and minor inconsistencies in CLI output messages and configuration key ordering.

[gemini-code-assist]

Volcengine (Volcano Engine Ark) is a cloud provider that requires authentication. Including it in this list—which is intended for self-hosted or local deployments that run without authentication—will cause the client to omit the Authorization header, resulting in authentication failures.

            ApiProvider::Sglang | ApiProvider::Vllm | ApiProvider::Ollama => Ok(String::new()),

[gemini-code-assist]

The deepseek-reasoner alias is mapped to DeepSeek-V4-Flash, but it is typically the canonical alias for high-capability reasoning models (like DeepSeek-V4-Pro). Mapping it to the Flash version may lead to unexpected behavior when users expect the full reasoning model.

[gemini-code-assist]

The environment variables for WanjieArk and Volcengine are being set regardless of which provider is active. This cross-contamination is unnecessary and could lead to confusion. Each provider should only have its own relevant environment variables set.

        if resolved_runtime.provider == ProviderKind::WanjieArk {
            cmd.env("WANJIE_ARK_API_KEY", api_key);
        }
        if resolved_runtime.provider == ProviderKind::Volcengine {
            cmd.env("VOLCENGINE_API_KEY", api_key);
        }

[gemini-code-assist]

Similar to the previous block, environment variables for WanjieArk and Volcengine are being set for both providers. These should be separated to avoid leaking keys into unrelated environment variables.

        if resolved_runtime.provider == ProviderKind::WanjieArk {
            cmd.env("WANJIE_ARK_API_KEY", api_key);
        }
        if resolved_runtime.provider == ProviderKind::Volcengine {
            cmd.env("VOLCENGINE_API_KEY", api_key);
        }

[gemini-code-assist]

The configuration keys for wanjie_ark and volcengine are interleaved. Grouping them by provider improves readability and maintainability.

Suggested change

	"providers.wanjie_ark.api_key" => self.providers.wanjie_ark.api_key.clone(),
	"providers.volcengine.api_key" => self.providers.volcengine.api_key.clone(),
	"providers.wanjie_ark.base_url" => self.providers.wanjie_ark.base_url.clone(),
	"providers.volcengine.base_url" => self.providers.volcengine.base_url.clone(),
	"providers.wanjie_ark.model" => self.providers.wanjie_ark.model.clone(),
	"providers.volcengine.model" => self.providers.volcengine.model.clone(),
	"providers.wanjie_ark.api_key" => self.providers.wanjie_ark.api_key.clone(),
	"providers.wanjie_ark.base_url" => self.providers.wanjie_ark.base_url.clone(),
	"providers.wanjie_ark.model" => self.providers.wanjie_ark.model.clone(),
	"providers.volcengine.api_key" => self.providers.volcengine.api_key.clone(),
	"providers.volcengine.base_url" => self.providers.volcengine.base_url.clone(),
	"providers.volcengine.model" => self.providers.volcengine.model.clone(),

[gemini-code-assist]

The suggested command uses deepseek instead of codewhale. This is inconsistent with the other provider setup messages in this file.

Suggested change

	("VOLCENGINE_API_KEY", "deepseek auth set --provider volcengine")
	("VOLCENGINE_API_KEY", "codewhale auth set --provider volcengine")

[Hmbown]

Independent review (Devin): merged against origin/main (54151a4) to verify conflicts and correctness.

Unresolved bug — Volcengine silently skips auth (HIGH). config.rs still returns Ok(String::new()) for Volcengine in the no-auth self-hosted group alongside Sglang/Vllm/Ollama. The second commit fixed reasoning_effort branching but left this path untouched. Any user without an explicit VOLCENGINE_API_KEY in env gets a silent empty bearer token and an HTTP 401 with no actionable error message. Needs its own anyhow::bail! arm (same pattern as WanjieArk/Openrouter/Fireworks).

model_completion_names mismatch (MEDIUM). Volcengine is bucketed with OFFICIAL_DEEPSEEK_MODELS (deepseek-v4-pro, deepseek-v4-flash — lowercase), but the registry canonical IDs are DeepSeek-V4-Pro / DeepSeek-V4-Flash (mixed case), and normalize_model_for_provider passes Volcengine through unchanged. The completion list served by the picker won't match what the API accepts.

Alias collision still present (LOW). deepseek-v4-flash and deepseek-chat appear as Volcengine-Flash aliases while also resolving against Deepseek/NvidiaNim/Openrouter entries; first-match wins and the ordering is registry-position-dependent. Consider prefixed-only aliases for the Volcengine entries.

Merge conflicts with origin/main (v0.8.48 surface, 3 files).

• crates/config/src/lib.rs: v0.8.48 adds CODEWHALE_PROVIDER/CODEWHALE_MODEL env var lookups and a Moonshot provider arm in model_for(); both conflict with the Volcengine arms.
• crates/cli/src/lib.rs: v0.8.48 refactors secret bridging to provider_env_vars() helper; conflicts with the per-provider cmd.env() blocks added here.
• crates/tui/src/config.rs: provider_passes_model_through conflict (Moonshot vs Volcengine arm). All three are straightforward and non-destructive — both sides should be kept.

Sibling provider PRs (#1864/#1868 SiliconFlow): touch the same match arms in config.rs, config/src/lib.rs, and client.rs — whoever lands second will need a rebase.

No tests added. No unit or integration tests cover Volcengine key resolution, model normalisation, or reasoning_effort dispatch. Adding a test parallel to the existing WanjieArk/Atlascloud tests would close the auth-regression risk.