fix(custom): support wire = "responses" | "anthropic" for kind="openai-compatible"

[whp233]

fix(custom): support wire = "responses" | "anthropic" for kind="openai-compatible"

Summary / 摘要

Custom provider (kind = "openai-compatible") currently only supports Chat Completions wire. 配置 wire = "responses" / "anthropic" 被忽略，始终走 ChatCompletions，导致需要 Responses 或 Anthropic Messages 的模型（如 muse-spark-1.2 on opencode.ai/zen/v1）无法在 custom provider 下工作。

Custom provider (kind="openai-compatible") only speaks Chat Completions, ignoring wire setting. Models requiring Responses / Anthropic wire (e.g. muse-spark-1.2 via opencode.ai/zen/v1) cannot be used as custom.

Reproduction / 复现

Minimal config.toml:

[providers.myspark]
kind = "openai-compatible"
base_url = "https://opencode.ai/zen/v1"
api_key_env = "OPENCODE_API_KEY"  # or OPENCODE_ZEN_API_KEY
model = "muse-spark-1.2"
wire = "responses"   # 期望走 OpenAI Responses；改成 "anthropic" / "messages" 同理
# wire = "anthropic"

[providers.myspark2]
kind = "openai-compatible"
base_url = "https://opencode.ai/zen/v1"
api_key_env = "OPENCODE_API_KEY"
model = "muse-spark-1.2"
wire = "anthropic"

Steps:

1. codewhale --provider myspark (or set as default)
2. Send any message / codewhale doctor --verbose
3. Observe request goes to /v1/chat/completions instead of /responses or /v1/messages

Expected / 期望

• wire = "responses" (also response, openai-responses) → WireFormat::Responses, requests hit /responses (OpenAI Responses API, like openai_codex).
• wire = "anthropic" / "messages" / "anthropic-messages" → WireFormat::AnthropicMessages, requests hit /v1/messages.
• wire omitted / "openai" / "chat" / "chat_completions" → WireFormat::ChatCompletions (default, backward compatible).
• Case-insensitive, trims whitespace.

Actual / 实际

Always WireFormat::ChatCompletions regardless of wire value. 对 opencode.ai/zen/v1 + muse-spark-1.2 这类 Responses-Anthropic hybrid 网关，请求 400/404 或 unsupported wire，因为服务端期望 Responses/Anthropic payload 而收到 Chat Completions schema.

Root cause / 根因

crates/config/src/provider.rs:1651-1653:

impl Provider for Custom {
    fn wire_policy(&self) -> WirePolicy {
        WirePolicy::Fixed(WireFormat::ChatCompletions) // ← fixed, ignores config
    }
}

• Custom::wire_policy() is Fixed(ChatCompletions), so WirePolicy::resolve() never consults providers.<name>.wire.
• crates/tui/src/config.rs & crates/tui/src/client.rs resolve wire_format via provider.wire_policy() + provider_wire_format_for_config(), but Custom never reaches the ModelAware / wire_config_prefers_anthropic path. providers.<name>.wire is parsed & stored in lib.rs (ProviderConfigToml::wire) yet never read for Custom.
• Other built-ins (e.g. OpencodeZen → ModelAware, OpenaiCodex → Fixed(Responses), Anthropic → Fixed(AnthropicMessages)) demonstrate the wire abstraction already exists; Custom is the outlier.

Proposed fix / 修复建议

Branch: fix/custom-wire-responses (or main with following changes)

1. crates/config/src/provider.rs — Custom::wire_policy
   Make Custom respect wire field: either ModelAware or Fixed with override via wire string. Suggested mapping (reuse WirePolicy::resolve keys):

   • responses → Responses
   • anthropic / messages / anthropic_messages / anthropic-messages → AnthropicMessages
   • chat / chat_completions / openai / empty → ChatCompletions (default)

2. crates/tui/src/config.rs — provider_wire_format_for_config / wire_config_prefers_anthropic
   Ensure ApiProvider::Custom path reads config.wire (like dual-wire vendors do). Normalize aliases (wire, wire_format, api_style already aliased in config_document.rs).

3. crates/tui/src/client.rs — dispatch & validation
   No wire-specific hardcoding for Custom; rely on resolved wire_format for match self.wire_format { ChatCompletions | Responses | AnthropicMessages } → correct body builder / endpoint. Keep validate_route / limits intact.

Backward compat: 无 wire 时保持 ChatCompletions，不影响现有 custom 用户。

Workaround / 临时绕过

Reuse the existing Responses-fixed built-in provider openai_codex + env override (since Custom is blocked):

[providers.openai_codex]
# base_url/model 留空，由 env 覆盖，避免写死官方 URL

export OPENAI_CODEX_BASE_URL="https://opencode.ai/zen/v1"
export OPENAI_CODEX_API_KEY="$OPENCODE_API_KEY"   # or OPENCODE_ZEN_API_KEY
# 可选：指定默认模型
export OPENAI_CODEX_MODEL="muse-spark-1.2"
codewhale --provider openai_codex --model muse-spark-1.2

或直接改 env 启动：

$env:OPENAI_CODEX_BASE_URL="https://opencode.ai/zen/v1"
$env:OPENAI_CODEX_API_KEY=$env:OPENCODE_API_KEY
codewhale --provider openai_codex

原理：OpenaiCodex::wire_policy = Fixed(Responses) 天然走 /responses，绕过 Custom 的 Fixed-Chat 限制。For Anthropic wire, use providers.anthropic similarly if gateway supports it.

缺点：占用 openai_codex 槽位，不能同时配多个 custom Responses 网关；env 覆盖对多路由不友好。

Related files / 相关文件

• crates/config/src/provider.rs — Custom struct & wire_policy() (root cause)
• crates/tui/src/config.rs — provider_wire_format_for_config(), wire_config_prefers_anthropic(), ProviderConfigToml::wire handling
• crates/tui/src/client.rs — provider_default_wire_format(), DeepSeekClient::wire_format dispatch (ChatCompletions / Responses / AnthropicMessages)
• crates/config/src/lib.rs — ProviderKind::Custom, ProviderConfigToml, wire field definition

Environment

• CodeWhale: main @ a0341b5a4 (pre-fix) / branch fix/custom-wire-responses
• Provider: kind = "openai-compatible" (Custom) → opencode.ai/zen/v1 / muse-spark-1.2
• OS: Windows 10 (also repro on Linux/macOS)

---

Fix branch: whp233/CodeWhale#fix/custom-wire-responses (commit 934a69a)
Closes #5713

[github-actions]

Thanks @whp233 for taking the time to contribute.

This repository is observing a maintainer-managed PR intake gate in dry-run mode, so this pull request is staying open. This note helps maintainers prepare the allowlist before any enforcement is considered.

Please read CONTRIBUTING.md for the expected contribution shape. A maintainer can grant recurring PR access by commenting /lgtm on a pull request.

[Hmbown]

Closing as superseded by #5716 — verified, not assumed: of #5716's 8 files, the 4 this PR touches (crates/tui/src/client.rs, crates/tui/src/config.rs, .github/workflows/build-windows.yml byte-identical diffs; crates/config/src/provider.rs the same hunk after rebase) are fully contained there, and #5716 adds the opencode-zen muse-spark routing, route-resolver integration, config example, and the dev proxy on top.

Full credit to @whp233 for the wire-dialect design — the wire = "responses" | "anthropic" | "chat" selector for kind = "openai-compatible" custom providers originates here and is the core of what #5716 carries forward. It aligns with the open design asks #5094/#5713.

Both PRs are red on the same checks (Lint, Tests ×3); the fix-forward is running on #5716's lane against current main (it is the superset). Nothing is lost by closing this one — the commits stay, and #5716's merge will carry the work.