feat(engines): add vLLM as 6th UI-connected engine — RT, meta agent, …

[yodakrkids]

…settings

vLLM uses OpenAI-compatible API, so it routes through the existing openai_compat.rs path alongside ollama / lmstudio.

Backend (5 files)

• agent_detect.rs: probe_vllm() — GET {endpoint}/v1/models
• openai_compat.rs: vllm_base_url(), discover_vllm(), engine_name + per-engine API key (VLLM_API_KEY) routing
• agents.rs: start_openai_compat_stream + run_eval_agent vllm branches
• executor.rs: RT run_participant / stream_participant vllm cases
• model_discovery.rs: ENGINES list + dispatch + fallback_models entry

Frontend (11 files)

• engineConfig.ts / EngineSelector.tsx / AgentAvatar.tsx / index.css: register vllm in shared engine registry + agent color token
• CreateRoundtableDialog.tsx: RT participant engine dropdown
• MetaAgentSelector.tsx: onboarding meta agent — vllm endpoint state, detect_available_agents wiring, HTTP endpoint input, install hint
• AgentsSection.tsx: agent profile engine dropdown + endpoint override (engineEndpoint:vllm setting)
• RuntimeSection.tsx: Insight Agent engine select
• types/index.ts: extend engine union comments
• buildSendInput.ts: forward customBaseUrl when engine === "vllm"
• initialSetupApply.ts: KNOWN_ENGINES set
• locales/{en,ko}/dialog.json: meta_agent.vllm_install_hint

Defaults

• endpoint: http://localhost:8000 (override via UI or VLLM_ENDPOINT env)
• optional VLLM_API_KEY (Bearer token if set)

Verified

• tsc --noEmit: clean
• cargo check: clean
• Vitest streaming-flow + metaAgentSelector-modelDiscovery: 29 passed
• Manual: vLLM visible in meta agent selector

Summary

Related plan / issue

Changes

Test plan

• npx tsc --noEmit
• npx vite build
• cd src-tauri && cargo check
• npx vitest run
• cd src-tauri && cargo test --lib
• Manual smoke test of the affected feature

Invariants touched

Screenshots / logs

Checklist

• PR title follows Conventional Commits (feat(scope): ...)
• Tests added or updated
• Docs updated (plan, README, how-to) if behavior changed
• No secrets, tokens, or personal paths in the diff

[gemini-code-assist]

Code Review

This pull request adds support for vLLM as an OpenAI-compatible engine, integrating it across backend detection, streaming, roundtable execution, and frontend settings. The review comments point out critical routing issues where vLLM non-streaming, streaming, and evaluation requests are incorrectly routed to the Ollama endpoint because they call functions hardcoded to Ollama's base URL. Additionally, a high-severity issue was identified in the vLLM detection probe, which fails to include the VLLM_API_KEY authorization header, potentially causing onboarding failures on authenticated vLLM instances.

[gemini-code-assist]

The fallback non-streaming run_participant execution for vllm currently routes through openai_compat::run(run_input), which is hardcoded to use ollama_base_url() (typically http://localhost:11434). This causes vLLM roundtable requests to be incorrectly routed to the Ollama endpoint. We should instead use openai_compat::stream_run_with_base with the correct vllm_base_url() by blocking on the current Tokio runtime handle.

            "vllm" => {
                let base_url = openai_compat::vllm_base_url();
                let res = tokio::runtime::Handle::try_current()
                    .map_err(|_| AppError::Agent("No tokio runtime available for vllm".into()))
                    .and_then(|rt| {
                        rt.block_on(async {
                            openai_compat::stream_run_with_base(run_input, base_url, |_| {}, |_| {}).await
                        })
                    });
                (res, "vllm")
            }

[gemini-code-assist]

In stream_participant, the streaming execution for vllm currently calls openai_compat::stream_run(run_input, on_progress, on_chunk), which is hardcoded to use ollama_base_url(). This causes vLLM streaming roundtable requests to be incorrectly routed to the Ollama endpoint. We should instead use openai_compat::stream_run_with_base with the correct base URL for the selected engine.

Suggested change

	let label = if engine_key_owned == "vllm" { "vllm" } else { "ollama" };
	(openai_compat::stream_run(run_input, on_progress, on_chunk).await, label)
	let base_url = if engine_key_owned == "vllm" {
	openai_compat::vllm_base_url()
	} else {
	std::env::var("OLLAMA_HOST").unwrap_or_else(|_| "http://localhost:11434".into())
	};
	let label = if engine_key_owned == "vllm" { "vllm" } else { "ollama" };
	(openai_compat::stream_run_with_base(run_input, base_url, on_progress, on_chunk).await, label)

[gemini-code-assist]

In run_eval_agent, the evaluation path for vllm currently routes through openai_compat::run(run_input), which is hardcoded to use ollama_base_url(). This causes vLLM evaluation requests to be incorrectly routed to the Ollama endpoint. We should instead use openai_compat::stream_run_with_base with the correct vllm_base_url() by blocking on the current Tokio runtime handle.

        "ollama" => openai_compat::run(run_input),
        "vllm" => {
            let base_url = openai_compat::vllm_base_url();
            let rt = tokio::runtime::Handle::try_current()
                .map_err(|_| AppError::Agent("No tokio runtime available for vllm".into()))?;
            rt.block_on(async {
                openai_compat::stream_run_with_base(run_input, base_url, |_| {}, |_| {}).await
            })
        }

[gemini-code-assist]

The probe_vllm function does not include the Authorization header with VLLM_API_KEY when sending the probe request. If the vLLM instance requires authentication (which is common for shared or cloud-hosted instances), the detection probe will fail with a 401 Unauthorized status, even if the key is configured in the environment. Adding the Authorization header ensures that authenticated vLLM instances are correctly detected during onboarding.

    eprintln!("[agent-detect] probe vllm: GET {}", url);
    let mut req = client.get(&url);
    if let Ok(token) = std::env::var("VLLM_API_KEY") {
        req = req.header("Authorization", format!("Bearer {}", token));
    }
    match req.send().await {

[hang-in]

@yodakrkids vLLM 6번째 엔진 추가 PR 감사드립니다. tunaFlow 의 openai_compat 라우팅 (ollama / lmstudio 와 같은 layer) 을 그대로 재사용하신 패턴이 합리적입니다. 코드 평가 + CI 통과 + Gemini review 확인 후 머지 진행하겠습니다 — 다음 patch release (v0.1.8-beta-5) 또는 minor release (v0.1.9-beta) 에 포함 예정. CI 가 fork PR 정책으로 메인테이너 approval 대기 중이라 곧 trigger 하겠습니다.

[hang-in]

@yodakrkids 안녕하세요. 정말 감사드립니다 — vLLM 6번째 엔진 추가 PR 받았습니다.

본 PR 의 변경분을 그대로 main 으로 가져가면서, Gemini code review 가 지적한 4 건 critical fix 를 함께 적용한 후속 PR #297 을 만들어 머지했습니다:

머지된 PR: #297 (squash commit cf8d8a5)

• yodakrkids 의 vLLM 추가 commit 은 9fc1a13 으로 cherry-pick — author 보존됨
• 그 위에 Gemini critical 4 건 fix commit 3 개를 쌓음

적용된 4 fix

1. executor.rs:55 — run_participant (non-streaming) vllm 분기가 openai_compat::run 호출 → 내부 ollama_base_url() 하드코딩 회귀. stream_run_with_base(input, vllm_base_url(), ...) + Handle::current().block_on(...) 으로 수정.
2. executor.rs:184 — stream_participant (streaming) 동일 회귀. ollama / vllm 분기 분리 후 stream_run_with_base 직접 await.
3. agents.rs:614 — run_eval_agent 동일 회귀. Handle::try_current() + block_on(stream_run_with_base(..., vllm_base_url(), ...)).
4. agent_detect.rs:275 — probe_vllm 의 Authorization 헤더 누락. VLLM_API_KEY env 있으면 Bearer <key> 헤더 추가.

해결된 충돌
main 의 PR #295 (explicit endpoint detect trigger — Enter / refresh button) 와의 MetaAgentSelector 충돌은 main 의 explicit-trigger 패턴 유지 + vllm 옵션 추가 형태로 해결했습니다. 디바운스 기반 자동 detect 는 외부 사용자 보고 (192.168.1.1 입력 중 . 칠 때마다 발동) 회피를 위해 제거되었습니다.

검증

• cargo check PASS
• cargo test --lib — 656 passed / 0 failed
• tsc --noEmit clean
• vitest run — 478 passed / 0 failed
• CI 3 종 (rust-check / frontend-check / eval) 모두 SUCCESS 후 머지

곧 v0.1.9-beta minor release publish 후 release URL + 회복 안내 댓글 별도로 드리겠습니다. 정말 좋은 기여 감사합니다.

[hang-in]

@yodakrkids vLLM PR 감사드립니다. Gemini code review 가 critical 4 항목 (executor.rs / agents.rs vLLM 분기가 ollama base URL 로 잘못 라우팅 + probe_vllm Authorization 누락) 지적해 주셨고, 본 PR 의 commit 들 위에 메인테이너가 4 fix follow-up commit 추가하여 PR #297 로 supersede 머지했습니다.

머지 commit: cf8d8a5 (PR #297)
Release: https://github.com/hang-in/tunaFlow/releases/tag/v0.1.9-beta (publish 후 URL)

본 PR 은 PR #297 로 supersede 됐으니 close 합니다. vLLM 외 다른 영역 PR 도 환영합니다 — 본 PR 의 6th engine 추가 작업 덕분에 minor bump (v0.1.8 → v0.1.9) 가 가능했습니다. 감사합니다.

[hang-in]

v0.1.9-beta publish 완료 — https://github.com/hang-in/tunaFlow/releases/tag/v0.1.9-beta

@yodakrkids 의 vLLM 6th engine 추가 (merge cf8d8a5 via PR #297) 가 첫 외부 contributor feature release 입니다. macOS DMG / Windows installer 자산 모두 빌드 완료. 다시 한 번 감사드립니다.