[Wibias]

Summary

• Backend for [Feature]: Multiple providers using the same model, create a combo to handle load balancing. #133: combo/<id> virtual models with failover and sticky weighted round-robin, per-target cooldowns, and hop-on-retryable upstream failures (429 / 5xx / permission-subscription 403).
• Catalog exposure + GET / PUT / DELETE /api/combos.
• defaultEffort (default medium): fills reasoning.effort when the client omits it; never overrides client-sent effort; skipped when the resolved target has no reasoning ladder (noReasoningModels / empty / unknown ladder); clamped to a supported rung when a ladder exists.
• Intentionally the runtime + API foundation — Combos CRUD GUI is a follow-up draft PR on top of this API.

Runtime

User asks for combo/<id> → pick a target → call upstream → on retryable failure cool that target and hop to the next member until success or exhaustion.

Test plan

• bun test tests/combos.test.ts
• bun test src/combos/effort.test.ts
• bun test tests/server-combo-failover-e2e.test.ts (A 403 → B 200 through handleResponses)
• Live proxy: PUT/GET defaultEffort round-trip after restart
• Live proxy: failover hop with mock private upstreams via /api/providers + combo/<id> → /v1/responses
• Confirm non-combo models are unchanged

[lidge-jun]

Thanks for building this — combos/load-balancing (#133) is a feature we want, and the review confirmed the direction. The implementation needed deeper changes than a review round-trip could carry, so we're rebuilding it as a small dependency-ordered stack directly on dev, preserving your authorship (your reconstruction commits carry you as author; redesigned parts carry Co-authored-by).

Review findings driving the rebuild (against head a4abda1, re-checked at 6824e7b):

• P0: after a 503 hop, a backup provider's 401 entered the XAI OAuth refresh branch and the refreshed XAI bearer was sent to the backup provider (token leak);
• failover didn't cover connection exceptions (502 before combo recovery), passthrough/runTurn/web-search returned early, and cross-adapter hops died with HTTP 500 ("Non-streaming not supported");
• "sticky round-robin" was weighted-random with rotation after the first success regardless of stickyLimit, and the test asserted picks.size >= 1;
• deleted combo members fell open onto the default provider; post-hop vision/effort safety wasn't recomputed; catalog advertised fabricated capabilities; usage attribution recorded only the final provider.
• The PR was also 58 commits behind dev with content conflicts in router.ts/responses.ts and overlapped the landed feat(gui): providers workspace redesign polish + freeTier + rate limits #139/Add React Doctor and clear GUI/src diagnostics #140 rebuild surface.

Slice 1 is already on dev (f48a5c7 authored by you + f9a297f/11984f9b/68cda367): validated combo CRUD, reserved combo/ namespace with backward compat, deterministic smooth-weighted round-robin with real sticky semantics, fail-closed disabled/deleted members, and a typed combo_unavailable 503. Slices 2 (per-target failover engine, OAuth scoping, cross-adapter hops) and 3 (member-derived catalog capabilities, per-attempt usage attribution) follow; #133 stays open until they land.

Closing this PR in favor of the dev stack — the defaultEffort follow-up commit (6824e7b) is folded into the slices too.