feat(sse): provider-family auto combos auto/glm, auto/minimax, auto/zai, auto/mimo, auto/gemma, auto/llama, auto/gemini (#6453)

[diegosouzapw]

Summary

New routable ids auto/<family> — auto/glm, auto/minimax, auto/mimo, auto/zai, auto/gemma, auto/llama, auto/gemini — that materialize an on-demand virtual combo spanning whatever installed backends currently expose that model family, degrading gracefully as backends rotate. This is a new axis alongside the existing task-mode (auto/best-coding) and category:tier (auto/coding:fast) combos: instead of grouping by routing intent, it groups by underlying model family.

• New pure module open-sse/services/autoCombo/modelFamily.ts:
  • detectModelFamily(modelId) — prefix-based classification for glm, minimax, mimo, gemma, llama, gemini.
  • zai alias rule: z.ai's hosted API serves the same glm-* model ids as every other GLM backend (glm provider, custom OpenAI/Anthropic-compatible connections, etc.), so auto/zai is deliberately not a model-name prefix rule — it's resolved by provider id instead. auto/zai means "route to my z.ai backend specifically"; auto/glm means "route to any connected provider currently serving a GLM model, z.ai included". Documented inline in modelFamily.ts.
• Wired into the existing on-demand recognition/materialization path (builtinCatalog.ts::createBuiltinAutoCombo / isRecognizedBuiltinAuto) — reuses createVirtualAutoCombo (no DB writes), extended with an optional family overlay on AutoComboSpec alongside the existing category/tier overlay.
• Advertised in the /v1/models catalog listing (src/app/api/v1/models/catalog.ts) alongside the existing auto/* entries via the new AUTO_FAMILY_IDS list.
• Unknown auto/<family> ids fall through to the same clean "Unknown built-in auto combo" 400 error every other unrecognized auto/* id already gets — no new error path, no raw error leakage.
• A family with zero matching connected/no-auth candidates degrades to an empty candidate pool (never silently falls back to the full pool, matching the existing :free/:pro tier-filter precedent) — routing among a subset when partially available.

Test plan

• New test file tests/unit/autoCombo/provider-family-combos.test.ts (vitest, real temp-DB backed like virtual-auto-combo.test.ts):
  • detectModelFamily pure classification for all 6 prefix families + null for unrelated ids + confirms zai is never detected from a model id.
  • isValidModelFamily / AUTO_FAMILY_IDS shape.
  • auto/glm materializes a combo spanning BOTH a glm provider connection and a zai provider connection (both serving glm-5.2), excluding an unrelated openai connection.
  • auto/zai materializes a combo with ONLY the zai-provider connection, even though a glm provider connection also serves a GLM model (proves the provider-override rule).
  • Degrade-gracefully path: auto/minimax excludes a connected-but-unrelated openai candidate while including only genuine minimax-family candidates.
  • auto/<unknownfamily> rejects with the same clean "Unknown built-in auto combo" error.
  • isRecognizedBuiltinAuto recognizes all 7 family ids.
  • 11/11 passing: npx vitest run --config vitest.mcp.config.ts tests/unit/autoCombo/provider-family-combos.test.ts
• npm run test:vitest — 242/243 passing (the 1 failure, open-sse/mcp-server/__tests__/audit.test.ts, is a pre-existing timeout flake under full-suite thread contention; passes 3/3 in isolation, file untouched by this PR).
• node --import tsx/esm --test tests/unit/virtual-auto-combo.test.ts tests/unit/auto-combos-enhanced-4235.test.ts tests/unit/auto-custom-provider-5873.test.ts — 14/14 passing (no regression to the existing auto-combo test suites).
• npm run typecheck:core — clean.
• npm run typecheck:noimplicit:core — the only 2 pre-existing errors (combo.ts, cliRuntime.ts) are untouched by this PR (base-red, confirmed via git diff --stat).
• npx eslint --suppressions-location config/quality/eslint-suppressions.json <changed files> — clean.
• node scripts/check/check-file-size.mjs — no new violations on touched files (4 pre-existing base-reds elsewhere, untouched by this PR).
• node scripts/check/check-complexity.mjs — verified 0 NEW violations introduced by this PR's edits to virtualFactory.ts/catalog.ts (both already had the same violated functions over the cyclomatic cap before this change; diffed base vs after via a same-directory temp-file eslint run — identical violation count, only body/line-number growth on already-flagged functions). The global count regression (2035→2048) is pre-existing repo-wide drift, not attributable to this PR's 3 touched files.
• npm run check:cycles — clean.

Closes #6453

[chatgpt-codex-connector]

You have reached your Codex usage limits for code reviews. You can see your limits in the Codex usage dashboard.

[gemini-code-assist]

Warning

You have reached your daily quota limit. Please wait up to 24 hours and I will start processing your requests again!