jeo-code 0.8.33 → 0.8.34

package/CHANGELOG.md

@@ -6,6 +6,27 @@ and this project adheres to [Semantic Versioning](https://semver.org/spec/v2.0.0
 
 The README mirrors the latest 5 entries — regenerate with `bun run changelog:sync`.
 
+## [0.8.34] - 2026-07-21
+_Provider/model-catalog sync against gjc's actual current source (a shallow clone of the real, public `Yeachan-Heo/gajae-code` repo at v0.11.5 — direct `packages/ai/src/models.json` comparison, not inference from release notes). The system prompt's philosophy, the `/route` command, and the `tencent` provider were explicitly out of scope and are unchanged._
+
+### Added
+- **`deepinfra` provider** (`src/ai/providers/openai-compatible-catalog.ts`) — `https://api.deepinfra.com/v1/openai`, `DEEPINFRA_API_KEY`, default `deepinfra/deepseek-ai/DeepSeek-V3.2`, plus known-model picker fallbacks (DeepSeek-R1-0528, Llama-3.3-70B-Instruct-Turbo, Kimi-K2.6, Qwen3-Coder-480B, gpt-oss-120b) — a real, popular inference cloud gjc's catalog has that jeo's didn't; one data-driven catalog row, auto-registered (no new adapter code, per this file's own "add a provider = one table row" design).
+- **`litellm` provider** — `http://localhost:4000/v1` (LiteLLM's own documented local-proxy default, expected to be overridden via `jeo provider add litellm --base-url <your-proxy>`), `LITELLM_API_KEY`. Deliberately has no `knownModels` list — unlike every other row, LiteLLM is a self-hosted proxy whose model catalog is entirely user-defined (whatever their own `config.yaml` routes to), so there is no globally-fixed list to offer.
+- `src/ai/types.ts`'s `ProviderName` and `src/auth/storage.ts`'s `AuthProvider`/`API_KEY_ONLY_PROVIDERS` gained both new names; `test/deepinfra-litellm-provider.test.ts` (+5 tests: registration, env var, base URL, routing, credential-gated readiness).
+
+### Scope notes (why the rest of gjc's current provider list wasn't ported wholesale)
+- **`tencent` — explicitly kept**, per instruction, even though gjc's current `models.json` has no such entry at all (jeo's Tencent Cloud MaaS integration predates and is independent of gjc's own catalog).
+- **`moonshot`** and **`firepass`** are gjc's OWN renamed/split entries for base URLs jeo already serves under different names — `moonshot`'s `https://api.moonshot.ai/v1` is jeo's existing `kimi` provider's base URL (jeo intentionally keeps ONE `kimi` provider spanning both the OAuth Kimi-Code subscription and the plain API-key path, rather than gjc's two separate provider identifiers); `firepass`'s `https://api.fireworks.ai/inference/v1` is byte-identical to jeo's existing `fireworks` entry. Already at parity, no action needed.
+- **`opencode` / `opencode-go` / `opencode-zen`** — these are the competing `opencode` CLI's own hosted proxy/gateway products; a jeo user would need an opencode account to use them, and integrating a rival agent's own product as an upstream provider has little practical value for jeo's userbase. Not ported.
+- **`azure-openai`, `amazon-bedrock`, `google-vertex`, `cursor`, `github-copilot`, `gitlab-duo`** — each requires a fundamentally new, non-trivial auth/wire-protocol adapter (Azure resource/deployment config, AWS SigV4 request signing, GCP OAuth + region/project config, Cursor's proprietary undocumented agent API, a GitHub Copilot-specific device-auth flow, GitLab OAuth) — none fit the existing "OpenAI-compatible base URL + API key" pattern every other row here uses; each is a real, separate feature in its own right, not a catalog row. Not ported in this pass to avoid fabricating auth flows without live verification.
+- **`cloudflare-ai-gateway` / `vercel-ai-gateway`** — both require a per-account/per-gateway URL embedded in the base (`.../v1/<account>/<gateway>/...`), which isn't a good fit for a fixed catalog row; already achievable today via jeo's existing `jeo provider add --base-url` custom-endpoint mechanism.
+- **`ollama-cloud`** — `src/ai/providers/ollama.ts`'s adapter never sends an `Authorization` header (by design, for the keyless local daemon); serving the hosted `https://ollama.com` cloud variant needs that adapter to accept a bearer credential, a real (if modest) code change beyond a catalog row. Deferred as a follow-up rather than rushed.
+- System prompt philosophy (`src/prompts/`) and the `/route` command (`src/commands/launch/route-slash.ts`, `src/agent/route-history.ts`) were not touched — out of scope per the request.
+
+### Verified
+- `test/deepinfra-litellm-provider.test.ts` — 5 pass (new).
+- `bun run typecheck` clean; full `bun test` 3102 pass / 0 fail across 305 files (an initial run under heavy concurrent system load hit many unrelated timeouts across subsystems this pass never touched — browser, daemon, tokenizer, tmux, search-gitignore — all confirmed pre-existing/load-induced via isolated re-runs, then a clean full re-run came back 0 fail).
+
 ## [0.8.33] - 2026-07-20
 _Real terminal-resize testing (repeated tmux resize-window + keystroke reproductions, byte-for-byte ANSI replay outside jeo/tmux to isolate root causes deterministically) uncovered and closed a progressive screen-corruption bug: resizing down to a narrow terminal (e.g. a ~20-column tmux pane) produced a growing stack of duplicate status-bar lines on every subsequent keystroke, and a separate, independently-reproducible scroll bug in the mid-turn live renderer._
 

package/README.ja.md

@@ -283,11 +283,11 @@ CI は `.github/workflows/npm-publish.yml` で公開します — GitHub リリ
 ## 変更履歴 (Changelog)
 
 <!-- CHANGELOG:START (auto-generated from CHANGELOG.md — run `bun run changelog:sync`) -->
+- **[0.8.34]** (2026-07-21) — Provider/model-catalog sync against gjc's actual current source (a shallow clone of the real, public `Yeachan-Heo/gajae-code` repo at v0.11.5 — direct `packages/ai/src/models.json` comparison, not inference from release notes). The system prompt's philosophy, the `/route` command, and the `tencent` provider were explicitly out of scope and are unchanged.
 - **[0.8.33]** (2026-07-20) — Real terminal-resize testing (repeated tmux resize-window + keystroke reproductions, byte-for-byte ANSI replay outside jeo/tmux to isolate root causes deterministically) uncovered and closed a progressive screen-corruption bug: resizing down to a narrow terminal (e.g. a ~20-column tmux pane) produced a growing stack of duplicate status-bar lines on every subsequent keystroke, and a separate, independently-reproducible scroll bug in the mid-turn live renderer.
 - **[0.8.32]** (2026-07-20) — gjc parity: TUI display-width/wrapping is now grapheme-cluster-aware for emoji sequences (VS16 presentation, skin-tone modifiers, keycaps, ZWJ-joined emoji like family/profession glyphs), fixing box-border and wrap misalignment ("깨짐") that the old per-code-point width summation produced whenever such a sequence appeared in a message — Korean/CJK-only text was already correctly handled and is unaffected.
 - **[0.8.31]** (2026-07-20) — `web_search` no longer silently degrades every non-Anthropic model to keyless DuckDuckGo scraping — OpenAI and Gemini sessions now get their own native, hosted search tool (matching the same active-model-gated, credential-required design Anthropic already had), with DuckDuckGo remaining the always-on terminal fallback for everyone else.
 - **[0.8.30]** (2026-07-20) — `bun run build` (the host-only dev binary build) has been silently broken for every user who ran it — a drift between the plain `package.json` `build` script and the actual working release-binary build command (`scripts/ci-release-build-binaries.ts`), which had already documented and fixed the exact same crash.
-- **[0.8.29]** (2026-07-20) — gajae-code (gjc) v0.10.2→v0.11.4 gap analysis: most of that range is gjc-specific architecture jeo intentionally does not replicate (SDK broker/session-index recovery, Gajae Pet, coordinator-mcp, Telegram/Discord rich-delivery internals, browser tab workers, worktree-subcommand removal, RPC durable model selection) — one genuinely applicable TUI bug was found and closed; the CLI empty-non-TTY-stdin fix (gjc #2586) was investigated and found already-correct by construction (jeo's one-shot gate forces true whenever stdin isn't a TTY, regardless of args, so the hang gjc fixed cannot occur here).
 
 See [CHANGELOG.md](CHANGELOG.md) for the full history.
 <!-- CHANGELOG:END -->

package/README.ko.md

@@ -284,11 +284,11 @@ CI는 `.github/workflows/npm-publish.yml`로 배포합니다 — GitHub 릴리
 ## 변경 이력 (Changelog)
 
 <!-- CHANGELOG:START (auto-generated from CHANGELOG.md — run `bun run changelog:sync`) -->
+- **[0.8.34]** (2026-07-21) — Provider/model-catalog sync against gjc's actual current source (a shallow clone of the real, public `Yeachan-Heo/gajae-code` repo at v0.11.5 — direct `packages/ai/src/models.json` comparison, not inference from release notes). The system prompt's philosophy, the `/route` command, and the `tencent` provider were explicitly out of scope and are unchanged.
 - **[0.8.33]** (2026-07-20) — Real terminal-resize testing (repeated tmux resize-window + keystroke reproductions, byte-for-byte ANSI replay outside jeo/tmux to isolate root causes deterministically) uncovered and closed a progressive screen-corruption bug: resizing down to a narrow terminal (e.g. a ~20-column tmux pane) produced a growing stack of duplicate status-bar lines on every subsequent keystroke, and a separate, independently-reproducible scroll bug in the mid-turn live renderer.
 - **[0.8.32]** (2026-07-20) — gjc parity: TUI display-width/wrapping is now grapheme-cluster-aware for emoji sequences (VS16 presentation, skin-tone modifiers, keycaps, ZWJ-joined emoji like family/profession glyphs), fixing box-border and wrap misalignment ("깨짐") that the old per-code-point width summation produced whenever such a sequence appeared in a message — Korean/CJK-only text was already correctly handled and is unaffected.
 - **[0.8.31]** (2026-07-20) — `web_search` no longer silently degrades every non-Anthropic model to keyless DuckDuckGo scraping — OpenAI and Gemini sessions now get their own native, hosted search tool (matching the same active-model-gated, credential-required design Anthropic already had), with DuckDuckGo remaining the always-on terminal fallback for everyone else.
 - **[0.8.30]** (2026-07-20) — `bun run build` (the host-only dev binary build) has been silently broken for every user who ran it — a drift between the plain `package.json` `build` script and the actual working release-binary build command (`scripts/ci-release-build-binaries.ts`), which had already documented and fixed the exact same crash.
-- **[0.8.29]** (2026-07-20) — gajae-code (gjc) v0.10.2→v0.11.4 gap analysis: most of that range is gjc-specific architecture jeo intentionally does not replicate (SDK broker/session-index recovery, Gajae Pet, coordinator-mcp, Telegram/Discord rich-delivery internals, browser tab workers, worktree-subcommand removal, RPC durable model selection) — one genuinely applicable TUI bug was found and closed; the CLI empty-non-TTY-stdin fix (gjc #2586) was investigated and found already-correct by construction (jeo's one-shot gate forces true whenever stdin isn't a TTY, regardless of args, so the hang gjc fixed cannot occur here).
 
 See [CHANGELOG.md](CHANGELOG.md) for the full history.
 <!-- CHANGELOG:END -->

package/README.md

@@ -291,11 +291,11 @@ Huge thanks to [gajae-code](https://github.com/Yeachan-Heo/gajae-code) for the i
 ## Changelog
 
 <!-- CHANGELOG:START (auto-generated from CHANGELOG.md — run `bun run changelog:sync`) -->
+- **[0.8.34]** (2026-07-21) — Provider/model-catalog sync against gjc's actual current source (a shallow clone of the real, public `Yeachan-Heo/gajae-code` repo at v0.11.5 — direct `packages/ai/src/models.json` comparison, not inference from release notes). The system prompt's philosophy, the `/route` command, and the `tencent` provider were explicitly out of scope and are unchanged.
 - **[0.8.33]** (2026-07-20) — Real terminal-resize testing (repeated tmux resize-window + keystroke reproductions, byte-for-byte ANSI replay outside jeo/tmux to isolate root causes deterministically) uncovered and closed a progressive screen-corruption bug: resizing down to a narrow terminal (e.g. a ~20-column tmux pane) produced a growing stack of duplicate status-bar lines on every subsequent keystroke, and a separate, independently-reproducible scroll bug in the mid-turn live renderer.
 - **[0.8.32]** (2026-07-20) — gjc parity: TUI display-width/wrapping is now grapheme-cluster-aware for emoji sequences (VS16 presentation, skin-tone modifiers, keycaps, ZWJ-joined emoji like family/profession glyphs), fixing box-border and wrap misalignment ("깨짐") that the old per-code-point width summation produced whenever such a sequence appeared in a message — Korean/CJK-only text was already correctly handled and is unaffected.
 - **[0.8.31]** (2026-07-20) — `web_search` no longer silently degrades every non-Anthropic model to keyless DuckDuckGo scraping — OpenAI and Gemini sessions now get their own native, hosted search tool (matching the same active-model-gated, credential-required design Anthropic already had), with DuckDuckGo remaining the always-on terminal fallback for everyone else.
 - **[0.8.30]** (2026-07-20) — `bun run build` (the host-only dev binary build) has been silently broken for every user who ran it — a drift between the plain `package.json` `build` script and the actual working release-binary build command (`scripts/ci-release-build-binaries.ts`), which had already documented and fixed the exact same crash.
-- **[0.8.29]** (2026-07-20) — gajae-code (gjc) v0.10.2→v0.11.4 gap analysis: most of that range is gjc-specific architecture jeo intentionally does not replicate (SDK broker/session-index recovery, Gajae Pet, coordinator-mcp, Telegram/Discord rich-delivery internals, browser tab workers, worktree-subcommand removal, RPC durable model selection) — one genuinely applicable TUI bug was found and closed; the CLI empty-non-TTY-stdin fix (gjc #2586) was investigated and found already-correct by construction (jeo's one-shot gate forces true whenever stdin isn't a TTY, regardless of args, so the hang gjc fixed cannot occur here).
 
 See [CHANGELOG.md](CHANGELOG.md) for the full history.
 <!-- CHANGELOG:END -->

package/README.zh.md

@@ -283,11 +283,11 @@ CI 通过 `.github/workflows/npm-publish.yml` 发布 — GitHub 发布 release
 ## 更新日志 (Changelog)
 
 <!-- CHANGELOG:START (auto-generated from CHANGELOG.md — run `bun run changelog:sync`) -->
+- **[0.8.34]** (2026-07-21) — Provider/model-catalog sync against gjc's actual current source (a shallow clone of the real, public `Yeachan-Heo/gajae-code` repo at v0.11.5 — direct `packages/ai/src/models.json` comparison, not inference from release notes). The system prompt's philosophy, the `/route` command, and the `tencent` provider were explicitly out of scope and are unchanged.
 - **[0.8.33]** (2026-07-20) — Real terminal-resize testing (repeated tmux resize-window + keystroke reproductions, byte-for-byte ANSI replay outside jeo/tmux to isolate root causes deterministically) uncovered and closed a progressive screen-corruption bug: resizing down to a narrow terminal (e.g. a ~20-column tmux pane) produced a growing stack of duplicate status-bar lines on every subsequent keystroke, and a separate, independently-reproducible scroll bug in the mid-turn live renderer.
 - **[0.8.32]** (2026-07-20) — gjc parity: TUI display-width/wrapping is now grapheme-cluster-aware for emoji sequences (VS16 presentation, skin-tone modifiers, keycaps, ZWJ-joined emoji like family/profession glyphs), fixing box-border and wrap misalignment ("깨짐") that the old per-code-point width summation produced whenever such a sequence appeared in a message — Korean/CJK-only text was already correctly handled and is unaffected.
 - **[0.8.31]** (2026-07-20) — `web_search` no longer silently degrades every non-Anthropic model to keyless DuckDuckGo scraping — OpenAI and Gemini sessions now get their own native, hosted search tool (matching the same active-model-gated, credential-required design Anthropic already had), with DuckDuckGo remaining the always-on terminal fallback for everyone else.
 - **[0.8.30]** (2026-07-20) — `bun run build` (the host-only dev binary build) has been silently broken for every user who ran it — a drift between the plain `package.json` `build` script and the actual working release-binary build command (`scripts/ci-release-build-binaries.ts`), which had already documented and fixed the exact same crash.
-- **[0.8.29]** (2026-07-20) — gajae-code (gjc) v0.10.2→v0.11.4 gap analysis: most of that range is gjc-specific architecture jeo intentionally does not replicate (SDK broker/session-index recovery, Gajae Pet, coordinator-mcp, Telegram/Discord rich-delivery internals, browser tab workers, worktree-subcommand removal, RPC durable model selection) — one genuinely applicable TUI bug was found and closed; the CLI empty-non-TTY-stdin fix (gjc #2586) was investigated and found already-correct by construction (jeo's one-shot gate forces true whenever stdin isn't a TTY, regardless of args, so the hang gjc fixed cannot occur here).
 
 See [CHANGELOG.md](CHANGELOG.md) for the full history.
 <!-- CHANGELOG:END -->

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "jeo-code",
-  "version": "0.8.33",
+  "version": "0.8.34",
   "description": "Clean, highly optimized AI coding agent using spec-first loop",
   "type": "module",
   "main": "src/cli.ts",

package/src/ai/providers/openai-compatible-catalog.ts

@@ -59,6 +59,13 @@ export const OPENAI_COMPAT_PROVIDERS: readonly OpenAICompatProviderDef[] = [
   { name: "zenmux", label: "ZenMux", baseUrl: "https://zenmux.ai/api/v1", apiKeyEnv: "ZENMUX_API_KEY", defaultModel: "zenmux/anthropic/claude-opus-4.6" },
   { name: "qianfan", label: "Qianfan", baseUrl: "https://qianfan.baidubce.com/v2", apiKeyEnv: "QIANFAN_API_KEY", defaultModel: "qianfan/deepseek-v3.2" },
   { name: "xiaomi", label: "Xiaomi", baseUrl: "https://api.xiaomimimo.com/v1", apiKeyEnv: "XIAOMI_API_KEY", defaultModel: "xiaomi/mimo-v2-flash" },
+  { name: "deepinfra", label: "DeepInfra", baseUrl: "https://api.deepinfra.com/v1/openai", apiKeyEnv: "DEEPINFRA_API_KEY", defaultModel: "deepinfra/deepseek-ai/DeepSeek-V3.2", knownModels: ["deepseek-ai/DeepSeek-V3.2", "deepseek-ai/DeepSeek-R1-0528", "meta-llama/Llama-3.3-70B-Instruct-Turbo", "moonshotai/Kimi-K2.6", "Qwen/Qwen3-Coder-480B-A35B-Instruct-Turbo", "openai/gpt-oss-120b"] },
+  // Self-hosted OpenAI-compatible proxy (LiteLLM) — the model catalog behind it is
+  // entirely user-configured (whatever the user's own litellm config.yaml routes to),
+  // so unlike every other row here there is no globally-fixed `knownModels` list to
+  // offer; `baseUrl` is LiteLLM's own documented local-proxy default and is expected
+  // to be overridden via `jeo provider add litellm --base-url <your-proxy-url>`.
+  { name: "litellm", label: "LiteLLM", baseUrl: "http://localhost:4000/v1", apiKeyEnv: "LITELLM_API_KEY", defaultModel: "litellm/gpt-4o-mini" },
   { name: "xiaomi-token-plan-ams", label: "Xiaomi Token Plan (Europe)", baseUrl: "https://token-plan-ams.xiaomimimo.com/v1", apiKeyEnv: "XIAOMI_TOKEN_PLAN_AMS_API_KEY", defaultModel: "xiaomi-token-plan-ams/mimo-v2.5", subscription: true },
   { name: "xiaomi-token-plan-cn", label: "Xiaomi Token Plan (China)", baseUrl: "https://token-plan-cn.xiaomimimo.com/v1", apiKeyEnv: "XIAOMI_TOKEN_PLAN_CN_API_KEY", defaultModel: "xiaomi-token-plan-cn/mimo-v2.5", subscription: true },
   { name: "xiaomi-token-plan-sgp", label: "Xiaomi Token Plan (Singapore)", baseUrl: "https://token-plan-sgp.xiaomimimo.com/v1", apiKeyEnv: "XIAOMI_TOKEN_PLAN_SGP_API_KEY", defaultModel: "xiaomi-token-plan-sgp/mimo-v2.5", subscription: true },

package/src/ai/types.ts

@@ -1,6 +1,6 @@
 import type { Credential } from "../auth";
 
-export type ProviderName = "anthropic" | "openai" | "gemini" | "antigravity" | "ollama" | "lmstudio" | "xai" | "kimi" | "groq" | "deepseek" | "mistral" | "openrouter" | "together" | "cerebras" | "fireworks" | "nvidia" | "alibaba-coding-plan" | "huggingface" | "nanogpt" | "qwen-portal" | "synthetic" | "venice" | "zenmux" | "qianfan" | "xiaomi" | "xiaomi-token-plan-ams" | "xiaomi-token-plan-cn" | "xiaomi-token-plan-sgp" | "minimax-code" | "minimax-code-cn" | "zai" | "minimax" | "tencent";
+export type ProviderName = "anthropic" | "openai" | "gemini" | "antigravity" | "ollama" | "lmstudio" | "xai" | "kimi" | "groq" | "deepseek" | "mistral" | "openrouter" | "together" | "cerebras" | "fireworks" | "nvidia" | "alibaba-coding-plan" | "huggingface" | "nanogpt" | "qwen-portal" | "synthetic" | "venice" | "zenmux" | "qianfan" | "xiaomi" | "xiaomi-token-plan-ams" | "xiaomi-token-plan-cn" | "xiaomi-token-plan-sgp" | "minimax-code" | "minimax-code-cn" | "zai" | "minimax" | "tencent" | "deepinfra" | "litellm";
 
 
 /** An image attached to a (user) message — base64 payload + IANA media type. */

package/src/auth/storage.ts

@@ -8,12 +8,12 @@ import { jeoEnv } from "../util/env";
 /** Providers with an interactive OAuth login + refresh flow. */
 export type OAuthProvider = "anthropic" | "openai" | "gemini" | "antigravity" | "kimi";
 /** Every provider jeo resolves a credential for: OAuth-capable ∪ API-key-only. */
-export type AuthProvider = OAuthProvider | "xai" | "groq" | "deepseek" | "mistral" | "openrouter" | "together" | "cerebras" | "fireworks" | "nvidia" | "alibaba-coding-plan" | "huggingface" | "nanogpt" | "qwen-portal" | "synthetic" | "venice" | "zenmux" | "qianfan" | "xiaomi" | "xiaomi-token-plan-ams" | "xiaomi-token-plan-cn" | "xiaomi-token-plan-sgp" | "minimax-code" | "minimax-code-cn" | "zai" | "minimax" | "tencent";
+export type AuthProvider = OAuthProvider | "xai" | "groq" | "deepseek" | "mistral" | "openrouter" | "together" | "cerebras" | "fireworks" | "nvidia" | "alibaba-coding-plan" | "huggingface" | "nanogpt" | "qwen-portal" | "synthetic" | "venice" | "zenmux" | "qianfan" | "xiaomi" | "xiaomi-token-plan-ams" | "xiaomi-token-plan-cn" | "xiaomi-token-plan-sgp" | "minimax-code" | "minimax-code-cn" | "zai" | "minimax" | "tencent" | "deepinfra" | "litellm";
 
 
 export const OAUTH_PROVIDERS: readonly OAuthProvider[] = ["anthropic", "openai", "gemini", "antigravity", "kimi"];
 /** API-key-only providers (no OAuth flow) — resolved from config.providers / `<NAME>_API_KEY`. */
-export const API_KEY_ONLY_PROVIDERS: readonly AuthProvider[] = ["xai", "groq", "deepseek", "mistral", "openrouter", "together", "cerebras", "fireworks", "nvidia", "alibaba-coding-plan", "huggingface", "nanogpt", "qwen-portal", "synthetic", "venice", "zenmux", "qianfan", "xiaomi", "xiaomi-token-plan-ams", "xiaomi-token-plan-cn", "xiaomi-token-plan-sgp", "minimax-code", "minimax-code-cn", "zai", "minimax", "tencent"];
+export const API_KEY_ONLY_PROVIDERS: readonly AuthProvider[] = ["xai", "groq", "deepseek", "mistral", "openrouter", "together", "cerebras", "fireworks", "nvidia", "alibaba-coding-plan", "huggingface", "nanogpt", "qwen-portal", "synthetic", "venice", "zenmux", "qianfan", "xiaomi", "xiaomi-token-plan-ams", "xiaomi-token-plan-cn", "xiaomi-token-plan-sgp", "minimax-code", "minimax-code-cn", "zai", "minimax", "tencent", "deepinfra", "litellm"];
 
 /** Narrow an AuthProvider to the OAuth-capable subset (xai has no OAuth flow). */
 export function isOAuthProvider(p: AuthProvider): p is OAuthProvider {