Skip to content

7.4 billion tokens per month. 34 free LLM providers. 635 free model endpoints. All behind one /v1 endpoint, plus any custom OpenAI-compatible endpoint. Smart routing, automatic failover, encrypted keys. Personal experimentation only.

TypeScript

fix(providers): remove the retired GitHub Models provider (#1557) (#1560)

Tashfeencece2df2af96
diff --git a/client/src/components/keys/shared.tsx b/client/src/components/keys/shared.tsx
index 001fcac8..a5e731a4 100644
--- a/client/src/components/keys/shared.tsx
+++ b/client/src/components/keys/shared.tsx
@@ -60,7 +60,6 @@ export const PLATFORMS: { value: Platform; label: string; url: string; keyless?:
{ value: 'nvidia', label: 'NVIDIA NIM', url: 'https://build.nvidia.com/settings/api-keys' },
{ value: 'mistral', label: 'Mistral', url: 'https://console.mistral.ai/api-keys/' },
{ value: 'openrouter', label: 'OpenRouter', url: 'https://openrouter.ai/keys' },
- { value: 'github', label: 'GitHub Models', url: 'https://github.com/settings/tokens' },
{ value: 'cohere', label: 'Cohere', url: 'https://dashboard.cohere.com/api-keys' },
{ value: 'cloudflare', label: 'Cloudflare Workers AI', url: 'https://dash.cloudflare.com' },
{ value: 'zhipu', label: 'Zhipu AI (Z.ai)', url: 'https://z.ai/manage-apikey/apikey-list' },
diff --git a/docs/en/api/01-rest-api.md b/docs/en/api/01-rest-api.md
index 5a89946d..828553f3 100644
--- a/docs/en/api/01-rest-api.md
+++ b/docs/en/api/01-rest-api.md
@@ -162,7 +162,7 @@ final = client.chat.completions.create(
print(final.choices[0].message.content)
```
-Works with `stream=True` as well — you'll get `delta.tool_calls` chunks followed by a `finish_reason: "tool_calls"` close. Under the hood, OpenAI-compatible providers (Groq, Cerebras, Mistral, OpenRouter, GitHub Models, HuggingFace, Cloudflare, Cohere compat) get the request passed through; Gemini requests get translated into Google's `functionDeclarations` / `functionResponse` shape and the response is translated back.
+Works with `stream=True` as well — you'll get `delta.tool_calls` chunks followed by a `finish_reason: "tool_calls"` close. Under the hood, OpenAI-compatible providers (Groq, Cerebras, Mistral, OpenRouter, HuggingFace, Cloudflare, Cohere compat) get the request passed through; Gemini requests get translated into Google's `functionDeclarations` / `functionResponse` shape and the response is translated back.
## Gemini Google Search grounding
@@ -334,8 +334,6 @@ curl http://localhost:3001/v1/embeddings \
| Family (`model`) | Dims | Providers (failover order) |
| --- | --- | --- |
| `gemini-embedding-001` *(default)* | 3072 | Google |
-| `text-embedding-3-large` | 3072 | GitHub Models |
-| `text-embedding-3-small` | 1536 | GitHub Models |
| `embed-v4.0` | 1536 | Cohere |
| `bge-m3` | 1024 | Cloudflare → Hugging Face |
| `qwen3-embedding-0.6b` | 1024 | Cloudflare |
diff --git a/docs/en/architecture/00-high-level-index.md b/docs/en/architecture/00-high-level-index.md
index f18c4613..9bf3e6c4 100644
--- a/docs/en/architecture/00-high-level-index.md
+++ b/docs/en/architecture/00-high-level-index.md
@@ -85,7 +85,6 @@ A self-hosted, single-user, personal-use setup was re-reviewed against each prov
| OpenRouter | ✅ Likely OK | April 2026 ToS sharpens the no-resale / no-competing-service clause; private single-user proxy still fine. |
| Cloudflare Workers AI | ⚠️ Ambiguous | No anti-proxy clause; covered by general Self-Serve Subscription Agreement. |
| NVIDIA NIM | ⚠️ Caution | Trial ToS §1.2 / §1.4: *"evaluation only, not production."* Free access is a recurring 40 RPM rate limit (the 2025 credit system was discontinued), but the evaluation-only scope stands. |
-| GitHub Models | ⚠️ Caution | Free tier explicitly scoped to *"experimentation"* and *"prototyping."* |
| Cohere | ❌ Avoid | Terms §14 still forbids *"personal, family or household purposes."* |
| Zhipu (open.bigmodel.cn) | ✅ Likely OK | Personal/non-commercial research carve-out still in the platform docs. |
| Z.ai (api.z.ai) | ⚠️ Caution | New row — Singapore entity (distinct from Zhipu CN). §III.3(l) anti-traffic-redirect clause could plausibly be read against a proxy; no explicit personal-use carve-out. |
diff --git a/docs/en/architecture/03-streaming-pipeline.md b/docs/en/architecture/03-streaming-pipeline.md
index c9ece226..140323cd 100644
--- a/docs/en/architecture/03-streaming-pipeline.md
+++ b/docs/en/architecture/03-streaming-pipeline.md
@@ -231,7 +231,6 @@ Thin adapter over shared fallback loop:
| `rememberReasoning(sessionKey, modelKey, reasoning)` | Store per-session trace |
| `rescueInlineToolCalls(text, tools)` | Convert inline dialect → structured `tool_calls` |
| `repairToolArguments(args, schemas)` | Fix double-encoded JSON against request schemas |
-| `truncateMessagesForGithub(messages)` | Trim history to GitHub Models input ceiling |
| `normalizeOutboundContent(chunk)` | Provider-specific response normalization |
| `sanitizeResponse(chunk)` | Redact secrets from streamed frames |
diff --git a/docs/en/providers/01-supported-platforms.md b/docs/en/providers/01-supported-platforms.md
index 2df8c030..7d3de037 100644
--- a/docs/en/providers/01-supported-platforms.md
+++ b/docs/en/providers/01-supported-platforms.md
@@ -27,7 +27,6 @@ Of the 43 built-in platforms, **10 use dedicated adapters** and **33 ride `OpenA
| `mistral` | Mistral | Keyed | OpenAI-compat | `api.mistral.ai/v1`. |
| `sambanova` | SambaNova | Retired (V23) | Not registered | Free tier permanently gone; kept in the type union for historical key rows. |
| `openrouter` | OpenRouter | Keyed | OpenAI-compat | Sends `HTTP-Referer`/`X-Title` headers; `:free` model suffix marks zero-cost routes. |
-| `github` | GitHub Models | Keyed | OpenAI-compat | Routes to `models.github.ai/inference`; catalog uses `<publisher>/<model>` ids. |
| `cohere` | Cohere | Keyed | Native (`CohereProvider`) | Uses Cohere's OpenAI-compatible endpoint. |
| `cloudflare` | Cloudflare Workers AI | Keyed (`account_id:token`) | Native (`CloudflareProvider`) | Compound credential format parsed by the adapter. |
| `zhipu` | Zhipu (Z.ai / bigmodel.cn) | Keyed | Native (`ZhipuProvider`) | Console autodetect: a rejected domestic key is re-probed against the global `api.z.ai` host; 60s timeout covers glm-4.7-flash hidden reasoning. |
diff --git a/docs/zh-cn/api/01-rest-api.md b/docs/zh-cn/api/01-rest-api.md
index 69630f30..a038edeb 100644
--- a/docs/zh-cn/api/01-rest-api.md
+++ b/docs/zh-cn/api/01-rest-api.md
@@ -141,7 +141,7 @@ final = client.chat.completions.create(
print(final.choices[0].message.content)
```
-配合 `stream=True` 也可以,你会先收到一串 `delta.tool_calls` 分块,然后是一个 `finish_reason: "tool_calls"` 收尾。底层实现上,兼容 OpenAI 的提供方(Groq、Cerebras、Mistral、OpenRouter、GitHub Models、HuggingFace、Cloudflare、Cohere 兼容模式)会直接透传请求;发往 Gemini 的请求会被翻译成 Google 的 `functionDeclarations` / `functionResponse` 结构,响应再翻译回来。
+配合 `stream=True` 也可以,你会先收到一串 `delta.tool_calls` 分块,然后是一个 `finish_reason: "tool_calls"` 收尾。底层实现上,兼容 OpenAI 的提供方(Groq、Cerebras、Mistral、OpenRouter、HuggingFace、Cloudflare、Cohere 兼容模式)会直接透传请求;发往 Gemini 的请求会被翻译成 Google 的 `functionDeclarations` / `functionResponse` 结构,响应再翻译回来。
## Gemini 的 Google 搜索接地
@@ -251,8 +251,6 @@ curl http://localhost:3001/v1/embeddings \
| 系列(`model`) | 维度 | 提供方(故障转移顺序) |
| --- | --- | --- |
| `gemini-embedding-001` *(默认)* | 3072 | Google |
-| `text-embedding-3-large` | 3072 | GitHub Models |
-| `text-embedding-3-small` | 1536 | GitHub Models |
| `embed-v4.0` | 1536 | Cohere |
| `bge-m3` | 1024 | Cloudflare → Hugging Face |
| `qwen3-embedding-0.6b` | 1024 | Cloudflare |
diff --git a/docs/zh-cn/architecture/00-high-level-index.md b/docs/zh-cn/architecture/00-high-level-index.md
index 9d90cfef..5f2329d0 100644
--- a/docs/zh-cn/architecture/00-high-level-index.md
+++ b/docs/zh-cn/architecture/00-high-level-index.md
@@ -85,7 +85,6 @@ FreeLLMAPI 是一个自托管的、兼容 OpenAI 的网关,把约 34 家提供
| OpenRouter | ✅ 大概率可以 | 2026 年 4 月的条款强化了禁止转售、禁止竞争性服务的条款;私有单用户代理仍然没问题。 |
| Cloudflare Workers AI | ⚠️ 模糊 | 没有反代理条款;受通用的自助订阅协议约束。 |
| NVIDIA NIM | ⚠️ 谨慎 | 试用条款 §1.2 / §1.4:*“仅供评估,不用于生产。”* 免费访问是持续的 40 RPM 限流(2025 年的积分制已停止),但仅供评估的范围依然有效。 |
-| GitHub Models | ⚠️ 谨慎 | 免费层明确限定为*“实验”*和*“原型验证”*。 |
| Cohere | ❌ 避免 | 条款 §14 仍禁止*“个人、家庭或家用目的”*。 |
| 智谱(open.bigmodel.cn) | ✅ 大概率可以 | 平台文档中仍保留个人、非商业研究的豁免。 |
| Z.ai(api.z.ai) | ⚠️ 谨慎 | 新条目 —— 新加坡实体(与智谱中国不同)。§III.3(l) 的反流量重定向条款有可能被解读为针对代理;没有明确的个人使用豁免。 |
diff --git a/docs/zh-cn/architecture/03-streaming-pipeline.md b/docs/zh-cn/architecture/03-streaming-pipeline.md
index 99ec3c4e..9a9c4e86 100644
--- a/docs/zh-cn/architecture/03-streaming-pipeline.md
+++ b/docs/zh-cn/architecture/03-streaming-pipeline.md
@@ -231,7 +231,6 @@ Codex CLI 和 Agents SDK 说 **Responses API**(有状态、`previous_response_
| `rememberReasoning(sessionKey, modelKey, reasoning)` | 按会话存轨迹 |
| `rescueInlineToolCalls(text, tools)` | 内联方言 → 结构化 `tool_calls` |
| `repairToolArguments(args, schemas)` | 对请求 schema 修双重编码 JSON |
-| `truncateMessagesForGithub(messages)` | 裁历史到 GitHub Models 输入天花板 |
| `normalizeOutboundContent(chunk)` | 提供方专属响应归一化 |
| `sanitizeResponse(chunk)` | 从流式帧脱敏机密 |
diff --git a/docs/zh-cn/providers/01-supported-platforms.md b/docs/zh-cn/providers/01-supported-platforms.md
index c819f535..479587a8 100644
--- a/docs/zh-cn/providers/01-supported-platforms.md
+++ b/docs/zh-cn/providers/01-supported-platforms.md
@@ -23,7 +23,6 @@
| `mistral` | Mistral | 带密钥 | OpenAI 兼容 | `api.mistral.ai/v1`。 |
| `sambanova` | SambaNova | 已退役(V23) | 未注册 | 免费额度永久消失;因历史密钥行而保留在类型联合中。 |
| `openrouter` | OpenRouter | 带密钥 | OpenAI 兼容 | 发送 `HTTP-Referer`/`X-Title` 头;`:free` 模型后缀标记零成本路由。 |
-| `github` | GitHub Models | 带密钥 | OpenAI 兼容 | 路由到 `models.github.ai/inference`;目录使用 `<publisher>/<model>` 形式的 id。 |
| `cohere` | Cohere | 带密钥 | 原生(`CohereProvider`) | 使用 Cohere 的 OpenAI 兼容端点。 |
| `cloudflare` | Cloudflare Workers AI | 带密钥(`account_id:token`) | 原生(`CloudflareProvider`) | 复合凭据格式由适配器解析。 |
| `zhipu` | Zhipu (Z.ai / bigmodel.cn) | 带密钥 | 原生(`ZhipuProvider`) | 控制台自动探测:国内密钥被拒后会改向全球 `api.z.ai` 主机重新探测;60s 超时覆盖 glm-4.7-flash 的隐藏推理。 |
diff --git a/server/src/__tests__/db/idempotency.test.ts b/server/src/__tests__/db/idempotency.test.ts
index fe8bbd2f..1050e11f 100644
--- a/server/src/__tests__/db/idempotency.test.ts
+++ b/server/src/__tests__/db/idempotency.test.ts
@@ -192,16 +192,13 @@ describe('Migration idempotency', () => {
]);
});
- it('caps GitHub GPT-4.1 at the routable free-tier context window (#426)', () => {
+ it('removes the retired GitHub Models provider (#1557)', () => {
process.env.ENCRYPTION_KEY = '0'.repeat(64);
const db = initDb(':memory:');
- const row = db.prepare(`
- SELECT context_window FROM models
- WHERE platform = 'github' AND model_id = 'openai/gpt-4.1'
- `).get() as { context_window: number };
-
- expect(row.context_window).toBe(8000);
+ for (const table of ['models', 'embedding_models', 'api_keys']) {
+ expect((db.prepare(`SELECT COUNT(*) AS n FROM ${table} WHERE platform = 'github'`).get() as { n: number }).n).toBe(0);
+ }
});
it('V14: cerebras deprecation disables qwen-3-235b and llama3.1-8b but keeps gpt-oss-120b enabled', () => {
diff --git a/server/src/__tests__/db/intelligence-tiers.test.ts b/server/src/__tests__/db/intelligence-tiers.test.ts
index 58c14b3a..f18d1ff6 100644
--- a/server/src/__tests__/db/intelligence-tiers.test.ts
+++ b/server/src/__tests__/db/intelligence-tiers.test.ts
@@ -43,7 +43,6 @@ describe('intelligence tier audit (migrateModelsV17)', () => {
it('demotes lapped models down from Large', () => {
expect(tier('google', 'gemini-2.5-flash')).toBe('Medium'); // AA 21
- expect(tier('github', 'gpt-4o')).toBe('Medium'); // AA 17
expect(tier('cohere', 'command-a-03-2025')).toBe('Medium'); // AA 13
expect(tier('ollama', 'devstral-2:123b')).toBe('Medium'); // AA 22
});
diff --git a/server/src/__tests__/db/migrate/endpoint-identity.test.ts b/server/src/__tests__/db/migrate/endpoint-identity.test.ts
index 8192b9aa..81377ea5 100644
--- a/server/src/__tests__/db/migrate/endpoint-identity.test.ts
+++ b/server/src/__tests__/db/migrate/endpoint-identity.test.ts
@@ -74,7 +74,8 @@ describe('custom model endpoint identity migration', () => {
).all(seededProfileId, seededModelId);
const fallbackBefore = chainOf();
const profileBefore = profileOf();
- const totalBefore = (db.prepare('SELECT COUNT(*) AS n FROM models').get() as { n: number }).n;
+ const countModels = () => (db.prepare("SELECT COUNT(*) AS n FROM models WHERE platform != 'github'").get() as { n: number }).n;
+ const totalBefore = countModels();
await runMigrations(db, 'up');
@@ -87,9 +88,9 @@ describe('custom model endpoint identity migration', () => {
expect(chainOf()).toEqual(fallbackBefore);
expect(profileOf()).toEqual(profileBefore);
expect(db.pragma('foreign_key_check')).toEqual([]);
- // Nothing was dropped on the way through (later migrations only add rows).
- expect((db.prepare('SELECT COUNT(*) AS n FROM models').get() as { n: number }).n)
- .toBeGreaterThanOrEqual(totalBefore);
+ // Nothing was dropped on the way through: later migrations only add rows,
+ // apart from the retired GitHub provider's (#1557).
+ expect(countModels()).toBeGreaterThanOrEqual(totalBefore);
} finally {
db.close();
}
diff --git a/server/src/__tests__/db/migrate/remove-github-provider.test.ts b/server/src/__tests__/db/migrate/remove-github-provider.test.ts
new file mode 100644
index 00000000..d7bb1f24
--- /dev/null
+++ b/server/src/__tests__/db/migrate/remove-github-provider.test.ts
@@ -0,0 +1,62 @@
+import Database from 'better-sqlite3';
+import { afterEach, describe, expect, it } from 'vitest';
+import { runMigrations } from '../../../db/migrate/runner.js';
+import { up, down } from '../../../db/migrations/20261017_000002_remove_github_provider.js';
+
+// GitHub Models was retired upstream (#549); leftover keys looked healthy and
+// custom GitHub models still joined the chain (#1557).
+
+const dbs: Database.Database[] = [];
+afterEach(() => {
+ for (const db of dbs.splice(0)) db.close();
+});
+
+async function migratedDb(): Promise<Database.Database> {
+ const db = new Database(':memory:');
+ dbs.push(db);
+ await runMigrations(db, 'up');
+ return db;
+}
+
+function githubCounts(db: Database.Database) {
+ const count = (sql: string) => (db.prepare(sql).get() as { n: number }).n;
+ return {
+ keys: count("SELECT COUNT(*) AS n FROM api_keys WHERE platform = 'github'"),
+ models: count("SELECT COUNT(*) AS n FROM models WHERE platform = 'github'"),
+ chain: count("SELECT COUNT(*) AS n FROM fallback_config WHERE model_db_id IN (SELECT id FROM models WHERE platform = 'github')"),
+ };
+}
+
+describe('remove_github_provider migration (#1557)', () => {
+ it('removes a leftover GitHub key and custom model, keeps other providers, and DOWN restores them', async () => {
+ const db = await migratedDb();
+ // The state a user upgrading from 0.15.0 has: a GitHub key and a custom
+ // GitHub model in the fallback chain.
+ const keyId = Number(db.prepare(`
+ INSERT INTO api_keys (platform, label, encrypted_key, iv, auth_tag, status, enabled)
+ VALUES ('github', 'gh', 'enc', 'iv', 'tag', 'healthy', 1)
+ `).run().lastInsertRowid);
+ const modelId = Number(db.prepare(`
+ INSERT INTO models (platform, model_id, display_name, intelligence_rank, speed_rank, enabled, source)
+ VALUES ('github', 'openai/gpt-4.1', 'GPT-4.1', 20, 7, 1, 'user')
+ `).run().lastInsertRowid);
+ db.prepare('INSERT INTO fallback_config (model_db_id, priority, enabled) VALUES (?, 999, 1)').run(modelId);
+ db.prepare(`
+ INSERT INTO api_keys (platform, label, encrypted_key, iv, auth_tag, status, enabled)
+ VALUES ('groq', 'groq', 'enc', 'iv', 'tag', 'healthy', 1)
+ `).run();
+ const otherModels = (db.prepare("SELECT COUNT(*) AS n FROM models WHERE platform != 'github'").get() as { n: number }).n;
+
+ up(db);
+ expect(githubCounts(db)).toEqual({ keys: 0, models: 0, chain: 0 });
+ expect(db.prepare("SELECT COUNT(*) AS n FROM api_keys WHERE platform = 'groq'").get()).toEqual({ n: 1 });
+ expect((db.prepare("SELECT COUNT(*) AS n FROM models WHERE platform != 'github'").get() as { n: number }).n).toBe(otherModels);
+ expect(db.pragma('foreign_key_check')).toEqual([]);
+
+ down(db);
+ expect(githubCounts(db)).toEqual({ keys: 1, models: 1, chain: 1 });
+ expect(db.prepare('SELECT id, status, enabled FROM api_keys WHERE platform = ?').get('github'))
+ .toEqual({ id: keyId, status: 'healthy', enabled: 1 });
+ expect(db.prepare("SELECT value FROM settings WHERE key = 'migration_github_removed_rows'").get()).toBeUndefined();
+ });
+});
diff --git a/server/src/__tests__/db/migrate/roundtrip.test.ts b/server/src/__tests__/db/migrate/roundtrip.test.ts
index b5cd1da8..f549c9da 100644
--- a/server/src/__tests__/db/migrate/roundtrip.test.ts
+++ b/server/src/__tests__/db/migrate/roundtrip.test.ts
@@ -46,6 +46,7 @@ const CUSTOM_ENDPOINT_GROUP_LABEL_FILENAME = '20260930_000001_custom_endpoint_gr
const SEARCH_ENGINES_FILENAME = '20261013_000001_search_engines.ts';
const DECISION_REQUESTS_FILENAME = '20261016_000001_decision_requests.ts';
const LLMTR_REENABLE_KEYS_FILENAME = '20261017_000001_llmtr_reenable_keys.ts';
+const REMOVE_GITHUB_PROVIDER_FILENAME = '20261017_000002_remove_github_provider.ts';
interface SchemaRow {
type: string;
@@ -138,6 +139,7 @@ describe('migration round trip', () => {
SEARCH_ENGINES_FILENAME,
DECISION_REQUESTS_FILENAME,
LLMTR_REENABLE_KEYS_FILENAME,
+ REMOVE_GITHUB_PROVIDER_FILENAME,
]);
} finally {
db.close();
diff --git a/server/src/__tests__/lib/content.test.ts b/server/src/__tests__/lib/content.test.ts
index cfb10c0f..a65e9929 100644
--- a/server/src/__tests__/lib/content.test.ts
+++ b/server/src/__tests__/lib/content.test.ts
@@ -1,5 +1,5 @@
import { describe, it, expect } from 'vitest';
-import { GITHUB_MAX_INPUT_TOKENS, contentToString, estimateInputTokens, flattenMessageContent, messageHasImage, normalizeOutboundContent, sanitizeResponse, stripImagesFromMessages, truncateMessagesForGithub } from '../../lib/content.js';
+import { contentToString, estimateInputTokens, flattenMessageContent, messageHasImage, normalizeOutboundContent, sanitizeResponse, stripImagesFromMessages } from '../../lib/content.js';
import type { ChatMessage } from '@freellmapi/shared/types.js';
describe('contentToString', () => {
@@ -204,73 +204,6 @@ describe('sanitizeResponse', () => {
});
});
-describe('truncateMessagesForGithub', () => {
- const msg = (role: ChatMessage['role'], content: string): ChatMessage => ({ role, content });
- // ~4000 tokens each under the chars/4 estimate the gateway uses everywhere.
- const big = (c: string) => c.repeat(16000);
-
- it('returns the same array when the request already fits', () => {
- const messages = [msg('system', 'be brief'), msg('user', 'hi')];
- expect(truncateMessagesForGithub(messages)).toBe(messages);
- });
-
- it('leaves an empty list alone', () => {
- const messages: ChatMessage[] = [];
- expect(truncateMessagesForGithub(messages)).toBe(messages);
- });
-
- it('drops the oldest turns, keeping the system prompt and the newest context', () => {
- const messages = [
- msg('system', 'be brief'),
- msg('user', big('a')),
- msg('assistant', big('b')),
- msg('user', big('c')),
- msg('user', 'the actual question'),
- ];
- const out = truncateMessagesForGithub(messages);
- expect(out).not.toBe(messages);
- expect(out[0]).toBe(messages[0]);
- expect(out[out.length - 1].content).toBe('the actual question');
- // Only one of the ~4000-token turns fits alongside the newest question.
- expect(out.map(m => m.content)).not.toContain(big('a'));
- expect(out.map(m => m.content)).toContain(big('c'));
- const total = out.reduce((sum, m) => sum + Math.ceil(String(m.content).length / 4), 0);
- expect(total).toBeLessThanOrEqual(GITHUB_MAX_INPUT_TOKENS);
- });
-
- it('keeps the system prompt verbatim wherever it sits in the list', () => {
- const system = msg('system', 'be brief');
- const messages = [msg('user', big('a')), system, msg('user', big('b'))];
- const out = truncateMessagesForGithub(messages);
- expect(out[0]).toBe(system);
- });
-
- it('truncates a single oversized message instead of sending nothing', () => {
- const out = truncateMessagesForGithub([msg('user', big('y').repeat(3))]);
- expect(out).toHaveLength(1);
- expect(String(out[0].content).length).toBe(GITHUB_MAX_INPUT_TOKENS * 4);
- });
-
- it('reserves the system prompt budget when truncating the newest message', () => {
- const system = msg('system', 'x'.repeat(400)); // 100 tokens
- const out = truncateMessagesForGithub([system, msg('user', big('z').repeat(3))]);
- expect(out[0]).toBe(system);
- expect(String(out[1].content).length).toBe((GITHUB_MAX_INPUT_TOKENS - 100) * 4);
- });
-
- it('honours a caller-supplied budget', () => {
- const messages = [msg('user', 'one'), msg('user', 'two'), msg('user', 'the newest')];
- const out = truncateMessagesForGithub(messages, 3); // 12 chars of room
- expect(out.map(m => m.content)).toEqual(['the newest']);
- });
-
- it('never rewrites multimodal content, only drops whole messages', () => {
- const image: ChatMessage = { role: 'user', content: [{ type: 'image_url', image_url: { url: 'data:x' } }] } as ChatMessage;
- const out = truncateMessagesForGithub([msg('user', big('a')), image], 1);
- expect(out).toEqual([image]);
- });
-});
-
describe('estimateInputTokens', () => {
const messages: ChatMessage[] = [{ role: 'user', content: 'x'.repeat(400) }];
diff --git a/server/src/__tests__/lib/sampling-params.test.ts b/server/src/__tests__/lib/sampling-params.test.ts
index cfa8aa94..05f1720d 100644
--- a/server/src/__tests__/lib/sampling-params.test.ts
+++ b/server/src/__tests__/lib/sampling-params.test.ts
@@ -148,7 +148,6 @@ describe('reasoning_effort (request-side reasoning control)', () => {
it('forwarded verbatim on openai-compat platforms without a droplist entry', () => {
expect(extendedBodyParams('groq', { reasoning_effort: 'medium' }).reasoning_effort).toBe('medium');
expect(extendedBodyParams('cerebras', { reasoning_effort: 'none' }).reasoning_effort).toBe('none');
- expect(extendedBodyParams('github', { reasoning_effort: 'high' }).reasoning_effort).toBe('high');
});
it('stripped for platforms with no support (never 400s a strict upstream)', () => {
@@ -206,10 +205,7 @@ describe('reasoning_effort normalization (#619 — off-scale values must not 400
expect(pickSamplingParams({ reasoning_effort: 'nonsense', seed: 3 })).toEqual({ seed: 3 });
});
- it('per-platform clamping: github and groq take only low/medium/high', () => {
- expect(extendedBodyParams('github', { reasoning_effort: 'none' }).reasoning_effort).toBe('low');
- expect(extendedBodyParams('github', { reasoning_effort: 'minimal' }).reasoning_effort).toBe('low');
- expect(extendedBodyParams('github', { reasoning_effort: 'medium' }).reasoning_effort).toBe('medium');
+ it('per-platform clamping: groq takes only low/medium/high', () => {
// Groq's enum is the same three values, so 'none'/'minimal' are clamped
// instead of being sent and 400ing the request off the platform.
expect(extendedBodyParams('groq', { reasoning_effort: 'none' }).reasoning_effort).toBe('low');
diff --git a/server/src/__tests__/lib/unified-max-tokens.test.ts b/server/src/__tests__/lib/unified-max-tokens.test.ts
index 0c7de2aa..7a8b032f 100644
--- a/server/src/__tests__/lib/unified-max-tokens.test.ts
+++ b/server/src/__tests__/lib/unified-max-tokens.test.ts
@@ -17,11 +17,13 @@ import {
resolveMaxTokens,
defaultMaxTokensFor,
maxTokensCapFor,
- GITHUB_MAX_OUTPUT_TOKENS,
UNIFIED_MAX_TOKENS_SETTING,
UNIFIED_MAX_TOKENS_AUTO,
} from '../../lib/sampling-params.js';
+// Moondream's documented output ceiling (PLATFORM_PARAM_POLICIES.moondream).
+const MOONDREAM_MAX_OUTPUT_TOKENS = 4096;
+
beforeEach(() => {
settingStore.clear();
});
@@ -100,32 +102,32 @@ describe('resolveMaxTokens under the cap', () => {
describe('per-platform max_tokens ceiling', () => {
it('is declared only for platforms whose API rejects large values', () => {
- expect(maxTokensCapFor('github')).toBe(GITHUB_MAX_OUTPUT_TOKENS);
+ expect(maxTokensCapFor('moondream')).toBe(MOONDREAM_MAX_OUTPUT_TOKENS);
expect(maxTokensCapFor('groq')).toBeUndefined();
expect(maxTokensCapFor('nonsense-platform')).toBeUndefined();
});
it('clamps an oversized request even with the operator cap off', () => {
expect(unifiedMaxTokensCap()).toBeNull();
- expect(resolveMaxTokens('github', 65536)).toBe(GITHUB_MAX_OUTPUT_TOKENS);
+ expect(resolveMaxTokens('moondream', 65536)).toBe(MOONDREAM_MAX_OUTPUT_TOKENS);
// Other platforms keep the historical pass-through behaviour.
expect(resolveMaxTokens('groq', 65536)).toBe(65536);
});
it('leaves a request at or below the platform ceiling alone', () => {
- expect(resolveMaxTokens('github', 128)).toBe(128);
- expect(resolveMaxTokens('github', GITHUB_MAX_OUTPUT_TOKENS)).toBe(GITHUB_MAX_OUTPUT_TOKENS);
+ expect(resolveMaxTokens('moondream', 128)).toBe(128);
+ expect(resolveMaxTokens('moondream', MOONDREAM_MAX_OUTPUT_TOKENS)).toBe(MOONDREAM_MAX_OUTPUT_TOKENS);
});
it('sends nothing when the client sent nothing and the platform has no floor', () => {
- expect(resolveMaxTokens('github', undefined)).toBeUndefined();
+ expect(resolveMaxTokens('moondream', undefined)).toBeUndefined();
});
it('takes the tighter of the platform ceiling and the operator cap', () => {
settingStore.set(UNIFIED_MAX_TOKENS_SETTING, 'auto'); // 32768 — looser
- expect(resolveMaxTokens('github', 65536)).toBe(GITHUB_MAX_OUTPUT_TOKENS);
+ expect(resolveMaxTokens('moondream', 65536)).toBe(MOONDREAM_MAX_OUTPUT_TOKENS);
settingStore.set(UNIFIED_MAX_TOKENS_SETTING, '128'); // tighter than the platform
- expect(resolveMaxTokens('github', 65536)).toBe(128);
+ expect(resolveMaxTokens('moondream', 65536)).toBe(128);
});
});
@@ -153,6 +155,6 @@ describe('contextBudget clamping', () => {
it('takes the tightest of context budget, platform cap, and operator cap', () => {
settingStore.set(UNIFIED_MAX_TOKENS_SETTING, '16000');
- expect(resolveMaxTokens('github', 64000, 48000)).toBe(Math.min(GITHUB_MAX_OUTPUT_TOKENS, 16000, 48000));
+ expect(resolveMaxTokens('moondream', 64000, 48000)).toBe(Math.min(MOONDREAM_MAX_OUTPUT_TOKENS, 16000, 48000));
});
});
diff --git a/server/src/__tests__/providers/openai-compat.test.ts b/server/src/__tests__/providers/openai-compat.test.ts
index 5d235a8d..2bb281cc 100644
--- a/server/src/__tests__/providers/openai-compat.test.ts
+++ b/server/src/__tests__/providers/openai-compat.test.ts
@@ -405,6 +405,27 @@ describe('OpenAICompatProvider', () => {
expect(await provider.validateKey('valid')).toBe(true);
});
+ it('validateKey treats a 200 that is not JSON as an error, not a healthy key (#1557)', async () => {
+ // Retired GitHub Models answered every path with 200 "OK" in plain text.
+ vi.spyOn(global, 'fetch').mockResolvedValueOnce({
+ ok: true,
+ status: 200,
+ headers: new Headers({ 'content-type': 'text/plain' }),
+ text: () => Promise.resolve('OK\n'),
+ } as any);
+ await expect(provider.validateKey('any')).rejects.toThrow(/did not answer JSON/);
+ });
+
+ it('validateKey accepts a JSON model list', async () => {
+ vi.spyOn(global, 'fetch').mockResolvedValueOnce({
+ ok: true,
+ status: 200,
+ headers: new Headers({ 'content-type': 'application/json' }),
+ text: () => Promise.resolve('{"object":"list","data":[]}'),
+ } as any);
+ expect(await provider.validateKey('valid')).toBe(true);
+ });
+
it('validateKey preserves the provider reason on confirmed 401', async () => {
vi.spyOn(global, 'fetch').mockResolvedValueOnce({
ok: false,
diff --git a/server/src/__tests__/routes/fallback.test.ts b/server/src/__tests__/routes/fallback.test.ts
index a8a4f047..e27d7a20 100644
--- a/server/src/__tests__/routes/fallback.test.ts
+++ b/server/src/__tests__/routes/fallback.test.ts
@@ -68,7 +68,8 @@ describe('Fallback API', () => {
const db = getDb();
const target = db.prepare(`
SELECT id, platform, model_id FROM models
- WHERE platform = 'github' AND model_id = 'openai/gpt-4.1'
+ WHERE platform = 'groq'
+ ORDER BY id LIMIT 1
`).get() as { id: number; platform: string; model_id: string };
const other = db.prepare(`
SELECT platform, model_id FROM models
diff --git a/server/src/__tests__/routes/proxy-github-limits.test.ts b/server/src/__tests__/routes/proxy-github-limits.test.ts
deleted file mode 100644
index 5b2699e0..00000000
--- a/server/src/__tests__/routes/proxy-github-limits.test.ts
+++ /dev/null
@@ -1,131 +0,0 @@
-import { describe, it, expect, beforeAll, beforeEach, afterEach, vi } from 'vitest';
-import type { Express } from 'express';
-import { createApp } from '../../app.js';
-import { initDb, getDb, getUnifiedApiKey } from '../../db/index.js';
-import { GITHUB_MAX_INPUT_TOKENS } from '../../lib/content.js';
-import { GITHUB_MAX_OUTPUT_TOKENS } from '../../lib/sampling-params.js';
-import { mintDashboardToken } from '../helpers/auth.js';
-
-// GitHub Models is the strictest free platform in the catalog: it 400s on a
-// max_tokens above its own ceiling and 413s on a long history instead of
-// truncating it. Both are pre-dispatch rejections, so without these guards a
-// github hop is a guaranteed wasted attempt on every oversized request — and
-// the same oversized body rides the rest of the chain.
-
-let dashToken = '';
-
-async function request(app: Express, path: string, body: any, headers: Record<string, string>) {
- const server = app.listen(0, '127.0.0.1');
- if (!server.listening) await new Promise<void>(resolve => server.once('listening', () => resolve()));
- const addr = server.address() as any;
- const res = await fetch(`http://127.0.0.1:${addr.port}${path}`, {
- method: 'POST',
- headers: { 'Content-Type': 'application/json', ...headers },
- body: JSON.stringify(body),
- });
- const text = await res.text();
- server.close();
- let json: any = null;
- try { json = JSON.parse(text); } catch { /* SSE body */ }
- return { status: res.status, body: json };
-}
-
-describe('github platform limits', () => {
- let app: Express;
- const sent: any[] = [];
-
- beforeAll(() => {
- process.env.ENCRYPTION_KEY = '0'.repeat(64);
- initDb(':memory:');
- app = createApp();
- dashToken = mintDashboardToken();
- });
-
- beforeEach(async () => {
- const db = getDb();
- db.prepare('DELETE FROM api_keys').run();
- db.prepare('DELETE FROM requests').run();
- db.prepare('DELETE FROM rate_limit_cooldowns').run();
- db.prepare('DELETE FROM rate_limit_usage').run();
- sent.length = 0;
- // Only a github key, so every route the chain can pick is a github one.
- const { status } = await request(app, '/api/keys',
- { platform: 'github', key: 'ghp_github_limits_test', label: 't' },
- { Authorization: `Bearer ${dashToken}` });
- expect(status).toBe(201);
-
- const origFetch = global.fetch;
- vi.spyOn(global, 'fetch').mockImplementation(async (url, init) => {
- const urlStr = typeof url === 'string' ? url : url.toString();
- if (!urlStr.includes('models.github.ai')) return origFetch(url as any, init);
- sent.push(JSON.parse(String((init as RequestInit).body)));
- return new Response(JSON.stringify({
- id: 'gh1', object: 'chat.completion', created: 1, model: 'gpt-4o',
- choices: [{ index: 0, message: { role: 'assistant', content: 'ok' }, finish_reason: 'stop' }],
- usage: { prompt_tokens: 10, completion_tokens: 1, total_tokens: 11 },
- }), { status: 200, headers: { 'Content-Type': 'application/json' } });
- });
- });
-
- afterEach(() => {
- vi.restoreAllMocks();
- });
-
- it('clamps an aggressive max_tokens to the platform ceiling instead of 400ing', async () => {
- const r = await request(app, '/v1/chat/completions', {
- model: 'gpt-4o', max_tokens: 65536,
- messages: [{ role: 'user', content: 'hello' }],
- }, { Authorization: `Bearer ${getUnifiedApiKey()}` });
-
- expect(r.status).toBe(200);
- expect(sent).toHaveLength(1);
- expect(sent[0].max_tokens).toBe(GITHUB_MAX_OUTPUT_TOKENS);
- });
-
- it('trims a long history to the platform input budget before dispatch', async () => {
- // 1900 tokens per turn under the chars/4 estimate: four of them plus the
- // question sit above the 7500-token input budget but still inside the
- // model's 8000-token context window, so routing picks github and the
- // guard — not the router — is what has to shrink the request. (max_tokens
- // 1 keeps the routing output reserve out of the way.)
- const turn = (c: string) => c.repeat(7600);
- const messages = [
- { role: 'system', content: 'be brief' },
- { role: 'user', content: turn('a') },
- { role: 'assistant', content: turn('b') },
- { role: 'user', content: turn('c') },
- { role: 'assistant', content: turn('d') },
- { role: 'user', content: 'the actual question' },
- ];
-
- const r = await request(app, '/v1/chat/completions', {
- model: 'gpt-4o', max_tokens: 1, messages,
- }, { Authorization: `Bearer ${getUnifiedApiKey()}` });
-
- expect(r.status).toBe(200);
- expect(sent).toHaveLength(1);
- const out = sent[0].messages as Array<{ role: string; content: string }>;
- // System prompt and the newest turn survive; the oldest turns are gone.
- expect(out[0].role).toBe('system');
- expect(out[out.length - 1].content).toBe('the actual question');
- expect(out.length).toBeLessThan(messages.length);
- expect(out.some(m => m.content === turn('a'))).toBe(false);
- const tokens = out.reduce((sum, m) => sum + Math.ceil(m.content.length / 4), 0);
- expect(tokens).toBeLessThanOrEqual(GITHUB_MAX_INPUT_TOKENS);
- });
-
- it('leaves a history that already fits untouched', async () => {
- const messages = [
- { role: 'system', content: 'be brief' },
- { role: 'user', content: 'first' },
- { role: 'assistant', content: 'second' },
- { role: 'user', content: 'third' },
- ];
- const r = await request(app, '/v1/chat/completions', {
- model: 'gpt-4o', max_tokens: 1, messages,
- }, { Authorization: `Bearer ${getUnifiedApiKey()}` });
-
- expect(r.status).toBe(200);
- expect(sent[0].messages.map((m: any) => m.content)).toEqual(messages.map(m => m.content));
- });
-});
diff --git a/server/src/__tests__/services/builtin-model-discovery.test.ts b/server/src/__tests__/services/builtin-model-discovery.test.ts
index 66f512b7..26249499 100644
--- a/server/src/__tests__/services/builtin-model-discovery.test.ts
+++ b/server/src/__tests__/services/builtin-model-discovery.test.ts
@@ -136,15 +136,6 @@ describe('built-in model discovery (#1348)', () => {
expect(builtinDiscoveryEligibility(getDb(), 'longcat')).toMatchObject({ eligible: false, reason: 'catalog_managed' });
});
- it('is off limits while the bundled baseline still holds catalog rows for the platform', () => {
- // A fresh DB seeds github rows until the first catalog sync prunes them.
- const github = rows('github');
- expect(github.length).toBeGreaterThan(0);
- expect(builtinDiscoveryEligibility(getDb(), 'github')).toMatchObject({ eligible: false, reason: 'catalog_managed' });
- getDb().exec("DELETE FROM fallback_config WHERE model_db_id IN (SELECT id FROM models WHERE platform = 'github'); DELETE FROM models WHERE platform = 'github'; DELETE FROM embedding_models WHERE platform = 'github';");
- expect(builtinDiscoveryEligibility(getDb(), 'github').eligible).toBe(true);
- });
-
it('can be switched off', () => {
process.env.BUILTIN_MODEL_DISCOVERY = 'off';
expect(builtinDiscoveryEligibility(getDb(), 'siliconflow')).toMatchObject({ eligible: false, reason: 'disabled' });
diff --git a/server/src/__tests__/services/catalog-sync.test.ts b/server/src/__tests__/services/catalog-sync.test.ts
index 70195964..bc994366 100644
--- a/server/src/__tests__/services/catalog-sync.test.ts
+++ b/server/src/__tests__/services/catalog-sync.test.ts
@@ -120,20 +120,6 @@ describe('applyCatalog', () => {
expect(fb).toBeTruthy();
});
- it('caps GitHub GPT-4.1 catalog context at the routable free-tier limit (#426)', () => {
- const models = existingAsCatalogModels();
- const target = models.find((m) => m.platform === 'github' && m.modelId === 'openai/gpt-4.1');
- expect(target).toBeDefined();
- target!.contextWindow = 128000;
-
- applyCatalog(getDb(), catalogOf(models));
-
- const row = getDb()
- .prepare("SELECT context_window FROM models WHERE platform = 'github' AND model_id = 'openai/gpt-4.1'")
- .get() as { context_window: number };
- expect(row.context_window).toBe(8000);
- });
-
it('updates metadata in place and respects the enabled policy', () => {
const models = existingAsCatalogModels();
const target = models.find((m) => m.modelId === 'brand-new-model')!;
diff --git a/server/src/__tests__/services/embeddings.test.ts b/server/src/__tests__/services/embeddings.test.ts
index b44d7eca..97e19cdb 100644
--- a/server/src/__tests__/services/embeddings.test.ts
+++ b/server/src/__tests__/services/embeddings.test.ts
@@ -442,17 +442,6 @@ describe('embeddings service', () => {
expect(String(fetchMock.mock.calls[0][0])).toContain('generativelanguage.googleapis.com');
});
- it('forwards dimensions to the github provider', async () => {
- addKey('github');
- const fetchMock = mockFetch(async () => okEmbeddingResponse(512));
-
- await runEmbeddings('text-embedding-3-small', ['hello'], 512);
-
- const body = JSON.parse(String((fetchMock.mock.calls[0][1] as RequestInit).body));
- expect(body.dimensions).toBe(512);
- expect(String(fetchMock.mock.calls[0][0])).toContain('models.github.ai');
- });
-
it('forwards dimensions to the openrouter provider', async () => {
addKey('openrouter');
const fetchMock = mockFetch(async () => okEmbeddingResponse(1024));
diff --git a/server/src/__tests__/services/router.test.ts b/server/src/__tests__/services/router.test.ts
index da57de9b..429290fc 100644
--- a/server/src/__tests__/services/router.test.ts
+++ b/server/src/__tests__/services/router.test.ts
@@ -136,42 +136,6 @@ describe('Router', () => {
expect(large.modelDbId).not.toBe(baseline.modelDbId);
});
- it('skips GitHub GPT-4.1 above its free-tier 8K request cap (#426)', () => {
- const db = getDb();
- const githubKey = encrypt('test-github-key');
- db.prepare(`
- INSERT INTO api_keys (platform, label, encrypted_key, iv, auth_tag, status, enabled)
- VALUES (?, ?, ?, ?, ?, ?, ?)
- `).run('github', 'github', githubKey.encrypted, githubKey.iv, githubKey.authTag, 'healthy', 1);
- const groqKey = encrypt('test-groq-key');
- db.prepare(`
- INSERT INTO api_keys (platform, label, encrypted_key, iv, auth_tag, status, enabled)
- VALUES (?, ?, ?, ?, ?, ?, ?)
- `).run('groq', 'groq', groqKey.encrypted, groqKey.iv, groqKey.authTag, 'healthy', 1);
-
- const github = db.prepare(`
- SELECT id, context_window FROM models
- WHERE platform = 'github' AND model_id = 'openai/gpt-4.1'
- `).get() as { id: number; context_window: number };
- const groq = db.prepare(`
- SELECT id FROM models
- WHERE platform = 'groq' AND context_window > 9000
- LIMIT 1
- `).get() as { id: number };
-
- db.prepare('UPDATE fallback_config SET priority = 1000, enabled = 1').run();
- db.prepare('UPDATE fallback_config SET priority = 1 WHERE model_db_id = ?').run(github.id);
- db.prepare('UPDATE fallback_config SET priority = 2 WHERE model_db_id = ?').run(groq.id);
- db.prepare(`
- UPDATE models SET tpm_limit = NULL, tpd_limit = NULL
- WHERE id IN (?, ?)
- `).run(github.id, groq.id);
-
- expect(github.context_window).toBe(8000);
- expect(routeRequest(7000).modelDbId).toBe(github.id);
- expect(routeRequest(9000).modelDbId).toBe(groq.id);
- });
-
it('prefers margin-fitting models but still serves raw-window fits (#956 review)', () => {
const db = getDb();
const groqKey = encrypt('test-groq-key');
diff --git a/server/src/db/migrate/defaults.ts b/server/src/db/migrate/defaults.ts
index b5f9b9e9..c5e3c068 100644
--- a/server/src/db/migrate/defaults.ts
+++ b/server/src/db/migrate/defaults.ts
@@ -41,6 +41,7 @@ import * as customEndpointGroupLabel from '../migrations/20260930_000001_custom_
import * as searchEngines from '../migrations/20261013_000001_search_engines.js';
import * as decisionRequests from '../migrations/20261016_000001_decision_requests.js';
import * as llmtrReenableKeys from '../migrations/20261017_000001_llmtr_reenable_keys.js';
+import * as removeGithubProvider from '../migrations/20261017_000002_remove_github_provider.js';
export interface MigrationModule {
up(db: Db): void;
@@ -94,6 +95,7 @@ export const QUOTA_SNAPSHOT_FRESHNESS_FILENAME = '20260915_000001_quota_snapshot
export const SEARCH_ENGINES_FILENAME = '20261013_000001_search_engines.ts';
export const DECISION_REQUESTS_FILENAME = '20261016_000001_decision_requests.ts';
export const LLMTR_REENABLE_KEYS_FILENAME = '20261017_000001_llmtr_reenable_keys.ts';
+export const REMOVE_GITHUB_PROVIDER_FILENAME = '20261017_000002_remove_github_provider.ts';
export const DEFAULT_MIGRATIONS: readonly DefaultMigration[] = [
{ filename: LEGACY_BASELINE_FILENAME, module: legacyBaseline },
@@ -138,4 +140,5 @@ export const DEFAULT_MIGRATIONS: readonly DefaultMigration[] = [
{ filename: SEARCH_ENGINES_FILENAME, module: searchEngines },
{ filename: DECISION_REQUESTS_FILENAME, module: decisionRequests },
{ filename: LLMTR_REENABLE_KEYS_FILENAME, module: llmtrReenableKeys },
+ { filename: REMOVE_GITHUB_PROVIDER_FILENAME, module: removeGithubProvider },
];
diff --git a/server/src/db/migrations/20261017_000002_remove_github_provider.ts b/server/src/db/migrations/20261017_000002_remove_github_provider.ts
new file mode 100644
index 00000000..a3dbc731
--- /dev/null
+++ b/server/src/db/migrations/20261017_000002_remove_github_provider.ts
@@ -0,0 +1,92 @@
+// Migration: remove the retired GitHub Models provider
+// Created: 2026-10-17
+//
+// DOWN: reversible (the removed rows are kept in settings)
+//
+// GitHub Models was retired on 30 July 2026 (#549). models.github.ai now
+// answers every request with 200 "OK" in plain text, so leftover GitHub keys
+// looked healthy, custom GitHub models joined the fallback chain, and every
+// request routed there failed (#1557). The adapter is gone, so these rows can
+// never serve again: remove the keys, models, embedding and media rows, model
+// overrides, and the chain and profile entries that point at the models.
+// Request history keeps its 'github' rows; that is a record, not config.
+
+import type { Db } from '../types.js';
+
+const SETTING_KEY = 'migration_github_removed_rows';
+
+type Row = Record<string, unknown>;
+
+// Insert order for DOWN: parents before the rows that reference them.
+const TABLES = ['api_keys', 'models', 'embedding_models', 'media_models', 'model_overrides', 'fallback_config', 'profile_models'] as const;
+type Table = (typeof TABLES)[number];
+
+function tableExists(db: Db, table: string): boolean {
+ return db.prepare("SELECT 1 FROM sqlite_master WHERE type = 'table' AND name = ?").get(table) !== undefined;
+}
+
+function selectRows(db: Db, table: Table): Row[] {
+ if (!tableExists(db, table)) return [];
+ if (table === 'fallback_config' || table === 'profile_models') {
+ return db.prepare(`
+ SELECT * FROM ${table} WHERE model_db_id IN (SELECT id FROM models WHERE platform = 'github')
+ `).all() as Row[];
+ }
+ return db.prepare(`SELECT * FROM ${table} WHERE platform = 'github'`).all() as Row[];
+}
+
+// Encrypted key columns may be BLOBs; JSON keeps them as tagged base64.
+function encode(row: Row): Row {
+ const out: Row = {};
+ for (const [k, v] of Object.entries(row)) {
+ out[k] = Buffer.isBuffer(v) ? { $buffer: v.toString('base64') } : v;
+ }
+ return out;
+}
+
+function decode(row: Row): Row {
+ const out: Row = {};
+ for (const [k, v] of Object.entries(row)) {
+ const tagged = v as { $buffer?: string } | null;
+ out[k] = tagged && typeof tagged === 'object' && typeof tagged.$buffer === 'string'
+ ? Buffer.from(tagged.$buffer, 'base64')
+ : v;
+ }
+ return out;
+}
+
+export function up(db: Db): void {
+ const removed: Partial<Record<Table, Row[]>> = {};
+ for (const table of TABLES) removed[table] = selectRows(db, table).map(encode);
+
+ // Children first: chain and profile rows reference models(id).
+ for (const table of ['fallback_config', 'profile_models'] as const) {
+ if (!tableExists(db, table)) continue;
+ db.prepare(`
+ DELETE FROM ${table} WHERE model_db_id IN (SELECT id FROM models WHERE platform = 'github')
+ `).run();
+ }
+ for (const table of ['models', 'embedding_models', 'media_models', 'model_overrides', 'api_keys'] as const) {
+ if (!tableExists(db, table)) continue;
+ db.prepare(`DELETE FROM ${table} WHERE platform = 'github'`).run();
+ }
+
+ db.prepare(`
+ INSERT INTO settings (key, value) VALUES (?, ?)
+ ON CONFLICT(key) DO UPDATE SET value = excluded.value
+ `).run(SETTING_KEY, JSON.stringify(removed));
+}
+
+export function down(db: Db): void {
+ const row = db.prepare('SELECT value FROM settings WHERE key = ?').get(SETTING_KEY) as { value: string } | undefined;
+ const removed: Partial<Record<Table, Row[]>> = row ? JSON.parse(row.value) : {};
+ for (const table of TABLES) {
+ for (const raw of removed[table] ?? []) {
+ const r = decode(raw);
+ const cols = Object.keys(r);
+ db.prepare(`INSERT OR IGNORE INTO ${table} (${cols.join(', ')}) VALUES (${cols.map(() => '?').join(', ')})`)
+ .run(...cols.map(c => r[c]));
+ }
+ }
+ db.prepare('DELETE FROM settings WHERE key = ?').run(SETTING_KEY);
+}
diff --git a/server/src/db/model-pricing.ts b/server/src/db/model-pricing.ts
index 474f9084..695f3155 100644
--- a/server/src/db/model-pricing.ts
+++ b/server/src/db/model-pricing.ts
@@ -45,11 +45,6 @@ export const MODEL_PRICING: PricingRow[] = [
['cohere', 'command-r-08-2024', 0.15, 0.60],
['cohere', 'command-r-plus-08-2024', 2.50, 10.00],
- // GitHub Models (OpenAI official prices)
- ['github', 'gpt-4o', 2.50, 10.00],
- ['github', 'openai/gpt-4.1', 2.00, 8.00],
- ['github', 'openai/gpt-5', 1.25, 10.00], // legacy
-
// Google AI Studio (official prices)
['google', 'gemini-2.5-flash', 0.30, 2.50],
['google', 'gemini-2.5-flash-lite', 0.10, 0.40],
diff --git a/server/src/lib/content.ts b/server/src/lib/content.ts
index e2c1eed7..c2411248 100644
--- a/server/src/lib/content.ts
+++ b/server/src/lib/content.ts
@@ -88,75 +88,6 @@ export function stripImagesFromMessages(messages: ChatMessage[]): ChatMessage[]
});
}
-// GitHub Models refuses a long history outright (413) rather than truncating
-// it the way most providers do, so an oversized body turns every github hop in
-// the fallback chain into a wasted attempt. 7500 is the input budget we trim
-// to: comfortably inside the ~8000-token ceiling observed live, with room for
-// the tool schemas and the system prompt the gateway prepends.
-export const GITHUB_MAX_INPUT_TOKENS = 7500;
-
-// The proxy's own chars/4 input estimate, per message — same arithmetic the
-// routing/budget paths use, so the guard below trims against the number the
-// rest of the gateway reasons about.
-function estimateMessageTokens(message: ChatMessage): number {
- return Math.ceil(contentToString(message.content).length / 4);
-}
-
-// Trim one message's text down to a token budget. Non-string content
-// (multimodal envelopes) is returned untouched — dropping or rewriting image
-// blocks here would corrupt the envelope, and text is the only thing this
-// guard is allowed to shorten.
-function truncateMessageText(message: ChatMessage, budget: number): ChatMessage {
- if (typeof message.content !== 'string') return message;
- const maxChars = Math.max(0, budget) * 4;
- return message.content.length <= maxChars
- ? message
- : { ...message, content: message.content.slice(0, maxChars) };
-}
-
-/**
- * Fit a message list inside GitHub Models' input ceiling before dispatch.
- * Keeps the leading system prompt verbatim (its instructions are the whole
- * point of prepending one) and then as many of the NEWEST messages as the
- * remaining budget allows, dropping the oldest turns first — the same trade a
- * client's own context window makes. A single message that blows the budget on
- * its own is truncated rather than dropped, so the turn is never dispatched
- * with nothing to answer.
- *
- * Returns the ORIGINAL array when the request already fits, which is both the
- * common case and the cheap one: callers can hand every github-routed request
- * through here without copying anything.
- */
-export function truncateMessagesForGithub(
- messages: ChatMessage[],
- budget: number = GITHUB_MAX_INPUT_TOKENS,
-): ChatMessage[] {
- if (messages.length === 0) return messages;
- const total = messages.reduce((sum, m) => sum + estimateMessageTokens(m), 0);
- if (total <= budget) return messages;
-
- const systemIdx = messages.findIndex(m => m.role === 'system');
- const systemMessage = systemIdx >= 0 ? messages[systemIdx] : null;
- const rest = messages.filter((_, i) => i !== systemIdx);
-
- const systemTokens = systemMessage ? estimateMessageTokens(systemMessage) : 0;
- let used = systemTokens;
- const kept: ChatMessage[] = [];
- // Newest → oldest: the most recent context is the context the answer needs.
- for (let i = rest.length - 1; i >= 0; i--) {
- const tokens = estimateMessageTokens(rest[i]);
- if (used + tokens > budget) break;
- kept.unshift(rest[i]);
- used += tokens;
- }
- // Nothing fit alongside the system prompt: keep the newest message in
- // truncated form instead of sending an empty conversation.
- if (kept.length === 0 && rest.length > 0) {
- kept.push(truncateMessageText(rest[rest.length - 1], budget - systemTokens));
- }
- return systemMessage ? [systemMessage, ...kept] : kept;
-}
-
// Harden the OUTBOUND envelope so strict OpenAI clients don't choke on the
// shape variations free-tier providers emit. Complements normalizeOutboundContent
// (which fixes array content) by fixing the *frame* fields, not the content:
diff --git a/server/src/lib/key-parser.ts b/server/src/lib/key-parser.ts
index e2c20e68..078062c7 100644
--- a/server/src/lib/key-parser.ts
+++ b/server/src/lib/key-parser.ts
@@ -95,7 +95,6 @@ export const PREFIX_MAP: Record<string, string> = {
NVIDIA_: 'nvidia',
MISTRAL_: 'mistral',
OPENROUTER_: 'openrouter',
- GITHUB_: 'github',
COHERE_: 'cohere',
TAVILY_: 'tavily',
EXA_: 'exa',
diff --git a/server/src/lib/sampling-params.ts b/server/src/lib/sampling-params.ts
index 0262a6b3..20b78daa 100644
--- a/server/src/lib/sampling-params.ts
+++ b/server/src/lib/sampling-params.ts
@@ -16,8 +16,7 @@
// Forward-by-default: most OpenAI-compatible providers ignore unknown body
// fields, so the default is to send everything and let the per-platform
// droplist name the documented exceptions (Mistral 422s on unknown keys,
-// Groq 400s on the logprobs family, Azure-backed GitHub Models rejects
-// non-Azure knobs). A provider that still 400s fails over like any other
+// Groq 400s on the logprobs family). A provider that still 400s fails over like any other
// provider-invalid request and shows up in the attempt trail — and its
// droplist entry is one line to add.
@@ -233,11 +232,6 @@ export interface PlatformParamPolicy {
maxTokensCap?: number;
}
-/** GitHub Models' own output-token ceiling: asking for more 400s ("max_tokens
- * is too large"), so the request never reaches the model. Wired into the
- * github policy below as maxTokensCap. */
-export const GITHUB_MAX_OUTPUT_TOKENS = 400;
-
// Keyed by Platform (not string) so a typo'd platform id fails tsc instead of
// silently no-op'ing the policy; the string-typed accessors below cast at the
// boundary since routes carry platform ids as plain strings.
@@ -274,7 +268,7 @@ export const PLATFORM_PARAM_POLICIES: Partial<Record<Platform, PlatformParamPoli
},
// Groq documents logprobs / top_logprobs / logit_bias as unsupported and
// rejects requests that include them. Its reasoning_effort enum is the plain
- // low/medium/high one as well: 'none' and 'minimal' are rejected with a 400
+ // low/medium/high one: 'none' and 'minimal' are rejected with a 400
// ("must be one of low, medium, or high", observed on openai/gpt-oss-120b),
// which cost the request a fallback hop onto another platform's copy of the
// same model. Clamped rather than dropped so the knob still arrives, at the
@@ -283,17 +277,6 @@ export const PLATFORM_PARAM_POLICIES: Partial<Record<Platform, PlatformParamPoli
drop: ['logprobs', 'top_logprobs', 'logit_bias'],
reasoningEfforts: ['low', 'medium', 'high'],
},
- // GitHub Models sits on Azure OpenAI, which 400s "Unrecognized request
- // argument" for knobs outside the OpenAI set. Its reasoning_effort enum is
- // the older low/medium/high one, so 'none'/'minimal' are clamped rather
- // than sent. Its free tier also refuses any max_tokens above
- // GITHUB_MAX_OUTPUT_TOKENS, so the cap is clamped here instead of being
- // spent as a wasted fallback hop.
- github: {
- drop: ['top_k', 'min_p', 'repetition_penalty'],
- reasoningEfforts: ['low', 'medium', 'high'],
- maxTokensCap: GITHUB_MAX_OUTPUT_TOKENS,
- },
// Gemini's generationConfig has no equivalents for these; the adapter
// translates the rest natively (topK, seed, penalties, responseSchema, and
// reasoning_effort → thinkingConfig — see toGeminiExtendedConfig).
diff --git a/server/src/providers/index.ts b/server/src/providers/index.ts
index 9e7c3791..42ebcc1d 100644
--- a/server/src/providers/index.ts
+++ b/server/src/providers/index.ts
@@ -196,14 +196,10 @@ register(new OpenAICompatProvider({
},
}));
-// GitHub Models — OpenAI-compatible. Catalog uses `<publisher>/<model>` ids
-// (e.g. `openai/gpt-4.1`); the old Azure endpoint rejects that prefix with
-// "Unknown model", so route to the current models.github.ai endpoint.
-register(new OpenAICompatProvider({
- platform: 'github',
- name: 'GitHub Models',
- baseUrl: 'https://models.github.ai/inference',
-}));
+// GitHub Models was retired upstream on 30 July 2026 (#549): models.github.ai
+// now answers every request with 200 "OK" in plain text, so a registered
+// adapter only produced fake-healthy keys and broken streams (#1557). The
+// provider is gone; migration 20261017_000002 removes leftover rows.
// Cohere - OpenAI-compatible via Cohere compatibility endpoint
register(new CohereProvider());
diff --git a/server/src/providers/openai-compat.ts b/server/src/providers/openai-compat.ts
index d4fb0f38..6d5d20bc 100644
--- a/server/src/providers/openai-compat.ts
+++ b/server/src/providers/openai-compat.ts
@@ -602,6 +602,20 @@ export class OpenAICompatProvider extends BaseProvider {
async validateKey(apiKey: string, quotaContext?: QuotaObservationContext): Promise<KeyValidationResult> {
const res = await this.fetchCatalogEndpoint(this.validateUrl ?? this.modelsUrl, apiKey, quotaContext);
+ // An OpenAI-compatible /models answers JSON. A 200 with anything else is a
+ // shut-down or parked service, not a working key: retired GitHub Models
+ // answered every path with 200 "OK" in plain text and every key looked
+ // healthy (#1557). Thrown, so health records an error without counting it
+ // toward auto-disable. Provider-specific validate URLs keep their own shape.
+ if (res.ok && !this.validateUrl && typeof res.text === 'function') {
+ const text = await res.text().catch(() => '');
+ try {
+ JSON.parse(text);
+ } catch {
+ throw providerHttpError(res, `${this.name} /models did not answer JSON (got ${JSON.stringify(text.trim().slice(0, 40))}); the service may be retired or down`);
+ }
+ return true;
+ }
return this.validationResult(res);
}
diff --git a/server/src/routes/keys.ts b/server/src/routes/keys.ts
index 9840ee30..37523715 100644
--- a/server/src/routes/keys.ts
+++ b/server/src/routes/keys.ts
@@ -41,7 +41,8 @@ export const keysRouter = Router();
// Active providers — must match providers/index.ts registrations + shared/types.ts Platform.
// Moonshot and MiniMax direct integrations were dropped in V4. HuggingFace
// was dropped in V4 and re-added in V13 via the router.huggingface.co route.
-// SambaNova was dropped in V23 (free tier permanently retired).
+// SambaNova was dropped in V23 (free tier permanently retired). GitHub Models
+// was dropped after its retirement on 30 July 2026 (#1557).
const PLATFORMS = [
'aclide',
'speka',
@@ -56,7 +57,7 @@ const PLATFORMS = [
'moondream',
'freeai',
'google', 'groq', 'cerebras', 'sail', 'electronhub', 'experiential', 'router9', 'septor', 'clod', 'speechify', 'blaze', 'lucidity', 'airforce', 'dreamprompting', 'waterfall', 'logfare', 'bai', 'radeon', 'nvidia', 'mistral',
- 'openrouter', 'github', 'cohere', 'tavily', 'exa', 'firecrawl', 'linkup', 'parallel', 'cloudflare', 'zhipu', 'ollama',
+ 'openrouter', 'cohere', 'tavily', 'exa', 'firecrawl', 'linkup', 'parallel', 'cloudflare', 'zhipu', 'ollama',
'kilo', 'pollinations', 'llm7', 'huggingface', 'opencode', 'ovh', 'agnes', 'reka', 'siliconflow',
'routeway', 'bazaarlink', 'ainative', 'aion', 'anyapi', 'requesty', 'navy', 'nara', 'sealion', 'orcarouter', 'unorouter', 'xkiro', 'modelscope',
'qianfan', 'volcengine', 'longcat', 'xfyun', 'aihorde', 'custom',
diff --git a/server/src/routes/proxy.ts b/server/src/routes/proxy.ts
index e4d69146..d06b3520 100644
--- a/server/src/routes/proxy.ts
+++ b/server/src/routes/proxy.ts
@@ -14,7 +14,7 @@ import { runImageGeneration, runVideoGeneration, runSpeech, runTranscription, Me
import multer from 'multer';
import { getDb } from '../db/index.js';
import { resolveAuth, prependSystemPrompt, type ResolvedAuth } from '../lib/system-prompt.js';
-import { contentToString, estimateInputTokens, messageHasImage, normalizeOutboundContent, sanitizeResponse, truncateMessagesForGithub } from '../lib/content.js';
+import { contentToString, estimateInputTokens, messageHasImage, normalizeOutboundContent, sanitizeResponse } from '../lib/content.js';
import { routeOutputBudget } from '../lib/output-cap.js';
import { resolveTaskType } from '../lib/task-type.js';
import { normalizeMessageImages } from '../lib/image-normalize.js';
@@ -1398,12 +1398,6 @@ proxyRouter.post('/completions', async (req: Request, res: Response) => {
requestedModel: attempt === 0 ? requestedModelLabel : undefined,
});
- // Same GitHub input ceiling as /chat/completions below: trim the
- // dispatched copy so a long legacy prompt doesn't 413 the github hop.
- const dispatchMessages = route.platform === 'github'
- ? truncateMessagesForGithub(messages)
- : messages;
-
if (stream) {
let totalOutputTokens = 0;
let headerSent = false;
@@ -1434,7 +1428,7 @@ proxyRouter.post('/completions', async (req: Request, res: Response) => {
try {
const gen = route.provider.streamChatCompletion(
route.apiKey,
- dispatchMessages,
+ messages,
route.modelId,
{ temperature, max_tokens, top_p, stop, contextBudget, signal: AbortSignal.any([clientAbort.signal, hedgeAbort.signal]) },
quotaContextForRoute(route, 'chat/completions'),
@@ -1534,7 +1528,7 @@ proxyRouter.post('/completions', async (req: Request, res: Response) => {
const result = await route.provider.chatCompletion(
route.apiKey,
- dispatchMessages,
+ messages,
route.modelId,
{ temperature, max_tokens, top_p, stop, contextBudget, signal: AbortSignal.any([clientAbort.signal, hedgeAbort.signal]) },
quotaContextForRoute(route, 'chat/completions'),
@@ -2353,16 +2347,6 @@ proxyRouter.post('/chat/completions', async (req: Request, res: Response) => {
outboundMessages = restoreSessionReasoning(outboundMessages, rememberedReasoning, route.platform);
}
- // GitHub Models 413s a history above its input ceiling instead of
- // truncating it, so a long conversation burns the github hop of every
- // chain it appears in. Trim the outbound copy to what the platform will
- // accept — scoped to this attempt, so the next candidate still sees the
- // client's full history. A no-op returning the same array when the
- // request already fits.
- if (route.platform === 'github') {
- outboundMessages = truncateMessagesForGithub(outboundMessages);
- }
-
if (stream) {
// — Stream turn-integrity (#231 audit) —
// The old loop forwarded upstream chunks verbatim and called any
diff --git a/server/src/services/builtin-model-discovery.ts b/server/src/services/builtin-model-discovery.ts
index c7a293fd..52536425 100644
--- a/server/src/services/builtin-model-discovery.ts
+++ b/server/src/services/builtin-model-discovery.ts
@@ -46,7 +46,7 @@ import { customModelSyncFreePatterns, customModelSyncIntervalMs } from './custom
* zero rows, enabled or disabled, in the signed live catalog (checked against
* 2026.09.27). Remove a platform here in the same change that adds it to the
* catalog; managedPlatforms in the catalog covers binaries already shipped. */
-export const BUILTIN_DISCOVERY_PLATFORMS: readonly Platform[] = ['github', 'longcat', 'siliconflow'];
+export const BUILTIN_DISCOVERY_PLATFORMS: readonly Platform[] = ['longcat', 'siliconflow'];
/** `BUILTIN_MODEL_DISCOVERY`: 'manual' (default) keeps only the dashboard's
* Fetch models action, so nothing is registered without the operator picking
diff --git a/server/src/services/catalog-sync.ts b/server/src/services/catalog-sync.ts
index 6db3e113..8f969d97 100644
--- a/server/src/services/catalog-sync.ts
+++ b/server/src/services/catalog-sync.ts
@@ -11,7 +11,6 @@ import type { Scheduler } from '../lib/scheduler.js';
import {
applyAllModelOverrides,
refreshModelOverrideBaselines,
- routableContextWindow,
applyModelOverrides,
deleteTombstonedCatalogModels,
isCatalogModelTombstoned,
@@ -571,7 +570,7 @@ function applyCatalogInner(db: Db, catalog: Catalog): NonNullable<SyncResult['co
tpm: m.limits.tpm,
tpd: m.limits.tpd,
monthlyTokenBudget: m.monthlyTokenBudget,
- contextWindow: routableContextWindow(m.platform, m.modelId, m.contextWindow),
+ contextWindow: m.contextWindow,
supportsVision: m.supportsVision ? 1 : 0,
supportsTools: m.supportsTools ? 1 : 0,
};
diff --git a/server/src/services/embeddings.ts b/server/src/services/embeddings.ts
index cf7a7cd8..fdccd7af 100644
--- a/server/src/services/embeddings.ts
+++ b/server/src/services/embeddings.ts
@@ -113,7 +113,6 @@ export const EMBEDDING_PLATFORMS = new Set([
'google',
'nvidia',
'openrouter',
- 'github',
'cloudflare',
'huggingface',
'cohere',
@@ -283,8 +282,6 @@ async function callProvider(row: EmbeddingModelRow, credential: ProviderCredenti
return openAiStyleEmbed('https://integrate.api.nvidia.com/v1/embeddings', row.platform, key, row.model_id, inputs, { input_type: 'query' }, dimensions);
case 'openrouter':
return openAiStyleEmbed('https://openrouter.ai/api/v1/embeddings', row.platform, key, row.model_id, inputs, {}, dimensions);
- case 'github':
- return openAiStyleEmbed('https://models.github.ai/inference/embeddings', row.platform, key, row.model_id, inputs, {}, dimensions);
case 'sealion':
return openAiStyleEmbed('https://api.sea-lion.ai/v1/embeddings', row.platform, key, row.model_id, inputs, {}, dimensions);
case 'speka':
diff --git a/server/src/services/model-state.ts b/server/src/services/model-state.ts
index f547a51c..ddc875ef 100644
--- a/server/src/services/model-state.ts
+++ b/server/src/services/model-state.ts
@@ -90,11 +90,6 @@ function modelRowValue(row: Record<string, unknown> | undefined, key: keyof Mode
return value ?? null;
}
-export function routableContextWindow(platform: string, modelId: string, contextWindow: number | null): number | null {
- if (platform === 'github' && modelId === 'openai/gpt-4.1') return 8000;
- return contextWindow;
-}
-
function cachedCatalogDefaults(db: Db, platform: string, modelId: string): StoredOverrides {
const row = db.prepare("SELECT value FROM settings WHERE key = 'catalog_applied_json'").get() as { value: string } | undefined;
if (!row) return {};
@@ -108,7 +103,7 @@ function cachedCatalogDefaults(db: Db, platform: string, modelId: string): Store
rpmLimit: model.limits.rpm, rpdLimit: model.limits.rpd,
tpmLimit: model.limits.tpm, tpdLimit: model.limits.tpd,
monthlyTokenBudget: model.monthlyTokenBudget ?? '',
- contextWindow: routableContextWindow(platform, modelId, model.contextWindow),
+ contextWindow: model.contextWindow,
supportsVision: model.supportsVision, supportsTools: model.supportsTools,
enabled: model.enabled,
};
diff --git a/server/src/services/provider-quota.ts b/server/src/services/provider-quota.ts
index fb1827e0..b94a8b60 100644
--- a/server/src/services/provider-quota.ts
+++ b/server/src/services/provider-quota.ts
@@ -141,7 +141,6 @@ export function inferPoolForPlatform(platform: Platform, modelId?: string | null
if (platform === 'sambanova') return 'sambanova::shared';
if (platform === 'nvidia') return 'nvidia::credit-pool';
if (platform === 'mistral') return 'mistral::experiment-pool';
- if (platform === 'github') return 'github::account';
if (platform === 'cohere') return 'cohere::trial-pool';
if (platform === 'cloudflare') return 'cloudflare::account';
if (platform === 'zhipu') return 'zhipu::account';
@@ -190,7 +189,7 @@ function isSharedPool(platform: Platform): boolean {
if (platform === 'plugsky') return true;
if (platform === 'inferbase' || platform === 'simplellm') return true;
if (['electronhub', 'experiential', 'router9', 'septor', 'clod', 'speechify', 'blaze', 'lucidity', 'airforce', 'dreamprompting', 'waterfall', 'logfare'].includes(platform)) return true;
- return ['openrouter', 'google', 'groq', 'cerebras', 'sail', 'bai', 'radeon', 'sambanova', 'nvidia', 'mistral', 'github', 'cohere', 'cloudflare', 'zhipu', 'ollama', 'kilo', 'pollinations', 'llm7', 'huggingface', 'opencode', 'routeway', 'bazaarlink', 'ainative', 'aion', 'requesty', 'navy', 'nara', 'sealion', 'orcarouter', 'unorouter', 'xkiro', 'anyapi', 'modelscope', 'aihorde'].includes(platform);
+ return ['openrouter', 'google', 'groq', 'cerebras', 'sail', 'bai', 'radeon', 'sambanova', 'nvidia', 'mistral', 'cohere', 'cloudflare', 'zhipu', 'ollama', 'kilo', 'pollinations', 'llm7', 'huggingface', 'opencode', 'routeway', 'bazaarlink', 'ainative', 'aion', 'requesty', 'navy', 'nara', 'sealion', 'orcarouter', 'unorouter', 'xkiro', 'anyapi', 'modelscope', 'aihorde'].includes(platform);
}
type HeaderSpec = { metric: QuotaMetric; limit: string; remaining?: string; reset?: string; strategy?: QuotaResetStrategy };
diff --git a/server/src/services/router.ts b/server/src/services/router.ts
index e956b769..57581f9f 100644
--- a/server/src/services/router.ts
+++ b/server/src/services/router.ts
@@ -1799,13 +1799,6 @@ function getModelChainRow(db: Db, modelDbId: number): ChainRow | undefined {
*/
export const CONTEXT_WINDOW_SAFETY_FACTOR = 1.25;
-// Platforms whose pre-dispatch trim guard already caps the dispatched input
-// below the live context ceiling (lib/content.ts truncateMessagesForGithub):
-// the guard — not the routing estimate — is what guarantees the fit there, so
-// applying the factor too would only make that guard unreachable. The margin
-// checks below treat these platforms as strict comparisons.
-const TRIM_GUARDED_PLATFORMS = new Set(['github']);
-
/** True when `estimatedTokens` fits the RAW advertised window (null window =
* unknown, never filtered — same convention as the auto-router). This is the
* comparison /v1/models publishes and the soft-preference fallback tier.
@@ -1819,17 +1812,15 @@ export function fitsContextWindowStrict(contextWindow: number | null | undefined
/** True when `estimatedTokens` plausibly fits `contextWindow` WITH the safety
* margin. The chars/4 heuristic portion is scaled by the factor; an explicit
* output reserve derived from the client's max_tokens (`routingReserveTokens`)
- * is already an exact count and is added UNSCALED (#956 review). Trim-guarded
- * platforms compare strictly — their guard guarantees the fit. Also enforces
+ * is already an exact count and is added UNSCALED (#956 review). Also enforces
* the operator's min-context-window floor (lib/min-context-window.ts): a model
* below the configured preset is excluded from every chain, while an unknown
* window (null) always passes. */
-export function fitsContextWindow(platform: string, contextWindow: number | null | undefined, estimatedTokens: number, exactOutputReserve = 0): boolean {
+export function fitsContextWindow(_platform: string, contextWindow: number | null | undefined, estimatedTokens: number, exactOutputReserve = 0): boolean {
if (!passesContextWindowFloor(contextWindow)) return false;
if (contextWindow == null) return true;
// Raw advertised comparison first — the margin can only shrink eligibility.
if (estimatedTokens > contextWindow) return false;
- if (TRIM_GUARDED_PLATFORMS.has(platform)) return true;
const reserve = Math.max(0, exactOutputReserve);
const heuristic = Math.max(0, estimatedTokens - reserve);
return heuristic * CONTEXT_WINDOW_SAFETY_FACTOR + reserve <= contextWindow;
@@ -2302,9 +2293,7 @@ export function routeRequest(estimatedTokens = 1000, skipKeys?: Set<string>, pre
// false (#956 review): e.g. `context 131072 < estimated 106000 x1.25 = 132500`.
const reserve = Math.max(0, exactOutputReserve);
const requiredWithMargin = Math.ceil(Math.max(0, estimatedTokens - reserve) * CONTEXT_WINDOW_SAFETY_FACTOR + reserve);
- diag.push(TRIM_GUARDED_PLATFORMS.has(entry.platform)
- ? `${label}: context ${entry.context_window} < estimated ${estimatedTokens}`
- : `${label}: context ${entry.context_window} < estimated ${estimatedTokens} x${CONTEXT_WINDOW_SAFETY_FACTOR} = ${requiredWithMargin}`);
+ diag.push(`${label}: context ${entry.context_window} < estimated ${estimatedTokens} x${CONTEXT_WINDOW_SAFETY_FACTOR} = ${requiredWithMargin}`);
continue;
}
diff --git a/shared/types.ts b/shared/types.ts
index 89ebd228..f48fb04a 100644
--- a/shared/types.ts
+++ b/shared/types.ts
@@ -126,7 +126,6 @@ export type Platform =
| 'mistral'
| 'sambanova'
| 'openrouter'
- | 'github'
| 'cohere'
// Web search APIs (#1174), not chat providers. Their keys serve POST
// /v1/search (and the agents' web_search tools) through the catalog's