feat: per-tier endpoint routing

.env.example

@@ -54,13 +54,23 @@ SMALLCODE_CLASSIC_TUI=false
 # ─── OpenRouter Example ──────────────────────────────────────────────────────
 # SMALLCODE_BASE_URL=https://openrouter.ai/api/v1
 # SMALLCODE_MODEL=openai/gpt-4o-mini
-# OPENAI_API_KEY=sk-or-v1-...
+# OPENROUTER_API_KEY=sk-or-v1-...
 
 # ─── Multi-Model Routing (optional) ──────────────────────────────────────────
-# Auto-pick model based on task complexity
+# Auto-pick model based on task complexity. Each tier may use its own endpoint.
 # SMALLCODE_MODEL_FAST=gemma-4-e4b
 # SMALLCODE_MODEL_DEFAULT=qwen3-8b
 # SMALLCODE_MODEL_STRONG=qwen3-14b
+# SMALLCODE_BASE_URL_FAST=http://localhost:1234/v1
+# SMALLCODE_BASE_URL_DEFAULT=http://localhost:1234/v1
+# SMALLCODE_BASE_URL_STRONG=https://openrouter.ai/api/v1
+#
+# Local default + OpenRouter strong model:
+# SMALLCODE_MODEL=qwen3:8b
+# SMALLCODE_BASE_URL=http://localhost:11434/v1
+# SMALLCODE_MODEL_STRONG=openai/gpt-4o-mini
+# SMALLCODE_BASE_URL_STRONG=https://openrouter.ai/api/v1
+# OPENROUTER_API_KEY=sk-or-v1-...
 
 # ─── RTK (Rust Token Killer) ─────────────────────────────────────────────────
 # RTK auto-rewrites bash commands (git, tests, ls, etc.) for 60-90% token savings.

CHANGELOG.md

@@ -1,5 +1,27 @@
 # Changelog
 
+## Unreleased
+
+### feat: per-tier endpoint routing
+
+`SMALLCODE_MODEL_FAST`, `SMALLCODE_MODEL_DEFAULT`, `SMALLCODE_MODEL_MEDIUM`,
+and `SMALLCODE_MODEL_STRONG` can now be paired with matching
+`SMALLCODE_BASE_URL_*` variables. `smallcode.toml` also supports
+`[models.fast]`, `[models.default]`, `[models.medium]`, and
+`[models.strong]` sections with `name` and `baseUrl`, so users can keep
+default work on localhost while routing larger tiers to OpenRouter.
+
+Complexity-based tier selection and adaptive failure-rate routing now resolve
+the matching endpoint per request (`activeModelTarget`). Auth headers follow
+the selected URL — OpenRouter keys and headers for cloud tiers, no auth for
+local endpoints. Primary `[model]` config remains env-first; TOML tier sections
+always merge; env tier vars override TOML.
+
+### Verification
+
+- 90/90 unit tests pass (`npm test`) — 7 new in `test/config_normalize.test.js`,
+  `test/model_routing.test.js`, and `test/provider_compat.test.js`
+
 ## [1.2.1] - 2026-05-24
 
 ### fix: provider compatibility — 7 bugs causing 400 errors on cloud/local LLMs

README.md

@@ -76,11 +76,40 @@ SMALLCODE_BASE_URL=http://localhost:1234/v1
 # Optional: escalation (auto-fallback to cloud on hard fail)
 # ANTHROPIC_API_KEY=sk-ant-...
 # OPENAI_API_KEY=sk-...
+# OPENROUTER_API_KEY=sk-or-v1-...
 # DEEPSEEK_API_KEY=sk-...
 ```
 
 See `.env.example` for all options. Also supports `smallcode.toml` for backwards compatibility.
 
+SmallCode can route each model tier to a different endpoint. This lets you keep
+fast/default work local while sending complex tasks to a larger OpenRouter model:
+
+```bash
+SMALLCODE_MODEL=qwen3:8b
+SMALLCODE_BASE_URL=http://localhost:11434/v1
+
+SMALLCODE_MODEL_STRONG=openai/gpt-4o-mini
+SMALLCODE_BASE_URL_STRONG=https://openrouter.ai/api/v1
+OPENROUTER_API_KEY=sk-or-v1-...
+```
+
+Equivalent `smallcode.toml`:
+
+```toml
+[model]
+provider = "openai"
+name = "qwen3:8b"
+baseUrl = "http://localhost:11434/v1"
+
+[models.strong]
+name = "openai/gpt-4o-mini"
+baseUrl = "https://openrouter.ai/api/v1"
+```
+
+Tier URLs are optional. If `SMALLCODE_BASE_URL_STRONG` or
+`[models.strong].baseUrl` is omitted, that tier uses the primary base URL.
+
 ## Architecture
 
 SmallCode is built with a modular architecture:
@@ -227,12 +256,13 @@ When 3 or more files are edited in a single agent turn, `coordinateMultiFileEdit
 When `patch` fails because `old_str` no longer exists in the file, `semanticMerge()` asks the model to merge the intended change into the current file content. Returns the complete corrected file. Replaces the hard error with a recovery attempt. TTL 1 min (content-specific).
 
 ### Adaptive Model Select (Rank 8)
-`AdaptiveModelRouter` in `src/model/adaptive_router.js` tracks per-model call/fail counts. When the primary model's failure rate exceeds 0.3 (medium) or 0.6 (strong), `chatCompletion` automatically overrides `body.model` with `SMALLCODE_MODEL_MEDIUM` or `SMALLCODE_MODEL_STRONG`. Requires at least 3 calls before routing decisions activate. Reset via `router.reset()`.
+`AdaptiveModelRouter` in `src/model/adaptive_router.js` tracks per-model call/fail counts. When the primary model's failure rate exceeds 0.3 (medium) or 0.6 (strong), `chatCompletion` automatically routes to `SMALLCODE_MODEL_MEDIUM` or `SMALLCODE_MODEL_STRONG` and their matching `SMALLCODE_BASE_URL_*` endpoint when configured. Requires at least 3 calls before routing decisions activate. Reset via `router.reset()`.
 
 ```bash
 # Optional: configure fallback models for adaptive routing
 SMALLCODE_MODEL_MEDIUM=qwen2.5-coder:32b
 SMALLCODE_MODEL_STRONG=gpt-4o
+SMALLCODE_BASE_URL_STRONG=https://openrouter.ai/api/v1
 ```
 
 ### Benchmark Harness

README_zh-CN.md

@@ -50,11 +50,40 @@ SMALLCODE_BASE_URL=http://localhost:1234/v1
 # 可选：模型升级（硬失败时自动回退到云端）
 # ANTHROPIC_API_KEY=sk-ant-...
 # OPENAI_API_KEY=sk-...
+# OPENROUTER_API_KEY=sk-or-v1-...
 # DEEPSEEK_API_KEY=sk-...
 ```
 
 查看 `.env.example` 了解所有选项。同时支持 `smallcode.toml` 以保持向后兼容。
 
+SmallCode 可以为每个模型层级配置不同 endpoint。这样可以让 fast/default
+任务继续使用本地模型，而复杂任务使用 OpenRouter 上的更大模型：
+
+```bash
+SMALLCODE_MODEL=qwen3:8b
+SMALLCODE_BASE_URL=http://localhost:11434/v1
+
+SMALLCODE_MODEL_STRONG=openai/gpt-4o-mini
+SMALLCODE_BASE_URL_STRONG=https://openrouter.ai/api/v1
+OPENROUTER_API_KEY=sk-or-v1-...
+```
+
+等价的 `smallcode.toml`：
+
+```toml
+[model]
+provider = "openai"
+name = "qwen3:8b"
+baseUrl = "http://localhost:11434/v1"
+
+[models.strong]
+name = "openai/gpt-4o-mini"
+baseUrl = "https://openrouter.ai/api/v1"
+```
+
+层级 URL 是可选的。如果省略 `SMALLCODE_BASE_URL_STRONG` 或
+`[models.strong].baseUrl`，该层级会使用主 `baseUrl`。
+
 ## 架构
 
 SmallCode 采用模块化架构：

bin/commands.js

@@ -60,6 +60,7 @@ module.exports = function createCommandHandler(config, conversationHistory, impr
         } else {
           const newModel = parts.slice(1).join(' ');
           config.model.name = newModel;
+          delete config.activeModelTarget;
           console.log(`  ${chalk.green('✓')} Switched to ${chalk.cyan(newModel)}`);
         }
         console.log('');
@@ -73,6 +74,7 @@ module.exports = function createCommandHandler(config, conversationHistory, impr
           console.log(chalk.gray('  Switch: /endpoint http://host:port/v1'));
         } else {
           config.model.baseUrl = parts[1];
+          delete config.activeModelTarget;
           console.log(`  ${chalk.green('✓')} Endpoint: ${chalk.gray(parts[1])}`);
         }
         console.log('');