smashimo/oh-my-pi
1
0
Fork 0
Code Issues Pull Requests Actions Packages Projects Releases Wiki Activity

Fix custom model precedence across load and refresh

Browse Source
This commit is contained in:
Zakhar Kogan
2026-03-23 15:09:19 +02:00
parent 1bf48ca208
commit fc6076e9bd
2 changed files with 396 additions and 50 deletions
Download Patch File Download Diff File Expand all files Collapse all files
packages/coding-agent/src/config/model-registry.ts
+163 -44
View File
@@ -356,7 +356,7 @@ export interface ProviderDiscoveryState {
/** Result of loading custom models from models.json */
interface CustomModelsResult {
models?: Model<Api>[];
models?: CustomModelOverlay[];
overrides?: Map<string, ProviderOverride>;
modelOverrides?: Map<string, Map<string, ModelOverride>>;
keylessProviders?: Set<string>;
@@ -551,6 +551,24 @@ interface CustomModelBuildOptions {
useDefaults: boolean;
}
type CustomModelOverlay = {
id: string;
provider: string;
api: Api;
baseUrl: string;
name?: string;
reasoning?: boolean;
thinking?: ThinkingConfig;
input?: ("text" | "image")[];
cost?: { input: number; output: number; cacheRead: number; cacheWrite: number };
contextWindow?: number;
maxTokens?: number;
headers?: Record<string, string>;
compat?: Model<Api>["compat"];
contextPromotionTarget?: string;
premiumMultiplier?: number;
};
function mergeCustomModelHeaders(
providerHeaders: Record<string, string> | undefined,
modelHeaders: Record<string, string> | undefined,
@@ -567,6 +585,68 @@ function mergeCustomModelHeaders(
return headers;
}
function buildCustomModelOverlay(
providerName: string,
providerBaseUrl: string,
providerApi: Api | undefined,
providerHeaders: Record<string, string> | undefined,
providerApiKey: string | undefined,
authHeader: boolean | undefined,
providerCompat: Model<Api>["compat"] | undefined,
modelDef: CustomModelDefinitionLike,
): CustomModelOverlay | undefined {
const api = modelDef.api ?? providerApi;
if (!api) return undefined;
return {
id: modelDef.id,
provider: providerName,
api,
baseUrl: modelDef.baseUrl ?? providerBaseUrl,
name: modelDef.name,
reasoning: modelDef.reasoning,
thinking: modelDef.thinking as ThinkingConfig | undefined,
input: modelDef.input as ("text" | "image")[] | undefined,
cost: modelDef.cost,
contextWindow: modelDef.contextWindow,
maxTokens: modelDef.maxTokens,
headers: mergeCustomModelHeaders(providerHeaders, modelDef.headers, authHeader, providerApiKey),
compat: mergeCompat(providerCompat, modelDef.compat),
contextPromotionTarget: modelDef.contextPromotionTarget,
premiumMultiplier: modelDef.premiumMultiplier,
};
}
function applyStandaloneCustomModelPolicies(model: CustomModelOverlay): CustomModelOverlay {
if (model.id !== "gpt-5.4" || model.provider === "github-copilot" || model.contextWindow !== undefined) {
return model;
}
return { ...model, contextWindow: 1_000_000 };
}
function finalizeCustomModel(model: CustomModelOverlay, options: CustomModelBuildOptions): Model<Api> {
const resolvedModel = options.useDefaults ? applyStandaloneCustomModelPolicies(model) : model;
const cost =
resolvedModel.cost ?? (options.useDefaults ? { input: 0, output: 0, cacheRead: 0, cacheWrite: 0 } : undefined);
const input = resolvedModel.input ?? (options.useDefaults ? ["text"] : undefined);
return enrichModelThinking({
id: resolvedModel.id,
name: resolvedModel.name ?? (options.useDefaults ? resolvedModel.id : undefined),
api: resolvedModel.api,
provider: resolvedModel.provider,
baseUrl: resolvedModel.baseUrl,
reasoning: resolvedModel.reasoning ?? (options.useDefaults ? false : undefined),
thinking: resolvedModel.thinking,
input: input as ("text" | "image")[],
cost,
contextWindow: resolvedModel.contextWindow ?? (options.useDefaults ? 128000 : undefined),
maxTokens: resolvedModel.maxTokens ?? (options.useDefaults ? 16384 : undefined),
headers: resolvedModel.headers,
compat: resolvedModel.compat,
contextPromotionTarget: resolvedModel.contextPromotionTarget,
premiumMultiplier: resolvedModel.premiumMultiplier,
} as Model<Api>);
}
function buildCustomModel(
providerName: string,
providerBaseUrl: string,
@@ -578,28 +658,18 @@ function buildCustomModel(
modelDef: CustomModelDefinitionLike,
options: CustomModelBuildOptions,
): Model<Api> | undefined {
const api = modelDef.api ?? providerApi;
if (!api) return undefined;
const withDefaults = options.useDefaults;
const cost = modelDef.cost ?? (withDefaults ? { input: 0, output: 0, cacheRead: 0, cacheWrite: 0 } : undefined);
const input = modelDef.input ?? (withDefaults ? ["text"] : undefined);
return enrichModelThinking({
id: modelDef.id,
name: modelDef.name ?? (withDefaults ? modelDef.id : undefined),
api,
provider: providerName,
baseUrl: modelDef.baseUrl ?? providerBaseUrl,
reasoning: modelDef.reasoning ?? (withDefaults ? false : undefined),
thinking: modelDef.thinking as ThinkingConfig | undefined,
input: input as ("text" | "image")[],
cost,
contextWindow: modelDef.contextWindow ?? (withDefaults ? 128000 : undefined),
maxTokens: modelDef.maxTokens ?? (withDefaults ? 16384 : undefined),
headers: mergeCustomModelHeaders(providerHeaders, modelDef.headers, authHeader, providerApiKey),
compat: mergeCompat(providerCompat, modelDef.compat),
contextPromotionTarget: modelDef.contextPromotionTarget,
premiumMultiplier: modelDef.premiumMultiplier,
} as Model<Api>);
const model = buildCustomModelOverlay(
providerName,
providerBaseUrl,
providerApi,
providerHeaders,
providerApiKey,
authHeader,
providerCompat,
modelDef,
);
if (!model) return undefined;
return finalizeCustomModel(model, options);
}
/**
@@ -610,6 +680,8 @@ export class ModelRegistry {
#customProviderApiKeys: Map<string, string> = new Map();
#keylessProviders: Set<string> = new Set();
#discoverableProviders: DiscoveryProviderConfig[] = [];
#customModelOverlays: CustomModelOverlay[] = [];
#providerOverrides: Map<string, ProviderOverride> = new Map();
#modelOverrides: Map<string, Map<string, ModelOverride>> = new Map();
#configError: ConfigError | undefined = undefined;
#modelsConfigFile: ConfigFile<ModelsConfig>;
@@ -676,6 +748,7 @@ export class ModelRegistry {
this.#customProviderApiKeys.clear();
this.#keylessProviders.clear();
this.#discoverableProviders = [];
this.#providerOverrides.clear();
this.#modelOverrides.clear();
this.#configError = undefined;
this.#providerDiscoveryStates.clear();
@@ -703,14 +776,17 @@ export class ModelRegistry {
this.#configError = configError;
this.#keylessProviders = keylessProviders;
this.#discoverableProviders = discoverableProviders;
this.#customModelOverlays = customModels;
this.#providerOverrides = overrides;
this.#modelOverrides = modelOverrides;
this.#addImplicitDiscoverableProviders(configuredProviders);
const builtInModels = this.#loadBuiltInModels(overrides, modelOverrides);
const cachedDiscoveries = this.#loadCachedDiscoverableModels();
const combined = this.#mergeCustomModels(builtInModels, [...customModels, ...cachedDiscoveries]);
const builtInModels = this.#applyHardcodedModelPolicies(this.#loadBuiltInModels(overrides, modelOverrides));
const cachedDiscoveries = this.#applyHardcodedModelPolicies(this.#loadCachedDiscoverableModels());
const resolvedDefaults = this.#mergeResolvedModels(builtInModels, cachedDiscoveries);
const combined = this.#mergeCustomModels(resolvedDefaults, this.#customModelOverlays);
this.#models = this.#applyHardcodedModelPolicies(combined);
this.#models = this.#applyModelOverrides(combined, this.#modelOverrides);
}
/** Load built-in models, applying provider and per-model overrides */
@@ -742,15 +818,50 @@ export class ModelRegistry {
});
}
#mergeResolvedModels(baseModels: Model<Api>[], replacementModels: Model<Api>[]): Model<Api>[] {
const merged = [...baseModels];
for (const replacementModel of replacementModels) {
const existingIndex = merged.findIndex(
m => m.provider === replacementModel.provider && m.id === replacementModel.id,
);
if (existingIndex >= 0) {
merged[existingIndex] = replacementModel;
} else {
merged.push(replacementModel);
}
}
return merged;
}
/** Merge custom models with built-in, replacing by provider+id match */
#mergeCustomModels(builtInModels: Model<Api>[], customModels: Model<Api>[]): Model<Api>[] {
#mergeCustomModels(builtInModels: Model<Api>[], customModels: CustomModelOverlay[]): Model<Api>[] {
const merged = [...builtInModels];
for (const customModel of customModels) {
const existingIndex = merged.findIndex(m => m.provider === customModel.provider && m.id === customModel.id);
if (existingIndex >= 0) {
merged[existingIndex] = customModel;
merged[existingIndex] = enrichModelThinking({
...merged[existingIndex],
id: customModel.id,
provider: customModel.provider,
api: customModel.api,
baseUrl: customModel.baseUrl,
name: customModel.name ?? merged[existingIndex].name,
reasoning: customModel.reasoning ?? merged[existingIndex].reasoning,
thinking: customModel.thinking ?? merged[existingIndex].thinking,
input: customModel.input ?? merged[existingIndex].input,
cost: customModel.cost ?? merged[existingIndex].cost,
contextWindow: customModel.contextWindow ?? merged[existingIndex].contextWindow,
maxTokens: customModel.maxTokens ?? merged[existingIndex].maxTokens,
headers: customModel.headers
? { ...merged[existingIndex].headers, ...customModel.headers }
: merged[existingIndex].headers,
compat: mergeCompat(merged[existingIndex].compat, customModel.compat),
contextPromotionTarget:
customModel.contextPromotionTarget ?? merged[existingIndex].contextPromotionTarget,
premiumMultiplier: customModel.premiumMultiplier ?? merged[existingIndex].premiumMultiplier,
} as Model<Api>);
} else {
merged.push(customModel);
merged.push(finalizeCustomModel(customModel, { useDefaults: true }));
}
}
return merged;
@@ -935,22 +1046,31 @@ export class ModelRegistry {
if (discovered.length === 0) {
return;
}
const merged = this.#mergeCustomModels(
this.#models,
const discoveredModels = this.#applyHardcodedModelPolicies(
discovered.map(model => {
const existing =
this.find(model.provider, model.id) ??
this.#models.find(candidate => candidate.provider === model.provider);
return existing
const existing = this.find(model.provider, model.id);
if (existing) {
return {
...model,
baseUrl: existing.baseUrl,
headers: existing.headers ? { ...existing.headers, ...model.headers } : model.headers,
};
}
const providerOverride = this.#providerOverrides.get(model.provider);
return providerOverride
? {
...model,
baseUrl: existing.baseUrl,
headers: existing.headers ? { ...existing.headers, ...model.headers } : model.headers,
baseUrl: providerOverride.baseUrl ?? model.baseUrl,
headers: providerOverride.headers
? { ...model.headers, ...providerOverride.headers }
: model.headers,
}
: model;
}),
);
this.#models = this.#applyHardcodedModelPolicies(this.#applyModelOverrides(merged, this.#modelOverrides));
const resolved = this.#mergeResolvedModels(this.#models, discoveredModels);
const combined = this.#mergeCustomModels(resolved, this.#customModelOverlays);
this.#models = this.#applyModelOverrides(combined, this.#modelOverrides);
}
async #discoverProviderModels(
@@ -1454,8 +1574,8 @@ export class ModelRegistry {
});
}
#parseModels(config: ModelsConfig): Model<Api>[] {
const models: Model<Api>[] = [];
#parseModels(config: ModelsConfig): CustomModelOverlay[] {
const models: CustomModelOverlay[] = [];
for (const [providerName, providerConfig] of Object.entries(config.providers)) {
const modelDefs = providerConfig.models ?? [];
@@ -1464,7 +1584,7 @@ export class ModelRegistry {
this.#customProviderApiKeys.set(providerName, providerConfig.apiKey);
}
for (const modelDef of modelDefs) {
const model = buildCustomModel(
const model = buildCustomModelOverlay(
providerName,
providerConfig.baseUrl!,
providerConfig.api as Api | undefined,
@@ -1473,7 +1593,6 @@ export class ModelRegistry {
providerConfig.authHeader,
providerConfig.compat,
modelDef as CustomModelDefinitionLike,
{ useDefaults: true },
);
if (!model) continue;
models.push(model);
@@ -1635,7 +1754,7 @@ export class ModelRegistry {
config.authHeader,
config.compat,
modelDef as CustomModelDefinitionLike,
{ useDefaults: false },
{ useDefaults: true },
);
if (!model) {
throw new Error(`Provider ${providerName}, model ${modelDef.id}: no "api" specified.`);
packages/coding-agent/test/model-registry.test.ts
+233 -6
View File
@@ -50,7 +50,13 @@ describe("ModelRegistry", () => {
/** Create minimal provider config */
function providerConfig(
baseUrl: string,
models: Array<{ id: string; name?: string; reasoning?: boolean; thinking?: ThinkingConfig }>,
models: Array<{
id: string;
name?: string;
reasoning?: boolean;
thinking?: ThinkingConfig;
contextWindow?: number;
}>,
api: string = "anthropic-messages",
) {
return {
@@ -64,7 +70,7 @@ describe("ModelRegistry", () => {
thinking: m.thinking,
input: ["text"],
cost: { input: 0, output: 0, cacheRead: 0, cacheWrite: 0 },
contextWindow: 100000,
contextWindow: m.contextWindow ?? 100000,
maxTokens: 8000,
})),
};
@@ -88,6 +94,19 @@ describe("ModelRegistry", () => {
fs.writeFileSync(modelsJsonPath, JSON.stringify({ providers }));
}
function mockOpenAiCompatibleModels(url: string, modelIds: string[]) {
return hookFetch(input => {
const requestUrl = String(input);
if (requestUrl === url) {
return new Response(JSON.stringify({ data: modelIds.map(id => ({ id })) }), {
status: 200,
headers: { "Content-Type": "application/json" },
});
}
throw new Error(`Unexpected URL: ${requestUrl}`);
});
}
describe("baseUrl override (no custom models)", () => {
test("overriding baseUrl keeps all built-in models", () => {
writeRawModelsJson({
@@ -422,13 +441,221 @@ describe("ModelRegistry", () => {
expect(registry.find("openai", "gpt-5.4")?.contextWindow).toBe(1_000_000);
});
test("custom gpt-5.4 replacement also applies the hardcoded context window policy", () => {
writeModelsJson({
openai: providerConfig("https://my-proxy.example.com/v1", [{ id: "gpt-5.4" }], "openai-responses"),
test("custom gpt-5.4 replacement keeps the hardcoded context window when contextWindow is omitted", () => {
writeRawModelsJson({
openai: {
baseUrl: "https://my-proxy.example.com/v1",
apiKey: "TEST_KEY",
api: "openai-responses",
models: [{ id: "gpt-5.4" }],
},
});
const registry = new ModelRegistry(authStorage, modelsJsonPath);
expect(registry.find("openai", "gpt-5.4")?.contextWindow).toBe(1_000_000);
const model = registry.find("openai", "gpt-5.4");
expect(model?.contextWindow).toBe(1_000_000);
expect(model?.baseUrl).toBe("https://my-proxy.example.com/v1");
});
test("custom-only gpt-5.4 provider keeps the hardcoded context window when contextWindow is omitted", () => {
writeRawModelsJson({
"my-proxy": {
baseUrl: "https://my-proxy.example.com/v1",
apiKey: "TEST_KEY",
api: "openai-responses",
models: [{ id: "gpt-5.4" }],
},
});
const registry = new ModelRegistry(authStorage, modelsJsonPath);
const model = registry.find("my-proxy", "gpt-5.4");
expect(model?.contextWindow).toBe(1_000_000);
expect(model?.baseUrl).toBe("https://my-proxy.example.com/v1");
});
test("custom gpt-5.4 replacement preserves its explicit context window", () => {
writeModelsJson({
openai: providerConfig(
"https://my-proxy.example.com/v1",
[{ id: "gpt-5.4", contextWindow: 256000 }],
"openai-responses",
),
});
const registry = new ModelRegistry(authStorage, modelsJsonPath);
expect(registry.find("openai", "gpt-5.4")?.contextWindow).toBe(256000);
});
test("modelOverrides can still patch a custom gpt-5.4 replacement", () => {
writeRawModelsJson({
openai: {
baseUrl: "https://my-proxy.example.com/v1",
apiKey: "TEST_KEY",
api: "openai-responses",
models: [
{
id: "gpt-5.4",
name: "gpt-5.4",
reasoning: false,
input: ["text"],
cost: { input: 0, output: 0, cacheRead: 0, cacheWrite: 0 },
contextWindow: 256000,
maxTokens: 128000,
},
],
modelOverrides: {
"gpt-5.4": {
contextWindow: 512000,
},
},
},
});
const registry = new ModelRegistry(authStorage, modelsJsonPath);
expect(registry.find("openai", "gpt-5.4")?.contextWindow).toBe(512000);
});
test("discoverable bundled replacement survives refresh", async () => {
writeModelsJson({
openai: providerConfig(
"https://my-proxy.example.com/v1",
[{ id: "gpt-5.4", name: "Proxy GPT-5.4", contextWindow: 256000 }],
"openai-responses",
),
});
const registry = new ModelRegistry(authStorage, modelsJsonPath);
expect(registry.find("openai", "gpt-5.4")?.name).toBe("Proxy GPT-5.4");
expect(registry.find("openai", "gpt-5.4")?.contextWindow).toBe(256000);
using _hook = mockOpenAiCompatibleModels("https://my-proxy.example.com/v1/models", ["gpt-5.4"]);
await registry.refreshProvider("openai", "online");
const model = registry.find("openai", "gpt-5.4");
expect(model?.name).toBe("Proxy GPT-5.4");
expect(model?.contextWindow).toBe(256000);
expect(model?.baseUrl).toBe("https://my-proxy.example.com/v1");
});
test("discoverable custom-only gpt-5.4 survives refresh", async () => {
writeRawModelsJson({
"custom-local": {
baseUrl: "http://127.0.0.1:8080",
apiKey: "TEST_KEY",
api: "openai-responses",
discovery: { type: "llama.cpp" },
models: [{ id: "gpt-5.4" }],
},
});
const registry = new ModelRegistry(authStorage, modelsJsonPath);
expect(registry.find("custom-local", "gpt-5.4")?.contextWindow).toBe(1_000_000);
using _hook = mockOpenAiCompatibleModels("http://127.0.0.1:8080/models", ["gpt-5.4"]);
await registry.refreshProvider("custom-local", "online");
const model = registry.find("custom-local", "gpt-5.4");
expect(model?.contextWindow).toBe(1_000_000);
expect(model?.baseUrl).toBe("http://127.0.0.1:8080");
});
test("discoverable custom compat survives refresh", async () => {
writeRawModelsJson({
openai: {
baseUrl: "https://my-proxy.example.com/v1",
apiKey: "TEST_KEY",
api: "openai-responses",
models: [
{
id: "gpt-5.4",
compat: {
extraBody: { source: "proxy" },
},
},
],
},
});
const registry = new ModelRegistry(authStorage, modelsJsonPath);
expect(registry.find("openai", "gpt-5.4")?.compat?.extraBody).toEqual({ source: "proxy" });
using _hook = mockOpenAiCompatibleModels("https://my-proxy.example.com/v1/models", ["gpt-5.4"]);
await registry.refreshProvider("openai", "online");
expect(registry.find("openai", "gpt-5.4")?.compat?.extraBody).toEqual({ source: "proxy" });
});
test("modelOverrides still apply after discoverable refresh", async () => {
writeRawModelsJson({
openai: {
baseUrl: "https://my-proxy.example.com/v1",
apiKey: "TEST_KEY",
api: "openai-responses",
models: [
{
id: "gpt-5.4",
contextWindow: 256000,
},
],
modelOverrides: {
"gpt-5.4": {
contextWindow: 512000,
},
},
},
});
const registry = new ModelRegistry(authStorage, modelsJsonPath);
expect(registry.find("openai", "gpt-5.4")?.contextWindow).toBe(512000);
using _hook = mockOpenAiCompatibleModels("https://my-proxy.example.com/v1/models", ["gpt-5.4"]);
await registry.refreshProvider("openai", "online");
expect(registry.find("openai", "gpt-5.4")?.contextWindow).toBe(512000);
});
test("newly discovered ids inherit provider fields, not another model's custom fields", async () => {
writeRawModelsJson({
openai: {
baseUrl: "https://provider.example.com/v1",
headers: { "X-Provider": "provider" },
apiKey: "TEST_KEY",
api: "openai-responses",
models: [
{
id: "gpt-5.4",
baseUrl: "https://special.example.com/v1",
headers: { "X-Model": "special" },
},
],
},
});
const registry = new ModelRegistry(authStorage, modelsJsonPath);
expect(registry.find("openai", "gpt-5.4")?.baseUrl).toBe("https://special.example.com/v1");
using _hook = mockOpenAiCompatibleModels("https://provider.example.com/v1/models", ["gpt-5.4", "gpt-5.5"]);
await registry.refreshProvider("openai", "online");
const discovered = registry.find("openai", "gpt-5.5");
expect(discovered?.baseUrl).toBe("https://provider.example.com/v1");
expect(discovered?.headers?.["X-Provider"]).toBe("provider");
expect(discovered?.headers?.["X-Model"]).toBeUndefined();
});
test("provider compat overlays preserve bundled model compat", () => {
writeRawModelsJson({
"minimax-code": {
baseUrl: "https://proxy.example.com/v1",
apiKey: "TEST_KEY",
api: "openai-completions",
compat: {
extraBody: { source: "proxy" },
},
models: [{ id: "MiniMax-M2.5" }],
},
});
const registry = new ModelRegistry(authStorage, modelsJsonPath);
const model = registry.find("minimax-code", "MiniMax-M2.5");
expect(model?.compat?.thinkingFormat).toBe("zai");
expect(model?.compat?.reasoningContentField).toBe("reasoning_content");
expect(model?.compat?.extraBody).toEqual({ source: "proxy" });
});
test("removing custom models from models.json keeps built-in provider models", async () => {