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

fix(compaction): preserve provider defaults for remote compaction

Browse Source
This commit is contained in:
can1357
2026-06-27 01:23:17 +02:00
parent 14cc9cba0f
commit 0b8206a46d
6 changed files with 141 additions and 10 deletions
Download Patch File Download Diff File Expand all files Collapse all files
packages/agent/src/compaction/openai.ts
+10 -2
View File
@@ -13,7 +13,7 @@
*/
import { ProviderHttpError } from "@oh-my-pi/pi-ai/errors";
import { parseTextSignature } from "@oh-my-pi/pi-ai/providers/openai-shared";
import { parseAzureDeploymentNameMap, parseTextSignature } from "@oh-my-pi/pi-ai/providers/openai-shared";
import { transformMessages } from "@oh-my-pi/pi-ai/providers/transform-messages";
import type { Api, AssistantMessage, FetchImpl, Message, Model } from "@oh-my-pi/pi-ai/types";
import {
@@ -144,6 +144,14 @@ function appendAzureApiVersion(endpoint: string): string {
return `${endpoint}${separator}api-version=${encodeURIComponent($env.AZURE_OPENAI_API_VERSION || DEFAULT_AZURE_API_VERSION)}`;
}
function resolveOpenAiCompactModel(model: Model): string {
const requestModel = model.remoteCompaction?.model ?? model.requestModelId ?? model.id;
const compactionApi = model.remoteCompaction?.api ?? model.api;
if (compactionApi !== "azure-openai-responses") return requestModel;
const mappedDeployment = parseAzureDeploymentNameMap($env.AZURE_OPENAI_DEPLOYMENT_NAME_MAP).get(requestModel);
return mappedDeployment ?? requestModel;
}
function resolveOpenAiCodexCompactEndpoint(baseUrl: string | undefined): string {
const rawBase = baseUrl && baseUrl.length > 0 ? baseUrl : CODEX_BASE_URL;
const normalizedBase = rawBase.endsWith("/") ? rawBase.slice(0, -1) : rawBase;
@@ -495,7 +503,7 @@ export async function requestOpenAiRemoteCompaction(
opts?: { fetch?: FetchImpl; timeoutMs?: number },
): Promise<OpenAiRemoteCompactionResponse> {
const endpoint = resolveOpenAiCompactEndpoint(model);
const requestModel = model.remoteCompaction?.model ?? model.requestModelId ?? model.id;
const requestModel = resolveOpenAiCompactModel(model);
const request: OpenAiRemoteCompactionRequest = {
model: requestModel,
input: trimOpenAiCompactInput(compactInput, model.contextWindow ?? Number.POSITIVE_INFINITY, instructions),
packages/agent/test/remote-compaction.test.ts
+8 -1
View File
@@ -286,6 +286,8 @@ test("uses configured OpenAI-compatible compaction for custom providers", async
});
test("uses Azure request shape for Azure Responses remote compaction", async () => {
const previousDeploymentMap = Bun.env.AZURE_OPENAI_DEPLOYMENT_NAME_MAP;
Bun.env.AZURE_OPENAI_DEPLOYMENT_NAME_MAP = "gpt-5-compact=azure-gpt-5-compact";
const model = makeAzureModel({
headers: { "x-custom-header": "custom" },
remoteCompaction: {
@@ -332,7 +334,12 @@ test("uses Azure request shape for Azure Responses remote compaction", async ()
expect(requestAuthorization).toBeUndefined();
expect(requestContentType).toBe("application/json");
expect(requestCustomHeader).toBe("custom");
expect(requestBody).toMatchObject({ model: "gpt-5-compact" });
expect(requestBody).toMatchObject({ model: "azure-gpt-5-compact" });
if (previousDeploymentMap === undefined) {
delete Bun.env.AZURE_OPENAI_DEPLOYMENT_NAME_MAP;
} else {
Bun.env.AZURE_OPENAI_DEPLOYMENT_NAME_MAP = previousDeploymentMap;
}
});
describe("requestOpenAiRemoteCompaction abort", () => {
packages/coding-agent/src/config/model-discovery.ts
+2 -1
View File
@@ -6,7 +6,7 @@
* discovery lives in pi-catalog's provider-models.
*/
import { type ApiKey, type FetchImpl, withAuth } from "@oh-my-pi/pi-ai";
import type { Api, Model } from "@oh-my-pi/pi-ai/types";
import type { Api, Model, RemoteCompactionConfig } from "@oh-my-pi/pi-ai/types";
import { buildModel } from "@oh-my-pi/pi-catalog/build";
import {
getBundledModelReferenceIndex,
@@ -95,6 +95,7 @@ export interface DiscoveryProviderConfig {
baseUrl?: string;
headers?: Record<string, string>;
compat?: ModelSpec<Api>["compat"];
remoteCompaction?: RemoteCompactionConfig<Api>;
discovery: ProviderDiscovery;
optional?: boolean;
}
packages/coding-agent/src/config/model-registry.ts
+27 -6
View File
@@ -146,8 +146,8 @@ export function mergeDiscoveredModel<TApi extends Api>(
baseUrl: providerOverride?.baseUrl ?? model.baseUrl ?? existing.baseUrl,
headers: existing.headers ? { ...existing.headers, ...model.headers } : model.headers,
transport: providerOverride?.transport ?? existing.transport ?? model.transport,
remoteCompaction: mergeRemoteCompactionConfig(
existing.remoteCompaction ?? model.remoteCompaction,
remoteCompaction: mergeProviderRemoteCompactionConfig(
mergeRemoteCompactionConfig(existing.remoteCompaction, model.remoteCompaction),
providerOverride?.remoteCompaction,
),
...(supportsTools !== undefined ? { supportsTools } : {}),
@@ -160,7 +160,7 @@ export function mergeDiscoveredModel<TApi extends Api>(
baseUrl: providerOverride.baseUrl ?? model.baseUrl,
headers: providerOverride.headers ? { ...model.headers, ...providerOverride.headers } : model.headers,
...(providerOverride.transport !== undefined ? { transport: providerOverride.transport } : {}),
remoteCompaction: mergeRemoteCompactionConfig(model.remoteCompaction, providerOverride.remoteCompaction),
remoteCompaction: mergeProviderRemoteCompactionConfig(model.remoteCompaction, providerOverride.remoteCompaction),
compat: model.compatConfig,
} as ModelSpec<TApi>);
}
@@ -377,6 +377,13 @@ function mergeRemoteCompactionConfig(
return { ...baseConfig, ...overrideConfig };
}
function mergeProviderRemoteCompactionConfig(
modelConfig: RemoteCompactionConfig<Api> | undefined,
providerConfig: RemoteCompactionConfig<Api> | undefined,
): RemoteCompactionConfig<Api> | undefined {
return mergeRemoteCompactionConfig(providerConfig, modelConfig);
}
/**
* Project a built model back to spec shape for the model-manager/cache
* boundary: sparse compat comes from `compatConfig`, never from the resolved
@@ -1111,12 +1118,25 @@ export class ModelRegistry {
)
: models;
const withRemoteCompaction = providerConfig.remoteCompaction
? withDecoderMetadata.map(model =>
buildModel({
...model,
remoteCompaction: mergeProviderRemoteCompactionConfig(
model.remoteCompaction,
providerConfig.remoteCompaction,
),
compat: model.compatConfig,
} as ModelSpec<Api>),
)
: withDecoderMetadata;
if (providerConfig.provider !== "ollama" || providerConfig.api !== "openai-responses") {
return withDecoderMetadata;
return withRemoteCompaction;
}
const contextLengthOverride = getOllamaContextLengthOverride();
return withDecoderMetadata.map(model => {
return withRemoteCompaction.map(model => {
const normalized =
model.api === "openai-completions"
? buildModel({
@@ -1249,6 +1269,7 @@ export class ModelRegistry {
baseUrl: providerConfig.baseUrl,
headers: resolvedProviderHeaders,
compat: mergeCompat(providerConfig.compat, disableStrictCompat),
remoteCompaction: providerConfig.remoteCompaction,
discovery: providerConfig.discovery,
optional: false,
});
@@ -1664,7 +1685,7 @@ export class ModelRegistry {
// Preserve the model's existing transport when the override omits one;
// providers without a `transport` field keep the default per-API dispatch.
...(override.transport !== undefined ? { transport: override.transport } : {}),
remoteCompaction: mergeRemoteCompactionConfig(entry.remoteCompaction, override.remoteCompaction),
remoteCompaction: mergeProviderRemoteCompactionConfig(entry.remoteCompaction, override.remoteCompaction),
};
}
#applyRuntimeProviderOverrides(models: Model<Api>[]): Model<Api>[] {
packages/coding-agent/test/model-registry-runtime-provider.test.ts
+41
View File
@@ -181,6 +181,47 @@ describe("ModelRegistry runtime provider registration", () => {
}
});
test("refreshRuntimeProviders preserves model-level remoteCompaction over provider defaults", async () => {
const providerName = "dynamic-compact-provider";
const providerEndpoint = "https://runtime.example.com/v1/responses/provider-compact";
const modelEndpoint = "https://runtime.example.com/v1/responses/model-compact";
registry.registerProvider(
providerName,
{
baseUrl: "https://runtime.example.com/v1",
apiKey: "RUNTIME_KEY",
api: "openai-responses",
remoteCompaction: {
enabled: true,
api: "openai-responses",
endpoint: providerEndpoint,
model: "provider-compact",
},
fetchDynamicModels: async () => [
{
...baseModel,
id: "dynamic-compact-model",
remoteCompaction: {
endpoint: modelEndpoint,
model: "model-compact",
},
},
],
},
"ext://runtime",
);
await registry.refreshRuntimeProviders("online");
const model = registry.find(providerName, "dynamic-compact-model");
expect(model?.remoteCompaction).toEqual({
enabled: true,
api: "openai-responses",
endpoint: modelEndpoint,
model: "model-compact",
});
});
test("registerProvider preserves explicit thinking and backfills wire facts", () => {
const config: ProviderConfigInput = {
baseUrl: "https://runtime.example.com/v1",
packages/coding-agent/test/model-registry.test.ts
+53
View File
@@ -1917,6 +1917,7 @@ describe("ModelRegistry", () => {
let specialCache: ModelRegistry;
let vertexAuthoritative: ModelRegistry;
let syntheticCacheLoad: ModelRegistry;
let cachedDiscoverableRemoteCompaction: ModelRegistry;
let vertexNonAuthoritative: ModelRegistry;
let vertexStale: ModelRegistry;
const vertexProjectModel = () =>
@@ -2123,6 +2124,49 @@ describe("ModelRegistry", () => {
),
},
);
cachedDiscoverableRemoteCompaction = readonlyRegistry(
{
providers: {
"cached-compact-proxy": {
baseUrl: "https://compact-proxy.example.com/v1",
apiKey: "TEST_KEY",
api: "openai-responses",
discovery: { type: "openai-models-list" },
remoteCompaction: {
enabled: true,
api: "openai-responses",
endpoint: "https://compact-proxy.example.com/v1/responses/provider-compact",
model: "provider-compact",
},
models: [],
},
},
},
{
seedCache: dbPath =>
writeModelCache(
"cached-compact-proxy:openai-models-list-context-v2",
Date.now(),
[
buildModel({
id: "cached-compact-model",
name: "Cached Compact Model",
api: "openai-responses",
provider: "cached-compact-proxy",
baseUrl: "https://compact-proxy.example.com/v1",
reasoning: true,
input: ["text"],
cost: { input: 0, output: 0, cacheRead: 0, cacheWrite: 0 },
contextWindow: 128_000,
maxTokens: 16_384,
}),
],
true,
"",
dbPath,
),
},
);
});
test("legacy cached discovery sentinels are ignored after nullable limit cutover", () => {
@@ -2151,6 +2195,15 @@ describe("ModelRegistry", () => {
expect(specialCache.find("openai-codex", "gpt-5.4-codex-pro")?.maxTokens).toBe(128_000);
});
test("applies provider remoteCompaction to cached configured discovery models", () => {
expect(cachedDiscoverableRemoteCompaction.find("cached-compact-proxy", "cached-compact-model")?.remoteCompaction).toEqual({
enabled: true,
api: "openai-responses",
endpoint: "https://compact-proxy.example.com/v1/responses/provider-compact",
model: "provider-compact",
});
});
test("replaces bundled google-vertex models with authoritative Vertex project discovery", () => {
const vertexModels = getModelsForProvider(vertexAuthoritative, "google-vertex");
expect(vertexModels.map(model => model.id)).toEqual(["zai-org/glm-4.7-maas"]);