fix(catalog): preserve Codex 872k context window for codex/gpt-6.1-sol (#20437)

This commit is contained in:
Rafa Martins
2026-11-17 23:42:59 -03:00
parent bc4b0c5f51
commit 71d1e05edf
6 changed files with 94 additions and 5 deletions

View File

@@ -0,0 +1 @@
- **fix(catalog):** keep Codex 872K context window for `codex/gpt-6.1-sol` when `models.dev` syncs the 1,050,000 public-API limit, and include `gpt-6.1-terra` in Codex `ultra` effort extension ([#20437](https://github.com/cheaper-inference/OmniRoute/issues/20437))

View File

@@ -49,7 +49,7 @@ const RESPONSES_REASONING_ENCRYPTED_CONTENT_INCLUDE = "reasoning.encrypted_conte
// include on this path. This compatibility policy makes no claim about the cause
// or resolution of upstream safety rejections.
const CODEX_MODEL_PATTERN =
/^(?:(?:codex|cx)\/)?(?:gpt-6(?:\.1)?-(?:astra|sol|luna)|gpt-5\.6-(?:sol|terra|luna))(?:-(?:none|low|medium|high|xhigh|max|ultra))?$/;
/^(?:(?:codex|cx)\/)?(?:gpt-6(?:\.1)?-(?:astra|sol|terra|luna)|gpt-5\.6-(?:sol|terra|luna))(?:-(?:none|low|medium|high|xhigh|max|ultra))?$/;
function isCodexModel(model: unknown): boolean {
if (typeof model !== "string") return false;
return CODEX_MODEL_PATTERN.test(model.trim().toLowerCase());

View File

@@ -987,7 +987,24 @@ export function getResolvedModelCapabilities(
const persistedContextWindow = usePersistedOverrides
? getContextOverride(resolved, snapshot)
: null;
const limitRows = { synced, registryModel, spec, authoritativeContextWindow };
// #20437: models.dev maps `openai` to both `openai` and `cx` (Codex), so a synced
// public-API context window (e.g. 1,050,000 for gpt-6.1-sol) can exceed Codex's
// actual OAuth window (872,000) in the provider registry. Do not let a larger
// synced limit override Codex's narrower registry window.
const effectiveSynced =
resolved.provider === "codex" &&
synced &&
typeof registryModel?.contextLength === "number" &&
typeof synced.limit_context === "number" &&
synced.limit_context > registryModel.contextLength
? { ...synced, limit_context: null }
: synced;
const limitRows = {
synced: effectiveSynced,
registryModel,
spec,
authoritativeContextWindow,
};
const contextLimit = pickContextWindow(limitRows, persistedContextWindow);
const contextWindow = contextLimit.value;

View File

@@ -18,7 +18,7 @@ export type VscodeCatalogModel = {
const STANDARD_EFFORT_SUFFIX_PATTERN = /-(xhigh|high|medium|low|none)$/i;
const CODEX_EXTENDED_EFFORT_SUFFIX_PATTERN =
/^(.*gpt-(?:5\.6-(?:sol|terra|luna)|6(?:\.1)?-(?:astra|sol|luna)))-(max|ultra)$/i;
/^(.*gpt-(?:5\.6-(?:sol|terra|luna)|6(?:\.1)?-(?:astra|sol|terra|luna)))-(max|ultra)$/i;
const DEFAULT_REASONING_EFFORT = "none";
const KNOWN_REASONING_EFFORTS = new Set(["none", "low", "medium", "high", "xhigh", "max", "ultra"]);
@@ -191,7 +191,7 @@ function getCodexGpt56DefaultReasoningEffort(model: VscodeCatalogModel) {
.trim()
.toLowerCase();
const match = providerModelId.match(
/^gpt-(?:5\.6-(sol|terra|luna)|6(?:\.1)?-(?:astra|sol|luna))(?:-(?:none|low|medium|high|xhigh|max|ultra))?$/
/^gpt-(?:5\.6-(sol|terra|luna)|6(?:\.1)?-(astra|sol|terra|luna))(?:-(?:none|low|medium|high|xhigh|max|ultra))?$/
);
if (!match) return undefined;
// The live catalog defaults GPT-5.6 Sol to low and every other model here to medium.

View File

@@ -41,7 +41,7 @@ export function extendCodexGpt56EffortValues(
}
const match = normalizedModel.match(
/^gpt-(?:5\.6-(sol|terra|luna)|6(?:\.1)?-(astra|sol|luna))(?:-(?:none|low|medium|high|xhigh|max|ultra))?$/
/^gpt-(?:5\.6-(sol|terra|luna)|6(?:\.1)?-(astra|sol|terra|luna))(?:-(?:none|low|medium|high|xhigh|max|ultra))?$/
);
if (!match) return values;

View File

@@ -14,6 +14,11 @@ import { extendCodexGpt56EffortValues } from "../../src/shared/reasoning/effortS
import { isCodexExtendedEffortBaseModel } from "../../src/shared/reasoning/codexExtendedEffort.ts";
import * as reasoningMetadata from "../../src/lib/vscode/reasoningMetadata.ts";
import { supportsVscodeServiceTierVariants } from "../../src/lib/vscode/serviceTierVariants.ts";
import {
getResolvedModelCapabilitiesWithLimitSources,
type ModelCapabilityResolutionSnapshot,
} from "../../src/lib/modelCapabilities.ts";
import { enrichCatalogModelEntry } from "../../src/lib/modelMetadataRegistry.ts";
const MODEL = "gpt-6.1-sol";
const EFFORTS = ["low", "medium", "high", "xhigh", "max", "ultra"];
@@ -175,3 +180,69 @@ test("Responses Lite keeps delegation for GPT-6.1 Sol ultra only", async () => {
assert.equal(captured[0].parallel_tool_calls, effort === "ultra");
}
});
test("#20437 codex/gpt-6.1-sol retains 872k Codex window even when synced models.dev carries 1,050,000 API limit", () => {
const syncedRow = {
tool_call: true,
reasoning: true,
attachment: true,
structured_output: true,
temperature: null,
modalities_input: JSON.stringify(["text", "image"]),
modalities_output: JSON.stringify(["text"]),
knowledge_cutoff: null,
release_date: null,
last_updated: null,
status: "stable",
family: "gpt-6",
open_weights: false,
limit_context: 1_050_000,
limit_input: 922_000,
limit_output: 128_000,
interleaved_field: null,
};
const snapshot: ModelCapabilityResolutionSnapshot = {
inputModalities: new Map(),
synced: {
openai: { [MODEL]: syncedRow },
cx: { [MODEL]: syncedRow },
},
maxTokenOverrides: new Map(),
maxInputTokenOverrides: new Map(),
reasoningEffortsOverrides: new Map(),
contextOverrides: new Map(),
customVisionOverrides: new Map(),
compatVisionOverrides: new Map(),
syncedAvailableModelVision: new Map(),
};
const codexCaps = getResolvedModelCapabilitiesWithLimitSources(`codex/${MODEL}`, snapshot);
assert.equal(codexCaps.capabilities.contextWindow, 872_000);
assert.equal(codexCaps.capabilities.maxInputTokens, 872_000);
assert.equal(codexCaps.limitSources.contextWindow, `registry:codex/${MODEL}`);
const openaiCaps = getResolvedModelCapabilitiesWithLimitSources(`openai/${MODEL}`, snapshot);
assert.equal(openaiCaps.capabilities.contextWindow, 1_050_000);
const enrichedCodex = enrichCatalogModelEntry(
{ id: `codex/${MODEL}`, owned_by: "codex", root: MODEL },
undefined,
{
modelsDevPricing: {},
userPricing: {},
capabilityResolutionSnapshot: snapshot,
}
);
assert.equal(enrichedCodex.context_length, 872_000);
assert.deepEqual(
extendCodexGpt56EffortValues("codex", "gpt-6.1-terra", [
"low",
"medium",
"high",
"xhigh",
"max",
]),
EFFORTS
);
});