fix(model): normalize client context-window suffixes (#9193)

Validated in local merge-train (devbox-vm-06-dev002) @ combined-tip (FAST gates — only pre-existing audit.test.ts flake).

changelog.d/fixes/9193-context-window-suffix.md

@@ -0,0 +1 @@
+- **fix(model):** normalize client context-window suffixes for combo routing. (thanks @b1nhm1nh)

open-sse/services/combo/comboStructure.ts

@@ -13,13 +13,10 @@
 
 import { getModelContextLimit } from "../../../src/lib/modelCapabilities";
 import { getComboModelString, normalizeComboStep } from "../../../src/lib/combos/steps.ts";
-import {
-  getProviderByAlias,
-  getProviderById,
-} from "../../../src/shared/constants/providers.ts";
+import { getProviderByAlias, getProviderById } from "../../../src/shared/constants/providers.ts";
 import { estimateTokens } from "../contextManager.ts";
 import { getResolvedModelCapabilities } from "../modelCapabilities.ts";
-import { parseModel } from "../model.ts";
+import { parseModel, stripContextWindowSuffix } from "../model.ts";
 import { dedupeTargetsByExecutionKey, isRecord } from "./comboData.ts";
 import { getTargetProvider, MAX_COMBO_DEPTH } from "./comboPredicates.ts";
 import { evaluateContextLimit } from "./contextOverrideGate.ts";
@@ -323,7 +320,8 @@ export function getComboModelsFromData(
   modelStr: string,
   combosData: ComboCollectionLike
 ): string[] | null {
-  const combo = getComboFromData(modelStr, combosData);
+  const baseModelStr = stripContextWindowSuffix(modelStr);
+  const combo = getComboFromData(baseModelStr || modelStr, combosData);
   if (!combo) return null;
   return combo.models.map((m) => normalizeModelEntry(m).model);
 }

open-sse/services/model.ts

@@ -16,6 +16,16 @@ type ResolvedModelTarget = {
   model: string | null;
 };
 
+// Client context-window tags are routing hints, not part of provider model IDs.
+const CONTEXT_WINDOW_SUFFIX_RE = /\[(\d+)([kKmM])?\]\s*$/;
+
+export function stripContextWindowSuffix(
+  modelStr: string | null | undefined
+): string | null | undefined {
+  if (typeof modelStr !== "string" || !modelStr) return modelStr;
+  return modelStr.replace(CONTEXT_WINDOW_SUFFIX_RE, "").trimEnd();
+}
+
 // Derive alias→provider mapping from the single source of truth (PROVIDER_ID_TO_ALIAS)
 // This prevents the two maps from drifting out of sync
 const ALIAS_TO_PROVIDER_ID: Record<string, string> = {};
@@ -428,12 +438,12 @@ export function parseModel(modelStr: string | null | undefined): ParsedModel {
     };
   }
 
-  // Extract [1m] suffix before parsing provider/model
+  // Extract the legacy [1m] marker while stripping all client context tags.
   let extendedContext = false;
-  let cleanStr = modelStr;
-  if (cleanStr.endsWith("[1m]")) {
+  const cleanStripped = stripContextWindowSuffix(modelStr) as string;
+  let cleanStr = cleanStripped;
+  if (/\[1m\]\s*$/i.test(modelStr)) {
     extendedContext = true;
-    cleanStr = cleanStr.slice(0, -4);
   }
   cleanStr = cleanStr.trim();
 
@@ -665,7 +675,9 @@ async function resolveModelByProviderInference(modelId: string, extendedContext:
 
   // Canonicalize candidates (deduplicate alias providers pointing to the same provider ID)
   const canonicalCandidates = Array.from(
-    new Set(candidatesToUse.map((p) => resolveProviderAlias(p)).filter((p): p is string => p !== null))
+    new Set(
+      candidatesToUse.map((p) => resolveProviderAlias(p)).filter((p): p is string => p !== null)
+    )
   );
 
   // Filter candidates by active connections configured in the database

src/sse/handlers/chat.ts

@@ -20,7 +20,8 @@ import {
   recordModelLockoutFailure,
   isDailyQuotaExhausted,
 } from "@omniroute/open-sse/services/accountFallback.ts";
-import { getModelInfo, getComboForModel } from "../services/model";
+import { getCombo, getComboForModel, getModelInfo } from "../services/model";
+import { stripContextWindowSuffix } from "@omniroute/open-sse/services/model.ts";
 import { resolveBareModelToConnectionDefault } from "@omniroute/open-sse/services/model.ts";
 import { errorResponse } from "@omniroute/open-sse/utils/error.ts";
 import { getImageModelEntry } from "@omniroute/open-sse/config/imageRegistry.ts";
@@ -391,6 +392,17 @@ async function handleChatImplementation(
   // resolveRoutingModel). The resolved model still passes through
   // enforceApiKeyPolicy below, so it cannot bypass per-key allowlists.
   let modelStr = resolveRoutingModel(request, body);
+  if (typeof modelStr === "string") {
+    // Preserve literal combo names such as "Claude [1m]". Context tags are
+    // stripped only when the exact request does not identify a combo.
+    const exactCombo = await getCombo(modelStr);
+    if (!exactCombo) {
+      modelStr = stripContextWindowSuffix(modelStr) || modelStr;
+      if (body?.model !== modelStr) {
+        body = { ...body, model: modelStr };
+      }
+    }
+  }
 
   // cc discovery alias (`claude/<provider>/<model>`, `claude/combo/<name>`):
   // resolve back to the real id before any combo lookup / resolveModelOrError()

src/sse/services/model.ts

@@ -13,10 +13,11 @@ import {
   parseModel,
   getModelInfoCore,
   splitSyncedEffortSuffix,
+  stripContextWindowSuffix,
 } from "@omniroute/open-sse/services/model.ts";
 import { REGISTRY } from "@omniroute/open-sse/config/providerRegistry.ts";
 
-export { parseModel };
+export { parseModel, stripContextWindowSuffix };
 
 /**
  * Reserved provider prefixes — built-in provider ids + aliases. User-defined
@@ -140,7 +141,10 @@ function resolveSyncedModelIdAndEffort(
   }
   if (findSyncedModelMeta(syncedModels, modelId)) return { modelId, effort: null };
 
-  for (const candidate of syncedModels as Array<{ id?: unknown; supportedThinkingEfforts?: unknown }>) {
+  for (const candidate of syncedModels as Array<{
+    id?: unknown;
+    supportedThinkingEfforts?: unknown;
+  }>) {
     if (typeof candidate?.id !== "string" || !Array.isArray(candidate.supportedThinkingEfforts)) {
       continue;
     }
@@ -396,13 +400,21 @@ export async function getCombo(modelStr) {
  */
 export async function getComboForModel(modelStr) {
   // 1. Existing behavior — exact combo name match
-  const combo = await getCombo(modelStr);
+  let combo = await getCombo(modelStr);
   if (combo) return combo;
 
+  // Client context tags are ignored only after exact lookup, preserving literal
+  // combo names such as "Claude [1m]" while allowing "Claude[500k]" to use "Claude".
+  const baseModelStr = stripContextWindowSuffix(modelStr);
+  if (baseModelStr && baseModelStr !== modelStr) {
+    combo = await getCombo(baseModelStr);
+    if (combo) return combo;
+  }
+
   // 2. NEW — check model-combo mappings table (pattern match)
   try {
     const { resolveComboForModel } = await import("@/lib/localDb");
-    const mapped = await resolveComboForModel(modelStr);
+    const mapped = await resolveComboForModel(baseModelStr || modelStr);
     if (mapped && (mapped as any).models?.length > 0) {
       return mapped;
     }

tests/unit/chat-route-edge-cases.test.ts

@@ -15,6 +15,7 @@ const {
   settingsDb,
   idempotencyLayerModule,
   semanticCacheModule,
+  combosDb,
 } = harness;
 
 const { getBackgroundDegradationConfig } =
@@ -66,6 +67,61 @@ test("handleChat resolves model alias before routing", async () => {
   assert.equal(seenModels[0], "gpt-4.1", "Model alias should be resolved to gpt-4.1");
 });
 
+test("handleChat strips client context-window tags before combo routing and dispatch", async () => {
+  await seedConnection("openai", { apiKey: "sk-openai-context-tag" });
+  await combosDb.createCombo({
+    name: "context-tag-combo",
+    models: ["openai/gpt-4.1"],
+  });
+
+  const seenModels = [];
+  globalThis.fetch = async (_url, init = {}) => {
+    const body = JSON.parse(String(init.body));
+    seenModels.push(body.model);
+    return buildOpenAIResponse("Context tag response");
+  };
+
+  const response = await handleChat(
+    buildRequest({
+      body: {
+        model: "context-tag-combo[500k]",
+        stream: false,
+        messages: [{ role: "user", content: "Use the context-tagged combo" }],
+      },
+    })
+  );
+
+  assert.equal(response.status, 200);
+  assert.deepEqual(seenModels, ["gpt-4.1"]);
+});
+
+test("handleChat preserves a literal context-tagged combo name", async () => {
+  await seedConnection("openai", { apiKey: "sk-openai-literal-context-tag" });
+  await combosDb.createCombo({
+    name: "literal [1m]",
+    models: ["openai/gpt-4.1"],
+  });
+
+  const seenModels = [];
+  globalThis.fetch = async (_url, init = {}) => {
+    seenModels.push(JSON.parse(String(init.body)).model);
+    return buildOpenAIResponse("Literal combo response");
+  };
+
+  const response = await handleChat(
+    buildRequest({
+      body: {
+        model: "literal [1m]",
+        stream: false,
+        messages: [{ role: "user", content: "Use the literal combo" }],
+      },
+    })
+  );
+
+  assert.equal(response.status, 200);
+  assert.deepEqual(seenModels, ["gpt-4.1"]);
+});
+
 test("Test 3: handleChat returns cached response directly for Semantic Cache hits", async () => {
   await seedConnection("openai", { apiKey: "sk-openai-semantic" });
   let fetchCount = 0;

tests/unit/combo-bracket-names.test.ts

@@ -65,7 +65,7 @@ test("getComboForModel treats an exact bracketed name as a combo before model su
   assert.equal(parsedAsModel.model, "Claude");
 });
 
-test("getComboForModel does not strip bracket suffix when no exact bracketed combo exists", async () => {
+test("getComboForModel falls back to the base combo when no exact context-tagged combo exists", async () => {
   await combosDb.createCombo({
     name: "Claude",
     models: [{ provider: "claude", model: "claude-sonnet-4-6" }],
@@ -74,7 +74,7 @@ test("getComboForModel does not strip bracket suffix when no exact bracketed com
   const resolved = await sseModelService.getComboForModel("Claude [1m]");
   const parsedAsModel = sseModelService.parseModel("Claude [1m]");
 
-  assert.equal(resolved, null);
+  assert.equal(resolved?.name, "Claude");
   assert.equal(parsedAsModel.extendedContext, true);
   assert.equal(parsedAsModel.model, "Claude");
 });

tests/unit/combo-routing-engine.test.ts

@@ -187,6 +187,14 @@ test("getComboFromData and getComboModelsFromData resolve combos from array and
   assert.deepEqual(models, ["openai/gpt-4o-mini", "claude/sonnet"]);
 });
 
+test("getComboModelsFromData strips context-window tags before matching a combo", () => {
+  const combos = [{ name: "alpha", models: ["openai/gpt-4o-mini"] }];
+
+  assert.deepEqual(getComboModelsFromData("alpha[500k]", combos), ["openai/gpt-4o-mini"]);
+  assert.deepEqual(getComboModelsFromData("alpha[1M]", combos), ["openai/gpt-4o-mini"]);
+  assert.equal(getComboModelsFromData("alpha[beta]", combos), null);
+});
+
 test("validateComboDAG rejects circular references and resolveNestedComboModels expands nested combos", () => {
   const combos = [
     { name: "root", models: ["child-a", "openai/gpt-4o-mini"] },

tests/unit/model-parse.test.ts

@@ -1,6 +1,14 @@
 import { test } from "node:test";
 import assert from "node:assert/strict";
-import { normalizeCrossProxyModelId, parseModel } from "../../open-sse/services/model.ts";
+import {
+  normalizeCrossProxyModelId,
+  parseModel,
+  stripContextWindowSuffix,
+} from "../../open-sse/services/model.ts";
+
+test("context-window suffix helper strips client tags without changing the base model", () => {
+  assert.strictEqual(stripContextWindowSuffix("my-glm52[500k]"), "my-glm52");
+});
 
 // [1m] extended context suffix — PR #311 (DavyMassoneto)
 test("[1m] suffix: strips suffix and sets extendedContext=true", () => {
@@ -9,6 +17,12 @@ test("[1m] suffix: strips suffix and sets extendedContext=true", () => {
   assert.strictEqual(result.extendedContext, true);
 });
 
+test("[1M] suffix: preserves the extended-context flag case-insensitively", () => {
+  const result = parseModel("claude-sonnet-4-6[1M]");
+  assert.strictEqual(result.model, "claude-sonnet-4-6");
+  assert.strictEqual(result.extendedContext, true);
+});
+
 test("[1m] suffix: normal model has extendedContext=false", () => {
   const result = parseModel("claude-sonnet-4-6");
   assert.strictEqual(result.model, "claude-sonnet-4-6");