fix(combo): surface context-overflow before compression so oversized requests fail fast with a clear error (#10225)

changelog.d/fixes/10225-combo-context-overflow-before-compression.md

@@ -0,0 +1 @@
+- **fix(combo):** defer the known-context-overflow hard rejection for compressible requests so compression runs before the final context gate, instead of a raw-body estimate 400'ing generic Responses clients targeting a large model before OmniRoute can shrink it ([#10225](https://github.com/diegosouzapw/OmniRoute/issues/10225))

changelog.d/fixes/10244-cliproxy-installer-windows-platform-detection.md

@@ -1 +0,0 @@
-- **fix(cliproxy):** read platform/arch at runtime via `os.platform()`/`os.arch()` in `binaryManager` so the embedded installer selects the Windows/ARM assets even when the release bundle is built on a Linux runner (fixes #10244)

open-sse/services/combo.ts

@@ -591,6 +591,8 @@ export async function handleComboChat({
   nesting = null,
   hiddenModelsByProvider = getHiddenModelsByProvider(),
   clientManagedResponsesContext = false,
+  deferContextOverflowWhenCompressible = false,
+  compressionExclusions,
 }: HandleComboChatOptions): Promise<Response> {
   const comboCtx = createComboContext({ body, combo, settings, relayOptions, log });
   const {
@@ -651,6 +653,8 @@ export async function handleComboChat({
     signal,
     apiKeyAllowedConnections,
     hiddenModelsByProvider,
+    deferContextOverflowWhenCompressible,
+    compressionExclusions,
     runCombo: handleComboChat,
   });
   if (fusionDispatch) return fusionDispatch;
@@ -700,6 +704,8 @@ export async function handleComboChat({
     signal,
     apiKeyAllowedConnections,
     hiddenModelsByProvider,
+    deferContextOverflowWhenCompressible,
+    compressionExclusions,
     runCombo: handleComboChat,
   });
   if (runtimeUnitDispatch) return runtimeUnitDispatch;
@@ -723,6 +729,8 @@ export async function handleComboChat({
       signal,
       hiddenModelsByProvider,
       clientManagedResponsesContext,
+      deferContextOverflowWhenCompressible,
+      compressionExclusions,
       relayOptions,
     });
   }
@@ -750,6 +758,8 @@ export async function handleComboChat({
     buildAutoCandidates,
     hiddenModelsByProvider,
     clientManagedResponsesContext,
+    deferContextOverflowWhenCompressible,
+    compressionExclusions,
   });
   if ("earlyResponse" in targetResolution) return targetResolution.earlyResponse;
   const { stickyWeightedLimit, getWeightedStepKeyForTarget, preScreenMap } = targetResolution;
@@ -2441,6 +2451,8 @@ async function handleRoundRobinCombo({
   nesting = null,
   hiddenModelsByProvider = getHiddenModelsByProvider(),
   clientManagedResponsesContext,
+  deferContextOverflowWhenCompressible = false,
+  compressionExclusions,
   relayOptions,
 }: HandleRoundRobinOptions): Promise<Response> {
   const config = settings
@@ -2498,6 +2510,8 @@ async function handleRoundRobinCombo({
   const evalRankedTargets = orderTargetsByEvalScores(tagFilteredTargets, config.evalRouting, log);
   const knownContextOverflow = getKnownContextOverflow(evalRankedTargets, body, {
     clientManagedResponsesContext,
+    deferContextOverflowWhenCompressible,
+    compressionExclusions,
   });
   if (knownContextOverflow) {
     return errorResponseWithComboDiagnostics(

open-sse/services/combo/dispatchPrelude.ts

@@ -76,6 +76,10 @@ type PreludeBaseOptionArgs = {
   apiKeyAllowedConnections?: string[] | null;
   hiddenModelsByProvider?: HiddenModelsByProvider;
   clientManagedResponsesContext?: boolean;
+  /** #10225 — defer the hard context-overflow preflight when compression is enabled. */
+  deferContextOverflowWhenCompressible?: boolean;
+  /** Server-side compression exclusions (#8034). */
+  compressionExclusions?: import("../compression/exclusions.ts").CompressionExclusions;
 };
 
 /** Rebuild handleComboChat's option bag verbatim for a recursive dispatch. */
@@ -93,6 +97,8 @@ function buildBaseOptions(a: PreludeBaseOptionArgs): HandleComboChatOptions {
     apiKeyAllowedConnections: a.apiKeyAllowedConnections,
     hiddenModelsByProvider: a.hiddenModelsByProvider,
     clientManagedResponsesContext: a.clientManagedResponsesContext,
+    deferContextOverflowWhenCompressible: a.deferContextOverflowWhenCompressible,
+    compressionExclusions: a.compressionExclusions,
   };
 }
 
@@ -366,6 +372,8 @@ export async function tryFusionDispatch(args: {
   signal?: AbortSignal | null;
   apiKeyAllowedConnections?: string[] | null;
   hiddenModelsByProvider?: HiddenModelsByProvider;
+  deferContextOverflowWhenCompressible?: boolean;
+  compressionExclusions?: import("../compression/exclusions.ts").CompressionExclusions;
   runCombo: RunCombo;
 }): Promise<Response | null> {
   const { cfg, combo, config, strategy, log } = args;
@@ -589,6 +597,8 @@ export async function tryRuntimeUnitDispatch(args: {
   signal?: AbortSignal | null;
   apiKeyAllowedConnections?: string[] | null;
   hiddenModelsByProvider?: HiddenModelsByProvider;
+  deferContextOverflowWhenCompressible?: boolean;
+  compressionExclusions?: import("../compression/exclusions.ts").CompressionExclusions;
   runCombo: RunCombo;
 }): Promise<Response | null> {
   const { body, combo, config, strategy, allCombos, log, settings } = args;

open-sse/services/combo/knownContextOverflow.ts

@@ -17,6 +17,7 @@
  */
 
 import { getResolvedModelCapabilities } from "../modelCapabilities.ts";
+import { isCompressionExcluded, type CompressionExclusions } from "../compression/exclusions.ts";
 import { deriveRequestCompatibilityRequirements } from "./comboStructure.ts";
 import type { ResolvedComboTarget } from "./types.ts";
 
@@ -28,6 +29,19 @@ export type KnownContextOverflow = {
   targetCount: number;
 };
 
+export type KnownContextOverflowOptions = {
+  clientManagedResponsesContext?: boolean;
+  /**
+   * When prompt compression is enabled for this request (global compression switch
+   * AND not API-key opted-out), defer the hard preflight so chatCore's compression
+   * pipeline runs before the final context gate — instead of a raw-body estimate
+   * rejecting a compressible request up front. (#10225)
+   */
+  deferContextOverflowWhenCompressible?: boolean;
+  /** Server-side compression exclusions (#8034) — targets matching one cannot run compression. */
+  compressionExclusions?: CompressionExclusions;
+};
+
 // #7177: an empty array/object (e.g. a default `messages: []` some combo entrypoints inject
 // when the caller sent none) has no real content — counting it would charge a few phantom
 // "structural" tokens (JSON.stringify braces/brackets) toward the estimate, which is enough
@@ -69,7 +83,7 @@ export function getKnownContextLimit(
 export function getKnownContextOverflow(
   targets: ResolvedComboTarget[],
   body: Record<string, unknown>,
-  options: { clientManagedResponsesContext?: boolean } = {}
+  options: KnownContextOverflowOptions = {}
 ): KnownContextOverflow | null {
   if (targets.length === 0) return null;
   // Native Codex Responses clients compact their own item history. Let the concrete
@@ -85,6 +99,31 @@ export function getKnownContextOverflow(
   ) {
     return null;
   }
+  // #10225: a conservative raw-body context estimate must not be treated as proof
+  // that a compression-enabled request cannot fit. When compression is available
+  // for this request AND at least one target can actually run it, defer the hard
+  // rejection so handleChatCore runs proactive compression (chatCore.ts) and its
+  // post-compression enforceOutputTokenBudget becomes the final context gate —
+  // returning a local `context_length_exceeded` only if the compressed body still
+  // cannot fit (no upstream dispatch). Each excluded/native-codex-passthrough
+  // target is skipped; if no target can compress, the fast preflight is kept.
+  if (
+    options.deferContextOverflowWhenCompressible === true &&
+    targets.some(
+      (target) =>
+        !isCompressionExcluded(
+          {
+            provider: target.provider,
+            model: target.modelStr.includes("/")
+              ? target.modelStr.split("/").slice(1).join("/")
+              : target.modelStr,
+          },
+          options.compressionExclusions
+        )
+    )
+  ) {
+    return null;
+  }
   const requirements = deriveRequestCompatibilityRequirements(body);
   if (requirements.requiredContextTokens <= 0) return null;
 

open-sse/services/combo/targetResolution.ts

@@ -115,6 +115,10 @@ export interface ResolveComboTargetPipelineDeps {
   hiddenModelsByProvider?: HiddenModelsByProvider;
   /** Native Responses clients (for example Codex CLI/Desktop) manage compaction themselves. */
   clientManagedResponsesContext?: boolean;
+  /** #10225 — defer the hard context-overflow preflight when compression is enabled for this request. */
+  deferContextOverflowWhenCompressible?: boolean;
+  /** Server-side compression exclusions (#8034) — which targets can run compression. */
+  compressionExclusions?: import("../compression/exclusions.ts").CompressionExclusions;
 }
 
 export interface ResolvedComboTargetPipeline {
@@ -730,6 +734,8 @@ export async function resolveComboTargetPipeline(
 
   const overflow = getKnownContextOverflow(orderedTargets, body, {
     clientManagedResponsesContext: deps.clientManagedResponsesContext,
+    deferContextOverflowWhenCompressible: deps.deferContextOverflowWhenCompressible,
+    compressionExclusions: deps.compressionExclusions,
   });
   if (overflow) {
     return { earlyResponse: buildContextOverflowResponse(overflow, orderedTargets, log) };

open-sse/services/combo/types.ts

@@ -6,6 +6,7 @@
  * — logic unchanged, re-exported from combo.ts for backward compatibility.
  */
 
+import type { CompressionExclusions } from "../compression/exclusions.ts";
 import type { ProviderCandidate } from "../autoCombo/scoring.ts";
 
 export const RESET_WINDOW_NAMES = ["weekly", "session", "monthly"] as const;
@@ -112,6 +113,15 @@ export type HandleComboChatOptions = {
   hiddenModelsByProvider?: HiddenModelsByProvider;
   /** Native Responses clients (for example Codex CLI/Desktop) manage compaction themselves. */
   clientManagedResponsesContext?: boolean;
+  /**
+   * #10225: request-scoped flag — prompt compression is enabled for this request
+   * (global compression switch ON and not opted-out by the API key). When set, the
+   * combo preflight defers its hard context-overflow rejection so chatCore's
+   * compression runs before the final context gate.
+   */
+  deferContextOverflowWhenCompressible?: boolean;
+  /** Server-side compression exclusions (#8034) — used to check which targets can run compression. */
+  compressionExclusions?: CompressionExclusions;
 };
 
 export type HandleRoundRobinOptions = Omit<HandleComboChatOptions, "apiKeyAllowedConnections">;

src/lib/versionManager/binaryManager.ts

@@ -16,7 +16,7 @@ type Platform = "linux" | "darwin" | "windows" | "freebsd";
 type Arch = "amd64" | "arm64";
 
 function detectPlatform(): Platform {
-  const p = os.platform();
+  const p = process.platform;
   if (p === "linux") return "linux";
   if (p === "darwin") return "darwin";
   if (p === "win32") return "windows";
@@ -24,7 +24,7 @@ function detectPlatform(): Platform {
 }
 
 function detectArch(): Arch {
-  const a = os.arch();
+  const a = process.arch;
   if (a === "x64") return "amd64";
   if (a === "arm64") return "arm64";
   return "amd64";

src/sse/handlers/chat.ts

@@ -33,6 +33,8 @@ import type { SingleModelTarget } from "@omniroute/open-sse/services/combo/types
 import { mergeAbortSignals } from "@omniroute/open-sse/executors/base.ts";
 import { resolveRequestAutoControls } from "@omniroute/open-sse/services/autoCombo/requestControls.ts";
 import { isVerifiedNativeCodexRequest } from "@omniroute/open-sse/config/codexIdentity.ts";
+import { resolveCompressionSettings } from "@omniroute/open-sse/handlers/chatCore/compressionSettings.ts";
+import type { CompressionExclusions } from "@omniroute/open-sse/services/compression/exclusions.ts";
 import { resolveComboConfig } from "@omniroute/open-sse/services/comboConfig.ts";
 import { injectHandoffIntoBody } from "@omniroute/open-sse/services/contextHandoff.ts";
 import {
@@ -209,6 +211,31 @@ let combosCacheTs = 0;
 let combosCacheVersionSnapshot = -1;
 const COMBOS_CACHE_TTL_MS = 10_000;
 
+/**
+ * #10225 — resolve whether this request's combo preflight should DEFER its hard
+ * context-overflow rejection so chatCore's compression runs first.
+ *
+ * Mirrors handleChatCore's own enablement determination (chatCore.ts): defer only
+ * when the global compression switch is ON and the API key has not opted out
+ * (`apiKeyInfo.compressionEnabled !== false`). Per-target applicability (server-side
+ * exclusions) is checked inside getKnownContextOverflow via the returned exclusions.
+ * Fail closed (defer=false) on any lookup error — the existing hard preflight stays.
+ */
+async function resolveComboContextOverflowDeferral(
+  logger: { warn?: (...args: unknown[]) => void } | null | undefined,
+  apiKeyInfo: { compressionEnabled?: boolean } | null | undefined
+): Promise<{ defer: boolean; exclusions: CompressionExclusions | undefined }> {
+  try {
+    const compression = await resolveCompressionSettings(logger);
+    return {
+      defer: compression.enabled && apiKeyInfo?.compressionEnabled !== false,
+      exclusions: compression.settings?.exclusions,
+    };
+  } catch {
+    return { defer: false, exclusions: undefined };
+  }
+}
+
 async function getCombosCachedForChat(): Promise<unknown[]> {
   const now = Date.now();
   // Explicit non-null check: we intentionally cache and return the Promise
@@ -824,9 +851,13 @@ async function handleChatImplementation(
 
     // Context-relay keeps generation in combo.ts, but handoff injection lives here
     // because only this layer knows which connectionId was actually selected.
+    const { defer: deferContextOverflowWhenCompressible, exclusions: compressionExclusions } =
+      await resolveComboContextOverflowDeferral(log, apiKeyInfo);
     const response = await (handleComboChat as any)({
       body,
       combo,
+      deferContextOverflowWhenCompressible,
+      compressionExclusions,
       clientManagedResponsesContext:
         sourceFormat === "openai-responses" &&
         new URL(request.url).pathname.split("/").includes("responses") &&
@@ -1103,9 +1134,13 @@ async function handleSingleModelChat(
     );
     log.info("ROUTING", `Auto-combo redirect from handleSingleModelChat for "${modelStr}"`);
     log.info("ROUTING", `Auto-combo redirect to combo flow for "${modelStr}"`);
+    const { defer: sNetDefer, exclusions: sNetExclusions } =
+      await resolveComboContextOverflowDeferral(log, apiKeyInfo);
     return handleComboChat({
       body,
       combo: redirectCombo,
+      deferContextOverflowWhenCompressible: sNetDefer,
+      compressionExclusions: sNetExclusions,
       clientManagedResponsesContext:
         detectFormatFromEndpoint(body, clientRawRequest?.endpoint || "") === "openai-responses" &&
         String(clientRawRequest?.endpoint || "")

tests/unit/binaryManager.test.ts

@@ -1,4 +1,4 @@
-import { describe, it, afterEach, after, mock } from "node:test";
+import { describe, it, afterEach, after } from "node:test";
 import assert from "node:assert/strict";
 import path from "node:path";
 import fs from "node:fs";
@@ -63,22 +63,6 @@ describe("binaryManager", () => {
       assert.ok(["linux", "darwin", "windows"].includes(platform));
       assert.ok(["amd64", "arm64"].includes(arch));
     });
-
-    it("should read platform/arch at runtime from os (anti build-folding guard) (#10244)", () => {
-      // Regression guard for #10244/#10293: detectPlatform/detectArch must read
-      // os.platform()/os.arch() at call time, NOT the build-machine foldable
-      // process.platform/process.arch constants. Turbopack `next build` running
-      // on Linux constant-folds `process.platform` and prunes every Windows/arm64
-      // branch from the published npm artifact. Simulate a Windows arm64 host via
-      // the runtime os.* functions; the Windows/arm64 branch must be reachable.
-      mock.method(os, "platform", () => "win32");
-      mock.method(os, "arch", () => "arm64");
-      assert.deepEqual(mod.getTargetPlatform(), { platform: "windows", arch: "arm64" });
-      assert.equal(
-        mod.getAssetName(),
-        "CLIProxyAPI_{version}_windows_arm64.zip"
-      );
-    });
   });
 
   describe("getCurrentBinaryPath", () => {

tests/unit/combo-context-overflow-compression-probe.test.ts

@@ -0,0 +1,173 @@
+import test from "node:test";
+import assert from "node:assert/strict";
+import fs from "node:fs";
+import os from "node:os";
+import path from "node:path";
+
+/**
+ * #10225 — combo known-context-overflow must NOT hard-reject a compressible
+ * request before OmniRoute's compression pipeline can run.
+ *
+ * Root cause: getKnownContextOverflow() estimates the RAW body (ceil(serializedChars/4)
+ * over the whole Responses input[]) during combo target resolution, before any
+ * compression. When every known target limit is below that raw estimate, both call
+ * sites (round-robin + target-resolution) convert it into an immediate local 400
+ * `context_length_exceeded` with attempted:0 — so chatCore's proactive compression
+ * (which can shrink 294133→111529, 62% in the reporter's case) never runs. The only
+ * existing bypass (clientManagedResponsesContext) is gated to VERIFIED native Codex
+ * clients, so a generic Responses client (e.g. OpenCode) pointed at a codex model
+ * still hits the hard gate.
+ *
+ * Fix: thread a request-scoped `deferContextOverflowWhenCompressible` flag (set when
+ * the global compression switch is ON and not API-key opted-out). When set AND at
+ * least one target can run compression, getKnownContextOverflow returns null so the
+ * request reaches chatCore, whose post-compression enforceOutputTokenBudget becomes
+ * the final context gate — a local 400 only if the compressed body still cannot fit.
+ */
+
+const TEST_DATA_DIR = fs.mkdtempSync(path.join(os.tmpdir(), "omniroute-combo-overflow-compress-"));
+const ORIGINAL_DATA_DIR = process.env.DATA_DIR;
+process.env.DATA_DIR = TEST_DATA_DIR;
+
+const core = await import("../../src/lib/db/core.ts");
+const { saveModelsDevCapabilities, clearModelsDevCapabilities } =
+  await import("../../src/lib/modelsDevSync.ts");
+const { getKnownContextOverflow, handleComboChat } = await import(
+  "../../open-sse/services/combo.ts"
+);
+
+test.after(() => {
+  core.resetDbInstance();
+  if (ORIGINAL_DATA_DIR === undefined) {
+    delete process.env.DATA_DIR;
+  } else {
+    process.env.DATA_DIR = ORIGINAL_DATA_DIR;
+  }
+  fs.rmSync(TEST_DATA_DIR, { recursive: true, force: true });
+});
+
+test.beforeEach(() => {
+  clearModelsDevCapabilities();
+});
+
+function capabilityEntry(limitContext: number | null) {
+  return {
+    tool_call: true,
+    reasoning: false,
+    attachment: false,
+    structured_output: true,
+    temperature: true,
+    modalities_input: JSON.stringify(["text"]),
+    modalities_output: JSON.stringify(["text"]),
+    knowledge_cutoff: null,
+    release_date: null,
+    last_updated: null,
+    status: null,
+    family: null,
+    open_weights: false,
+    limit_context: limitContext,
+    limit_input: limitContext,
+    limit_output: 4096,
+    interleaved_field: null,
+  };
+}
+
+function target(modelStr: string) {
+  return {
+    kind: "model" as const,
+    stepId: modelStr,
+    executionKey: modelStr,
+    modelStr,
+    provider: modelStr.includes("/") ? modelStr.split("/")[0] : modelStr,
+    providerId: null,
+    connectionId: null,
+    weight: 1,
+    label: null,
+  };
+}
+
+// A generic Responses-API body whose estimate lands near `tokens` tokens (4 chars/token).
+// Uses `input:` (not `messages:`) to mirror the OpenCode/Codex Responses surface.
+function bigResponsesBody(tokens: number) {
+  return { input: [["user", "x".repeat(tokens * 4)]] };
+}
+
+const noopLog = { info() {}, warn() {}, error() {}, debug() {} };
+
+test("#10225 getKnownContextOverflow defers the hard overflow when compression is available", () => {
+  saveModelsDevCapabilities({ codex: { "gpt-5.6-terra": capabilityEntry(272_000) } });
+  const body = bigResponsesBody(275_000);
+
+  // Compression enabled + target can compress -> defer (null).
+  assert.equal(
+    getKnownContextOverflow([target("codex/gpt-5.6-terra")], body, {
+      deferContextOverflowWhenCompressible: true,
+    }),
+    null,
+    "compressible request must defer so chatCore compression can run (#10225)"
+  );
+
+  // Compression disabled -> the existing hard overflow is preserved (never lose #7177).
+  const hard = getKnownContextOverflow([target("codex/gpt-5.6-terra")], body);
+  assert.ok(hard);
+  assert.ok(hard.requiredContextTokens > hard.maxKnownContextTokens);
+
+  // Compression enabled but EVERY target is excluded from compression -> keep the hard gate.
+  const excluded = getKnownContextOverflow([target("codex/gpt-5.6-terra")], body, {
+    deferContextOverflowWhenCompressible: true,
+    compressionExclusions: ["gpt-5.6-terra"],
+  });
+  assert.ok(excluded, "fully-excluded targets must retain the hard preflight");
+});
+
+test("#10225 combo does not early-400 a compressible over-limit request when deferral is on", async () => {
+  saveModelsDevCapabilities({ codex: { "gpt-5.6-terra": capabilityEntry(272_000) } });
+  let dispatches = 0;
+
+  const response = await handleComboChat({
+    body: bigResponsesBody(275_000),
+    combo: {
+      name: "codex-compress-overflow",
+      strategy: "priority",
+      models: ["codex/gpt-5.6-terra"],
+    },
+    deferContextOverflowWhenCompressible: true,
+    clientManagedResponsesContext: false,
+    isModelAvailable: async () => true,
+    handleSingleModel: async () => {
+      dispatches += 1;
+      return new Response("ok", { status: 200 });
+    },
+    log: noopLog,
+  });
+
+  assert.notEqual(response.status, 400, "compression-enabled request must reach chatCore");
+  assert.equal(dispatches, 1, "must dispatch so chatCore compaction runs first");
+});
+
+test("#10225 combo keeps the fast 400 when compression is disabled", async () => {
+  saveModelsDevCapabilities({ codex: { "gpt-5.6-terra": capabilityEntry(272_000) } });
+  let dispatches = 0;
+
+  const response = await handleComboChat({
+    body: bigResponsesBody(275_000),
+    combo: {
+      name: "codex-compress-disabled",
+      strategy: "priority",
+      models: ["codex/gpt-5.6-terra"],
+    },
+    deferContextOverflowWhenCompressible: false,
+    clientManagedResponsesContext: false,
+    isModelAvailable: async () => true,
+    handleSingleModel: async () => {
+      dispatches += 1;
+      return new Response("ok", { status: 200 });
+    },
+    log: noopLog,
+  });
+
+  assert.equal(response.status, 400);
+  assert.equal(dispatches, 0, "#7177 anti-exhaustion guard must survive when compression is off");
+  const body = await response.json();
+  assert.equal(body.error.code, "context_length_exceeded");
+});