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))

open-sse/services/combo.ts

@@ -591,6 +591,11 @@ export async function handleComboChat({
   nesting = null,
   hiddenModelsByProvider = getHiddenModelsByProvider(),
   clientManagedResponsesContext = false,
+  deferContextOverflowWhenCompressible = false,
+  compressionExclusions,
+  sourceFormat = null,
+  endpointPath = null,
+  requestHeaders = null,
 }: HandleComboChatOptions): Promise<Response> {
   const comboCtx = createComboContext({ body, combo, settings, relayOptions, log });
   const {
@@ -651,6 +656,11 @@ export async function handleComboChat({
     signal,
     apiKeyAllowedConnections,
     hiddenModelsByProvider,
+    deferContextOverflowWhenCompressible,
+    compressionExclusions,
+    sourceFormat,
+    endpointPath,
+    requestHeaders,
     runCombo: handleComboChat,
   });
   if (fusionDispatch) return fusionDispatch;
@@ -700,6 +710,11 @@ export async function handleComboChat({
     signal,
     apiKeyAllowedConnections,
     hiddenModelsByProvider,
+    deferContextOverflowWhenCompressible,
+    compressionExclusions,
+    sourceFormat,
+    endpointPath,
+    requestHeaders,
     runCombo: handleComboChat,
   });
   if (runtimeUnitDispatch) return runtimeUnitDispatch;
@@ -723,6 +738,11 @@ export async function handleComboChat({
       signal,
       hiddenModelsByProvider,
       clientManagedResponsesContext,
+      deferContextOverflowWhenCompressible,
+      compressionExclusions,
+      sourceFormat,
+      endpointPath,
+      requestHeaders,
       relayOptions,
     });
   }
@@ -750,6 +770,11 @@ export async function handleComboChat({
     buildAutoCandidates,
     hiddenModelsByProvider,
     clientManagedResponsesContext,
+    deferContextOverflowWhenCompressible,
+    compressionExclusions,
+    sourceFormat,
+    endpointPath,
+    requestHeaders,
   });
   if ("earlyResponse" in targetResolution) return targetResolution.earlyResponse;
   const { stickyWeightedLimit, getWeightedStepKeyForTarget, preScreenMap } = targetResolution;
@@ -2441,6 +2466,11 @@ async function handleRoundRobinCombo({
   nesting = null,
   hiddenModelsByProvider = getHiddenModelsByProvider(),
   clientManagedResponsesContext,
+  deferContextOverflowWhenCompressible = false,
+  compressionExclusions,
+  sourceFormat = null,
+  endpointPath = null,
+  requestHeaders = null,
   relayOptions,
 }: HandleRoundRobinOptions): Promise<Response> {
   const config = settings
@@ -2498,6 +2528,11 @@ async function handleRoundRobinCombo({
   const evalRankedTargets = orderTargetsByEvalScores(tagFilteredTargets, config.evalRouting, log);
   const knownContextOverflow = getKnownContextOverflow(evalRankedTargets, body, {
     clientManagedResponsesContext,
+    deferContextOverflowWhenCompressible,
+    compressionExclusions,
+    sourceFormat,
+    endpointPath,
+    requestHeaders,
   });
   if (knownContextOverflow) {
     return errorResponseWithComboDiagnostics(

open-sse/services/combo/dispatchPrelude.ts

@@ -76,6 +76,14 @@ 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;
+  /** #10503 — request-shape facts for the target-aware deferral check (see knownContextOverflow.ts). */
+  sourceFormat?: string | null;
+  endpointPath?: string | null;
+  requestHeaders?: Headers | Record<string, unknown> | null;
 };
 
 /** Rebuild handleComboChat's option bag verbatim for a recursive dispatch. */
@@ -93,6 +101,11 @@ function buildBaseOptions(a: PreludeBaseOptionArgs): HandleComboChatOptions {
     apiKeyAllowedConnections: a.apiKeyAllowedConnections,
     hiddenModelsByProvider: a.hiddenModelsByProvider,
     clientManagedResponsesContext: a.clientManagedResponsesContext,
+    deferContextOverflowWhenCompressible: a.deferContextOverflowWhenCompressible,
+    compressionExclusions: a.compressionExclusions,
+    sourceFormat: a.sourceFormat,
+    endpointPath: a.endpointPath,
+    requestHeaders: a.requestHeaders,
   };
 }
 
@@ -366,6 +379,11 @@ export async function tryFusionDispatch(args: {
   signal?: AbortSignal | null;
   apiKeyAllowedConnections?: string[] | null;
   hiddenModelsByProvider?: HiddenModelsByProvider;
+  deferContextOverflowWhenCompressible?: boolean;
+  compressionExclusions?: import("../compression/exclusions.ts").CompressionExclusions;
+  sourceFormat?: string | null;
+  endpointPath?: string | null;
+  requestHeaders?: Headers | Record<string, unknown> | null;
   runCombo: RunCombo;
 }): Promise<Response | null> {
   const { cfg, combo, config, strategy, log } = args;
@@ -589,6 +607,11 @@ export async function tryRuntimeUnitDispatch(args: {
   signal?: AbortSignal | null;
   apiKeyAllowedConnections?: string[] | null;
   hiddenModelsByProvider?: HiddenModelsByProvider;
+  deferContextOverflowWhenCompressible?: boolean;
+  compressionExclusions?: import("../compression/exclusions.ts").CompressionExclusions;
+  sourceFormat?: string | null;
+  endpointPath?: string | null;
+  requestHeaders?: Headers | Record<string, unknown> | null;
   runCombo: RunCombo;
 }): Promise<Response | null> {
   const { body, combo, config, strategy, allCombos, log, settings } = args;

open-sse/services/combo/knownContextOverflow.ts

@@ -17,6 +17,8 @@
  */
 
 import { getResolvedModelCapabilities } from "../modelCapabilities.ts";
+import { isCompressionExcluded, type CompressionExclusions } from "../compression/exclusions.ts";
+import { shouldUseNativeCodexPassthrough } from "../../handlers/chatCore/passthroughHelpers.ts";
 import { deriveRequestCompatibilityRequirements } from "./comboStructure.ts";
 import type { ResolvedComboTarget } from "./types.ts";
 
@@ -28,6 +30,29 @@ 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;
+  /**
+   * #10503: the exact request-shape facts chatCore.ts uses to decide
+   * `shouldUseNativeCodexPassthrough` (open-sse/handlers/chatCore/passthroughHelpers.ts) —
+   * threaded down so the deferral decision below can be target-aware instead of
+   * relying on the looser `clientManagedResponsesContext` proxy. Reused verbatim
+   * (not re-derived) so the combo-layer decision can never drift from chatCore's own.
+   */
+  sourceFormat?: string | null;
+  endpointPath?: string | null;
+  requestHeaders?: Headers | Record<string, unknown> | null;
+};
+
 // #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 +94,7 @@ 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 +110,55 @@ export function getKnownContextOverflow(
   ) {
     return null;
   }
+  // #10225 / #10499-sweep #10503: 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).
+  //
+  // Target-awareness is load-bearing here: a target is only a valid reason to defer
+  // when handleChatCore will ACTUALLY attempt compression for it. Two classes are
+  // excluded from "can compress" even though `isCompressionExcluded` (operator
+  // exclusions) says nothing about them:
+  //   - Operator-excluded targets (#8034, existing `isCompressionExcluded` check).
+  //   - Native Codex Responses passthrough targets: chatCore.ts unconditionally sets
+  //     `compressionExcluded = nativeCodexPassthrough || ...` for these, computed via
+  //     `shouldUseNativeCodexPassthrough()` (chatCore/passthroughHelpers.ts) — called
+  //     here with the SAME request-shape facts (sourceFormat/endpointPath/headers)
+  //     chatCore itself uses, reused verbatim rather than re-derived from the looser
+  //     `clientManagedResponsesContext` flag (which always requires a VERIFIED native
+  //     client; chatCore's own gate does NOT for provider==="codex" — see
+  //     shouldUseNativeCodexPassthrough's `provider === "codex" || isVerifiedNativeCodexRequest`
+  //     short-circuit). Deferring on such a target's account would let an oversized
+  //     body sail straight through to `fetch()` uncompressed instead of being caught
+  //     by either preflight — silently defeating the whole point of this feature.
+  // If NO target can compress, the fast raw-body preflight is kept (unchanged).
+  if (
+    options.deferContextOverflowWhenCompressible === true &&
+    targets.some((target) => {
+      const isNativeCodexPassthroughTarget = shouldUseNativeCodexPassthrough({
+        provider: target.provider,
+        sourceFormat: options.sourceFormat,
+        endpointPath: options.endpointPath,
+        body,
+        headers: options.requestHeaders,
+      });
+      if (isNativeCodexPassthroughTarget) return false;
+      return !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 +115,14 @@ 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;
+  /** #10503 — request-shape facts for the target-aware deferral check (see knownContextOverflow.ts). */
+  sourceFormat?: string | null;
+  endpointPath?: string | null;
+  requestHeaders?: Headers | Record<string, unknown> | null;
 }
 
 export interface ResolvedComboTargetPipeline {
@@ -730,6 +738,11 @@ export async function resolveComboTargetPipeline(
 
   const overflow = getKnownContextOverflow(orderedTargets, body, {
     clientManagedResponsesContext: deps.clientManagedResponsesContext,
+    deferContextOverflowWhenCompressible: deps.deferContextOverflowWhenCompressible,
+    compressionExclusions: deps.compressionExclusions,
+    sourceFormat: deps.sourceFormat,
+    endpointPath: deps.endpointPath,
+    requestHeaders: deps.requestHeaders,
   });
   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 +113,25 @@ 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;
+  /**
+   * #10503: request-shape facts (mirroring chatCore.ts's own resolution) threaded
+   * down to getKnownContextOverflow so the deferral decision can be target-aware —
+   * a native-Codex-Responses-passthrough target must never count as "compressible"
+   * (chatCore disables compression for it unconditionally). See
+   * knownContextOverflow.ts::KnownContextOverflowOptions for the full rationale.
+   */
+  sourceFormat?: string | null;
+  endpointPath?: string | null;
+  requestHeaders?: Headers | Record<string, unknown> | null;
 };
 
 export type HandleRoundRobinOptions = Omit<HandleComboChatOptions, "apiKeyAllowedConnections">;

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 +851,20 @@ 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,
+      // #10503: same request-shape facts chatCore.ts resolves for itself
+      // (resolveChatCoreRequestFormat), so getKnownContextOverflow's target-aware
+      // deferral check can never drift from chatCore's own native-codex-passthrough
+      // decision. See knownContextOverflow.ts::KnownContextOverflowOptions.
+      sourceFormat,
+      endpointPath: new URL(request.url).pathname,
+      requestHeaders: request.headers,
       clientManagedResponsesContext:
         sourceFormat === "openai-responses" &&
         new URL(request.url).pathname.split("/").includes("responses") &&
@@ -1103,11 +1141,22 @@ 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);
+    // #10503: same request-shape facts chatCore.ts resolves for itself — threaded
+    // down so getKnownContextOverflow's target-aware deferral check can never drift
+    // from chatCore's own native-codex-passthrough decision.
+    const sNetSourceFormat = detectFormatFromEndpoint(body, clientRawRequest?.endpoint || "");
     return handleComboChat({
       body,
       combo: redirectCombo,
+      deferContextOverflowWhenCompressible: sNetDefer,
+      compressionExclusions: sNetExclusions,
+      sourceFormat: sNetSourceFormat,
+      endpointPath: clientRawRequest?.endpoint || "",
+      requestHeaders: clientRawRequest?.headers,
       clientManagedResponsesContext:
-        detectFormatFromEndpoint(body, clientRawRequest?.endpoint || "") === "openai-responses" &&
+        sNetSourceFormat === "openai-responses" &&
         String(clientRawRequest?.endpoint || "")
           .split("/")
           .includes("responses") &&