fix(combo): context-aware fallback ignores model_context_override (#7933)

* fix(combo): context-aware fallback ignores model_context_override

filterTargetsByRequestCompatibility resolved a target's context capacity through the override-free getResolvedModelCapabilities, so a persisted model_context_override (Feature 5004) never reached the compatibility filter. A provider whose catalog maxInputTokens is a deliberately small client-facing hint (below the real window so coding agents auto-compact, #6191) is then dropped from the fallback pool for large-context requests even when an operator recorded its true larger capacity. With only that provider left after Claude quota is exhausted, the combo returns a hard 503 with no fallback.

evaluateContextLimit now consults the raw override (getModelContextOverride, null when unset) before catalog limits, only when an override exists - so a genuinely-too-small maxInputTokens is still enforced for non-overridden models. Consistent with two other override-aware call sites in this file. Registry values (#6191 client hint) untouched. Tests: override rescues small-catalog target; without override too-small still dropped; existing #6191/#7039 tests pass (14/14).

* fix(quality): extract context-fit evaluation to keep comboStructure.ts under cap

The model_context_override fix grew open-sse/services/combo/comboStructure.ts
past the 800-line file-size cap. Extract evaluateContextLimit() (the
override-then-catalog context-fit check) into a new leaf,
open-sse/services/combo/contextOverrideGate.ts, so comboStructure.ts only
keeps the two call-site wires. No behavior change.

Co-authored-by: diegosouzapw <8016841+diegosouzapw@users.noreply.github.com>

---------

Co-authored-by: tmone <25759142+tmone@users.noreply.github.com>
Co-authored-by: diegosouzapw <8016841+diegosouzapw@users.noreply.github.com>

open-sse/services/combo/comboStructure.ts

@@ -18,6 +18,7 @@ import { getResolvedModelCapabilities } from "../modelCapabilities.ts";
 import { parseModel } from "../model.ts";
 import { dedupeTargetsByExecutionKey, isRecord } from "./comboData.ts";
 import { getTargetProvider, MAX_COMBO_DEPTH } from "./comboPredicates.ts";
+import { evaluateContextLimit } from "./contextOverrideGate.ts";
 import { hasEstimableContent } from "./knownContextOverflow.ts";
 import {
   normalizeModelEntry,
@@ -487,57 +488,13 @@ function exceedsKnownOutputLimit(
   return maxOutputTokens < requestedOutputTokens;
 }
 
-/**
- * Decide whether a target's known context limit accommodates the request.
- *
- * `maxInputTokens` is an **input-only** cap — the requested output reserve is
- * already enforced separately against `maxOutputTokens` (see
- * `exceedsKnownOutputLimit`), so it must NOT be re-counted here. Comparing
- * `maxInputTokens` against `estimatedInputTokens + requestedOutputTokens`
- * double-counted the output reserve and shrank the effective input allowance
- * (#7039).
- *
- * `contextWindow` is the total window, so input + output must both fit.
- *
- * Returns `true` when the known limit accommodates the request, `false` when
- * it is known to be too small, and `null` when no limit metadata is known.
- */
-function evaluateContextLimit(
-  capabilities: { maxInputTokens?: number | null; contextWindow?: number | null },
-  requirements: { estimatedInputTokens: number; requiredContextTokens: number }
-): boolean | null {
-  const hasMaxInput = capabilities.maxInputTokens != null;
-  const hasContextWindow = capabilities.contextWindow != null;
-
-  // Neither limit is known — cannot judge.
-  if (!hasMaxInput && !hasContextWindow) return null;
-
-  // The input-only cap must accommodate the estimated input.
-  const inputFits = hasMaxInput
-    ? capabilities.maxInputTokens! >= requirements.estimatedInputTokens
-    : true;
-
-  // The total window must accommodate input + requested output. The output
-  // reserve is enforced separately via `maxOutputTokens`, but when a model
-  // exposes both `maxInputTokens` and `contextWindow` the two must not be
-  // checked in isolation: a request whose input fits `maxInputTokens` but whose
-  // input + output exceeds `contextWindow` must still be rejected (#7039
-  // follow-up — shared-window models where `maxInputTokens` defaults to the
-  // total window size).
-  const totalFits = hasContextWindow
-    ? capabilities.contextWindow! >= requirements.requiredContextTokens
-    : true;
-
-  return inputFits && totalFits;
-}
-
 function hasKnownCompatibleContextLimit(
   target: ResolvedComboTarget,
   requirements: RequestCompatibilityRequirements
 ): boolean {
   if (requirements.requiredContextTokens <= 0) return false;
   const capabilities = getResolvedModelCapabilities(target.modelStr);
-  return evaluateContextLimit(capabilities, requirements) === true;
+  return evaluateContextLimit(capabilities, requirements, target.modelStr) === true;
 }
 
 function hasOnlyContextWindowFailures(reasons: string[]): boolean {
@@ -576,7 +533,7 @@ function getTargetCompatibilityFailures(
     failures.push("output_tokens");
   }
 
-  const contextVerdict = evaluateContextLimit(capabilities, requirements);
+  const contextVerdict = evaluateContextLimit(capabilities, requirements, target.modelStr);
   if (requirements.requiredContextTokens > 0 && contextVerdict === false) {
     failures.push("context_window");
   }

open-sse/services/combo/contextOverrideGate.ts

@@ -0,0 +1,94 @@
+/**
+ * Context-fit evaluation for combo routing's compatibility filter, extracted
+ * from comboStructure.ts to keep that file under the file-size cap (PR
+ * #7933's model_context_override fix pushed it over).
+ *
+ * evaluateContextLimit() is the single chokepoint both compatibility-check
+ * call sites in comboStructure.ts (hasKnownCompatibleContextLimit,
+ * getTargetCompatibilityFailures) go through. It first consults a persisted
+ * per-model context override, then falls back to the catalog's
+ * maxInputTokens/contextWindow limits.
+ *
+ * Override rationale (Feature 5004): the catalog's `maxInputTokens` can be a
+ * deliberately smaller *client-facing* hint (e.g. set below the true window so
+ * coding agents auto-compact — #6191); using it to filter fallback targets
+ * wrongly drops otherwise-capable providers for large prompts, collapsing the
+ * pool to one provider and producing a hard 503 with no fallback once that
+ * provider's quota is exhausted. An operator-set or auto-discovered override
+ * reflects the real capacity, so it supersedes both catalog limits. Uses the
+ * raw override (`getModelContextOverride` returns `null` when none is set) —
+ * NOT `getModelContextLimitForModelString`, which falls back to
+ * `contextWindow` and would therefore bypass the `maxInputTokens` cap for
+ * every model, not just overridden ones.
+ */
+
+import { getModelContextOverride } from "../../../src/lib/db/modelContextOverrides";
+import { parseModel } from "../model.ts";
+
+/**
+ * Resolve the context-fit verdict from a persisted per-model override, if one
+ * is set. Returns `undefined` when there is no `modelStr` or no override
+ * exists, so the caller falls through to the catalog-based check; otherwise
+ * returns the fit verdict for the override itself.
+ */
+function resolveContextOverrideVerdict(
+  modelStr: string | undefined,
+  requiredContextTokens: number
+): boolean | undefined {
+  if (!modelStr) return undefined;
+  const parsed = parseModel(modelStr);
+  const override = getModelContextOverride(parsed.provider, parsed.model);
+  if (override == null) return undefined;
+  return override >= requiredContextTokens;
+}
+
+/**
+ * Decide whether a target's known context limit accommodates the request.
+ *
+ * `maxInputTokens` is an **input-only** cap — the requested output reserve is
+ * already enforced separately against `maxOutputTokens` (see
+ * `exceedsKnownOutputLimit` in comboStructure.ts), so it must NOT be
+ * re-counted here. Comparing `maxInputTokens` against `estimatedInputTokens +
+ * requestedOutputTokens` double-counted the output reserve and shrank the
+ * effective input allowance (#7039).
+ *
+ * `contextWindow` is the total window, so input + output must both fit.
+ *
+ * Returns `true` when the known limit accommodates the request, `false` when
+ * it is known to be too small, and `null` when no limit metadata is known.
+ */
+export function evaluateContextLimit(
+  capabilities: { maxInputTokens?: number | null; contextWindow?: number | null },
+  requirements: { estimatedInputTokens: number; requiredContextTokens: number },
+  modelStr?: string
+): boolean | null {
+  const overrideVerdict = resolveContextOverrideVerdict(
+    modelStr,
+    requirements.requiredContextTokens
+  );
+  if (overrideVerdict !== undefined) return overrideVerdict;
+
+  const hasMaxInput = capabilities.maxInputTokens != null;
+  const hasContextWindow = capabilities.contextWindow != null;
+
+  // Neither limit is known — cannot judge.
+  if (!hasMaxInput && !hasContextWindow) return null;
+
+  // The input-only cap must accommodate the estimated input.
+  const inputFits = hasMaxInput
+    ? capabilities.maxInputTokens! >= requirements.estimatedInputTokens
+    : true;
+
+  // The total window must accommodate input + requested output. The output
+  // reserve is enforced separately via `maxOutputTokens`, but when a model
+  // exposes both `maxInputTokens` and `contextWindow` the two must not be
+  // checked in isolation: a request whose input fits `maxInputTokens` but whose
+  // input + output exceeds `contextWindow` must still be rejected (#7039
+  // follow-up — shared-window models where `maxInputTokens` defaults to the
+  // total window size).
+  const totalFits = hasContextWindow
+    ? capabilities.contextWindow! >= requirements.requiredContextTokens
+    : true;
+
+  return inputFits && totalFits;
+}

tests/unit/combo-context-window-filter.test.ts

@@ -18,6 +18,8 @@ const { saveModelsDevCapabilities, clearModelsDevCapabilities } =
   await import("../../src/lib/modelsDevSync.ts");
 const { filterTargetsByRequestCompatibility, getKnownContextOverflow, handleComboChat } =
   await import("../../open-sse/services/combo.ts");
+const { setModelContextOverride, removeModelContextOverride } =
+  await import("../../src/lib/db/modelContextOverrides.ts");
 
 test.after(() => {
   core.resetDbInstance();
@@ -364,3 +366,53 @@ test("maxInputTokens defaulting to contextWindow still rejects when input + outp
     ["unit-7039-window/huge"]
   );
 });
+
+// A persisted model_context_override (Feature 5004 — operator-set or
+// auto-discovered) reflects a target's real usable window and must win over the
+// static catalog limit for server-side routing. Otherwise a provider whose
+// catalog `maxInputTokens` is a deliberately smaller client-facing hint (set
+// below the true window so coding agents auto-compact, #6191) gets wrongly
+// dropped for large-context requests, collapsing the fallback pool.
+test("model_context_override lets a small-catalog target survive a large-context request", () => {
+  saveModelsDevCapabilities({
+    "unit-override": {
+      big: capabilityEntry(1_000_000),
+      capped: capabilityEntry(8_000),
+    },
+  });
+  setModelContextOverride("unit-override", "capped", 1_000_000);
+  try {
+    const out = filterTargetsByRequestCompatibility(
+      [target("unit-override/capped"), target("unit-override/big")],
+      largeContextBody(),
+      noopLog
+    );
+    assert.deepEqual(
+      out.map((entry) => entry.modelStr).sort(),
+      ["unit-override/big", "unit-override/capped"]
+    );
+  } finally {
+    removeModelContextOverride("unit-override", "capped");
+  }
+});
+
+test("without an override the small-catalog target is still dropped for the large request", () => {
+  saveModelsDevCapabilities({
+    "unit-override": {
+      big: capabilityEntry(1_000_000),
+      capped: capabilityEntry(8_000),
+    },
+  });
+  // No override: capped (8K) is genuinely too small and must be filtered out,
+  // guarding the override read-path from masking a real too-small target.
+  const out = filterTargetsByRequestCompatibility(
+    [target("unit-override/capped"), target("unit-override/big")],
+    largeContextBody(),
+    noopLog
+  );
+
+  assert.deepEqual(
+    out.map((entry) => entry.modelStr),
+    ["unit-override/big"]
+  );
+});