fix(context): qualified key beats bare min when provider is explicit; bare wins for inferred provider

cursoragent · cursoragent · commit f8cf752c5970 · 2026-03-12T09:22:12.000Z
diff --git a/src/agents/context.lookup.test.ts b/src/agents/context.lookup.test.ts
@@ -195,30 +195,33 @@ describe("lookupContextTokens", () => {
     expect(result).toBe(200_000);
   });
 
-  it("resolveContextTokensForModel: bare key wins over qualified; OpenRouter raw entry not shadowed by Google config", async () => {
-    // applyConfiguredContextWindows writes "gemini-2.5-pro" → 2M (bare key, Google config).
-    // Discovery writes "google/gemini-2.5-pro" → 999k (OpenRouter raw qualified entry).
-    // Lookup order: bare first → finds 2M immediately, never reaching the
-    // OpenRouter raw "google/gemini-2.5-pro" qualified key.
-    mockDiscoveryDeps([{ id: "google/gemini-2.5-pro", contextWindow: 999_000 }], {
-      google: {
-        models: [{ id: "gemini-2.5-pro", contextWindow: 2_000_000 }],
+  it("resolveContextTokensForModel: config direct scan prevents OpenRouter qualified key collision for Google provider", async () => {
+    // When provider is explicitly "google" and cfg has a Google contextWindow
+    // override, the config direct scan returns it before any cache lookup —
+    // so the OpenRouter raw "google/gemini-2.5-pro" qualified entry is never hit.
+    // Real callers (status.summary.ts) always pass cfg when provider is explicit.
+    mockDiscoveryDeps([{ id: "google/gemini-2.5-pro", contextWindow: 999_000 }]);
+
+    const cfg = {
+      models: {
+        providers: {
+          google: { models: [{ id: "gemini-2.5-pro", contextWindow: 2_000_000 }] },
+        },
       },
-    });
+    };
 
     const { resolveContextTokensForModel } = await import("./context.js");
     await new Promise((r) => setTimeout(r, 0));
 
-    // Google provider: bare key "gemini-2.5-pro" → 2M (config-written) is found
-    // first; OpenRouter's "google/gemini-2.5-pro" qualified entry is never hit.
+    // Google with explicit cfg: config direct scan wins before any cache lookup.
     const googleResult = resolveContextTokensForModel({
+      cfg: cfg as never,
       provider: "google",
       model: "gemini-2.5-pro",
     });
     expect(googleResult).toBe(2_000_000);
 
-    // OpenRouter provider with slash model id: bare lookup "google/gemini-2.5-pro"
-    // → 999k (OpenRouter raw discovery entry, still intact).
+    // OpenRouter provider with slash model id: bare lookup finds the raw entry.
     const openrouterResult = resolveContextTokensForModel({
       provider: "openrouter",
       model: "google/gemini-2.5-pro",
@@ -297,4 +300,27 @@ describe("lookupContextTokens", () => {
     });
     expect(explicitResult).toBe(2_000_000);
   });
+
+  it("resolveContextTokensForModel: qualified key beats bare min when provider is explicit (original #35976 fix)", async () => {
+    // Regression: when both "gemini-3.1-pro-preview" (bare, min=128k) AND
+    // "google-gemini-cli/gemini-3.1-pro-preview" (qualified, 1M) are in cache,
+    // an explicit-provider call must return the provider-specific qualified value,
+    // not the collided bare minimum.
+    mockDiscoveryDeps([
+      { id: "github-copilot/gemini-3.1-pro-preview", contextWindow: 128_000 },
+      { id: "gemini-3.1-pro-preview", contextWindow: 128_000 },
+      { id: "google-gemini-cli/gemini-3.1-pro-preview", contextWindow: 1_048_576 },
+    ]);
+
+    const { resolveContextTokensForModel } = await import("./context.js");
+    await new Promise((r) => setTimeout(r, 0));
+
+    // Qualified "google-gemini-cli/gemini-3.1-pro-preview" → 1M wins over
+    // bare "gemini-3.1-pro-preview" → 128k (cross-provider minimum).
+    const result = resolveContextTokensForModel({
+      provider: "google-gemini-cli",
+      model: "gemini-3.1-pro-preview",
+    });
+    expect(result).toBe(1_048_576);
+  });
 });
diff --git a/src/agents/context.ts b/src/agents/context.ts
@@ -345,30 +345,38 @@ export function resolveContextTokensForModel(params: {
     }
   }
 
-  // Try bare key first. Bare keys contain the most-recently-written value
-  // (config overrides via applyConfiguredContextWindows run last) and are
-  // always correct for models registered with bare IDs.
+  // When provider is explicitly given and the model ID is bare (no slash),
+  // try the provider-qualified cache key BEFORE the bare key.  Discovery
+  // entries are stored under qualified IDs (e.g. "google-gemini-cli/
+  // gemini-3.1-pro-preview → 1M"), while the bare key may hold a cross-
+  // provider minimum (128k).  Returning the qualified entry gives the correct
+  // provider-specific window for /status and session context-token persistence.
+  //
+  // Guard: only when params.provider is explicit (not inferred from a slash in
+  // the model string). For model-only callers (e.g. status.ts log-usage
+  // fallback with model="google/gemini-2.5-pro"), the inferred provider would
+  // construct "google/gemini-2.5-pro" as the qualified key which accidentally
+  // matches OpenRouter's raw discovery entry — the bare lookup is correct there.
+  if (params.provider && ref && !ref.model.includes("/")) {
+    const qualifiedResult = lookupContextTokens(
+      `${normalizeProviderId(ref.provider)}/${ref.model}`,
+    );
+    if (qualifiedResult !== undefined) {
+      return qualifiedResult;
+    }
+  }
+
+  // Bare key fallback.  For model-only calls with slash-containing IDs
+  // (e.g. "google/gemini-2.5-pro") this IS the raw discovery cache key.
   const bareResult = lookupContextTokens(params.model);
   if (bareResult !== undefined) {
     return bareResult;
   }
 
-  // Bare key miss. As a last resort, try the provider-qualified key so that
-  // discovery entries stored under qualified IDs are reachable. The pi-ai
-  // registry can return entries like "google-gemini-cli/gemini-3.1-pro-preview"
-  // for the google-gemini-cli provider; without this fallback, resolving
-  // { provider: "google-gemini-cli", model: "gemini-3.1-pro-preview" } with
-  // no bare-key hit would always miss.
-  //
-  // Collision risk: the same keyspace holds raw slash-containing model IDs
-  // (e.g. OpenRouter's "google/gemini-2.5-pro"). This fallback is only reached
-  // when the bare key misses, which means native discovery for the requested
-  // provider produced no bare entry. In that situation the only available
-  // context-window data is what another provider stored under the same
-  // qualified key — in practice identical (OpenRouter mirrors native limits) —
-  // so returning it is strictly better than falling back to DEFAULT_CONTEXT_TOKENS.
-  // If native discovery IS running, it writes a bare key, and we never reach here.
-  if (ref && !ref.model.includes("/")) {
+  // When provider is implicit, try qualified as a last resort so inferred
+  // provider/model pairs (e.g. model="google-gemini-cli/gemini-3.1-pro")
+  // still find discovery entries stored under that qualified ID.
+  if (!params.provider && ref && !ref.model.includes("/")) {
     const qualifiedResult = lookupContextTokens(
       `${normalizeProviderId(ref.provider)}/${ref.model}`,
     );
