fix(combo): resolve effort variants via base model capabilities (#11034)

⭐5 — Variantes de modelo com sufixo de effort (ex: model-max/model-xhigh) são sintetizadas no catálogo a partir de supportedThinkingEfforts do modelo base, sem linha própria no registry — getResolvedModelCapabilities resolvia NULL para tool/vision/context, o que fazia combos com tools reordenar silenciosamente essas variantes para trás. Fix: quando todas as fontes diretas falham, resolve as capacidades a partir do modelo base após strip do sufixo. TDD, 65 testes focados verdes.

changelog.d/fixes/command-code-effort-capabilities.md

@@ -0,0 +1 @@
+- fix(combo): resolve effort-suffixed command-code variants (e.g. `deepseek-v4-flash-max`) to their base model for capability lookups, so tool-bearing combo requests keep the declared priority order instead of reordering behind models with confirmed capabilities

src/lib/modelCapabilities.ts

@@ -254,6 +254,39 @@ function leafModelId(modelId: string | null | undefined): string | null {
   return leaf && leaf !== modelId ? leaf : null;
 }
 
+/**
+ * Effort suffixes the catalog synthesizes as `<base>-<tier>` variant ids from a
+ * base model's `supportedThinkingEfforts` (mirrors REGISTERED_EFFORT_SUFFIXES
+ * in open-sse/utils/registeredEffortVariants.ts, plus `minimal` for muse).
+ */
+const EFFORT_VARIANT_SUFFIXES = [
+  "none",
+  "minimal",
+  "low",
+  "medium",
+  "high",
+  "xhigh",
+  "max",
+] as const;
+
+/**
+ * Strip a trailing effort-tier suffix off a model id (e.g.
+ * `deepseek-v4-flash-max` → `deepseek-v4-flash`). Longest token first so
+ * `xhigh` is matched before `high`. Returns null when no known suffix matches
+ * or the id would be left empty.
+ */
+function stripKnownEffortSuffix(modelId: string): string | null {
+  const normalized = String(modelId || "").trim();
+  if (!normalized) return null;
+  for (const suffix of EFFORT_VARIANT_SUFFIXES) {
+    const token = `-${suffix}`;
+    if (normalized.length > token.length && normalized.endsWith(token)) {
+      return normalized.slice(0, -token.length);
+    }
+  }
+  return null;
+}
+
 function getStaticSpec(modelId: string | null, rawModel: string | null): ModelSpec | undefined {
   if (modelId) {
     const byCanonical = getModelSpec(modelId);
@@ -704,15 +737,38 @@ export function getResolvedModelCapabilities(
   // persisted override never feeds back into the comparison that (re)writes it.
   const usePersistedOverrides = options?.persistedOverrides !== false;
   const resolved = resolveCapabilityInput(input);
-  const spec = getStaticSpec(resolved.model, resolved.rawModel);
+  let spec = getStaticSpec(resolved.model, resolved.rawModel);
-  const registryModel = getRegistryModel(resolved.provider, resolved.model);
+  let registryModel = getRegistryModel(resolved.provider, resolved.model);
-  const synced = getSyncedCapabilityForResolved(
+  let synced = getSyncedCapabilityForResolved(
     resolved.provider,
     resolved.model,
     resolved.rawModel,
     snapshot
   );
 
+  // Effort-suffixed variants (e.g. command-code `deepseek-v4-flash-max`,
+  // `meta/muse-spark-1.2-contributor-xhigh`) are synthesized in the catalog
+  // from the base model's `supportedThinkingEfforts`; they have no registry
+  // row, synced row, or static spec of their own. Without a base-model
+  // fallback the variant resolves with NULL tool/vision/context capabilities,
+  // so a tool-bearing combo request treats the target as incompatible and
+  // silently reorders it behind models with confirmed capabilities. Resolve
+  // the variant's capabilities from its base model when every direct source
+  // misses.
+  if (!spec && !registryModel && !synced && resolved.provider && resolved.model) {
+    const baseModelId = stripKnownEffortSuffix(resolved.model);
+    if (baseModelId && baseModelId !== resolved.model) {
+      spec = getStaticSpec(baseModelId, resolved.rawModel);
+      registryModel = getRegistryModel(resolved.provider, baseModelId);
+      synced = getSyncedCapabilityForResolved(
+        resolved.provider,
+        baseModelId,
+        resolved.rawModel,
+        snapshot
+      );
+    }
+  }
+
   const modalitiesInput = parseModalities(synced?.modalities_input);
   const modalitiesOutput = parseModalities(synced?.modalities_output);
   const lookupKey =

tests/unit/command-code-registry-vision.test.ts

@@ -74,3 +74,53 @@ test("MiniMax M3 via command-code keeps existing vision capability (no regressio
   const caps = getResolvedModelCapabilities("command-code/MiniMaxAI/MiniMax-M3");
   assert.equal(caps.supportsVision, true);
 });
+
+test("command-code effort-suffixed variants resolve capabilities from their base model", async () => {
+  // Effort variants (e.g. `-max`, `-xhigh`) are synthesized from the base's
+  // supportedThinkingEfforts and have no registry/synced row of their own.
+  // Without the base-model fallback they resolve NULL tool/vision/context,
+  // which makes a tool-bearing combo request drop them behind confirmed
+  // targets (observed: orchestrator tried opencode-go/mimo-v2.5-max at
+  // position 2 while command-code deepseek sat unused at its declared
+  // priority position 2).
+  //
+  // Seed the models.dev capability store (the source getResolvedModelCapabilities
+  // reads for tool/vision/context) with the base models, then verify the
+  // effort-suffixed variants inherit those capabilities via the base fallback.
+  const modelsDevSync = await import("../../src/lib/modelsDevSync.ts");
+  modelsDevSync.saveModelsDevCapabilities({
+    "command-code": {
+      "deepseek/deepseek-v4-flash": {
+        tool_call: true,
+        reasoning: true,
+        attachment: false,
+        limit_context: 1000000,
+        limit_input: 1000000,
+        limit_output: 131072,
+      },
+      "meta/muse-spark-1.2-contributor": {
+        tool_call: true,
+        reasoning: true,
+        attachment: true,
+        limit_context: 1048576,
+        limit_input: 1048576,
+        limit_output: 1048576,
+      },
+    },
+  });
+
+  const cases: Array<[string, boolean]> = [
+    ["command-code/deepseek/deepseek-v4-flash-max", false], // text-only base
+    ["command-code/deepseek/deepseek-v4-flash-high", false],
+    ["command-code/meta/muse-spark-1.2-contributor-xhigh", true], // vision base
+    ["command-code/meta/muse-spark-1.2-contributor-high", true],
+  ];
+  for (const [modelId, vision] of cases) {
+    const caps = getResolvedModelCapabilities(modelId);
+    assert.equal(caps.provider, "command-code", `${modelId} provider`);
+    assert.equal(caps.supportsTools, true, `${modelId} must inherit tool support from base`);
+    assert.equal(caps.supportsVision, vision, `${modelId} must inherit vision from base`);
+    assert.equal(typeof caps.contextWindow, "number", `${modelId} must inherit a context window`);
+    assert.equal(caps.supportsThinking, true, `${modelId} must inherit reasoning from base`);
+  }
+});