board #11409

open-sse/config/opencodeZenGoSharedModels.ts

@@ -12,5 +12,15 @@
 export const OPENCODE_ZEN_GO_SHARED_MODELS = Object.freeze([
   { id: "kimi-k2.7-code", name: "Kimi K2.7 Code" },
   { id: "qwen3.5-plus", name: "Qwen3.5 Plus", targetFormat: "claude", supportsVision: false },
-  { id: "qwen3.6-plus", name: "Qwen3.6 Plus", targetFormat: "claude", supportsVision: false },
+  {
+    id: "qwen3.6-plus",
+    name: "Qwen3.6 Plus",
+    targetFormat: "claude",
+    supportsVision: false,
+    // #10788: effort-tier aliases exist as explicit registry rows; declare the
+    // vocabulary on the shared base row so variant synthesis and the sanitizer
+    // agree on it for both opencode-go and opencode-zen.
+    supportsReasoning: true,
+    supportedThinkingEfforts: ["high", "max"],
+  },
 ]);

open-sse/config/providers/registry/nvidia/index.ts

@@ -18,7 +18,17 @@ export const nvidiaProvider: RegistryEntry = {
   passthroughModels: true,
   models: [
     // #6108: z-ai/glm-5.1 EOL'd 2026-07-02 (direct probe returns 410) — dropped.
-    { id: "z-ai/glm-5.2", name: "GLM 5.2" },
+    // #10788: NVIDIA's hosted GLM-5.2 exposes a BINARY thinking switch
+    // (chat_template_kwargs.enable_thinking), not effort tiers — see
+    // mapNvidiaGlm52ReasoningParams. Declaring an empty tier list keeps the
+    // catalog from synthesizing unresolvable -low/-high/-max variant ids while
+    // still marking the model reasoning-capable.
+    {
+      id: "z-ai/glm-5.2",
+      name: "GLM 5.2",
+      supportsReasoning: true,
+      supportedThinkingEfforts: [],
+    },
     // #3329/#6108: minimaxai/minimax-m3 stays excluded from the nvidia tier — it
     // still 404s here for most callers; the single 200 probe in #6108 was not
     // reproducible enough to override the #3329 guard. Re-add only once NVIDIA

open-sse/config/providers/registry/opencode/go/index.ts

@@ -19,9 +19,15 @@ export const opencode_goProvider: RegistryEntry = {
     // `kimi-k2.7-code` (the live API rejects the plain `kimi-k2.7` alias for
     // `/chat/completions`, even though the docs config example uses it).
     // GLM-5.2 — base model + effort-tier aliases (#6922).
-    // OpencodeExecutor rewrites the alias to the canonical id and injects
-    // reasoning_effort, mirroring the deepseek-v4-pro-* pattern.
-    { id: "glm-5.2", name: "GLM-5.2", supportsReasoning: true },
+    // #10788: the tier vocabulary is declared on the base row so the catalog's
+    // variant synthesis (#9485) and the effort sanitizer share one source of
+    // truth with OpencodeExecutor's EFFORT_TIERS.
+    {
+      id: "glm-5.2",
+      name: "GLM-5.2",
+      supportsReasoning: true,
+      supportedThinkingEfforts: ["high", "max"],
+    },
     { id: "glm-5.2-high", name: "GLM-5.2 (high effort)", supportsReasoning: true },
     { id: "glm-5.2-max", name: "GLM-5.2 (max effort)", supportsReasoning: true },
 
@@ -34,11 +40,16 @@ export const opencode_goProvider: RegistryEntry = {
     { id: "kimi-k2.6", name: "Kimi K2.6" },
     { id: "kimi-k2.5", name: "Kimi K2.5" },
     // #8353: Kimi K3 base + max-effort alias from the OpenCode Go registry.
-    { id: "kimi-k3", name: "Kimi K3", supportsReasoning: true },
+    { id: "kimi-k3", name: "Kimi K3", supportsReasoning: true, supportedThinkingEfforts: ["max"] },
     { id: "kimi-k3-max", name: "Kimi K3 (max effort)", supportsReasoning: true },
     // MiMo-V2.5 — base model + effort-tier aliases (#6922).
     { id: "mimo-v2.5-pro", name: "MiMo-V2.5-Pro", supportsReasoning: true },
-    { id: "mimo-v2.5", name: "MiMo-V2.5", supportsReasoning: true },
+    {
+      id: "mimo-v2.5",
+      name: "MiMo-V2.5",
+      supportsReasoning: true,
+      supportedThinkingEfforts: ["high", "max"],
+    },
     { id: "mimo-v2.5-high", name: "MiMo-V2.5 (high effort)", supportsReasoning: true },
     { id: "mimo-v2.5-max", name: "MiMo-V2.5 (max effort)", supportsReasoning: true },
     // #3110: MiniMax M3 via OpenCode Go tier
@@ -59,7 +70,14 @@ export const opencode_goProvider: RegistryEntry = {
     // so combo routing skips them when the request contains image blocks,
     // preventing image content from reaching a vision-incapable upstream.
     // #8353: effort-tier aliases from the OpenCode Go registry.
-    { id: "qwen3.7-max", name: "Qwen3.7 Max", targetFormat: "claude", supportsVision: false },
+    {
+      id: "qwen3.7-max",
+      name: "Qwen3.7 Max",
+      targetFormat: "claude",
+      supportsVision: false,
+      supportsReasoning: true,
+      supportedThinkingEfforts: ["high", "max"],
+    },
     {
       id: "qwen3.7-max-high",
       name: "Qwen3.7 Max (high effort)",
@@ -79,6 +97,8 @@ export const opencode_goProvider: RegistryEntry = {
       name: "Qwen3.7 Plus",
       targetFormat: "claude",
       supportsVision: false,
+      supportsReasoning: true,
+      supportedThinkingEfforts: ["high", "max"],
     },
     {
       id: "qwen3.7-plus-high",
@@ -111,7 +131,13 @@ export const opencode_goProvider: RegistryEntry = {
       supportsReasoning: true,
     },
     // #8353: hy3 is the Go-tier base id (distinct from hy3-preview / hy3-free).
-    { id: "hy3", name: "Hunyuan3", contextLength: 256000, supportsReasoning: true },
+    {
+      id: "hy3",
+      name: "Hunyuan3",
+      contextLength: 256000,
+      supportsReasoning: true,
+      supportedThinkingEfforts: ["none", "low", "high"],
+    },
     {
       id: "hy3-none",
       name: "Hunyuan3 (none effort)",
@@ -201,7 +227,12 @@ export const opencode_goProvider: RegistryEntry = {
       targetFormat: "openai-responses",
     },
     // #8353: Grok 4.5 + effort tiers from the OpenCode Go registry.
-    { id: "grok-4.5", name: "Grok 4.5", supportsReasoning: true },
+    {
+      id: "grok-4.5",
+      name: "Grok 4.5",
+      supportsReasoning: true,
+      supportedThinkingEfforts: ["low", "medium", "high"],
+    },
     { id: "grok-4.5-low", name: "Grok 4.5 (low effort)", supportsReasoning: true },
     { id: "grok-4.5-medium", name: "Grok 4.5 (medium effort)", supportsReasoning: true },
     { id: "grok-4.5-high", name: "Grok 4.5 (high effort)", supportsReasoning: true },

open-sse/config/providers/registry/opencode/zen/index.ts

@@ -79,18 +79,35 @@ export const opencode_zenProvider: RegistryEntry = {
     },
 
     // ── DeepSeek ────────────────────────────────────────────────
-    { id: "deepseek-v4-pro", name: "DeepSeek V4 Pro" },
-    { id: "deepseek-v4-flash", name: "DeepSeek V4 Flash" },
+    // #10788: same tier vocabulary as opencode-go's DeepSeek rows — the Zen
+    // upstream accepts the identical effort set on these models.
+    {
+      id: "deepseek-v4-pro",
+      name: "DeepSeek V4 Pro",
+      supportsReasoning: true,
+      supportedThinkingEfforts: ["none", "low", "high", "max"],
+    },
+    {
+      id: "deepseek-v4-flash",
+      name: "DeepSeek V4 Flash",
+      supportsReasoning: true,
+      supportedThinkingEfforts: ["none", "low", "high", "max"],
+    },
 
     // ── GLM / Z.AI ─────────────────────────────────────────────
-    { id: "glm-5.2", name: "GLM-5.2" },
+    {
+      id: "glm-5.2",
+      name: "GLM-5.2",
+      supportsReasoning: true,
+      supportedThinkingEfforts: ["high", "max"],
+    },
 
     // ── MiniMax ────────────────────────────────────────────────
     // #3110: MiniMax M3 — frontier coding model with 1M context
     { id: "minimax-m3", name: "MiniMax M3", contextLength: 1048576, supportsVision: true },
 
     // ── Kimi / Moonshot ────────────────────────────────────────
-    { id: "kimi-k3", name: "Kimi K3" },
+    { id: "kimi-k3", name: "Kimi K3", supportsReasoning: true, supportedThinkingEfforts: ["max"] },
     // kimi-k2.7-code declared identically on opencode-go — see OPENCODE_ZEN_GO_SHARED_MODELS.
 
     // ── Qwen ───────────────────────────────────────────────────

open-sse/executors/opencode.ts

@@ -895,10 +895,21 @@ export class OpencodeExecutor extends BaseExecutor {
       const mb = modifiedBody as Record<string, unknown>;
       const parsed = parseEffortLevel(model);
       if (parsed) {
-        mb.model = parsed.baseModel;
-        if (mb.reasoning_effort === undefined) {
-          mb.reasoning_effort = parsed.effort;
+        const deepseekFamily =
+          parsed.baseModel === "deepseek-v4-pro" || parsed.baseModel === "deepseek-v4-flash";
+        if (deepseekFamily) {
+          // DeepSeek via opencode-go proxies the native DeepSeek contract, which
+          // accepts a flat reasoning_effort field (#4647).
+          mb.model = parsed.baseModel;
+          if (mb.reasoning_effort === undefined) {
+            mb.reasoning_effort = parsed.effort;
+          }
         }
+        // #10788: every other family's ONLY native effort mechanism is the
+        // -<tier> suffix in the model id itself (the ids `opencode models
+        // opencode-go --verbose` lists). The opencode-go ChatCompletionRequest
+        // carries no flat reasoning_effort field, so rewriting to the base id
+        // silently dropped the tier — forward the aliased id verbatim instead.
       }
     }
     // #1543 / upstream PR #1099: thinking-mode upstreams routed through OpenCode

tests/unit/opencode-go-effort-aliases-6922.test.ts

@@ -96,35 +96,43 @@ test("#6922 parseEffortLevel: base model without tier → null", () => {
   assert.strictEqual(parseEffortLevel("glm-5.2"), null);
 });
 
-// ─── transformRequest: end-to-end model-id rewrite + reasoning_effort inject ──
+// ─── transformRequest: #6922 wiring, updated by #10788 ─────────────────────
 //
-// parseEffortLevel is a pure function, but the actual bug (#6922) surfaces
-// through OpencodeExecutor.transformRequest — the caller that rewrites the
-// outbound model id and injects reasoning_effort. These tests exercise that
-// public entry point directly so a broken wiring (e.g. parseEffortLevel
-// correct but never called, or its result dropped) would fail here even if
-// the parseEffortLevel-only tests above stayed green.
+// parseEffortLevel is a pure function, but the original bug (#6922) surfaces
+// through OpencodeExecutor.transformRequest. Since #10788, glm-5.2 / mimo-v2.5
+// (non-DeepSeek families) forward the effort-suffixed alias VERBATIM — the
+// suffix is their only native effort mechanism and opencode-go has no flat
+// reasoning_effort field to receive a rewritten tier. Only DeepSeek V4 keeps
+// the base-rewrite + field-injection contract.
 
 const CREDENTIALS = { apiKey: "k" } as Record<string, unknown>;
 
-test("#6922 transformRequest: glm-5.2-high → model rewritten to glm-5.2, reasoning_effort injected", () => {
+test("#6922/#10788 transformRequest: glm-5.2-high forwards the alias verbatim", () => {
   const executor = new OpencodeExecutor("opencode-go");
   const body = { model: "glm-5.2-high", messages: [{ role: "user", content: "hi" }] };
 
   const out = executor.transformRequest("glm-5.2-high", body, true, CREDENTIALS);
 
-  assert.equal(out.model, "glm-5.2", "model id must be rewritten to the base id");
-  assert.equal(out.reasoning_effort, "high", "reasoning_effort must be injected from the alias");
+  assert.equal(out.model, "glm-5.2-high", "alias id must reach the wire untouched");
+  assert.equal(
+    out.reasoning_effort,
+    undefined,
+    "no flat reasoning_effort may be injected for non-DeepSeek families"
+  );
 });
 
-test("#6922 transformRequest: mimo-v2.5-max → model rewritten to mimo-v2.5, reasoning_effort injected", () => {
+test("#6922/#10788 transformRequest: mimo-v2.5-max forwards the alias verbatim", () => {
   const executor = new OpencodeExecutor("opencode-go");
   const body = { model: "mimo-v2.5-max", messages: [{ role: "user", content: "hi" }] };
 
   const out = executor.transformRequest("mimo-v2.5-max", body, true, CREDENTIALS);
 
-  assert.equal(out.model, "mimo-v2.5", "model id must be rewritten to the base id");
-  assert.equal(out.reasoning_effort, "max", "reasoning_effort must be injected from the alias");
+  assert.equal(out.model, "mimo-v2.5-max", "alias id must reach the wire untouched");
+  assert.equal(
+    out.reasoning_effort,
+    undefined,
+    "no flat reasoning_effort may be injected for non-DeepSeek families"
+  );
 });
 
 test("#6922 transformRequest: does not clobber an already-set reasoning_effort", () => {
@@ -137,7 +145,7 @@ test("#6922 transformRequest: does not clobber an already-set reasoning_effort",
 
   const out = executor.transformRequest("glm-5.2-high", body, true, CREDENTIALS);
 
-  assert.equal(out.model, "glm-5.2", "model id is still rewritten to the base id");
+  assert.equal(out.model, "glm-5.2-high", "non-DeepSeek alias id is left untouched");
   assert.equal(
     out.reasoning_effort,
     "caller-supplied",

tests/unit/opencode-go-effort-aliases-8353.test.ts

@@ -27,7 +27,15 @@ const { parseEffortLevel, OpencodeExecutor } =
 const { REGISTRY } = (await import("../../open-sse/config/providerRegistry.ts")) as {
   REGISTRY: Record<
     string,
-    { models?: Array<{ id: string; name?: string; targetFormat?: string }> }
+    {
+      models?: Array<{
+        id: string;
+        name?: string;
+        targetFormat?: string;
+        supportsReasoning?: boolean;
+        supportedThinkingEfforts?: string[];
+      }>;
+    }
   >;
 };
 
@@ -189,29 +197,46 @@ test("#8353 parseEffortLevel: MiniMax M3 has no effort-tier aliases", () => {
 
 const CREDENTIALS = { apiKey: "k" } as Record<string, unknown>;
 
+// #10788: DeepSeek V4 keeps the base-rewrite + reasoning_effort injection (the
+// opencode-go DeepSeek contract accepts the flat field); every other family
+// must receive the effort-suffixed alias VERBATIM, because the suffix is their
+// only native effort mechanism and opencode-go has no flat reasoning_effort.
 const TRANSFORM_SAMPLES = [
-  { alias: "deepseek-v4-flash-low", base: "deepseek-v4-flash", effort: "low" },
-  { alias: "grok-4.5-medium", base: "grok-4.5", effort: "medium" },
-  { alias: "hy3-none", base: "hy3", effort: "none" },
-  { alias: "kimi-k3-max", base: "kimi-k3", effort: "max" },
-  { alias: "qwen3.7-plus-max", base: "qwen3.7-plus", effort: "max" },
-  { alias: "qwen3.7-max-high", base: "qwen3.7-max", effort: "high" },
+  {
+    alias: "deepseek-v4-flash-low",
+    wireModel: "deepseek-v4-flash",
+    effort: "low" as const,
+    note: "DeepSeek rewrites to base + reasoning_effort (#4647)",
+  },
+  { alias: "grok-4.5-medium", wireModel: "grok-4.5-medium", effort: null },
+  { alias: "hy3-none", wireModel: "hy3-none", effort: null },
+  { alias: "kimi-k3-max", wireModel: "kimi-k3-max", effort: null },
+  { alias: "qwen3.7-plus-max", wireModel: "qwen3.7-plus-max", effort: null },
+  { alias: "qwen3.7-max-high", wireModel: "qwen3.7-max-high", effort: null },
   {
     alias: "muse-spark-1.2-contributor-xhigh",
-    base: "muse-spark-1.2-contributor",
-    effort: "xhigh",
+    wireModel: "muse-spark-1.2-contributor-xhigh",
+    effort: null,
   },
 ] as const;
 
-for (const { alias, base, effort } of TRANSFORM_SAMPLES) {
-  test(`#8353 transformRequest: ${alias} → model=${base}, reasoning_effort=${effort}`, () => {
+for (const { alias, wireModel, effort, note } of TRANSFORM_SAMPLES) {
+  test(`#8353/#10788 transformRequest: ${alias} → model=${wireModel}`, () => {
     const executor = new OpencodeExecutor("opencode-go");
     const body = { model: alias, messages: [{ role: "user", content: "hi" }] };
 
     const out = executor.transformRequest(alias, body, true, CREDENTIALS);
 
-    assert.equal(out.model, base, "model id must be rewritten to the base id");
-    assert.equal(out.reasoning_effort, effort, "reasoning_effort must be injected from the alias");
+    assert.equal(out.model, wireModel);
+    if (effort === null) {
+      assert.equal(
+        out.reasoning_effort,
+        undefined,
+        "non-DeepSeek families must not receive a flat reasoning_effort field"
+      );
+    } else {
+      assert.equal(out.reasoning_effort, effort, note);
+    }
   });
 }
 
@@ -228,3 +253,42 @@ test("#8353 transformRequest: does not clobber an already-set reasoning_effort",
   assert.equal(out.model, "deepseek-v4-flash");
   assert.equal(out.reasoning_effort, "caller-supplied");
 });
+
+// ─── #10788: base-model tier declarations match the executor vocabulary ────
+
+test("#10788 registry base rows declare the same tiers EFFORT_TIERS parses", () => {
+  const expectedTiers: Record<string, string[]> = {
+    "glm-5.2": ["high", "max"],
+    "mimo-v2.5": ["high", "max"],
+    "grok-4.5": ["low", "medium", "high"],
+    hy3: ["none", "low", "high"],
+    "kimi-k3": ["max"],
+    "qwen3.6-plus": ["high", "max"],
+    "qwen3.7-max": ["high", "max"],
+    "qwen3.7-plus": ["high", "max"],
+  };
+  for (const providerId of ["opencode-go", "opencode-zen"]) {
+    const entry = REGISTRY[providerId];
+    assert.ok(entry?.models, `${providerId} must expose models`);
+    for (const [base, tiers] of Object.entries(expectedTiers)) {
+      if (providerId === "opencode-zen" && !entry.models.some((m) => m.id === base)) continue;
+      const row = entry.models.find((m) => m.id === base);
+      assert.ok(row, `${providerId} must declare base model ${base}`);
+      assert.ok(row.supportsReasoning, `${providerId}/${base} must be reasoning-capable`);
+      assert.deepEqual(
+        [...(row.supportedThinkingEfforts ?? [])].sort(),
+        [...tiers].sort(),
+        `${providerId}/${base} tier vocabulary must match EFFORT_TIERS`
+      );
+    }
+  }
+});
+
+test("#10788 nvidia z-ai/glm-5.2 declares reasoning with an empty tier list (binary switch)", () => {
+  const entry = REGISTRY["nvidia"];
+  assert.ok(entry?.models, "nvidia must expose models");
+  const row = entry.models.find((m) => m.id === "z-ai/glm-5.2");
+  assert.ok(row, "nvidia z-ai/glm-5.2 must exist");
+  assert.equal(row.supportsReasoning, true);
+  assert.deepEqual(row.supportedThinkingEfforts, []);
+});