feat(anthropic): add an opt-in thinking display policy per provider and model

docs-site/src/content/docs/reference/adapters.md

@@ -395,6 +395,11 @@ MiMo model Command Code serves.
   (`reasoning.summary: "none"`) keeps the provider default. Native Opus 5.5 on `api.anthropic.com`
   defaults to `display: "updates"`; explicit `reasoning.summary: "concise"` or `"detailed"`
   selects `"summarized"` without the updates beta. See [Opus 5.5 progress updates](/guides/providers/#claude-opus-55-progress-updates).
+  The provider or exact-model `thinkingDisplay` setting overrides that default:
+  `summarized` forces `display: "summarized"`, `updates` keeps progress commentary, and
+  `omitted` drops the display field so a Chat client receives no reasoning deltas during the
+  think. An explicit request summary still wins. See
+  [provider configuration](/reference/configuration/providers/).
 - **Structured output:** Responses `text.format` and Chat Completions `response_format` requests
   with `type: "json_schema"` become Anthropic `output_config.format`. The format merges into an
   existing adaptive-thinking output configuration, preserving a compatible `output_config.effort`.

docs-site/src/content/docs/reference/configuration/providers.md

@@ -388,6 +388,8 @@ Providers can expose a built-in shorthand, such as `agy` for `google-antigravity
 | `modelSuppressSyntheticMax?` | `Record<string, boolean>` | Catalog-only per-model switch. `true` prevents OpenCodex from adding a missing synthetic `max` rung while retaining a provider-declared `max` and continuing to add `ultra` to reasoning-capable ladders. A preserved row that already contains `max` keeps it until healthy discovery rebuilds that row, because the catalog does not persist whether an older `max` was synthetic. If a configured default names a suppressed missing `max`, the catalog falls back to the highest real rung below it. Codex uses the same ladder membership for the picker and explicit `spawn_agent` effort validation, so an explicit `max` spawn can fail client-side before proxy clamping; `ultra` remains the supported harness path. |
 | `modelSupportsReasoningSummaries?` | `Record<string, boolean>` | Set a model to `false` to stop advertising summaries and strip summary-delivery fields. |
 | `modelReasoningSummaryDelivery?` | `Record<string, "sequential" \| "sequential_cutoff" \| "concurrent" \| "concurrent_cutoff">` | Per-model Responses delivery enum; rewrites an existing delivery field. |
+| `thinkingDisplay?` | `"auto" \| "summarized" \| "updates" \| "omitted"` | Anthropic adaptive `thinking.display` policy for this provider. Unset and `auto` keep the built-in selection: native Claude Opus 5.5 on `api.anthropic.com` sends `updates` (progress commentary), and other adaptive models send `summarized`. `summarized` forces reasoning summaries for both. `updates` is honored only where the adapter already sent it — native Opus 5.5 — and is treated as `auto` for every other model. `omitted` sends adaptive thinking without a display field, matching hidden thinking; the stream then carries signature-only thinking blocks, so a Chat client sees only stream heartbeats during a long think. An explicit request `reasoning.summary` of `concise` or `detailed` still forces `summarized`. The dashboard provider editor and `PATCH /api/providers` accept `null` to clear it. `ocx config set providers.<name>.thinkingDisplay summarized` writes it directly. |
+| `modelThinkingDisplay?` | `Record<string, "auto" \| "summarized" \| "updates" \| "omitted">` | Per-model override of `thinkingDisplay`, keyed by the exact upstream model id (for example `claude-opus-5-5`). An entry always wins over the provider-wide value, including an explicit `auto` that returns one model to the built-in behavior; a model with no entry uses the provider value. The same value rules apply, including `updates` degrading to `auto` on models that never received that display. `PATCH /api/providers` merges entries and accepts `null` for one entry or the whole map to clear it. |
 | `modelAdapters?` | `Record<string, string>` | Per-model `openai-chat` or `openai-responses` wire override for mixed-wire gateways. Explicit entries beat registry defaults. The OpenCode Go preset selects Responses for `gpt-5.6-luna` while leaving sibling models on their documented wires; DeepSeek can select native Responses for `deepseek-v4-flash`; Alibaba Token Plan (Beijing) pins `qwen3.8-flash`, `qwen3.7-plus`, and `glm-5.3` to its native Responses API for Responses inbound only (live-verified on that gateway, including plaintext reasoning replay; Chat and Anthropic inbound stay on Chat, and the rest of the family can be opted in here); and GitHub Copilot declares Responses-only defaults for the following models (`gpt-5.3-codex`, `gpt-5.4`, `gpt-5.4-mini`, `gpt-5.5`, `gpt-5.6-luna`, `gpt-5.6-sol`, `gpt-5.6-terra`, `gpt-6.1-sol`, `gpt-6.1-terra`, `gpt-6-astra`, `grok-4.5`, `grok-4.6`, `mai-code-1.1-flash`, `mai-code-1-flash-picker`) because those models reject `/chat/completions` for agent traffic. Models without a built-in default (for example `gpt-5.4-nano`) can be opted in here. Single-wire upstream pins and canonical ChatGPT forward reject overrides. |
 | xAI Chat Completions (dashboard / CLI) | switch | Grok 4.5/4.6 OAuth Responses requests default to Responses. Existing Chat overrides are migrated once on upgrade; later Chat choices are preserved. Turn on to select Chat for both models, off to select Responses. CLI: `ocx provider edit xai --xai-chat on` or `--xai-chat off` (running proxy required). Mixed means only one model currently uses Chat. Other overrides and tier policy stay unchanged. API-key and translated Chat/Anthropic defaults are unchanged. Grok 4.7 defaults to Responses on OAuth through its registry wire default and can use Chat through an explicit `modelAdapters["grok-4.7"] = "openai-chat"` override. |
 | `xaiResponsesXSearch?` | `boolean` | Disabled by default. On an xAI Responses destination, append the provider-hosted `x_search` declaration only when a live `web_search` tool survives final request normalization. Existing declarations are not duplicated, caller `tool_choice`/`allowed_tools` selectors are never widened, and this is separate from the web-search sidecar's `search.xSearch` options. |

scripts/test-layout/layout.json

@@ -202,7 +202,7 @@
     "anthropic-empty-content.test.ts": "adapters/anthropic",
     "anthropic-eof-tolerance.test.ts": "adapters/anthropic",
     "anthropic-error-body.test.ts": "adapters/anthropic",
-    "anthropic-error-stop-reason.test.ts": "adapters/anthropic", "anthropic-thinking-usage.test.ts": "adapters/anthropic",
+    "anthropic-error-stop-reason.test.ts": "adapters/anthropic", "anthropic-thinking-usage.test.ts": "adapters/anthropic", "anthropic-thinking-display-policy.test.ts": "adapters/anthropic",
     "anthropic-fast-opt-in.test.ts": "adapters/anthropic",
     "anthropic-fast-speed.test.ts": "adapters/anthropic",
     "anthropic-hardening.test.ts": "adapters/anthropic",
@@ -777,7 +777,7 @@
     "config-non-object-backup.test.ts": "config",
     "config-ownership-uninstall.test.ts": "config", "config-ownership-concurrent-record.test.ts": "config", "config-ownership-removal.test.ts": "config",
     "config-rebase-provenance-writers.test.ts": "config",
-    "config-save-boundary.test.ts": "config",
+    "config-save-boundary.test.ts": "config", "config-thinking-display.test.ts": "config",
     "config-spend-ceilings.test.ts": "config",
     "config-spend-pool-collision.test.ts": "config",
     "config-user-edits.test.ts": "config", "hub-disable-live-save.test.ts": "config",
@@ -1217,7 +1217,7 @@
     "management-provider-reset-replay.test.ts": "server",
     "management-provider-synthetic-max.test.ts": "server",
     "management-provider-upstream-websocket.test.ts": "server",
-    "management-provider-context-inference.test.ts": "server", "management-provider-validation.test.ts": "server",
+    "management-provider-context-inference.test.ts": "server", "management-provider-validation.test.ts": "server", "management-provider-thinking-display.test.ts": "server",
     "management-provider-verbosity.test.ts": "server",
     "management-route-registry.test.ts": "server",
     "protocol-routes.test.ts": "server",

src/adapters/anthropic.ts

@@ -1109,6 +1109,18 @@ export function createAnthropicAdapter(provider: OcxProviderConfig, cacheRetenti
       const summary = (parsed._rawBody as { reasoning?: { summary?: unknown } } | undefined)?.reasoning?.summary;
       // Codex sends "auto" by default, so only an explicit concise/detailed request opts out of updates (#7728).
       const summaryRequested = summary === "concise" || summary === "detailed";
+      // Operator policy for the adaptive `thinking.display` field (#8363). Unset and "auto" keep
+      // the built-in selection. Any model entry wins over the provider-wide value, including an
+      // explicit "auto" that opts one model back out. "updates" is only sendable where this
+      // adapter already sent it (native Opus 5.5), so anywhere else it is treated as auto rather
+      // than inventing a display value the model never received.
+      const modelThinkingDisplay = modelRecordValue(provider.modelThinkingDisplay, parsed.modelId);
+      const configuredThinkingDisplay = modelThinkingDisplay !== undefined
+        ? modelThinkingDisplay
+        : provider.thinkingDisplay;
+      const thinkingDisplayPolicy = configuredThinkingDisplay === "updates" && !nativeOpus55
+        ? "auto"
+        : configuredThinkingDisplay ?? "auto";
       const thinkingToggleWire = declaresThinkingToggleWire(provider, parsed.modelId);
       const reasoningControlsUnsupported = reasoningControlsAreUnsupported(provider, parsed.modelId);
       const capabilities = resolveReasoningCapabilities(provider, parsed.modelId);
@@ -1145,7 +1157,16 @@ export function createAnthropicAdapter(provider: OcxProviderConfig, cacheRetenti
           // thinking blocks, so a Chat client sees minutes of heartbeats and no reasoning delta
           // during a long think (#5824). Native Opus 5.5 selects its display below;
           // other models request summarized thinking unless the caller hides it.
-          body.thinking = parsed.options.hideThinkingSummary
+          // Under the default policy this stays the pre-existing selection, including its local
+          // hideThinkingSummary behavior; a configured "summarized" overrides that hiding.
+          // An explicit request summary forces summarized on every adaptive model (#7881), so it
+          // also outranks a configured "omitted". "updates" already degraded to auto above
+          // because only native Opus 5.5 can send it.
+          // Local hideThinkingSummary only speaks for the built-in policy, where it keeps its
+          // pre-existing meaning; an operator choice replaces that local preference.
+          const hiddenThinkingDisplay = (thinkingDisplayPolicy === "omitted" && !summaryRequested)
+            || (thinkingDisplayPolicy === "auto" && parsed.options.hideThinkingSummary);
+          body.thinking = hiddenThinkingDisplay
             ? { type: "adaptive" }
             : { type: "adaptive", display: "summarized" };
           const effort = adaptiveEffort(effectiveReasoning);
@@ -1193,7 +1214,12 @@ export function createAnthropicAdapter(provider: OcxProviderConfig, cacheRetenti
       }
 
       if (nativeOpus55 && !reasoningControlsUnsupported && thinkingEnabled !== false) {
-        body.thinking = { type: "adaptive", display: summaryRequested ? "summarized" : "updates" };
+        // An explicit per-request summary (concise/detailed) still forces summarized (#7881).
+        body.thinking = summaryRequested || thinkingDisplayPolicy === "summarized"
+          ? { type: "adaptive", display: "summarized" }
+          : thinkingDisplayPolicy === "omitted"
+            ? { type: "adaptive" }
+            : { type: "adaptive", display: "updates" };
       }
       // One discriminator governs the display beta and response classification, including reuse.
       progressUpdates = (body.thinking as { display?: string } | undefined)?.display === "updates";

src/config/provider-validation.ts

@@ -13,6 +13,7 @@ import {
   isWirePinnedModel,
   MODEL_ADAPTER_OVERRIDE_ALLOWED,
   REASONING_SUMMARY_DELIVERY_VALUES,
+  THINKING_DISPLAY_VALUES,
   UPSTREAM_HTTP_VERSION_VALUES,
   type OcxProviderConfig,
   type ModelCapabilities,
@@ -29,6 +30,7 @@ const SENSITIVE_PROVIDER_HEADERS = new Set([
   "x-amz-security-token",
 ]);
 const REASONING_SUMMARY_DELIVERY_SET = new Set<string>(REASONING_SUMMARY_DELIVERY_VALUES);
+const THINKING_DISPLAY_SET = new Set<string>(THINKING_DISPLAY_VALUES);
 const DISPLAY_NAME_CONTROL_CHARS = /[\u0000-\u001f\u007f-\u009f\u2028\u2029]/;
 const MAX_MODEL_DISPLAY_NAME_LENGTH = 128;
 
@@ -385,6 +387,33 @@ export function reasoningSummaryDeliveryRecordConfigError(
   return null;
 }
 
+/** Validate `thinkingDisplay`: one closed vocabulary shared by config and management writes. */
+export function thinkingDisplayConfigError(value: unknown, field = "thinkingDisplay"): string | null {
+  if (value === undefined) return null;
+  if (typeof value !== "string" || !THINKING_DISPLAY_SET.has(value)) {
+    return `${field} must be one of: ${THINKING_DISPLAY_VALUES.join(", ")}`;
+  }
+  return null;
+}
+
+/** Validate `modelThinkingDisplay`: a plain map of model id to the same closed vocabulary. */
+export function modelThinkingDisplayConfigError(value: unknown, field = "modelThinkingDisplay"): string | null {
+  if (value === undefined) return null;
+  if (!value || typeof value !== "object" || Array.isArray(value)) return `${field} must be a plain object`;
+  const prototype = Object.getPrototypeOf(value);
+  if (prototype !== Object.prototype && prototype !== null) return `${field} must be a plain object with own properties`;
+  for (const [key, entry] of Object.entries(value)) {
+    const safeKey = JSON.stringify(redactSecretString(key));
+    if (!key.trim()) return `${field} keys must be nonblank model ids`;
+    if (["__proto__", "prototype", "constructor"].includes(key)) {
+      return `${field} keys must not be reserved object keys`;
+    }
+    const entryError = thinkingDisplayConfigError(entry, `${field}.${safeKey}`);
+    if (entryError) return entryError;
+  }
+  return null;
+}
+
 /** Validate a provider's per-model wire override map against runtime routing rules. */
 export function modelAdapterRecordConfigError(
   value: unknown,

src/config/schema/config-schema.ts

@@ -60,6 +60,8 @@ import {
   providerHeadersConfigError,
   providerForwardClientHeadersConfigError,
   reasoningSummaryDeliveryRecordConfigError,
+  thinkingDisplayConfigError,
+  modelThinkingDisplayConfigError,
   credentialPinnedProviderConfigError,
 } from "../provider-validation";
 import {
@@ -767,6 +769,26 @@ export const configSchema = z.object({
         message: reasoningSummaryDeliveryError,
       });
     }
+    const thinkingDisplayError = thinkingDisplayConfigError(
+      (provider as { thinkingDisplay?: unknown }).thinkingDisplay,
+    );
+    if (thinkingDisplayError) {
+      ctx.addIssue({
+        code: "custom",
+        path: ["providers", redactSecretString(name), "thinkingDisplay"],
+        message: thinkingDisplayError,
+      });
+    }
+    const modelThinkingDisplayError = modelThinkingDisplayConfigError(
+      (provider as { modelThinkingDisplay?: unknown }).modelThinkingDisplay,
+    );
+    if (modelThinkingDisplayError) {
+      ctx.addIssue({
+        code: "custom",
+        path: ["providers", redactSecretString(name), "modelThinkingDisplay"],
+        message: modelThinkingDisplayError,
+      });
+    }
     const minMaxOutputError = positiveIntegerConfigError(
       (provider as { minMaxOutputTokens?: unknown }).minMaxOutputTokens,
       "minMaxOutputTokens",

src/config/schema/leaf-validators.ts

@@ -29,6 +29,7 @@ import {
   MODEL_ADAPTER_OVERRIDE_ALLOWED,
   pinnedWireAdapter,
   PROVIDER_WEB_SEARCH_BRIDGE_BACKENDS,
+  THINKING_DISPLAY_VALUES,
   UPSTREAM_HTTP_VERSION_VALUES,
   type OcxProviderConfig,
   type OcxConfig,
@@ -360,6 +361,10 @@ export const providerConfigSchema = z.object({
   inferContextWindow: z.boolean().optional(),
   modelSupportsServiceTier: z.record(z.string().min(1), z.boolean()).optional(),
   modelSuppressSyntheticMax: z.record(z.string().min(1), z.boolean()).optional(),
+  // Validated against a closed vocabulary here as well as at the write boundaries, so a
+  // hand-edited typo cannot load silently and leave the operator with the built-in display (#8363).
+  thinkingDisplay: z.enum(THINKING_DISPLAY_VALUES).optional(),
+  modelThinkingDisplay: z.record(z.string().min(1), z.enum(THINKING_DISPLAY_VALUES)).optional(),
   promoteAdditionalTools: z.boolean().optional(),
   preserveResponsesReasoningContent: z.boolean().optional(),
   preserveResponsesInputItemIds: z.boolean().optional(),

src/providers/model-rename-fields.ts

@@ -103,6 +103,8 @@ export const PROVIDER_MODEL_RENAME_ROLES = {
   modelSupportsVerbosity: "record",
   supportsVerbosity: "none",
   modelReasoningSummaryDelivery: "record",
+  thinkingDisplay: "none",
+  modelThinkingDisplay: "record",
   modelPreferHostedTools: "record",
   supportsOpenAiWebSearchToolFields: "none",
   xaiResponsesXSearch: "none",

src/providers/resolved-model-policy.ts

@@ -43,6 +43,7 @@ export type StaticProviderPolicyField =
   | "reasoningDetailsModels" | "thinkingToggleModels" | "thinkingBudgetModels" | "showThinkingSummary"
   | "escapeBuiltinToolNames" | "googleMode" | "project" | "location" | "modelCapabilities"
   | "modelAutoCompactTokenLimits" | "modelSuppressSyntheticMax" | "modelReasoningSummaryDelivery"
+  | "thinkingDisplay" | "modelThinkingDisplay"
   | "codexToolMode" | "modelAdapters";
 type StaticProviderPolicyShape = Pick<OcxProviderConfig, StaticProviderPolicyField>;
 export type ResolvedProviderStaticPolicy = Readonly<Partial<StaticProviderPolicyShape>> & Readonly<{
@@ -182,6 +183,8 @@ export function resolveModelPolicy(input: ResolveModelPolicyInput): ResolvedMode
     "supportsVerbosity", "responsesItemIdRepair",
     "showThinkingSummary", "escapeBuiltinToolNames", "googleMode", "project", "location",
   ] as const) putScalar(key, entry?.[key] as StaticProviderPolicyShape[typeof key] | undefined);
+  // Operator-only display policy; no registry entry declares it (#8363).
+  put("thinkingDisplay", provider.thinkingDisplay, provider.thinkingDisplay !== undefined ? "operator" : "unknown");
   const legacyClinePassLadder = entry !== undefined && input.providerName === "cline-pass"
     && provider.reasoningWireFormat === "gateway-object"
     && sameStringArray(provider.reasoningEfforts, ["low"]);
@@ -218,6 +221,7 @@ export function resolveModelPolicy(input: ResolveModelPolicyInput): ResolvedMode
   for (const key of [
     "modelAutoCompactTokenLimits", "modelSuppressSyntheticMax",
     "modelReasoningSummaryDelivery", "codexToolMode", "modelAdapters",
+    "modelThinkingDisplay",
   ] as const) putScalar(key, undefined);
   put("modelCapabilities", mergeReasoningModelCapabilities(entry?.modelCapabilities, provider.modelCapabilities),
     provider.modelCapabilities !== undefined ? "operator-capability" : entry?.modelCapabilities !== undefined ? "registry" : "unknown");

src/server/auth-cors.ts

@@ -29,6 +29,8 @@ import {
   providerHeadersConfigError,
   providerForwardClientHeadersConfigError,
   reasoningSummaryDeliveryRecordConfigError,
+  thinkingDisplayConfigError,
+  modelThinkingDisplayConfigError,
   upstreamHttpVersionConfigError,
   credentialPinnedProviderConfigError,
 } from "../config/provider-validation";
@@ -949,6 +951,10 @@ export function providerManagementConfigError(
     raw.modelSupportsReasoningSummaries,
   );
   if (reasoningSummaryDeliveryError) return `provider ${name} ${reasoningSummaryDeliveryError}`;
+  const thinkingDisplayError = thinkingDisplayConfigError(raw.thinkingDisplay);
+  if (thinkingDisplayError) return `provider ${name} ${thinkingDisplayError}`;
+  const modelThinkingDisplayError = modelThinkingDisplayConfigError(raw.modelThinkingDisplay);
+  if (modelThinkingDisplayError) return `provider ${name} ${modelThinkingDisplayError}`;
   const modelAdaptersError = modelAdapterRecordConfigError(raw.modelAdapters, "modelAdapters", name, typed);
   if (modelAdaptersError) return `provider ${name} ${modelAdaptersError}`;
   const credentialPinError = credentialPinnedProviderConfigError(name, typed);
@@ -1175,6 +1181,8 @@ const PROVIDER_CONFIG_FIELD_POLICY = {
   modelSupportsVerbosity: "editor",
   supportsVerbosity: "editor",
   modelReasoningSummaryDelivery: "editor",
+  thinkingDisplay: "editor",
+  modelThinkingDisplay: "editor",
   modelPreferHostedTools: "editor",
   supportsOpenAiWebSearchToolFields: "editor",
   xaiResponsesXSearch: "editor",