feat: opt-in Claude 1M beta header with safe 200k fallback

README.md

@@ -120,8 +120,8 @@ opencode run "Hello" --model=google/antigravity-claude-opus-4-6-thinking --varia
 | `antigravity-gemini-3-pro` | low, high | Gemini 3 Pro with thinking |
 | `antigravity-gemini-3.1-pro` | low, high | Gemini 3.1 Pro with thinking (rollout-dependent) |
 | `antigravity-gemini-3-flash` | minimal, low, medium, high | Gemini 3 Flash with thinking |
-| `antigravity-claude-sonnet-4-6` | — | Claude Sonnet 4.6 |
-| `antigravity-claude-opus-4-6-thinking` | low, max | Claude Opus 4.6 with extended thinking |
+| `antigravity-claude-sonnet-4-6` | — | Claude Sonnet 4.6 (200k base context) |
+| `antigravity-claude-opus-4-6-thinking` | low, max | Claude Opus 4.6 with extended thinking (200k base context) |
 
 **Gemini CLI quota** (separate from Antigravity; used when `cli_first` is true or as fallback):
 
@@ -140,6 +140,7 @@ opencode run "Hello" --model=google/antigravity-claude-opus-4-6-thinking --varia
 > - When a Gemini quota pool is exhausted, the plugin automatically falls back to the other pool.
 > - Claude and image models always use Antigravity.
 > Model names are automatically transformed for the target API (e.g., `antigravity-gemini-3-flash` → `gemini-3-flash-preview` for CLI).
+> - Claude Antigravity limits are 200k by default. Experimental opt-in long-context beta header attempts are available via `antigravity.json` (`claude_long_context_beta`) and automatically fall back to stable 200k if rejected.
 
 **Using variants:**
 ```bash
@@ -190,12 +191,12 @@ Add this to your `~/.config/opencode/opencode.json`:
           }
         },
         "antigravity-claude-sonnet-4-6": {
-          "name": "Claude Sonnet 4.6 (Antigravity)",
+          "name": "Claude Sonnet 4.6 (Antigravity, 200k base)",
           "limit": { "context": 200000, "output": 64000 },
           "modalities": { "input": ["text", "image", "pdf"], "output": ["text"] }
         },
         "antigravity-claude-opus-4-6-thinking": {
-          "name": "Claude Opus 4.6 Thinking (Antigravity)",
+          "name": "Claude Opus 4.6 Thinking (Antigravity, 200k base)",
           "limit": { "context": 200000, "output": 64000 },
           "modalities": { "input": ["text", "image", "pdf"], "output": ["text"] },
           "variants": {

assets/antigravity.schema.json

@@ -120,6 +120,17 @@
       "type": "boolean",
       "description": "Enable Claude prompt auto-caching by adding top-level cache_control when absent."
     },
+    "claude_long_context_beta": {
+      "default": false,
+      "type": "boolean",
+      "description": "Enable experimental Claude long-context beta header injection for Claude 4.6 models. If rejected, requests auto-fallback to stable 200k behavior."
+    },
+    "claude_long_context_beta_header": {
+      "default": "context-1m-2025-08-07",
+      "type": "string",
+      "minLength": 1,
+      "description": "Claude long-context beta header value. Override if provider beta token changes."
+    },
     "proactive_token_refresh": {
       "default": true,
       "type": "boolean",

docs/CONFIGURATION.md

@@ -35,6 +35,8 @@ Settings that affect how the model thinks and responds.
 | Option | Default | Description |
 |--------|---------|-------------|
 | `keep_thinking` | `false` | Preserve Claude's thinking blocks across turns. **Warning:** enabling may degrade model stability. |
+| `claude_long_context_beta` | `false` | Experimental: attempt Claude 4.6 long-context beta header (provider entitlement required) |
+| `claude_long_context_beta_header` | `"context-1m-2025-08-07"` | Beta header value used when `claude_long_context_beta` is enabled |
 | `session_recovery` | `true` | Auto-recover from tool_result_missing errors |
 | `auto_resume` | `false` | Auto-send resume prompt after recovery |
 | `resume_text` | `"continue"` | Text to send when auto-resuming |
@@ -51,6 +53,24 @@ When `false` (default), thinking is stripped:
 - **Pros:** More stable model behavior, smaller context
 - **Cons:** Model may be less coherent, forgets previous reasoning
 
+### About `claude_long_context_beta`
+
+Claude models on the Antigravity path remain **200k context by default**.
+
+When `claude_long_context_beta` is enabled:
+- Claude 4.6 requests include the configured `anthropic-beta` token (`claude_long_context_beta_header`)
+- If the provider rejects that beta header, the plugin retries once automatically without it
+- The request falls back to the stable 200k path and logs the rejection reason
+
+Example:
+
+```json
+{
+  "claude_long_context_beta": true,
+  "claude_long_context_beta_header": "context-1m-2025-08-07"
+}
+```
+
 ---
 
 ## Account Rotation
@@ -168,6 +188,7 @@ These settings are `false` by default:
 | Setting | Default | What it does |
 |---------|---------|--------------|
 | `keep_thinking` | `false` | Preserve Claude thinking (may degrade stability) |
+| `claude_long_context_beta` | `false` | Opt-in experimental Claude 1M beta header attempt |
 | `auto_resume` | `false` | Auto-continue after recovery |
 
 ---

docs/MODEL-VARIANTS.md

@@ -104,12 +104,15 @@ Claude models use token-based thinking budgets:
 | `low` | 8192 | Light thinking |
 | `max` | 32768 | Maximum thinking |
 
+> **Context Limit Note:** Claude models on Antigravity are configured with a **200k base context**.
+> If you enable `claude_long_context_beta` in `antigravity.json`, the plugin can attempt an experimental long-context beta header and automatically falls back to 200k if rejected.
+
 ### Claude Example
 
 ```json
 {
   "antigravity-claude-opus-4-6-thinking": {
-    "name": "Claude Opus 4.6 Thinking (Antigravity)",
+    "name": "Claude Opus 4.6 Thinking (Antigravity, 200k base)",
     "limit": { "context": 200000, "output": 64000 },
     "modalities": { "input": ["text", "image", "pdf"], "output": ["text"] },
     "variants": {

script/build-schema.ts

@@ -41,6 +41,10 @@ const optionDescriptions: Record<string, string> = {
     "Enable tool hallucination prevention for Claude models. Injects parameter signatures and strict usage rules.",
   claude_prompt_auto_caching:
     "Enable Claude prompt auto-caching by adding top-level cache_control when absent.",
+  claude_long_context_beta:
+    "Enable experimental Claude long-context beta header injection for Claude 4.6 models. If rejected, requests auto-fallback to stable 200k behavior.",
+  claude_long_context_beta_header:
+    "Claude long-context beta header value. Override if provider beta token changes.",
   proactive_token_refresh:
     "Enable proactive background token refresh before expiry, ensuring requests never block.",
   proactive_refresh_buffer_seconds:

src/plugin.ts

@@ -30,6 +30,7 @@ import {
 import {
   buildThinkingWarmupBody,
   isGenerativeLanguageRequest,
+  isUnsupportedClaudeLongContextBetaError,
   prepareAntigravityRequest,
   transformAntigravityResponse,
 } from "./plugin/request";
@@ -50,6 +51,7 @@ import { checkAccountsQuota } from "./plugin/quota";
 import { initDiskSignatureCache } from "./plugin/cache";
 import { createProactiveRefreshQueue, type ProactiveRefreshQueue } from "./plugin/refresh-queue";
 import { initLogger, createLogger } from "./plugin/logger";
+import { scrubTextForLog } from "./plugin/logging-utils";
 import { initHealthTracker, getHealthTracker, initTokenTracker, getTokenTracker } from "./plugin/rotation";
 import { initAntigravityVersion } from "./plugin/version";
 import { executeSearch } from "./plugin/search";
@@ -67,6 +69,7 @@ const MAX_OAUTH_ACCOUNTS = 10;
 const MAX_WARMUP_SESSIONS = 1000;
 const MAX_WARMUP_RETRIES = 2;
 const CAPACITY_BACKOFF_TIERS_MS = [5000, 10000, 20000, 30000, 60000];
+const CLAUDE_LONG_CONTEXT_REJECTION_REASON_MAX_CHARS = 240;
 
 function getCapacityBackoffDelay(consecutiveFailures: number): number {
   const index = Math.min(consecutiveFailures, CAPACITY_BACKOFF_TIERS_MS.length - 1);
@@ -86,6 +89,8 @@ const log = createLogger("plugin");
 const rateLimitToastCooldowns = new Map<string, number>();
 const RATE_LIMIT_TOAST_COOLDOWN_MS = 5000;
 const MAX_TOAST_COOLDOWN_ENTRIES = 100;
+const CLAUDE_LONG_CONTEXT_FALLBACK_TOAST_SESSIONS = new Set<string>();
+const MAX_CLAUDE_LONG_CONTEXT_FALLBACK_TOAST_SESSIONS = 500;
 
 // Track if "all accounts blocked" toasts were shown to prevent spam in while loop
 let softQuotaToastShown = false;
@@ -117,6 +122,25 @@ function shouldShowRateLimitToast(message: string): boolean {
   return true;
 }
 
+function shouldShowClaudeLongContextFallbackToast(sessionKey: string): boolean {
+  if (CLAUDE_LONG_CONTEXT_FALLBACK_TOAST_SESSIONS.has(sessionKey)) {
+    return false;
+  }
+
+  CLAUDE_LONG_CONTEXT_FALLBACK_TOAST_SESSIONS.add(sessionKey);
+  if (
+    CLAUDE_LONG_CONTEXT_FALLBACK_TOAST_SESSIONS.size >
+    MAX_CLAUDE_LONG_CONTEXT_FALLBACK_TOAST_SESSIONS
+  ) {
+    const first = CLAUDE_LONG_CONTEXT_FALLBACK_TOAST_SESSIONS.values().next().value;
+    if (first !== undefined) {
+      CLAUDE_LONG_CONTEXT_FALLBACK_TOAST_SESSIONS.delete(first);
+    }
+  }
+
+  return true;
+}
+
 function resetAllAccountsBlockedToasts(): void {
   softQuotaToastShown = false;
   rateLimitToastShown = false;
@@ -1537,6 +1561,7 @@ export const createAntigravityPlugin = (providerId: string) => async (
 
           let lastFailure: FailureContext | null = null;
           let lastError: Error | null = null;
+          let disableClaudeLongContextBetaForRetry = false;
           const abortSignal = init?.signal ?? undefined;
 
           // Helper to check if request was aborted
@@ -2021,6 +2046,9 @@ export const createAntigravityPlugin = (providerId: string) => async (
                   {
                     claudeToolHardening: config.claude_tool_hardening,
                     claudePromptAutoCaching: config.claude_prompt_auto_caching,
+                    claudeLongContextBetaEnabled: config.claude_long_context_beta,
+                    claudeLongContextBetaHeader: config.claude_long_context_beta_header,
+                    disableClaudeLongContextBetaForRetry,
                     fingerprint: account.fingerprint,
                   },
                 );
@@ -2330,6 +2358,55 @@ export const createAntigravityPlugin = (providerId: string) => async (
                   }
                 }
 
+                const canBeClaudeLongContextRejection =
+                  response.status === 400
+                  || response.status === 403
+                  || response.status === 422;
+
+                if (
+                  prepared.claudeLongContextBetaApplied
+                  && !disableClaudeLongContextBetaForRetry
+                  && canBeClaudeLongContextRejection
+                ) {
+                  const errorBodyText = await response.clone().text().catch(() => "");
+                  if (
+                    isUnsupportedClaudeLongContextBetaError(
+                      response.status,
+                      errorBodyText,
+                      prepared.claudeLongContextBetaHeader,
+                    )
+                  ) {
+                    disableClaudeLongContextBetaForRetry = true;
+
+                    const sessionKey = prepared.sessionId
+                      ?? `${account.index}:${prepared.effectiveModel ?? "claude"}`;
+
+                    if (shouldShowClaudeLongContextFallbackToast(sessionKey)) {
+                      await showToast(
+                        "Claude long-context beta rejected by provider. Falling back to stable 200k path.",
+                        "warning",
+                      );
+                    }
+
+                    const reasonPreview = scrubTextForLog(
+                      errorBodyText,
+                      CLAUDE_LONG_CONTEXT_REJECTION_REASON_MAX_CHARS,
+                    );
+                    pushDebug(
+                      `claude-long-context-beta rejected status=${response.status} header=${prepared.claudeLongContextBetaHeader ?? "unknown"} reason=${reasonPreview}`,
+                    );
+                    log.debug("claude-long-context-beta-rejected", {
+                      status: response.status,
+                      model: prepared.effectiveModel,
+                      header: prepared.claudeLongContextBetaHeader,
+                      reasonPreview,
+                    });
+
+                    i -= 1;
+                    continue;
+                  }
+                }
+
                 const shouldRetryEndpoint = (
                   response.status === 403 ||
                   response.status === 404 ||

src/plugin/config/models.ts

@@ -68,12 +68,12 @@ export const OPENCODE_MODEL_DEFINITIONS: OpencodeModelDefinitions = {
     },
   },
   "antigravity-claude-sonnet-4-6": {
-    name: "Claude Sonnet 4.6 (Antigravity)",
+    name: "Claude Sonnet 4.6 (Antigravity, 200k base)",
     limit: { context: 200000, output: 64000 },
     modalities: DEFAULT_MODALITIES,
   },
   "antigravity-claude-opus-4-6-thinking": {
-    name: "Claude Opus 4.6 Thinking (Antigravity)",
+    name: "Claude Opus 4.6 Thinking (Antigravity, 200k base)",
     limit: { context: 200000, output: 64000 },
     modalities: DEFAULT_MODALITIES,
     variants: {

src/plugin/config/schema.test.ts

@@ -11,7 +11,12 @@ describe("cli_first config", () => {
   it("documents cli_first in the JSON schema", () => {
     const schemaPath = new URL("../../../assets/antigravity.schema.json", import.meta.url);
     const schema = JSON.parse(readFileSync(schemaPath, "utf8")) as {
-      properties?: Record<string, { type?: string; default?: unknown; description?: string }>;
+      properties?: Record<string, {
+        type?: string
+        default?: unknown
+        description?: string
+        minLength?: number
+      }>;
     };
 
     const cliFirst = schema.properties?.cli_first;
@@ -46,3 +51,43 @@ describe("claude_prompt_auto_caching config", () => {
     expect(claudePromptAutoCaching?.description?.length ?? 0).toBeGreaterThan(0);
   });
 });
+
+describe("claude_long_context_beta config", () => {
+  it("includes claude_long_context_beta defaults in DEFAULT_CONFIG", () => {
+    expect(DEFAULT_CONFIG).toHaveProperty("claude_long_context_beta", false);
+    expect(DEFAULT_CONFIG).toHaveProperty("claude_long_context_beta_header", "context-1m-2025-08-07");
+  });
+
+  it("documents claude_long_context_beta in the JSON schema", () => {
+    const schemaPath = new URL("../../../assets/antigravity.schema.json", import.meta.url);
+    const schema = JSON.parse(readFileSync(schemaPath, "utf8")) as {
+      properties?: Record<string, { type?: string; default?: unknown; description?: string }>;
+    };
+
+    const claudeLongContextBeta = schema.properties?.claude_long_context_beta;
+    expect(claudeLongContextBeta).toBeDefined();
+    expect(claudeLongContextBeta).toMatchObject({
+      type: "boolean",
+      default: false,
+    });
+    expect(typeof claudeLongContextBeta?.description).toBe("string");
+    expect(claudeLongContextBeta?.description?.length ?? 0).toBeGreaterThan(0);
+  });
+
+  it("documents claude_long_context_beta_header in the JSON schema", () => {
+    const schemaPath = new URL("../../../assets/antigravity.schema.json", import.meta.url);
+    const schema = JSON.parse(readFileSync(schemaPath, "utf8")) as {
+      properties?: Record<string, { type?: string; default?: unknown; description?: string }>;
+    };
+
+    const claudeLongContextBetaHeader = schema.properties?.claude_long_context_beta_header;
+    expect(claudeLongContextBetaHeader).toBeDefined();
+    expect(claudeLongContextBetaHeader).toMatchObject({
+      type: "string",
+      default: "context-1m-2025-08-07",
+      minLength: 1,
+    });
+    expect(typeof claudeLongContextBetaHeader?.description).toBe("string");
+    expect(claudeLongContextBetaHeader?.description?.length ?? 0).toBeGreaterThan(0);
+  });
+});

src/plugin/config/schema.ts

@@ -223,6 +223,26 @@ export const AntigravityConfigSchema = z.object({
    * @default false
    */
   claude_prompt_auto_caching: z.boolean().default(false),
+
+  /**
+   * Enable experimental Claude long-context beta header injection.
+   *
+   * When enabled, Claude 4.6 models attempt a provider beta header for 1M context.
+   * If the provider rejects the beta header, the request automatically falls back
+   * to the stable 200k path.
+   *
+   * @default false
+   */
+  claude_long_context_beta: z.boolean().default(false),
+
+  /**
+   * Claude long-context beta header value.
+   *
+   * Override this if provider beta token changes in the future.
+   *
+   * @default "context-1m-2025-08-07"
+   */
+  claude_long_context_beta_header: z.string().trim().min(1).default("context-1m-2025-08-07"),
 
   // =========================================================================
   // Proactive Token Refresh (ported from LLM-API-Key-Proxy)
@@ -460,6 +480,8 @@ export const DEFAULT_CONFIG: AntigravityConfig = {
   tool_id_recovery: true,
   claude_tool_hardening: true,
   claude_prompt_auto_caching: false,
+  claude_long_context_beta: false,
+  claude_long_context_beta_header: "context-1m-2025-08-07",
   proactive_token_refresh: true,
   proactive_refresh_buffer_seconds: 1800,
   proactive_refresh_check_interval_seconds: 300,