fix(claude): only enable 1M context beta when model has [1m] suffix (xvirobotics#273)

Both executors unconditionally set:

  queryOptions.betas = ['context-1m-2025-08-07'];

The comment acknowledged this was a no-op on OAuth / Pro-Max (where
betas are ignored and the `[1m]` suffix is the actual lever), but on
API-key auth the SDK honors the beta header — so every API-key user
got 1M context for *every* model, including plain `opus-4-7`.

Visible symptom: `/model claude-opus-4-7` showed a 1000k context
window in the card footer (and billed at 2× the rate) instead of the
expected 200k. The user only intended 1M on the explicit
`claude-opus-4-7[1m]` form.

Fix: only emit the beta flag when the model name contains `[1m]`.
That makes the suffix the single canonical signal for both auth
paths — SDK parses it directly, and we re-send the matching beta as
belt-and-braces for any auth mode that needs the explicit header.

Both `executor.ts` and `persistent-executor.ts` had the same
unconditional line; both fixed. Comment notes they must stay in
sync.

No new tests — the SDK's beta dispatch is mocked away in vitest, and
the change is one conditional. All 291 existing tests still pass.

Co-authored-by: Flood Sung <floodsung@xvirobotics.ai>
Co-authored-by: Claude Opus 4.7 <noreply@anthropic.com>

Author: 3 people

src/engines/claude/executor.ts

@@ -365,9 +365,20 @@ export class ClaudeExecutor {
       queryOptions.resume = sessionId;
     }
 
-    // Beta flags are ignored by the SDK on OAuth/Pro-Max auth. For 1M context,
-    // use the model-name suffix `[1m]` (e.g. `claude-opus-4-7[1m]`) instead.
-    queryOptions.betas = ['context-1m-2025-08-07'];
+    // 1M context window is opt-in via the `[1m]` model-name suffix
+    // (`claude-opus-4-7[1m]`). The suffix is the canonical signal — the SDK
+    // parses it directly on both OAuth and API-key auth paths.
+    //
+    // We *also* set the matching `betas` flag when the suffix is present,
+    // belt-and-braces: harmless when the SDK already inferred it, and a
+    // safety net if a future SDK rev only honors the explicit beta header
+    // for some auth modes. Without the suffix, leave `betas` unset —
+    // setting it unconditionally on API-key auth had the side-effect of
+    // forcing every model (e.g. plain `opus-4-7`) into 1M context at 2×
+    // the price, which the user never asked for.
+    if (this.config.claude.model?.includes('[1m]')) {
+      queryOptions.betas = ['context-1m-2025-08-07'];
+    }
 
     return queryOptions;
   }

src/engines/claude/persistent-executor.ts

@@ -365,8 +365,21 @@ export class PersistentClaudeExecutor extends EventEmitter {
         append: '\n\n' + appendSections.join('\n\n'),
       };
     }
-    // Beta flags (parity with legacy executor)
-    queryOptions.betas = ['context-1m-2025-08-07'];
+    // 1M context window is opt-in via the `[1m]` model-name suffix
+    // (`claude-opus-4-7[1m]`). The suffix is the canonical signal — the SDK
+    // parses it directly on both OAuth and API-key auth paths.
+    //
+    // We *also* set the matching `betas` flag when the suffix is present,
+    // belt-and-braces: harmless when the SDK already inferred it, and a
+    // safety net if a future SDK rev only honors the explicit beta header
+    // for some auth modes. Without the suffix, leave `betas` unset —
+    // setting it unconditionally on API-key auth had the side-effect of
+    // forcing every model (e.g. plain `opus-4-7`) into 1M context at 2×
+    // the price, which the user never asked for. (Parity with legacy
+    // executor; both spots must stay in sync.)
+    if (this.options.model?.includes('[1m]')) {
+      queryOptions.betas = ['context-1m-2025-08-07'];
+    }
 
     // Hooks: AskUserQuestion (mirrored from legacy executor — required so
     // that questions can be answered by users via Feishu cards) + Agent