fix: agent_input_token_budget wrongly treated as a secret and unsettable from chat (#1294)

* fix: don't classify agent_input_token_budget as a secret (token must be a suffix)

* test: agent_input_token_budget is settable from chat

tests/test_manage_settings_token_budget.py

@@ -0,0 +1,22 @@
+"""Regression: agent_input_token_budget must be settable from chat (not flagged secret)."""
+import asyncio
+import json
+
+import src.settings as settings_mod
+from src.tool_implementations import do_manage_settings
+
+
+def test_set_token_budget_is_not_refused_as_secret(monkeypatch):
+    store = {}
+    monkeypatch.setattr(settings_mod, "load_settings", lambda: dict(store))
+    monkeypatch.setattr(settings_mod, "save_settings", lambda s: store.update(s))
+
+    result = asyncio.run(do_manage_settings(json.dumps({
+        "action": "set", "key": "agent_input_token_budget", "value": 8000,
+    })))
+
+    # The "token" substring used to flag this int setting as a credential and
+    # refuse to set it (even though there's a deliberate "token budget" alias).
+    assert "credential" not in result.get("response", "").lower(), result
+    assert result.get("exit_code") == 0, result
+    assert store.get("agent_input_token_budget") == 8000